Urban rail vehicle key component temperature detection system and method based on deep learning

Abstract

The application provides a kind of urban rail vehicle key component temperature detection system and method based on deep learning. Four infrared cameras are used to continuously collect thermal images of key components under the vehicle, including axle box, traction motor and gear box. The corresponding thermal images are labeled and segmented. The segmented images are used for model training. The trained model is used to extract the thermal patterns of key components for detection and analysis. This method addresses the issues of missed detection, image classification, speed adaptability and inaccurate positioning during the detection process of key components of urban rail vehicles. It uses a thermal imager to take continuous shots and combines with a deep learning model. The model is trained using a semantic segmentation algorithm to analyze and identify positive and negative samples. Finally, it can automatically and accurately identify the detected components, classify the images and locate the fault axis.

Description

Technical Field

[0001] The present invention relates to the field of rail transit technology, and in particular to a temperature detection method for key components of urban rail vehicles based on deep learning. Background Art

[0002] The axlebox, gearbox, and traction motor of a vehicle's running gear are critical components that require daily maintenance. Failure in any of these components can result in a serious safety incident. The operating temperature of the equipment can be used as a performance indicator to monitor its performance.

[0003] To ensure the safe operation of these key rotating machinery, an online infrared thermal imager was developed to monitor their temperature. Combined with intelligent methods such as deep learning, this approach has ushered in the era of big data maintenance for key components of the vehicle's running gear, reducing manual maintenance costs. However, the rational collection of data on key components and the accurate location of faults have become challenges hindering intelligent temperature monitoring of key vehicle components.

[0004] Commonly used temperature monitoring methods for key components of the running gear of urban rail vehicles include point temperature measurement and thermal imaging timed snapshot measurement. Point temperature measurement uses infrared laser points to obtain corresponding temperature values, which can easily result in incomplete detection. The hot spots of the tested components of the running gear are usually unevenly distributed in different locations, and point measurement can easily result in omissions of key monitoring areas. The thermal imaging timed snapshot measurement mode relies on the matching of passing magnets and shooting angles to capture snapshots. Although it can overcome the problem of uneven hot spots, it requires a large number of cameras and has serious image omission issues, which also poses hidden risks.

[0005] With the development of deep image learning theory, image screening technology has gradually been applied to the intelligent maintenance of urban rail vehicles. Deep image learning models primarily improve the accuracy of image screening by constructing a reasonable network model, defining relevant input parameters, and training the model using positive and negative samples. Currently, widely used models in deep image learning include supervised convolutional neural networks (CNNs), unsupervised generative adversarial networks (GANs), and deep belief networks (DBNs). These models all continuously improve their image recognition capabilities by inputting a large number of samples, achieving the goal of accurately distinguishing and locating running gear parts under inspection.

[0006] The current urban rail vehicle running gear temperature monitoring detection method adopts point temperature measurement method and timing snapshot two ways, the point temperature measurement method is that the dotting type infrared measuring instrument is installed beside the track, the measurement is started when the vehicle passes through, and a point temperature change graph is drawn, only the temperature of a line on the running gear is detected. The timing snapshot method is that the thermal imager obtains the vehicle array temperature, after the equipment detects that the vehicle passes through the detection area, the thermal imager is started to wait for triggering to shoot the magnetic steel to snap multiple temperature pictures, the highest temperature picture of the shooting position is regarded as the detected component and the related temperature information is saved.

[0007] The temperature detection of the key components of the urban rail vehicle running needs to ensure the missed detection rate and comprehensive coverage of the detected parts, and can accurately locate the specific shaft position in the case of failure. The detection equipment has informationization and intelligentization characteristics, and the detection equipment also has certain economy.

[0008] From the above points, the point temperature detection equipment cannot comprehensively cover the area of the detected parts of the running gear. The key components are the shaft boxes on both sides, the traction motor and the gear box. The heating conditions of each detected part are uneven. The point temperature detection can only measure the temperature of a line, and the missed detection may occur. Even if the detection equipment is increased to make up for this defect, the cost will also be increased. The timing snapshot detection method determines the picture snapshot position through timing and magnetic steel cooperation, but the vehicle does not pass through the detection area at a constant speed. The instability of the passing speed will cause the loss of the corresponding collected pictures, the detection has randomness, and the missed detection rate requirement cannot be met. SUMMARY

[0009] The purpose of the present application is to solve the problems of missed detection, picture loss and high detection cost in the prior art temperature detection of key components of the rail vehicle. A kind of urban rail vehicle key component temperature detection system and method based on deep learning are provided to reduce the cost and improve the accuracy of key component temperature detection.

[0010] To achieve the above purpose, the technical scheme adopted by the present application is:

[0011] A kind of urban rail vehicle key component temperature detection system based on deep learning, comprising:

[0012] First infrared camera and second infrared camera: respectively arranged at the outer side of the two sides of the vehicle running track. Respectively used for continuously collecting the thermal image of the first side shaft box and the thermal image of the second side shaft box of the vehicle bottom;

[0013] Third infrared camera and fourth infrared camera: respectively arranged at the inner side of the two sides of the vehicle running track, respectively used for continuously collecting the thermal image of the first side gear box and the thermal image of the first side traction motor, and the thermal image of the second side gear box and the thermal image of the second side traction motor;

[0014] The data preprocessing unit: acquires thermal images collected by the first infrared camera, the second infrared camera, the third infrared camera and the fourth infrared camera, and performs labeling processing: draws a first marking box, performs frame selection marking in the image collected by the first infrared camera, draws a second marking box, performs frame selection marking in the image collected by the second infrared camera, draws a third marking box and a fourth marking box, and performs frame selection marking in the image collected by the third infrared camera and / or the image collected by the fourth infrared camera, respectively; the first marking box is set according to the first side axle box image coverage area, the second marking box is set according to the second side axle box image coverage area, the third marking box is set according to the traction motor image coverage area, and the fourth marking box is set according to the gear box image coverage area;

[0015] The data training unit: inputs the image marked by the frame selection marking into the convolution network, performs image training, and obtains a training model, which can filter the first side axle box feature image, the second side axle box feature image, the traction motor feature image and the gear box feature image based on the input image;

[0016] The data analysis unit: inputs the image to be analyzed into the trained model, and extracts the first side axle box feature image, the second side axle box feature image, the traction motor feature image and the gear box feature image in the image to be analyzed.

[0017] In some embodiments of the present application, the data preprocessing unit is further configured to group the images collected by the infrared cameras and label the grouped objects.

[0018] In some embodiments of the present application, the data preprocessing unit is further configured to group the images according to the bogie at the bottom of the vehicle, and divide the first side axle box, the traction motor, the gear box and the second side axle box corresponding to each group of bogies into a group.

[0019] In some embodiments of the present application, the data preprocessing unit is further configured to remove interference images collected by each infrared camera.

[0020] In some embodiments of the present application, the image collection direction of the first infrared camera, the second infrared camera, the third infrared camera and the fourth infrared camera forms an angle of 30° with the horizontal plane and is arranged towards the direction of the bottom of the vehicle.

[0021] Some embodiments of the present application further provide a sound track vehicle key component temperature detection method based on deep learning, which is realized by using the temperature detection system described above, and includes the following steps:

[0022] Thermal image acquisition step: acquiring the thermal images of the first side axle box, the first side traction motor and the first side gear box, the second side traction motor and the second side gear box, and the second side axle of the train through each thermal imager;

[0023] The image marking step includes: drawing a first marking box, marking an image in the thermal image collected by the first infrared camera, drawing a second marking box, marking an image in the thermal image collected by the second infrared camera, drawing a third marking box, marking an image in the thermal image collected by the third infrared camera and / or the fourth infrared camera, and drawing a fourth marking box, marking an image in the thermal image collected by the third infrared camera and / or the fourth infrared camera.

[0024] The training step includes: inputting the marked image into a training model to train the model, and the trained model can extract a first side axle box image, a second side axle box image, a first side traction motor image, a second side traction motor image, a first side gear box image and a second side gear box image.

[0025] The image extraction step includes: inputting a to-be-processed thermal image into the trained model to extract an image.

[0026] In some embodiments of the present application, the method further includes the following steps:

[0027] The images of each vehicle are divided into two groups based on the bogie as a reference, and the two groups of images are marked and processed respectively.

[0028] In some embodiments of the present application, the method further includes the following steps: removing the non-vehicle temperature pictures collected by each infrared camera before and after the vehicle.

[0029] In some embodiments of the present application, the method further includes the following steps: after the image extraction, the axle position information and the measured temperature information of the image are bound and stored in a database.

[0030] The method for detecting the temperature of key components of a city rail vehicle based on deep learning has the following advantages:

[0031] (1) The collection mode of the thermal imager is to collect the whole vehicle after the device is triggered, and the pictures are collected at the highest frame rate of the thermal imager, which can ensure that the captured pictures can fully cover the detected components, so that the probability of missing pictures on the collection side is reduced to 0%. The number of vehicle bottom cameras is reduced, which can adapt to a certain range of speed changes. The highest adaptive speed of the vehicle speed depends on the shooting frame rate of the thermal imager. Since it is continuous shooting, the variable speed of the vehicle will not affect image collection.

[0032] (2) The collected full-quantity pictures are screened and positioned by deep learning, which can ensure that the screened pictures can be positioned to the corresponding axle position information, so that the maintenance personnel can directly view the fault positioning on site, and the practicability of the device is improved. The final key detection part is displayed in the form of data and image, which solves the temperature unevenness in device detection, and makes the detection data more convincing.

[0033] (3) can improve the efficiency in the vehicle daily inspection, can improve the artificial detection efficiency, the maintenance personnel only need to check whether there is over temperature or temperature rise alarm component, the vehicle can check the detection condition after returning to the warehouse, reduce the operation time, at the same time reduce the labor consumption.

[0034] (4) the infrared detection camera of surface array type is adopted, different field of view cameras can be configured according to different sizes of the detected parts, the increase of equipment caused by insufficient coverage area of point temperature detection is reduced, and the hardware cost of equipment is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0036] Figure 1 The present application provides a track vehicle key component temperature detection method flow chart;

[0037] Figure 2 The gear box and traction motor thermal image schematic diagram obtained by the single side infrared camera;

[0038] Figure 3 The axle box thermal image schematic diagram obtained by the single side infrared camera;

[0039] Figure 4a The gear box thermal image original drawing collected by the infrared camera;

[0040] Figure 4b The gear box image obtained after segmentation;

[0041] Figure 5a The traction motor thermal image original drawing collected by the infrared camera;

[0042] Figure 5b The traction motor image obtained after segmentation;

[0043] Figure 6a The axle box thermal image original drawing collected by the infrared camera;

[0044] Figure 6b The axle box image obtained after segmentation;

[0045] Figure 7 The present application provides a temperature detection system infrared camera arrangement schematic diagram;

[0046] Figure 8 The model training convergence algorithm schematic diagram.

[0047] In the above figure:

[0048] 101 - first side rail, 102 - second side rail;

[0049] 201 - first infrared camera, 202 - second infrared camera, 203 - third infrared camera, 204 - fourth infrared camera. DETAILED DESCRIPTION

[0050] In order to make the technical problems to be solved by the present application, technical solutions and beneficial effects more clearly, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and not to limit the present application.

[0051] The first embodiment of the present application first provides a key component temperature detection system for urban rail vehicles based on deep learning, which is arranged on the train bottom and walking track. The system continuously collects thermal images of the train bottom components at high frequency, extracts the main images of the components, and analyzes the temperature of the train bottom components.

[0052] Firstly, the types of key components on the train bottom involved in the present application are introduced, including axle box, gear box, and traction motor. Taking actual passing as an example, the vehicle model is B-type vehicle, the number of carriages is 6-formation, the head and tail vehicles are trailers, and the bogie only has an axle box without traction motor and gear box. The middle four carriages are motor cars with axle boxes, traction motors, and gear boxes. The axle boxes of each carriage are evenly distributed, with a total of 8 axle boxes on the left and right, and 48 axle boxes on the whole train to be detected; each power carriage has two traction motors and two gear boxes on each bogie, which are symmetrically distributed, with a total of 16 gear boxes and 16 traction motors on the whole train.

[0053] The train bottom of every two carriages has two bogies, which can be defined as the front first bogie and the rear second bogie. In the present application, the positions of the two bogies are used to classify the key components on the train bottom.

[0054] The arrangement of the axle boxes on the train bottom is as follows: arranged on both sides, with the train bottom track as the reference, the axle boxes on the first side are located outside the first side rail 101, the axle boxes on the second side are located outside the second side rail 102, and the axle boxes on both sides are symmetrically distributed.

[0055] The traction motors and gear boxes are arranged on both sides of the train bottom, located inside the train walking track, and the arrangement on both sides is different. Taking the direction of the train head as the reference: the first side has the traction motor in front and the gear box in back, and the second side has the gear box in front and the traction motor in back.

[0056] The structure of the temperature detection system is as follows.

[0057] REFERENCE Figure 7 The structure of the temperature detection system is as follows.Figure 7 as shown.

[0058] The first infrared camera 201 and the second infrared camera 202 are respectively arranged on the outer sides of the two vehicle walking tracks. They are respectively used for continuously collecting thermal images of the first side axle box and the second side axle box of the vehicle bottom. The first infrared camera 201 and the second infrared camera 202 are arranged horizontally, and can also be arranged in a staggered manner.

[0059] The third infrared camera 203 and the fourth infrared camera 204 are respectively arranged on the inner sides of the two vehicle walking tracks. They are respectively used for continuously collecting thermal images of the first side gear box and the first side traction motor of the vehicle bottom, and the second side gear box and the second side traction motor. The third infrared camera 203 and the fourth infrared camera 204 are arranged horizontally, and can also be arranged in a staggered manner.

[0060] In some embodiments of the present application, the image collection directions of the first infrared camera 201, the second infrared camera 202, the third infrared camera 203 and the fourth infrared camera 204 form an angle of 30° with the horizontal plane, and are arranged towards the direction of the vehicle bottom, which is convenient for shooting.

[0061] The first infrared camera 201 is arranged at a position corresponding to the position of the first side axle box, and is used for collecting images of the first side axle box. The second infrared camera 202 is arranged at a position corresponding to the position of the second side axle box, and is used for collecting images of the second side axle box. The third infrared camera 203 is arranged at a position corresponding to the positions of the first side traction motor and the gear box, and is used for collecting images of the first side traction motor and the gear box. The fourth infrared camera 204 is arranged at a position corresponding to the positions of the second side traction motor and the gear box, and is used for collecting images of the second side traction motor and the gear box. As described above, due to the different arrangement sequences of the traction motors and the gear boxes on the two sides, the image collection sequences of the third infrared camera 203 and the fourth infrared camera 204 are different. For example, the third infrared camera 203 collects images of the traction motor, the gear box, the traction motor and the gear box in a staggered manner, and the fourth infrared camera 204 collects images of the gear box, the traction motor, the gear box and the traction motor in a staggered manner.

[0062] The thermal images collected above are stored in a data storage unit. The collected full-quantity temperature pictures are stored according to different camera arrangement positions. According to the model of the vehicle, the possible corresponding camera positions of the key components of the running gear, such as the axle box, the traction motor and the gear box, are confirmed, and the non-vehicle temperature pictures before and after the vehicle is connected are removed to reduce interference.

[0063] The data preprocessing unit is realized based on a data processor. The thermal images collected by the first infrared camera 201, the second infrared camera 202, the third infrared camera 203 and the fourth infrared camera 204 are subjected to labeling processing.

[0064] Since all points of the picture have temperature, the temperature of the key parts is detected, and only the temperature of the key useful part is extracted. The purpose of the labeling process is to draw the identification box on the collected picture according to the detection points of different infrared cameras and the actual possible display area of the measured object, so as to improve the identification efficiency of the key parts. And the collected picture is divided into detection sets according to the frequency of the measured items.

[0065] The labeling process is realized by the following steps.

[0066] The first marking box is drawn, and the image collected by the first infrared camera 201 is marked by framing. The first marking box is set according to the image coverage area of the first side axle box.

[0067] The second marking box is drawn, and the image collected by the second infrared camera is marked by framing. The second marking box is set according to the image coverage area of the second side axle box.

[0068] The third marking box and the fourth marking box are drawn, and the image collected by the third infrared camera and / or the image collected by the fourth infrared camera is marked by framing. The third marking box is set according to the image coverage area of the traction motor, and the fourth marking box is set according to the image coverage area of the gear box. Since the traction motor and the gear box are staggered, the third marking box and the fourth marking box are used in the labeling process of the images collected by the third infrared camera and the fourth infrared camera 204.

[0069] The first marking box, the second marking box, the third marking box and the fourth marking box are all square boxes. In this embodiment, a square box is used.

[0070] The data training unit inputs the image marked by the marking box into the convolution network for image training, and obtains a training model. The training model can filter the first side axle box feature image, the second side axle box feature image, the traction motor feature image and the gear box feature image based on the input image.

[0071] The data analysis unit: the trained model can be used for subsequent analysis of the image of the key parts of the vehicle. When detection is needed again, the image to be analyzed is input into the trained model, and the first side axle box feature image, the second side axle box feature image, the first side traction motor feature image, the first side gear box, the second side traction motor feature image and the second side gear box feature image in the image to be analyzed are extracted.

[0072] In some embodiments of the present application, the data preprocessing unit is further configured to group the images collected by the infrared camera and label the grouped objects. Specifically, the data preprocessing unit can group the images according to the bogie, and divide the first side axle box, the traction motor, the gear box and the second side axle box corresponding to each group of bogies into a group.

[0073] Specifically, after grouping, in the trailer: the first bogie corresponds to the first group, including the two axle boxes on the first side close to the first bogie and the two axle boxes on the second side close to the first bogie; the second bogie corresponds to the second group, corresponding to the two axle boxes on the first side close to the second bogie and the two axle boxes on the second side close to the second bogie. In the motor car: the first bogie corresponds to the first group, including the two axle boxes on the first side close to the first bogie and the two axle boxes on the second side close to the first bogie, and the first side traction motor and the first side gear box; the second bogie corresponds to the second group, corresponding to the two axle boxes on the first side close to the second bogie and the two axle boxes on the second side close to the second bogie, and the first side gear box and the second side traction motor.

[0074] The second embodiment of the present application further provides a sound track vehicle key component temperature detection method based on deep learning, which is realized by using the temperature detection system of the first embodiment, and the flow is as follows Figure 1 , and specifically includes the following steps.

[0075] S1: thermal image acquisition step: acquiring the thermal images of the first side axle box, the first side traction motor and the first side gear box, the second side traction motor and the second side gear box, and the second side axle of the train through each thermal imager.

[0076] With the highest speed of the vehicle being 40km / h, the highest frequency of the thermal imager is selected to be 50Hz, the front end of the image starts to collect after the vehicle contacts the magnetic steel, and the temperature picture is collected and uploaded at the highest collection frequency. The average number of picture collected by a single infrared camera in a single passing is 1000, and the total number of temperature pictures collected by four cameras is 4000. This method can take pictures of all components under the vehicle due to the high shooting frequency. However, more than 4000 pictures do not have complete key component images on each picture, and therefore the following image labeling and model training process is needed.

[0077] S2: image labeling step: drawing a first marking box, labeling the image in the thermal image collected by the first infrared camera, drawing a second marking box, labeling the image in the thermal image collected by the second infrared camera, drawing a third marking box, labeling the image in the thermal image collected by the third infrared camera and / or the fourth infrared camera, and drawing a fourth marking box, labeling the image in the thermal image collected by the third infrared camera and / or the fourth infrared camera.

[0078] In the image labeling step, the pictures with complete measured parts in the pictures collected in a single passing are selected, the pictures are subjected to binaryzation processing, and then a semantic segmentation algorithm is used for processing to extract the frame features, which are used as the standard. Other pictures are processed according to this step to segment the frame.

[0079] In some embodiments of the present application, in order to facilitate data processing, the following steps are further included: taking the bogie as the reference, the image of each vehicle is divided into two groups, and the two groups of images are labeled respectively. As shown in Reference Example 1, the types of pictures included in each group of images in the motor car or trailer are known, and will not be repeated here.

[0080] The specific implementation of labeling the first group of images corresponding to the first bogie is as follows. The speed of the vehicle passing through the magnetic steel is extracted (the faster the vehicle speed, the fewer the pictures, the slower the vehicle speed, the more the pictures), the actual length of the vehicle is determined according to the vehicle type, and still taking the 6-axle vehicle as an example, each group of pictures is divided into 12 groups of picture sets with the bogie as the dividing point. The actual size of the actual detection component in the image is extracted for labeling (x1, y1), (x2, y2), which respectively correspond to the coordinates of the upper left corner and the lower right corner of the image, and the range is further narrowed. The bogie coordinate system is established, and facing the direction of the vehicle passing through, the picture labeling box coordinates of the left axle box are (180, 38), (350, 260), the picture labeling box coordinates of the right axle box are (80, 30), (250, 260), the traction motor marking box coordinates are (140, 58), (380, 278), and the gear box marking box coordinates are (20, 58), (260, 278). When labeling the second group of images, the traction motor marking box coordinates and the gear box marking box coordinates need to be replaced, and will not be repeated here.

[0081] Reference Figure 4a 、 Figure 4b 、 Figure 5a 、 Figure 5b 、 Figure 6a 、 Figure 6b , respectively shows the photographed pictures and the extracted key components.

[0082] In some embodiments of the present application, the following steps are further included: before step S2, a data denoising preprocessing step is further included. The collected full-quantity temperature pictures are stored according to different camera placement positions, and according to the model of the vehicle, the possible camera positions corresponding to the key components of the running gear, such as the axle box, the traction motor and the gear box, are confirmed, and the non-vehicle temperature pictures before and after the vehicle is removed to reduce interference. The collection of images is shown in Figure 2 , the pictures are continuously collected, and the interference pictures at the start and end of the pictures are removed.

[0083] S3: training step: input the labeled images into the training model, perform model training, and the trained model can extract the first side axle box image, the second side axle box image, the first side traction motor image, the second side traction motor image, the first side gear box image and the second side gear box image.

[0084] The pre-processed temperature picture information is input into a convolutional network. In this case, the FCN full convolutional network is used to establish the network model, and the related parameters are selected as shown in the following table.

[0085]

[0086]

[0087] The convolutional neural network is trained. In this embodiment, 4000 pictures of one pass are selected as the training set, and subsequent passes are used as the test set. The specific situation during training needs to be checked by the target loss function. In this example, the cross-entropy loss function is used for evaluation:

[0088] L = ylogy' + (1-y)log(1-y')

[0089] In the formula, y is the actual detected gear box, traction motor or axle box value, and y' is the predicted value of the model. If the cross-entropy loss is larger, the corresponding gradient is larger, and the training speed is faster. According to the actual training situation, the evaluation result is as follows: Figure 8 The batch-size size is 16, the input image size is 384x288, the learning rate is 0.0001, and the iteration number is 200 epochs. As can be seen from the figure, the training tends to be stable after 119 iterations.

[0090] S4: Image extraction step: input the thermal image to be processed into the trained model to extract the image.

[0091] Specifically, after the training is completed, the images collected in the next pass are input into the trained model, and after the processing of the semantic segmentation algorithm, the complete contour pattern of the intermediate variable of the detection item can be obtained. In all the pass pictures, the pictures with complete features in each group of pictures are accurately selected.

[0092] S5: The axle box, traction motor and gear box temperature pictures selected by the model are bound with the axle position name and temperature information according to the divided picture set. Since the motors and gear boxes of each bogie of the motor train are symmetrically distributed, transposition processing is required. The front three carriages and the rear three carriages of the urban rail vehicle have an axial symmetry relationship, and center symmetry processing is required in the left and right position relationship to ensure the accuracy of positioning.

[0093] S6: Output presentation: the pictures with the selected and determined positions are arranged into JSON format and uploaded to the database for saving and presented on the front end.

[0094] Compared with the prior art, the application discloses a kind of urban rail vehicle key component temperature detection method based on deep learning, the present method is aimed at the problem of missed detection, picture classification, speed adaptability and inaccurate positioning in the detection process of urban rail vehicle key component, proposes to use thermal imager to shoot continuous shooting, cooperates with deep learning model, through the analysis and recognition ability of model to positive and negative samples by semantic segmentation algorithm, finally realizes the functions of automatic and accurate identification of detected component, picture classification and positioning fault axis.

[0095] The above merely describes preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A temperature detection system for key components of urban rail vehicles based on deep learning, characterized in that: include: The first infrared camera and the second infrared camera are respectively arranged on the outside of the vehicle running tracks on both sides, and are used to continuously collect thermal images of the first side axle box and the second side axle box under the vehicle; The third infrared camera and the fourth infrared camera are respectively arranged on the inner side of the vehicle running track on both sides, and are used to continuously collect thermal images of the first side gearbox and the first side traction motor under the vehicle, as well as thermal images of the second side gearbox and the second side traction motor; A data preprocessing unit acquires thermal images captured by the first infrared camera, the second infrared camera, the third infrared camera, and the fourth infrared camera, and performs annotation processing: defining a first marking frame, marking the image captured by the first infrared camera, defining a second marking frame, marking the image captured by the second infrared camera, defining a third marking frame and a fourth marking frame, and marking the image captured by the third infrared camera and / or the image captured by the fourth infrared camera, respectively; The first marking frame is set according to the area covered by the first side axle box image, the second marking frame is set according to the area covered by the second side axle box image, the third marking frame is set according to the area covered by the traction motor image, and the fourth marking frame is set according to the area covered by the gear box image; Data training unit: inputting the marked image into the convolutional network for image training to obtain a training model, wherein the training model can filter the first side axle box feature image, the second side axle box feature image, the traction motor feature image, and the gear box feature image based on the input image; Data analysis unit: inputs the image to be analyzed into the trained model, and extracts the first side axle box feature image, the second side axle box feature image, the traction motor feature image and the gear box feature image from the image to be analyzed.

2. The temperature detection system for key components of urban rail vehicles based on deep learning according to claim 1, characterized in that: The data pre-processing unit is further configured to group the images captured by the infrared camera and perform labeling on the grouped objects.

3. The temperature detection system for key components of urban rail vehicles based on deep learning as claimed in claim 2, characterized in that: The data pre-processing unit is further configured to group the images according to the vehicle bottom bogies, and divide the first side axle box, traction motor, gear box and second side axle box corresponding to each group of bogies into one group.

4. The temperature detection system for key components of urban rail vehicles based on deep learning according to claim 1, characterized in that: The data pre-processing unit is further configured to remove interference images collected by each infrared camera.

5. The temperature detection system for key components of urban rail vehicles based on deep learning according to claim 1, characterized in that: The image acquisition directions of the first infrared camera, the second infrared camera, the third infrared camera and the fourth infrared camera are set at an angle of 30° to the horizontal plane towards the bottom of the vehicle.

6. A method for detecting the temperature of key components of a rail vehicle based on deep learning, implemented based on the temperature detection system according to any one of claims 1 to 5, characterized in that: The following steps are involved: Thermal image acquisition step: using each thermal imager to acquire thermal images of the axle box on the first side, the traction motor and the gear box on the first side, the traction motor and the gear box on the second side, and the axial direction on the second side of the train; Image annotation step: delineating a first marking frame, marking the image in the thermal image captured by the first infrared camera, delineating a second marking frame, marking the image in the thermal image captured by the second infrared camera, delineating a third marking frame, marking the image in the thermal image captured by the third infrared camera and / or the fourth infrared camera, delineating a fourth marking frame, marking the image in the thermal image captured by the third infrared camera and / or the fourth infrared camera; Training step: inputting the labeled images into the training model for model training, wherein the trained model can extract the first side axle box image, the second side axle box image, the first side traction motor image, the second side traction motor image, the first side gear box image, and the second side gear box image; Image extraction step: Input the thermal image to be processed into the trained model to extract the image.

7. The method for detecting temperature of key components of a rail vehicle based on deep learning according to claim 6, characterized in that: Further comprising the steps of: Taking the bogie as the benchmark, the images of each vehicle are divided into two groups, and the two groups of images are labeled separately.

8. The method for detecting temperature of key components of a rail vehicle based on deep learning according to claim 6, wherein: Further comprising the steps of: Remove non-vehicle temperature images collected by each infrared camera before and after the vehicle is picked up.

9. The method for detecting temperature of key components of a rail vehicle based on deep learning according to claim 6, wherein: Further comprising the steps of: After the image is extracted, the axial position information of the image and the measured temperature information are bound and stored in the database.

Images