Performance evaluation of enterprises’ innovation capacity based on fuzzy system model and convolutional neural network

Abstract

Influenced by national policies and macro-economic environment, large domestic enterprises is actively promoting strategic transformation to enhance their core competitiveness, and performance evaluation of enterprises’ innovation capacity has become a hot topic in recent years. This paper proposes a performance evaluation method of enterprises’ innovation capacity based on deep learning fuzzy system model and convolutional neural network analysis of innovation network. First of all, on account of the characteristics of breakthrough innovation and drawing on the traditional innovation performance evaluation model, this paper constructs a breakthrough innovation performance evaluation index system for enterprises from the six dimensions of main resource input, technology out-turn, process management, product performance, social value and commercial Value. Secondly, the introduction of machine learning of fuzzy convolutional neural network to assess the advancement execution of enterprises is of great significance for enterprise managers to find out the problems and causes of enterprises’ innovation, optimize the allocation of enterprises’ resources and further improve the innovation performance of enterprises. The experimental results show to verify the adequacy of the algorithm.

Keywords

Innovation network fuzzy system model convolutional neural network (CNN)

1 Introduction

In today’s turbulent global economy, innovation plays an important role in maintaining the competitive advantage and long-term development of the enterprises [1]. Choosing and pursuing a right innovation project is one of the key factors for enterprises to achieve success [2, 3]. Therefore, in the fierce market competition, in order to maintain long-term development and enhance the core competitiveness of enterprises, enterprises need to continuously carry out a series of innovative projects, such as technological innovation, product innovation and management innovation [4]. At present, the R & D and manufacture of complex products represented by airline passenger aircrafts, high-speed trains and new energy vehicles are becoming the mainstream form of product innovation, which is the key to the achievement of core competitiveness of the nation and enterprises in the international market. Large complex product development projects are often seen as a portfolio of innovative projects containing a range of sub projects and programs [5]. As these projects compete for enterprise resources, enterprises often face the challenge of scarce resources [6]. In order to better coordinate these sub-projects and programs, enterprises often manage innovation activities by adopting the project portfolio management whole-process integration method, which also highlights the important role of project portfolio management [7]. The objectives of portfolio management not only include shortening the product development cycle, reducing development costs, and ensuring timely delivery of products or services that are consistent with business strategy within limited resources SFC-Subtractive Fuzzy Clustering [8 –10], but also include creating a project portfolio to help enterprises achieve long-term growth target [11 –13]. Multi-NFClass-Multi Neuro-Fuzzy Classification model the combination of fuzzy logic and neural network is a successive way to integrate the Model system [14]. An enterprise innovation capability performance evaluation method based on innovative network convolutional neural network analysis with deep learning fuzzy based system (Vaibhav Kumar, and Garg, 2018) [16] is proposed in this paper to establish an enterprise breakthrough innovation performance evaluation index system from six dimensions of main resource input, technology out-turn, process management, product performance [17 –20] and business evaluation based on the features of breakthrough innovation and by learning from the traditional advancement execution evaluation model with fuzzy function [21 –25] and the convolutional neural network is introduced to make an analysis on the advancement execution of enterprises, and the great significance for the further improvement in the innovation performance of enterprises [26]. The correlated state of art, evaluation index framework and proposed method has been discussed in section-1 and section-2 correspondingly. Then the convolutional neural network structure and deep learning fuzzy CNN classifier model were described with fuzzy set formulas in section-3 and section-4 respectively. The state of art research, experimental analysis and results shown in section-5 and section-6 [27].

2 Development of innovation performance evaluation index framework

On the basis of extensively learning from relevant performance evaluation researches, based on the connotation and features of breakthrough innovation, and by reference to breakthrough advancement recognition indicators, an omnibearing and full-process breakthrough advancement execution evaluation index framework was established [28]. (1) Breakthrough innovation input performance indicator: Depending on various differences between breakthrough advancement and incremental innovation, there are significant differences in the resources and conditions that are required by enterprises during the development and use of the two kinds of innovations. Incremental innovation activities are carried out depending on the existing methods and resources abilities of the enterprises [29], the span breakthrough advancement may make the first innovation out of date, and likewise it causes damage to existing resources and abilities. Which is to make better breakthrough advancement execution, continuous and huge amount of capital and ability and other advanced assets are required to be invested in. In addition, in terms of the composition of R&D members [30], knowledge differences are required in order to avoid inertial thinking and collective blindness, breakthrough innovation is required to be taken as the common vision of the enterprise in order to coordinate the conflict between the acquisition of innovative resources and the distribution of interests [31].

(2) Breakthrough innovation technology performance indicator: Breakthrough innovation technology gets over the path dependence, has a discontinuity in technology evolution, the existing technology and economic rules has changed traces in transition of technology performance thus the leading edge and position of technology can be maintained. Transition is a process of quantitative change to qualitative change. The accumulation and diffusion of technology, organizational learning and knowledge fusion within the enterprise are indirect technological innovation performance, which has an important impact on the final direct technical output. Therefore, four indicators, namely, patent quantity, specialized execution, authoritative learning, and innovation administration are adopted in the measurement of the breakthrough innovation technology performance dimension.

(3) Breakthrough Advancement product performance index: Innovation is not a product. The presence of a product is to fulfill the requirements of clients. A thing that can’t address the issues of clients is not a product even if it is technically perfect. Unlike continuous innovation products, breakthrough innovation products are hard to become famous among mainstream clients at first, so existing customer relationship networks can be rarely borrowed, but they have other features that a couple of radical (and normally new) users like so as to substitute or cause crash on existing products with related superior performance.

(4) Breakthrough innovation process performance indicator: breakthrough innovation process performance revolve the executive improve of enterprise breakthrough advancement activities, including the following four parts: First, the independence of the Research and Development team, for example, the duality of breakthrough advancement organizations is proved; second, failure sufferance, if the managers can normally understand failure in the procedure of enterprise advancement, the initiative spirit of employees can be better protected; third, the source of innovativeness, the features of breakthrough innovation determines the ideas come from technology personnel, instead of users or executives; fourth, process innovation, it means whether the enterprises can adequately update creation equipment and procedures.

(5) Breakthrough advancement business performance indicator: Breakthrough advancement business execution can be calculated from two perspectives of market and innovation with four index. Market index include market clarity and market fitness. Most of the breakthrough advancement products are first applied in new or non-essential markets. The innovation orientation markets are often unclear and difficult to predict. Moreover, innovation orientation markets often do not coincide with the mainstream market, they are not complementary to and perfection of the mainstream market, and they are a kind of substitution.

(6) Breakthrough innovation social performance indicator: He Shankan accepts that all mankind subjective exercises are a procedure of always seeking after social qualities in advancement. Be that as it may, the positive and negative effects of mechanical development on social advancement nearly occur simultaneously. This requires seeing the applications used in technological advancement achievements from perspective of the responsibilities that human beings should undertake, examining all scientific and technological innovations with ecological values, and solving contradictions between material and ecological civilizations.

3 Convolutional neural network structure

By constructing multiple hidden layers, traditional neural network algorithms can also learn complicated classified curved surfaces through a large number of samples. Therefore, traditional neural networks are used in speech recollection and image recognition. In the traditional pattern recognition method, an algorithm is first designed generally for extraction of features other hand common Histogram of Oriented Gradients (HOG) features, Local. Binary Patterns (LBP) features and scale invariant feature transform (SIFT) features, at that point the highlights extricated are passed into the following trainable classifier to prepare the classifier. At last, tests are brought into the classifier for classification. In this mode, in light of the fact that the removed highlights are commonly little, a completely associated multi-layer system can be structured as a classifier. Another idea is to consider making the feature extraction and classification by directly using neural networks, and omit the previous feature extraction process. The back-propagation algorithm can train the previous layers of neural networks into feature extractors, and use the latter layers to make classification: but this method has a major drawback [4, 5].

First of all, if the neural network is used for feature extraction directly, the hidden layer nodes must be not too small, and the incoming speech spectrum and image pixels are often large in order to extract features good enough. Suppose the input layer data has 1000 neurons (generally far more than 1000), the first hidden layer has 100 neurons. Since the neural network is fully connected, there are more than 100,000 connecting weights between the two layers only. Even the training speed is not taken into consideration, if there are insufficient training samples, these parameters are difficult to fit the network; secondly, the fully connected network learns each sample, and the incoming of each sample will affect the update of parameters, and there is often a great similarity between the incoming data. For example, in image recognition, the difference between two frames of images may be only a small azimuthal movement, and thus the fully connected network cannot capture this information, and the training process cannot be optimized according to the sample, so this is a time-consuming and laborious process. Convolutional neural networks can alleviate these problems by taking advantage of the local features of the data through certain means.

The convolutional neural network is a multi-layer perceptron (MLP) inspired by the visual neural mechanism. Like traditional neural networks, it adopts backpropagation algorithm to train the network. The difference is that its network structure has been simplified and improved. If the neural network is considered as an algorithm for training and generation of classifiers, the convolutional neural network can be generalized as an algorithm for extracting features and training to generate classifiers. The three basic concepts of convolutional neural networks are: local perceptual fields, weight sharing, and time or space downsampling. This structural feature makes the network highly adaptable to the changes such as input noise, deformation and distortion.

In a convolutional neural network model, neurons can be divided into two categories, one is the S element used for feature extraction, and the other is the C element that is resistant to deformation. There are two important parameters in the S element, namely the threshold parameters and receptive fields; the receptive field extracts the space from the input layer as the input, and the threshold parameter controls the degree of response of the output to the input. Similarly, a convolutional neural network has multi-layer network structure, every one of its layer is actually made out of different element maps, each of which indicates a feature; every feature map has many dependent neurons. Correspondingly, the network layer of the convolutional neural network is split into a convolutional layer and a downsampling layer, also called downsampling or subsampling layer; it is not linear mapping between the network layers, it is a downsampling layer from the convolutional layer to the downsampling layer, and it is a convolution extract operation from the downsampling layer to the convolutional layer.

Figure 1 the simplified convolutional neural network structure:

Fig. 1

Convolutional Neural Network Structure Diagram.

In the Fig. 1 Input is the input layer, through 3 trainable digital filters added with bias, and through a Sigmoid function, C₁ is obtained, and C₁ is a convolution layer. Because there are 3 filters, there are 3 feature maps, and each feature map represents a set of learned features; the process from C1 to S2 is a downsampling process, and the specific practice is summating 4 consecutive pixels (2×2 regions) in C₁, adding bias to weighted value, and mapping to a point in S2, each feature map in C1 is separately downsampled, so S2 also contains 3 feature maps, and then it is convolutional filtering from S₂ to C₃ and it is downsampling from C₃ to S₄. The features are pulled into a straight line to act on the neural network. Figure 2 shows the connection process:

Fig. 2

Convolutional Neural Network Connection Process.

f_x here is a digital filter, b_x is a bias, C_x is a feature map of the convolutional layer; W_x₊₁ is the weight value of the downsampling, b_x₊₁ is the corresponding weight, and the downsampling layer S_x₊₁ is obtained.

It can be seen that Layer C is used as the convolution layer for feature extraction. Each neuron is connected to the small receptive field on the upper layer, and then the receptive field is moved, and the new receptive field is mapped to another neuron on Layer C. Sigmoid is adopted to make the process have displacement invariance. As long as the size of input layer and the size of local receptive field are determined, then the size of Layer C is also determined. Layer S is downsampling layer whose purpose is to change multiple pixels of Layer C into one pixel.

Because the weight values on the mapping surface are shared, that is to say, the weight of each neuron is the same, the entire network parameters will be greatly reduced and the complexity will be reduced; the method of a combination of feature extraction and downsampling is adopted for the network, and the downsampling is locally averaged. Such a structure makes the network highly resistant to distortion.

In summary, it is simpler to implement SIFT algorithm than CNN, fewer design parameters can be used, and it has less requirements for computational performance and storage, and thus it can be used for image recognition tasks that require real-time processing, but it lacks robustness to nonlinear transformation. CNN is a popular computing model for image processing based on neural network, needs to be trained and adjusted to obtain better judgment ability, has higher requirements for computation and storage in neural network training process, and is very complicated in its implementation process. Figure 3 Therefore, it is necessary to make a balance between computing power and results in the actual projects in order to choose a more appropriate scheme. We used network architecture to make output of the neurons.

Fig. 3

Convolutional Neural Network Architecture.

In the above Fig. 3 the input given to convolution layer it establishes the new connection and the hidden layer are presents and output neurons connected with loop formation to produce a membership function. The connection loads of the convolutional neurons situated in one element map are shared, and the position of the nearby feature turns out to be less significant. Such a way enable the shift invariance to be achieved. The output neural value $F_{n}^{l}$ of the n^th feature map of the convolutional layer l can be determined as follows Equation (1)

$\begin{matrix} F_{n}^{l} (x, y) = f (\sum_{m \in S_{n}^{l}} \sum_{i = 0}^{k^{l} - 1} \sum_{j = 0}^{K^{l} - 1} φ_{mn}^{l} (i, j) F_{m}^{l - 1} \\ (x + i, y + i) + c_{n}^{l}) \end{matrix}$ (1)

Where Equation (1) the x, y are coordinates of a neuron inside the element map; $S_{n}^{l}$ is the set of element maps of the previous layer l–1 that are associated with the n-th element map of the layer l; $φ_{mn}^{l}$ is the matrix of coefficients k^l is the size of the responsive field of the neurons of the lth layer; $c_{n}^{l}$ bias label for the nth element map of the lth layer.

4 Deep learning fuzzy CNN classifier model

We proposed Deep learning Fuzzy convolution layer incorporates three expert stages, to be specific fuzzy convolution stage, pooling stage and nonlinearity stages The Fuzzy convolutional stage is a procedure of applying fuzzy convolutional filters, Equation (2) $y_{ij} = \sum_{l = 0}^{p - 1} \sum_{k = 0}^{r - 1} S_{ω} y (i + l) (j + k)$ (2)

To Evaluate the Convolutional neural network layer error let x_n is the input and ${\hat{x}}_{n}$ is the output C is number of sample analysed. $E = - \frac{1}{C} \sum_{n = 1}^{c} [x_{n} \log ({\hat{x}}_{n}) + (1 - x_{n}) \log (1 - {\hat{x}}_{n})]$ (3)

The back propogation alogorithm is used to identify the fuzzy layer weight function used Equation as (4), $W_{fn} (j + 1) = W_{fn} (j) - α_{nn} \frac{\partial F}{\partial W_{fn}}$ (4)

The Defuzzification membershiop function M_y (k)Equation (5) to be used to get output stage, $M_{y} 2 * 1 (k + 1) = M_{y} (k) + b_{vy} \nabla M_{y}$ (5)

The fuzzification membershiop function M_ω (k) Equation (5) to be used to get convolution layers weight calculated, $M_{ω} (k + 1) = M_{ω} (k) + α_{M ω} \nabla W_{μ y}$ (6)

Where, $δ_{j} = {(W_{μ j})}^{T} δ_{j} f (y_{j})$ (7) $\nabla W_{μ y} = \sum x_{ij} {* δ}_{j}$ (8)

The mean and variance of fuzzification layer membership function, where α_Mxis the learning rate of fuzzification layer. $M_{x} (k + 1) = M_{x} (k) + α_{Mx} \nabla W_{μ x}$ (9) ${σ M}_{x} (k + 1) = σ_{Mx} (k) + σ_{Mx} \nabla M_{x}$ (10)

Equation (10) for each element in the input is assigned as multiple linear labels based on membership function. The fuzzy membership function calculates the grades that describes the membership of the input node.

As shown in the above Fig. 4 represents the Accuracy rate analysis for varies method like Subtractive Fuzzy Clustering (SFC), Multi Neuro Fuzzy class (Multi-NFclass) and Proposed method deep learning fuzzy logic enterprises model based on dataset range. Deep learning approach based on fuzzy logic is more accurate when compared to other methods.

Fig. 4

Accuracy of Fuzzy Membership Function.

Using deep learning fuzzy logic Fig. 5 shows the perfomance of the enterprises model compared to Subtractive Fuzzy Clustering (SFC), Multi Neuro Fuzzy class (Multi-NFclass) and Deep learning Fuzzy logic system Performance analysis. Deep Learning fuzzy sets performance is high when compared to other methods.

Fig. 5

Performance of Fuzzy membership function.

As shown in the above graph Fig. 6 represents the error rate for the fuzzy logic for enterprises model error has reduced compared to Subtractive Fuzzy Clustering (SFC), Multi Neuro Fuzzy class (Multi-NFclass) because of the deep learning fuzzy logic system.

Fig. 6

Error Rate for fuzzy set.

As shown in the above Fig. 7 represents the precision rate for varies method like Subtractive Fuzzy Clustering (SFC), Multi Neuro Fuzzy class (Multi-NFclass) and Proposed method deep learning fuzzy logic enterprises model based on dataset range. Deep learning approach based on fuzzy is more accurate when compared to other methods.

Fig. 7

Precision Rate for fuzzy set.

5 Experimental analysis

In the process of data collection, there are two enterprises whose enterprise innovation index values are missing in a single year, which respectively are the data of Yanzhou Coal Mining Company Limited in 2012 and the data of Jizhong Energy Group Co., Ltd. in 2013. These two missing values are made up by interpolation method based on the values of neighboring two years. At the same time, when observing the original data, it is found that a small number of indexes are negative, for example, the net profit growth rate is negative. The processing method of this paper is to convert the negative index to 0 and continue to participate in subsequent data processing. According to the steps described above, the learning samples are selected and the network model is trained and tested. After the network operates for 167 epochs, the model squared error and MSE are reduced below the target fault, then simulation impacts of the system is good, as shown in Fig. 4. Besides, the output value of the model after training is very close to the expected value, achieving the accuracy requirement, and then the network can be used for simulation.

The network, with its training completed, has formed a structure with a stable weight threshold, and the trained network is saved at this time. When new data is imported, the corresponding output can be automatically obtained, that is, the overall performance of total score of the enterprise. The saved model is used then, and the sample data of No. 1, No. 7, No. 13, No. 19 and No. 25 are input into the model to obtain the test result. The fault as shown in Fig. 8. The expected output and actual output, the fault is comparatively within acceptable range, which satisfies the reality and logic, indicates that the method has good speculation capacity and can be generally utilized in the comprehensive performance evaluation of enterprises.

Fig. 8

Network training results.

It can be found that the absolute error between the actual output and the expected output obtained by the convolutional neural network simulation is controlled at about 0.01, the maximum value of the relative error is 14.21%, the minimum value is 0.23%, and the average relative error is 5.23%. This situation is acceptable in the comprehensive performance evaluation of enterprises. It can be seen that the neural network has achieved the evaluation results of high efficiency and small error, and can be promoted and applied in the future comprehensive perfomance evaluation work of enterprises. In the future, when a comprehensive performance evaluation is to be made for other enterprises, the network model can be used, all the above-mentioned index data of the enterprise can be input into it, and the total comprehensive performance score can be quickly obtained through simply running the network model.

The first is innovation performance, as can be seen from the comparison in Figs. 9 and 10. The changes in the innovation performance of enterprises are basically the same as the situation of the general evaluation results of the comprehensive performance.

Fig. 9

Linear Regression Analysis of Actual Output and Expected Output.

Fig. 10

Innovation Performance Evaluation Result Diagram.

This is mainly caused by the more importance attached to the innovation performance in weight distribution. Nevertheless, even in this case, there are still many “abnormal points”. The “abnormal point” here refers to the situation where the evaluation ranking or the change trend of the comprehensive performance is consistent with that of the innovation performance. It is these abnormal points that are the key to studying the gap and finding the crux of the enterprises. For example, in 2011, the innovation performance scores of the three enterprises, namely, China Shenhua Company, Lu’an Environmental Energy Development Co., Ltd. and Yanzhou Coal Mining Company Limited were very close, but they have a large difference in their comprehensive performance scores with distinct hierarchies, which indicated that the main factors affecting the comprehensive performance rankings of the three enterprise become the non-innovative performance score. The radius of the problem is narrowed through such analysis, and the research efficiency can be improved after focusing on the core influential factors. For another example, during the period from 2011 to 2013, the comprehensive performance score of China National Coal Group Corp. was higher than that of Jizhong Energy Group Co., Ltd., while its innovation performance score was lower than that of Jizhong Energy Group Co., Ltd., which indicated that the comprehensive performance score of China National Coal Group Corp. was improved by its good non-innovation performance, and thus it should pay more attention to the performance of its innovation management in the future.

6 Conclusion

The enterprise comprehensive performance evaluation system covers many influential factors. The index system established has complex contents and diverse hierarchies, it must be organized through hierarchy collating to make the index system be in order, and then the logical advantages of the convolutional neural network method can be fully taken in this paper. Deep learning of Fuzzy Convolutional neural network with extremely strong nonlinear mapping capabilities can be used for efficiently and accurately discovering the laws. Therefore, the analytic hierarchy process and the convolutional neural network algorithm are combined in this paper to use the convolutional neural network to effectively make weight allocation and quantitative processing of all indexes, and transform the subjective human judgment into objective statistical data, and thus the rationality, intuitiveness and credibility of conclusion can be enhanced through analysis.

References

J.J.

, Guo

and Huang

G.H.

, Achieving the objective of ecological planning for arid inland river basin under uncertainty based on ecological risk assessment, Stochastic Environmental Research & Risk Assessment30(5) (2016), 1485–1501

Frederick

C.B.

, Limiting the uncertainty in risk assessment by the development of physiologically based pharmacokinetic and pharmacodynamic models, Toxicology Letters68(1–2) (1993), 159–175.

Yang

M.Y.

and Han

, Prediction and Management Research of Hi-Tech Enterprises’ Marketing Risks – An Empirical Analysis of 35 Hi-Tech Enterprises of Communication and Relevance Equipment Manufacturing Industry[C], International Conference on Management & Service Science, 2011.

Shakeel

P.M.

, Baskar

, Dhulipala

V.S.

, Mishra

and Jaber

M.M.

, Maintaining security and privacy in health care system using learning based Deep-Q-Networks, Journal of medical systems42(10) (2018), 186. https://doi.org/10.1007/s10916-018-1045-z.

Rehman , Noor , Ali , Abbas , Shah

Ali

, Inayat

Syed

, Ali

Irfan

, Muhammad and Park

, Select this result for bulk action variable precision multi decisionλ-soft dominance based rough sets and their applications in conflict problems, Journal of Intelligent & Fuzzy Systems36(6) (2019), 5345–5360.

Kavousi-Fard

, Modeling Uncertainty in Tidal Current Forecast Using Prediction Interval-Based SVR, IEEE Transactions on Sustainable EnergyPP(99) (2016), 1–1.

Trisal

S.K.

and Kaul

, K-RCC: A novel approach to reduce the computational complexity of KNN algorithm for detecting human behavior on social networks, Journal of Intelligent & Fuzzy Systems36(6) (2019), 5475–5497.

Dulesov

A.S.

, Karandeev

D.Y.

and Dulesova

N.V.

, Reliability analysis of distribution network of mining enterprises electrical power supply based on measure of information uncertainty[C], Iop Conference Series: Earth & Environmental Science, 2017.

Z.M.

, Xu

S.W.

, Cui

L.G.

, et al., The Short-Term Forecast Model of Pork Price Based on CNN-GA, Advanced Materials Research628 (2013), 350–358.

10.

Baskar

, Dhulipala

V.R.S.

, Shakeel

P.M.

, Sridhar

K.P.

and Kumar

, Hybrid fuzzy based spearman rank correlation for cranial nerve palsy detection in MIoT environment, Health Technology, (2019). https://doi.org/10.1007/s12553-019-00294-8

11.

Landman

, Toro

K.G.

, Commo

S.A.

, et al., Prediction Interval Development for Wind-Tunnel Balance Check-Loading, Journal of Aircraft52(3) (2015), 1–6.

12.

Coccia

and Todini

, Recent developments in predictive uncertainty assessment based on the model conditional processor approach,15,10-10-28), Hydrology and Earth System Sciences7(6) (2011), 3253–3274.

13.

Kavousifard

, Khosravi

and Nahavandi

, A New Fuzzy-Based Combined Prediction Interval for Wind Power Forecasting, IEEE Transactions on Power Systems31(1) (2015), 18–26.

14.

Kokkinos

, Bouwknegt

, Verhaelen

, et al., Virological fit-for-purpose risk assessment in a leafy green production enterprise, Food Control51 (2015), 333–339.

15.

Jiménez Moreno

Robinson

, Avilés Sanchez

Oscar

, Diana

, Ovalle RGB-D Training for Convolutional Neural Network with final fuzzy layer for depth weighting, Contemporary Engineering Science, 2017.

16.

Shakeel

P.M.

, Baskar

, Sampath

and Jaber

M.M.

, Echocardiography image segmentation using feed forward artificial neural network (FFANN) with fuzzy multi-scale edge detection (FMED), International Journal of Signal and Imaging Systems Engineering11(5) (2019), 270–278.

17.

Nguyen

Tuan-Linh

, Kavuri

Swathi

, Lee

Minho

, A Fuzzy Convolutional Neural Network for text sentiment analysis, Journal of Intelligent & Fuzzy System, 2018.

18.

Popko

E.A.

and Weinstein

I.A.

, Fuzzy Logic Module of convolutional neural network for Handwriten Digits Recognition, Journal of Physics, 2016.

19.

Sun

C.-C.

, A performance evaluation model by integrating fuzzy AHP and fuzzy TOPSIS methods, Expert systems with Applications37 (2010), 7745–7754.

20.

Manogaran

, Shakeel

P.M.

, Fouad

, Nam

, Baskar

, Chilamkurti

and Sundarasekar

, Wearable IoT Smart-Log Patch: An Edge Computing-Based Bayesian Deep Learning Network System for Multi Access Physical Monitoring System, Sensors19(13) (2019), 3030.

21.

Ross

T.J.

, Fuzzy logic with engineering applications (West Sussex, UK: John Wiley & Sons), 2010.

22.

Meng

, Lu

, Wang

, Li

, Jiang

and Liu

, Encoding Source Language, with Convolutional Neural Networks for machine Translation, 2015.

23.

Mitra

, Fuzzy Multi-layer Perception, Inferencing and Rule Generation, 1995.

24.

, Wu

and Yuan

, Sentiment Analysis with the exploration of overall opinion sentence, in Proceedings of the international joint conference on neural networks, 2017.

25.

G.D.

and Wu

P.-J.

, Type-2 fuzzy neural network for voice activity detection, 2016 International Conference on Fuzzy Theory and Its Applications (Fuzzy), (2016).

26.

Dong

, Sun

, Shi

, et al., A novel valuation pruning optimization fuzzing test model based on mutation tree for industrial control systems, Applied Soft Computing70 (2018), 896–902.

27.

Lanet

J.-L.

, Le Bouder

, Benattou

, et al., When time meets test, International Journal of Information Security17(4) (2018), 395–409.

28.

Xie

, Zhou

, Pan

, et al., An Intercomparison Study of Two Proximate Damped Ly alpha Systems with Residual Flux upon the Ly alpha Absorption Trough toward Quasars, Astrophysical Journal858(1) (2018), 32.

29.

Chen

, Cui

, Ma

, et al., A systematic review of fuzzing techniques, Computers & Security75 (2018), 118–137.

30.

Zhang

, Ye

, Bi

, et al., FFUZZ: Towards full system high coverage fuzz testing on binary executables, Plos One13(5) (2018).