Abnormal detection method, system and storage medium for loose bolts in subway gearboxes

Abstract

Abnormal detection method, system and storage medium for loose bolts in subway gearboxes. The present invention relates to abnormal detection method, system and storage medium for loose bolts in gearboxes. The purpose of the present invention is to solve the problem of poor stability and accuracy in the identification of loose bolts in gearboxes by existing methods. The specific process is: 1. Establishing a sample data set for training a target detection model; obtaining a trained target detection model; 2. Establishing a sample data set for training an anti-loosening line posture estimation network model; 3. Obtaining a trained anti-loosening line posture estimation network model; 4. Inputting the image to be tested into the trained target detection model to obtain a gearbox bolt sub-image; inputting the gearbox bolt sub-image into the trained anti-loosening line posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolts; judging whether the bolt is faulty based on the coordinates, and outputting the location of the fault if there is a fault. The present invention is used in the field of abnormal detection of subway gearbox bolts.

Description

Technical Field

[0001] The invention relates to a method, a system and a storage medium for detecting abnormality of loosening bolts in a gear box. Background Art

[0002] The anti-loosening line has the function of intuitively displaying the looseness of the bolts and can clearly show the looseness of the bolts. Once the bolts are loose, it may cause the gearbox to loosen, thereby endangering driving safety. In the detection of loose bolt faults, existing methods often use manual inspection images to perform fault detection. Since the detection results will be affected by the subjective factors of the inspectors during the detection process, it is easy to have problems such as missed detection and wrong detection of faults, which affect driving safety. The use of automatic detection based on neural networks instead of manual vehicle safety detection can improve the efficiency of vehicle safety inspections, eliminate interference from human factors, and reduce labor costs. However, whether manual or automatic detection based on neural networks has the problem of poor stability and accuracy in identifying loose gearbox bolts. Summary of the Invention

[0003] The purpose of the present invention is to solve the problem of poor stability and accuracy in identifying loose gear box bolts in existing methods, and to propose an abnormal detection method, system and storage medium for loose subway gear box bolts.

[0004] The specific process of the abnormal detection method for loose bolts in subway gear boxes is as follows:

[0005] Step 1: Scan the moving subway to obtain an image dataset;

[0006] Establish a sample dataset for training the target detection model based on the acquired image dataset;

[0007] Build a target detection model, which is a neural network model;

[0008] Train the target detection model based on the sample data set for training the target detection model to obtain a trained target detection model;

[0009] Step 2: Establish a sample data set for training the anti-loosening wire posture estimation network model; the specific process is as follows:

[0010] Step 21: Input the sample data set of the training target detection model into the trained target detection model to locate the gearbox bolts, and cut out the gearbox bolt sub-graph according to the positioning information;

[0011] Step 22: Amplify the cut gearbox bolt subgraph to obtain an amplified gearbox bolt subgraph dataset;

[0012] Step 23: Label the amplified gearbox bolt subgraph to obtain a label file that corresponds one-to-one with the amplified gearbox bolt subgraph dataset. The amplified gearbox bolt subgraph dataset and the label file that corresponds one-to-one with the amplified gearbox bolt subgraph dataset are used as sample datasets for training the anti-loosening wire posture estimation network model.

[0013] The specific process of marking the amplified gearbox bolt sub-drawing in steps 2 and 3 is as follows:

[0014] The bolts in the gearbox bolt sub-diagram are divided into two cases and marked:

[0015] Case A: Select the anti-loosening line at the center of the bolt head as the first marked point P1, and select the intersection of the anti-loosening line and the edge of the bolt head as the second marked point P2;

[0016] Case B: Select the end of the anti-loosening wire as the first marked point P1, and select the intersection of the anti-loosening wire and the edge of the bolt head as the second marked point P2;

[0017] Step 3: Constructing a network model for estimating the posture of the anti-loosening wire, which is a neural network model;

[0018] Training the anti-loosening line posture estimation network model based on the sample data set for training the anti-loosening line posture estimation network model to obtain a trained anti-loosening line posture estimation network model;

[0019] Step 4: Input the image to be tested into the trained object detection model to obtain the gearbox bolt sub-image;

[0020] Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt;

[0021] Based on the coordinates of the marked points P1 and P2 on the bolt, determine whether the bolt is faulty. If there is a fault, output the fault location.

[0022] Preferably, in step 1, the moving subway is scanned to obtain an image dataset; the specific process is:

[0023] High-definition linear array imaging devices are installed on both sides and at the bottom of the rails. The subway triggers the sensor to start the imaging device to scan the moving subway, obtain high-definition linear array images and store them.

[0024] Preferably, in step 1, a sample dataset for training the target detection model is established based on the acquired image dataset; the specific process is:

[0025] Based on the subway train's wheelbase, vehicle model, and prior knowledge, an image of the gearbox bolt position is captured from the image captured in step 1 to obtain an original image dataset, data amplify the original image dataset, label the amplified dataset, and obtain a label file that corresponds one-to-one to the amplified dataset;

[0026] The amplified dataset and the labeled files corresponding to the amplified dataset are used as sample datasets for training the target detection model.

[0027] Preferably, in step 1, a target detection model is constructed, and the target detection model is a neural network model; the specific process is:

[0028] The target detection model includes the ResNet50 feature extraction network, RPN network and RoIpooling layer;

[0029] The ResNet50 feature extraction network is selected as the backbone network of the target detection model, and the amplified dataset is used as the input of the ResNet50 feature extraction network to obtain the output feature map of the ResNet50 backbone network;

[0030] The output feature map of the ResNet50 backbone network is used as the input of the RPN network, and the RPN network generates candidate boxes;

[0031] The candidate box generated by the RPN network and the feature map output by the ResNet50 backbone network are input into the RoIpooling layer to obtain the candidate box feature map of the gearbox bolt.

[0032] Preferably, in step 1, the target detection model is trained based on the sample data set for training the target detection model to obtain a trained target detection model; the specific process is:

[0033] The OHEM method is used to train the target detection model.

[0034] Preferably, in step 3, a network model for estimating the posture of the anti-loosening wire is constructed, and the network model for estimating the posture of the anti-loosening wire is a neural network model;

[0035] The anti-loosening line posture estimation network model is trained based on the sample data set for training the anti-loosening line posture estimation network model to obtain a trained anti-loosening line posture estimation network model. The specific process is as follows:

[0036] The HRNET network is selected as the backbone network of the anti-loosening wire posture estimation network model. The sample data set for training the anti-loosening wire posture estimation network model is input into the HRNET network, and the coordinates of the marked points P1 and P2 on the bolt are output until convergence to obtain the trained anti-loosening wire posture estimation network model.

[0037] Preferably, in step 4, the image to be tested is input into a trained target detection model to obtain a gearbox bolt sub-image;

[0038] Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt;

[0039] Determine whether the bolt is faulty based on the coordinates of the marked points P1 and P2 on the bolt. If a fault occurs, output the fault location. The specific process is as follows:

[0040] The specific process is:

[0041] Step 41: Install high-definition linear array imaging equipment on both sides and the bottom of the rails to obtain high-definition linear array images; intercept the gearbox position image from the high-definition linear array image based on the wheelbase, vehicle model, and prior knowledge of the subway train;

[0042] Step 42: Input the intercepted gearbox position image into the trained object detection model, output the candidate frame feature map of the gearbox bolt, and obtain the gearbox bolt sub-map;

[0043] Step 43: Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt;

[0044] Step 44

[0045] For the bolt in situation A, determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows:

[0046] P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the angle between points P1 and P2 and the horizontal direction. When θ-θ1<α, the bolt is loose. Otherwise, the bolt is not loose.

[0047] Where θ is the angle between points P1 and P2 and the horizontal direction; θ1 is the reserved angle; α is the loose threshold;

[0048] For the bolts in case B, there are two situations:

[0049] Case 1: Determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows:

[0050] P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the angle between points P1 and P2 and the horizontal direction. When θ-θ2<β, the bolt is loose. Otherwise, the bolt is not loose.

[0051] Among them, θ2 is the reserved angle; β is the loose threshold;

[0052] Case 2: Determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows:

[0053] By judging whether the distance between points P1 and P2 is greater than the pre-stored distance difference, the specific implementation method is as follows: P1 coordinates (x1, y1), P2 coordinates (x2, y2), through the formula Calculate the distance between points P1 and P2. When D-D1>g, the bolt is loose. Otherwise, the bolt is not loose.

[0054] Where D1 is the pre-stored distance between two points; g is the distance threshold;

[0055] If one of Case 1 and Case 2 is judged to be a loose bolt, then the bolt is loose.

[0056] Preferably, if there is a fault, the fault location is output; the specific process is:

[0057] The fault location in the HD linear array image is calculated by mapping the gearbox bolt sub-image to the gearbox position image, and vice versa.

[0058] After calculating the position of the fault in the high-definition linear array image, the fault component information is uploaded to the alarm platform and the fault is displayed on the display interface.

[0059] The subway gearbox bolt loosening anomaly detection system is used to perform the subway gearbox bolt loosening anomaly detection method.

[0060] A storage medium stores at least one instruction, and the at least one instruction is loaded by a processor to run an abnormality detection system for loose bolts in a subway gear box.

[0061] Beneficial effects of the present invention:

[0062] The present invention utilizes an image recognition method to reduce the manual workload of subway vehicle inspection and can significantly improve the accuracy of inspection. The present invention uses a posture estimation method instead of a deep learning segmentation algorithm to automatically identify loose gearbox bolts, thereby improving the stability and accuracy of the overall algorithm. The present invention uses a TopDown method to design a network, first using a Faster R-CNN network framework to locate the bolts, and then using an HRNet network to locate the anti-loosening wire, thereby achieving better detection accuracy. When constructing an anti-loosening wire dataset, the present invention designs a labeling method that can better train a posture estimation model, better locate the anti-loosening wire, and better design monitoring logic when detecting faults, thereby improving detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Flowchart of the present invention;

[0064] Figure 2 For gearbox bolts Figure 1 ;

[0065] Figure 3 For gearbox bolts Figure 2 ;

[0066] Figure 4 For gearbox bolts Figure 1 's labeled diagram;

[0067] Figure 5 For gearbox bolts Figure 2 's labeled diagram. DETAILED DESCRIPTION

[0068] It should be noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0069] Specific implementation method 1: Combination Figure 1 The present embodiment describes the specific process of the method for detecting the loose bolts in a subway gearbox as follows:

[0070] Step 1: Scan the moving subway to obtain an image dataset;

[0071] Establish a sample dataset for training the target detection model based on the acquired image dataset;

[0072] Build a target detection model, which is a neural network model;

[0073] Train the target detection model (gearbox bolt positioning deep learning network model) based on the sample data set for training the target detection model to obtain a trained target detection model;

[0074] Step 2: Establish a sample data set for training the anti-loosening wire posture estimation network model; the specific process is as follows:

[0075] Step 21: Input the sample data set of the training target detection model into the trained target detection model to locate the gearbox bolts, and cut out the gearbox bolt sub-image according to the positioning information;

[0076] Step 22: Amplify the cut gearbox bolt subgraph to obtain an amplified gearbox bolt subgraph dataset;

[0077] Step 23: Label the amplified gearbox bolt subgraph to obtain a label file that corresponds one-to-one with the amplified gearbox bolt subgraph dataset. The amplified gearbox bolt subgraph dataset and the label file that corresponds one-to-one with the amplified gearbox bolt subgraph dataset are used as sample datasets for training the anti-loosening wire posture estimation network model.

[0078] The specific process of marking the amplified gearbox bolt sub-drawing in steps 2 and 3 is as follows:

[0079] The bolts in the gearbox bolt sub-diagram are divided into two cases and marked:

[0080] Case A: Select the anti-loosening line at the center of the bolt head as the first marking point P1, and select the intersection of the anti-loosening line and the edge of the bolt head as the second marking point P2; Figure 2 、 Figure 4 ;

[0081] Case B: Select the end of the anti-loosening wire as the first marking point P1, and select the intersection of the anti-loosening wire and the edge of the bolt head as the second marking point P2; Figure 3 、 Figure 5 ;

[0082] This marking method ensures that each bolt has only two marking points, reducing the false alarm rate of marking points during detection, and ensuring better identifiability of marking points in different bolts, thereby improving the accuracy of the model.

[0083] The design of specific marking points can be three, four, etc. Through experimental comparison, a marking method that only marks two points is designed.

[0084] 1. Bolts are divided into two categories: Figure 2 、 Figure 3 .

[0085] 2 generals Figure 2 Marked as Figure 4 As shown, Figure 3 Marked as Figure 5 The specific method for selecting annotation points is as follows:

[0086] exist Figure 2 Select the vertex of the anti-loosening line at the center of the bolt as the first marking point P1, select the intersection point of the anti-loosening line and the edge of the bolt as P2, and mark these two points.

[0087] exist Figure 3 Select the end of the anti-loosening line as the first marking point P1, select the intersection of the anti-loosening line and the edge of the bolt as P2, and mark these two points.

[0088] Step 3: Constructing a network model for estimating the posture of the anti-loosening wire, which is a neural network model;

[0089] Training the anti-loosening line posture estimation network model based on the sample data set for training the anti-loosening line posture estimation network model to obtain a trained anti-loosening line posture estimation network model;

[0090] Step 4: Input the image to be tested into the trained target detection model (gearbox bolt positioning network model) to obtain the gearbox bolt sub-image;

[0091] Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt;

[0092] Based on the coordinates of the marked points P1 and P2 on the bolt, determine whether the bolt is faulty. If there is a fault, output the fault location.

[0093] Specific embodiment 2: This embodiment differs from specific embodiment 1 in that in step 1, the moving subway is scanned to obtain an image dataset; the specific process is as follows:

[0094] High-definition linear array imaging devices are installed on both sides and at the bottom of the rails. The subway triggers the sensor to start the imaging device to scan the moving subway, obtain high-definition linear array images and store them.

[0095] Other steps and parameters are the same as those in the first embodiment.

[0096] Specific embodiment three: This embodiment differs from specific embodiments one or two in that, in step one, a sample dataset for training the target detection model is established based on the acquired image dataset; the specific process is as follows:

[0097] Based on the subway train's wheelbase, vehicle model, and prior knowledge, an image of the gearbox bolt position is captured from the image captured in step 1 to obtain an original image dataset, data amplify the original image dataset, label the amplified dataset, and obtain a label file that corresponds one-to-one to the amplified dataset;

[0098] The amplified dataset and the labeled files corresponding to the amplified dataset are used as sample datasets for training the target detection model (gearbox bolt positioning deep learning network model).

[0099] Other steps and parameters are the same as those in the first or second embodiment.

[0100] Specific embodiment 4: This embodiment differs from specific embodiments 1 to 3 in that a target detection model is constructed in step 1, and the target detection model is a neural network model; the specific process is as follows:

[0101] The target detection model includes the ResNet50 feature extraction network, RPN network and RoIpooling layer;

[0102] The ResNet50 feature extraction network is selected as the backbone network of the target detection model, and the amplified dataset is used as the input of the ResNet50 feature extraction network to obtain the output feature map of the ResNet50 backbone network;

[0103] The output feature map of the ResNet50 backbone network is used as the input of the RPN network, and the RPN network generates candidate boxes;

[0104] The candidate boxes generated by the RPN network and the feature maps output by the ResNet50 backbone network are input into the RoIpooling layer to obtain the candidate box feature maps of the gearbox bolts for subsequent target classification and positioning.

[0105] The other steps and parameters are the same as those in the first to third embodiments.

[0106] Specific embodiment 5: This embodiment differs from any one of specific embodiments 1 to 4 in that, in step 1, the target detection model is trained based on the sample data set for training the target detection model to obtain a trained target detection model; the specific process is as follows:

[0107] The OHEM method is used to train the target detection model (the number of training times is set to 30 epochs (epoch means that all samples in the training set are trained once). OHEM is a method for mining difficult samples. Its specific implementation is as follows:

[0108] Step 4.1: First, perform a forward propagation of FasterRCNN (input sample passes through fasterrcnn) to obtain a separate loss value for each ROI;

[0109] Step 42: Perform non-maximum suppression (NMS) on each ROI;

[0110] Step 43: Sort each ROI after non-maximum suppression (NMS) by loss value from large to small, and select the top 30% of the loss values ​​from large to small as input for classification and bounding box regression;

[0111] This method can better learn difficult-to-learn features and improve the accuracy of the detection model.

[0112] The other steps and parameters are the same as those in the first to fourth embodiments.

[0113] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that, in step 3, a network model for estimating the posture of the anti-loosening wire is constructed, and the network model for estimating the posture of the anti-loosening wire is a neural network model;

[0114] The anti-loosening line posture estimation network model is trained based on the sample data set for training the anti-loosening line posture estimation network model to obtain a trained anti-loosening line posture estimation network model. The specific process is as follows:

[0115] The HRNET network is selected as the backbone network of the anti-loosening wire posture estimation network model. The sample data set for training the anti-loosening wire posture estimation network model is input into the HRNET network, and the coordinates of the marked points P1 and P2 on the bolt are output until convergence to obtain the trained anti-loosening wire posture estimation network model.

[0116] The anti-loosening line pose estimation network model changes the link between high and low resolutions from series to parallel, and introduces interaction between high and low resolutions to improve model performance.

[0117] Generally, deep learning target recognition methods are used to detect loose bolts, while the present invention uses a posture estimation method for the first time to identify loose bolts by detecting anti-loosening lines.

[0118] The other steps and parameters are the same as those in the first to fifth embodiments.

[0119] Specific embodiment seven: This embodiment differs from any one of specific embodiments one to six in that, in step four, the image to be tested is input into a trained target detection model to obtain a gearbox bolt sub-image;

[0120] Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt;

[0121] Determine whether the bolt is faulty based on the coordinates of the marked points P1 and P2 on the bolt. If a fault occurs, output the fault location. The specific process is as follows:

[0122] Step 41: Install high-definition linear array imaging equipment on both sides and the bottom of the rails to obtain high-definition linear array images; intercept the gearbox position image from the high-definition linear array image based on the wheelbase, vehicle model, and prior knowledge of the subway train;

[0123] Step 42: Input the intercepted gearbox position image into the trained object detection model, output the candidate frame feature map of the gearbox bolt, and obtain the gearbox bolt sub-map;

[0124] Step 43: Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt;

[0125] Step 44

[0126] For situation A ( Figure 2) bolt, based on the coordinates of the marked points P1 and P2 on the bolt, determine whether the bolt is loose; the specific process is:

[0127] P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the angle between points P1 and P2 and the horizontal direction. When θ-θ1<α, the bolt is loose. Otherwise, the bolt is not loose.

[0128] Where θ is the angle between points P1 and P2 and the horizontal direction; θ1 is the reserved angle; α is the loose threshold;

[0129] For situation B ( Figure 3 ) bolts, there are two situations:

[0130] Case 1: Determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows:

[0131] P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the angle between points P1 and P2 and the horizontal direction. When θ-θ2<β, the bolt is loose. Otherwise, the bolt is not loose.

[0132] Among them, θ2 is the reserved angle; β is the loose threshold;

[0133] However, since the bolt loosening cannot be detected by the pre-stored angle θ2 and the looseness threshold β when the deflection angle is around 180°, case 2 is designed;

[0134] Case 2: Determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows:

[0135] By judging whether the distance between points P1 and P2 is greater than the pre-stored distance difference, the specific implementation method is as follows

[0136] P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the distance between points P1 and P2. When D-D1>g, the bolt is loose. Otherwise, the bolt is not loose.

[0137] Where D1 is the pre-stored distance between two points; g is the distance threshold;

[0138] If one of Case 1 and Case 2 is judged to be a loose bolt, then the bolt is loose.

[0139] The other steps and parameters are the same as those in the first to sixth embodiments.

[0140] Specific embodiment eight: This embodiment differs from specific embodiments one to seven in that, if a fault occurs, the fault location is output; the specific process is as follows:

[0141] The fault location in the HD linear array image is calculated by mapping the gearbox bolt sub-image to the gearbox position image, and vice versa.

[0142] After calculating the position of the fault in the high-definition linear array image, the fault component information is uploaded to the alarm platform and the fault is displayed on the display interface.

[0143] The other steps and parameters are the same as those in the first to seventh embodiments.

[0144] Specific embodiment 9: In this embodiment, the system for detecting the abnormality of loose bolts in a subway gear box is used to execute the method for detecting the abnormality of loose bolts in a subway gear box.

[0145] Specific embodiment 10: This embodiment is a storage medium, in which at least one instruction is stored. The at least one instruction is loaded by a processor and runs a subway gearbox bolt loosening anomaly detection system.

[0146] It should be understood that any method described herein may be provided as a computer program product, software, or computerized method, which may include a non-transitory machine-readable medium having instructions stored thereon, the instructions being used to program a computer system or other electronic device. The storage medium may include, but is not limited to, magnetic storage media, optical storage media; magneto-optical storage media including: read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers; or other types of media suitable for storing electronic instructions.

[0147] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for detecting abnormal loosening of subway gearbox bolts, characterized by: The specific process of the method is: Step 1: Scan the moving subway to obtain an image dataset; Establish a sample dataset for training the target detection model based on the acquired image dataset; Build a target detection model, which is a neural network model; Train the target detection model based on the sample data set for training the target detection model to obtain a trained target detection model; The specific process is: The target detection model includes the ResNet50 feature extraction network, RPN network and RoI pooling layer; The ResNet50 feature extraction network is selected as the backbone network of the target detection model, and the amplified dataset is used as the input of the ResNet50 feature extraction network to obtain the output feature map of the ResNet50 backbone network; The output feature map of the ResNet50 backbone network is used as the input of the RPN network, and the RPN network generates candidate boxes; The candidate boxes generated by the RPN network and the feature maps output by the ResNet50 backbone network are input into the RoI pooling layer to obtain the candidate box feature maps of the gearbox bolts; Step 2: Establish a sample data set for training the anti-loosening wire posture estimation network model; the specific process is as follows: Step 21: Input the sample data set of the training target detection model into the trained target detection model to locate the gearbox bolts, and cut out the gearbox bolt sub-graph according to the positioning information; Step 22: Amplify the cut gearbox bolt subgraph to obtain an amplified gearbox bolt subgraph dataset; Step 23: Label the amplified gearbox bolt subgraph to obtain a label file that corresponds one-to-one with the amplified gearbox bolt subgraph dataset. The amplified gearbox bolt subgraph dataset and the label file that corresponds one-to-one with the amplified gearbox bolt subgraph dataset are used as sample datasets for training the anti-loosening wire posture estimation network model. The specific process of marking the amplified gearbox bolt sub-drawing in steps 2 and 3 is as follows: The bolts in the gearbox bolt sub-diagram are divided into two cases and marked: Case A: Select the anti-loosening line at the center of the bolt head as the first marked point P1, and select the intersection of the anti-loosening line and the edge of the bolt head as the second marked point P2; Case B: Select the end of the anti-loosening wire as the first marked point P1, and select the intersection of the anti-loosening wire and the edge of the bolt head as the second marked point P2; Step 3: Constructing a network model for estimating the posture of the anti-loosening wire, which is a neural network model; Training the anti-loosening line posture estimation network model based on the sample data set for training the anti-loosening line posture estimation network model to obtain a trained anti-loosening line posture estimation network model; The specific process is: The HRNET network is selected as the backbone network of the anti-loosening wire posture estimation network model. The sample data set for training the anti-loosening wire posture estimation network model is input into the HRNET network, and the coordinates of the marked points P1 and P2 on the bolt are output until convergence, thus obtaining the trained anti-loosening wire posture estimation network model. Step 4: Input the image to be tested into the trained object detection model to obtain the gearbox bolt sub-image; Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt; Based on the coordinates of the marked points P1 and P2 on the bolt, determine whether the bolt is faulty. If there is a fault, output the fault location.

2. The method for detecting abnormal loosening of subway gearbox bolts according to claim 1, characterized in that: In step 1, the moving subway is scanned to obtain an image dataset. The specific process is as follows: High-definition linear array imaging devices are installed on both sides and at the bottom of the rails. The subway triggers the sensor to start the imaging device to scan the moving subway, obtain high-definition linear array images and store them.

3. The method for detecting abnormal loosening of subway gearbox bolts according to claim 2, characterized in that: In step 1, a sample dataset for training the target detection model is established based on the acquired image dataset; the specific process is as follows: Based on the subway train's wheelbase, vehicle model, and prior knowledge, an image of the gearbox bolt position is captured from the image captured in step 1 to obtain an original image dataset, data amplify the original image dataset, label the amplified dataset, and obtain a label file that corresponds one-to-one to the amplified dataset; The amplified dataset and the labeled files corresponding to the amplified dataset are used as sample datasets for training the target detection model.

4. The method for detecting abnormal loosening of subway gearbox bolts according to claim 3, characterized in that: In step 1, the target detection model is trained based on the sample data set for training the target detection model to obtain a trained target detection model. The specific process is as follows: The OHEM method is used to train the target detection model.

5. The method for detecting abnormal loosening of subway gearbox bolts according to claim 4, characterized in that: In the step 4, the image to be tested is input into the trained target detection model to obtain a gearbox bolt sub-image; Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt; Determine whether the bolt is faulty based on the coordinates of the marked points P1 and P2 on the bolt. If a fault occurs, output the fault location. The specific process is as follows: The specific process is: Step 41: Install high-definition linear array imaging equipment on both sides and the bottom of the rails to obtain high-definition linear array images; intercept the gearbox position image from the high-definition linear array image based on the wheelbase, vehicle model, and prior knowledge of the subway train; Step 42: Input the intercepted gearbox position image into the trained object detection model, output the candidate frame feature map of the gearbox bolt, and obtain the gearbox bolt sub-map; Step 43: Input the gearbox bolt sub-image into the trained anti-loosening wire posture estimation network model to obtain the coordinates of the marked points P1 and P2 on the bolt; Step 44 For the bolt in situation A, determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows: P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the angle between points P1 and P2 and the horizontal direction. When |θ-θ1| < α, the bolt is loose. Otherwise, the bolt is not loose. Where θ is the angle between points P1 and P2 and the horizontal direction; θ1 is the reserved angle; α is the loose threshold; For the bolts in case B, there are two situations: Case 1: Determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows: P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the angle between points P1 and P2 and the horizontal direction. When |θ-θ2| < β, the bolt is loose. Otherwise, the bolt is not loose. Among them, θ2 is the reserved angle; β is the loose threshold; Case 2: Determine whether the bolt is loose based on the coordinates of the marked points P1 and P2 on the bolt. The specific process is as follows: By judging whether the distance between points P1 and P2 is greater than the pre-stored distance difference, the specific implementation method is as follows P1 coordinates (x1, y1), P2 coordinates (x2, y2), by formula Calculate the distance between points P1 and P2. When D-D1>g, the bolt is loose. Otherwise, the bolt is not loose. Where D1 is the pre-stored distance between two points; g is the distance threshold; If one of Case 1 and Case 2 is judged to be a loose bolt, then the bolt is loose.

6. The method for detecting abnormal loosening of subway gearbox bolts according to claim 5, characterized in that: If there is a fault, the fault location is output; the specific process is: The fault location in the HD linear array image is calculated by mapping the gearbox bolt sub-image to the gearbox position image, and vice versa. After calculating the position of the fault in the high-definition linear array image, the fault component information is uploaded to the alarm platform and the fault is displayed on the display interface.

7. Abnormal detection system for loose bolts in subway gearboxes, characterized by: The system is used to execute the abnormality detection method for loosening bolts of a subway gear box according to any one of claims 1 to 6.

8. A storage medium, characterized in that: The storage medium stores at least one instruction, and the processor loads and runs the subway gearbox bolt loosening anomaly detection system according to claim 7.

Images