Locomotive wheel automatic measurement system and method based on image recognition

Abstract

The invention discloses a locomotive wheel automatic measurement system and method based on image recognition, and the method comprises the following steps: S1, rotating a wheel, adjusting the position of equipment, and focusing the side surface, tread and hole inner surface of the wheel; s2, acquiring an image and point cloud data, preprocessing and fusing, and generating a three-dimensional feature data set; s3, searching and optimizing a network structure by using a neural architecture, and extracting wheel features; s4, embedding a focused attention network, and enhancing the features of each region; s5, wheel segmentation, size extraction and defect identification are carried out based on the enhanced features; and S6, judging whether the wheel is qualified or not, and outputting measurement parameters and defect analysis results. According to the invention, by combining the neural architecture search and the focused attention network, automatic and efficient measurement and defect detection of locomotive wheels are realized, and the detection precision and efficiency are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of automatic locomotive wheel measurement technology, and in particular to an automatic locomotive wheel measurement system and method based on image recognition. Background Technology

[0002] With the rapid development of modern railway transportation, the inspection and maintenance of locomotive wheels has become a crucial link in ensuring the safe operation of trains. Traditional wheel inspection methods mostly rely on manual operation, assessing wheel dimensions, wear, and defects through manual measurement and visual inspection. While these methods were effective under past technological conditions, their efficiency and accuracy have gradually become inadequate for modern high-speed railways due to increased railway transport loads and operating speeds. Manual inspection is not only labor-intensive but also susceptible to human factors, leading to significant measurement errors. Furthermore, it cannot achieve real-time, comprehensive wheel condition monitoring and data recording, and lacks early warning of potential risks.

[0003] In recent years, with the development of computer vision and deep learning technologies, automated detection technology has been gradually applied to wheel inspection, using image recognition and deep network models to automatically measure various parameters of the wheel. However, existing automated detection methods still face some problems. While most methods have made initial progress in wheel size extraction, they often neglect the fine-grained processing of features in different areas of the wheel (such as the rim, tread, and bores), resulting in insufficient detection accuracy under complex working conditions, especially when wear is severe or there are defects such as micro-cracks on the surface. In addition, existing deep learning models usually require a large amount of data for training, and the model search process is time-consuming, lacking adaptability to different wheel types and measurement tasks.

[0004] Existing technologies for wheel inspection suffer from drawbacks such as low efficiency, insufficient accuracy, and poor adaptability, failing to meet the high requirements of high-speed railways and heavy-haul transportation.

[0005] Therefore, how to provide an automatic locomotive wheel measurement system and method based on image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an automatic locomotive wheel measurement system and method based on image recognition. This invention fully utilizes neural architecture search algorithms and focused attention mechanisms, detailing how to automatically optimize the deep network structure in wheel measurement tasks for feature enhancement and defect detection. Through adaptive network structure search and attention enhancement techniques, the size, wear, and defects of various regions of the wheel are accurately identified. This invention possesses the advantages of high precision, high efficiency, and strong adaptability, achieving real-time wheel condition monitoring, reducing manual intervention, and maintaining good detection performance even under complex working conditions.

[0007] An automatic locomotive wheel measurement method based on image recognition according to an embodiment of the present invention includes the following steps:

[0008] S1. Drive the wheel to rotate at a constant speed using a wheel rotation device, and adjust the positions of the wheel side scanning device, wheel tread scanning device and wheel hole scanning device by moving and adjusting the device respectively, so as to focus on the side, tread and inner surface of the wheel to be tested.

[0009] S2. Simultaneous laser scanning or visual imaging is performed on the wheel side, tread and hole surfaces using wheel side scanning equipment, wheel tread scanning equipment and wheel hole scanning equipment respectively to obtain multi-view images or point cloud data of the wheel under test, and preprocessing and fusion are performed to generate a unified wheel three-dimensional feature dataset.

[0010] S3. Using the neural architecture search algorithm with the wheel 3D feature dataset as input, automatically search and optimize the deep network structure and hyperparameters suitable for the automatic extraction of wheel features, and obtain the optimized deep feature extraction network structure.

[0011] S4. Embed the focus attention network into the deep feature extraction network structure, and use the focus attention network to enhance the features of the wheel side edge region, tread wear region and wheel hole inner diameter boundary region, and output enhanced features.

[0012] S5. Based on enhanced features, automatic segmentation and dimensional parameter extraction of the wheel side profile, tread wear area and wheel bore inner diameter boundary are realized respectively, and the defect area of ​​the wheel surface is automatically identified.

[0013] S6. Based on the extracted dimensional parameters and defect identification results, automatically determine whether the wheel is qualified, and output the final wheel measurement parameters and defect analysis results.

[0014] Optionally, S2 specifically includes:

[0015] S21. Start the wheel side scanning device, wheel tread scanning device and wheel hole scanning device respectively. Driven by the wheel rotation device, perform synchronous laser scanning or visual imaging on the side, tread and inner surface of the hole of the wheel to be tested to obtain multi-view images or point cloud data, and mark the source of each type of data respectively.

[0016] S22. Perform noise suppression and filtering on the multi-view images or point cloud data of the side, tread and inner surface of the hole, including the removal of environmental noise, scanning error and invalid information from background interference.

[0017] S23. Perform spatial registration on the processed multi-view images or point cloud data according to the spatial geometric features of each detected part of the wheel, and convert the data obtained from different perspectives and different devices into the same spatial coordinate system.

[0018] S24. The multi-view images or point cloud data after spatial registration are fused, and the data information of different detection devices and detection angles are comprehensively utilized to generate a three-dimensional feature point cloud or multi-view fused image covering the entire surface of the wheel.

[0019] S25. Normalize the obtained 3D feature point cloud or multi-view fused image data, including the unified standardization of data amplitude, size and resolution, and finally generate a unified wheel 3D feature dataset.

[0020] Optionally, S3 specifically includes:

[0021] S31. Use the obtained unified wheel 3D feature dataset as the input dataset, and the input dataset is D. 3D ;

[0022] S32. For the input dataset D 3D Perform feature analysis to extract the task feature vector T task The search space is dynamically adjusted based on wheel type, task complexity, and spatial distribution, and the adaptive search space is defined as S. adapt (T task );

[0023] S33. Constructing the multi-objective fitness function F multi ;

[0024] S34, in the adaptive search space S adapt (T task Within the scope of N, candidate network structures are generated through sampling, mutation, and combination. For each candidate network structure N... i Preliminary training and performance evaluation were conducted to obtain the corresponding fitness score F. multi (N i );

[0025] S35. For each candidate network structure, if F is satisfied in the early stage of training... multi (N i ) <F thresh Or M mem (N i M max If this occurs, the early stopping mechanism is immediately triggered, terminating the training and search process and eliminating candidate network structures N. i ;

[0026] S36. Save the candidate network structures and performance metrics that meet the multi-objective fitness requirements to the architecture knowledge base K. arch This forms a mapping relationship between the historical preferred structure and the task characteristics;

[0027] S37. When performing an architecture search for a new task or a new wheel model, first compare and analyze the characteristics of the current task with those of historical tasks, and prioritize the architecture knowledge base K. arch The historical network structure with the highest structural similarity is retrieved as the starting point for the search, and structural fine-tuning and local optimization are performed to accelerate the acquisition of highly adaptable network structures.

[0028] S38. After multiple rounds of dynamic search, early stopping screening, and historical structure reuse, the final output is in the current wheel 3D feature dataset D. 3D and multi-objective fitness function F multi The optimal deep feature extraction network structure and hyperparameters are described below.

[0029] Optionally, S4 specifically includes:

[0030] S41. Take the obtained optimal deep feature extraction network structure as the basic network structure, and set it as N. opt The obtained 3D feature dataset is then input into N. opt Perform preliminary feature extraction;

[0031] S42. Based on the preliminary feature extraction results, the detection area is divided according to the wheel structure, including the side edge area, the tread wear area, and the hole inner diameter boundary area. A dynamically configurable focus attention module is independently embedded in each area to obtain the primary enhancement features of each area.

[0032] S43. For the primary enhancement features of each detection region, attention weights are automatically learned and allocated in both spatial and channel dimensions to highlight key structural details within each region and generate high-response features for each region.

[0033] S44. High response features for all detection areas, interact and fuse information at different spatial scales and semantic levels to form a fused comprehensive enhanced feature;

[0034] S45. Based on the comprehensive enhanced features, using data from consecutive moments during wheel rotation or multi-frame scanning, joint sparse activation is performed on spatial location and channel dimension to dynamically respond to real-world changing regions and generate temporal sparse enhanced features.

[0035] S46. Enhancement features for temporal sparsity The normalized attention weights of each detection region are jointly optimized globally, with regional balance and inter-regional correlation as objectives, while a global regularization optimization function L is introduced. opt Adaptive global adjustment of attention weight allocation for different detection regions;

[0036] S47. The enhanced feature F after global optimization enhanced The output is used for wheel contour segmentation, dimensional parameter extraction, and automatic defect identification.

[0037] Optionally, S5 specifically includes:

[0038] S51. The obtained enhanced features are used to automatically segment the wheel side edge region, tread wear region and wheel bore inner diameter boundary region to obtain the segmentation mask for each detection region.

[0039] S52. Based on the segmentation mask of each detection area, the image edge detection and contour extraction method is applied to extract the wheel side contour boundary, tread wear boundary and hole inner diameter boundary respectively, and obtain the structural boundary line of each detection area.

[0040] S53. Based on the extracted structural boundary lines, and combined with the correspondence between pixels or point clouds in the segmentation mask and actual spatial coordinates, spatial geometric calculation methods are used to calculate the wheel side diameter, tread width, tread wear depth, and wheel hole inner diameter to obtain the size parameters of each detection area.

[0041] S54. Based on enhanced features and segmentation masks, surface anomaly analysis is performed on each detection area. Defect feature matching and region classification methods are used to detect and mark crack, spalling, and pit defect areas, and record the defect location, area and type information.

[0042] S55. Summarize and organize the dimensional parameters and defect identification results of each detection area, and output the detection analysis results including the wheel side diameter, tread width, tread wear depth, hole inner diameter, and defect type, location, and area.

[0043] Optionally, S6 specifically includes:

[0044] S61. Compare the obtained dimensional parameters of each detection area with the preset wheel size qualification standard, and determine whether the wheel side diameter, tread width, tread wear depth and wheel hole inner diameter meet the specified standard value range, and record them as dimensional judgment conditions.

[0045] S62. Compare the obtained defect identification results with the preset defect judgment criteria, and determine whether the detected defect types such as cracks, chips, and dents, as well as their area and location, meet the wheel safety standards, and record them as defect judgment conditions.

[0046] S63. Based on the size judgment conditions and defect judgment conditions, comprehensively judge whether the wheel is a qualified product. When the size parameters of all inspection areas meet the qualified standards and all defects do not exceed the defect judgment standards, output the qualified judgment result of the wheel; otherwise, output the unqualified judgment result of the wheel.

[0047] S64. Compile the final wheel qualification results, dimensional parameters of each inspection area, defect type and location information, and generate an inspection analysis report.

[0048] An automatic locomotive wheel measurement system based on image recognition according to an embodiment of the present invention includes the following modules:

[0049] The rotation adjustment module is used to drive the wheel to rotate and adjust the position of each scanning device to focus on the side of the wheel, the tread, and the inner surface of the hole;

[0050] The acquisition and fusion module is used to simultaneously acquire multi-view images or point cloud data of the wheel, and perform preprocessing and fusion to generate a three-dimensional feature dataset of the wheel.

[0051] The neural architecture search module is used to automatically search and optimize the deep feature extraction network structure and hyperparameters using a wheel 3D feature dataset as input.

[0052] The focus attention enhancement module is used to embed a focus attention network into the deep feature extraction network structure to enhance the features of key regions and output enhanced features;

[0053] The segmentation and recognition module is used to automatically segment the detection area, extract size parameters, and identify defects based on enhanced features;

[0054] The judgment output module is used to automatically determine whether the wheel is qualified and output the inspection results based on the size parameters and defect identification results.

[0055] The beneficial effects of this invention are:

[0056] This invention overcomes the shortcomings of traditional wheel detection methods by introducing a deep fusion of a neural architecture search algorithm and a focused attention network, achieving efficient, accurate, and intelligent automatic wheel measurement and defect detection. Traditional wheel detection methods largely rely on manual operation, resulting in low efficiency and susceptibility to human factors, failing to meet the high requirements of high-speed rail and heavy-haul transportation. This invention, however, improves the accuracy and efficiency of wheel detection through adaptive optimization of the deep network structure and automatic selection of the optimal feature extraction scheme.

[0057] This invention enhances the response of wheel features in various regions through a focused attention mechanism, particularly in key areas such as the wheel flange, tread wear, and bore diameter. It automatically improves the detection accuracy of minute defects and wear, avoiding details easily overlooked in traditional detection methods. Through multi-task, multi-modal deep learning, the system can quickly and accurately segment different regions of the wheel and extract dimensional parameters under complex working conditions, ensuring high accuracy of the measurement results.

[0058] This invention improves measurement accuracy, shortens detection time, and increases detection efficiency, providing a more reliable and efficient wheel monitoring technology for the railway transportation industry, and has broad practical application value. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of an automatic locomotive wheel measurement method based on image recognition proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the structure of an automatic locomotive wheel measurement system based on image recognition proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 An automatic locomotive wheel measurement method based on image recognition includes the following steps:

[0064] S1. Drive the wheel to rotate at a constant speed using a wheel rotation device, and adjust the positions of the wheel side scanning device, wheel tread scanning device and wheel hole scanning device by moving and adjusting the device respectively, so as to focus on the side, tread and inner surface of the wheel to be tested.

[0065] S2. Simultaneous laser scanning or visual imaging is performed on the wheel side, tread and hole surfaces using wheel side scanning equipment, wheel tread scanning equipment and wheel hole scanning equipment respectively to obtain multi-view images or point cloud data of the wheel under test, and preprocessing and fusion are performed to generate a unified wheel three-dimensional feature dataset.

[0066] S3. Using the neural architecture search algorithm with the wheel 3D feature dataset as input, automatically search and optimize the deep network structure and hyperparameters suitable for the automatic extraction of wheel features, and obtain the optimized deep feature extraction network structure.

[0067] S4. Embed the focus attention network into the deep feature extraction network structure, and use the focus attention network to enhance the features of the wheel side edge region, tread wear region and wheel hole inner diameter boundary region, and output enhanced features.

[0068] S5. Based on enhanced features, automatic segmentation and dimensional parameter extraction of the wheel side profile, tread wear area and wheel bore inner diameter boundary are realized respectively, and the defect area of ​​the wheel surface is automatically identified.

[0069] S6. Based on the extracted dimensional parameters and defect identification results, automatically determine whether the wheel is qualified, and output the final wheel measurement parameters and defect analysis results.

[0070] In this embodiment, S2 specifically includes:

[0071] S21. Start the wheel side scanning device, wheel tread scanning device and wheel hole scanning device respectively. Driven by the wheel rotation device, perform synchronous laser scanning or visual imaging on the side, tread and inner surface of the hole of the wheel to be tested to obtain multi-view images or point cloud data, and mark the source of each type of data respectively.

[0072] S22. Perform noise suppression and filtering on the multi-view images or point cloud data of the side, tread and inner surface of the hole, including the removal of environmental noise, scanning error and invalid information from background interference.

[0073] S23. Perform spatial registration on the processed multi-view images or point cloud data according to the spatial geometric features of each detected part of the wheel, and convert the data obtained from different perspectives and different devices into the same spatial coordinate system.

[0074] S24. The multi-view images or point cloud data after spatial registration are fused, and the data information of different detection devices and detection angles are comprehensively utilized to generate a three-dimensional feature point cloud or multi-view fused image covering the entire surface of the wheel.

[0075] S25. Normalize the obtained 3D feature point cloud or multi-view fused image data, including the unified standardization of data amplitude, size and resolution, and finally generate a unified wheel 3D feature dataset.

[0076] In this embodiment, S3 specifically includes:

[0077] S31. Use the obtained unified wheel 3D feature dataset as the input dataset, and the input dataset is D. 3D ;

[0078] S32. For the input dataset D 3D Perform feature analysis to extract the task feature vector T task The search space is dynamically adjusted based on wheel type, task complexity, and spatial distribution, and the adaptive search space is defined as S. adapt (T task );

[0079] S33. Constructing the multi-objective fitness function F multi :

[0080]

[0081] Among them, A acc For feature extraction accuracy, S time For reasoning speed, M mem For memory consumption, M max The maximum available memory is represented by α1, α2, and α3, which are weighting coefficients.

[0082] S34, in the adaptive search space S adapt (T task Within the scope of N, candidate network structures are generated through sampling, mutation, and combination. For each candidate network structure N... i Preliminary training and performance evaluation were conducted to obtain the corresponding fitness score F. multi (N i );

[0083] S35. For each candidate network structure, if F is satisfied in the early stage of training... multi (N i ) <F thresh Or M mem (N i M max If this occurs, the early stopping mechanism is immediately triggered, terminating the training and search process and eliminating candidate network structures N. i ;

[0084] S36. Save the candidate network structures and performance metrics that meet the multi-objective fitness requirements to the architecture knowledge base K. archThis forms a mapping relationship between the historical preferred structure and the task characteristics;

[0085] S37. When performing an architecture search for a new task or a new wheel model, first compare and analyze the characteristics of the current task with those of historical tasks, and prioritize the architecture knowledge base K. arch The historical network structure with the highest structural similarity is retrieved as the starting point for the search, and structural fine-tuning and local optimization are performed to accelerate the acquisition of highly adaptable network structures.

[0086] S38. After multiple rounds of dynamic search, early stopping screening, and historical structure reuse, the final output is in the current wheel 3D feature dataset D. 3D and multi-objective fitness function F multi The optimal deep feature extraction network structure and hyperparameters are described below.

[0087] In this embodiment, S4 specifically includes:

[0088] S41. Take the obtained optimal deep feature extraction network structure as the basic network structure, and set it as N. opt The obtained 3D feature dataset is then input into N. opt Perform preliminary feature extraction;

[0089] S42. Based on the preliminary feature extraction results, the detection area is divided according to the wheel structure, including the side edge area, the tread wear area, and the hole inner diameter boundary area. A dynamically configurable focus attention module is independently embedded in each area to obtain the primary enhancement features of each area.

[0090] S43. For the primary enhancement features of each detection region, attention weights are automatically learned and allocated in both spatial and channel dimensions to highlight key structural details within each region and generate high-response features for each region.

[0091] S44. High response features for all detection areas, interact and fuse information at different spatial scales and semantic levels to form a fused comprehensive enhanced feature;

[0092] S45. Based on the comprehensive enhanced features, using data from consecutive moments during wheel rotation or multi-frame scanning, joint sparse activation is performed on spatial location and channel dimension to dynamically respond to real-world changing regions and generate temporal sparse enhanced features.

[0093]

[0094] Among them, F t (x) represents the fused feature of the t-th frame. For spatially sparse masks, For timing consistency mask;

[0095] S46. Enhancement features for temporal sparsity The normalized attention weights of each detection region are jointly optimized globally, with regional balance and inter-regional correlation as objectives, while a global regularization optimization function L is introduced. opt Adaptive global adjustment of attention weight allocation for different detection regions:

[0096]

[0097] Where K is the total number of detection areas, p k C represents the normalized attention weight for the k-th region. ij For regional feature correlation, λ and γ are adjustment coefficients, and p i p represents the normalized attention weights for the i-th region. j Let be the normalized attention weight for the j-th region;

[0098] S47. The enhanced feature F after global optimization enhanced The output is used for wheel contour segmentation, dimensional parameter extraction, and automatic defect identification.

[0099] In this embodiment, S5 specifically includes:

[0100] S51. The obtained enhanced features are used to automatically segment the wheel side edge region, tread wear region and wheel bore inner diameter boundary region to obtain the segmentation mask for each detection region.

[0101] S52. Based on the segmentation mask of each detection area, the image edge detection and contour extraction method is applied to extract the wheel side contour boundary, tread wear boundary and hole inner diameter boundary respectively, and obtain the structural boundary line of each detection area.

[0102] S53. Based on the extracted structural boundary lines, and combined with the correspondence between pixels or point clouds in the segmentation mask and actual spatial coordinates, spatial geometric calculation methods are used to calculate the wheel side diameter, tread width, tread wear depth, and wheel hole inner diameter to obtain the size parameters of each detection area.

[0103] S54. Based on enhanced features and segmentation masks, surface anomaly analysis is performed on each detection area. Defect feature matching and region classification methods are used to detect and mark crack, spalling, and pit defect areas, and record the defect location, area and type information.

[0104] S55. Summarize and organize the dimensional parameters and defect identification results of each detection area, and output the detection analysis results including the wheel side diameter, tread width, tread wear depth, hole inner diameter, and defect type, location, and area.

[0105] In this embodiment, S6 specifically includes:

[0106] S61. Compare the obtained dimensional parameters of each detection area with the preset wheel size qualification standard, and determine whether the wheel side diameter, tread width, tread wear depth and wheel hole inner diameter meet the specified standard value range, and record them as dimensional judgment conditions.

[0107] S62. Compare the obtained defect identification results with the preset defect judgment criteria, and determine whether the detected defect types such as cracks, chips, and dents, as well as their area and location, meet the wheel safety standards, and record them as defect judgment conditions.

[0108] S63. Based on the size judgment conditions and defect judgment conditions, comprehensively judge whether the wheel is a qualified product. When the size parameters of all inspection areas meet the qualified standards and all defects do not exceed the defect judgment standards, output the qualified judgment result of the wheel; otherwise, output the unqualified judgment result of the wheel.

[0109] S64. Compile the final wheel qualification results, dimensional parameters of each inspection area, defect type and location information, and generate an inspection analysis report.

[0110] refer to Figure 2 An automatic locomotive wheel measurement system based on image recognition includes the following modules:

[0111] The rotation adjustment module is used to drive the wheel to rotate and adjust the position of each scanning device to focus on the side of the wheel, the tread, and the inner surface of the hole;

[0112] The acquisition and fusion module is used to simultaneously acquire multi-view images or point cloud data of the wheel, and perform preprocessing and fusion to generate a three-dimensional feature dataset of the wheel.

[0113] The neural architecture search module is used to automatically search and optimize the deep feature extraction network structure and hyperparameters using a wheel 3D feature dataset as input.

[0114] The focus attention enhancement module is used to embed a focus attention network into the deep feature extraction network structure to enhance the features of key regions and output enhanced features;

[0115] The segmentation and recognition module is used to automatically segment the detection area, extract size parameters, and identify defects based on enhanced features;

[0116] The judgment output module is used to automatically determine whether the wheel is qualified and output the inspection results based on the size parameters and defect identification results.

[0117] Example 1:

[0118] To verify the feasibility of this invention in practice, it was applied to a locomotive maintenance factory of a railway transportation company. This factory is responsible for regularly inspecting and repairing locomotive wheels. During wheel inspection, the factory needs to ensure that the size, wear condition, and presence of defects such as cracks on each wheel can be accurately assessed. Because the factory needs to inspect a large number of wheels quickly, traditional manual inspection methods are inefficient, have a high error rate, and are not sensitive enough to detecting minor defects, significantly impacting the safety and efficiency of railway transportation.

[0119] The factory adopted the automated inspection method described in this invention. With the help of deep learning technology and focused attention mechanism, it can automatically measure the size of wheels and detect defects, and output the inspection results in a very short time, which not only improves the inspection accuracy, but also greatly shortens the inspection cycle.

[0120] During wheel inspection, the wheel is driven to rotate at a constant speed by a wheel rotation device, while the angle and position of the scanning device are adjusted by a moving adjustment device to ensure that the side, tread, and bore surfaces are fully scanned. Each area of ​​the wheel is scanned by a side scanning device, a tread scanning device, and a bore scanning device, respectively, to acquire images or point cloud data of the wheel.

[0121] The data collected by the scanning device undergoes noise removal and data fusion through a data preprocessing and fusion module to generate a unified 3D feature dataset for wheels. This dataset is then fed into a deep learning network based on neural architecture search, which automatically optimizes and extracts the best network structure suitable for automatic wheel feature extraction. This adaptive adjustment of the network allows the system to automatically adjust the network structure for different wheel models and scanning conditions to achieve optimal feature extraction results.

[0122] The focused attention network is embedded in the deep feature extraction network to enhance features in the side edge region of the wheel, the tread wear region, and the inner diameter boundary region of the bore. After processing by the focused attention network, minute wear and cracks on the wheel can be identified more accurately.

[0123] Based on enhanced feature data, the system can automatically segment and extract dimensional parameters of the wheel side profile, tread wear area, and wheel bore inner diameter boundary, and perform defect detection to identify defect areas such as cracks, chips, and wear on the wheel surface, and finally automatically determine whether the wheel is qualified.

[0124] Table 1 Comparison of Wheel Inspection Efficiency and Accuracy

[0125]

[0126]

[0127] Table 1 shows significant differences in efficiency and accuracy among different wheel inspection methods. Traditional manual inspection takes 30 minutes per wheel, with a total inspection time of 500 hours. The automated inspection system, however, takes only 5 minutes, reducing the total inspection time to 83.3 hours, an efficiency improvement of approximately 6 times. The optimized automated inspection system further reduces the inspection time to 66.7 hours. In comparison, manual-assisted automated inspection and rapid automated inspection take 166.7 hours and 50 hours respectively, still shorter than traditional manual inspection but shorter than the fully automated system.

[0128] Regarding defect identification accuracy, the automated inspection system achieved 97%, while the optimized system improved to 98%, significantly higher than the 85% of traditional manual inspection. Rapid automated inspection performed best, achieving an accuracy of 99%. In terms of wheel size measurement error, traditional manual inspection had an error of 2-5 mm, while the automated inspection system reduced the error to 0.5-1 mm, and the optimized system reduced it to 0.3-0.6 mm, demonstrating a substantial improvement in precision.

[0129] Overall, automated inspection systems outperform traditional manual inspection in terms of efficiency, accuracy, and precision. Furthermore, with optimization, the system's performance continues to improve, demonstrating significant advantages.

[0130] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automatic locomotive wheel measurement method based on image recognition, characterized in that, Includes the following steps: S1. Drive the wheel to rotate at a constant speed using a wheel rotation device, and adjust the positions of the wheel side scanning device, wheel tread scanning device and wheel hole scanning device by moving and adjusting the device respectively, so as to focus on the side, tread and inner surface of the wheel to be tested. S2. Simultaneous laser scanning or visual imaging is performed on the wheel side, tread and hole surfaces using wheel side scanning equipment, wheel tread scanning equipment and wheel hole scanning equipment respectively to obtain multi-view images or point cloud data of the wheel under test, and preprocessing and fusion are performed to generate a unified wheel three-dimensional feature dataset. S3. Using the neural architecture search algorithm with the wheel 3D feature dataset as input, automatically search and optimize the deep network structure and hyperparameters suitable for the automatic extraction of wheel features, and obtain the optimized deep feature extraction network structure. S4. Embed the focus attention network into the deep feature extraction network structure, and use the focus attention network to enhance the features of the wheel side edge region, tread wear region and wheel hole inner diameter boundary region, and output enhanced features. S5. Based on enhanced features, automatic segmentation and dimensional parameter extraction of the wheel side profile, tread wear area and wheel bore inner diameter boundary are realized respectively, and the defect area of ​​the wheel surface is automatically identified. S6. Based on the extracted dimensional parameters and defect identification results, automatically determine whether the wheel is qualified, and output the final wheel measurement parameters and defect analysis results.

2. The automatic locomotive wheel measurement method based on image recognition according to claim 1, characterized in that, S2 specifically includes: S21. Start the wheel side scanning device, wheel tread scanning device and wheel hole scanning device respectively. Driven by the wheel rotation device, perform synchronous laser scanning or visual imaging on the side, tread and inner surface of the hole of the wheel to be tested to obtain multi-view images or point cloud data, and mark the source of each type of data respectively. S22. Perform noise suppression and filtering on the multi-view images or point cloud data of the side, tread and inner surface of the hole, including the removal of environmental noise, scanning error and invalid information from background interference. S23. Perform spatial registration on the processed multi-view images or point cloud data according to the spatial geometric features of each detected part of the wheel, and convert the data obtained from different perspectives and different devices into the same spatial coordinate system. S24. The multi-view images or point cloud data after spatial registration are fused, and the data information of different detection devices and detection angles are comprehensively utilized to generate a three-dimensional feature point cloud or multi-view fused image covering the entire surface of the wheel. S25. Normalize the obtained 3D feature point cloud or multi-view fused image data, including the unified standardization of data amplitude, size and resolution, and finally generate a unified wheel 3D feature dataset.

3. The automatic locomotive wheel measurement method based on image recognition according to claim 1, characterized in that, S3 specifically includes: S31. Use the obtained unified wheel 3D feature dataset as the input dataset, and the input dataset is D. 3D ; S32. For the input dataset D 3D Perform feature analysis to extract the task feature vector T task The search space is dynamically adjusted based on wheel type, task complexity, and spatial distribution, and the adaptive search space is defined as S. adapt (T task ); S33. Constructing the multi-objective fitness function F multi ; S34, in the adaptive search space S adapt (T task Within the scope of N, candidate network structures are generated through sampling, mutation, and combination. For each candidate network structure N... i Preliminary training and performance evaluation were conducted to obtain the corresponding fitness score F. multi (N i ); S35. For each candidate network structure, if F is satisfied in the early stage of training... multi (N i ) <F thresh Or M mem (N i M max If this occurs, the early stopping mechanism is immediately triggered, terminating the training and search process and eliminating candidate network structures N. i ; S36. Save the candidate network structures and performance metrics that meet the multi-objective fitness requirements to the architecture knowledge base K. arch This forms a mapping relationship between the historical preferred structure and the task characteristics; S37. When performing an architecture search for a new task or a new wheel model, first compare and analyze the characteristics of the current task with those of historical tasks, and prioritize the architecture knowledge base K. arch The historical network structure with the highest structural similarity is retrieved as the starting point for the search, and structural fine-tuning and local optimization are performed to accelerate the acquisition of highly adaptable network structures. S38. After multiple rounds of dynamic search, early stopping screening, and historical structure reuse, the final output is in the current wheel 3D feature dataset D. 3D and multi-objective fitness function F multi The optimal deep feature extraction network structure and hyperparameters are described below.

4. The automatic locomotive wheel measurement method based on image recognition according to claim 1, characterized in that, S4 specifically includes: S41. Take the obtained optimal deep feature extraction network structure as the basic network structure, and set it as N. opt The obtained 3D feature dataset is then input into N. opt Perform preliminary feature extraction; S42. Based on the preliminary feature extraction results, the detection area is divided according to the wheel structure, including the side edge area, the tread wear area, and the hole inner diameter boundary area. A dynamically configurable focus attention module is independently embedded in each area to obtain the primary enhancement features of each area. S43. For the primary enhancement features of each detection region, attention weights are automatically learned and allocated in both spatial and channel dimensions to highlight key structural details within each region and generate high-response features for each region. S44. High response features for all detection areas, interact and fuse information at different spatial scales and semantic levels to form a fused comprehensive enhanced feature; S45. Based on the comprehensive enhanced features, using data from consecutive moments during wheel rotation or multi-frame scanning, joint sparse activation is performed on spatial location and channel dimension to dynamically respond to real-world changing regions and generate temporal sparse enhanced features. S46. Enhancement features for temporal sparsity The normalized attention weights of each detection region are jointly optimized globally, with regional balance and inter-regional correlation as objectives, while a global regularization optimization function L is introduced. opt Adaptive global adjustment of attention weight allocation for different detection regions; S47. The enhanced feature F after global optimization enhanced The output is used for wheel contour segmentation, dimensional parameter extraction, and automatic defect identification.

5. The automatic locomotive wheel measurement method based on image recognition according to claim 1, characterized in that, S5 specifically includes: S51. The obtained enhanced features are used to automatically segment the wheel side edge region, tread wear region and wheel bore inner diameter boundary region to obtain the segmentation mask for each detection region. S52. Based on the segmentation mask of each detection area, the image edge detection and contour extraction method is applied to extract the wheel side contour boundary, tread wear boundary and hole inner diameter boundary respectively, and obtain the structural boundary line of each detection area. S53. Based on the extracted structural boundary lines, and combined with the correspondence between pixels or point clouds in the segmentation mask and actual spatial coordinates, spatial geometric calculation methods are used to calculate the wheel side diameter, tread width, tread wear depth, and wheel hole inner diameter to obtain the size parameters of each detection area. S54. Based on enhanced features and segmentation masks, surface anomaly analysis is performed on each detection area. Defect feature matching and region classification methods are used to detect and mark crack, spalling, and pit defect areas, and record the defect location, area and type information. S55. Summarize and organize the dimensional parameters and defect identification results of each detection area, and output the detection analysis results including the wheel side diameter, tread width, tread wear depth, hole inner diameter, and defect type, location, and area.

6. The automatic locomotive wheel measurement method based on image recognition according to claim 1, characterized in that, S6 specifically includes: S61. Compare the obtained dimensional parameters of each detection area with the preset wheel size qualification standard, and determine whether the wheel side diameter, tread width, tread wear depth and wheel hole inner diameter meet the specified standard value range, and record them as dimensional judgment conditions. S62. Compare the obtained defect identification results with the preset defect judgment criteria, and determine whether the detected defect types such as cracks, chips, and dents, as well as their area and location, meet the wheel safety standards, and record them as defect judgment conditions. S63. Based on the size judgment conditions and defect judgment conditions, comprehensively judge whether the wheel is a qualified product. When the size parameters of all inspection areas meet the qualified standards and all defects do not exceed the defect judgment standards, output the qualified judgment result of the wheel; otherwise, output the unqualified judgment result of the wheel. S64. Compile the final wheel qualification results, dimensional parameters of each inspection area, defect type and location information, and generate an inspection analysis report.

7. An automatic locomotive wheel measurement system based on image recognition, comprising the automatic locomotive wheel measurement method based on image recognition as described in any one of claims 1 to 6, characterized in that, Includes the following modules: The rotation adjustment module is used to drive the wheel to rotate and adjust the position of each scanning device to focus on the side of the wheel, the tread, and the inner surface of the hole; The acquisition and fusion module is used to simultaneously acquire multi-view images or point cloud data of the wheel, and perform preprocessing and fusion to generate a three-dimensional feature dataset of the wheel. The neural architecture search module is used to automatically search and optimize the deep feature extraction network structure and hyperparameters using a wheel 3D feature dataset as input. The focus attention enhancement module is used to embed a focus attention network into the deep feature extraction network structure to enhance the features of key regions and output enhanced features; The segmentation and recognition module is used to automatically segment the detection area, extract size parameters, and identify defects based on enhanced features; The judgment output module is used to automatically determine whether the wheel is qualified and output the inspection results based on the size parameters and defect identification results.

Images