An automatic identification method for loose suspension components under a high-speed train

Abstract

This invention discloses an automatic identification method for loose suspension components under high-speed trains, belonging to the field of railway intelligent operation and maintenance and machine vision technology. The method includes: acquiring high-definition images of the train underside using an image acquisition device deployed in the maintenance trench; preprocessing the images and segmenting them into regions of interest (ROIs) containing fasteners and anti-loosening marks; inputting the ROIs into an improved key point detection network to identify the coordinates of the bolt head center point, nut vertex, and anti-loosening mark line endpoints; constructing a geometric analysis model including benchmark determination, vector calculation, double-line verification, and deviation calculation modules to calculate angular offset or displacement; calculating the consistency rate of loosening determination through a combination of single-frame judgment and multi-frame verification; when the consistency rate exceeds a threshold and the offset direction is consistent, a loosening is finally determined, and a graded warning is implemented based on the degree of offset. This invention achieves fully automatic and high-precision identification of loose fasteners under the train underside, effectively reducing the missed detection rate and false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of railway intelligent operation and maintenance and machine vision technology, specifically to an automatic identification method for loose suspension components under a high-speed train. Background Technology

[0002] The suspension system of a high-speed train is the core guarantee for the safe operation of the train. It contains a large number of critical fasteners (such as bolts and nuts). These fasteners play an important role in fixing suspension components such as shock absorbers, anti-roll torsion bars, and traction rods, and in transmitting loads. Under the long-term high-speed operation and continuous vibration and impact of the train, the fasteners are very prone to loosening. If they are not detected and dealt with in time, they may lead to major safety accidents such as suspension components falling off and train derailment. Therefore, accurate detection of the loosening status of the fasteners of the suspension components under the train is a key link in the operation and maintenance of high-speed trains.

[0003] Currently, the inspection of the fasteners of the undercarriage suspension components after a high-speed train enters the depot mainly adopts the following two methods:

[0004] I. Manual visual inspection

[0005] Maintenance personnel, armed with flashlights, meticulously inspected the anti-loosening markings (continuous lines drawn with a paint pen) on each fastener in the cramped, dimly lit, and structurally complex space under the vehicle. By observing whether these markings were misaligned or broken, they determined if the bolts and nuts were loose. This method has the following drawbacks:

[0006] First, the inspection efficiency is extremely low. A train set contains tens of thousands of fasteners, and it is time-consuming and labor-intensive to check them one by one manually. The maintenance work takes a long time and seriously affects the turnaround efficiency of the train set.

[0007] Secondly, the rate of missed detection is high. The environment under the vehicle is dark and cramped, and there is oil and pipe obstruction, which can easily cause visual fatigue for the human eye. The rate of missed detection increases significantly, especially after long-term continuous operation.

[0008] Third, it is highly subjective and lacks consistency. The judgment results are affected by subjective factors such as the maintenance personnel's experience, eyesight, and working conditions at the time. It is difficult to guarantee the consistency of test results from different personnel and different batches.

[0009] II. Simple Machine Vision Inspection

[0010] In some maintenance scenarios, cameras were deployed to capture images of the vehicle's underside. However, these only performed simple dimensional measurements or surface defect detection, failing to accurately identify subtle displacements of anti-loosening markings. This resulted in the following shortcomings:

[0011] First, the robustness is poor and the anti-interference ability is weak. The environment under the vehicle has problems such as oil stains reflecting light, uneven lighting, and specular reflection of metal surfaces. Existing solutions lack targeted image preprocessing algorithms, and the quality of the acquired images is unstable, resulting in low recognition rate and high false alarm rate.

[0012] Second, it cannot identify minute displacements and makes it difficult to detect early loosening. Existing solutions can only detect obvious misalignment of anti-loosening marks (such as complete breakage or large displacement of the marking line), and cannot effectively identify early latent minute loosenings (such as angular displacement of 1-3 degrees or pixel-level displacement), thus failing to achieve early warning of faults.

[0013] Third, the lack of a geometric analysis model and the single dimension of judgment mean that the existing solution relies only on a single image feature for judgment and has not built a geometric analysis model for anti-loosening marks. It is unable to comprehensively calculate multi-dimensional parameters such as angle offset, displacement, and double-line distance difference, resulting in insufficient reliability of the judgment results.

[0014] Fourth, the lack of multi-frame verification logic makes it susceptible to transient interference. When the train moves in the maintenance trench, there are transient interferences such as vehicle vibration and dust obstruction. The existing solution uses a single-frame image for judgment, which is very easy to misjudge transient interference as looseness, resulting in a high false alarm rate and affecting maintenance efficiency.

[0015] Fifth, the testing process has not formed an automated closed loop. The existing solution can only achieve image acquisition or simple recognition, and cannot automatically generate maintenance reports. The test results need to be manually confirmed, which further increases the operation and maintenance costs and workload.

[0016] Therefore, there is an urgent need for an automatic identification method for loosening status of train undercarriage suspension components that can overcome interference from the complex environment under the train, accurately identify minute displacements of anti-loosening marks, and possess a geometric analysis model and multi-frame verification logic, in order to address the shortcomings of the existing technologies. Summary of the Invention

[0017] To address the problem mentioned in the background art regarding the inaccurate state detection of the suspension components under the train, this invention provides an automatic identification method for the loose state of the suspension components under the train.

[0018] The above-mentioned objective of this application is achieved through the following technical solution:

[0019] An automatic identification method for loose suspension components under a high-speed train, comprising the following steps:

[0020] By deploying an image acquisition device in the maintenance trench, high-definition visible light images of the undercarriage suspension components are captured when the train passes through the inspection area at low speed.

[0021] The acquired raw images are preprocessed to segment out rectangular boxes containing fasteners and anti-loosening marks from the complex background under the vehicle, which are then used as regions of interest.

[0022] The region of interest is input into a pre-trained key point detection network to identify and locate the coordinates of the bolt head center point, the nut polygon vertex, and the endpoint of the anti-loosening mark line;

[0023] A geometric analysis model is constructed based on the coordinates of each key point in the positioning to calculate the angular offset or feature point displacement of the anti-loosening mark line relative to the fastener fixing reference. The geometric analysis model includes a benchmark determination module, a vector calculation module, a dual-line verification module, and a deviation calculation module.

[0024] The calculated angular offset or feature point displacement is compared with a preset safety threshold to determine whether the fastener is loose and generate an inspection report.

[0025] By adopting the above technical solution, firstly, the image acquisition device deployed in the maintenance trench automatically triggers shooting when the EMU passes at low speed, realizing automated acquisition of images of the undercarriage suspension components, eliminating the need for manual shooting one by one and greatly improving detection efficiency; secondly, through preprocessing and region of interest segmentation, interference from complex backgrounds under the train (such as pipes, cables, oil stains, shadows, etc.) is effectively filtered out, allowing subsequent key point detection to focus on fasteners and anti-loosening marking areas, improving the targeting and accuracy of recognition; then, a pre-trained key point detection network is introduced, which can accurately locate the center point of the bolt head and the polygonal top of the nut. Key coordinates, such as the endpoints of the anti-loosening marking lines, provide a reliable data foundation for geometric analysis. A geometric analysis model is constructed, including a benchmark determination module, a vector calculation module, a dual-line verification module, and a deviation calculation module. This model comprehensively calculates the angular offset or feature point displacement of the anti-loosening marking lines relative to a fixed reference object from multiple dimensions, overcoming the problem of insufficient reliability of a single judgment dimension. Finally, by comparing with a preset safety threshold, the system automatically determines whether the fastener is loose and generates a maintenance report, forming a complete automated closed loop of "collection—processing—identification—calculation—judgment—reporting," providing maintenance personnel with accurate maintenance basis.

[0026] In a preferred embodiment, this application can be further configured such that the key point detection network adopts an improved YOLOv5 network or a CenterNet network, and its loss function includes a regularization term for key point offset, which is used to improve the detection accuracy of small displacements of the anti-loosening mark.

[0027] By adopting the above technical solution, a regularization term for keypoint offset is added to the loss function based on the YOLOv5 or CenterNet network. This regularization term can constrain the fitting accuracy of small displacements during network training, so that the positioning error of the network for the endpoints of the anti-loosening marker is controlled at the sub-pixel level. Compared with conventional keypoint detection networks, it improves the detection sensitivity of early small misalignments of anti-loosening markers (such as 1-3 degree angular offset or pixel-level displacement), and solves the technical problem that existing solutions are unable to detect latent loosening faults.

[0028] In a preferred embodiment, this application can be further configured such that the benchmark determination module specifically includes:

[0029] Using the center point of the bolt head output by the key point detection network as the origin of the coordinate system, determine whether the bolt has chamfered or cotter pin features;

[0030] If it exists, then use this feature to determine the absolute vertical reference or the absolute horizontal reference;

[0031] If it does not exist, the line connecting the corner point of the nut and the center point of the bolt is used as the relative reference line.

[0032] By adopting the above technical solution, a unified geometric analysis coordinate system is first established with the center point of the bolt head as the origin. Then, depending on whether the bolt has chamfered or cotter pin features, an adaptive datum determination strategy is selected: when such features exist, their geometric constraints are used to determine an absolute vertical or horizontal datum, improving the stability and reproducibility of the datum; when such features do not exist, the line connecting the corner point of the nut and the center point of the bolt is used as the relative datum line to ensure the universality of the solution. This adaptive datum determination mechanism enables the geometric analysis model to be compatible with fasteners of different types and structures, improving the applicability of this method on different EMU models and suspension components.

[0033] In a preferred embodiment, this application can be further configured such that the vector calculation module specifically includes:

[0034] Obtain the coordinates of the two ends of the anti-loosening mark line output by the key point detection network, and calculate the coordinates of the midpoint of the anti-loosening mark line;

[0035] Calculate the angle between the vector pointing from the origin to the midpoint and the baseline, and use it as the first loosening judgment parameter;

[0036] Calculate the vertical offset distance of the midpoint of the anti-loosening mark line relative to the baseline, and use it as the second loosening judgment parameter.

[0037] By adopting the above technical solution, the coordinates of the midpoint of the anti-loosening mark line are first calculated, condensing the position information of the mark line into a feature point, simplifying subsequent geometric calculations. Then, the angle between the vector from the origin to the midpoint and the baseline (as the first loosening judgment parameter) and the vertical offset distance of the midpoint relative to the baseline (as the second loosening judgment parameter) are calculated respectively. The two parameters comprehensively characterize the offset state of the anti-loosening mark from two dimensions: angle and distance, and complement each other: the angular offset is sensitive to rotational loosening, and the vertical offset distance is sensitive to translational loosening; effectively improving the comprehensiveness and robustness of loosening identification.

[0038] In a preferred embodiment, this application can be further configured such that the dual-line verification module specifically includes:

[0039] For nuts containing two anti-loosening marking lines, the first distance value of the first anti-loosening marking line relative to the bolt edge and the second distance value of the second anti-loosening marking line relative to the bolt edge are obtained respectively.

[0040] Calculate the change in the difference between the first distance value and the second distance value, as well as the rate of change of the difference.

[0041] When the change in the difference exceeds the preset first distance threshold or the rate of change in the difference exceeds the preset second distance threshold, a double-line verification anomaly flag is generated to help determine the loose state of the fastener.

[0042] By adopting the above technical solution, for nuts with two anti-loosening marking lines, the distance values ​​of the two marking lines relative to the bolt edge are calculated, and the change amount and rate of change of the distance difference are further calculated. When the change amount or rate of change of the difference exceeds a preset threshold, a double-line verification anomaly mark is generated. For fasteners with double marking lines, the relative positional relationship of the two marking lines is an important basis for judging looseness. If the two marking lines are offset synchronously, it may be that the whole rotation is loose. If the two marking lines are offset asynchronously or the relative distance changes significantly, it may indicate a more complex loosening mode or local detachment of the marking lines. This provides additional auxiliary features for loosening judgment and further improves the reliability of the judgment results.

[0043] In a preferred embodiment, this application can be further configured such that the deviation calculation module specifically includes:

[0044] Obtain the initial angle or initial displacement value of the fastener in a standard installation state;

[0045] Calculate the angle difference between the currently calculated included angle and the initial angle value, and use it as a core loosening determination parameter;

[0046] Alternatively, the displacement difference between the currently calculated vertical offset distance and the initial displacement value can be used as an auxiliary loosening judgment parameter.

[0047] When the absolute value of the angle difference exceeds the preset angle threshold, the fastener is determined to be in a suspected loose state.

[0048] By adopting the above technical solution, the initial angle value or initial displacement value of the fastener in the standard installation state is first obtained as a reference; then the difference between the current detection value and the initial value is calculated, where the angle difference is used as the core loosening judgment parameter and the displacement difference is used as the auxiliary loosening judgment parameter; by comparing with the preset angle threshold, when the absolute value of the angle difference exceeds the standard, it is judged as a suspected loosening state; by using relative deviation rather than absolute quantity as the judgment basis, the influence of individual differences of different fasteners (such as the initial line position of the anti-loosening mark is not completely consistent) on the judgment result is effectively eliminated.

[0049] In a preferred embodiment, this application can be further configured such that: the determination of whether the fastener is loose includes single-frame determination logic: calculating the angle offset or feature point displacement of the current single-frame image and comparing it with a preset safety threshold, and outputting a single-frame suspected looseness determination result.

[0050] By adopting the above technical solution, the loosening determination of the current single frame image can be realized: the complete angle offset or feature point displacement of each frame image is calculated and compared with the preset safety threshold, and the single frame suspected loosening determination result is output in real time; the single frame determination logic has the characteristics of low computation and fast response speed, which can meet the real-time processing needs of the EMU passing through the detection area at low speed, and provides basic determination data for subsequent multi-frame verification.

[0051] In a preferred embodiment, this application can be further configured to include multi-frame verification logic.

[0052] Obtain multiple angular offsets or feature point displacements calculated for the same fastener in multiple consecutive frames of images;

[0053] Statistical analysis is performed on the multiple angular offsets or feature point displacements to calculate the consistency rate of loosening determination;

[0054] If the consistency rate exceeds the preset consistency rate threshold and the offset direction of each frame is consistent, it will be finally determined as "loose".

[0055] If the consistency rate is lower than the preset consistency rate threshold, it is determined to be "instantaneous interference" and the alarm is ignored.

[0056] By adopting the above technical solution, a multi-frame verification mechanism is introduced on the basis of single-frame judgment: the calculation results of multiple angular offsets or displacements of the same fastener in consecutive multi-frame images are obtained, statistically analyzed, and the loosening judgment consistency rate is calculated; when the consistency rate exceeds the preset threshold and the offset direction of each frame is consistent, it is finally judged as "loose"; when the consistency rate is lower than the threshold, it is judged as "instantaneous interference" and the alarm is ignored; it effectively eliminates single-frame misjudgment caused by instantaneous factors such as train body vibration, dust obstruction, and oil stain reflection, and greatly reduces the false alarm rate; compared with the traditional single-frame detection scheme, this multi-frame verification scheme improves the detection accuracy by about 30% and reduces the false alarm rate by more than 50%.

[0057] In a preferred example, this application can be further configured to include tiered warning logic.

[0058] When the absolute value of the angle offset is within the first angle threshold range, it is determined as a level one warning and a reminder is generated;

[0059] When the absolute value of the angle offset is within the second angle threshold range, it is determined to be a level two warning and a planned maintenance suggestion is generated.

[0060] When the absolute value of the angular offset is within the third angular threshold range, it is determined to be a level three warning, and an immediate shutdown and maintenance alarm is generated.

[0061] By adopting the above technical solution, a three-level early warning mechanism is set up according to the magnitude of different angular offsets: the first level (small angle offset) generates a notice to remind maintenance personnel to pay attention; the second level (medium angle offset) generates a planned maintenance suggestion to be included in the routine maintenance plan; the third level (large angle offset) generates an immediate shutdown maintenance alarm, requiring immediate handling. This hierarchical early warning logic directly links the degree of looseness with the maintenance response level, avoiding the information loss caused by the binary judgment of "either loose or tight", providing maintenance personnel with differentiated handling basis, and helping to rationally allocate maintenance resources and prioritize the handling of high-risk faults.

[0062] In a preferred embodiment, this application can be further configured such that the generated maintenance report includes:

[0063] Location information of loose fasteners, quantitative value of looseness, captured images, multi-frame verification results, and graded warning levels;

[0064] The inspection report is pushed to the maintenance terminal via wireless network, and the specific location of the loose fastener is highlighted on the map of the maintenance terminal.

[0065] By adopting the above technical solution, the generated inspection report includes complete information such as the location information of loose fasteners, the quantitative value of the loosening degree (angle or displacement), captured images, multi-frame verification results, and graded warning levels. At the same time, the report is pushed to the maintenance terminal in real time via wireless network, and the specific location of the loose fasteners is highlighted on the maintenance terminal's map. This achieves visualization, traceability, and location of the inspection results. Maintenance personnel do not need to search through the huge vehicle undercarriage structure one by one. They can quickly reach the fault point directly according to the map markings, improving the efficiency and accuracy of maintenance work, and truly realizing a fully intelligent closed loop of "detection-judgment-early warning-location-repair".

[0066] In summary, this application includes at least one of the following beneficial technical effects:

[0067] 1. By using machine vision technology to replace manual visual inspection, the entire process from image acquisition, preprocessing, key point recognition, geometric calculation, looseness judgment to maintenance report generation is automated, eliminating the need for manual inspection of fasteners one by one. Compared with manual inspection, the inspection time of a train set can be shortened by more than 80%, significantly reducing the workload of maintenance personnel and improving the turnover efficiency of train sets entering the depot for maintenance.

[0068] 2. An improved YOLOv5 or CenterNet keypoint detection network is adopted, and a regularization term with keypoint offset is added to the loss function, which can control the positioning error of the endpoints of the anti-loosening marker line to the sub-pixel level. Combined with the constructed geometric analysis model (including the benchmark determination module, vector calculation module, dual-line verification module and deviation calculation module), the angular offset (accurate to within 1 degree) and feature point displacement (pixel level) can be accurately calculated, which can detect early latent loosening faults in advance and avoid the fault from escalating and causing safety accidents.

[0069] 3. Employing a logic combining single-frame judgment and multi-frame verification, the system statistically analyzes the angular offsets of the same fastener across multiple consecutive frames to calculate the consistency rate of loosening judgment. Only when the consistency rate exceeds a preset threshold (e.g., 80%) and the offset directions are consistent across frames is the fastener ultimately judged as "loose"; otherwise, it is judged as "instantaneous interference" and the alarm is ignored. This effectively eliminates single-frame misjudgments, reducing the false alarm rate by more than 50%. Furthermore, the detection results are not affected by subjective factors such as the maintenance personnel's experience, eyesight, or sense of responsibility, and the consistency of detection results across different batches and operations by different personnel is strong.

[0070] 4. A three-level early warning mechanism is set up according to different ranges of angular offset (Level 1 early warning generates attention reminder, Level 2 early warning generates planned maintenance suggestion, and Level 3 early warning generates immediate shutdown and maintenance alarm), directly linking the degree of loosening with the maintenance response level; the generated maintenance report includes the location information of the loose fastener, the quantitative value of the degree of loosening, the captured image, the multi-frame verification result and the graded early warning level, and is pushed to the maintenance terminal via wireless network, highlighting the specific location of the loose fastener on the electronic map; maintenance personnel can quickly locate the fault point according to the map markings, realizing a fully intelligent closed loop of "detection-judgment-early warning-location-maintenance". Attached Figure Description

[0071] Figure 1 This is a flowchart of an embodiment of the automatic identification method for loose suspension components under a high-speed train according to this application;

[0072] Figure 2 is a side view showing the relative positions of the image acquisition device, the EMU, and the maintenance trench in an embodiment of this application;

[0073] Figure 3 is a front view of the installation position of the image acquisition device on the side wall of the trench in an embodiment of this application;

[0074] Figure 4 is a flowchart of the implementation of the dual-line verification module of the image acquisition device in this embodiment of the application;

[0075] Figure 5 is a schematic diagram illustrating the principle of anti-loosening mark recognition and geometric analysis in the embodiments of this application;

[0076] Figure 6 is a flowchart of the implementation of the multi-frame motion trajectory verification logic in the embodiments of this application. Detailed Implementation

[0077] The following is in conjunction with the appendix Figure 1-6 This application will be described in further detail.

[0078] In one embodiment, such as Figure 1-3 As shown, this application discloses an automatic identification method for the loose state of suspension components under a high-speed train, which specifically includes the following steps:

[0079] S10: By deploying an image acquisition device in the maintenance trench, high-definition visible light images of the undercarriage suspension components are captured when the train passes through the inspection area at low speed.

[0080] In this embodiment, the maintenance trench refers to the trench structure set below the EMU when it enters the depot for maintenance; the image acquisition device refers to a high-speed industrial camera deployed on the side wall of the maintenance trench or beside the track; low-speed passage refers to the EMU passing through the detection area at a speed of 5-15 km / h, which ensures the clarity of image acquisition without excessively affecting the turnaround efficiency of the EMU for maintenance; triggered shooting refers to the simultaneous triggering of the camera and supplementary lighting equipment to take pictures when the EMU is detected to have arrived at the detection area by photoelectric sensors or wheel axle sensors.

[0081] Specifically, linear camera arrays are deployed on both sides of the maintenance trench, such as... Figure 2 and Figure 3 As shown, each array contains 3-4 high-speed industrial cameras, supplemented by high-frequency LED strobe lights. When the train enters the detection area at a speed of approximately 10 km / h, the photoelectric sensors beside the track detect the train's arrival and immediately send a trigger signal to the camera control system. Based on the train's speed and the preset shooting interval, the system controls the cameras to continuously acquire high-definition visible light images of both sides of the train's undercarriage and key areas of the bogie at a scanning frequency of 2-3 lines per millimeter. The system simultaneously records the train's speed information at the current acquisition moment, which is obtained through radar speed measurement or encoder ranging, providing data support for motion compensation and interference removal in subsequent image processing.

[0082] S20: Preprocess the acquired raw image to segment out a rectangular frame containing fasteners and anti-loosening marks from the complex background under the vehicle, which is then used as the region of interest.

[0083] In this embodiment, preprocessing refers to a series of enhancement and denoising operations performed on the original image, mainly including: using the MSRCP algorithm to eliminate reflections and shadows in the image and enhance details in dark areas; using Gaussian filtering or median filtering to remove image noise; the MSRCP algorithm, namely the multi-scale Retinex algorithm with chroma preservation, is an image enhancement algorithm that can effectively compress the dynamic range and enhance the detail information in shadow areas while preserving the color fidelity of the image, and is suitable for complex environments such as the underside of a car where there is uneven lighting and metallic reflection; converting the image to the HSV color space, which refers to a color model that describes color in three dimensions: hue, saturation, and brightness. Compared with the RGB space, the HSV space can separate color information from brightness information, making it easier to perform target segmentation based on color features; the region of interest (ROI) refers to the rectangular area containing fasteners and anti-loosening marks segmented from the original image.

[0084] Specifically, the first step is to preprocess the acquired raw images: the MSRCP algorithm is used to eliminate specular reflections on the metal surface and shadows caused by pipe obstructions, while enhancing the dark details in the areas where bolts, nuts, and anti-loosening marks are located, making the color characteristics of the anti-loosening marks more vivid; then, Gaussian filtering is used to remove random noise generated during image acquisition; the preprocessed image is converted from the RGB color space to the HSV color space, and taking advantage of the characteristic that anti-loosening marks are usually high-saturation colors, such as red, yellow, and white, corresponding hue threshold ranges are set, such as red hue range of 0-10 degrees and 160-180 degrees, and saturation thresholds, such as S greater than 0.5, to perform coarse color segmentation and initially extract candidate regions for anti-loosening marks; then, the Canny edge detection algorithm is combined to extract the edge contour information of the undercarriage components, and contour analysis and morphological operations, such as closing operations to fill holes and opening operations to remove small noise, are used to finally segment the minimum bounding rectangle containing bolts, nuts, and anti-loosening marks from the complex undercarriage background as the region of interest.

[0085] S30: Input the region of interest into a pre-trained key point detection network to identify and locate the coordinates of the bolt head center point, the nut polygon vertex, and the endpoint of the anti-loosening mark line.

[0086] In this embodiment, the keypoint detection network refers to a deep learning model used to identify and locate the coordinates of specific keypoints in an image. This application uses an improved YOLOv5 network or a CenterNet network. YOLOv5 is a convolutional neural network based on single-stage object detection, characterized by fast detection speed and high accuracy. Through improvement, it can simultaneously output object detection boxes and keypoint coordinates. CenterNet is a keypoint detection network based on the target center point, transforming the object detection problem into a keypoint detection problem, and has high localization accuracy for small targets. The bolt head center point refers to the coordinates of the geometric center of the head of a hexagonal head bolt or internal hexagonal bolt. The nut polygon vertices refer to the coordinates of the six corner points of the hexagonal nut, used to determine the rotation angle and geometric position of the nut. The endpoints of the anti-loosening mark line refer to the starting and ending coordinates of the anti-loosening mark line drawn with a paint pen, used to calculate the direction and midpoint position of the mark line.

[0087] Specifically, the region of interest image extracted in step S20 is input into a pre-trained keypoint detection network. During the training phase, this network has learned the typical morphological features of fasteners and their anti-loosening marks through a large number of labeled samples, and can accurately regress the sub-pixel coordinates of the bolt head center point, the six vertices of the nut, and the two endpoints of the anti-loosening mark line. To improve the robustness of localization, a regularization constraint term for keypoint offset is introduced into the network structure, which effectively suppresses coordinate drift caused by image blurring, partial occlusion, or changes in illumination. The output keypoint coordinates will be used as input to the subsequent geometric analysis module to construct the reference vector and the current state vector, and then calculate the angle offset and displacement.

[0088] S40: Construct a geometric analysis model based on the coordinates of each key point of the positioning, and calculate the angular offset or feature point displacement of the anti-loosening mark line relative to the fastener fixing reference. The geometric analysis model includes a benchmark determination module, a vector calculation module, a double-line verification module, and a deviation calculation module.

[0089] In this embodiment, the geometric analysis model refers to a mathematical model constructed based on the coordinates of key points for calculating the offset of the anti-loosening mark, including four sub-modules: a reference determination module, a vector calculation module, a double-line verification module, and a deviation calculation module; the fixed reference object refers to the fixed component used as a reference benchmark in the judgment of fastener loosening, such as the edge of the bolt head, the corner of the nut, the chamfer feature of the bolt, etc.; the angular offset refers to the difference in angle between the current direction of the anti-loosening mark line and the initial direction under the standard installation state; the feature point displacement refers to the vertical offset distance of a feature point such as the midpoint on the anti-loosening mark line relative to the fixed reference object.

[0090] Specifically, based on the coordinates of each key point extracted in step S30, the four sub-modules of the geometric analysis model are called sequentially to perform calculations:

[0091] The reference determination module first establishes a local coordinate system with the center point of the bolt head as the origin. It then uses image recognition to determine whether the bolt has chamfer or cotter pin features. Chamfer features are represented as a bright ring on the edge of the bolt head, while cotter pin features are represented as a metal wire passing through the bolt. If chamfer features are present, the perpendicular line from the chamfer edge is used as the absolute vertical reference. If cotter pin features are present, the axial direction of the cotter pin is used as the absolute horizontal reference. If neither of these features is present, the line connecting the nut corner point and the bolt center point, i.e., the corner point closest to the bolt center among the six corner points, is used as the relative reference line.

[0092] The vector calculation module obtains the coordinates of the two ends of the anti-loosening mark line, P1(x1,y1) and P2(x2,y2), and calculates the coordinates of the midpoint M as ((x1+x2) / 2, (y1+y2) / 2); calculates the vector OM pointing from the origin O to the midpoint M; calculates the angle θ between the vector OM and the baseline, with a value ranging from 0 degrees to 180 degrees, as the first loosening judgment parameter; and calculates the vertical offset distance d of the midpoint M relative to the baseline, in pixels, as the second loosening judgment parameter.

[0093] The dual-line verification module is designed for nuts with two anti-loosening marking lines, such as those used in certain fastener models. It calculates the vertical distance L1 from the first marking line to the bolt edge and the vertical distance L2 from the second marking line to the bolt edge. It also calculates the difference change ΔL = |L1 - L2| and the difference change rate v = d(ΔL) / dt, which requires time series calculation of multiple frames. When ΔL exceeds a preset first distance threshold (e.g., 3 pixels) or v exceeds a preset second distance threshold (e.g., 1 pixel per frame), a dual-line verification anomaly is generated.

[0094] The deviation calculation module retrieves the initial angle value θ0 and initial displacement value d0 of the fastener in the standard installation state from the database; calculates the current angle difference Δθ = |θ - θ0| as the core loosening judgment parameter; calculates the current displacement difference Δd = |d - d0| as the auxiliary loosening judgment parameter; when Δθ exceeds the preset angle threshold, such as 3 degrees, the fastener is judged to be in a suspected loose state.

[0095] S50: Compare the calculated angular offset or feature point displacement with the preset safety threshold to determine whether the fastener is loose and generate an inspection report.

[0096] In this embodiment, the safety threshold refers to a pre-set critical value used to determine whether the fastener is loose, including an angle threshold such as 3 degrees and a displacement threshold such as 5 pixels; exceeding this threshold is considered to pose a risk of loosening.

[0097] Specifically, the angle offset Δθ and displacement difference Δd calculated in step S40 are compared with preset safety thresholds. If Δθ is greater than or equal to 3 degrees or Δd is greater than or equal to 5 pixels, the fastener is determined to be in a loose state. If neither exceeds the threshold, but Δθ is between 2 and 3 degrees or Δd is between 3 and 5 pixels, it is marked as a potential loosening trend and included in the continuous tracking and monitoring queue. The system synchronously retrieves the detection results of the most recent 5 consecutive frames of images for temporal consistency verification. Only when at least 3 frames meet the loosening judgment conditions is the loosening state finally confirmed to avoid misjudgment caused by instantaneous vibration or image jitter. After confirming loosening, the system automatically associates the fastener's equipment number, installation location, historical maintenance records, and current mileage, generates a structured maintenance report, and triggers the corresponding level of warning signal according to preset rules.

[0098] In one embodiment, in step S30, the region of interest is input into a pre-trained keypoint detection network. The keypoint detection network adopts an improved YOLOv5 network or a CenterNet network, and its loss function adds a regularization term for keypoint offset to improve the detection accuracy of small displacements of the anti-loosening marker.

[0099] In this embodiment, the keypoint detection network refers to a deep learning-based target detection and localization model used to identify and output the pixel coordinates of key locations such as the center of the bolt head, the corner of the nut, and the endpoint of the anti-loosening mark line from the input image. The improved YOLOv5 network adds a keypoint prediction branch to the original YOLOv5, enabling it to simultaneously perform target detection and keypoint regression. The improved CenterNet network uses the target center point as the keypoint for detection, naturally adapting to keypoint localization tasks. The loss function is a mathematical function used during model training to measure the difference between the predicted and true values, and the training objective is to minimize this function value. The regularization term is an additional penalty term added to the loss function to constrain model parameters and prevent overfitting. The keypoint offset regularization term in this application specifically penalizes the keypoint coordinate prediction error, and its form is λ × Σ(Δp)², where Δp is the Euclidean distance offset between the predicted key point coordinates and the manually labeled real coordinates, and λ is the regularization coefficient; micro displacement refers to the pixel-level positional change of the endpoint of the anti-loosening mark line, which usually corresponds to an angular offset of 1-3 degrees or a sub-millimeter-level translation in the actual physical space. This micro change is an important indicator of early loosening faults.

[0100] Specifically, an improved CenterNet network is used as the keypoint detection network. During the training phase, the region of interest image input to the network is 384×384 pixels in size, and the annotation information includes three types of keypoints: the center point of the bolt head (1), the vertices of the nut polygon (6, arranged clockwise), and the endpoints of the anti-loosening marking line (2, corresponding to the start and end points of the marking line respectively). The improved loss function L_total consists of a weighted sum of three parts: L_total = L_center + L_size + λ × L_offset, where L_center is the center point classification loss using the focal loss function, L_size is the target size regression loss using the L1 loss function, and L_offset is the keypoint offset regularization loss using the smoothing L1 loss function; the regularization coefficient λ is set to 0.001. In each training iteration, after the network outputs the predicted keypoint coordinates, L_offset... The algorithm calculates the sum of squares of the offsets between each predicted coordinate and its corresponding true coordinate and adds them to the total loss, thereby guiding the network weights to update in the direction of reducing key point localization errors during backpropagation. The training dataset contains over 100,000 labeled samples, covering images of undercarriage bolts under different vehicle models, lighting conditions, oil stain levels, and wear states. After 200 epochs of training, the learning rate was initially set to 0.001, and the Adam optimizer was used. After training, the network achieved an average localization error of 0.3 pixels for the endpoints of the anti-loosening markers, which is better than 1.2 pixels without regularization. The detection sensitivity for small angular offsets within 3 degrees increased from 62% to 89%, and the recognition accuracy for small displacements within 2 pixels increased from 58% to 85%. During the inference phase, the processing time for a single region of interest image is approximately 18-22 milliseconds, meeting the real-time detection requirements when high-speed trains pass at low speeds.

[0101] In one embodiment, such as Figure 5 As shown, in step S40, the reference determination module specifically includes:

[0102] S4110: Using the center point of the bolt head output by the key point detection network as the origin of the coordinate system, determine whether the bolt has chamfered or cotter pin features.

[0103] In this embodiment, the origin of the coordinate system refers to the center point of the bolt head as the origin of the geometric analysis coordinate system; the chamfer refers to the bevel formed by cutting the edge of the bolt head, which appears as a bright ring around the bolt head in the image; the cotter pin refers to the metal pin that passes through the tail of the bolt or the side of the nut to prevent the nut from loosening, which appears as a thin metal wire passing through the bolt in the image.

[0104] Specifically, the system establishes a local coordinate system with the center point of the bolt head as the origin; the 80×80 pixel area around the origin is input into the feature classification branch to identify chamfer or cotter pin features. If the probability of chamfer existence is greater than 0.7, it is determined that chamfer features exist; if the probability of cotter pin existence is greater than 0.7, it is determined that cotter pin features exist; if both are less than 0.7, it is determined that there are no absolute reference features.

[0105] S4111: If it exists, use this feature to determine the absolute vertical reference or the absolute horizontal reference.

[0106] In this embodiment, the absolute vertical reference refers to a reference line that is consistent with the direction of gravity or the axial direction of the bolt, and has a fixed directional attribute that is not affected by the individual installation angle differences of the bolt; the absolute horizontal reference refers to a reference line that is perpendicular to the absolute vertical reference.

[0107] Specifically, if a chamfer feature is detected by S4110, the system fits the chamfered circular edge, calculates the vector direction from the center of the circle to the center point of the bolt as the absolute vertical reference direction, and draws a straight line in this direction through the origin to obtain the absolute vertical reference line, the perpendicular of which is the absolute horizontal reference line; if a cotter pin feature is detected, the system extracts the direction of the cotter pin's central axis, queries the relative angle relationship between the cotter pin and the bolt from the configuration database according to the bolt model, determines the reference direction, and then draws the reference line; the system records the reference determination method and the reference direction angle value.

[0108] S4112: If it does not exist, the line connecting the corner point of the nut and the center point of the bolt shall be used as the relative reference line.

[0109] In this embodiment, the relative baseline refers to a reference line established based solely on the geometric features of the current fastener itself, without relying on absolute direction; the nut corner point refers to the coordinates of the six vertices of the hexagonal nut; the principle for selecting the relative baseline is to select the corner point closest to the bolt center point from the six nut corner points, and connect this corner point with the bolt center point to form a line.

[0110] Specifically, the system acquires the coordinates of six nut vertices, calculates the distance from each vertex to the bolt center point, selects the vertex with the smallest distance as the reference corner point, and connects this corner point with the bolt center point to form a straight line as the relative reference line. The system records the reference determination method as the relative reference line and the index number of the selected reference corner point. If insufficient nut corner points (less than 4) cannot be acquired due to occlusion, the system determines that the image quality of that frame is unqualified and marks it as needing to be re-acquired.

[0111] In one embodiment, such as Figure 5 As shown, in step S40, the vector calculation module specifically includes:

[0112] S4120: Obtain the coordinates of the two ends of the anti-loosening mark line output by the key point detection network, and calculate the coordinates of the midpoint of the anti-loosening mark line.

[0113] In this embodiment, the coordinates of the two ends of the anti-loosening mark line refer to the coordinates of the starting point and the ending point of the anti-loosening mark line; the coordinates of the midpoint refer to the geometric center point of the line connecting the two ends. Condensing the mark line into a single feature point simplifies subsequent geometric calculations.

[0114] Specifically, the coordinates of the two endpoints P1 and P2 of the anti-loosening mark line are extracted from the output of the key point detection network. First, the validity of the two endpoint coordinates is checked: check whether the coordinates are within the image boundary range, and check whether the distance between the two points is within a reasonable range (e.g., 5-200 pixels). If the distance is too short or too long, it may be a false detection by the network. In this case, the anti-loosening mark is marked as abnormal and a re-detection mechanism is triggered. After the verification is passed, the midpoint coordinates M = ((x1+x2) / 2, (y1+y2) / 2) are calculated. In the case of two anti-loosening mark lines, the midpoints M1 and M2 of the two mark lines are calculated separately. Subsequent vector calculation and deviation calculation are processed separately based on the results of the two mark lines. Finally, the average value is taken or the judgment is made according to the rules of the dual-line verification module.

[0115] S4121: Calculate the angle between the vector pointing from the origin to the midpoint and the baseline, and use it as the first loosening judgment parameter.

[0116] In this embodiment, the vector pointing from the origin to the midpoint refers to the direction line from the center point of the bolt head to the midpoint of the anti-loosening mark line; the reference line refers to the reference straight line passing through the center point of the bolt head, which is determined by the reference determination module based on the chamfer feature, cotter pin feature, or nut corner position; the included angle refers to the angle formed between the above vector and the reference line. When the fastener becomes loose, the anti-loosening mark line will rotate relative to the nut or bolt, causing the included angle to change. Therefore, the included angle is used as the first loosening judgment parameter.

[0117] Specifically, the system obtains the direction of the baseline from the baseline determination module and the vector direction from the origin to the midpoint of the anti-loosening mark from the vector calculation module. In the image coordinate system, the system takes the center point of the bolt head as the vertex and draws two rays along the baseline direction and the vector direction respectively. It calculates the minimum angle between these two rays, which ranges from 0 degrees to 180 degrees. The calculated angle value is compared with the initial angle value pre-recorded under the standard installation state of the fastener. The difference between the two is the angle offset. The system stores the angle value in the detection result of the current frame and marks the measurement timestamp and image frame number of the angle for subsequent multi-frame verification and deviation calculation. For fasteners with two anti-loosening mark lines, the system calculates the angle corresponding to each mark line separately.

[0118] S4122: Calculate the vertical offset distance of the midpoint of the anti-loosening mark line relative to the baseline, and use it as the second loosening judgment parameter.

[0119] In this embodiment, the vertical offset distance refers to the vertical distance from the midpoint of the anti-loosening mark line to the baseline, expressed in pixels. It characterizes the degree of lateral translation of the anti-loosening mark line relative to the baseline and is used to detect translational loosening.

[0120] Specifically, the system calculates the vertical distance d from point M to the baseline based on the coordinates of the midpoint M and the direction vector V_ref of the baseline. If the distance value is zero, it means that the midpoint falls on the baseline. The larger the distance, the farther the deviation. The accuracy of d can reach 0.1 pixels. For fasteners with two anti-loosening marking lines, d1 and d2 are calculated separately and the average value is taken. The system stores d in the detection result and compares it with the initial displacement value to obtain the displacement difference Δd, which complements the angular offset and is used together for loosening judgment.

[0121] In one embodiment, in step S40, as Figure 4-5 The dual-line verification module specifically includes:

[0122] S4130: For a nut containing two anti-loosening marking lines, obtain the first distance value of the first anti-loosening marking line relative to the bolt edge and the second distance value of the second anti-loosening marking line relative to the bolt edge.

[0123] In this embodiment, a nut with two anti-loosening marking lines refers to a nut with two separate painted marking lines drawn at the connection between the nut and the bolt during fastener installation to prevent loosening; the first distance value refers to the shortest distance from the midpoint of the first anti-loosening marking line to the edge of the bolt; the second distance value refers to the shortest distance from the midpoint of the second anti-loosening marking line to the edge of the bolt; the bolt edge refers to the outer contour boundary of the bolt head, which is presented as a circular or hexagonal edge line in the image, serving as a fixed reference for distance measurement.

[0124] Specifically, the system first identifies two anti-loosening marking lines on the same nut using a key point detection network, labeling them as marking line A and marking line B respectively. For each marking line, the system extracts the midpoint of the marking line as a feature point. Simultaneously, the system extracts the edge contour of the bolt head using an edge detection algorithm and fits the circular or hexagonal boundary line of the bolt edge. The system calculates the shortest distance L1 from the midpoint of marking line A to the bolt edge and the shortest distance L2 from the midpoint of marking line B to the bolt edge, with L1 and L2 in pixels. The system stores L1 and L2 in the detection data record of the fastener and notes the corresponding marking line number, measurement time, and image frame number.

[0125] S4131: Calculate the change in the difference between the first distance value and the second distance value, as well as the rate of change of the difference.

[0126] In this embodiment, the difference change refers to the absolute value of the difference between the first distance value L1 and the second distance value L2, and the calculation formula is ΔL = |L1 - L2|; the difference change rate refers to how fast the difference changes over time, that is, the rate of change of the difference relative to the number of frames in a series of consecutive frames, and the calculation formula is v = d(ΔL) / dt.

[0127] Specifically, the system calculates the difference ΔL_current = |L1 - L2| for the current frame, reads the difference ΔL_previous from the detection history, and calculates the change in difference = |ΔL_current - ΔL_previous|. Simultaneously, it caches the difference data for five consecutive frames, fits the change curve using the least squares method, and calculates the curve slope as the rate of change in difference. If the number of data points is insufficient, the rate of change is not calculated temporarily and is marked as pending accumulation. The system stores the calculation results in a temporary cache for further comparison. For fasteners detected for the first time, both the change in difference and the rate of change are initialized to zero.

[0128] S4132: When the difference change exceeds the preset first distance threshold or the difference change rate exceeds the preset second distance threshold, a double-line verification anomaly mark is generated to assist in determining the loose state of the fastener.

[0129] In this embodiment, the first distance threshold is a pre-set maximum allowable value for the difference change, which is usually set according to the bolt specifications and the accuracy of the anti-loosening mark line drawing, for example, set to 3 pixels or 0.5 mm; the second distance threshold is a pre-set maximum allowable value for the difference change rate, used to capture abnormal situations of rapid difference changes, for example, set to 1 pixel per frame; the dual-line verification anomaly flag is a Boolean flag, which is set to true when the difference change or change rate exceeds the limit, indicating that the relative position of the two anti-loosening mark lines has changed abnormally, which helps to support the conclusion that the fastener is loose.

[0130] Specifically, the system compares the difference change with a first distance threshold (3 pixels). If it exceeds this threshold, the dual-line verification anomaly flag is immediately set to true, and the anomaly type is recorded as difference change exceeding the limit. If it does not exceed the limit, the difference change rate is further compared with a second distance threshold (1 pixel per frame). If it exceeds this threshold, the dual-line verification anomaly flag is also set to true, and the anomaly type is recorded as difference change rate exceeding the limit. If neither exceeds the limit, the dual-line verification anomaly flag remains false. In the final loosening determination, if the angle offset or displacement offset is determined to be suspected loosening and the dual-line verification anomaly flag is true, the system will increase the confidence level of the determination by one level. If the angle offset and displacement offset do not exceed the limit but the dual-line verification anomaly flag is true, the system can still trigger a level one warning to prompt maintenance personnel to pay attention.

[0131] In one embodiment, such as Figure 5 As shown, in step S40, the deviation calculation module specifically includes:

[0132] S4140: Obtain the initial angle or initial displacement value of the fastener in a standard installation state.

[0133] In this embodiment, the standard installation state refers to the state in which the fastener is correctly installed according to the design specifications, is fully tightened, and has not loosened in any way. This state is usually determined when the train set leaves the factory or during the first inspection after a major overhaul. The initial angle value refers to the angle between the line connecting the midpoint of the anti-loosening mark and the center point of the bolt and the reference line relative to the baseline when the fastener is in the standard installation state. The initial displacement value refers to the vertical offset distance from the midpoint of the anti-loosening mark to the baseline when the fastener is in the standard installation state. Both of these serve as reference benchmarks for each subsequent inspection, are stored in the database, and are bound to the unique identifier of the fastener.

[0134] Specifically, the system queries historical inspection records in the database based on the unique identifier of the current fastener (train number, car number, bogie number, and position coordinates). If a record containing a standard installation status marker is found, the initial angle value θ0 and the initial displacement value d0 are read. If the fastener is being inspected for the first time and there is no standard installation status record in the database, the system uses the included angle θ calculated in the current frame as the initial angle value θ0 and the vertical offset distance d calculated in the current frame as the initial displacement value d0, and creates an initial reference record in the database with a status of pending confirmation. This record will be officially activated after manual review and approval.

[0135] S4141: Calculate the angle difference between the currently calculated included angle and the initial angle value, and use it as the core loosening judgment parameter.

[0136] In this embodiment, the angle difference refers to the difference between the included angle θ_current calculated at the current detection time and the initial angle θ0 in the standard installation state. The calculation formula is Δθ = θ_current - θ0, and the absolute value represents the magnitude of the angle change. The angle difference is the core basis for judging whether the fastener has become loose due to rotation, because when the nut rotates relative to the bolt, the anti-loosening mark line drawn between the two will be deflected by the same angle, causing θ_current to deviate from θ0. The larger the angle difference, the more serious the rotational loosening.

[0137] Specifically, the system obtains the included angle θ_current calculated in the current frame from step S4121, and obtains the initial angle value θ0 of the fastener from step S4140; the system calculates the difference between the two, Δθ = |θ_current - θ0|, accurate to 0.1 degrees; for fasteners with two anti-loosening marking lines, the system calculates the angle differences Δθ1 and Δθ2 corresponding to the two marking lines respectively, and then takes the average of the two as the final Δθ, while recording the degree of difference between the two differences for subsequent analysis; the system stores the calculated Δθ in the detection result of the current frame and marks the calculation timestamp. This Δθ will be used as a core parameter for subsequent threshold comparison and loosening judgment; the system also appends Δθ to the detection history sequence of the fastener for consistency rate and offset direction statistics in multi-frame verification logic.

[0138] S4142: Or calculate the displacement difference between the currently calculated vertical offset distance and the initial displacement value, as an auxiliary loosening judgment parameter.

[0139] In this embodiment, the displacement difference refers to the difference between the vertical offset distance d_current calculated at the current detection time and the initial displacement value d0 under the standard installation state. The calculation formula is Δd = |d_current - d0|, usually in pixels. The displacement difference is used as an auxiliary loosening judgment parameter, mainly to capture the translational movement that may accompany the loosening of fasteners. For example, when the nut moves slightly in the axial direction of the bolt, or when the anti-loosening mark line is displaced laterally due to vibration, Δd will increase significantly.

[0140] Specifically, the system obtains the vertical offset distance d_current calculated in the current frame from the vector calculation module, obtains the initial displacement value d0 of the fastener from S4140, calculates the absolute value of the difference between the two Δd = |d_current - d0|, and retains one decimal place. For fasteners with two anti-loosening marking lines, Δd1 and Δd2 are calculated separately and the average value is taken. The system stores Δd in the detection result, which is not used as an independent basis for loosening judgment, but triggers a low-level attention reminder or increases the attention level in the comprehensive judgment when Δθ does not exceed the standard but Δd exceeds the standard.

[0141] S4143: When the absolute value of the angle difference exceeds the preset angle threshold, the fastener is determined to be in a suspected loose state.

[0142] In this embodiment, the preset angle threshold refers to a pre-set critical angle difference used to distinguish between the normal state and the suspected loose state of fasteners. It is usually set according to the EMU operation and maintenance specifications and the accuracy requirements of the anti-loosening mark detection, for example, it is set to 3 degrees. The suspected loose state refers to the state in which the fastener is judged to be loose based on the current detection data, but has not yet been verified by multiple frames or manually confirmed. It is a preliminary judgment result of looseness. This result will trigger the subsequent multi-frame verification process: if the consistency rate of multi-frame verification meets the standard and the offset direction is consistent, it will be upgraded to the final confirmed loose state and an alarm will be generated; if the verification does not meet the standard, it will be judged as instantaneous interference and the alarm will be ignored.

[0143] Specifically, the system obtains the angle difference Δθ calculated by S4141 and compares it with a preset angle threshold T_angle (e.g., 3 degrees). If Δθ > T_angle, the system marks the fastener as potentially loose in the current frame, the suspected looseness flag is set to true, and the specific parameter values ​​triggering the judgment and the judgment timestamp are recorded. If Δθ ≤ T_angle, the suspected looseness flag is set to false. For fasteners in a suspected loose state, the system automatically sends the detection data of that frame to the multi-frame verification module for further analysis. In addition, when Δθ does not exceed the threshold but Δd exceeds the preset displacement threshold (e.g., 5 pixels) or the double-line verification anomaly flag is true, the system can also selectively trigger a low-level attention alert, but will not determine it as a suspected loose state.

[0144] In one embodiment, step S50, namely determining whether the fastener is loose, includes single-frame determination logic: calculating the angle offset or feature point displacement of the current single-frame image and comparing it with a preset safety threshold, and outputting a single-frame suspected looseness determination result.

[0145] In this embodiment, the single-frame judgment logic refers to a method for quickly and initially judging whether a fastener has become loose based solely on the currently acquired frame of image data. This logic does not rely on historical detection data or multi-frame image sequences, has a fast response speed and low computational load, and is suitable for real-time processing scenarios when a high-speed train passes through the detection area at low speed. A single-frame image refers to a complete image captured by the image acquisition device at a certain moment. The angle offset refers to the difference between the included angle of the anti-loosening mark line calculated in the current frame and the initial included angle under standard installation conditions, i.e., Δθ output by the deviation calculation module. The feature point displacement refers to the difference between the vertical offset distance of the midpoint of the anti-loosening mark line calculated in the current frame and the initial displacement value under standard installation conditions, i.e., Δd output by the deviation calculation module. The preset safety threshold is a pre-set critical value used to distinguish between normal and abnormal states, mainly including two types: angle threshold and displacement threshold. The single-frame suspected loosening judgment result refers to the preliminary conclusion output after the current frame judgment is completed, including two states: normal and suspected loosening. This result will be sent to the multi-frame verification logic for further verification.

[0146] Specifically, in the single-frame judgment logic, the system first obtains the core loosening judgment parameter Δθ (angle offset) and auxiliary loosening judgment parameter Δd (displacement offset) calculated for the current frame from the deviation calculation module. Simultaneously, it obtains the status of the dual-line verification anomaly flag from the dual-line verification module. The system compares Δθ with a preset angle safety threshold. The default value of the angle safety threshold is 3 degrees, which can be configured in the system according to different fastener types and importance, with an adjustment range of 1 to 10 degrees. If Δθ exceeds the angle safety threshold, the system immediately marks the current frame as a suspected loosening frame and records the trigger parameter as angle offset exceeding the limit and the specific Δθ value. If Δθ does not exceed the angle safety threshold, the system further judges the auxiliary parameters: when Δθ... If d exceeds the preset displacement safety threshold (default value is 5 pixels, approximately 0.5-1.5 mm) or the double-line verification anomaly flag is true, the system will also mark the current frame as a single frame suspected of looseness, but with a low warning level, and record the trigger parameters as displacement offset exceeding the limit or double-line verification anomaly; if Δθ does not exceed the angle safety threshold, and Δd does not exceed the displacement safety threshold, and the double-line verification anomaly flag is false, the system will determine the current frame as normal; finally, the system outputs the single-frame judgment result of the current frame, including the judgment conclusion (normal or suspected looseness), judgment basis (specific parameters and thresholds that trigger the judgment), detection timestamp, image frame number, fastener unique identifier, etc., and this judgment result is temporarily stored in the cache for multi-frame verification logic to read and process. The advantage of single-frame judgment logic is its fast processing speed, with a single-frame processing time controlled within 50 milliseconds, which can meet the real-time detection requirements when the EMU passes at a speed of 5-15 km / h. However, single-frame judgment may be affected by instantaneous interference such as vibration under the vehicle, oil stain reflection, and dust obstruction, which may lead to misjudgment. Therefore, the single-frame judgment result is not used as the final maintenance basis, but as input data for multi-frame verification.

[0147] In one embodiment, such as Figure 6 As shown, in step S50, the determination of whether the fastener is loose also includes multi-frame verification logic:

[0148] S5110: Obtain multiple angular offsets or feature point displacements calculated for the same fastener in multiple consecutive frames of images.

[0149] In this embodiment, consecutive multi-frame images refer to multiple images continuously captured by the image acquisition device at different time points and shooting positions when the EMU passes through the detection area at low speed, typically 5 to 10 frames; the same fastener refers to the same bolt or nut matched and identified in consecutive multi-frame images through the fastener's unique identifier (including train number, carriage number, bogie number, and position coordinates); angular offset refers to the Δθ value calculated in each frame image, i.e., the difference between the included angle of the current frame and the initial angle; feature point displacement refers to the Δd value calculated in each frame image, i.e., the difference between the vertical offset distance of the current frame and the initial displacement value; the purpose of obtaining multiple offsets is to accumulate sufficient data samples to provide a basis for subsequent statistical analysis, and to eliminate possible random errors or instantaneous interference in a single frame through mutual verification of multi-frame data.

[0150] Specifically, the system first establishes a temporary buffer queue to store the detection results of the same fastener across multiple consecutive frames. After processing each frame and obtaining its Δθ and Δd values, the system stores the detection results into the corresponding buffer queue based on the fastener's unique identifier. The buffer queue has a preset capacity of 10 frames and operates on a first-in, first-out (FIFO) principle, automatically discarding the oldest frame when the queue is full. The system continues acquiring data until the fastener passes through the detection area, typically collecting between 8 and 15 frames. For each frame, the system records the following information: angle offset Δθ, displacement offset Δd, offset direction (clockwise or counterclockwise, determined by the sign of Δθ), double-line check anomaly flag, and detection timestamp. If a frame cannot obtain valid Δθ and Δd due to poor image quality or keypoint recognition failure, the system marks it as an invalid frame, excluding it from subsequent statistics, but retains the invalid frame record for quality traceability. Once the cumulative number of valid frames reaches 5, the system triggers the S5111 statistical analysis process.

[0151] S5111: Perform statistical analysis on the multiple angular offsets or feature point displacements to calculate the consistency rate of loosening judgment.

[0152] In this embodiment, statistical analysis refers to the process of statistically processing parameters such as Δθ value and offset direction of multi-frame data in the buffer queue and calculating various statistical indicators; the loosening judgment consistency rate refers to the proportion of frames judged as suspected loosening in a series of consecutive images to the total number of valid frames, and the calculation formula is: consistency rate = number of suspected loosening frames ÷ total number of valid frames × 100%; where, suspected loosening frames refer to frames in a single frame judgment where Δθ exceeds the angle safety threshold or Δd exceeds the displacement safety threshold or the double-line verification is abnormal; the consistency rate reflects the stability and consistency of multi-frame detection results and is a key indicator for judging real loosening and instantaneous interference.

[0153] Specifically, the system reads all valid frame data for the fastener from the cache queue. First, it counts the total number of valid frames N, for example, N=10. Then, it iterates through each frame and, based on the comparison between Δθ and the angle safety threshold (3 degrees), combined with the comparison between Δd and the displacement safety threshold and the double-line verification anomaly flag, determines whether the frame is a suspected loose frame. The number of suspected loose frames M is counted, and the looseness determination consistency rate R = M ÷ N × 100% is calculated. At the same time, the system counts the number of frames C with clockwise offset and the number of frames D with counterclockwise offset among all suspected loose frames. If C is much greater than D, the majority offset direction is determined to be clockwise; if D is much greater than C, the majority offset direction is determined to be counterclockwise; if C and D are similar, the offset directions are determined to be inconsistent. The statistical results are summarized into a multi-frame verification report for use in subsequent determination steps.

[0154] S5112: If the consistency rate exceeds the preset consistency rate threshold and the offset direction of each frame is consistent, then it is finally determined to be "loose".

[0155] In this embodiment, the consistency rate threshold is a pre-set critical consistency value used to distinguish between genuine loosening and transient interference, with a default setting of 80%. Consistent offset direction means that the offset direction of all suspected loosening frames is either clockwise or counterclockwise, which is an important characteristic of genuine loosening—when the nut actually rotates loose, its rotation direction is unique; while transient interference (such as vibration or dust obstruction) may cause the anti-loosening mark to shift in random directions in different frames. Only when the consistency rate exceeds the threshold and the offset direction is consistent will the system finally confirm a loose state. This dual-condition design effectively reduces the false alarm rate.

[0156] Specifically, the system obtains the loosening judgment consistency rate R and the offset direction statistics results calculated by S5111; it compares R with the preset consistency rate threshold T_consist (default 80%). If R ≥ 80%, the first condition is met. At the same time, it checks the consistency of the offset direction: it counts the offset directions of all suspected loose frames. If the proportion of a certain direction (clockwise or counterclockwise) reaches more than 80%, it is considered that the offset direction is consistent; if the proportions of two directions are close (e.g., each accounts for 50%), it is considered that the offset direction is inconsistent; when R ≥ 80% and the offset direction is consistent, the system sets the final loosening judgment result to "confirmed loosening"; at this time, the system determines the graded warning level based on the average value of Δθ of all valid frames: an average Δθ between 3 and 5 degrees is a level 1 warning, between 5 and 10 degrees is a level 2 warning, and 10 degrees and above is a level 3 warning; the system generates the final maintenance warning information, including fastener position, degree of looseness, warning level, multi-frame verification statistics, etc., and prepares to push it to the maintenance terminal.

[0157] S5113: If the consistency rate is lower than the preset consistency rate threshold, it is determined to be "instantaneous interference" and the alarm is ignored.

[0158] In this embodiment, transient interference refers to accidental factors in the environment under the train that cause anomalies in a single frame image. Common transient interferences include: camera shake caused by the vibration of the train body when it passes, transient occlusion of the lens by oil or dust, occasional reflections on metal surfaces, and transient occlusion of pipes or cables. A consistency rate below the threshold means that only a few frames in a series of consecutive frames are judged as suspected looseness, while most frames are judged as normal. This indicates that the anomaly is not repetitive and is very likely caused by transient interference. Ignoring alarms means that the system does not generate maintenance warnings or push looseness alarm information to the maintenance terminal to avoid wasting maintenance resources and affecting operation and maintenance efficiency due to false alarms.

[0159] Specifically, the system obtains the loosening judgment consistency rate R calculated by S5111 and compares R with the preset consistency rate threshold T_consist (80%). If R < 80%, the system determines that the current detection result belongs to transient interference and does not confirm it as real loosening. At this time, the system does not generate maintenance warning and does not push alarm information to the maintenance terminal. However, in order to facilitate subsequent analysis and system optimization, the system will record this transient interference event in the detection log, including: fastener identification, detection time, Δθ and Δd values ​​of each frame, consistency rate R, and the inferred cause of interference (such as judging whether it is an image quality problem based on the proportion of invalid frames, or judging whether it is vibration interference based on the randomness of the offset direction). If the same fastener repeatedly shows a consistency rate close to the threshold but not meeting the standard in multiple consecutive warehouse inspections (for example, the consistency rate of three consecutive inspections is between 70% and 80%), the system can trigger a manual review suggestion to prompt maintenance personnel to pay special attention to the fastener in actual maintenance.

[0160] In one embodiment, step S50, i.e., determining whether the fastener is loose, further includes tiered warning logic:

[0161] S5120: When the absolute value of the angle offset is within the first angle threshold range, it is determined as a level one warning and a reminder is generated.

[0162] In this embodiment, the absolute value of the angle offset refers to the absolute value of Δθ obtained by the deviation calculation module, in degrees, representing the rotation angle of the anti-loosening mark line relative to its initial position; the first angle threshold range refers to the preset minimum loose angle range, indicating that the fastener has a slight angle offset, but has not yet reached the level requiring immediate repair. Typically, the first angle threshold range is set to be greater than or equal to 3 degrees and less than 5 degrees; Level 1 warning is the lowest level of warning, indicating that a slight loosening has been detected, but the degree of loosening has not yet posed a direct threat to driving safety; Attention reminder refers to the system-generated prompt information, informing maintenance personnel that the fastener has a slight loosening and suggesting that they pay attention to it during routine inspections, but not requiring immediate handling. This level of warning is mainly used for the early detection of hidden dangers and trend tracking.

[0163] Specifically, the system obtains the final angular offset Δθ of the current fastener after multi-frame verification from the deviation calculation module (usually the average Δθ of multiple valid frames). The system compares Δθ with a preset angle threshold range. The default range of the first angle threshold range is 3 degrees ≤ Δθ < 5 degrees. When Δθ falls into this range, the system determines it as a level one warning. The system generates a notification, which includes: fastener location information (train number, carriage number, bogie number, component name and coordinates), current loosening angle value (accurate to 0.1 degrees), level one warning indicator, detection timestamp and corresponding captured image. The alert is pushed to the maintenance personnel's terminal device via wireless network and displayed as a "needs attention" status in the maintenance task list, but without triggering an audible alarm or forced pop-up. During routine maintenance, maintenance personnel can use this alert to perform additional visual inspections or retests on the fastener. Simultaneously, the system stores this Level 1 alert record in the fastener's health file for subsequent loosening trend analysis. If the same fastener is repeatedly detected within the Level 1 alert range but does not continue to deteriorate, the system can automatically extend the inspection cycle for the fastener or maintain its "attention" status. If Δθ increases in subsequent inspections and enters a higher alert range, the alert level is upgraded.

[0164] S5121: When the absolute value of the angle offset is within the second angle threshold range, it is determined as a level two warning and a planned maintenance suggestion is generated.

[0165] In this embodiment, the second angle threshold range refers to the intermediate angle range between slight loosening and severe loosening, indicating that the fastener has a moderate degree of loosening, posing a certain safety hazard, and should be addressed in planned maintenance. It is usually set to be greater than or equal to 5 degrees and less than 10 degrees. The second-level warning is an intermediate warning level, indicating that the loosening has reached the point where attention is needed and it should be included in the maintenance plan, but has not yet reached an emergency state that endangers driving safety. The planned maintenance suggestion refers to the specific maintenance guidance information generated by the system, informing maintenance personnel that the fastener needs to be tightened and suggesting that it be prioritized for inspection and repair in the most recent planned maintenance (such as return to the warehouse for maintenance or regular maintenance). The purpose is to incorporate the loosening hazard into the regular operation and maintenance process and prevent it from deteriorating further.

[0166] Specifically, the system obtains the confirmed final angle offset Δθ and compares it with the second angle threshold range. The default range of the second angle threshold range is 5 degrees ≤ Δθ < 10 degrees. When Δθ falls into this range, the system determines it as a level two warning. The system generates a planned maintenance suggestion, which includes: fastener location information (train number, carriage number, bogie number, component name and three-dimensional coordinates), current loosening angle value (accurate to 0.1 degrees), level two warning mark, detection timestamp, captured image, and specific handling suggestions (such as "Please use a torque wrench to retighten the bolt according to the standard torque value and redraw the anti-loosening mark"). The planned maintenance suggestion is pushed to the maintenance scheduling system and maintenance personnel terminal equipment via wireless network, marked as a planned maintenance in the maintenance task list, and automatically added to the work list of the most recent planned maintenance. Maintenance dispatchers can arrange maintenance work orders and assign maintenance personnel based on this suggestion. At the same time, the system will store the level 2 warning record in the health file of the fastener and highlight it on the dashboard of the EMU operation and maintenance management system. Unlike the level 1 warning, the level 2 warning will trigger the system's sound prompt and task forced confirmation mechanism to ensure that maintenance dispatchers are aware of it and arrange for its handling. If the same fastener is detected in the level 2 warning range multiple times in a row and Δθ shows an increasing trend, the system will automatically upgrade it to the level 3 warning.

[0167] S5122: When the absolute value of the angle offset is within the third angle threshold range, it is determined as a level three warning and an immediate shutdown and maintenance alarm is generated.

[0168] In this embodiment, the third angle threshold range refers to the maximum loose angle range, indicating that the fasteners have become severely loose, posing a direct threat to train safety and requiring immediate attention. Typically, the third angle threshold range is set to be greater than or equal to 10 degrees. Level 3 warning is the highest warning level, indicating that the loosening has reached a dangerous state. Failure to address this promptly may result in suspension components detaching, equipment failure, or even train safety accidents. The immediate shutdown and maintenance alarm refers to an emergency alarm message generated by the system, requiring the train to immediately stop operation or stop leaving the depot. Operation can only resume after maintenance is completed and safety is confirmed.

[0169] Specifically, the system obtains the confirmed final angle offset Δθ and compares it with the third angle threshold range. The default range of the third angle threshold range is Δθ ≥ 10 degrees. When Δθ falls into this range, the system determines it as a level three warning. The system generates an immediate shutdown and maintenance alarm, which includes: fastener location information (train number, carriage number, bogie number, component name, and 3D coordinates), the current loose angle value (accurate to 0.1 degrees), the level three warning indicator, the detection timestamp, the captured image, and clear processing instructions. The level three alarm is simultaneously pushed through multiple channels: it is displayed in bright red on the large screen of the maintenance dispatch center and triggers an audible and visual alarm. After receiving the level three alarm, maintenance personnel must immediately go to the site to tighten the fasteners, replace them if necessary, and re-capture images for retesting. Only after confirming that Δθ has fallen back to a safe range (usually less than 3 degrees) can the level three warning be lifted and the train resume normal operation. The system records the level three warning and processing results in detail and includes them in the safety event database for subsequent safety analysis and accountability tracing.

[0170] In one embodiment, step S50, i.e., the generated maintenance report, includes:

[0171] S5130: Location information of loose fasteners, quantitative value of looseness, captured images, multi-frame verification results, and graded warning levels.

[0172] In this embodiment, the maintenance report refers to the comprehensive inspection result document generated after the system completes the automatic identification of the loosening state; the location information includes the train group number, carriage number, bogie number, component name, and three-dimensional coordinates of the undercarriage; the loosening degree quantification value refers to the specific value of the angle offset Δθ, accurate to 0.1 degrees; the captured image refers to the original high-definition image of the fastener when it becomes loose, with key point positions, baselines, offset angles, and other information superimposed on the image; the multi-frame verification result includes the total number of valid frames, the number of suspected loosening frames, the consistency rate, the consistency of the offset direction, and the final judgment conclusion; the graded warning level refers to Level 1 attention reminder, Level 2 planned maintenance suggestion, or Level 3 immediate shutdown maintenance alarm.

[0173] Specifically, after completing single-frame judgment, multi-frame verification, and graded early warning, the system automatically generates a maintenance report. The report includes six parts: The first part is the fastener identification information, including train group number, carriage number, bogie number, component name, fastener number, and three-dimensional coordinates of the undercarriage; the second part is the quantitative value of the loosening degree, including the current angular offset Δθ, the initial angle value θ0, the current vertical offset distance d, and the displacement difference Δd, while also providing the change compared to the last inspection; the third part is the captured image and annotations. The system selects the image with the best quality and the most obvious offset characteristics as the attached image to the report, and overlays the bolt head center point (red dot) and anti-loosening marking line (blue line segment) on the image. The system includes six parts: 1) Marking the midpoint of the line (green dot), the baseline (yellow dashed line), and the offset angle; 2) Multi-frame verification results, including the total number of valid frames, the number of suspected loose frames, the consistency rate of looseness determination, offset direction statistics, the consistency determination conclusion of offset direction, and the final confirmed determination conclusion; 3) Graded warning levels and response suggestions, including the warning level, warning level icon, the threshold range of the angle offset, the suggested handling method, and the handling time limit requirements; 4) Detection metadata, including detection time, detection equipment number, detection location communication channel number, algorithm version number, etc.; The system stores the generated maintenance report in the maintenance database and generates a unique report number for easy subsequent query and traceability.

[0174] S5131: The inspection report is pushed to the maintenance terminal via wireless network and the specific location of the loose fastener is highlighted on the map of the maintenance terminal.

[0175] In this embodiment, the wireless network refers to the high-speed communication network deployed within the EMU maintenance depot, typically employing 5G or Wi-Fi technology; the maintenance terminal refers to the smart device carried by maintenance personnel, such as a tablet, smartphone, or dedicated industrial handheld terminal, with a maintenance operation management system APP installed on the device; the map refers to an electronic schematic diagram of the EMU's undercarriage structure, usually a two-dimensional plan view or a three-dimensional model, marking the layout and coordinate grid of each component under the carriage; highlighting refers to marking the location of loose fasteners on the electronic map with eye-catching colors and icons, accompanied by flashing animations or halo effects, enabling maintenance personnel to intuitively and quickly locate the fault point.

[0176] Specifically, the system executes the following push and display process: First, the system encapsulates the maintenance report into a data packet, encrypts it using the TLS encryption protocol, and the data packet also contains the coordinate mapping information of the fastener on the electronic map; then, the system pushes the encrypted maintenance report data packet to all online maintenance terminals through the 5G base station or Wi-Fi access point in the maintenance warehouse. The push adopts a publish-subscribe model. Level 3 warning reports are pushed to all on-duty maintenance personnel and require mandatory confirmation of receipt; Level 2 warnings are pushed to the on-duty maintenance team leader; and Level 1 warnings are pushed to the maintenance task queue for personnel to retrieve themselves; after receiving the maintenance report, the maintenance terminal automatically decodes it and pops up a prompt in the maintenance operation APP. Level 1 warnings are a normal notification bar prompt, Level 2 warnings are a pop-up prompt with vibration, and Level 3 warnings are a full-screen red flashing with continuous vibration until manual confirmation; in the map module of the maintenance operation APP, the system maps the fastener's location according to the report. The system displays a highlighted marker on the corresponding location on the map. The marker appears as a red circular pulse light ring expanding outwards from the center of the marker, flashing for 3 seconds before stabilizing into a red dot. Level 3 warning markers are surrounded by a red border and a warning icon; Level 2 warnings are indicated by an orange dot, and Level 1 by a yellow dot. Users can quickly access a detailed report page by clicking on the marker. For Level 3 and Level 2 warnings, the map module also provides navigation assistance, calculating the optimal path to the fault location based on the maintenance personnel's current location, drawing a blue guide line, and displaying the estimated walking distance and time. After the maintenance personnel arrive at the site and complete the tightening, they click the task confirmation button on the terminal to upload the results. Upon receiving confirmation, the system updates the maintenance report status to "completed" and changes the corresponding highlighted marker on the map to a green checkmark icon. For Level 3 warnings, management personnel must also review and confirm in the system before the train can be released from the depot lock.

[0177] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0178] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. The actual structure is not limited to this. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A method for automatically identifying the loose state of suspension components under a high-speed train, characterized in that, The automatic identification method for loose suspension components under the EMU includes the following steps: By deploying an image acquisition device in the maintenance trench, high-definition visible light images of the undercarriage suspension components are captured when the train passes through the inspection area at low speed. The acquired raw images are preprocessed to segment out rectangular boxes containing fasteners and anti-loosening marks from the complex background under the vehicle, which are then used as regions of interest. The region of interest is input into a pre-trained key point detection network to identify and locate the coordinates of the bolt head center point, the nut polygon vertex, and the endpoint of the anti-loosening mark line; A geometric analysis model is constructed based on the coordinates of each key point in the positioning to calculate the angular offset or feature point displacement of the anti-loosening mark line relative to the fastener fixing reference. The geometric analysis model includes a benchmark determination module, a vector calculation module, a dual-line verification module, and a deviation calculation module. The calculated angular offset or feature point displacement is compared with a preset safety threshold to determine whether the fastener is loose and generate an inspection report.

2. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The key point detection network adopts an improved YOLOv5 network or CenterNet network, and its loss function adds a regularization term with key point offset to improve the detection accuracy of small displacements of the anti-loosening mark.

3. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The benchmark determination module specifically includes: Using the center point of the bolt head output by the key point detection network as the origin of the coordinate system, determine whether the bolt has chamfered or cotter pin features; If it exists, then use this feature to determine the absolute vertical reference or the absolute horizontal reference; If it does not exist, the line connecting the corner point of the nut and the center point of the bolt is used as the relative reference line.

4. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The vector calculation module specifically includes: Obtain the coordinates of the two ends of the anti-loosening mark line output by the key point detection network, and calculate the coordinates of the midpoint of the anti-loosening mark line; Calculate the angle between the vector pointing from the origin to the midpoint and the baseline, and use it as the first loosening judgment parameter; Calculate the vertical offset distance of the midpoint of the anti-loosening mark line relative to the baseline, and use it as the second loosening judgment parameter.

5. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The dual-line verification module specifically includes: For nuts containing two anti-loosening marking lines, the first distance value of the first anti-loosening marking line relative to the bolt edge and the second distance value of the second anti-loosening marking line relative to the bolt edge are obtained respectively. Calculate the change in the difference between the first distance value and the second distance value, as well as the rate of change of the difference. When the change in the difference exceeds the preset first distance threshold or the rate of change in the difference exceeds the preset second distance threshold, a double-line verification anomaly flag is generated to help determine the loose state of the fastener.

6. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The deviation calculation module specifically includes: Obtain the initial angle or initial displacement value of the fastener in a standard installation state; Calculate the angle difference between the currently calculated included angle and the initial angle value, and use it as a core loosening determination parameter; Alternatively, the displacement difference between the currently calculated vertical offset distance and the initial displacement value can be used as an auxiliary loosening judgment parameter. When the absolute value of the angle difference exceeds the preset angle threshold, the fastener is determined to be in a suspected loose state.

7. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The determination of whether a fastener is loose includes single-frame determination logic: calculating the angle offset or feature point displacement of the current single-frame image and comparing it with a preset safety threshold, and outputting a single-frame suspected looseness determination result.

8. The method for automatically identifying the loose state of suspension components under a high-speed train according to claim 7, characterized in that, It also includes multi-frame verification logic: Obtain multiple angular offsets or feature point displacements calculated for the same fastener in multiple consecutive frames of images; Statistical analysis is performed on the multiple angular offsets or feature point displacements to calculate the consistency rate of loosening determination; If the consistency rate exceeds the preset consistency rate threshold and the offset direction of each frame is consistent, it will be finally determined as "loose". If the consistency rate is lower than the preset consistency rate threshold, it is determined to be "instantaneous interference" and the alarm is ignored.

9. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, It also includes a tiered early warning logic: When the absolute value of the angle offset is within the first angle threshold range, it is determined as a level one warning and a reminder is generated; When the absolute value of the angle offset is within the second angle threshold range, it is determined to be a level two warning and a planned maintenance suggestion is generated. When the absolute value of the angular offset is within the third angular threshold range, it is determined to be a level three warning, and an immediate shutdown and maintenance alarm is generated.

10. The method for automatically identifying the loose state of suspension components under a high-speed train as described in claim 1, characterized in that, The generated maintenance report includes: Location information of loose fasteners, quantitative value of looseness, captured images, multi-frame verification results, and graded warning levels; The inspection report is pushed to the maintenance terminal via wireless network, and the specific location of the loose fastener is highlighted on the map of the maintenance terminal.

Images