Automatic wheel surface image acquisition and damage analysis system for locomotive non-pulling wheel turning repair scene

Abstract

An automatic wheel surface image acquisition and damage analysis system for a locomotive non-pulling wheel turning repair scene comprises a mechanical module unit, an image acquisition unit, an algorithm analysis unit and a main control unit, and all the units cooperatively work through physical connection, electrical connection and data connection. The image acquisition and analysis system is a systematic scheme formed by a large amount of research, debugging and collaborative optimization aiming at a core pain point of a railway maintenance non-pulling wheel turning repair scene. Through deep coupling of image acquisition, light source adaptation, algorithm analysis and a master control scheduling unit, the scene adaptation problem of a general technology is solved, and systematic efficiency is improved through unit collaboration.

Description

Technical Field

[0001] This invention relates to the field of railway maintenance technology, and in particular to an automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without removing the wheels. Background Technology

[0002] In the field of rail transit locomotive operation and maintenance, wheel turning machines are core equipment for locomotive wheel repair. Wheel turning is a necessary procedure to remove existing damage to the wheel surface, restore the standard wheel profile, and ensure the safe operation of locomotives. However, the current quality inspection of wheel surface damage after turning is still at the level of traditional manual operation, and there is no supporting automated image acquisition and intelligent damage recognition and classification equipment, resulting in significant shortcomings in the control of turning quality.

[0003] In existing technologies, wheel turning workers often use non-professional equipment such as handheld mobile phones and portable cameras to collect images of the wheel surface. This not only results in inconsistent image resolution and contrast due to differences in equipment performance, but also leads to a chaotic style and blurred key damage features due to the lack of standardized shooting angles and lighting conditions. This makes it impossible to form a standardized damage image database, and it is also difficult to support subsequent unified and accurate damage identification, thus creating hidden dangers for the traceability and evaluation of turning quality.

[0004] The inspection of the surface quality of the wheel after turning is entirely done manually. It requires the use of a wheel turning machine that rotates at each angle. The operators visually inspect the wheel tread, flange and other key parts for damage. This process is not only cumbersome and time-consuming, but also significantly slows down the overall maintenance efficiency of wheel turning.

[0005] Manual damage assessment has significant subjective limitations. Inspection results rely excessively on the experience of operators, making it prone to omissions and misjudgments of minor damage due to factors such as operator fatigue and lack of experience. This hinders effective control over the quality of locomotive turning and could even lead to damaged wheels being put into service, posing a risk to train safety. Furthermore, inconsistencies in damage assessment across different units hinder comprehensive damage statistics across the entire railway locomotive maintenance sector.

[0006] Currently, in the field of wheel-on-wheel repair for rail transit locomotives, there is an urgent need for a dedicated system that can achieve standardized image acquisition and intelligent damage analysis, filling the technological gap in automated detection of wheel quality after repair and improving the overall operation and maintenance level and safety assurance capabilities of the repair process. Summary of the Invention

[0007] To overcome existing shortcomings, this invention proposes an automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal.

[0008] An automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal includes a mechanical module unit, an image acquisition unit, an algorithm analysis unit, and a main control unit. These units work collaboratively through physical, electrical, and data connections.

[0009] The mechanical module unit includes system A and system B. System A is installed on the front side of the non-drop wheel lathe via a mechanical module fixing device, and system B is installed on the rear side of the non-drop wheel lathe via a mechanical module fixing device. Each system contains two mechanical modules, corresponding to the left and right wheels respectively. The mechanical modules drive the X-axis, Y-axis, and Z-axis linear slides via servo motors. Each axis is connected by a synchronous transmission device and supported by a fixed bracket. The mechanical module unit is electrically connected to the main control unit via a cable harness and receives displacement commands and enable / disable signals.

[0010] The image acquisition unit is rigidly connected to the end of the mechanical module unit via a camera mounting bracket, and includes a 3-axis linear array camera, a 2-axis linear array camera, a 3-axis fixed-focus lens, a 2-axis fixed-focus lens, a 3-axis main light source, a 2-axis main light source, a 3-axis auxiliary light source, and a 2-axis auxiliary light source. The linear array camera is threadedly connected to the fixed-focus lens via an M42 interface. The main light source and auxiliary light source are bolted to the camera mounting bracket via a bracket. The light source angle is manually adjusted via an auxiliary light source adjuster. The image acquisition unit is connected to the main control unit via Gigabit Ethernet to receive start / stop / parameter adjustment commands.

[0011] The algorithm analysis unit is deployed on an industrial control all-in-one computer. It communicates with the image acquisition unit and the main control unit through the internal bus of the main control unit, receives image data and returns damage identification results.

[0012] The main control unit integrates a PLC control system and a Windows system. It works in coordination with the mechanical module unit, image acquisition unit, and algorithm analysis unit through LAN interface, COM interface, and Ethercat master station to achieve fully automated operation of the entire system.

[0013] In systems A and B of the aforementioned mechanical module units, each mechanical module is configured as a 2-axis or 3-axis structure depending on the type of non-load-bearing turning machine:

[0014] The 2-axis mechanical module includes an X-axis linear slide and a Z-axis linear slide. The X-axis linear slide is driven by an X-axis servo motor and moves laterally through the X-axis mechanical module. The Z-axis linear slide is driven by a Z-axis servo motor and moves longitudinally through the Z-axis mechanical module. The X-axis and Z-axis are connected by fixed brackets for X-axis and Z-axis.

[0015] The 3-axis mechanical module includes X-axis, Y-axis, and Z-axis linear slides. The Y-axis linear slide is driven by a Y-axis servo motor and achieves vertical movement through the Y-axis mechanical module; the X-axis linear slide is driven by an X-axis servo motor and achieves lateral movement through the X-axis mechanical module; the Z-axis linear slide is driven by a Z-axis servo motor and achieves longitudinal movement through the Z-axis mechanical module; the Y-axis and Z-axis are connected by fixed brackets for Y-axis and Z-axis, and the Z-axis and X-axis are connected by fixed brackets for Z-axis and X-axis.

[0016] The mechanical module has a positioning error of ≤1.0mm and is compatible with three wheel size ranges: 1250-1150mm, 1050-975mm, and 1050-950mm.

[0017] In the mechanical module unit, the servo motors are powered by 220V. The X-axis and Y-axis servo motors do not have brakes, while the Z-axis servo motor has a brake. Each servo motor drives the slide table through a synchronous transmission device, which includes a synchronous pulley, a synchronous belt, and a protective cover. The mechanical module unit also includes a counterweight to balance the motion inertia.

[0018] The image acquisition unit uses a Hikvision MV-CL042-91GM linear scan camera with a 4096×2-line CMOS sensor; a fixed-focus lens with a focal length of 40mm and a maximum aperture of F2.8; a main light source with dimensions of 200 * 200mm and a power of 60.6W, and an auxiliary light source with dimensions of 150 * 100mm and a power of 26.4W; the brightness of the light source is steplessly adjusted by the main control unit to suppress reflections on the metal surface of the wheel.

[0019] The algorithm analysis unit has a built-in damage recognition algorithm based on the YOLOv8n-SMC model. The industrial control all-in-one computer is equipped with an i5-12400 processor, 32G memory, and 128G SSD + 2T SSD hard drive. The algorithm analysis unit processes images in parallel through multi-threading to identify four types of damage: cracks, peeling, diagonal cracks, and iron filings scratches.

[0020] The main control unit expands external devices through multiple interfaces, including a DP interface for display connection, an HDMI interface for backup display, and a USB interface for data transmission, ensuring system scalability and stability.

[0021] The method includes the following steps:

[0022] Step 1, System Startup and Enablement: The main control unit sends an enable signal to the servo motor of the mechanical module unit, enabling the servo motor to enter the working ready state and resetting the mechanical module to the initial position;

[0023] Step 2, Mechanical module movement steps: Based on the received locomotive model, locomotive entry position and turning and repair sequence signal, the main control unit selects system A or system B of the mechanical module unit to work, and issues displacement command to control the mechanical module to drive the image acquisition unit to move to the preset image acquisition position.

[0024] Step 3, Image Acquisition Step: The line scan camera acquires the shooting signal from the main control unit, starts scanning at a preset frequency, and uses a high-brightness light source for supplementary lighting to achieve full-coverage image acquisition of the wheel surface, and transmits the acquired image to the main control unit in real time;

[0025] Step 4, Equipment Reset Procedure: After image acquisition is completed, the main control unit sends a disabling signal to the servo motor, causing the mechanical module to reset to its initial position;

[0026] Step 5, Damage Identification and Classification: The main control unit forwards the acquired images to the algorithm analysis unit, which performs preprocessing, multi-threaded parallel detection, damage detection and filtering, priority determination and visual annotation on the images through a customized damage identification model, and outputs the damage location, type and confidence level.

[0027] Step 6, Result Output: The main control unit receives the recognition result and displays a visual image on the industrial control all-in-one computer screen. If damage is detected, an alarm is triggered, and the damage image and data are stored.

[0028] The mechanical module movement in step two specifically includes:

[0029] The main control unit controls the operation of either system A or system B of the mechanical module unit according to the type of the non-dismounting wheel lathe and the locomotive model. System A is installed on the front side of the lathe, and system B is installed on the rear side of the lathe.

[0030] In the movement steps of the mechanical module, the main control unit automatically selects the movement trajectory and number of axes of the mechanical module according to the type of non-falling wheel lathe and the model of the testing machine:

[0031] For the 2-axis mechanical module, the image acquisition unit is moved laterally and longitudinally by the coordinated movement of the X-axis linear slide and the Z-axis linear slide, which can adapt to the range of wheel sizes.

[0032] For a 3-axis mechanical module, the vertical movement degree of freedom is increased by linking the linear slides of the X-axis, Y-axis and Z-axis, thus avoiding spatial interference.

[0033] The image acquisition unit is precisely positioned within the depth-of-field coverage area of ​​the wheel surface, and is compatible with wheel sizes of 1250-1150mm, 1050-975mm and 1050-950mm.

[0034] The motion trajectory of the mechanical module is automatically matched based on the built-in model parameter library, and the positioning error is ≤1.0mm.

[0035] In the image acquisition step, the linear scan camera works in conjunction with a high-brightness light source to optimize image quality:

[0036] The line scan camera scans the surface of the rotating wheel line by line at a maximum line frequency of 80kHz to avoid motion blur.

[0037] The main light source illuminates the wheel tread at a low angle, while the auxiliary light source illuminates the root of the wheel flange at an angle, suppressing metallic reflection through diffused light.

[0038] Fixed-focus lenses achieve fast and accurate focusing through the linkage of a manual focus ring and a Z-axis linear slide.

[0039] The image acquisition specifically includes:

[0040] The line scan camera uses the Hikvision MV-CL042-91GM model. The high-brightness light source includes a main light source and an auxiliary light source. The main light source is used for supplemental lighting of the wheel tread surface, and the auxiliary light source is used for supplemental lighting of the wheel flange root. The brightness of the light source is steplessly adjusted by the main control unit, and the angle is adjusted by the auxiliary light source adjuster to suppress reflection on the wheel surface and improve the contrast of damage.

[0041] During image acquisition, the line scan camera rotates synchronously with the wheel to avoid motion blur and ensure image clarity.

[0042] The damage identification and classification steps include the following sub-steps:

[0043] Image preprocessing: Convert the acquired BMP format images to 1024x1250 resolution JPG format while maintaining 95% image quality;

[0044] Multi-threaded parallel detection: The images to be detected are divided into multiple batches of 10 images each, and multiple batches of parallel detection are achieved through a 3-thread pool;

[0045] Damage detection and filtering: Differentiated confidence thresholds are used, with a detection confidence threshold of 0.4 for cracks and diagonal cracks, and a detection confidence threshold of 0.3 for peeling and iron filings scratches, and an overlap frame filtering threshold of 0.3 is set.

[0046] Damage priority determination: Output the damage results of a single image according to the priority of peeling > crack > oblique crack > iron filings scratch;

[0047] Visual annotation and result output: Configure exclusive annotation colors for four types of damage, generate detection result images and return the results in JSON format.

[0048] In the damage identification and classification steps, the YOLOv8n custom model used is optimized in the following ways:

[0049] A C2f-CCFM cross-scale feature fusion module is introduced into layers 2 and 4 of the backbone network to enhance cross-scale feature interaction;

[0050] Optimize the head network structure, add a P2 / 4 scale detection head, and form a reciprocating feature fusion path;

[0051] The model achieved an accuracy of 96.15%, a recall of 97.47%, and a mean precision of 96.00% on a locomotive wheel surface damage dataset.

[0052] The method also includes a collaborative management step for damage data:

[0053] The main control unit associates and stores the damage identification results with the locomotive's entry position and axle sequence information, generates an inspection report, and uploads it to the server;

[0054] The system supports querying and statistical analysis of historical damage data, which is used for traceability of turning quality and support for operation and maintenance decisions.

[0055] The method achieves fully automated closed-loop control through a main control unit, including:

[0056] The main control unit monitors the lathe status in real time and dynamically adjusts the image acquisition parameters;

[0057] The algorithm analysis unit works in tandem with the mechanical module unit to ensure seamless integration of image acquisition and damage recognition, improving overall efficiency by more than 50%.

[0058] The image acquisition and analysis system of this invention is not a simple aggregation of existing technologies, but a systematic solution developed through extensive research, debugging, and collaborative optimization, addressing the core pain points of railway locomotive wheel-turning repair scenarios. Through deep coupling of image acquisition, light source adaptation, algorithm analysis, and the main control and scheduling unit, it not only solves the scenario adaptation problem of general technologies but also produces an unexpected effect of 1+1>2 through unit collaboration.

[0059] In the railway locomotive industry, the models of wheel lathes are diverse, and the locomotive models are varied. Directly applying existing data acquisition equipment can easily lead to problems such as structural interference and incomplete detection coverage. This invention, based on on-site surveys and operational condition analysis of over 95% of the lathes and locomotive models in major locomotive depots nationwide, adopts a modular mechanical structure and a configurable parameter system to achieve high adaptability: the data acquisition unit mounting base can be adjusted in multiple dimensions, including the angle of the camera and light source; the fixed-focus lens has a depth of field covering multiple locomotive models; the system has a built-in locomotive model parameter library, and the main control unit automatically identifies and matches the acquisition parameters and detection range.

[0060] After turning, the metal surface of a wheel is prone to strong reflections. Minor damage is barely noticeable due to minimal surface grayscale differences, making it difficult for traditional imaging components to balance reflection suppression with the capture of fine defects. This invention, developed through extensive optical experiments, presents a collaborative solution: a line-scan camera, a high-brightness soft light source, and a fixed-focus lens. The primary and secondary soft light sources have undergone thousands of adjustments to determine their parameters, effectively suppressing reflections and ensuring sufficient illumination at the wheel rim root. The fixed-focus lens is precisely selected and calibrated to match the camera's scanning characteristics, achieving full coverage of the wheel's diameter throughout its entire lifespan. Precise synchronization between the camera and wheel rotation prevents motion blur. This unit collaboration achieves micron-level high-resolution, high-contrast imaging, accurately capturing minute defects and resolving the pain point of reflection-induced blur. This effect requires extensive optical parameter adjustments and synchronous calibration, which cannot be achieved through simple component assembly.

[0061] Based on the existing regulations and on-site maintenance experience of various railway bureaus in China, the types of surface damage to locomotive wheels are systematically classified, and four core categories are identified: cracks, diagonal cracks, peeling, and metal scraping. Standardized definition criteria are also formulated for each type of damage. Based on the above classification criteria, a dedicated image dataset of locomotive wheel surface damage containing more than 30,000 samples has been constructed (with annotation completed).

[0062] In industrial settings, a balance needs to be struck between "accurate identification of multiple types of small damage" and "real-time detection." Direct application of existing YOLO models often results in missed or false detections of small targets and insufficient speed. This invention is a customized development based on the YOLOv8n model, adding a CCFM cross-scale module and optimizing the head network (parameters determined after hundreds of debugging sessions). Differential confidence thresholds and overlapping box filtering parameters are set for four types of damage, combined with lightweight model optimization. The synergistic effect yields unexpected results: when the C2f-CCFM model and the optimized head network work together, the model achieves optimal performance, with accuracy, recall, and mean precision reaching 96.79%, 97.97%, and 96.11% respectively, significantly outperforming traditional models. Comparative and ablation experiments validate this, and single-round detection takes ≤30 seconds, meeting real-time requirements in the field.

[0063] In conjunction with a series of innovations in the model inference phase, specific thresholds for each type of problem have been set based on their impact on locomotive safety and the difficulty of detection. The priority of damage assessment has also been optimized.

[0064] This effect requires massive data construction and multiple model optimizations, and cannot be achieved by simply applying a general model, reflecting the difficulty and innovation of research and development.

[0065] Traditional manual inspection is inefficient, subjective, and costly. A key challenge in scenario-based implementation is achieving coordinated automation across all stages to prevent inspection interruptions. This invention achieves a fully automated closed-loop process through a core control unit: synchronous lathe status triggers parameter adjustments, data is automatically transmitted to the algorithm unit for recognition, results are visualized in real-time, and reports are generated and uploaded to the backend. This synergistic effect produces systemic benefits: replacing manual labor increases efficiency by over 50%, ensures consistent and reliable inspections, and reduces labor and training costs. This automation requires communication protocol adaptation, scheduling logic design, and extensive integration testing, which cannot be achieved through simple equipment assembly, highlighting its systemic advantages. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the mechanical module unit structure.

[0067] Figure 2 This is a schematic diagram of a 3-axis mechanical module.

[0068] Figure 3 This is a schematic diagram of a 2-axis mechanical module.

[0069] Figure 4 This is a diagram of the YOLOv8n-SMC network structure. Detailed Implementation

[0070] The following describes in detail, with reference to the accompanying drawings and specific embodiments, an automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal, provided by the present invention.

[0071] according to Figure 1-3 As shown, an automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal includes a mechanical module unit, an image acquisition unit, an algorithm analysis unit, and a main control unit. These units work collaboratively through physical, electrical, and data connections.

[0072] The mechanical module unit includes system A and system B. System A is installed on the front side of the non-drop wheel lathe via a mechanical module fixing device, and system B is installed on the rear side of the non-drop wheel lathe via a mechanical module fixing device. Each system contains two mechanical modules, corresponding to the left and right wheels respectively. The mechanical modules drive the X-axis, Y-axis, and Z-axis linear slides via servo motors. Each axis is connected by a synchronous transmission device and supported by a fixed bracket. The mechanical module unit is electrically connected to the main control unit via a cable harness and receives displacement commands and enable / disable signals.

[0073] The image acquisition unit is rigidly connected to the end of the mechanical module unit via a camera mounting bracket, and includes a 3-axis linear array camera 15, a 2-axis linear array camera 151, a 3-axis fixed-focus lens (16), a 2-axis fixed-focus lens 161, a 3-axis main light source 17, a 2-axis main light source 171, a 3-axis auxiliary light source 18, and a 2-axis auxiliary light source 181. The linear array camera is threadedly connected to the fixed-focus lens via an M42 interface. The main light source and auxiliary light source are bolted to the camera mounting bracket via a bracket. The light source angle is manually adjusted via an auxiliary light source adjuster. The image acquisition unit is connected to the main control unit via Gigabit Ethernet and receives start / stop / parameter adjustment commands.

[0074] The algorithm analysis unit is deployed on an industrial control all-in-one computer. It communicates with the image acquisition unit and the main control unit through the internal bus of the main control unit, receives image data and returns damage identification results.

[0075] The main control unit integrates a PLC control system and a Windows system. It works in coordination with the mechanical module unit, image acquisition unit, and algorithm analysis unit through LAN interface, COM interface, and Ethercat master station to achieve fully automated operation of the entire system.

[0076] In systems A and B of the aforementioned mechanical module units, each mechanical module is configured as a 2-axis or 3-axis structure depending on the type of non-load-bearing turning machine:

[0077] The 2-axis mechanical module includes an X-axis linear slide and a Z-axis linear slide. The X-axis linear slide is driven by an X-axis servo motor 91 and moves laterally through the X-axis mechanical module 111. The Z-axis linear slide is driven by a Z-axis servo motor 61 and moves longitudinally through the Z-axis mechanical module 71. The X-axis and Z-axis are connected by X-axis and Z-axis fixed brackets 51.

[0078] The 3-axis mechanical module includes X-axis, Y-axis, and Z-axis linear slides. The Y-axis linear slide is driven by the Y-axis servo motor 2 and moves vertically through the Y-axis mechanical module 4. The X-axis linear slide is driven by the X-axis servo motor 9 and moves laterally through the X-axis mechanical module 11. The Z-axis linear slide is driven by the Z-axis servo motor 6 and moves longitudinally through the Z-axis mechanical module 7. The Y-axis and Z-axis are connected by the Y-axis and Z-axis fixed brackets 5, and the Z-axis and X-axis are connected by the Z-axis and X-axis fixed brackets 8.

[0079] The mechanical module has a positioning error of ≤1.0mm and is compatible with three wheel size ranges: 1250-1150mm, 1050-975mm, and 1050-950mm.

[0080] In the mechanical module unit, the servo motors are powered by 220V. The X-axis and Y-axis servo motors do not have brakes, while the Z-axis servo motor has a brake. Each servo motor drives the slide table through a synchronous transmission device, which includes a synchronous pulley, a synchronous belt, and a protective cover. The mechanical module unit also includes a counterweight 13 to balance the motion inertia.

[0081] The image acquisition unit uses a Hikvision MV-CL042-91GM linear scan camera with a 4096×2-line CMOS sensor; a fixed-focus lens with a focal length of 40mm and a maximum aperture of F2.8; a main light source with dimensions of 200 * 200mm and a power of 60.6W, and an auxiliary light source with dimensions of 150 * 100mm and a power of 26.4W; the brightness of the light source is steplessly adjusted by the main control unit to suppress reflections on the metal surface of the wheel.

[0082] The algorithm analysis unit has a built-in damage recognition algorithm based on the YOLOv8n-SMC model. The industrial control all-in-one computer is equipped with an i5-12400 processor, 32G memory, and 128G SSD + 2T SSD hard drive. The algorithm analysis unit processes images in parallel through multi-threading to identify four types of damage: cracks, peeling, diagonal cracks, and iron filings scratches.

[0083] The main control unit expands external devices through multiple interfaces, including a DP interface for display connection, an HDMI interface for backup display, and a USB interface for data transmission, ensuring system scalability and stability.

[0084] The method includes the following steps:

[0085] Step 1, System Startup and Enablement: The main control unit sends an enable signal to the servo motor of the mechanical module unit, enabling the servo motor to enter the working ready state and resetting the mechanical module to the initial position;

[0086] Step 2, Mechanical module movement steps: Based on the received locomotive model, locomotive entry position and turning and repair sequence signal, the main control unit selects system A or system B of the mechanical module unit to work, and issues displacement command to control the mechanical module to drive the image acquisition unit to move to the preset image acquisition position.

[0087] Step 3, Image Acquisition Step: The line scan camera acquires the shooting signal from the main control unit, starts scanning at a preset frequency, and uses a high-brightness light source for supplementary lighting to achieve full-coverage image acquisition of the wheel surface, and transmits the acquired image to the main control unit in real time;

[0088] Step 4, Equipment Reset Procedure: After image acquisition is completed, the main control unit sends a disabling signal to the servo motor, causing the mechanical module to reset to its initial position;

[0089] Step 5, Damage Identification and Classification: The main control unit forwards the acquired images to the algorithm analysis unit, which performs preprocessing, multi-threaded parallel detection, damage detection and filtering, priority determination and visual annotation on the images through a customized damage identification model, and outputs the damage location, type and confidence level.

[0090] Step 6, Result Output: The main control unit receives the recognition result and displays a visual image on the industrial control all-in-one computer screen. If damage is detected, an alarm is triggered, and the damage image and data are stored.

[0091] The mechanical module movement in step two specifically includes:

[0092] The main control unit controls the operation of either system A or system B of the mechanical module unit according to the type of the non-dismounting wheel lathe and the locomotive model. System A is installed on the front side of the lathe, and system B is installed on the rear side of the lathe.

[0093] In the movement steps of the mechanical module, the main control unit automatically selects the movement trajectory and number of axes of the mechanical module according to the type of non-falling wheel lathe and the model of the testing machine:

[0094] For the 2-axis mechanical module, the image acquisition unit is moved laterally and longitudinally by the coordinated movement of the X-axis linear slide and the Z-axis linear slide, which can adapt to the range of wheel sizes.

[0095] For a 3-axis mechanical module, the vertical movement degree of freedom is increased by linking the linear slides of the X-axis, Y-axis and Z-axis, thus avoiding spatial interference.

[0096] The image acquisition unit is precisely positioned within the depth-of-field coverage area of ​​the wheel surface, and is compatible with wheel sizes of 1250-1150mm, 1050-975mm and 1050-950mm.

[0097] The motion trajectory of the mechanical module is automatically matched based on the built-in model parameter library, and the positioning error is ≤1.0mm.

[0098] In the image acquisition step, the linear scan camera works in conjunction with a high-brightness light source to optimize image quality:

[0099] The line scan camera scans the surface of the rotating wheel line by line at a maximum line frequency of 80kHz to avoid motion blur.

[0100] The main light source illuminates the wheel tread at a low angle, while the auxiliary light source illuminates the root of the wheel flange at an angle, suppressing metallic reflection through diffused light.

[0101] Fixed-focus lenses achieve fast and accurate focusing through the linkage of a manual focus ring and a Z-axis linear slide.

[0102] The image acquisition specifically includes:

[0103] The line scan camera uses the Hikvision MV-CL042-91GM model. The high-brightness light source includes a main light source and an auxiliary light source. The main light source is used for supplemental lighting of the wheel tread surface, and the auxiliary light source is used for supplemental lighting of the wheel flange root. The brightness of the light source is steplessly adjusted by the main control unit, and the angle is adjusted by the auxiliary light source adjuster to suppress reflection on the wheel surface and improve the contrast of damage.

[0104] During image acquisition, the line scan camera rotates synchronously with the wheel to avoid motion blur and ensure image clarity.

[0105] The damage identification and classification steps include the following sub-steps:

[0106] Image preprocessing: Convert the acquired BMP format images to 1024x1250 resolution JPG format while maintaining 95% image quality;

[0107] Multi-threaded parallel detection: The images to be detected are divided into multiple batches of 10 images each, and multiple batches of parallel detection are achieved through a 3-thread pool;

[0108] Damage detection and filtering: Differentiated confidence thresholds are used, with a detection confidence threshold of 0.4 for cracks and diagonal cracks, and a detection confidence threshold of 0.3 for peeling and iron filings scratches, and an overlap frame filtering threshold of 0.3 is set.

[0109] Damage priority determination: Output the damage results of a single image according to the priority of peeling > crack > oblique crack > iron filings scratch;

[0110] Visual annotation and result output: Configure exclusive annotation colors for four types of damage, generate detection result images and return the results in JSON format.

[0111] In the damage identification and classification steps, the YOLOv8n custom model used is optimized in the following ways:

[0112] A C2f-CCFM cross-scale feature fusion module is introduced into layers 2 and 4 of the backbone network to enhance cross-scale feature interaction;

[0113] Optimize the head network structure, add a P2 / 4 scale detection head, and form a reciprocating feature fusion path;

[0114] The model achieved an accuracy of 96.15%, a recall of 97.47%, and a mean precision of 96.00% on a locomotive wheel surface damage dataset.

[0115] The method also includes a collaborative management step for damage data:

[0116] The main control unit associates and stores the damage identification results with the locomotive's entry position and axle sequence information, generates an inspection report, and uploads it to the server;

[0117] The system supports querying and statistical analysis of historical damage data, which is used for traceability of turning quality and support for operation and maintenance decisions.

[0118] The method achieves fully automated closed-loop control through a main control unit, including:

[0119] The main control unit monitors the lathe status in real time and dynamically adjusts the image acquisition parameters;

[0120] The algorithm analysis unit works in tandem with the mechanical module unit to ensure seamless integration of image acquisition and damage recognition, improving overall efficiency by more than 50%.

[0121] In the actual system workflow and control strategy, the system workflow includes five main stages: equipment initialization, mechanical positioning, image acquisition, damage analysis, and result output. The specific implementation process is as follows:

[0122] Equipment initialization phase: After the system is powered on, the main control unit sends enable signals to each servo driver via the EtherCAT bus. The mechanical module performs a homing operation, and each axis moves sequentially to the mechanical origin position. During initialization, the system automatically detects the status of each unit, including camera connection, light source brightness, network communication, etc., to ensure that the system is in a ready state.

[0123] Mechanical positioning stage: Based on the received locomotive model, axle sequence information, and lathe type, the main control unit retrieves the corresponding motion trajectory parameters from the preset parameter library. The system automatically selects system A or system B for operation by real-time analysis of the locomotive entry position signal. During positioning, the servo motor uses an S-curve acceleration and deceleration algorithm, resulting in smooth and shock-free movement. Taking a 3-axis mechanical module as an example, the X-axis first moves to the specified lateral position, then the Y-axis is adjusted vertically, and finally the Z-axis precisely adjusts the focusing distance. The entire positioning process can be completed within 15 seconds, with a positioning error ≤1.0mm.

[0124] Image acquisition phase: Once the mechanical module reaches the predetermined position, the main control unit sends a trigger signal to the line scan camera. The camera maintains strict synchronization with the wheel's rotation, achieving equidistant image acquisition through encoder feedback. During acquisition, the main light source provides basic illumination, while the auxiliary light source provides supplementary lighting to the wheel rim root. The light source brightness is automatically adjusted according to the wheel surface condition. Acquiring a single wheel takes approximately 10-15 seconds, and acquiring the entire wheelset can be completed within 30 seconds.

[0125] In the damage identification algorithm and data analysis, the algorithm analysis unit adopts the improved YOLOv8n-SMC model based on YOLOv8n, and its network structure is as follows: Figure 4 As shown, by introducing the C2f-CCFM module and optimizing the head network, the performance of small target detection is significantly improved.

[0126] The data processing workflow includes three steps: image preprocessing, model inference, and post-processing. In the preprocessing stage, the acquired BMP images are uniformly converted to 1024×1250 pixel JPG format, maintaining 95% image quality. Model inference employs multi-threaded parallel processing, grouping the images to be detected into batches of 10, and using a 3-thread pool for parallel computation. In the post-processing stage, a non-maximum suppression algorithm is applied, with an overlapping bounding box threshold of 0.3, ensuring that only the detection bounding box with the highest confidence is retained for the same damaged area.

[0127] The damage classification strategy sets differentiated thresholds based on the characteristic differences of four typical damage types: a confidence threshold of 0.4 for cracks and diagonal cracks, and a threshold of 0.3 for peeling and metal filings scratches. Detection results are sorted by priority (peeling > cracks > diagonal cracks > metal filings scratches) to meet the hierarchical management requirements of the maintenance site. Visual annotation uses a dedicated color system: cracks are yellow (255,255,0), peelings are magenta (255,0,255), diagonal cracks are cyan (0,255,255), and metal filings scratches are green (0,255,0).

[0128] The innovation of this system is reflected in the following aspects:

[0129] The mechanical structure, through its modular 2-axis / 3-axis mechanical modules, solves the installation challenge of limited space around the lathe. The mechanical modules feature adaptive positioning, automatically adjusting the data acquisition position according to wheel size, adapting to a full range of wheel sizes from 950-1250mm in diameter.

[0130] The optical system employs a dual-source soft light scheme, effectively suppressing reflections from metal surfaces through low-angle illumination and diffused light design. The light source brightness is infinitely adjustable, and combined with the global shutter technology of the line scan camera, it ensures high-contrast images are obtained under various working conditions.

[0131] The proposed YOLOv8n-SMC model performs excellently in locomotive wheel damage detection tasks, as shown in Table 2. Its accuracy, recall, and mAP reach 96.79%, 97.97%, and 96.11%, respectively, which are significantly better than traditional target detection algorithms.

[0132] The system integration, through intelligent scheduling by the main control unit, achieves full automation of mechanical motion, image acquisition, and damage analysis. The system seamlessly integrates with existing lathe repair processes, completing quality inspection without disrupting production schedules, and improving inspection efficiency by more than 50% compared to manual methods.

[0133] Example 1

[0134] This example demonstrates the system's application in a non-dismounting wheel turning and repair scenario at a railway locomotive depot. The depot is equipped with a CLG-3000 non-dismounting wheel turning machine and needs to inspect the surface quality of wheelsets on HXD3 electric locomotives after turning and repair. The locomotive wheels have a diameter of 1250mm, and damage detection and a report must be completed within 30 minutes after turning and repair. Traditional manual inspection suffers from low efficiency and high subjectivity. This system achieves efficient and accurate inspection through automation.

[0135] The system hardware was implemented strictly according to the documentation. The mechanical module unit adopted a dual-system layout of System A (installed on the front of the lathe) and System B (installed on the rear). Due to space constraints, a 3-axis mechanical module was selected in this example. The module drive motors were configured as follows: 400W brakeless servo motors for the X and Y axes, and a 400W brake-equipped motor for the Z axis, with positioning error controlled within ±0.5mm. The image acquisition unit used a Hikvision MV-CL042-91GM line scan camera, paired with an MVL AF4028M fixed-focus lens (40mm focal length, 200-1051mm working distance). The main light source power was 60.6W, and the auxiliary light source power was 26.4W, with brightness steplessly adjustable via the main control unit. The main control unit industrial computer was equipped with an i5-12400 processor, 32GB of memory, running Windows 10 for user interface interaction and Linux for real-time control.

[0136] During system initialization, the main control unit sends an enable signal to the servo driver via the EtherCAT bus, and the mechanical module performs a homing operation. Each axis moves to the mechanical origin using an S-curve acceleration / deceleration algorithm, taking approximately 10 seconds in total. During initialization, the system automatically detects the camera connection status (tested via LAN port ping), calibrates the light source brightness (default setting is 70%), and verifies network communication stability.

[0137] Once the locomotive wheelset enters the lathe station, the main control unit receives the locomotive model (HXD3), axle sequence, and storage end position signals. Based on the preset parameter library, the system automatically selects the right mechanical module of System A for operation. The mechanical module's movement trajectory is as follows:

[0138] The X-axis is moved laterally to a position 300mm from the wheel tread (based on a wheel diameter of 1250mm).

[0139] The Y-axis was moved vertically downwards by 150mm to avoid interference with the vehicle body;

[0140] Adjust the Z-axis focus distance to 800mm to ensure the depth of field covers the curved surface of the wheel.

[0141] After positioning is complete, the line scan camera triggers data acquisition. The camera synchronizes with the rotary encoder of the turning machine, scanning the wheel surface at an 80kHz line frequency. The main light source illuminates the tread at a 45° angle, while the auxiliary light source is angled at 30° towards the wheel rim root, with brightness dynamically adjusted to 85% to suppress metallic reflection. Acquired images are transmitted in real-time in BMP format; acquisition time for a single wheel is 12 seconds, and acquisition time for the entire wheelset is 24 seconds.

[0142] By using a line scan camera and a dual soft light source in synergy to create an imaging system, the optical design at the hardware level suppresses the high-brightness reflection of metal at the source, thus achieving clear imaging of the surface features of the wheel, which is an advantage of the line scan camera.

[0143] Linear scan cameras offer significant advantages in imaging scenarios related to wheel turning and repair on locomotives without wheel removal. Their photosensitive elements are arranged in a one-dimensional linear pattern, allowing for precise synchronization with wheel rotation for continuous scanning. This enables the acquisition of ultra-high-resolution two-dimensional images that clearly reveal key features such as micron-level surface textures and minute cracks, meeting the requirements for subsequent high-precision defect detection and dimensional measurement. Furthermore, the larger photosensitive area per pixel enhances photosensitive sensitivity, effectively suppressing noise and improving the image signal-to-noise ratio in complex reflective environments. Simultaneously, their scanning imaging method directly outputs one-dimensional image data corresponding to the wheel's circumference, eliminating the need for cropping and stitching area scan images. This significantly reduces the amount of data required for subsequent image processing, improving the overall operating efficiency of the detection system. It perfectly adapts to the dynamic continuous detection conditions of low-speed wheel rotation during wheel turning and repair operations, effectively avoiding problems such as motion blur and image distortion that easily occur when area scan cameras capture images dynamically.

[0144] Dual Soft Light Source Synergy Solution: This innovative solution employs a dual soft light source synergy, abandoning the traditional single direct-light hard light source. Both the main and auxiliary light sources utilize professional soft light sources with built-in diffusers, transforming direct light into uniform and soft diffused light, significantly reducing specular reflection intensity and preventing localized overexposure areas on the metal surface. The main light source illuminates the core detection area of ​​the wheel tread at a preset low angle to the camera's imaging path. This low-angle soft light illumination causes diffuse reflection rather than specular reflection on the smooth surface of the tread, ensuring sufficient illumination while avoiding... Reflected light enters the camera lens directly, and the light intensity can be dynamically adjusted according to the wheel material and rotation speed. The auxiliary light source is specifically positioned at a preset angle on the outer side of the wheel rim, obliquely illuminating the curved surface at the root of the wheel rim, a traditional blind spot for supplementary lighting. It fills the blind spot with diffused light, eliminates shadows, and avoids the formation of concentrated reflection points on the curved metal surface, thus complementing the main light source. In addition, the light intensity and activation sequence of the main and auxiliary light sources can be adjusted in conjunction with the control system, dynamically matching the camera scanning speed and the real-time status of the wheel. This provides stronger environmental adaptability and scene robustness, enabling clear imaging of the entire wheel tread and rim area without reflection.

[0145] For the confined space in the scenario of wheel turning without dismounting, the core limitation of this scenario lies in the extremely narrow working space around the wheel turning machine. Due to the space constraints of the main structure of the turning machine, the reserved gap under the locomotive body, and surrounding auxiliary equipment (such as hydraulic system and transmission mechanism), traditional large-scale inspection mechanical structures cannot be adapted. Therefore, it is necessary to design a compact servo motor + mechanical module to ensure that it can complete the full coverage acquisition of the entire surface of the wheel within the limited space, while avoiding interference with the turning machine, locomotive body and surrounding equipment.

[0146] To address the aforementioned space constraints, this solution adopts a core strategy of "zoning layout + compact multi-axis linkage". The designed mechanical module unit specifically includes System A and System B, which achieve efficient space utilization through front and rear partitioning installation: System A and System B are installed on the front and rear sides of the non-drop wheel lathe, respectively. Depending on the different lathe turning sequence, System A or System B is selected for operation to ensure no blind spots in detection.

[0147] To further adapt to the spatial and inspection needs of different scenarios, the mechanical module adopts a modular design. Depending on the specific model of the wheel turning machine and the locomotive type, it can be flexibly configured as a 2-axis or 3-axis mechanical module: The 2-axis mechanical module has two degrees of freedom of movement, lateral and longitudinal translation, which can meet the inspection coverage requirements of conventional spaces and conventional vehicle models, with a more streamlined structure and smaller footprint; The 3-axis mechanical module adds vertical movement to the 2-axis module, hiding the equipment under the rails, effectively saving space. It can adapt to special vehicle models with more confined spaces and more complex wheel structures, and adjusts the acquisition angle by rotation to avoid spatial interference while ensuring the rationality of the acquisition angle.

[0148] When the mechanical module is in standby mode, the overall clearance of the locomotive needs to be considered to ensure that the locomotive can pass through the wheel lathe normally without falling off.

[0149] One of the core innovations of this solution lies in the deep integration of the adaptive positioning function of the mechanical module with the scanning characteristics of the line scan camera, specifically addressing the issue of image clarity on curved wheel surfaces. Line scan cameras rely on continuous, uniform scanning motion to achieve high-definition imaging, but the wheel tread and rim are both curved surfaces. If the module positioning and camera scanning are not synchronized, it can easily lead to uneven image resolution and blurred edges in different areas of the curved surface. To address this, the system uses a collaborative control module to feed back the adaptive positioning data of the mechanical module (including the radius of curvature of the wheel surface and its real-time contact position) to the line scan camera control unit in real time. The camera dynamically adjusts the scanning frequency and exposure parameters based on this data—in areas with greater curvature, the scanning frequency is increased and the exposure time is optimized to ensure sufficient pixel density and clear details in that area; in areas with gentler curvature, the scanning frequency is appropriately reduced to improve detection efficiency. This collaborative mechanism of "adaptive positioning-dynamic scanning adaptation" fundamentally solves the problem of blurred curved surface imaging caused by the disconnect between mechanical positioning and image acquisition in traditional technologies, fully demonstrating the synergy and non-obviousness of technological integration.

[0150] In the entire process of locomotive wheel turning without removal from the lathe, there is a significant time window constraint in the post-turning measurement stage, which constitutes the core technical challenge for improving inspection efficiency in this scenario. Specifically, after the turning operation is completed, the wheel needs to undergo surface damage inspection before being removed from the turning machine; simultaneously, the turning machine drives the wheel to maintain a constant speed rotation, which is a necessary prerequisite for subsequent contact-based dimensional measurements. This solution innovatively utilizes this existing equipment condition, eliminating the need for additional drive mechanisms. It directly relies on the constant speed rotation of the wheel driven by the turning machine to complete the wheel surface image acquisition, fully utilizing the inherent working conditions in the production process while avoiding the space occupation and production rhythm disruption caused by additional drive devices. This achieves efficient integration of the inspection process with existing production processes. The inspection time for a single wheelset is ≤30 seconds.

[0151] In damage identification and algorithm analysis, the acquired images are forwarded from the main control unit to the algorithm analysis unit. The preprocessing module converts the images to a uniform 1024×1250 pixel JPG format (95% quality). Damage identification uses an improved YOLOv8n-SMC model, the network structure of which is as follows: Figure 4 As shown, the small target detection capability is enhanced by using the C2f-CCFM module.

[0152] The model inference stage employs multi-threaded parallel processing: the 30 images to be detected are divided into 3 batches (10 images per batch), and parallel computation is performed using a thread pool. Damage detection thresholds are set according to the documentation: the confidence threshold for cracks and diagonal cracks is 0.4, and the threshold for peeling and metal filings scratches is 0.3. In this example, a typical damage was detected: a peeling damage (ID1) with a confidence of 0.92 was found in the wheel tread area, and a crack (ID0) with a confidence of 0.45 was found at the wheel flange root. The algorithm outputs results according to a priority rule (peeling > crack) and filters redundant detection boxes using non-maximum suppression (overlapping box threshold 0.3).

[0153] After receiving the JSON-formatted results, the main control unit generates a visual report on the industrial computer interface. Damaged areas are marked with specific colors: peeling damage is outlined in purplish-red (255,0,255), and cracks are outlined in yellow (255,255,0). The system automatically triggers audible and visual alarms and stores the damage images and data to a 2TB SSD, while simultaneously uploading them to the locomotive depot's central server. The inspection report includes the damage type, location coordinates, confidence level, and handling recommendations. The overall processing time is ≤30 seconds, meeting real-time on-site requirements.

[0154] In this embodiment, the system's detection efficiency is improved by 55% compared to manual detection (manual detection takes approximately 5 minutes per round). Comparative experiments verify the algorithm's performance: the YOLOv8n-SMC model achieves an accuracy of 96.79%, a recall of 97.97%, and an mAP of 96.11% in this scenario, outperforming the traditional YOLOv8n model (92.06% accuracy). Specific data comparisons are shown in the table below:

[0155]

[0156] Table 1 Comparative test results

[0157] The system's mechanical compatibility has been verified with over 95% of models. In this example, the HXD3 wheel size (1250mm) perfectly matches the 1250-1150mm range of the mechanical module. Regarding optical imaging quality, the main and auxiliary dual light sources work together to eliminate reflections, improving the signal-to-noise ratio of the damaged area image to 35dB, ensuring that micron-level cracks are clearly visible.

[0158] Addressing the challenges of small, numerous, and easily overlooked surface damage on locomotive wheels after resurfacing, coupled with the low efficiency and error-prone nature of manual inspection, this chapter utilizes the advantages of machine vision technology. It employs the high-performance YOLOv8n model as a baseline for optimization research. Considering the baseline model's insufficient accuracy in detecting small cracks, an improved strategy is proposed, integrating the C2f-CCFM cross-scale feature fusion module and optimizing the head network structure to construct the YOLOv8n-SMC model. This aims to enhance the model's cross-scale feature recognition capabilities, improve the detection accuracy of four types of tread damage, and adapt to the detection needs within a short time window after resurfacing. Experimental test data comes from over 1000 wheels of AI-type locomotives at the locomotive depot. A dedicated dataset of over 30,000 locomotive wheel surface images was constructed through on-site collection and annotation. Mirroring and translation methods were used to enhance the data and improve the model's generalization ability.

[0159] The core of the improved scheme lies in the collaborative optimization of two modules: On the one hand, the C2f-CCFM module is introduced into the second and fourth layers of the backbone network. By leveraging its cross-connection fusion mechanism, the information barriers of features at different scales are broken down, the loss of shallow detail information is reduced, and the efficient complementarity between shallow fine spatial features and deep high-level semantic features is achieved, thereby improving the feature extraction effect of micro-cracks. On the other hand, a P2 / 4 scale detection head is added to the head network, and a complete downsampling feedback path is added to form a reciprocating feature fusion structure, which enhances the sensitivity and positioning accuracy of surface damage on wheels of different sizes.

[0160] Cross-scale feature fusion module (CCFM) is a feature enhancement structure optimized for small target detection, especially capable of capturing subtle features in complex scenes such as cracks and scratches. The core idea of ​​CCFM is to solve the problem in deep convolutional neural networks where deep features are rich in semantic information but suffer from severe loss of spatial details. CCFM essentially adopts a common PAFPN structure. Specifically, this structure first uses several layers of 1×1 convolution operations to uniformly map the number of channels of all features to the same value; then, it sequentially performs feature fusion work in two parts: top-down and bottom-up. CCFM is an optimization and improvement based on variant D, specifically by embedding several fusion blocks composed of convolutional layers in the fusion path. The main function of these fusion blocks is to effectively fuse adjacent features to generate new features. The analysis process is described in equations (1)-(3).

[0161] (1)

[0162] (2)

[0163] (3)

[0164] In the formula: Q, K, and V represent the query matrix, key matrix, and value matrix in self-attention, respectively; S3, S4, and S5 are the features of the last three stages; Flatten is the inverse operation; AIFI is multi-head self-attention; Reshape means restoring the feature shape to the same shape as S5.

[0165] The CCFM and C2f feature extraction modules are integrated to construct a high-performance C2f-CCFM module, which replaces the original C2f module in layers 2 and 4 of the backbone network.

[0166] Figure 4 As shown, the traditional YOLOv8n model uses three detector heads: P3 / 8, P4 / 16, and P5 / 32. To improve the detection sensitivity and localization accuracy of wheel tread surface damage, the YOLOv8n-SMC model first adds a P2 / 4 scale detector head to the head network. Through an additional upsampling module, the detection intensity is refined to a high-resolution feature map with a 4x downsampling. Simultaneously, the YOLOv8n-SMC model retains the upsampling fusion path of the traditional model while adding a complete downsampling feedback path. This reciprocating structure allows for more thorough interaction and complementarity when dealing with wheel surface damage of different sizes, improving the robustness of feature representation.

[0167] The second layer, located in the shallow layer of the backbone network, can accurately capture small-scale wheel surface damage features with less feature compression. The deeper fourth layer, through multiple rounds of convolution and feature fusion, integrates shallow detailed information with mid-level semantic information, carrying stronger high-level semantic information. Introducing the C2f-CCFM module at these two key locations can, on the one hand, reduce the loss of shallow detailed information by leveraging the residual connection characteristics of the C2f structure, and on the other hand, break down the information barriers between features of different scales through the cross-connection fusion mechanism of CCFM. It can fully utilize its powerful cross-scale information interaction capability to achieve more efficient complementary fusion of shallow fine spatial features and deep high-level semantic features, ensuring more comprehensive support for wheel tread and wheel surface damage features.

[0168] To verify the universality of the algorithm, different types of computer vision models were selected to conduct comparative experiments on the locomotive wheel tread surface damage dataset. The evaluation metrics selected were accuracy P, recall R, and mean precision mAP, and the calculation methods are shown in equations (4) and (5).

[0169] (4)

[0170] (5)

[0171] In the formula: TP is the number of wheel surface damages identified in the image containing wheel surface damage; FP is the number of wheel surface damages misidentified in the image without wheel surface damage; FN is the number of wheel surface damages missed in the image containing wheel surface damage.

[0172] The curve PR is plotted with P and R. The area under the curve is the AP value. The mAP value is the mean of AP for all categories. The calculation method is shown in equations (6) and (7).

[0173] (6)

[0174] (7)

[0175] A comparative experiment was conducted using mainstream object detection algorithms and YOLOv8n-SMC. The selected models included YOLOv3n, YOLOv5n, YOLOv5np6, YOLOv6n, and YOLOv8n, and the dataset used was the locomotive wheel surface damage dataset. The comparison results are shown in the table below.

[0176] Table 2 Comparative test results

[0177] As shown in the table above, a comparison of different metrics reveals that YOLOv8n-SMC exhibits a significant advantage over other object detection algorithms on the locomotive wheel surface damage dataset. The mean accuracy, recall, and average precision of YOLOv8n-SMC are 96.79%, 97.97%, and 96.11%, respectively, which are approximately 4 percentage points higher than the baseline model YOLOv8n. Experiments demonstrate that this approach effectively improves detection accuracy in wheel surface damage detection tasks.

[0178] Given the critical safety requirement of detecting surface damage on locomotive wheels, model performance optimization should prioritize recall, maximizing the identification of potential damage targets to avoid missed detections. Once the recall rate meets the safety threshold, precision can be further optimized to achieve a step-by-step improvement in detection performance.

[0179] To further verify the effectiveness of the proposed improvement scheme, an ablation experiment was conducted on a locomotive wheel surface damage dataset. The effects of adding the two optimization schemes individually and simultaneously on the YOLOv8n model were verified, and the results are shown in the table below, where "√" indicates the application of the scheme.

[0180] Table 3 Ablation experiment results based on modified YOLOv8n-SMC

[0181] As shown in the table above, for the locomotive wheel surface damage dataset, the addition of the C2f-CCFM module and the optimization of the head network structure both effectively improve the detection performance of wheel surface damage. Specifically, the C2f-CCFM module improves the accuracy and robustness of detection, increasing the mean accuracy and mean precision to 94.39% and 95.03%, respectively; the optimized head network structure enhances the recall capability of damage features, increasing the recall rate to 96.92% and the mean precision to 94.18%.

[0182] When the C2f-CCFM model is combined with the optimized head network, the model achieves optimal performance, with mean accuracy, recall, and mean precision reaching 96.79%, 97.97%, and 96.11%, respectively. The above analysis indicates that the two modules have a complementary and synergistic effect, significantly improving the overall performance of wheel surface damage detection, outperforming other models in terms of mean accuracy, recall, and mean precision.

[0183] Customized innovation in the inference stage of wheel surface damage recognition algorithm:

[0184] Multi-threaded parallel detection: To improve the detection efficiency of batch images and adapt to the performance of on-site edge computing devices, the images to be detected are split into multiple batches of 10 images each. Parallel detection of multiple batches is achieved through a 3-thread pool. Each thread independently completes model loading and image detection, avoiding process delays caused by single-threaded processing. Overall, it can support efficient parallel detection of 30 or more images. Single-round detection time ≤ 30 seconds.

[0185] Damage detection and filtering: Based on the actual severity of damage to the locomotive wheel surface and the level of attention paid to it on site, the damage underreporting rate is effectively reduced during the model inference stage. The threshold for damage feature identification is set separately as needed.

[0186] Cracks and diagonal cracks are early-stage damage to the surface of wheels. Their safety hazards are relatively controllable and have a certain margin of error in detection. Therefore, the detection confidence threshold for this type of damage is set at 0.4.

[0187] Peeling damage is in the middle to late stage of wheel surface damage evolution. If it is not identified and dealt with in time, it can easily cause serious driving safety hazards. In addition, it will significantly increase the amount of cutting during subsequent turning operations, resulting in resource waste. Therefore, it is necessary to promote early intervention and accurate early warning in the detection process. Thus, its detection confidence threshold is lowered to 0.3.

[0188] Although iron filings and scratches have a relatively low impact on structural safety, they are characterized by their hidden shape and ease of being missed. Furthermore, such defects directly affect the turning and finishing precision and final quality. Therefore, the detection confidence threshold is also set to 0.3 to achieve refined and differentiated accurate detection of different types of damage.

[0189] In terms of visualization of the detection results, a dedicated color-coding system is configured for the four types of damage: cracks are marked in yellow (255,255,0), peeling is marked in magenta (255,0,255), oblique cracks are marked in cyan (0,255,255), and iron filings scratches are marked in green (0,255,0). Simultaneously, an overlap box filtering threshold of 0.3 is set, and a non-maximum suppression algorithm is used to ensure that only the detection boxes with the highest confidence are retained for the same damage area, effectively eliminating redundant annotations and improving the readability and accuracy of the detection results.

[0190] Damage Priority Determination: Based on actual locomotive on-site maintenance cases and integrating existing theoretical research findings on the correlation between damage types and locomotive operational safety, a differentiated damage priority system is constructed: peeling > crack > diagonal crack > metal filings scratch. Through a priority-driven detection result output mechanism, the damage determination conclusion of a single image can accurately match the hierarchical control requirements of the maintenance process, providing a scientific basis for subsequent maintenance decisions.

[0191] Finally, it should be noted that the above embodiments are only used to describe the technical solutions of the present invention and not to limit the technical methods. The present invention can be extended to other modifications, variations, applications and embodiments, and therefore all such modifications, variations, applications and embodiments are considered to be within the scope of the present invention.

Claims

1. An automated wheel surface image acquisition and damage analysis system for locomotive wheel turning and repair scenarios without wheel removal, characterized in that, It includes a mechanical module unit, an image acquisition unit, an algorithm analysis unit, and a main control unit. These units work together through physical, electrical, and data connections. The mechanical module unit includes system A and system B. System A is installed on the front side of the wheelless turning machine through a mechanical module fixing device, and system B is installed on the rear side of the wheelless turning machine through a mechanical module fixing device. Each system contains 2 mechanical modules, corresponding to the left and right wheels respectively. The mechanical module drives the X-axis, Y-axis and Z-axis linear slides through servo motors. Each axis is connected by a synchronous transmission device and supported by a fixed bracket. The mechanical module unit is electrically connected to the main control unit through a cable harness and receives displacement commands and enable / disable signals. The image acquisition unit is rigidly connected to the end of the mechanical module unit via a camera mounting bracket, and includes a 3-axis linear array camera (15), a 2-axis linear array camera (151), a 3-axis fixed-focus lens (16), a 2-axis fixed-focus lens (161), a 3-axis main light source (17), a 2-axis main light source (171), a 3-axis auxiliary light source (18), and a 2-axis auxiliary light source (181). The linear array camera is threadedly connected to the fixed-focus lens via an M42 interface. The main light source and auxiliary light source are bolted to the camera mounting bracket via a bracket. The light source angle is manually adjusted via an auxiliary light source adjuster. The image acquisition unit is connected to the main control unit via Gigabit Ethernet to receive start / stop / parameter adjustment commands. The algorithm analysis unit is deployed on an industrial control all-in-one computer. It communicates with the image acquisition unit and the main control unit through the internal bus of the main control unit, receives image data and returns damage identification results. The main control unit integrates a PLC control system and a Windows system. It works in coordination with the mechanical module unit, image acquisition unit, and algorithm analysis unit through LAN interface, COM interface, and Ethercat master station to achieve fully automated operation of the entire system.

2. The automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal, as described in claim 1, is characterized in that... In systems A and B of the aforementioned mechanical module units, each mechanical module is configured as a 2-axis or 3-axis structure depending on the type of non-load-bearing turning machine: The 2-axis mechanical module includes an X-axis linear slide and a Z-axis linear slide. The X-axis linear slide is driven by an X-axis servo motor (91) and moves laterally through the X-axis mechanical module (111). The Z-axis linear slide is driven by a Z-axis servo motor (61) and moves longitudinally through the Z-axis mechanical module (71). The X-axis and Z-axis are connected by a fixed bracket (51) for the X-axis and Z-axis. The 3-axis mechanical module includes X-axis, Y-axis and Z-axis linear slides. The Y-axis linear slide is driven by the Y-axis servo motor (2) and moves vertically through the Y-axis mechanical module (4). The X-axis linear slide is driven by the X-axis servo motor (9) and moves laterally through the X-axis mechanical module (11). The Z-axis linear slide is driven by the Z-axis servo motor (6) and moves longitudinally through the Z-axis mechanical module (7). The Y-axis and Z-axis are connected by the Y-axis and Z-axis fixed brackets (5), and the Z-axis and X-axis are connected by the Z-axis and X-axis fixed brackets (8). The mechanical module has a positioning error of ≤1.0mm and is compatible with three wheel size ranges: 1250-1150mm, 1050-975mm, and 1050-950mm.

3. The automated wheel surface image acquisition and damage analysis system for locomotive wheel turning and repair scenarios according to claim 1, characterized in that, In the mechanical module unit, the servo motor is powered by 220V. The X-axis and Y-axis servo motors are not equipped with brakes, while the Z-axis servo motor is equipped with a brake. Each servo motor drives the slide table through a synchronous transmission device, which includes a synchronous pulley, a synchronous belt, and a protective cover. The mechanical module unit also includes a counterweight (13) to balance the motion inertia.

4. The automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal, as described in claim 1, is characterized in that... The image acquisition unit uses a Hikvision MV-CL042-91GM linear scan camera with a 4096×2-line CMOS sensor; a fixed-focus lens with a focal length of 40mm and a maximum aperture of F2.8; a main light source with a size of 200 * 200mm and a power of 60.6W, and an auxiliary light source with a size of 150 * 100mm and a power of 26.4W; the brightness of the light source is steplessly adjusted by the main control unit to suppress reflections from the metal surface of the wheel.

5. The automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal, as described in claim 1, is characterized in that... The algorithm analysis unit has a built-in damage recognition algorithm based on the YOLOv8n-SMC model. The industrial control all-in-one computer is equipped with an i5-12400 processor, 32G memory, and 128G SSD + 2T SSD hard drive. The algorithm analysis unit processes images in parallel through multi-threading to identify four types of damage: cracks, peeling, diagonal cracks, and iron filings scratches.

6. The automated wheel surface image acquisition and damage analysis system for locomotive wheel turning without wheel removal, as described in claim 1, is characterized in that... The main control unit expands external devices through multiple interfaces, including a DP interface for display connection, an HDMI interface for backup display, and a USB interface for data transmission, ensuring system scalability and stability.

7. The automated wheel surface image acquisition and damage analysis method for locomotive wheel turning and repair scenarios without wheel removal, as described in claim 1, is characterized in that... The method includes the following steps: Step 1, System Startup and Enablement: The main control unit sends an enable signal to the servo motor of the mechanical module unit, enabling the servo motor to enter the working ready state and resetting the mechanical module to the initial position; Step 2, Mechanical module movement steps: Based on the received locomotive model, locomotive entry position and turning and repair sequence signal, the main control unit selects system A or system B of the mechanical module unit to work, and issues displacement command to control the mechanical module to drive the image acquisition unit to move to the preset image acquisition position. Step 3, Image Acquisition Step: The line scan camera acquires the shooting signal from the main control unit, starts scanning at a preset frequency, and uses a high-brightness light source for supplementary lighting to achieve full-coverage image acquisition of the wheel surface, and transmits the acquired image to the main control unit in real time; Step 4, Equipment Reset Procedure: After image acquisition is completed, the main control unit sends a disabling signal to the servo motor, causing the mechanical module to reset to its initial position; Step 5, Damage Identification and Classification: The main control unit forwards the acquired images to the algorithm analysis unit, which performs preprocessing, multi-threaded parallel detection, damage detection and filtering, priority determination and visual annotation on the images through a customized damage identification model, and outputs the damage location, type and confidence level. Step 6, Result Output: The main control unit receives the recognition result and displays a visual image on the industrial control all-in-one computer screen. If damage is detected, an alarm is triggered, and the damage image and data are stored.

8. The method according to claim 7, characterized in that, The mechanical module movement in step two specifically includes: The main control unit controls the operation of either system A or system B of the mechanical module unit according to the type of the non-dismounting wheel lathe and the locomotive model. System A is installed on the front side of the lathe, and system B is installed on the rear side of the lathe. In the movement steps of the mechanical module, the main control unit automatically selects the movement trajectory and number of axes of the mechanical module according to the type of non-falling wheel lathe and the model of the testing machine: For the 2-axis mechanical module, the image acquisition unit is moved laterally and longitudinally by the coordinated movement of the X-axis linear slide and the Z-axis linear slide, which can adapt to the range of wheel sizes. For a 3-axis mechanical module, the vertical movement degree of freedom is increased by linking the linear slides of the X-axis, Y-axis and Z-axis, thus avoiding spatial interference. The image acquisition unit is precisely positioned within the depth-of-field coverage area of ​​the wheel surface, and is compatible with wheel sizes of 1250-1150mm, 1050-975mm and 1050-950mm. The motion trajectory of the mechanical module is automatically matched based on the built-in model parameter library, and the positioning error is ≤1.0mm.

9. The method according to claim 7, characterized in that, In the image acquisition step, the linear scan camera works in conjunction with a high-brightness light source to optimize image quality: The line scan camera scans the surface of the rotating wheel line by line at a maximum line frequency of 80kHz to avoid motion blur. The main light source illuminates the wheel tread at a low angle, while the auxiliary light source illuminates the root of the wheel flange at an angle, suppressing metallic reflection through diffused light. Fixed-focus lenses achieve fast and accurate focusing through the linkage of a manual focus ring and a Z-axis linear slide.

10. The method according to claim 9, characterized in that, The image acquisition specifically includes: The line scan camera uses the Hikvision MV-CL042-91GM model. The high-brightness light source includes a main light source and an auxiliary light source. The main light source is used for supplemental lighting of the wheel tread surface, and the auxiliary light source is used for supplemental lighting of the wheel flange root. The brightness of the light source is steplessly adjusted by the main control unit, and the angle is adjusted by the auxiliary light source adjuster to suppress reflection on the wheel surface and improve the contrast of damage. During image acquisition, the line scan camera rotates synchronously with the wheel to avoid motion blur and ensure image clarity.

11. The method according to claim 7, characterized in that, The damage identification and classification steps include the following sub-steps: Image preprocessing: Convert the acquired BMP format images to 1024x1250 resolution JPG format while maintaining 95% image quality; Multi-threaded parallel detection: The images to be detected are divided into multiple batches of 10 images each, and multiple batches of parallel detection are achieved through a 3-thread pool; Damage detection and filtering: Differentiated confidence thresholds are used, with a detection confidence threshold of 0.4 for cracks and diagonal cracks, and a detection confidence threshold of 0.3 for peeling and iron filings scratches, and an overlap frame filtering threshold of 0.3 is set. Damage priority determination: Output the damage results of a single image according to the priority of peeling > crack > oblique crack > iron filings scratch; Visual annotation and result output: Configure exclusive annotation colors for four types of damage, generate detection result images and return the results in JSON format.

12. The method according to claim 7, characterized in that, In the damage identification and classification steps, the YOLOv8n custom model used is optimized in the following ways: A C2f-CCFM cross-scale feature fusion module is introduced into layers 2 and 4 of the backbone network to enhance cross-scale feature interaction; The head network structure was optimized, and a P2 / 4 scale detection head was added to form a reciprocating feature fusion path.

13. The method according to claim 7, characterized in that, The method also includes a collaborative management step for damage data: The main control unit associates and stores the damage identification results with the locomotive's entry position and axle sequence information, generates an inspection report, and uploads it to the server; The system supports querying and statistical analysis of historical damage data, which is used for traceability of turning quality and support for operation and maintenance decisions.

14. The method according to claim 7, characterized in that, The method achieves fully automated closed-loop control through a main control unit, including: The main control unit monitors the lathe status in real time and dynamically adjusts the image acquisition parameters; The algorithm analysis unit works in tandem with the mechanical module unit to ensure seamless integration of image acquisition and damage identification.

Images