CN112488049A - Fault identification method for foreign matter clamped between traction motor and shaft of motor train unit - Google Patents

一种动车组牵引电机及轴间夹挂异物的故障识别方法，涉及图像处理技术领域，针对现有技术中针对牵引电机及轴间夹挂的异物进行故障检测存在准确性差的问题，包括以下步骤：步骤一：获取货车2D线阵灰度图像；步骤二：根据货车2D线阵灰度图像截取车轴及牵引电机区域子图；步骤三：对车轴及牵引电机区域子图中的异物进行标记；步骤四：将标记后的图像划分为多个区域块，并将划分后的图像作为训练集训练神经网络；步骤五：将待检测图像输入训练好的神经网络中，得到每个区域块的预测得分；步骤六：根据每个区域块的预测得分判定动车组牵引电机及轴间是否夹挂异物。

A fault identification method for a traction motor of an EMU and a foreign object clamped between axles relates to the technical field of image processing. In view of the problem of poor accuracy in fault detection for the foreign object clamped between the traction motor and the axle in the prior art, the method includes the following steps Step 1: Obtain a 2D linear grayscale image of the truck; Step 2: Intercept the sub-images of the axle and traction motor area according to the 2D linear grayscale image of the truck; Step 3: Mark the foreign objects in the sub-images of the axle and the traction motor area; Step 4: Divide the marked image into multiple regional blocks, and use the divided image as a training set to train the neural network; Step 5: Input the image to be detected into the trained neural network to obtain the prediction of each regional block Score; Step 6: Determine whether foreign objects are caught between the traction motor and the axle of the EMU according to the predicted score of each area block.

Fault identification method for foreign matter clamped between traction motor and shaft of motor train unit

Technical Field

The invention relates to the technical field of image processing, in particular to a fault identification method for foreign matters clamped between a traction motor and a shaft of a motor train unit.

Background

In the direction of railway safety, the traditional method is that after a photo is taken by a detection device, the fault point of a train is found through manual observation. This method allows fault detection during vehicle travel without requiring parking. However, the artificial observation has the defects of easy fatigue, high strength, training requirement and the like. More and more things can be replaced by machines at the present stage, and the machines have the characteristics of low cost, unified rule and no fatigue within 24 hours, so that the image recognition technology is used for replacing the traditional manual detection, and the feasibility is realized.

The foreign bodies clamped between the traction motor and the shaft are various in types and different in size. It is difficult to find a common feature using conventional image algorithms. Therefore, fault identification is carried out by using a deep learning neural network method, and the conditions of low accuracy and more false alarms exist.

Disclosure of Invention

The purpose of the invention is: aiming at the problem that the fault detection of foreign matters clamped and hung between a traction motor and a shaft in the prior art is poor in accuracy, the fault identification method for the foreign matters clamped and hung between the traction motor and the shaft of the motor train unit is provided.

The technical scheme adopted by the invention to solve the technical problems is as follows:

a fault identification method for foreign matters clamped between a traction motor and a shaft of a motor train unit comprises the following steps:

the method comprises the following steps: acquiring a 2D linear array gray image of the truck;

step two: intercepting regional subgraphs of the axle and the traction motor according to the 2D linear array gray image of the truck;

step three: marking foreign matters in the intercepted axle and traction motor region subgraphs;

step four: carrying out feature extraction on the marked axle and traction motor regional subgraphs to obtain a feature map, dividing the feature map into a plurality of regional blocks, and taking the feature map obtained after the regional blocks are divided as a training set to train a neural network;

step five: inputting the image to be detected into the trained neural network to obtain the prediction score of each region block corresponding to the image to be detected;

step six: and judging whether foreign matters are clamped between the traction motor of the motor train unit and the shaft corresponding to the image to be detected or not according to the prediction score of each region block corresponding to the image to be detected.

And further, intercepting the regional subgraphs of the axle and the traction motor in the step two according to the prior knowledge and wheelbase information provided by hardware and a frame to carry out regional subgraph interception on the axle and the traction motor.

Further, the core of the neural network is Resnet50, and layer4 output characteristic diagram in Resnet 50.

Further, the specific steps of dividing the feature map into a plurality of area blocks in the fourth step are as follows: firstly, a feature map is divided into four rectangular areas in the length direction of the feature map, the rectangular areas are overlapped to obtain four small area blocks, and then every two adjacent small area blocks are divided into one large area block to obtain three large area blocks.

Further, the training process of the neural network specifically includes: firstly, labeling a feature map obtained by dividing the region blocks, wherein labels are represented as [ x1, x2], wherein x1 represents the probability that foreign matters exist in the region blocks, x2 represents the probability that foreign matters do not exist in the region blocks, according to the divided 7 region blocks, if an object completely falls into one of the region blocks, the label is [1, 0], if the region block does not contain the object, the label is [0, 1], if the region block contains the object but is incomplete, if the proportion of the area of the object in the region block to the total area of the object is more than 0.1, the label of the region block is [0.9, 0.1], if the object exists in the region block but is not complete, the label of the region block is [0.1, 0.9], and finally training the neural network according to the labels and the feature map corresponding to the labels.

Further, the loss function of the neural network is:

label in the above formula₀Is the size of the first element in label, label₁Is the size of the second element in label, pre₁For the size of the first element of the prediction result, pre₀For the size of the second element of the prediction result, N is the number of the foreign objects contained in the 7 region blocks of the training sample, and M is the number of the samples.

Further, the predicted score is obtained through a CSSPPL module in the neural network, and the CSSPPL module specifically executes the following steps:

firstly, carrying out adaptive average pooling, convolution, activation function and convolution processing on features, carrying out adaptive maximum pooling, convolution, activation function and convolution processing on the features, adding the two processing results, activating by using a Sigmoid function to obtain a result F1, multiplying the F1 by the features to obtain a feature I, carrying out convolution and Sigmoid processing on the feature I to obtain F2, multiplying the F2 by the feature I to obtain a feature II, adjusting the size of the feature II by using an SPP module, and processing the feature II processed by the SPP module by using a Linear layer and a Softmax activation function to obtain an area prediction score.

Further, the SPP module resizes feature two to [2048, 16 ].

Further, the specific steps of judging whether foreign matters are clamped between the traction motor of the motor train unit and a shaft or not according to the prediction score of each region block in the sixth step are as follows:

if the prediction probability that foreign matters exist in any one of the prediction results of the 7 area blocks is greater than 0.85, determining that the foreign matters are hung in the clamping mode;

if the prediction probabilities of the foreign matters existing in the prediction results of the 7 area blocks are all smaller than 0.85, judging the 3 large area blocks, if the prediction probability of the foreign matters existing in any one of the 3 large area blocks is larger than 0.6, searching whether an area block with the prediction probability larger than 0.6 exists in two small area blocks contained in the large area block, and if so, judging that the foreign matters are clamped;

and if the conditions are not met, judging the operation to be normal.

Further, in the third step, the label of the foreign matters in the intercepted axle and traction motor region subgraphs is carried out through labelImg.

The invention has the beneficial effects that:

1. the image automatic identification mode is used for replacing manual detection, the fatigue problem that manual detection repeatedly looks at pictures for a long time can be solved, the same fault is unified, and the detection efficiency and the accuracy are improved.

2. The regional classification network is designed, so that the classification accuracy can be increased.

3. A label representation mode when the interception of the sample target is incomplete is designed, and the condition of network overfitting is reduced.

4. And designing a loss function to enable the obtained network to have stronger generalization.

Drawings

FIG. 1 is a schematic diagram of region block partitioning;

FIG. 2 is a schematic diagram of a CSSPPL module;

fig. 3 is an overall flowchart of a fault identification method for a motor train unit traction motor and foreign matter clamped between shafts provided by the application.

Detailed Description

It should be noted that, in the present invention, the embodiments disclosed in the present application may be combined with each other without conflict.

The first embodiment is as follows: specifically describing the embodiment with reference to fig. 3, the method for identifying the fault of the foreign matter clamped between the traction motor of the motor train unit and the shaft of the motor train unit comprises the following steps:

the method comprises the following steps: acquiring a 2D linear array gray image of the truck;

step two: intercepting regional subgraphs of the axle and the traction motor according to the 2D linear array gray image of the truck;

step three: marking foreign matters in the intercepted axle and traction motor region subgraphs;

step four: carrying out feature extraction on the marked axle and traction motor regional subgraphs to obtain a feature map, dividing the feature map into a plurality of regional blocks, and taking the feature map obtained after the regional blocks are divided as a training set to train a neural network;

step five: inputting the image to be detected into the trained neural network to obtain the prediction score of each region block corresponding to the image to be detected;

step six: and judging whether foreign matters are clamped between the traction motor of the motor train unit and the shaft corresponding to the image to be detected or not according to the prediction score of each region block corresponding to the image to be detected.

Raw image data acquisition

And (4) building high-speed imaging equipment at a fixed detection station to obtain a 2D high-definition linear array gray image of the truck. The long-time image collection obtains more sample data size to there are various natural interference such as the image of illumination, rainwater, mud stain etc. in the assurance data image, guarantee the variety of data, the final model has better robustness like this.

Sample image collection

After an original image is obtained, partial images of the axle and the traction motor are intercepted according to the priori knowledge and axle distance information provided by hardware and a frame. According to the size of the shot image, the size of the shot image is fixed, the vertical direction is 832 pixels, the horizontal direction is 256 pixels, the sub-image can contain a fault occurrence area, and calculation is more convenient when the image size is sampled by multiples of 32.

In actual vehicle passing, normal data are far larger than fault data, so fault simulation needs to be carried out manually, a real foreign matter crop database in the whole vehicle is collected, various foreign matters are placed between a traction motor and an axle, and a data set image is constructed.

Sample marking

In the traditional classification network, only the target type needs to be given, but the regional classification network of the invention needs to know the attribute of each region for a training sample, so the position of a foreign object needs to be labeled, and a labelImg is adopted to label a data set.

And (5) correspondingly making the data set images and the labeled data one by one as training samples.

Model building and training

Network construction

According to the characteristics of foreign matters clamped between a traction motor and a shaft (the positions of faults are limited and the change interval is small), if a target detection network is selected for identification, the program operation time is slow, the fault positions are limited, and the target positions do not need to be detected, so that a classification network is selected for identifying whether the foreign matters exist. However, the phenomenon of low accuracy is found in the identification process, and analysis shows that different carriages of different vehicles at different positions between the traction motor and the shaft have differences, the size of the foreign matters is uncertain, and the network cannot well find out definite classification characteristics. Aiming at the characteristic, the invention improves and provides the regional classification network, which can better duplicate the detail part in the view, is more sensitive to foreign matters, has better recognition effect on the foreign matters with different sizes and improves the recognition accuracy rate on the premise of ensuring the operation speed.

In order to avoid repeated feature extraction processes, the algorithm performs region division to obtain different region features when obtaining the last convolution layer of the feature map.

The second embodiment is as follows: the second step is to intercept the axle and traction motor region subgraph according to the prior knowledge and the wheelbase information provided by hardware and a frame.

The third concrete implementation mode: the second embodiment is further described in the present embodiment, and the difference between the second embodiment and the present embodiment is that the core of the neural network is Resnet50, and layer4 output characteristic diagram in Resnet 50.

The algorithm adopts Resnet50 as a backbone and takes the output of layer4 in Resnet50 as a characteristic diagram. And carrying out region division on the feature map to obtain region features.

The fourth concrete implementation mode: this embodiment is a further description of the third embodiment, and the difference between this embodiment and the third embodiment is that the specific steps of dividing the feature map into a plurality of region blocks in the fourth step are: firstly, a feature map is divided into four rectangular areas in the length direction of the feature map, the rectangular areas are overlapped to obtain four small area blocks, and then two adjacent small area blocks are divided into a large area block to obtain three large area blocks.

The characteristic diagram area division is shown in figure 1, a characteristic diagram of [8, 26] is obtained after characteristic extraction, firstly, the characteristic diagram is divided into 4 blocks of 8 x 8 in figure 1 left, the sizes and the shapes of bottles, plastic bags, toilet paper and the like are different due to a plurality of types of foreign matters, and in order to avoid that the characteristics are not obvious because small foreign matters are split into different blocks, two lines of overlapping areas exist between the adjacent blocks. The overlapped area of the two rows corresponds to the original image, namely, the foreign matter with the height of 64 pixels, so that the foreign matter with the height of less than 64 pixels can completely appear in one or two blocks, and the foreign matter with the height of more than 64 pixels can be split into different blocks but contain enough characteristics for classification. This division works better for small targets.

In 4 blocks obtained by dividing the complaint, adjacent blocks are combined to obtain 3 large blocks shown on the right side of the figure 1, and the blocks can completely contain larger foreign object targets and are more favorable for detecting the large targets.

The fifth concrete implementation mode: the fourth embodiment is further described, and the difference between the fourth embodiment and the first embodiment is that the training process of the neural network specifically includes: firstly, label labeling is carried out on a feature map obtained by dividing a region block, label is represented as [ x1, x2], wherein x1 is represented as the probability of foreign matters existing in the region block, x2 is the probability of foreign matters not existing in the region block, according to 7 divided region blocks, label is [1, 0] if an object completely falls into one of the region blocks, label is [0, 1] if the region block does not contain a point object, when the object exists in the region block but is incomplete, the proportion of the area of the object in the region block to the total area of the object is more than 0.1, the label of the region block is [0.9, 0.1], and the label of the region block is [0.1, 0.9], and finally, a neural network is trained according to the label and the feature map corresponding to the label.

Network training

When the area classification network is used, the whole image is divided into 7 areas, although the overlapped areas exist, the possibility of being segmented exists for the original complete target, so the invention provides an incomplete target label representation mode, which gives complete foreign matters and different labels of the incomplete foreign matters to punish the incomplete target so as to train the network and further reduce the over-fitting condition. The label generation method is expressed by using a one-hot form, such as [ x1, x2] where x1 represents the probability of the existence of the foreign object in the block, and x2 represents the probability of the absence of the foreign object in the block. According to 7 regions divided by the sample image, if an object completely falls in one region, the label of the region is [1, 0], if no object is contained in the region, the label of the region is [0, 1], and if the object exists in a certain region but is not complete and is commonly owned by other regions, the label of the region is [0.9, 0.1 ].

The sixth specific implementation mode: the present embodiment is a further description of a fifth embodiment, and the difference between the present embodiment and the fifth embodiment is that the loss function of the neural network is:

label in the above formula₀Is the size of the first element in label, label₁Is the size of the second element in label, pre₁For the size of the first element of the prediction result, pre₀For the size of the second element of the prediction result, N is the number of the foreign objects contained in the 7 region blocks of the training sample, and M is the number of the samples, and 7 is taken.

The loss function is customized, and because the samples have the conditions of unbalanced types and various foreign body samples with different difficulty degrees, the loss function used by the invention is as follows

Label in the above formula₀Is the size of the first element in label, label₁Is the size of the second element in label. pre₁For the size of the first element of the prediction result, pre₀The size of the second element of the prediction. N is the number of pieces of the training sample containing the foreign matter, and M is the number of samples multiplied by 7.

The seventh embodiment: the present embodiment is further described with respect to a sixth embodiment, and the difference between the present embodiment and the sixth embodiment is that the predicted score is obtained by a CSSPPL module in the neural network, and the CSSPPL module specifically executes the following steps:

firstly, carrying out adaptive average pooling, convolution, activation function and convolution processing on features, carrying out adaptive maximum pooling, convolution, activation function and convolution processing on the features, adding the two processing results, activating by using a Sigmoid function to obtain a result F1, multiplying the F1 by the features to obtain a feature I, carrying out convolution and Sigmoid processing on the feature I to obtain F2, multiplying the F2 by the feature I to obtain a feature II, adjusting the size of the feature II by using an SPP module, and processing the feature II processed by the SPP module by using a Linear layer and a Softmax activation function to obtain an area prediction score.

4 small blocks and 3 large blocks are obtained after the region division, and 7 region blocks are obtained. The CSSPPL module was built as shown in fig. 2, with the addition of an attention mechanism, making the image more focused on areas where foreign objects are present. The 7 area blocks are input to the CSAL module, respectively, to obtain a prediction score of whether or not foreign matter is present at the end.

The whole network structure:

the Feature is divided into sub-features of 7 regions by using the output of the resnet50 network layer4 layer as a Feature map Feature extracted by the network. The scores of the seven regions are obtained by inputting the 7 regions into the CSSPPL module. The cssprpl module structure is that firstly, feature is subjected to [ adaptive avgpool2d (1), Conv2d (kernel _ size ═ 3), Relu, Conv2d (kernel _ size ═ 3) ] and [ adaptive maxpool2d (1), Conv2d (kernel _ size ═ 3), Relu, Conv2d (kernel _ size ═ 3) ], after the two results are added, a result obtained by activating using a Sigmoid function is F1, F1 and feature are multiplied to obtain feature1, feature1 is subjected to [ Conv2d (kernel _ size ═ 7), momoid ] to obtain F2, F2 and feature1 are multiplied to obtain feature2, and after the region is divided, a result is added to a region map _ max and a prediction layer is added to obtain a region map _ prediction layer 2048, and a prediction layer is added to obtain a region map _ 3.

The specific implementation mode is eight: the present embodiment is further described with respect to the seventh embodiment, and the difference between the present embodiment and the seventh embodiment is that the SPP module adjusts the size of the second feature to [2048, 16 ].

The specific implementation method nine: the fifth embodiment is further described with respect to a seventh embodiment, and the difference between the fifth embodiment and the seventh embodiment is that the specific step of determining whether the foreign matter is caught between the traction motor of the motor train unit and the shaft according to the prediction score of each zone block in the sixth step is as follows:

if the prediction probability that foreign matters exist in any one of the prediction results of the 7 area blocks is greater than 0.85, determining that the foreign matters are hung in the clamping mode;

if the prediction probabilities of foreign matters existing in the prediction results of the 7 area blocks are all smaller than 0.85, judging the 3 large area blocks, if the prediction probability of the foreign matters existing in any one of the 3 large area blocks is larger than 0.6, searching whether an area block with the probability of predicting the foreign matters larger than 0.6 exists in two small area blocks contained in the large area block, and if so, judging that the foreign matters are clamped;

and if the conditions are not met, judging the operation to be normal.

Fault judgment of foreign matter clamped between traction motor and shaft of motor train unit

When the motor train unit passes through the detection base station, the camera acquires a linear array image. And intercepting partial images of the axle and the traction motor by using prior knowledge, hardware data and the like. And (4) placing the image into a region classification model for prediction to obtain a region classification network prediction score.

If the foreign matter probability predicted by any one of the 7 prediction results is greater than 0.85, outputting the whole image area as an alarm, and uploading the alarm to a platform; or when the probability of the predicted foreign matter in the 3 large blocks is greater than 0.6, searching whether a Block with the probability of the predicted foreign matter greater than 0.6 exists in the two small blocks contained in the 3 large blocks, if so, outputting the whole image area as an alarm, and uploading the alarm to the platform. And if the conditions are not met, judging the operation to be normal. The overall implementation flow chart is shown in fig. 3.

The detailed implementation mode is ten: this embodiment is a further description of the first embodiment, and the difference between this embodiment and the first embodiment is that the label of the foreign matter in the cut-out axle and traction motor region sub-map in step three is performed by labelImg.

It should be noted that the detailed description is only for explaining and explaining the technical solution of the present invention, and the scope of protection of the claims is not limited thereby. It is intended that all such modifications and variations be included within the scope of the invention as defined in the following claims and the description.

1.一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于包括以下步骤：1. a fault identification method for a traction motor of an EMU and a foreign body clamped between axles, is characterized in that comprising the following steps: 步骤一：获取货车2D线阵灰度图像；Step 1: Obtain a 2D linear grayscale image of the truck; 步骤二：根据货车2D线阵灰度图像截取车轴及牵引电机区域子图；Step 2: According to the 2D linear grayscale image of the truck, the sub-image of the axle and traction motor area is intercepted; 步骤三：对截取到的车轴及牵引电机区域子图中的异物进行标记；Step 3: Mark the foreign objects in the sub-images of the intercepted axle and traction motor area; 步骤四：将标记后的车轴及牵引电机区域子图进行特征提取，得到特征图，然后将特征图划分为多个区域块，并将区域块划分后的特征图作为训练集训练神经网络；Step 4: Perform feature extraction on the marked axle and traction motor region sub-maps to obtain a feature map, then divide the feature map into a plurality of regional blocks, and use the feature map divided by the regional blocks as a training set to train a neural network; 步骤五：将待检测图像输入训练好的神经网络中，得到待检测图像对应的每个区域块的预测得分；Step 5: Input the image to be detected into the trained neural network, and obtain the prediction score of each area block corresponding to the image to be detected; 步骤六：根据待检测图像对应的每个区域块的预测得分判定待检测图像对应的动车组牵引电机及轴间是否夹挂异物。Step 6: According to the prediction score of each area block corresponding to the to-be-detected image, determine whether foreign objects are caught between the EMU traction motor and the axles corresponding to the to-be-detected image. 2.根据权利要求1所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述步骤二中截取车轴及牵引电机区域子图根据先验知识及硬件、框架提供的轴距信息进行车轴及牵引电机区域子图截取。2. the fault identification method of a kind of EMU traction motor according to claim 1 and the foreign matter clipping between axles, it is characterized in that in the described step 2, intercepting axle and traction motor area sub-graph according to prior knowledge and hardware, frame The provided wheelbase information is used to intercept the sub-images of the axle and traction motor area. 3.根据权利要求1所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述神经网络的核心为Resnet50，Resnet50中layer4输出特征图。3 . The fault identification method for a traction motor of an EMU and a foreign body clamped between axles according to claim 1 , wherein the core of the neural network is Resnet50, and layer4 in Resnet50 outputs a feature map. 4 . 4.根据权利要求3所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述步骤四中将特征图划分为多个区域块的具体步骤为：首先将特征图以该特征图的长度方向划分为四个矩形区域，所述矩形区域间重叠设置，得到四个小区域块，然后将每相邻的两个小区域块划分为一个大区域块，得到三个大区域块。4. The fault identification method of a traction motor of an EMU and a foreign body clamped between axles according to claim 3, wherein the specific step of dividing the feature map into a plurality of regional blocks in the step 4 is: The feature map is divided into four rectangular regions in the length direction of the feature map, and the rectangular regions are overlapped to obtain four small region blocks, and then each adjacent two small region blocks are divided into a large region block to obtain Three large area blocks. 5.根据权利要求4所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述神经网络的训练过程具体为：首先将区域块划分后的特征图进行标签标记，标签表示为[x1，x2]，其中x1表示为区域块中存在异物的概率，x2为区域块中不存在异物的概率，根据所分的7个区域块，若目标完整落在其中的一个区域块中，则标签为[1，0]，若区域块中不包含目标则标签为[0，1]，当区域块中存在目标但不完整时，若该区域块中目标面积占目标总面积的比例大于0.1，则该区域标签为[0.9，0.1]，否者该区域块的标签为[0.1，0.9]，最后根据标签及该标签对应的特征图训练神经网络。5 . The fault identification method for a traction motor of an EMU and a foreign object clamped between axles according to claim 4 , wherein the training process of the neural network is specifically: firstly, label the feature map after the area block is divided. 6 . Mark, the label is expressed as [x1, x2], where x1 is the probability that there is a foreign body in the area block, x2 is the probability that there is no foreign body in the area block, according to the divided 7 area blocks, if the target completely falls in it In an area block, the label is [1, 0]. If the area block does not contain a target, the label is [0, 1]. When there is a target in the area block but it is incomplete, if the target area in the area block accounts for the target area If the ratio of the total area is greater than 0.1, the label of the area is [0.9, 0.1], otherwise the label of the area block is [0.1, 0.9], and finally the neural network is trained according to the label and the feature map corresponding to the label. 6.根据权利要求5所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述神经网络的损失函数为：6. The fault identification method of a traction motor of an EMU according to claim 5 and a foreign body clamped between axles, wherein the loss function of the neural network is:

上式中label₀为label中第一个元素的大小，label₁为label中第二个元素的大小，pre₁为预测结果第一个元素的大小，pre₀为预测结果第二个元素的大小，N为训练样本的7个区域块中包含异物的个数，M为样本数。In the above formula, label ₀ is the size of the first element in the label, label ₁ is the size of the second element in the label, pre ₁ is the size of the first element of the predicted result, and pre ₀ is the size of the second element of the predicted result. , N is the number of foreign objects contained in the 7 regions of the training sample, and M is the number of samples. 7.根据权利要求6所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述预测得分通过神经网络中的CSSPPL模块得到，所述CSSPPL模块具体执行如下步骤：7. the fault identification method of a kind of EMU traction motor according to claim 6 and inter-axle clamping foreign matter, it is characterized in that described prediction score obtains by the CSSPPL module in the neural network, and described CSSPPL module specifically executes the following steps : 首先对特征进行自适应平均池化、卷积、激活函数、卷积处理，同时对特征进行自适应最大池化、卷积、激活函数、卷积处理，然后将两个处理的结果相加后使用Sigmoid函数进行激活得到结果F1，之后将F1与特征相乘得到特征一，之后对特征一进行卷积、Sigmoid处理得到F2，将F2与特征一相乘得到特征二，然后利用SPP模块将特征二的大小进行调整，然后利用Linear层以及Softmax激活函数对SPP模块处理后的特征二进行处理后得到区域预测得分。First, adaptive average pooling, convolution, activation function, and convolution processing are performed on the features, and adaptive max pooling, convolution, activation function, and convolution processing are performed on the features at the same time, and then the results of the two processing are added. Use the Sigmoid function to activate the result F1, then multiply F1 with the feature to obtain feature 1, then perform convolution on feature 1 and sigmoid processing to obtain F2, multiply F2 and feature 1 to obtain feature 2, and then use the SPP module to convert the feature. The size of 2 is adjusted, and then the feature 2 processed by the SPP module is processed by the Linear layer and the Softmax activation function to obtain the regional prediction score. 8.根据权利要求7所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述SPP模块将特征二的大小调整为[2048，16]。8 . The fault identification method for the traction motor of an EMU and the foreign object caught between the axles according to claim 7 , wherein the SPP module adjusts the size of feature 2 to [2048, 16]. 9 . 9.根据权利要求7所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述步骤六中根据每个区域块的预测得分判定动车组牵引电机及轴间是否夹挂异物的具体步骤为：9. The fault identification method of a traction motor of an EMU and inter-axle clamping foreign objects according to claim 7, wherein in the step 6, the traction motor of the EMU and the inter-axle are determined according to the predicted score of each area block. The specific steps for whether to clamp a foreign body are as follows: 若7个区域块的预测结果中任意一个存在异物的预测概率大于0.85，则判定为夹挂异物；If the predicted probability of the presence of a foreign body in any of the prediction results of the 7 area blocks is greater than 0.85, it is determined that the foreign body is trapped; 若7个区域块的预测结果中存在异物的预测概率都小于0.85，则对3个大区域块进行判定，若3个大区域块中任意一个存在异物的预测概率大于0.6时，则查找该大区域块所包含的两个小的区域块中是否存在预测概率大于0.6的区域块，若存在，则判定为夹挂异物；If the predicted probability of foreign objects in the prediction results of the 7 regional blocks is all less than 0.85, the three large regional blocks are judged. If the predicted probability of foreign objects in any of the three large regional blocks is greater than 0.6, the large Whether there is an area block with a predicted probability greater than 0.6 in the two small area blocks included in the area block, if there is, it is determined as a foreign object trapped; 若以上条件均不满足则判定为正常。If the above conditions are not met, it is judged as normal. 10.根据权利要求1所述的一种动车组牵引电机及轴间夹挂异物的故障识别方法，其特征在于所述步骤三中对截取到的车轴及牵引电机区域子图中的异物进行标记通过labelImg进行。10. A fault identification method for a traction motor of an EMU and a foreign body clamped between axles according to claim 1, characterized in that in said step 3, the foreign body in the intercepted axle and traction motor area sub-picture is marked via labelImg.