Early warning model construction of intelligent healthcare medical equipment based on deep learning—Taking medical infrared imager as an example

Abstract

To realize the fault early warning function of healthcare medical equipment, this study constructs an equipment fault early warning model and combines particle swarm optimization and long-short term memory network to test the performance. The study obtains the optimal value of data feature vectors through particle swarm optimization algorithm and uses long-short term memory prediction model to predict and classify feature signals. In addition, the study uses the binning method to denoise the collected data and normalizes the denoised data so that each feature data was distributed between 0 and 1. The results showed that the fitting between actual values and predicted values was good. The maximum values of Precision, Recall, and F1 of the designed warning model were 97.98%, 97.82%, and 97.68%, respectively, which were significantly better than the control model. This indicated that the warning model designed by the research had good performance. The combination of the particle swarm optimization algorithm and the long short-term memory network model offered unique advantages in the medical field. The particle swarm optimization algorithm could efficiently identify key features, avoid local optima, and improve the model’s generalization ability. Long short-term memory networks could accurately capture the dynamic trends of faults and adapt to the temporal nature of medical data. Combining the two could meet the high-precision, real-time, and adaptive requirements of medical equipment fault warnings, effectively improving their accuracy.

Keywords

particle swarm optimization long-short term memory convolutional neural network failure prediction normalization infrared imager

Introduction

Due to the change of the concept of public pension, many countries have also begun to advocate the development of healthcare and medical undertakings. The healthcare equipment has been continuously improved, and the degree of scientific and technological development has been continuously deepened.¹ The proper functioning of medical equipment is essential to ensuring patient safety and effective treatment. Accurately and promptly predicting equipment failures can prevent medical accidents, reduce costs, and improve the quality and efficiency of medical services. Therefore, exploring a high-precision, highly adaptable, real-time method for warning of medical equipment faults is particularly important. The application of computers in the medical field is becoming more and more widespread, and daily maintenance and fault early warning of healthcare medical equipment are also important. The commonly used monitoring methods in the world include time-frequency analysis, fuzzy recognition, signal analysis, Wavelet Transform (WT), and neural networks.² The time-frequency analysis method has the advantage of clearly displaying the time-frequency characteristics of a signal. However, it has the disadvantage of being computationally complex. The method’s correlation with fault prediction is reflected in its ability to detect changes in the time-frequency characteristics of a signal when a fault is detected in the equipment. The advantage of the fuzzy recognition method is that it can effectively handle problems of ambiguity and uncertainty. However, the results depend on the rationality, accuracy, and relevance of the fuzzy rules. The method’s relevance to fault prediction lies in its ability to identify early fuzzy fault characteristics. The advantage of the signal analysis method is that it can obtain information directly from equipment operation signals. However, it is susceptible to noise interference. Its relevance to fault prediction lies in its ability to detect fault characteristics in a timely manner. The WT method is sensitive to sudden changes and singular points in a signal. However, it is computationally complex. The method is relevant to fault prediction because it can capture sudden changes and singular points in a signal caused by equipment faults. The advantage of the neural network method is that it can automatically extract complex fault features. However, it requires a large amount of sample data for training and takes a long time to train. Its relevance to fault prediction stems from its ability to learn the relationship between equipment operating status and faults, thereby enabling fault prediction and diagnosis. With the wide application of neural network, more and more experts and scholars are committed to introducing neural network into equipment fault early warning models and have made many breakthroughs.³ Among them, the performance of Convolutional Neural Network (CNN) and Support Vector Machine (SVM) is better, which has been unanimously recognized by the academic community.⁴ However, most of the current researches tend to identify the fault type after the fault occurs, and the relevant research on fault early warning is relatively few.⁵ The timely detection of potential medical equipment malfunctions is contingent upon the normal functioning of the equipment itself and the accurate assessment of the patient’s condition. Therefore, it is imperative to implement effective mechanisms for the early identification of such malfunctions.

At present, research on medical equipment fault warning mainly focuses on three aspects: (1) Combining microcontroller modules and software. (2) Utilizing rough neural networks. (3) Using machine learning algorithms. However, the microcontroller module cannot detect faults in advance or provide an early warning. Rough neural networks are prone to overfitting and have a longer learning time. In addition, the feature extraction of traditional machine learning algorithms mainly relies on human intervention, and the adaptability and learning ability of the algorithms are poor. Deep learning algorithms represent an advancement over traditional machine learning algorithms, exhibiting enhanced learning capacity and adaptability. Nevertheless, the field of medical equipment fault warning remains relatively under-researched. The majority of existing studies have focused on hardware devices and traditional neural networks, with relatively limited application of deep learning and other technologies. This has resulted in a notable decline in warning accuracy.

With the rapid development of artificial intelligence and big data technology, deep learning algorithms have shown strong application potential in various fields. In the field of medical equipment fault warning, deep learning algorithms have gradually received attention. Particle Swarm Optimization (PSO) and Long Short-Term Memory (LSTM) are swarm intelligence optimization algorithms and special recurrent neural networks (RNNs), respectively. They have unique advantages in feature optimization and time series processing. PSO is a swarm intelligence optimization algorithm, and its application in medical equipment fault diagnosis mainly includes four aspects: optimizing feature selection, optimizing model parameters, optimizing neural network weights, and improving the robustness of diagnostic models.⁶ The PSO algorithm simulates the foraging behavior of bird flocks. It initializes a group of random particles and updates their velocity and position based on individual and global extremums. This process approaches an optimal solution. The PSO algorithm has the advantages of strong applicability, fast convergence, and good parallelism, and it can effectively avoid local optima. Additionally, an LSTM is a special type of RNN that uses gating mechanisms (input, forget, and output gates) to selectively transmit information, effectively solving the gradient vanishing and exploding problems of traditional RNNs. It can effectively store information for a long time and reduce feature loss. It is particularly suitable for processing the fault diagnosis of medical equipment that relies on long-term sequence data.⁷

Therefore, in order to fill the gap of deep learning technology in the field of medical equipment fault warning and improve the accuracy of warning, the research combines PSO and LSTM to build fault diagnosis models of healthcare medical equipment. It tests the performance of a medical infrared imager as an example. Currently, most methods that combine PSO and LSTM use PSO to optimize LSTM hyperparameters, such as the learning rate and number of neurons, to improve model performance. However, the innovative combination approach presented in this article differs significantly from these methods. The study uses the PSO algorithm to optimize and screen the input features, extracting the most valuable feature vectors for fault warning. These vectors are then input into an LSTM network for deep learning, achieving accurate, early detection of medical equipment faults. The novelty of the research lies in its deep integration of the PSO feature optimization ability and the LSTM time series processing advantages. This improves feature quality by intelligently screening the PSO algorithm and reducing the LSTM network’s input dimension. The research effectively integrates the feature optimization capability of PSO and the time series processing capability of LSTM. This integration achieves fault warnings for medical infrared imagers, improves warning accuracy, reduces missed and false alarms, and prevents medical accidents caused by delayed detection of device faults. At the same time, it also provides a new technical approach and method in the field of medical equipment fault warning and expands the application of deep learning in medical equipment fault alarm. This is of great significance in promoting intelligent maintenance and management of medical equipment.

References 1 and 2 provide background knowledge on the development of medical equipment and fault detection methods. References 3 and 4 demonstrate the application of neural networks in device fault warning, providing a foundation for the deep learning application in this study. Reference 5 points out that current research is mostly focused on type recognition after the occurrence of faults, highlighting the innovation of this study in early warning of faults. References 6 and 7, respectively, demonstrate the advantages of the PSO algorithm and the LSTM network in fault diagnosis. These references provide the theoretical basis and technical support for combining the two algorithms in this study.

The research is divided into four sections. The first section summarizes the current status of research on the diagnosis and analysis of faults in medical equipment, and introduces the research methods. The second section introduces the identification and preprocessing of fault signals, extracts the optimal value of data using PSO algorithm, and establishes an LSTM fault prediction model. The third section takes the medical infrared thermal imaging device as an example to simulate and test the LSTM fault prediction model, and analyze the experimental data. The fourth section is the conclusion section of the study. It includes a summary and generalization of the experimental results and highlights the limitations and future prospects of the research. It also analyzes the application of advanced optimization algorithms in other fields.

Literature review

At present, the healthcare industry in developed countries has developed very mature. The maintenance and fault detection of various medical equipment has become a hotpot. Experts and scholars have conducted a lot of research on intelligent fault early warning of medical equipment and achieved some results. These achievements have been described according to anti-noise models, hybrid models, time series methods, optimization models, and other methods.

On the anti-noise model, Jin et al. proposed a new Anti-noise Multi-scale CNN (AM-CNN) to solve composite fault diagnosis problems under different noise intensities. According to the principle of noise superposition, the residual loss was expressed and the loss function was constructed. Experimental data showed that AM-CNN accuracy was improved by 39.93% under noiseless working conditions.⁸ Wang et al. proposed a model by granular computing. CNN with super parameter optimization was used for the fault features, which were used as granularity input by granularity calculation to determine the fault type. The experimental results verified its effectiveness.⁹ However, although the anti-noise model shows improvement in anti-noise performance and diagnostic accuracy, it lacks the ability to effectively recognize weak features in the early stages of faults.

On the mixed model, Huo et al. used Fuzzy C-Means Clustering (FCMC) to classify the fault detection. By comparing with CNN method for mechanical equipment fault detection, the experimental results confirmed its superiority.¹⁰ Han et al. used the short-term wavelet entropy calculation method to extract fault information. LSTM output was set as SVM input to obtain fault diagnosis results by adaptive classification. Experimental results verified the effectiveness of the algorithm.¹¹ Sheela et al. designed an artificial intelligence algorithm based on hybrid PSO-SVM for the diagnosis of COVID-19, which was used to analyze and judge computed tomography images. The results showed that the proposed work showed a specificity of 0.85, a sensitivity of 0.956, and an accuracy of 95.78%.¹² Amooei et al. designed two new models based on CNN-LSTM networks for the diagnosis of neurodegenerative diseases. This study used the first CNN-LSTM model to process spectrogram images and combined the second CNN-LSTM model with WT to perform secondary processing on spectrogram images. The results demonstrated that the classification accuracy achieved using only approximate sub-bands was 95.37%, while that attained using three sub-bands was 94.04%, and the accuracy obtained when all sub-bands were included was 94.53%.¹³

In the time series method, Liu et al. introduced an intelligent method by LSTM network. Without the circuit diagram and the unknown signal direction of the circuit board, the symptoms and port electrical signals were collected. Moreover, the fault diagnosis, classification, and identification experiments were carried out by using the fused and screened multimodal features. The final experimental results verified its effectiveness.¹⁴ Hou et al. proposed a method by blind deconvolution for impact response enhancement and chirp Z-transform for characteristics matches. The experimental results confirmed the effectiveness of the algorithm.¹⁵ Hybrid models and time series methods have certain dynamic analysis and fault diagnosis capabilities. However, there is still room for improvement in the timeliness and accuracy of fault warnings. Time series methods can capture the evolution process of faults, but they are insufficiently sensitive to weak features in the early stages of faults.

In optimizing the model, Yang et al. used Genetic Algorithm (GA) and SVM for data parameter optimization, preference weight adjustment of the two indicators in the optimization direction, and the best parameter selection after iteration. Results showed good accuracy.¹⁶ To overcome that it was difficult to collect analog signals for fault detection, Su et al. proposed a fault diagnosis method and optimized fault detection model through Gray Wolf Optimization (GWO) algorithm. Results showed that the fault diagnosis time of the research algorithm was shortened by about 90%.¹⁷ Optimization models often focus on improving diagnostic efficiency and performance, rather than fault warning functions.

On other models, Wang et al. proposed a new CNN for equipment fault classification. Through convolution and subsampling operations, the learning process was iterated, and data features were automatically obtained. Finally, t-distribution random neighborhood embedding technology could visualize learned features. Experimental data proved the superiority of this method.¹⁸

In summary, many experts and scholars have proposed effective methods for diagnosing equipment faults. However, the focus has mainly been on diagnosing the type of fault after it occurs. Research on early warning systems for faulty equipment before a fault occurs is relatively limited. It is on this basis that the research takes the medical infrared imager as an example and combines PSO and LSTM to build the prediction model of faulty equipment so as to realize normalized monitoring and early warning of medical and health equipment. The novelty of the research lies in the combination of PSO algorithm and LSTM. PSO algorithm is used to obtain the optimal values of data feature vectors, and LSTM prediction model is used to predict and classify feature signals. To compare this study with other literature, the study compares the methods, results, and contributions of different literature and this study. The comparison results are shown in Table 1.

Table 1.

Comparison of methods, results, and contributions between different literature and this study.

Serial number	Method	Result	Contribution
8	AM-CNN	Accuracy was improved by 39.93% under noiseless working conditions	Improved the accuracy of fault diagnosis
9	CNN-GC	The accuracy rate is 99.8%	Implemented the transformation from numerical space to granularity space
10	FCMC	Has good fault detection performance and practicality	Improved fault detection efficiency
11	Combining LSTM and SVM based on short-time wavelet entropy	Has better robustness, adaptability, and accuracy	Reduce the number of electrical signal samples and achieve diagnosis of multiple types of faults
12	PSO-SVM	The accuracy rate is 95.78%	Save doctors approximately 50% of their time
13	Two new models based on CNN-LSTM network	The classification result is 94.53%	Promoted early diagnosis of neurodegenerative diseases
14	Method based on LSTM network	The average accuracy of classification diagnosis reaches 0.9709	Fusion and screening of multimodal features
15	Method based on optimizing maximum second-order cyclostationarity blind deconvolution and linear frequency modulation Z-transform	Can effectively detect wheel bearing faults	Realized the diagnosis of early composite fault defects in wheelset bearings
16	GA-SVM	The accuracy rate is 98.53%	Implemented optimization of dissolved oxygen parameters
17	Based on deep belief network and GWO algorithm	The accuracy has been improved to 99.68%	Optimized SVM parameters
18	CNN based on minimum entropy deconvolution	Improved the performance of CNN in fault diagnosis	Applicable to problems with unknown fault mechanisms
This study	PSO-LSTM	Accuracy, recall rate, and F1 reached 97.98%, 97.82%, and 97.68%, respectively	Improve the accuracy of early warning and reduce missed or false alarms

As shown in Table 1, although numerous studies have been conducted in the past, most of them have focused on diagnosing types after faults occur, with little research on providing early warnings before faults occur. This study combines PSO and LSTM to construct a predictive model for early warning of faults in medical infrared thermal imagers, successfully filling this gap. At the same time, building on the strengths of previous research, the study leverages the PSO algorithm’s feature optimization capabilities and the LSTM time series’ processing advantages to advance medical equipment fault warning technology.

Method description

In order to carry out real-time and accurate fault early warning analysis on healthcare medical equipment, two methods of sensor and Universal Serial Bus (USB) interface are used to collect the original data, and then the data are preprocessed by noise reduction and normalization. Finally, the optimal value of the data feature vector is extracted and substituted into the built prediction model for identification.

Fault signal identification and preprocessing

Once the concept of health physiotherapy is put forward, it has been vigorously promoted and developed worldwide. A variety of intelligent medical equipment also emerged, covering all aspects of the field of healthcare, of which the most widely used medical equipment is the medical infrared imager. It uses infrared thermal imaging technology combined with optical components and imaging circuit components to detect the infrared radiation of patients. Moreover, it visually diagnoses the temperature distribution of patients through signal processing, photoelectric conversion, and other steps so as to provide a diagnostic basis for the treatment of the disease.¹⁹ Taking the medical infrared thermal imager as an example, this study used deep learning neural networks to build an intelligent healthcare equipment fault early warning model and perform performance tests. First, the study aims to optimize the data feature vector through PSO algorithm and use the optimized data as high-quality input for subsequent fault warning models. Second, a fault warning model based on LSTM is designed, which utilizes the advantages of LSTM in time series data processing to predict and classify the characteristic signals of intelligent healthcare equipment. The flowchart of building the fault warning model is shown in Figure 1.

Figure 1.

Flowchart for building a fault warning model.

As shown in Figure 1, the whole fault early warning model is divided into four steps, namely, signal acquisition, signal preprocessing, prediction of fault time sequence characteristics, and signal classification diagnosis. First, the fault feature points are determined by collecting the fault information of the equipment. Then, aiming at noise problems and data chaos in original data, the preprocessing of noise reduction and normalization operations are carried out. PSO can extract optimal solution from feature data. Processed data are divided into training set and test set, and the time series analysis and prediction are carried out through LSTM. Therefore, it can complete the identification and classification of fault signals and achieve the role of medical equipment failure early warning.

Considering the complex working environment of medical equipment, the causes of equipment failure often come from the comprehensive influence of multiple factors. Therefore, to ensure multi-directional early warning of equipment failure, this experiment studies the use of sensors to sample and extract original equipment data, and analyzes the historical failure conditions to identify and collect the failure related indicators. Next, the raw data collected by various sensors are preprocessed.

In Figure 2, the image preprocessing part has three parts: signal denoising, normalization, and extraction of optimal feature data. The research adopts the box method for noise reduction, and the noise signal is generally divided into two types, namely, data loss and data mutation.²⁰ For missing data, the mean value of its adjacent data can be used to make up. For data mutation, it is necessary to locate the mutation data point first. The specific judgment formula (1) is as follows.²¹

{\begin{cases} \frac{Δ y_{i}}{Δ x_{i}} > 10 \frac{Δ y_{i - 1}}{Δ x_{i - 1}} \\ \frac{Δ y_{i}}{Δ x_{i}} < 10 \frac{Δ y_{i + 1}}{Δ x_{i + 1}} \end{cases} i = 2, 3, 4, . . ., n

(1)

Figure 2.

Data preprocessing flowchart.

In formula (1), $Δ x$ and $Δ y$ are the points change amount and $i$ is $i$ th point. If change at the point $i$ satisfies formula (1) at the same time, it is determined that the point is sudden change data, and the smoothing method is used to correct the adjacent data. The original data after noise reduction will be further unified. Therefore, each characteristic data is distributed between 0 and 1, so as to improve convergence speed. The specific calculation formula is shown in formula (2).

x * = \frac{x - x_{\min}}{x_{\max} - x_{\min}}

(2)

In formula (2), $x$ is the initial value, $x_{\max}$ and $x_{\min}$ represent the peak value and valley value of data, respectively. The peak and valley values of the data are within a reasonable range. Considering that in the multi-feature data of medical equipment, the proportion of fault feature data and non-fault feature data is seriously misaligned. The feature data contains a vast amount of redundant information due to the fault data constituting only a small portion. To enhance the accuracy of the fault equipment prediction model, the PSO algorithm is employed for extracting the optimal solutions from the feature data. The advantages of the PSO algorithm are strong applicability, fast convergence speed, good parallelism, strong robustness, low sensitivity to population size, few parameters, and less susceptibility to deadlock caused by local optima. Figure 3 illustrates the schematic diagram of the PSO algorithm.

Figure 3.

Schematic diagram of PSO algorithm.

In Figure 3, the PSO algorithm follows a series of steps. It begins by initializing the original data and determining the size of the particle swarm and the maximum number of algorithm iterations. Each particle is assigned randomly generated speed and position attributes. The individual extreme values are then calculated, and the global optimal solution is obtained. Subsequently, the particle attributes are updated using the speed and position attributes to identify the global optimal solution position for each particle. Formula (3) represents the update calculation formula for the particle speed.

v_{i} = ω \times v_{i} + c_{1} \times r_{1} \times (p B e s t_{i} - x_{i}) + c_{2} \times r_{2} \times (g B e s t_{i} - x_{i})

(3)

In formula (3), $v_{i}$ and $x_{i}$ are $i$ ’s velocity and position. $ω$ is the inertia weight of non-negative numbers. $c_{1}$ and $c_{2}$ represents different learning factors. $r_{1}$ and $r_{2}$ are the random functions between intervals [0, 1]. $p B e s t_{i}$ represents the particle optimal solution and $g B e s t_{i}$ represents the global optimal solution. Through further calculation of formula (3), the updated calculation formula of particle position is shown in formula (4).

x_{i} = x_{i} + v_{i}

(4)

Next, the inertia weight $ω$ is dynamically adjusted by the linear decreasing weight strategy. The formula for calculating the inertia weight $ω$ is as shown in formula (5).

ω_{(i)} = \frac{(ω_{i n i} - ω_{e n d}) (I_{m a x} - i)}{I_{m a x}} + ω_{e n d}

(5)

In formula (5), $ω_{i n i}$ is the initial $ω$ value, $ω_{e n d}$ is the last generation inertia weight, and $I_{m a x}$ is maximum iterations. Next, the learning factor is optimized to have ability of dynamic change in formula (6).

{\begin{cases} c_{1} = 2 - \frac{i}{I_{m a x}} \times 0.5 \\ c_{2} = 2 - \frac{I_{m a x} - i}{I_{m a x}} \times 0.5 \end{cases}

(6)

Through the dynamic change of formula (6), the particles tend to the particle optimal solution in the early iteration process $c_{1} > c_{2}$ , and the particles tend to the global optimal solution in the later iteration process $c_{1} < c_{2}$ . By optimizing the weights and learning factors, the final expression of the individual extreme value function is shown in formula (7).

f = (x_{n} - x_{1}) \cdot \int_{l}^{n} \frac{Δ y}{y} d x

(7)

In formula (7), $x_{1}$ and $x_{n}$ are initial and end values of selected feature set timing signal. $Δ y$ is the variation of adjacent timing signals and $y$ is the eigenvalue of the corresponding timing signal. The flowchart of PSO algorithm is shown in Figure 4.

Figure 4.

Flowchart of PSO algorithm.

In Figure 4, the process of the PSO algorithm is mainly divided into four steps. The first step is to randomly initialize the particle population. The second step is to calculate the individual extremum of each particle and obtain the global optimal solution. The third step is to update the velocity and position attributes of each particle. The fourth step is to ascertain whether the termination condition has been satisfied. If the optimal solution is deemed satisfactory, it should be outputted and the process concluded. Otherwise, the second step should be revisited.

Medical infrared imager equipment failure warning model construction

This study uses PSO and LSTM to build an equipment fault prediction model, obtains the optimal value of the original data characteristic signal through the PSO algorithm, and uses the LSTM prediction model to complete the prediction and classification of the characteristic signal to achieve the medical equipment fault early warning function. The study demonstrates the optimization of feature signal values using the PSO algorithm. Now, the study will construct a fault warning model based on LSTM. LSTM uses gate design to store effective information for a long time, reducing feature loss and avoiding gradient disappearance and explosion. It performs better in predictive analysis that relies on long-term sequence data. The working principle of the equipment fault early warning model by LSTM is shown in Figure 5.

In Figure 5, the whole prediction has two parts, the PSO algorithm and the LSTM algorithm. For the selection of hyperparameters in PSO algorithm, the study will compare the fitness values of different combinations of hyperparameters through experiments. In addition, for the hyperparameters of LSTM, the study selects the softsign activation function and solves the training loss function through matrix multiplication. The LSTM algorithm is evolved from RNN. As a common deep learning network, the RNN not only has memory but also can share parameters.²² It can recognize and analyze past memories and process and transmit information in the form of a chain link. This makes it very suitable for processing nonlinear data features based on sequences. RNN has input, hidden, and output layers.

Figure 5.

Working principal diagram of LSTM warning model.

As shown in Figure 6, $X$ represents input data, $O$ represents output data, $S$ represents neuron memory information, and $U$ , $V$ , and $W$ are weight values of input samples, output, and input. $t$ is a time series, respectively. The expression of memory information $S_{t}$ is shown in formula (8).

S_{t} = f (U x_{t} + W S_{t - 1})

(8)

Figure 6.

RNN network structure expansion diagram.

In formula (8), $f$ is the activation function and $S_{t - 1}$ is the memory information for the previous node. The output $O_{t}$ at the moment $t$ can be obtained by outputting the memory information $S_{t}$ calculated in formula (8) with weights $V$ to the next neuron. The specific calculation formula is as follows (9).

O_{t} = g (V S_{t})

(9)

In formula (8), $g$ is the activation function. The commonly used activation functions include $Tanh$ , $R e l u$ , $S o f t \max$ , and $S i g m o i d$ . Here is the function $S o f t \max$ . However, in the long-term training of RNN network, the relevant weights are easy to accumulate errors with time, and eventually lead to gradient burst and gradient disappearance.²³ To solve this problem, Hochreiter and Schmidhube proposed the LSTM algorithm in 1997. Therefore, it can filter the data based on RNN and freely select the cell information through the gating mechanism. Figure 7 shows LSTM cell structure.

Figure 7.

Schematic diagram of LSTM cell structure.

As shown in Figure 7, $x_{t}$ is the input vector, $h_{t}$ represents output vector, and $C_{t}$ is the state information of the memory cell. $σ$ is the $S i g m o i d$ activation function. $f_{t}$ represents the forget gate, $O_{t}$ represents the output information of the output gate, $Tanh$ is the hyperbolic tangent function, and $i_{t}$ is the input gate. In memory cells, the screening function of the gate structure is realized through the $S i g m o i d$ activation function, and output result of $S i g m o i d$ activation function is 0 or 1, where 0 means not to pass and 1 means to pass.²⁴ The three gates in LSTM cells are forgetting, input, and output gate. The expression of forgetting gate $f$ is shown in formula (10).

f_{t} = σ (w_{f} \cdot [h_{t - 1}, x_{t}] + b_{f})

(10)

As shown in formula (10), the forgetting gate determines the number of cells retained in the input vector. $[h_{t - 1}, x_{t}]$ indicates the information connecting hidden layer and input, $h_{t - 1}$ indicates hidden layer state, and $b_{f}$ is the offset vector of forgetting gate. The information retained in cell unit is updated through input gate. The expression of the input door $i$ is shown in formula (11).

i_{t} = σ (w_{i} \cdot [h_{t - 1}, x_{t}] + b_{i})

(11)

In formula (11), $b_{i}$ represents offset vector of input gate. ${\tilde{C}}_{t}$ is the input gate and obtains candidate cell information by filtering the input information. The calculation formula of candidate cell information ${\tilde{C}}_{t}$ is shown in formula (12).

{\tilde{C}}_{t} = \tanh (W_{c} \cdot [h_{t - 1}, x_{t}] + b_{c})

(12)

Formula (12) $W_{c}$ represents the cell state matrix, and $b_{c}$ is the cell state offset vector. The candidate cell information ${\tilde{C}}_{t}$ is updated by calculating the $Tanh$ activation function to obtain $C_{t}$ , and its calculation formula is shown in formula (13).

C_{t} = f_{t} * C_{t - 1} + i_{t} * {\tilde{C}}_{t}

(13)

In formula (13), $C_{t - 1}$ represents the cell information before update. The updated cell information $C_{t}$ is transmitted to output gate through $Tanh$ , and the expression of the output gate $o$ is as shown in formula (14).

o_{t} = σ (w_{o} \cdot [h_{t - 1}, x_{t}] + b_{o})

(14)

In formula (14), $b_{o}$ is the offset vector of the output gate. After three gated screening, the cell state characteristics are calculated by $Tanh$ layer to obtain a vector of [−1, 1], and then the output $h_{t}$ of the whole cycle unit is obtained.²⁵ Formula (15) is the output $h_{t}$ .

h_{t} = o_{t} * \tanh (C_{t})

(15)

Numerical experiments

Taking the medical infrared thermal imager equipment of a hospital as an example, the performance test of the tmt-9000b equipment fault early warning model is carried out. The main parameters include the working band of 8–14 um, the input voltage range of 5–12 v, the imaging distance of 0.5–3 m, the ambient temperature of 10–30°C, the temperature measurement range of 30–42°C, and the temperature measurement accuracy of <0.4°C.

To validate the performance of the fault warning model designed in the study, the data required for performance validation is collected first. When collecting raw data, the study follows the collection method of Minghui et al.²⁶ The research adopts two ways of sensor and USB interface to collect original signal data, and collects raw data on the operating status of the device from June 2022 to December 2022. Finally, 26357 data are obtained. The cross-validation method used in the experiment is simple cross-validation, which directly divides the dataset into a training set and a validation set. The most commonly utilized partitioning ratios for datasets include 8:1:1, 7:3, 6:2:2, 98:1:1, and so forth. The specific ratio employed is contingent upon the size of the dataset and the requirements of the machine learning project. When the dataset size is small, a ratio of 7:3 or 8:2 is usually used to divide it into training and testing sets. Due to the fact that the raw data used in the experiment is collected by oneself and the quantity is relatively small, the original data is split into a training set and a test set in an 8:2 ratio. Specifically, the training set consists of 21,085 data points, while the test set contains 5,272 data points. In addition to the previously mentioned TMT-9000B medical thermal imaging camera device, temperature sensors, voltage sensors, and current sensors are used in the acquisition of data. These sensors are installed in key parts of the medical infrared camera and are used to monitor the operational status of the device in real time. Additionally, to obtain historical fault labels, the study examines equipment maintenance records containing information such as the time and type of equipment failure and the repair measures taken. The study also invites domain experts to annotate these data. Among the 21085 data collected in the study, the data types are all medical infrared thermal imaging equipment operation status data. Each sample contains 16 key parameters such as temperature, voltage, current, imaging distance, and ambient temperature during equipment operation, with a data size of approximately 2.5 GB.

Next, noise reduction and normalization are performed on the original data, and the comparison between the original signal and noise reduction waveform is shown in Figure 8. Figure 8(a) shows the waveform of the original signal, while Figure 8(b) shows the waveform after denoising. As shown in Figure 8, in the original signal, the waveform vibration amplitude is relatively large. For example, the average waveform vibration amplitude interval^1,9 in the first 4 seconds of the original signal is calculated to be 5.46. Whereas, the average waveform vibration amplitude interval in the first 4 seconds of the denoised waveform is [4, 6], and the average waveform is 5.23. In Figure 8, the trend of the waveform and the positions of inflection points are roughly the same before and after noise reduction. Therefore, it can be concluded that the study of noise reduction methods can reduce the interference of noise on the premise of ensuring original data characteristics. Additionally, combining amplitude and data calculations reveals that the original signal’s signal-to-noise ratio is 28.5 dB, with an information entropy of 0.92. After denoising, however, the signal-to-noise ratio increases to 41.1 dB, while the information entropy decreases to 0.70. This also indicates that denoising effectively suppresses noise and enhances signal regularity.

Figure 8.

Comparison between original signal and denoised waveform.

Next, the study normalizes the original data after noise reduction according to a certain proportion to make its wave value range within the [0, 1] range. Taking the measured temperature as an example, the normalized results of the measured temperature characteristic data are shown in Figure 9. As shown in Figure 9, the temperature measured in the previous 10 seconds is taken as an example. In Figure 9(a), excluding the time of the 0th second, the Maximum and Minimum Values (MMV) of the original data in the first 10 seconds are 40.3 and 34.2, respectively. The results after normalization are shown in Figure 9(b). At this time, the data MMV in the first 10 seconds is 1.00 and 0.29, respectively. Thus, the results are distributed in the range of [0, 1], which greatly reduces calculation and improves operation efficiency on the premise of ensuring operation accuracy.

Figure 9.

Comparison of normalized results of temperature characteristic data.

To obtain an optimal solution of data and reduce data redundancy, the PSO algorithm is trained according to the initial learning rate, time window step, training times, and dropout, and the fitness value of the chromosome is calculated. The fitness value is a comprehensive index based on the Root Mean Square Error (RMSE) and the convergence speed of the validation set. It can comprehensively reflect the prediction accuracy and training efficiency of the model. Generally speaking, the higher the fitness value, the higher the prediction accuracy of the model and the faster the convergence speed. When setting the range of PSO hyperparameters, the study refers to previous research.²⁷ The final results of each parameter are shown below. In Table 2, the initial learning rate of the optimal solution is mainly concentrated around 0.010. The time window step size is better in the range of 8–12, and the training times are concentrated in the range of 210–230. The dropout parameter value is between 0.25 and 0.35. Through the comprehensive comparison of various parameters, the chromosome with 0.985 fitness value is more suitable for constructing the LSTM prediction model. The initial learning rate used in the final study is 0.01, the time window step is 8, training times is 225, and the dropout value is 0.29.

Table 2.

PSO parameter results.

Initial learning rate	Time window step size	Training frequency	Drop-out	Fitness value
0.005	12	210	0.25	0.962
0.005	16	213	0.32	0.957
0.010	8	225	0.29	0.985
0.010	12	215	0.35	0.972
0.015	10	228	0.29	0.963
0.015	13	231	0.31	0.951

To further verify the stability of the PSO hyperparameters, an ablation experiment is conducted. In this experiment, the value of one hyperparameter is changed while the others are fixed at their optimal values. The changes in the model’s performance are then observed. Therefore, the ablation experiment results of PSO hyperparameters are shown in Table 3.

Table 3.

Experimental results of PSO hyperparameter ablation.

	RMSE	Accuracy	Recall	F1-score
Ablation experiment of initial learning rate
Initial learning rate
0.005	0.021	95.23%	94.88%	95.02%
0.010	0.015	97.55%	97.13%	97.23%
0.015	0.023	95.75%	95.98%	96.17%
Ablation experiment of time window step size
Time window step size
8	0.013	97.69%	97.48%	97.13%
10	0.021	96.97%	97.19%	96.49%
12	0.020	97.47%	97.06%	97.07%
13	0.015	97.26%	96.96%	96.99%
16	0.017	97.46%	97.21%	96.77%
Ablation experiment of training frequency
Training frequency
210	0.012	97.38%	96.89%	96.55%
213	0.014	97.24%	97.05%	96.48%
215	0.015	96.89%	97.46%	96.94%
225	0.011	97.49%	97.69%	97.18%
228	0.019	96.97%	97.07%	97.02%
231	0.020	97.33%	97.43%	96.72%
Ablation experiment of dropout
Dropout
0.25	0.017	97.40%	96.81%	96.59%
0.29	0.012	97.57%	97.05%	97.04%
0.31	0.014	97.25%	96.90%	96.82%
0.32	0.015	97.15%	96.86%	96.92%
0.35	0.018	96.89%	96.89%	96.65%

As shown in Table 3, the model performs best with an initial learning rate of 0.010. Low or high learning rates can decrease model performance. For example, the RMSEs of the models corresponding to initial learning rates of 0.005, 0.010, and 0.015 are 0.021, 0.015, and 0.023, respectively, with an accuracy of 95.23%, 97.55%, and 95.75%, respectively. Additionally, the ablation experiments on the time window, step size, training frequency, and dropout reveal that the parameter values corresponding to better model performance are 8, 225, and 0.29, respectively. The results of the study suggest that the PSO hyperparameter combination chosen for the study (an initial learning rate of 0.010, a time window step size of 8, a training frequency of 225, and a dropout of 0.29) has a significant impact on model performance, confirming the rationality and stability of the parameter selection process. Referring to relevant research, the hyperparameters and training configurations of the LSTM model are shown in Table 4.^28,29

Table 4.

LSTM model hyperparameters and training configuration table.

Hyperparameters/Configuration	Value	Hyperparameters/Configuration	Value
Input feature dimension	8	Number of epochs	100
Hidden layer size	128	Learning rate	0.010
LSTM layers	2	Dropout rate	0.29
Batch size	64	Activation function	Softsign
Optimizer	Adam	\	\

As shown in Table 4, the LSTM model has a hidden layer size of 128, two LSTM layers, and a dropout rate of 0.29. These parameters are set to optimize the model’s learning efficiency and prediction accuracy, thereby achieving effective early warning of medical equipment failures.

To test LSTM performance, optimal value data obtained by PSO is substituted into it. The loss function of the dataset is calculated, and the predicted value is compared. The LSTM performance comparison chart is shown in Figure 10. Figure 10(a) shows the loss function curve of the training set and verification set under LSTM. By observing the trend of the curve, it can be found that when iterations reach 43, the loss function of the training set and verification set stops falling sharply and gradually stabilizes. The loss functions after stabilization are 0.03 and 0.19, respectively. Next, real value and predicted value are compared at the normalized temperature, and the comparison result is shown in Figure 10(b). When the number of samples is 20, normalized temperature reaches a maximum value of 0.62. Moreover, the curve changes of real and predicted values are basically consistent, which proves LSTM prediction accuracy.

Figure 10.

LSTM prediction model performance comparison chart.

To verify the impact of PSO optimization input on the performance of LSTM model, a brief ablation experiment is conducted. The experimental results are shown in Table 5.

Table 5.

Experimental results on the effect of PSO optimization input on the performance of LSTM model in ablation experiments.

LSTM	PSO	RMSE	Accuracy	Recall	F1-score
√	√	0.012	97.68%	97.55%	97.49%
√	×	0.154	91.20%	90.01%	91.72%

As shown in Table 5, optimizing the input features using PSO significantly reduces the RMSE of the model compared to a model with unoptimized features. The accuracy, recall, and F1-score of the model are also significantly improved after optimizing the input features. These results suggest that PSO optimization can substantially enhance model performance. To verify performance of the model used in this study, a variety of similar prediction models is compared. The prediction models include RNN, Back Propagation (BP)-GA, Bidirectional LSTM (Bi-LSTM), and Attention-LSTM, which are trained and predicted in the same way. Finally, the absolute error comparison of the prediction of the operating temperature characteristics of each model is shown in Figure 11.

Figure 11.

Comparison chart of absolute prediction errors among variable models.

In Figure 11, the worst performing model is BP-GA with an Absolute Error Range (AER) of [0.09, 0.23], the RNN model with an AER of [0.04, 0.09], Bi-LSTM with an AER of [0.01, 0.03], and attention LSTM with an AER of [0.01, 0.02]. The best performing model is the PSO-LSTM prediction model used in the study with an AER of [0, 0.01] and stable performance. It is suggested that prediction model accuracy is better than similar prediction models. The smaller error of the PSO-LSTM prediction model can be attributed to the utilization of the PSO algorithm to identify the optimal solution for preprocessed data. It is then inputted into the LSTM warning model for learning and training. This approach effectively reduces the error of the PSO-LSTM prediction model. This is similar to the research results of Arya et al.³⁰ To comprehensively the comprehensive performance and accurately compare deviation between prediction and actual results of each model, RMSE and Mean Absolute Percent Error (MAPE) are used to calculate the prediction error corresponding to each feature. Table 6 shows final performance results.

Table 6.

Variable models performance results.

Model	Training TRMSE	Testing TRMSE	Training MAPE	Testing MAPE
RNN	0.059	0.115	0.091	0.179
BP-GA	0.092	0.156	0.121	0.235
Bi-LSTM	0.029	0.071	0.041	0.081
Attention-LSTM	0.023	0.041	0.035	0.077
PSO-LSTM	0.015	0.025	0.03	0.051

In Table 6, BP-GA prediction performance is the worst, and each index is the maximum value among the models, followed by RNN. The model indicators based on the LSTM model have little difference. However, Bi-LSTM prediction RMSE in test set is 0.071, which is much larger than attention LSTM prediction and PSO-LSTM prediction. The average absolute percentage error of attention LSTM prediction in test set is too large, and the specific value is 0.077. The differences between the RMSE of training set, average absolute percentage error of training set, RMSE of test set, and average absolute percentage error of test set of PSO-LSTM proposed in this study are 0.015, 0.030, 0.025, and 0.051, respectively. They are the lowest values of all indicators, which proves the performance superiority of PSO-LSTM proposed in this study.

To better validate the performance of the PSO-LSTM model proposed in the study, a comparison is made between the warning results of the model, including the precision, recall, and F1-score. The comparison results are shown in Table 7. In Table 7, the maximum values of precision, recall, and F1-score all occur in the PSO-LSTM model designed for the study, with values of 97.98%, 97.82%, and 97.68%, respectively, followed by the Attention LSTM model and Bi LSTM model. The minimum values of precision, recall, and F1-score all appear in the BP-GA model, indicating that the warning effect of this model is relatively poor and there is still room for further optimization. Overall, the PSO-LSTM model designed in this study performs well in warning precision, with a high overall model rating and better performance. The PSO-LSTM model demonstrates superior warning precision due to its capacity to identify the optimal value of input data to the LSTM, enhance data quality, and thereby optimize the training effectiveness and performance of the PSO-LSTM model. This is similar to the study of Manoj et al.³¹

Table 7.

Comparison of precision, recall, and F1-score for different models.

Model	Warning results
Model	Precision	Recall	F1-score
RNN	92.61%	92.91%	92.73%
BP-GA	89.71%	88.21%	88.94%
Bi-LSTM	93.64%	93.51%	93.99%
Attention-LSTM	95.50%	95.33%	95.60%
PSO-LSTM	97.98%	97.82%	97.68%

To more accurately evaluate the performance of the PSO-LSTM model in medical equipment fault warnings and reduce the risk of overfitting, the study adopts the K-fold cross-validation method. The dataset is divided into 10 subsets, with each subset used for validation and the remaining 9 subsets used for training. The results of the K-fold cross-validation are shown in Table 8.

Table 8.

Results of K-fold cross-validation.

Folded number	RMSE	Precision	Recall	F1
1	0.015	97.85%	97.58%	97.49%
2	0.014	97.92%	97.64%	97.57%
3	0.015	97.88%	97.61%	97.53%
4	0.013	97.95%	97.67%	97.61%
5	0.014	97.90%	97.63%	97.52%
6	0.013	97.93%	97.66%	97.56%
7	0.014	97.89%	97.62%	97.51%
8	0.013	97.96%	97.69%	97.62%
9	0.014	97.91%	97.64%	97.55%
10	0.013	97.94%	97.68%	97.59%
Average	0.014	97.91%	97.64%	97.56%
Standard deviation	0.0008	0.03%	0.03%	0.04%

In Table 8, the PSO-LSTM model performs well in all indicators on the 10 subsets. The average values of RMSE, precision, recall, and F1-score are 0.014, 97.91%, 97.64%, and 97.56%, respectively, with standard deviations of 0.0008, 0.03%, 0.03%, and 0.04%, respectively. This indicates that the PSO-LSTM model generalizes well and is robust across different subsets of data, providing stable performance. To improve the PSO-LSTM model’s practical application value, it is integrated into a hospital’s existing information system, and its effectiveness is analyzed. The actual application effects of the PSO-LSTM model are shown in Table 9.

Table 9.

Practical application effect of PSO-LSTM model.

Hardware compatibility	Delay			Sensor reliability
Over 95%	Model average latency	Data transmission delay	Model inference delay	Accuracy of fault detection	Reliability of data collection
Over 95%	54 ms	<30 ms	<55 ms	98.85%	99.32%

As shown in Table 9, the PSO-LSTM model is highly compatible with existing hospital information systems and has an adaptation rate of over 95%. In terms of latency, the model’s average latency is 54 milliseconds. Data transmission latency is less than 30 milliseconds, and model inference latency is less than 55 milliseconds. These results meet real-time monitoring requirements. In terms of sensor reliability, the model achieves a fault detection accuracy of 98.85% and a data collection reliability of 99.32%. These results indicate that the model is highly stable and reliable in practical applications. These results validate the PSO-LSTM model’s effectiveness and practicality in medical equipment fault warning.

Conclusions

Aiming at fault detection of healthcare medical equipment, the fault diagnosis model of healthcare medical equipment is built by combining the PSO algorithm and LSTM. The performance test is carried out by taking the medical infrared imager as an example. In data preprocessing part, the average waveform vibration amplitude interval^1,9 in the first 4 seconds of the original signal was 5.46, while the average waveform vibration amplitude interval^4,6 in the first 4 seconds of the denoised waveform was 5.23. It showed that the noise reduction effect was good. Original data MMV in the first 10 seconds were 40.3 and 34.2, respectively, and data MMV in the first 10 seconds after normalization were 1.00 and 0.29, respectively. Thus, the results were distributed in the [0, 1] interval, which saved a lot of data computation. By observing the loss function curves under LSTM, it was evident that the model tended to be stable when the number of iterations reached 43. After stabilization, the loss functions of two sets were 0.03 and 0.19, respectively. By comparing the performance of PSO-LSTM with a variety of similar prediction models, the differences of prediction model in the training set RMSE and MAPE, test set RMSE, and MAPE were 0.015, 0.030, 0.025, and 0.051, respectively, which were the lowest values of all indicators. The AER of the PSO-LSTM model was between [0, 0.01], and the performance was stable. The performance superiority of the PSO-LSTM model was proved. The precision, recall, and F1-score of the PSO-LSTM model were 97.98%, 97.82%, and 97.68%, respectively, which were significantly higher than those of the control model. In conclusion, data preprocessing had the potential to effectively denoise raw data. The PSO-LSTM prediction model demonstrated the capacity to effectively warn of medical equipment failures with excellent performance. These results suggest that the PSO-LSTM model can accurately detect potential equipment failures and respond in real time. These studies also demonstrate that the model provides excellent technical support for monitoring medical equipment in real-world settings. This significantly improves maintenance and management processes, enhances the quality and efficiency of medical services, and controls medical costs.

Based on the government’s emphasis on risk warning management of adverse events in medical devices, the requirements for risk warning of medical device equipment will become increasingly high and strict in the future, which will also prompt further exploration in risk warning of medical device equipment in future research. To implement the PSO-LSTM prediction model in a practical setting, it is essential to integrate the model as a functional plugin within the hospital’s existing information system and to ensure compatibility with the specific version of the model in use. Sufficient testing and monitoring are also required to ensure that the model goes live normally. In a real testing environment, the model also requires sensors and USB interfaces to obtain data and uses PSO-LSTM function plugins to complete device warnings. In addition, in order to enhance the practical value of the model, some basic training can be provided to relevant medical personnel, including the data collection process and methods, the usage process of PSO-LSTM functional modules, and the analysis and interpretation of results. Therefore, medical personnel can use the model and analyze the predicted results of the model.

However, there are also certain shortcomings in the research. First, affected by the experimental conditions, the study collected only two datasets for model training and experimental analysis, lacking the performance test of diversified data. In the actual working environment, the factors affecting the fault diagnosis of medical equipment are often complex and changeable, so future research will consider more diversified sample data for further research. Secondly, the PSO-LSTM prediction model also has certain limitations. On the one hand, the PSO algorithm is prone to getting stuck in local optima, which can lead to poor feature optimization performance and affect the accuracy of the PSO-LSTM model. Future research could explore introducing multiple population PSO or multi-strategy PSO to enhance its global search capability. This would optimize the feature selection process and improve the accuracy of the PSO-LSTM model. On the other hand, the LSTM model requires a large amount of data and is prone to overfitting. This can lead to unstable performance of the PSO-LSTM model on small datasets, reducing its generalization ability and reliability. Future research could employ transfer learning and data augmentation techniques to increase the size of the dataset, thereby enhancing the generalization ability and reliability of PSO-LSTM models. Third, although PSO algorithm performs well in feature vector optimization, its parameter selection (such as learning rate and time window step size) has a significant impact on model performance. Future research can optimize the parameters of PSO algorithm to further improve the optimization results of feature vectors, thereby enhancing the accuracy of medical equipment fault diagnosis. In addition, other feature selection methods can be further explored, such as model-based feature importance assessment, to reduce redundant features. In summary, future research should explore introducing multi-population PSO algorithms or transfer learning techniques to improve model performance. At the same time, combining multiple feature selection methods can reduce redundant features and improve the optimization efficiency of feature vectors, thereby enhancing model performance.

Advanced optimization algorithms have played an important role in fields such as online learning, scheduling, multi-objective optimization, transportation, medicine, and data classification. Chen et al. designed an adaptive fast fireworks algorithm to improve the computational efficiency of large-scale black box optimization algorithms. This study constructed mechanisms for expressing fast explosions and competition cooperation between fireworks, achieving optimization of neural network controllers.³² Dulebenets scholars designed a new adaptive polyploid meme algorithm to address the complex decision scheduling problems faced by cross docking terminal operators. This algorithm used hybridization techniques to facilitate the search process and directly relied on the concept of polyploidy, with better performance than well-known advanced metaheuristic algorithms.³³ Safaeian et al. designed a new three-objective optimization model to solve the scheduling decision problem of carpooling. This study constructed a multi-objective deer algorithm to intelligently find effective Pareto solutions, achieving intelligent planning of carpooling systems with better performance than other methods.³⁴

It is evident that advanced optimization algorithms demonstrate efficacy in the domains of scheduling and multi-objective optimization. Furthermore, these methods can be employed to address the decision-making challenge in this study, thereby enhancing the efficacy of the equipment warning model. Moreover, future research may wish to compare the proposed method with more advanced optimization algorithms, which could be used to address issues in other fields.

Footnotes

ORCID iD

Shan Li

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Zhang

, et al. Intrinsic component filtering for fault diagnosis of rotating machinery. Chin J Aeronaut 2021; 34(1): 397–409.

Althubaiti

Elasha

Teixeira

. Fault diagnosis and health management of bearings in rotating equipment based on vibration analysis – a review. J vibroeng 2022; 24(1): 46–74.

Shamayleh

Awad

Farhat

. IOT based predictive maintenance management of medical equipment. J Med Syst 2020; 44: 1–12.

Shi

Huang

Wang

, et al. An information integration approach to spacecraft fault diagnosis. Enterp Inf Syst 2021; 15(8): 1128–1161.

Zhan

Yang

. A novel inverter fault-tolerant topology for faults in pmsm drive systems. Adv Ind Eng Manag. 2024; 13(2).

Chen

Xue

, et al. Quantum-behaved particle swarm optimization of convolutional neural network for fault diagnosis. J Exp Theor Artif Intell 2024; 36(7): 1035–1051.

Sulaiman

Menanti

NAS

Yusop

, et al. Fault detection and diagnosis of air-conditioning system using long short-term memory recurrent neural network. Prz Elektrotechn 2023; 99(9): 113–117.

Jin

Qin

Zhang

, et al. A multi-scale revolutionary neural network for bearing compound fault diagnosis under variable noise conditions. Sci China Technol Sci 2022; 65(11): 2551–2563.

Wang

Yang

. Bearing fault diagnosis algorithm based on granular computing. Granul Comput 2023; 8(2): 333–344.

10.

Huo

Lin

. Intelligent fault diagnosis method of mechanical equipment based on fuzzy pattern recognition. J Intell Fuzzy Syst 2020; 38(4): 3657–3664.

11.

Han

Ding

, et al. Short time wavelet entropy integrating improved LSTM for fault diagnosis of modular multilevel converter. IEEE Trans Cybern 2021; 52(8): 7504–7512.

12.

Sheela

Arun

. Hybrid PSO-SVM algorithm for Covid-19 screening and quantification. Int J Inf Technol 2022; 14(4): 2049–2056.

13.

Sadhin

Rayhan

. A review of convolutional neural network in emerging trends and opportunities in precision agriculture. Acta inform Malays 2023; 7(2): 67–73.

14.

Liu

Lang

Zhang

, et al. Intelligent fault diagnosis of medical equipment based on long short term memory network. J Biomed Eng 2021; 38(2): 361–368.

15.

Hou

Luo

, et al. Comparative study on the use of acoustic emission and vibration analyses for the bearing fault diagnosis of high speed trains. Struct Health Monit 2022; 21(4): 1518–1540.

16.

Yang

, et al. Studies on fault diagnosis of dissolved oxygen sensor based on ga-svm. Math Biosci Eng 2021; 18(1): 386–399.

17.

Cao

Zeng

, et al. Application of DBN and GWO SVM in analog circuit fault diagnosis. Sci Rep 2021; 11(1): 1–14.

18.

Wang

Xiang

. A minimum entropy deconvolution enhanced revolutionary neural networks for fault diagnosis of axial piston pumps. Soft Comput 2020; 24(4): 2983–2997.

19.

Choudakkanavar

Mangai

Bansal

. MFCC based ensemble learning method for multiple fault diagnosis of roller bearing. Int J Inf Technol 2022; 14(5): 2741–2751.

20.

Zhu

Fan

, et al. Design of medical equipment system based on neural network algorithm and network feature. J Intell Fuzzy Syst 2021; 40(4): 6815–6825.

21.

Nambu

Ito

. Studies on noise reduction of optically pumped magnetometers by digital signal processing. Electron Commun Jpn 2024; 144(8): 331–336.

22.

Liu

Zhao

Wang

, et al. LSTM-GAN-AE: a promising approach for fault diagnosis in machine health monitoring. IEEE Trans Instrum Meas 2021; 71: 1–13.

23.

Tsai

Wang

Chung

, et al. Multisensor fusion time – frequency analysis of thrower blade fault diagnosis based on deep learning. IEEE Sens J 2022; 22(20): 19761–19771.

24.

Khan

Ullah

Pamucar

, et al. Performance measure using a multi attribute decision making approach based on complex t-spherical fuzzy power aggregation operators. J Comput Cogn Eng 2022; 1(3): 138–146.

25.

Chen

. Research on internet security situation awareness prediction technology based on Improved RBF neural network algorithm. J Comput Cogn Eng 2022; 1(3): 103–108.

26.

Minghui

Lin

, et al. Digital twin model of gas turbine and its application in warning of performance fault. Chin J Aeronaut 2023; 36(3): 449–470.

27.

Ghannadiasl

Ghaemifard

. Parameter selection for PSO-based hybrid algorithms and its effect on crack detection in cantilever beams. Num Meth Civ Eng 2024; 9(2): 17–28.

28.

Wang

Liu

. Street lamp status warning system based on internet of things technology. J Elect Res App 2024; 8(4): 154–160.

29.

Ding

. Industrial mechanical equipment fault detection and high-performance data analysis technology based on the Internet of Things. Intell Decis Technol 2024; 18(4): 3171–3184.

30.

Arya

Malik

. Software fault prediction using K-mean-based machine learning approach. Int J Perform Eng 2023; 19(2): 133–143.

31.

Manoj

Jose

. An integrated EMVO and ARBFN algorithms for output power forecasting and fault prediction in solar PV systems. J Electr Eng Technol 2023; 18(5): 3443–3455.

32.

Chen

Tan

. SF-FWA: a self-adaptive fast fireworks algorithm for effective large-scale optimization. Swarm Evol Comput 2023; 80(80): 101314.

33.

Dulebenets

. An adaptive polyploid memetic algorithm for scheduling trucks at a cross-docking terminal. Inf Sci 2021; 565(19): 390–421.

34.

Safaeian

Khayamim

Ozguven

, et al. Sustainable decisions in a ridesharing system with a tri-objective optimization approach. Transport Res Transport Environ 2023; 125: 103958.