Railway wagon release valve pull rod clamping stagnation fault image recognition method

Abstract

The invention discloses a rail wagon release valve pull rod clamping stagnation fault image recognition method, relates to the technical field of graphic image processing, and aims to solve the problem of low release valve fault recognition accuracy caused by the fact that a pull rod is generally recognized as a foot when a release valve foot is recognized in the prior art. According to the method, the deep learning algorithm is applied to automatic recognition of clamping stagnation of the release valve pull rod, misrecognized feet are eliminated by judging the coordinates of the upper left point and the lower right point of the detection frame, the fault is judged through the slope, and the fault recognition accuracy of the release valve is improved.

Description

Technical Field

[0001] The invention relates to the technical field of graphic image processing, and in particular to a method for identifying an image of a railway freight car relief valve tie rod jamming fault. Background Art

[0002] The train braking system is an important part of the train, and its performance and braking capacity are directly related to whether the train can run safely. The semi-automatic relief handle is a device used to operate the truck braking system. It can trigger the semi-automatic relief function of the braking system, which is usually used to reduce the braking force of the vehicle in an emergency, so that the vehicle can move or leave the dangerous area more easily. However, when the relief valve pull rod is stuck, the semi-automatic relief handle will be offset or even fall off, affecting driving safety. In the prior art, when identifying the foot of the relief valve, the pull rod is usually identified as the foot, which leads to a low accuracy rate in identifying relief valve faults. Summary of the invention

[0003] The purpose of the present invention is to propose a method for image recognition of a railway freight car relief valve pull rod jam fault, in view of the problem that in the prior art, when identifying the relief valve foot, the pull rod is usually identified as the foot, resulting in low relief valve fault recognition accuracy.

[0004] The technical solution adopted by the present invention to solve the above technical problems is:

[0005] A method for image recognition of a railway freight car relief valve tie rod stuck fault includes the following steps:

[0006] Acquire a linear array image of the side of a railway freight car containing a relief valve component to be identified, and input the linear array image of the side of a railway freight car containing a relief valve component to be identified into a pre-trained coarse positioning model to obtain a relief valve area sub-image, then input the relief valve area sub-image into a pre-trained fine positioning model to obtain a relief valve leg detection frame, and finally determine the relief valve pull rod stuck fault according to the relief valve leg detection frame;

[0007] The specific steps to determine the release valve tie rod jam fault according to the release valve leg detection frame are as follows:

[0008] Step 1: Determine whether the number of relief valve leg detection frames is greater than 2. If so, obtain the coordinates of the upper left points of all detection frames and calculate the distances between the coordinates of the upper left points of all detection frames. The two detection frames with the closest distance are the final relief valve leg detection frames, and then execute step 3. If not, execute step 2:

[0009] Step 2: Determine whether the number of the relief valve leg detection frames is equal to 2. If so, execute step 3. Otherwise, it is a relief valve pull rod stuck fault.

[0010] Step 3: Obtain the coordinates of the lower right point of the relief valve leg detection frame respectively, and determine whether the inclination angle of the lower right point coordinates of the two relief valve leg detection frames is greater than 10°. If it is greater than, it is a relief valve pull rod stuck fault, otherwise it is not a relief valve pull rod stuck fault.

[0011] Furthermore, the pre-trained coarse positioning model is obtained by the following steps:

[0012] Step 1: Obtain a linear array image of the side of a railway freight car containing a relief valve component, and mark the relief valve area in the image to obtain a relief valve area sub-image, thereby constructing a coarse positioning training set;

[0013] Step 2: Use the data in the coarse positioning training set to train the coarse positioning detection model;

[0014] The pre-trained fine positioning model is obtained through the following steps:

[0015] Step 3: Mark the relief valve legs in the relief valve area sub-image to construct a precise positioning training set;

[0016] Step 4: Use the data in the precise positioning training set to train the precise positioning detection model.

[0017] Furthermore, the specific steps of step one are:

[0018] Acquire a linear array image of the side of a railway freight car including a relief valve component, perform data amplification on the linear array image of the side of a railway freight car including the relief valve component, and then mark the relief valve area in the image after data amplification to obtain a relief valve area sub-image, thereby constructing a coarse positioning training set;

[0019] The specific steps of step three are:

[0020] The relief valve area sub-map is augmented with data, and the augmented relief valve area sub-map is labeled to construct a precise positioning training set.

[0021] Furthermore, the coarse positioning detection model and the fine positioning detection model are Res-DenseNet network models, which include three feature extraction modules, Block1, Block2, Add module, two fully connected layers and softmax;

[0022] The feature extraction module includes two convolutional layers, a pooling layer, an activation function layer, and two batch normalization layers;

[0023] The Res-DenseNet network model specifically performs the following steps:

[0024] After the input image passes through the two feature extraction modules in sequence, it enters Block1 and Block2 respectively. The feature maps output by Block1 and Block2 are added in the Add module. The added feature maps pass through the feature extraction module, fully connected layer, fully connected layer, and softmax in sequence before being output.

[0025] Furthermore, the kernel size of the convolutional layer is 3*3, and the kernel size of the pooling layer is 2*2.

[0026] Furthermore, the Add module specifically performs the following steps:

[0027] The input x path is divided into two. The x in one path is not processed. The x in the other path passes through the convolution layer, BN layer, Relu function, convolution layer, BN layer in sequence, and is added to x. The addition result is output after passing through the Relu function.

[0028] Furthermore, the loss function of the Res-DenseNet network model is expressed as:

[0029]

[0030] Among them, b and Represent the coordinates of the predicted box and the coordinates of the true label, respectively, ciou and λ L1 The coefficient representing the loss weight, L L1 With L ciou They are L1Loss and CIOULoss respectively.

[0031] Furthermore, It is expressed as:

[0032]

[0033] Where N represents the total number of samples, i represents the i-th sample, and b i represents the true value of the i-th sample, Represents the predicted value of the i-th sample.

[0034] Furthermore, It is expressed as:

[0035]

[0036] Among them, IOU represents the intersection-over-union ratio of the predicted box and the true labeled box, that is, the distance between the center point of the predicted box and the center point of the label, q represents the diagonal length of the minimum area of ​​the predicted box and the label, υ is a parameter representing the aspect ratio, α represents the positive weight coefficient, q,h, Represent the width and height of the predicted box and the width and height of the true marked box respectively.

[0037] Furthermore, the activation function of the Res-DenseNet network model is the HSwish function, expressed as:

[0038]

[0039] The beneficial effects of the present invention are:

[0040] This application applies a deep learning algorithm to the automatic identification of a stuck relief valve rod, and by judging the coordinates of the upper left and lower right points of the detection frame, it eliminates the misidentified feet, and by judging the slope, it judges the fault, thereby improving the accuracy of relief valve fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall flow chart of this application;

[0042] Figure 2 This is the structure diagram of the Res-DenseNet network model;

[0043] Figure 3 This is the Add module structure diagram;

[0044] Figure 4 This is a comparison chart of the effects of the ReLU function and the HSwish function;

[0045] Figure 5 This is a schematic diagram of a fault identification embodiment of the present application. DETAILED DESCRIPTION

[0046] It should be particularly noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0047] Specific implementation method 1: refer to Figure 1 Specifically describing this embodiment, a method for image recognition of a railway freight car relief valve tie rod stuck fault in this embodiment includes the following steps:

[0048] Acquire a linear array image of the side of a railway freight car containing a relief valve component to be identified, and input the linear array image of the side of a railway freight car containing a relief valve component to be identified into a pre-trained coarse positioning model to obtain a relief valve area sub-image, then input the relief valve area sub-image into a pre-trained fine positioning model to obtain a relief valve leg detection frame, and finally determine the relief valve pull rod stuck fault according to the relief valve leg detection frame;

[0049] The specific steps to determine the release valve tie rod jam fault according to the release valve leg detection frame are as follows:

[0050] Step 1: Determine whether the number of relief valve leg detection frames is greater than 2. If so, obtain the coordinates of the upper left points of all detection frames and calculate the distances between the coordinates of the upper left points of all detection frames. The two detection frames with the closest distance are the final relief valve leg detection frames, and then execute step 3. If not, execute step 2:

[0051] Step 2: Determine whether the number of the relief valve leg detection frames is equal to 2. If so, execute step 3. Otherwise, it is a relief valve pull rod stuck fault.

[0052] Step 3: Obtain the coordinates of the lower right point of the relief valve leg detection frame respectively, and determine whether the inclination angle of the lower right point coordinates of the two relief valve leg detection frames is greater than 10°. If it is greater than, it is a relief valve pull rod stuck fault, otherwise it is not a relief valve pull rod stuck fault.

[0053] Linear image acquisition

[0054] High-definition equipment is installed at the bottom of the truck track to take pictures of the truck passing at high speed and obtain images of the side of the truck. Line scanning can achieve seamless splicing of images and generate two-dimensional images with a large field of view and high precision.

[0055] Specific implementation method 2: This implementation method is a further explanation of specific implementation method 1. The difference between this implementation method and specific implementation method 1 is that the pre-trained coarse positioning model is obtained by the following steps:

[0056] Step 1: Obtain a linear array image of the side of a railway freight car containing a relief valve component, and mark the relief valve area in the image to obtain a relief valve area sub-image, thereby constructing a coarse positioning training set;

[0057] Step 2: Use the data in the coarse positioning training set to train the coarse positioning detection model;

[0058] The pre-trained fine positioning model is obtained through the following steps:

[0059] Step 3: Mark the relief valve legs in the relief valve area sub-image to construct a precise positioning training set;

[0060] Step 4: Use the data in the precise positioning training set to train the precise positioning detection model.

[0061] Coarse positioning of components

[0062] The stitched truck side images from the S1 and S7 cameras are roughly positioned based on the coarse positioning model, and the local area image containing the relief valve is captured from the large image, which can effectively reduce the time required for fault identification and improve the identification accuracy.

[0063] Create the original sample data set

[0064] Truck parts may be affected by natural or man-made conditions such as rain, mud, oil, and black paint. In addition, images taken at different sites may be different. Therefore, in the process of collecting training image datasets, diversity should be ensured, and images under various conditions and at different sites should be collected as much as possible.

[0065] Each sample image data set includes: an original image set and a label information set. The original image set is a rough positioning image of the relief valve component taken by the device. The label information set is the rectangular sub-region information containing the component, which is obtained by manual labeling. There is a one-to-one correspondence between the original image set and the label information data set, that is, each image corresponds to a label data.

[0066] Specific implementation method three: This implementation method is a further description of specific implementation method two. The difference between this implementation method and specific implementation method two is that the specific steps of step one are:

[0067] Acquire a linear array image of the side of a railway freight car including a relief valve component, perform data amplification on the linear array image of the side of a railway freight car including the relief valve component, and then mark the relief valve area in the image after data amplification to obtain a relief valve area sub-image, thereby constructing a coarse positioning training set;

[0068] The specific steps of step three are:

[0069] The relief valve area sub-map is augmented with data, and the augmented relief valve area sub-map is labeled to construct a precise positioning training set.

[0070] Dataset augmentation

[0071] Although the sample data set includes images under various conditions, it is still necessary to augment the sample data set to improve the stability of the algorithm. Use horizontal flipping to enhance the diversity of the images; adjust the image brightness and contrast to reduce the impact of uneven lighting and highlight the edge features of defects; use histogram equalization to enhance the image contrast; add Gaussian blur and random noise to improve the generalization ability of the model.

[0072] Model Training

[0073] Firstly, the images of the truck at the J station of the S1 and S7 cameras are spliced ​​together to construct a coarse positioning training model dataset, and the coarse positioning detection model is trained. The relief valve area sub-map is obtained through reasoning, and the fine positioning training model dataset is constructed. The fine positioning detection model is trained, and then the semi-automatic relief handle is detected. Figure 1 The model training process.

[0074] Specific implementation method 4: This implementation method is a further explanation of specific implementation method 3. The difference between this implementation method and specific implementation method 3 is that the coarse positioning detection model and the fine positioning detection model are Res-DenseNet network models, and the Res-DenseNet network model includes three feature extraction modules, Block1, Block2, Add module, two fully connected layers and softmax;

[0075] The feature extraction module includes two convolutional layers, a pooling layer, an activation function layer, and two batch normalization layers;

[0076] The Res-DenseNet network model specifically performs the following steps:

[0077] After the input image passes through the two feature extraction modules in sequence, it enters Block1 and Block2 respectively. The feature maps output by Block1 and Block2 are added in the Add module. The added feature maps pass through the feature extraction module, fully connected layer, fully connected layer, and softmax in sequence before being output.

[0078] Res-DenseNet Network

[0079] DenseNet is a convolutional neural network with dense connections. It is an improvement on the research of ResNet. In this network, the input of each layer is the union of the outputs of the previous layers, so each layer can use the information of all the previous layers. Compared with ResNet, it enhances the reuse of bypass features. In addition, the output of the current layer will also be passed to all subsequent layers as input, which can make the network deeper and easier to train, and further alleviate the problems of gradient disappearance and model degradation. Res-DenseNet combines ResNet and DenseNet, integrating the advantages of both, and can fully extract and utilize image features, avoid gradient disappearance, and shorten training time.

[0080] Compared with the traditional convolutional neural network, Res-DenseNet introduces a fusion module, where Block 1 is ResNet and Block 2 is DenseNet. This method can ensure that the network has a certain degree of complexity and obtain a relatively ideal accuracy through training. On the other hand, the feature fusion method is used to add the feature maps output by the two modules to achieve the integrity of feature extraction. The add method is used to increase the information contained in the feature map without increasing the dimension of the feature map. In this way, the amount of information in the feature map can be increased without increasing the dimension.

[0081] Compared with the traditional convolutional neural network, the feature extraction has been enhanced. The improved Res-DenseNet structure is as follows Figure 2 Before being passed to the two blocks, the image will first pass through two modules. Each module contains two convolutional layers, an activation function layer (Rectified Linear Unit, ReLU) and two batch normalization layers (BatchNormalization, BN). The kernel size of the convolutional layer is 3*3, and the kernel size of the pooling layer is 2*2. The Add module structure is as follows: Figure 3 As shown, its purpose is to add the feature maps output by Block1 and Block2, and the addition method is similar to the ResBlock structure in ResNet.

[0082] Specific implementation mode five: This implementation mode is a further explanation of specific implementation mode four. The difference between this implementation mode and specific implementation mode four is that the kernel size of the convolution layer is 3*3, and the kernel size of the pooling layer is 2*2.

[0083] Specific implementation method 6: This implementation method is a further description of the specific implementation method 5. The difference between this implementation method and the specific implementation method 5 is that the Add module specifically performs the following steps:

[0084] The input x path is divided into two. The x in one path is not processed. The x in the other path passes through the convolution layer, BN layer, Relu function, convolution layer, BN layer in sequence, and is added to x. The addition result is output after passing through the Relu function.

[0085] Specific implementation method 7: This implementation method is a further explanation of specific implementation method 6. The difference between this implementation method and specific implementation method 6 is that the loss function of the Res-DenseNet network model is expressed as:

[0086]

[0087] Among them, b and Represent the coordinates of the predicted box and the coordinates of the true label, respectively, ciou and λ L1 The coefficient representing the loss weight, L L1 With λ ciou They are L1Loss and CIOULoss respectively.

[0088] Improvement of loss function

[0089] To further improve the detection effect, CIOULoss is combined with L1Loss. The combined regression loss function can consider multiple angles, make a more accurate estimate of the distance between the predicted box and the true label, and solve the problem of L1Loss gradient non-smoothness. In addition, it also has better convergence. The definition of the joint regression loss function is:

[0090]

[0091] Among them, b and Represent the coordinates of the predicted box and the coordinates of the true label, respectively, ciou and λ L1 The coefficient representing the loss weight, L L1 With L ciou They are L1Loss and CIOULoss loss functions, respectively, and their definitions are:

[0092]

[0093] IOU represents the intersection-over-union ratio of the predicted box and the true labeled box, which represents the distance between the center point of the predicted box and the center point of the label. q represents the diagonal length of the minimum area of ​​the predicted box and the label. υ is a parameter representing the aspect ratio, which is used to measure the consistency of the aspect ratio of the two. α is a positive weight coefficient. The definitions of the two are as follows, where w, h, Represent the width and height of the predicted box and the width and height of the real marked box respectively.

[0094]

[0095] Specific implementation method 8: This implementation method is a further description of specific implementation method 7. The difference between this implementation method and specific implementation method 7 is It is expressed as:

[0096]

[0097] Where N represents the total number of samples, i represents the i-th sample, and b i represents the true value of the i-th sample, Represents the predicted value of the i-th sample.

[0098] Specific implementation method 9: This implementation method is a further description of specific implementation method 8. The difference between this implementation method and specific implementation method 8 is It is expressed as:

[0099]

[0100] Among them, IOU represents the intersection-over-union ratio of the predicted box and the true labeled box, that is, the distance between the center point of the predicted box and the center point of the label, q represents the diagonal length of the minimum area of ​​the predicted box and the label, υ is a parameter representing the aspect ratio, α represents the positive weight coefficient, w, h, Represent the width and height of the predicted box and the width and height of the true marked box respectively.

[0101] Specific implementation method ten: This implementation method is a further explanation of specific implementation method nine. The difference between this implementation method and specific implementation method nine is that the activation function of the Res-DenseNet network model is the HSwish function, which is expressed as:

[0102]

[0103] Improvements in activation functions

[0104] ReLU is a widely used activation function. Through ReLU, in the target detection model, the activation function can set all negative values ​​to 0, while the rest of the values ​​remain unchanged. The ReLU calculation formula is:

[0105] ReLU(x)=max(0,x)

[0106] From the above formula, we can see that if the negative gradient is input into the ReLU function, the output will always be zero, which will cause the negative gradient to be permanently shielded on the function. This phenomenon is also called the "neuron death" problem. Therefore, the HSwish function is introduced, which is defined as:

[0107]

[0108] It can be seen that when the input is negative, the output will no longer be set to 0. The comparison results are as follows Figure 4 As shown in the figure, the HSwish activation function can avoid the problem of ReLU losing information when the input is negative, so that the model can better learn image features.

[0109] Example

[0110] See the flowchart Figure 5 Filtering vehicle models means that the module only recognizes specific vehicle models. The semi-automatic relief handle on the relief valve of some vehicle models is not visible, so there is no need for detection.

[0111] The reasons for splicing the S1 and S7 workstation diagrams have been mentioned above.

[0112] Brightness adjustment and histogram equalization belong to image preprocessing. Some images are too dark, which is not conducive to detection. Therefore, improving the contrast by adjusting the brightness and histogram equalization can provide effective help for subsequent detection.

[0113] The three judgment branches represent different cropping strategies for images of different widths. If the width of the image is greater than 10,000 pixels, 5,000 pixels are taken from the middle to the left and right to obtain a sub-image of 10,000 pixels. If the width of the image is between 8,000 and 10,000 pixels, 4,000 pixels are taken from the middle to the left and right to obtain a sub-image of 8,000 pixels. If the width of the image is less than 8,000 pixels, the image is not cropped.

[0114] The preliminarily processed sub-image is resized to a size of 1600*420 and input into the coarse positioning model.

[0115] The coarse positioning model captures the relief valve area to obtain the relief valve sub-image, and then resizes the image to 512*320.

[0116] Input the relief valve sub-map into the precise positioning model. If two legs of the semi-automatic relief handle are detected, determine whether the inclination angle of the handle is greater than 10° (k is the slope, and the slope corresponding to 10° is 0.176) based on the two legs. If it is greater, an alarm is issued; if more than two legs are identified, determine which two legs are the real legs of the handle based on the coordinates of the upper left corner of the inference result, and repeat the previous calculation; if less than two legs are identified, an alarm is issued directly to avoid missed reports.

[0117] This application changes the pre-trained weights of the detection model. Based on ResNet, the DenseNet network is combined with ResNet to improve the network's feature extraction effect on the semi-automatic relief handle and improve the fault detection effect. This application optimizes the loss function to solve the problem of the initial loss function L1 Loss gradient being not smooth. This application optimizes the activation function and changes the ReLU activation function to the HSwish activation function, which is conducive to the model's learning of image features.

[0118] It should be noted that the specific implementation is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the scope of protection of the rights. Any partial changes made according to the claims and description of the present invention should still fall within the scope of protection of the present invention.

Claims

1. A method for image recognition of a railway freight car relief valve tie rod stuck fault, characterized in that The following steps are involved: Acquire a linear array image of the side of a railway freight car containing a relief valve component to be identified, and input the linear array image of the side of a railway freight car containing a relief valve component to be identified into a pre-trained coarse positioning model to obtain a relief valve area sub-image, then input the relief valve area sub-image into a pre-trained fine positioning model to obtain a relief valve leg detection frame, and finally determine the relief valve pull rod stuck fault according to the relief valve leg detection frame; The specific steps of judging the jamming fault of the relief valve tie rod according to the relief valve leg detection frame are as follows: Step 1: Determine whether the number of relief valve leg detection frames is greater than 2. If so, obtain the coordinates of the upper left points of all detection frames and calculate the distances between the coordinates of the upper left points of all detection frames. The two detection frames with the closest distance are the final relief valve leg detection frames, and then execute step 3. If not, execute step 2: Step 2: Determine whether the number of the relief valve leg detection frames is equal to 2. If so, execute step 3. Otherwise, it is a relief valve pull rod stuck fault. Step 3: Obtain the coordinates of the lower right point of the relief valve leg detection frame respectively, and determine whether the inclination angle of the lower right point coordinates of the two relief valve leg detection frames is greater than 10°. If it is greater than, it is a relief valve pull rod stuck fault, otherwise it is not a relief valve pull rod stuck fault.

2. According to claim 1, a railway freight car relief valve tie rod stuck fault image recognition method is characterized in that The pre-trained coarse positioning model is obtained by the following steps: Step 1: Obtain a linear array image of the side of a railway freight car containing a relief valve component, and mark the relief valve area in the image to obtain a relief valve area sub-image, thereby constructing a coarse positioning training set; Step 2: Use the data in the coarse positioning training set to train the coarse positioning detection model; The pre-trained precise positioning model is obtained by the following steps: Step 3: Mark the relief valve legs in the relief valve area sub-image to construct a precise positioning training set; Step 4: Use the data in the precise positioning training set to train the precise positioning detection model.

3. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 2 is characterized in that The specific steps of step one are: Acquire a linear array image of the side of a railway freight car including a relief valve component, perform data amplification on the linear array image of the side of a railway freight car including the relief valve component, and then mark the relief valve area in the image after data amplification to obtain a relief valve area sub-image, thereby constructing a coarse positioning training set; The specific steps of step three are: The relief valve area sub-map is augmented with data, and the augmented relief valve area sub-map is labeled to construct a precise positioning training set.

4. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 3 is characterized in that The coarse positioning detection model and the fine positioning detection model are Res-DenseNet network models, and the Res-DenseNet network model includes three feature extraction modules, Block1, Block2, Add module, two fully connected layers and softmax; The feature extraction module includes two convolutional layers, a pooling layer, an activation function layer and two batch normalization layers; The Res-DenseNet network model specifically performs the following steps: After the input image passes through the two feature extraction modules in sequence, it enters Block1 and Block2 respectively. The feature maps output by Block1 and Block2 are added in the Add module. The added feature maps pass through the feature extraction module, fully connected layer, fully connected layer, and softmax in sequence before being output.

5. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 4 is characterized in that The kernel size of the convolution layer is 3*3, and the kernel size of the pooling layer is 2*2.

6. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 5 is characterized in that The Add module specifically performs the following steps: The input x path is divided into two. The x in one path is not processed. The x in the other path passes through the convolution layer, BN layer, Relu function, convolution layer, BN layer in sequence, and is added to x. The addition result is output after passing through the Relu function.

7. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 6 is characterized in that The loss function of the Res-DenseNet network model is expressed as: Among them, b and Represent the coordinates of the predicted box and the coordinates of the true label, respectively, ciou and λ L1 The coefficient representing the loss weight, L L1 With L ciou They are L1 Loss and CIOU Loss respectively.

8. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 7 is characterized in that Said It is expressed as: Where N represents the total number of samples, i represents the i-th sample, and b i represents the true value of the i-th sample, Represents the predicted value of the i-th sample.

9. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 8 is characterized in that Said It is expressed as: Among them, IOU represents the intersection-over-union ratio of the predicted box and the true labeled box, that is, the distance between the center point of the predicted box and the center point of the label, q represents the diagonal length of the minimum area of ​​the predicted box and the label, υ is a parameter representing the aspect ratio, and α represents the positive weight coefficient. Represent the width and height of the predicted box and the width and height of the true marked box respectively.

10. The method for image recognition of a railway freight car relief valve tie rod stuck fault according to claim 9, characterized in that The activation function of the Res-DenseNet network model is the HSwish function, which is expressed as:

Images