Damage detection method of automobile hub based on image texture feature

Abstract

With the rapid growth in the number of motor vehicles worldwide, the general public is beginning to attach importance to the quality inspection of wheels before they leave the factory. The current wheel defect detection systems are often cumbersome to operate and have low practical performance. Therefore, this research will use dynamic image segmentation, image texture feature extraction and Back Propagation neural network classification based on wheel image defect feature analysis algorithm to achieve automatic intelligent detection of automotive wheel defects. In this study, an intelligent detection system for automotive wheel defects is also designed, and finally the performance of the detection system is tested experimentally to illustrate its practicality. The experimental results show that the proposed intelligent detection system for automotive wheel defects based on image texture features identifies defects in wheel castings with a correct rate of 96% and a false positive rate of only 2%. This illustrates that the detection system proposed in this study has a high recognition rate and can provide a useful reference for the automotive industry inspection.

Keywords

Image texture features automotive wheels dynamic image segmentation defect detection

1. Introduction

With the rapid development of economy and society and the gradual improvement of people’s living standard, people’s demand for automobiles is also increasing. After the Second World War, aluminium alloy wheels began to be used in ordinary cars [1]. And as an important safety component of the car, the performance and service life of the wheel hub largely determine the safety and reliability of the whole car, and at the same time, it is also concerned by many scholars [2, 3]. Due to the rapid development of digital image processing and intelligent pattern recognition technology, X-ray based automatic damage detection and identification of castings can greatly improve the production efficiency and automation of cast products. At the same time, the method is also widely used in the quality inspection and damage identification technology of automotive wheels [4]. At present, some foreign companies have completed the development of several sets of automatic online wheel inspection systems, however, the programming operation of these software at the early stage of inspection is tedious and complicated, and the accuracy of wheel damage identification cannot meet the requirements, the automatic detection and identification function of the software is basically on hold, and the price of the software is quite expensive. In China, there is currently no complete set of automatic wheel damage detection and identification system applied to engineering projects. A reliable, practical and economical online automatic wheel detection technology is urgently needed at home and abroad [5]. Therefore, this research will be based on image texture features of the automotive wheel damage detection method. Firstly, the image target region is obtained through pre-processing and image segmentation techniques, then the image texture features are extracted, and then the damage is classified through fuzzy neural networks. Finally, the detection results are judged according to relevant criteria. The automotive wheel damage detection method based on image texture features is recognised in terms of its practicality.

There have been several studies abroad on wheel detection. Schaefer and Purschke [6] has proposed the MODAN filter method. The MODAN filter is a median filter and requires adjustment of the filter template. It is used to remove damage structures from the inspection image. As a result, the algorithm has been successfully applied to casting damage detection systems. Kehoe and Parker [7] proposed an intelligent casting damage detection method that can automatically identify the type of casting damage. The damage image is first segmented using an adaptive threshold segmentation algorithm, and then the region is further accurately segmented by mathematical morphology of expansion corrosion. Then some feature parameters are extracted from the segmented damage images. Finally, the damage to the casting is classified by a damage recognition algorithm. The advantage of this method is that it enables automatic detection of casting damage with high accuracy, but it is difficult to build a database of all damage. Mery and Filbert [8] proposed to simulate damage by marking different grey values on real X-ray images. The algorithm acquires damage images at the location of the casting damage multiple times and then performs damage matching in each radiographic image. The algorithm therefore requires a large number of calculations prior to detection.

After researching the relevant studies mentioned above, it can be seen that most wheel detection systems nowadays use manual detection. This results in problems such as low recognition accuracy, complex algorithms and long detection times. Therefore, it is of great significance to design an intelligent recognition system with good performance for the automotive industry.

2. Damage detection method of automobile hub based on image texture feature

2.1 Imaging detection system of automobile hub based on X-ray

There have been several types of casting imaging inspection systems for X-rays that have been widely used, such as image intensifier ray real-time imaging inspection systems, line array ray imaging inspection systems, etc. However, the characteristics of the X-rays and the wheel castings themselves lead to problems such as lower contrast, more texture, and other problems in the generated X-rays wheel images, which eventually cause defects to be difficult to identify. Therefore, this research has adopted the frame integration method for noise reduction of X-ray wheel images to achieve better imaging results. The specific optimised inspection system is composed of an X-ray source, a digital flat panel detector and its control equipment, and an automotive wheel workpiece, as shown in Fig. 1.

Figure 1.

Image detection system of automobile hub based on X-ray.

The X-ray machine emits rays through the automobile wheel hub workpiece and projects them on the scintillator film of the flat panel detector. The acquisition and digitization of the projection image of the automobile wheel hub are completed and transmitted to the industrial computer.

In the X-ray based automobile hub imaging and detection system, there are some problems in the X-ray image of automobile hub, such as high noise, blurred image edge, low spatial contrast between target area and background area. In addition, the structure of hub aluminum castings is complex and there are many types of damage. These adverse factors directly affect the accuracy of hub damage detection and identification [9]. In order to accurately detect the specific location of hub damage and further classify the damage, the hub ray image must be preprocessed to provide a good basis for image processing and pattern recognition. The usual method of image preprocessing is image noise reduction.

2.2 Image denoising method based on high frame integration algorithm

The random noise of X-ray image of automobile wheel hub has a wide spectrum. If the traditional filter is used for filtering and noise reduction, the spectrum of some useful information in the whole automobile wheel hub image will be lost. Therefore, according to the randomness of the signal noise of the X-ray image of the automobile hub, the noise reduction effect can be effectively achieved by using the noise reduction method based high speed frame integral, that is, the superposition and average of multi-frame images, and the loss of some components in the spectrum can be avoided [10, 11].

Suppose a random noise in the image of automobile wheel hub collected in Subsection 2.1 is additive noise, and the output image of automobile wheel hub after denoising is $f({a,b})$ , in which the useful information is $g({a,b})$ and the noise is $m({a,b})$ , then:

$\displaystyle m({a,b})=f({a,b})-g({a,b})$ (1)

The wheel hub of the vehicle to be tested maintains its original position, and $N$ frames of images are collected in a certain period of time. After superposition, it is divide by the acquisition times $n$ to obtain the average value ${f}^{\prime}({a,b})$ of $f({a,b})$ , then:

$\displaystyle{f}^{\prime}({a,b})=\sum\limits_{j=1}^{N}{[{ng_{j}({a,b})+nm_{j}(% {a,b})}]}=\sum\limits_{j=1}^{N}\left({\frac{Ng({a,b})}{n}+\frac{m_{j}({a,b})}{% n}}\right)$ (2)

where $m_{j}({a,b})$ is random noise and $j$ is the number of denoising times. If the value of $n$ is greater than $N$ , it indicates that the random noise in the image of automobile hub has been effectively suppressed.

Then the average signal-to-noise ratio of multi-frame automobile hub images is:

$\displaystyle Q({a,b})^{\prime}=g^{2}({a,b})\left/\left[{\sum\limits_{j=1}^{N}% {m_{j}({a,b})/n}}\right]^{2}=NQ({a,b})\right.$ (3)

The signal-to-noise ratio $Q({a,b})$ of a single-frame automobile hub image is:

$\displaystyle Q({a,b})=g^{2}({a,b})/m^{2}({a,b})$ (4)

where, Eq. (4) considers the characteristics that the noise is uncorrelated and the mean value is zero. Therefore, it can be seen from Eqs (3) and (4) that the noise reduction method based on frame integral after multi-frame average improves the signal-to-noise ratio by $N$ times.

Theoretically, the more the frames superimposed by multiple frames are, the better the noise reduction effect of automobile wheel hub image is [12]. However, in specific experiments, it is found that when the number of superimposed frames reaches a certain value, increasing the number of superimposed frames has little impact on the image quality, mainly because increasing the number of frames also increases the uncertainty of sampling time and image position [13]. While increasing the number of superimposed frames, the time of image acquisition of automobile wheel hub will be correspondingly prolonged. However, from the perspective of production efficiency, the manufacturer hopes that the shorter the time of image acquisition is, the better it is, on the premise of ensuring the image quality of automobile wheel hub, which can effectively improve the detection efficiency of products. Therefore, it is very important to select an appropriate number of superimposed frames [14].

It is necessary to judge whether the number of superimposed frames used in the multi-frame superimposed noise reduction algorithm not only obtains good denoising effect, but also obtains the image of automobile hub in the shortest time. In this paper, the variance value of the image superimposed by each frame number is compared for judgment [15]. Let the gray value of the $h$ -th frame image in the automobile hub image sequence at point $({x,y})$ be $g_{\textit{hxy}}$ , the number of superimposed frames be $M$ , and the average gray value of the automobile hub image obtained after superimposing $H$ frames at point $({x,y})$ be $g_{\textit{Hxy}}$ .

The equation of variance $U_{H}$ is:

$\displaystyle g_{\textit{Hxy}}=\sum\limits_{m=1}^{M}{g_{\textit{hxy}}/M}$ (5) $\displaystyle U_{H}=\sum\limits_{m=1}^{M}{({g_{\textit{Hxy}}-g_{\textit{hxy}}}% )^{2}}/M$ (6)

2.3 Image segmentation method based on inverse coarse entropy thresholding algorithm

For the denoised automobile hub image $f({a,b})$ in Section 2.2, in order to accurately detect the hub damage in the image, it is necessary to segment its target area, so as to improve the detection accuracy of automobile hub target damage. In this paper, the image segmentation method of automobile hub based on reciprocal rough entropy thresholding algorithm is used to segment and obtain the target information of automobile hub.

According to the properties of reciprocal rough entropy, the reciprocal rough entropy takes the maximum value when the uncertainty in image $f({a,b})$ is the largest, that is, when $S_{OT}=S_{BT}=1$ , $BA_{T}$ obtains its maximum value of $\frac{1}{2}$ . $S_{OT}$ and $S_{BT}$ are the roughness of the target and background of the vehicle hub image in turn, and $BA_{T}$ is the reciprocal roughness entropy. For the single threshold segmentation of automobile wheel hub image, it is hoped that the uncertainty of the target and background area after segmentation is the smallest, that is, the roughness of the target and background is $S_{OT}=S_{BT}=0$ . Based on this property, the reciprocal rough entropy can be used to select the segmentation threshold $L$ of image $f({a,b})$ , that is, the best threshold $L^{\ast}$ should be located at the value of the minimum reciprocal rough entropy.

$\displaystyle L^{\ast}=\mathop{\arg\min}\limits_{L}BA_{T}=\mathop{\arg\min}% \limits_{L}\{{[{({S_{OT}+S_{BT}})/({S_{OT}+S_{BT}})^{2}}]/2}\}L$ (7)

The target and background in the image $f({a,b})$ can be approximately represented by rough set. It can be seen from Eq. (7) that at the optimal threshold, $|{{Z}_{T}}|=|{\bar{Z}_{T}}|$ , $|{{A}_{T}}|=|{\bar{A}_{T}}|$ , ${Z}_{T}$ and $\bar{Z}_{T}$ are the upper and lower approximate values of the target respectively, and ${A}_{T}$ and $\bar{A}_{T}$ are the upper and lower approximate values of the background respectively. At this time, $L^{\ast}$ is located at the boundary between the target and the background. The form of reciprocal rough entropy is simpler than logarithmic rough entropy and exponential rough entropy.

Equation (7) is a single threshold reciprocal rough entropy algorithm, which can be extended to the case of multiple thresholds. The threshold $\{{l_{1},l_{2},\ldots,l_{m}}\}$ is set to divide the gray range of the automobile wheel hub image into $m+1$ intervals, and divide the pixels in the image into $m+1$ classes $D_{0},D_{1},D_{2},\ldots,D_{m}$ . The total number of particles is set as $D_{i}({i=0,1,\ldots,m})$ , the lower approximation of class $D_{i}({i=0,1,\ldots,m})$ is defined as ${D}_{i}$ , and the upper approximation is defined as $\bar{D}_{i}$ , then the roughness of class $D_{i}$ is defined as:

$\displaystyle S_{D_{i}}=\frac{|{\bar{D}_{i}}|}{|{\bar{D}_{i}}|}-\frac{|{{D}_{i% }}|}{|{\bar{D}_{i}}|}$ (8)

At this time, the reciprocal rough entropy of the automobile hub image $f({a,b})$ is defined as:

$\displaystyle BA=\frac{1}{m+1}\sum\limits_{i=0}^{m}{\frac{S_{D_{i}}}{1+S_{D_{i% }}}}$ (9)

The best segmentation threshold is at the lowest reciprocal rough entropy.

In the thresholding segmentation algorithm based on rough entropy, firstly, the image $f({a,b})$ needs to be granulated. Different granulation methods directly affect the final image segmentation results. At present, some granulation algorithms have been proposed. In the maximum rough entropy thresholding algorithm, the image $f({a,b})$ is divided into non overlapping particles with equal size, and the particle size is selected as half of the minimum peak width of the image histogram [16]. However, the histogram of image $f({a,b})$ is not a smooth curve, and the peaks in the histogram cannot accurately represent the region in the image. Therefore, in order to select the particle size more accurately, this paper selects the particle size based on the uniformity histogram of the image. The uniformity histogram can describe the consistency region in the image $f({a,b})$ , and each peak represents a kind of uniformity region in the image. In the uniformity histogram, the two peaks are clearly distinguished, and the width of the small peak is more accurate and clear [17]. In this paper, the rough entropy thresholding algorithm uses non overlapping particles to granulate the image. The width of the particle is half of the minimum peak in the uniformity histogram.

After granulating the image $f({a,b})$ , the rough entropy of the image can be calculated. The steps of image segmentation of automobile hub based on reciprocal rough entropy thresholding algorithm are given below.

(1) The uniformity histogram of image $f({a,b})$ is counted, and 1/2 of the minimum peak in the uniformity histogram is selected as the width of particles. After the particle size is determined, the image $f({a,b})$ can be segmented and granulated [18].

(2) The threshold $L$ is initialized, and the initial value of the threshold $L$ is selected as the minimum gray value of the image $f({a,b})$ .

(3) The approximation matrix in the upper and lower approximation sets of the target and background is calculated according to the threshold $L$ , and the reciprocal rough entropy corresponding to the threshold $L$ is calculated.

(4) Repeat step 3 until the threshold $L$ is equal to the maximum gray value of image $f({a,b})$ .

(5) According to Eq. (7), the gray value when the minimum rough entropy is selected as the best segmentation threshold.

(6) According to the optimal threshold $L^{\ast}$ , the pixels in the image are classified to obtain the segmented image.

$\displaystyle f^{\ast}({a,b})=\left\{{{\begin{array}[]{ll}{0,}&{f({a,b})<L^{% \ast}}\\ {255,}&{f({a,b})\geqslant L^{\ast}}\\ \end{array}}}\right.$ (10)

where $V({a,b})$ is the gray value of pixel $({a,b})$ in the denoised image before segmentation; The gray value at the pixel $({a,b})$ of the segmented image is $V^{\ast}({a,b})$ .

2.4 Extraction of image texture features based on incomplete tree wavelet decomposition

2.4.1 Consistency discrimination of image target area

In Section 2.3, the gray distribution of target texture features in the segmented automobile hub image $f^{\ast}({a,b})$ often has strong regularity. In order to reduce the amount of computation of incomplete tree wavelet decomposition and improve the extraction speed of target texture features, the target region of image $f^{\ast}({a,b})$ is uniformly discriminated before wavelet decomposition [19]. The specific implementation method is as follows: divide $f^{\ast}({a,b})$ into four sub bands (upper left, upper right, lower left and lower right). For the four subbands, Prewitt operator is used to convolute in the horizontal and vertical directions of the 3 $\times$ 3 window, to calculate the transformation amount in the horizontal and vertical directions, take the arctangent of their ratio as a direction value of the window, and finally count the proportion of each direction value in the whole subband to describe the direction characteristics of the image. There are four characteristic values in total. If the maximum value of the difference is less than the set threshold, it is considered that the target area of the image $f^{\ast}({a,b})$ has strong consistency, and each sub block is approximately equivalent. Only one subband can be selected to represent the whole image $f^{\ast}({a,b})$ for incomplete wavelet tree structure decomposition, otherwise it needs to be decomposed [20].

2.4.2 Incomplete tree wavelet decomposition

Firstly, $f^{\ast}({a,b})$ is decomposed by wavelet, and four sub-band target texture images are obtained. For each subband target texture image, the average energy $P_{j}$ of each subband is calculated according to the following equation:

$\displaystyle P_{j}=\sum\limits_{a=1}^{N}{\sum\limits_{b=1}^{M}{|{v_{j}({a,b})% }|}}/NM$ (11)

where $N M$ is the size of sub-band target texture; $v_{j}({a,b})$ is the $j$ -th subband coefficient.

Secondly, it is to obtain the maximum value $P_{\max}$ of the average energy of each sub-band at the same scale, and then judge whether $P_{j}<\varepsilon\cdot P_{\max}$ is true, where $\varepsilon$ is a natural number between 0 and 1. The specific selection principle of parameter $\varepsilon$ is to retain the subband with larger average energy and remove the subband with smaller average energy.

2.4.3 Texture feature extraction

The method of decomposing the sub-band containing $f^{\ast}({a,b})$ information selectively can improve the retrieval accuracy and efficiency. However, the structure lacks the capture of directionality when decomposing the image subband, and it is difficult to distinguish textures with similar spatial and frequency domain characteristics and different directionality [21]. Therefore, in addition to extracting the above average energy coefficient, a feature that can characterize the directivity is also used.

After one wavelet decomposition of image $f^{\ast}({a,b})$ or image subband, four subbands are obtained. The ratio of the average energy $P_{HL}$ of vertical subband to the average energy $P_{LH}$ of horizontal subband is defined as the tangent value of the direction feature of the upper subband, which is defined as follows:

$\displaystyle\varpi=\arctan\left.\left({\frac{P_{HL}}{P_{LH}}}\right)\right/2$ (12)

In the incomplete tree wavelet decomposition structure of the image, a total of $m$ subbands are decomposed by wavelet. For each subband, the texture feature $\varpi_{m}$ can be obtained by using Eq. (12).

2.5 A fuzzy neural network-based approach to image damage recognition

The texture feature $\varpi_{m}$ of automobile wheel hub image extracted in Section 2.4 is used as the recognition sample of damage recognition method for automobile wheel hub image based on fuzzy neural network to complete damage recognition of automobile wheel hub.

Since fuzzy reasoning can be completed through the learning of neural network, the fuzzy neural network adopted in this paper consists of three parts:

(1)
Fuzzy quantization function;
(2)
Neural network structure;
(3)
Deblurring part.

Details are shown in Fig. 2.

Figure 2.
Structure diagram of fuzzy neural network.

The first layer of the network is the input layer, and each neuron of the input layer corresponds to an input variable, which is the texture feature of the automobile hub image. The neuron of this layer directly transmits the input texture feature of the automobile hub image to the neuron of the second layer, and its weight is 1. The second layer of the network is the fuzzy quantization layer, which is used to fuzz the input texture feature value of automobile hub image according to the membership function on the defined fuzzy subset, and convert the input texture feature of automobile hub image into the corresponding membership, so as to realize the fuzzification processing. The third layer is the hidden layer, which is used to realize the mapping from the fuzzy value of input variable to the fuzzy value of output variable. The fourth layer is the output layer, and its nodes correspond to the damage mode. The above four-layer fuzzy neural network can effectively realize the mapping from fuzzy input to fuzzy output, and through the learning algorithm of neural network, the mapping can approach the fuzzy nonlinear function relationship [22].

The general BP neural network algorithm is adopted, including two steps: forward propagation and back propagation. BP algorithm belongs to $\gamma$ -learning law and is a guided algorithm. Assuming that there are $m$ learning example samples, the input vector is $\varpi_{m}=({\varpi_{m1},\varpi_{m2},\ldots,\varpi_{mn}})$ , the output vector is $X=({X_{1},X_{2},\ldots,X_{n}})$ , the output of unit $j$ is $Q_{pj}$ , the network weight is $w_{ji}$ , the error signal is $e$ , and the expected output vector is $W_{m}$ . The learning process is as follows:

(1) Initialize the weight $w_{ji}$ and threshold $\mu_{j}$ of the network, and define them as random numbers in the $-$ 1 $\sim$ 1 interval.

(2) Input the learning set $\{{\varpi_{mn},X_{n}}\}$ of texture feature sample of automobile wheel hub image.

(3) The transfer function is a Sigmoid function. If the output of the $j$ -th unit of the $H$ -th layer is $X_{jm}$ and the input of the $j$ -th unit of the $H+1$ layer is $mk_{pj}=\sum{({w_{ji}X_{jm}-\mu_{j}})}$ , the output is:

$\displaystyle X_{jm}=z_{i}({mk_{pj}})=z_{i}\left[{\sum{({w_{ji}X_{jm}-\mu_{j}}% )}}\right]=\frac{1}{1+\exp\left[{-\sum{({w_{ji}X_{jm}-\mu_{j}})}}\right]}$ (13)

where $z_{i}$ is the transfer function.

Calculate network output error:

$\displaystyle e=\sum{({W_{m}-X_{jm}})}^{2}/2$ (14)

(4) If the number of iterations is the maximum, the learning is over, otherwise, the error back propagation is carried out, and enter step (5).

(5) Calculate the error of each unit of the network layer by layer in reverse:

$\displaystyle e=X_{jm}({W_{m}-X_{jm}})$ (15)

(6) Calculate the correction amount of each weight and cell threshold:

$\displaystyle\Delta w_{ji}({m+1})=eX_{jm}+w_{ji}(m)$ (16) $\displaystyle\Delta\mu_{j}({m+1})=e+\mu_{j}(m)$ (17)

Then turn to step (2).

Training the above fuzzy neural network model with 100 training samples and BP neural network algorithm until satisfactory training results are obtained [23]. After the training, given the texture feature test samples of automobile hub to be identified, the network model can give the output of automobile hub damage detection. According to the output value of automobile hub damage detection and its definition (Table 1), the decision result of fuzzy neural network of automobile hub damage detection can be obtained.

Table 1
Definition of pattern classes for output

Output value Loss level

0.5 $\sim$ 1.5 1

1.6 $\sim$ 2.5 2

2.6 $\sim$ 3.5 3

3.6 $\sim$ 4.5 4

3. Experimental analysis

Output value	Loss level
0.5 $\sim$ 1.5	1
1.6 $\sim$ 2.5	2
2.6 $\sim$ 3.5	3
3.6 $\sim$ 4.5	4

In order to test the effect of the method in this paper, the X-ray image data of damaged wheel hub is detected on the computer of Windows XP system. Figure 3 is taken as an example to test the denoising effect of this method, and the results are shown in Fig. 4.

Figure 3.

Image of automobile hub before denoising.

Figure 4.

After denoising by the proposed method.

Comparing Figs 3 and 4, it can be seen that after denoising the automobile hub images of crack damage and pore damage, the image quality of automobile hub can be optimized, the definition can be improved, and the texture details of automobile hub image are highlighted, indicating that the proposed method has a good denoising effect.

Two images of pore damage, two images of contraction cavities damage, two images of shrinkage damage, two images of crack damage and two images of inclusion damage are set. Under this condition, it can analyze the feature extraction effect of the method in this paper on these 10 images of automobile wheel hub, and make statistical and comparative analysis of the test data, as shown in Table 2.

Table 2

Statistics of hub feature extraction results

Damage type	Actual perimeter area ratio	This method extracts the
results	Aspect ratio	This method extracts the
results	Rectangularity	This method extracts the
results	Average contrast	This method extracts the results
Stoma 1	1.1551	1.1550	2.6077	2.6077	0.7027	0.7027	16.0333	16.0333
Stoma 2	1.2055	1.2054	1.8691	1.8691	0.7654	0.7654	10.3232	10.3232
Shrinkage cavity 1	1.3223	1.3222	2.3342	2.3342	0.7421	0.7421	11.2311	11.2311
Shrinkage cavity 2	1.2345	1.2345	2.3542	2.3542	0.8112	0.8112	11.2322	11.2322
Shrinkage porosity 1	1.6544	1.6544	1.9823	1.9822	0.7634	0.7634	11.0992	11.0992
Shrinkage porosity 2	1.2333	1.2333	1.3452	1.3452	0.6953	0.6953	10.2832	10.2832
Inclusion 1	1.0923	1.0923	2.3244	2.3244	0.7234	0.7234	10.2123	10.2123
Inclusion 2	1.3321	1.3321	2.1341	2.1341	0.7434	0.7434	10.2322	10.2322
Crack 1	1.3421	1.3421	1.1232	1.1232	0.7723	0.7723	11.3221	11.3221
Crack 2	1.2032	1.2032	1.0923	1.0923	0.8923	0.8923	11.2453	11.2453

By analyzing the data in Table 2, it can be seen that the method in this paper has a good effect on the feature extraction of these 10 automobile wheel hub images. Only the feature of the actual perimeter area ratio of the automobile wheel hub images for air hole 1, air hole 2 and shrinkage hole 1 and the feature of the aspect ratio of shrinkage porosity 1 have a deviation from the actual value, but the deviation value is 0.0001, so the extraction accuracy is very high.

Combined with the 10 automobile wheel hub images described in Table 2, one image of each damage type is randomly extracted as the detection target of the method in this paper. This method detects the damage after extracting the image texture features and the detection results are shown in Fig. 5.

Figure 5.

Damage detection results of automobile hub by this method.

Figure 6.

Comparison curve of recognition time for two automotive wheel damage detection systems.

As shown in Fig. 5, the detection results of damage quantity and damage level in the damage image of automobile hub by the method in this paper are consistent with the actual situation. This method can accurately detect a variety of damage of automobile hub.

Finally, a comparison experiment was conducted to compare the recognition time of the current more mature VisioBank-based automotive wheel damage recognition system with the image texture feature-based automotive wheel damage detection system proposed in this study, and the resulting comparison results are shown in Fig. 6.

Observing Fig. 6, it can be concluded that the recognition time of the automotive wheel damage detection system based on image texture features proposed in this study in the comparison experiment is 18.8% faster than the VisioBank-based automotive wheel damage recognition system, and the detection time is superior in all damage areas.

4. Discussion

The production of automobile hub needs a very complex process, which generally includes design, melting, degassing, deterioration, refining, die casting, cleaning and other steps. As long as one of the processes or one of the processes is wrong, it will cause casting damage. Air hole, contraction cavity, shrinkage porosity, inclusion and crack are the most common five kinds of damage. Their causes and appearance shapes are different, and the gray characteristics on X-ray images are also very different. All these differences constitute the basis of automobile hub damage detection. They are introduced in detail below.

Porosity is the damage caused by the gas dissolved in the liquid metal of the casting hub when it is cooled and not discharged in time. The water in the casting mold, the volatile matter contained in the binder, the oil stain on the raw materials and tools will be chemically generated at high temperature, but will generate atomic hydrogen, which will be absorbed by the aluminum alloy in the molten state at high temperature; As the solubility of the gas in the casting body decreases sharply with the formation of the gas in the casting body, the solubility of the gas in the casting body decreases sharply. Pores generally appear in the place with large wall thickness of the wheel hub, which has the characteristics of small size, nearly circular shape, scattered and independent existence. When the aluminum liquid begins to solidify, the dissolved gas in the position with small wall thickness continues to separate out from the aluminum alloy melt. In the position with large wall thickness, because the outer surface of the casting has solidified, the dissolved gas in the internal alloy liquid cannot escape out of the casting in time to form pores. The volume of these gases that cannot escape is generally small, and even the parts with large volume are often divided by liquid aluminum to form dispersed pores; The two-dimensional shape of the pores is similar to that of the spherical shape, which is a significant response in the X-ray direction.

Shrinkage cavity and shrinkage porosity are the same type of damage. Both are caused by the volume shrinkage of the alloy during the cooling and solidification of the casting. The large and concentrated holes are called shrinkage holes, and the small and dispersed holes are called shrinkage porosity. The solidification process of casting includes liquid shrinkage, solidification shrinkage and solid shrinkage of alloy; When the sum of liquid shrinkage and solidification shrinkage is greater than solid shrinkage, it is possible to leave holes in the casting, and these holes often appear in the last solidification part of the casting. Shrinkage cavity and porosity usually occur at the uneven wall thickness of the casting, because these places are generally also the last solidification part. The gray scale inside the shrinkage cavity is uniform, the damage area is large, and the edge is flat; The internal gray level of shrinkage porosity is uneven, the damage area is small, and the edge is uneven. Shrinkage porosity is sometimes called small dispersed shrinkage cavity.

Inclusions, also known as foreign matters, are particles different from the composition of the base metal in the casting. They are the general name of various metal and non-metallic inclusions, including high-density foreign matters and low-density foreign matters. In the process of metal melting, pouring and solidification, chemical reactions between liquid metal components or between liquid metal and furnace gas will produce inclusions; When the temperature of molten metal decreases and the solubility decreases, inclusions will be precipitated, and foreign impurities will produce inclusions.

The crack is formed when it is lower than or higher than the solidus temperature of the alloy, and the result of liquid solidification shrinkage. In other words, the cold crack is caused by the casting force acting on the casting exceeding the allowable range of strength or plasticity of the casting itself when the casting is cooled to low temperature. Therefore, increasing certain casting stress and reducing certain strength and plasticity will promote the occurrence of crack damage. The hot crack is produced by hammering the section of the broken sample of the aluminum alloy on the solidus. The characteristics of the crack ray image are in the shape of thin lines and long strips, and generally the two ends are pointed.

5. Conclusion

The aim of this image texture feature based automotive wheel damage detection is an innovative automotive wheel damage detection system currently developed to address the lack of accuracy of current automotive wheel defect automatic detection and identification systems. The research is based on X-ray image pre-processing and image segmentation, followed by defect detection, classification and rating of five common wheel defect types, and the design of an automotive wheel damage detection system combined with neural network algorithms. The experimental results illustrate that the automotive wheel damage detection method based on image texture features proposed in this study achieves a high correct rate of 96% in identifying defects in wheel castings and a false positive rate of only 2%. It shows that the optimised detection method can obtain extremely accurate detection rate detection and can be effectively applied in the automotive industry. However, as automotive wheel damage detection and identification is a comprehensive discipline with a large workload and short time frame, the following aspects still need to be studied in depth in future research work.

(1) An interest-based target region tracking algorithm is designed. The wheel damage detection method proposed in this paper can greatly improve the accuracy of wheel damage detection. However, due to the large error of the domestic mechanical system, the problem of partial deviation of wheel detection often occurs during automatic wheel detection, resulting in the inability to accurately extract the target region of interest during damage detection. Therefore, it is necessary to design the target tracking algorithm for the region of interest in the future.

(2) The fuzzy neural network recognition algorithm needs to be improved. Although the method in this paper can detect common wheel damage very well, it requires a large number of training samples, resulting in a long execution time. Therefore, further research will focus on how to further improve the method and increase the computational speed.

Footnotes

Acknowledgments

This thesis is supported by the enterprise practical training project for young teachers of Higher Vocational Colleges (No.: 2022QYSJ070).

References

Jang

Kang

Lim

. A feasible strain-history extraction method using machine learning for the durability evaluation of automotive parts. J Mech Sci Technol. 2021; 35(11): 5117-5125.

Diao

Jia

Liu

. Structural damage identification using modified Hilbert-Huang transform and support vector machine. J Civil Struct Health Monit. 2021; 11(4): 1155-1174.

Guo

Zhou

Zhang

. Structural nonlinear damage detection using improved Dempster-Shafer theory and time domain model. J Vibroeng. 2019; 21(6): 1679-1693.

Wang

Beck

. Sparse Bayesian factor analysis for structural damage detection under unknown environmental conditions. Mech Syst Signal Process. 2021; 154(11): 107563.

Weng

Zhu

Xia

. A review on dynamic substructuring methods for model updating and damage detection of large-scale structures. Adv Struct Eng. 2020; 23(3): 584-600.

Schaefer

Purschke

. Full automated X-ray system: Reliable and economic. Qualit Aat Zuverl Aassigkeit. 1991; 36(7).

Kehoe

Parker

. An intelligent knowledge based approach for the automated radiographic inspection of castings. NDT&E Int. 1992; 25(1): 23-36.

Mery

Filbert

. Improvement in automated aluminum casting inspection by finding correspondence of potential flaws in multiple radioscopic images. In: Proceedings of the 15th WCNDT. 2000 November. Rome, Italy. 2000. pp. 631-645.

Chang

Kim

Lee

. Hierarchical neural network for damage detection using modal parameters. Struct Eng & Mech. 2019; 70(4): 457-466.

10.

Min

. Simulation of wall structure damage detection based on model parameter identification. Comput Simul. 2020; 37(11): 404-407+416.

11.

Jayachitra

Priyadarshini

. Smart monitoring system for detection of damage in structural parts by EMI and ANSYS. Int J Eng Trends Technol. 2021; 69(2): 134-138.

12.

Zhou

Sun

. A variable-tensioned transfer matrix method for modal analysis of marine risers and its application to structural damage detection. KSCE J Civil Eng. 2021; 25(6): 2151-2165.

13.

Chen

Pan

Chen

. Structural damage detection via combining weighted strategy with trace Lasso. Adv Struct Eng. 2019; 22(3): 597-612.

14.

Feng

Zhang

Wang

. Structural damage detection using deep convolutional neural network and transfer learning. KSCE J Civil Eng. 2019; 23(10): 4493-4502.

15.

Ghasemkheili

Davoodi

. Structural damage detection using frequency response function. Int J Eng Technol. 2019; 11(6): 37-59.

16.

Yue

Chen

. Comparative studies on structural damage detection using Lp norm regularisation. Int J Lifecycle Perform Eng. 2019; 3(2): 171-186.

17.

Nie

Guo

. Structural damage detection using wavelet packet transform combining with principal component analysis. Int J Lifecycle Perform Eng. 2019; 3(2): 149-170.

18.

Teng

Chen

Gong

. Structural damage detection using convolutional neural networks combining strain energy and dynamic response. Mecc. 2020; 55(4): 945-959.

19.

Miao

Yang

Wang

. Study of multiple-site structural damage detection application using vibration response. Int J Appl Electromagn Mech. 2020; 64(1-4): 1411-1419.

20.

Zhao

Yan

Yang

. A novel artificial bee colony algorithm for structural damage detection. Adv Civil Eng. 2020; (6): 1-21.

21.

. Photogrammetry-based structural damage detection by tracking a visible laser line. Struct Health Monit. 2020; 19(1): 322-336.

22.

Lin

Zhan

. Experimental investigation on multi-objective multi-type sensor optimal placement for structural damage detection. Struct Health Monit. 2019; 18(3): 882-901.

23.

Fathnejat

Ahmadi-Nedushan

. An efficient two-stage approach for structural damage detection using meta-heuristic algorithms and group method of data handling surrogate model. Front Struct Civil Eng. 2020; 14(4): 907-929.