Railway track slide plate unsoldering displacement detection method and system

Abstract

The invention relates to the technical field of image processing, in particular to a railway track slide plate unsoldering displacement detection method and system. The method comprises the following steps: synchronously acquiring multi-dimensional data of the vibration frequency, the vibration amplitude, the slide plate stress and the slide plate image of the slide plate through a plurality of sensors; grouping the multi-dimensional data through a time sequence and a track section position according to a train running direction and a load type, and generating a basic data set containing time-space relevance; acquiring a vibration signal, synchronously acquiring an image sequence, and establishing a mapping relation between the vibration signal and an image feature; and when the vibration waveform characteristic offset exceeds a preset threshold value or the area growth rate of the image deformation region exceeds a reference value, independently triggering primary early warning. According to the railway track slide plate desoldering displacement detection method and system, desoldering displacement detection and risk early warning based on multi-modal data fusion and dynamic correlation analysis are achieved, and train operation accidents caused by slide plate faults are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for detecting the desoldering and displacement of a railway track slide bed. Background Art

[0002] Railway turnout systems use slide plates to support the point rails and enable track changes. The quality of their welds directly impacts train safety. Current detection methods for slide plate desoldering or displacement suffer from the following drawbacks: They rely on manual inspections or single sensors, making it difficult to detect small cracks or early desoldering. They cannot comprehensively determine the location and direction of desoldering, resulting in delayed repair responses. Furthermore, they lack a dynamic correlation between the slide plate's force direction and sensor data, making detection results less reliable.

[0003] For example, patent publication number CN118155067A, titled "Method, Device, and Apparatus for Detecting Turnout Slide Plate Wear Based on Improved YOLOv5," describes a method that includes: obtaining an image of the turnout slide plate to be detected; inputting the image into a turnout slide plate wear detection model, and outputting an image of the worn turnout slide plate; the turnout slide plate wear detection model uses images of the worn turnout slide plate as a training dataset to train an improved YOLOv5 network model. However, this method can only detect wear of the turnout slide plate, not desoldering or displacement of the slide plate. Summary of the Invention

[0004] In response to the problem that railway track slide bed plate detection in the existing technology cannot identify the desoldering and displacement of the slide bed plate, the present invention provides a railway track slide bed plate desoldering and displacement detection method and system, which realizes desoldering and displacement detection and risk warning based on multimodal data fusion and dynamic correlation analysis, and avoids train operation accidents caused by slide bed plate failure.

[0005] To achieve the above technical objectives, the present invention provides a technical solution, which is a method for detecting the desoldering and displacement of a railway track slide bed, comprising the following steps: a slide plate data acquisition step, synchronously collecting multi-dimensional data of the slide plate vibration frequency, vibration amplitude, slide plate force and slide plate image through multiple sensors; The basic data set generation step is to group the multidimensional data according to the train running direction, load type, passing time sequence and track section location to generate a basic data set containing spatiotemporal correlation; Data matrix construction and anomaly detection steps: obtain vibration signals, synchronously collect image sequences, establish a mapping relationship between vibration signals and image features, construct a multidimensional data matrix containing the slide plate vibration frequency, vibration amplitude, slide plate force and slide plate image, perform time synchronization comparison between the current detection group data and the historical normal state data set, calculate the vibration waveform similarity based on the dynamic time warping algorithm, and detect the deformation area on the slide plate surface based on the image feature matching algorithm; The desoldering abnormality judgment step calculates the vibration waveform feature offset and the image deformation area area growth rate through intra-group comparison. When the vibration waveform feature offset exceeds the preset threshold or the image deformation area area growth rate exceeds the baseline value, the primary warning is triggered independently. When the two detection results form a spatial position correlation, it is determined to be a desoldering abnormality event.

[0006] In this technical solution, multiple sensors simultaneously collect multidimensional data, including slide plate vibration frequency, vibration amplitude, slide plate force, and slide plate image data. This data comprehensively reflects the slide plate's operating status from multiple perspectives, addressing the potential limitations of a single data source and improving the accuracy and reliability of slide plate desoldering and displacement detection. Multidimensional data is grouped by train direction, load type, transit time sequence, and track section location to generate a basic dataset with temporal and spatial correlations. This allows for a better understanding of the slide plate's operating characteristics under different operating conditions, providing a more accurate basis for subsequent detection and early warning. A mapping relationship is established between vibration signals and image features. Vibration waveform similarity is calculated using a dynamic time warping algorithm, and surface deformation areas are detected using an image feature matching algorithm. This allows for simultaneous detection of slide plate anomalies from both the vibration signal and image features, improving the comprehensiveness and accuracy of detection. Vibration waveform feature offsets and image deformation area growth rates are calculated through intra-group comparisons. Preset thresholds and baseline values are set, and when these thresholds are exceeded, primary warnings are triggered independently. This improves the accuracy and timeliness of early warnings and avoids false alarms and missed alerts. When the two test results form a spatial correlation, a desoldering anomaly is identified. This spatial correlation further improves the reliability of the test results and eliminates misjudgments caused by accidental factors. This allows for timely detection of desoldering and displacement issues on the slideway plate and issues an early warning, providing railway maintenance personnel with ample time to carry out repairs and address the issue. This effectively avoids train accidents caused by slideway plate failures and improves railway safety.

[0007] The present invention is further configured to include: a desoldering feature classification step, determining the abnormal starting point based on the vibration frequency, vibration amplitude, force of the slide bed plate and the slide bed plate image, dividing the desoldering position into regions based on the train formation information when the abnormal event occurs, identifying the desoldering features, and classifying the desoldering features.

[0008] In this technical solution, the starting point of the abnormality is determined by integrating multi-dimensional data such as the vibration frequency, vibration amplitude, force on the slide plate, and the slide plate image. Compared with a single data source, this can more comprehensively and accurately capture the starting moment of the slide plate desoldering and displacement abnormality, improving the accuracy and timeliness of abnormality detection, helping to identify potential risks at the early stage of the problem and avoid further deterioration of the abnormality. The desoldering location is divided into regions based on the train formation information at the time of the abnormal event, taking into account the impact of the train operation status and load on the slide plate, making the regional division more scientific and targeted. Under different formations and loads, the force conditions of the slide plate are different, and the area and characteristics of the desoldering may also vary. This can better reflect the actual situation and provide accurate regional positioning for subsequent repairs and maintenance.

[0009] The present invention is further configured such that: in the basic data set generation step, the acquisition through time series includes: Install visual sensors on the train. When the train passes through a tunnel for the first time, it constructs a 3D map of the tunnel using visual SLAM technology and records the location information of key feature points. When the train passes through the tunnel, the visual sensor captures real-time image information inside the tunnel. Using the visual SLAM algorithm, feature points are extracted from the real-time image and matched with feature points in the pre-built map. Based on the real-time train position determined by the visual SLAM algorithm and combined with train speed information, the time when the train wheels pass through the slide bed is predicted. With the predicted passing time as the center, the same period of time is extended forward and backward to form a fixed time window; Taking the time when the train first arrives at the slide bed detection area as the benchmark, calculate the time difference between each subsequent data point and the benchmark time. According to the train speed information, multiply the time difference by the inverse of the speed to obtain the time code value. According to the track section division, assign unique spatial identification codes to different sections.

[0010] In this technical solution, traditional positioning methods fail in areas such as tunnels where GPS signals cannot cover. By installing visual sensors and utilizing visual SLAM technology, a three-dimensional map can be constructed in the tunnel and the train position can be located in real time, effectively solving the positioning problem in the tunnel. Visual SLAM technology can provide high-precision positioning information by extracting and matching feature points in the image. Compared with traditional positioning methods, visual SLAM is not affected by factors such as electromagnetic interference and multipath effects, and the positioning results are more stable and reliable. By combining train speed information and visual SLAM positioning results, the time when the train wheels pass through the slide bed can be accurately predicted, and a fixed time window can be formed with this as the center. At the same time, based on the time when the train first arrives at the slide bed detection area, the time difference is calculated and encoded, and combined with the spatial identification code of the track section, a precise association between time and space is achieved, providing a reliable foundation for subsequent data analysis.

[0011] The present invention is further configured such that: in the basic data set generation step, grouping the multidimensional data according to train running direction, load type, passing time sequence and track section position includes: The train running direction and load type are used as the first-level labels, wherein the running direction includes up and down, and the load type includes empty and loaded; The train passing time sequence and track section are used as secondary labels, and the track section includes straight segments, curved segments and switch areas; According to the historical desoldering event statistics, weights are assigned to different tag combinations.

[0012] In this technical solution, a multi-level data grouping system is constructed by using the train's running direction and load type as the first-level labels, and the train's passing sequence and track section as the second-level labels. This allows for more detailed data division and classification of the complex railway operating environment and train status information. The selected labels cover the key factors that affect the desoldering and displacement of the slide plate. The train's running direction affects the contact force and friction between the wheelset and the slide plate; the load type directly determines the pressure the train exerts on the slide plate; the train's passing sequence reflects the train's operating status within a specific time period; and the track sections have different geometric features and stress conditions. Historical desoldering events reflect the actual situation of slide plate failures under different operating conditions. By assigning weights to different label combinations, those label combinations that are more likely to cause desoldering events can be highlighted.

[0013] The present invention is further configured such that in the data matrix construction and anomaly detection steps, establishing a mapping relationship between the vibration signal and the image feature includes: The collected vibration signal of the slide plate is processed in segments according to a certain signal duration; According to the sliding window technology, the vibration signal is processed with a certain step size to continuously extract time domain features; Perform fast Fourier transform on each vibration signal to convert the time domain signal into a frequency domain signal; The texture features of the slide bed image are extracted based on the local binary pattern algorithm; Extract key point features of the slide bed image based on the scale-invariant feature transformation algorithm; Based on the deep learning method, a convolutional neural network model is used to extract features from the image; A regression model between vibration signal features and image features is established based on the support vector regression algorithm, with the time domain features and frequency domain features of the vibration signal as input and the texture features and key point features of the image as output.

[0014] In this technical solution, the vibration signal is segmented into segments of fixed duration and processed using a sliding window technique with a fixed step size to continuously extract time-domain features. This technique captures the time-series variation of the vibration signal, such as amplitude fluctuations and periodicity, providing rich time-domain information for subsequent analysis and helping to fully understand the dynamic response of the slide plate. The time-domain signal is converted into a frequency-domain signal using a fast Fourier transform, revealing the distribution of the vibration signal across different frequency components. A local binary pattern algorithm is used to extract texture features from the slide plate image. These texture features reflect microstructural variations on the slide plate surface, such as surface roughness and scratches. These features are closely related to issues such as wear, desoldering, and displacement, providing important visual information for detection. A scale-invariant feature transform algorithm is used to extract key point features from the image. Key point features are invariant to transformations such as scaling, rotation, and translation, enabling accurate identification of important feature points in the slide plate image, such as bolt locations and edge contours. This provides a stable image feature foundation for subsequent correlation analysis with vibration signal features. A convolutional neural network model is used to extract features from images. Convolutional neural networks have powerful feature learning capabilities and can automatically learn deeper, more abstract features from images, including potential features of slideway desoldering and displacement, further enriching the representation of image features. A support vector regression algorithm is used to establish a regression model between vibration signal features and image features. This model uses the time and frequency domain features of the vibration signal as input and the texture and key point features of the image as output. This model can explore the inherent correlation between the vibration signal and image features. By continuously optimizing the model parameters, the accuracy and stability of the mapping relationship are improved, providing a more scientific basis for slideway desoldering and displacement detection.

[0015] The present invention is further configured such that in the data matrix construction and anomaly detection step, the construction of a multidimensional data matrix including the vibration frequency, vibration amplitude, force on the slide plate and the image of the slide plate comprises: For the synchronously collected vibration frequency, vibration amplitude, and slide plate force data, adaptive median filtering is used to remove impulse noise, and wavelet threshold denoising is used to eliminate high-frequency interference. Different normalized weights are assigned to the vibration frequency, vibration amplitude, and slide plate force. The data are linearly mapped to the [0, 1] interval using the maximum and minimum normalization method to form a fused vibration-force feature vector. The features of the slide bed image are extracted by a lightweight convolutional neural network, and the extracted feature vector is normalized by the Z-Score method to make the feature mean 0 and the variance 1. Taking the time when the train first arrives at the slide plate detection area as the benchmark, calculate the time difference between each subsequent data point and the benchmark time. Based on the train speed information, multiply the time difference by the inverse of the speed to obtain the time code value. Based on the track section division, unique spatial identification codes are assigned to different sections. A four-dimensional data matrix is constructed, in which the four dimensions correspond to time, space, vibration-force characteristics, and image characteristics, respectively. According to the spatiotemporal coding rules, the preprocessed vibration-force feature vectors and image feature vectors are filled into the corresponding positions of the matrix, and the missing data points are supplemented by the weighted average interpolation method based on spatiotemporal neighboring data.

[0016] In this technical solution, adaptive median filtering is used to remove impulse noise, and wavelet threshold denoising eliminates high-frequency interference, significantly improving the quality of vibration frequency, vibration amplitude, and slide plate force data. Impulse noise and high-frequency interference can affect the authenticity and reliability of data. Denoising can make the data smoother and more accurate, laying a good foundation for subsequent feature extraction and data analysis. Different normalization weights are assigned to the vibration frequency, vibration amplitude, and slide plate force, and the data are linearly mapped to the [0, 1] interval using the maximum-minimum normalization method, eliminating the influence of different dimensions and orders of magnitude on data analysis. The denoised and normalized vibration frequency, vibration amplitude, and slide plate force data are fused into a vibration-force feature vector, effectively integrating different types of features. Vibration and force information reflects the operating state of the slide plate from different perspectives. The fused feature vector provides a more comprehensive description of the slide plate's dynamic characteristics, providing richer information for subsequent fusion with image features. Slide plate image features are extracted using a lightweight convolutional neural network, and Z-score normalization is used to normalize the feature distribution, stabilizing the feature mean to 0 and the variance to 1. Using the time when the train first arrives at the slide plate inspection area as a benchmark, the time difference between each subsequent data point and the benchmark is calculated. Based on the train's speed information, the time difference is multiplied by the inverse of the speed to obtain a time code value. This considers the influence of train speed on time, making the time information more accurately reflect the train's operating status on the track and enhancing the correlation between the data and time. Based on the track segment division, each section is assigned a unique spatial identifier, clarifying the spatial location of the data. By combining the time code and spatial identifier, a data structure with spatiotemporal correlation is constructed, facilitating the analysis of slide plate state changes at different temporal and spatial locations, providing strong support for fault location and prediction. A data matrix encompassing four dimensions—temporal, spatial, vibration-stress, and image features—is constructed, integrating multi-source data. This comprehensively describes the slide plate's status, encompassing temporal, spatial, and physical characteristics. This provides richer information for slide plate desoldering and displacement detection, improving detection integrity. Missing data points are supplemented using a weighted average interpolation method based on spatiotemporal neighboring data, ensuring data continuity and integrity.

[0017] The present invention is further configured such that: in the data matrix construction and anomaly detection steps, the calculation of the vibration waveform similarity according to the dynamic time warping algorithm includes: According to the DTW algorithm, a distance matrix is constructed. The matrix elements represent the distance between two vibration waveforms at different time points. When calculating the cumulative distance matrix, the path slope is restricted to a certain range. When calculating the path cost, if the vibration amplitude change between adjacent time points exceeds a preset threshold, the cost of this path point is additionally weighted to obtain the cumulative distance reflecting the similarity of the two vibration waveforms.

[0018] In this technical solution, a distance matrix is constructed based on the dynamic time warping algorithm, with the matrix elements accurately representing the distance between two vibration waveforms at different time points. Distance calculations based on waveform decomposition at specific time points fully consider the waveform's local characteristics along the time axis, providing a detailed and accurate data foundation for subsequent similarity calculations and avoiding the potential bias associated with similarity assessments based solely on the overall waveform profile. When calculating the cumulative distance matrix, the path slope is constrained to a certain range, ensuring a more reasonable match between the two vibration waveforms along the time axis and preventing mismatches caused by excessive distortion of the time axis. When calculating the path cost, additional weighting is applied to path points where the vibration amplitude change between adjacent time points exceeds a preset threshold. When a slide plate desoldering or shifting fault occurs, the vibration waveform amplitude often exhibits abnormal changes. By applying additional weighting to key change points, the algorithm focuses on these fault-related features, effectively capturing the slide plate's fault characteristics and improving fault detection sensitivity. By processing the path slope and path cost described above, the DTW algorithm can more accurately identify patterns in the vibration waveform associated with slide plate desoldering or shifting faults.

[0019] The present invention is further configured such that: in the data matrix construction and anomaly detection steps, detecting the deformation area on the surface of the slide bed plate according to the image feature matching algorithm comprises: The key point features of the current detection image and the historical normal state image are extracted respectively according to the scale-invariant feature transformation algorithm; By calculating the Euclidean distance between the two sets of image feature descriptors, we filter out feature point pairs with a matching degree higher than a preset threshold, generate an initial matching point set, and eliminate mismatched points in the initial matching point set using a random sampling consistency algorithm. According to the coordinate information of the matching point pairs, the geometric transformation parameters between the current image and the historical normal image are calculated through the affine transformation model, and the current image is mapped to the coordinate system of the historical normal image so that the two have the same geometric reference; The grayscale value difference between the corresponding position of the current image and the historical normal image after mapping is compared pixel by pixel, and a grayscale difference threshold is set. When the proportion of pixels in a certain area whose grayscale value difference exceeds the threshold reaches a preset ratio, the area is determined to be a deformation area of the slide bed surface.

[0020] In this technical solution, a scale-invariant feature transformation algorithm is used to extract key point features of the current detection image and the historical normal state image respectively. By calculating the Euclidean distance between the two sets of image feature descriptors to screen feature point pairs with high matching, and using a random sampling consistency algorithm to exclude mismatched points, the accuracy of feature matching is further improved and detection errors caused by mismatching are reduced. After mapping the current image to the coordinate system of the historical normal image, the grayscale value differences of the corresponding positions are compared pixel by pixel. The grayscale value differences can intuitively reflect the changes in the image surface. By setting a reasonable grayscale difference threshold and a preset ratio of the number of pixels, it is possible to accurately determine whether there is a deformation area on the slide bed surface. It has high sensitivity and accuracy and can effectively detect small deformations on the slide bed surface. Based on the coordinate information of the matching point pairs, the geometric transformation parameters between the current image and the historical normal image are calculated through the affine transformation model, and the current image is mapped to the coordinate system of the historical normal image, achieving precise alignment of the two images. When comparing the grayscale value differences, they can accurately correspond to the same position, improving the accuracy of fault location. Even if the image has a certain degree of geometric deformation such as translation, rotation and scaling, the affine transformation model can correct it, ensuring the accuracy of the detection results.

[0021] Another technical solution provided by the present invention is a railway track slide bed plate desoldering and displacement detection system, comprising: Data acquisition module: used to synchronously collect multi-dimensional data of the slide bed vibration frequency, vibration amplitude, slide bed force and slide bed image through multiple sensors; Data grouping module: groups the multidimensional data according to train running direction, load type, passing time sequence and track section position to generate a basic data set containing spatiotemporal correlation; Relationship building and comparison module: This module acquires vibration signals, synchronously collects image sequences, establishes a mapping relationship between vibration signals and image features, and constructs a multidimensional data matrix containing the slide plate's vibration frequency, vibration amplitude, slide plate force, and slide plate image. It also performs a time-synchronized comparison of the current detection group data with the historical normal state data set, calculates the vibration waveform similarity using a dynamic time warping algorithm, and detects the deformation area on the slide plate surface using an image feature matching algorithm. Early warning and judgment module: Through intra-group comparison, the vibration waveform feature offset and the image deformation area growth rate are calculated. When the vibration waveform feature offset exceeds the preset threshold or the image deformation area growth rate exceeds the baseline value, a primary early warning is triggered independently. When the two detection results form a spatial correlation, it is determined to be a desoldering abnormality event; Abnormal analysis module: Determine the abnormal starting point based on the vibration frequency, vibration amplitude, force and image of the slide bed plate. Based on the train formation information when the abnormal event occurs, divide the desoldering location into regions, identify the desoldering features, and classify the desoldering features.

[0022] In this technical solution, the data acquisition module uses multiple sensors to synchronously collect multidimensional data, including the slide plate's vibration frequency, amplitude, forces acting on the slide plate, and images. This data covers various aspects of the slide plate's mechanical motion and appearance. The data grouping module groups the data by train direction, load type, transit time, and track section location, generating a basic dataset with temporal and spatial correlations, making the data more organized and analyzable. The relationship building and comparison module further establishes a mapping relationship between vibration signals and image features, constructing a multidimensional data matrix. This multi-dimensional, multi-level integration approach comprehensively reflects the slide plate's operating status, significantly improving the comprehensiveness of the inspection. The relationship building and comparison module performs a time-synchronized comparison of the current inspection data set with a historical normal state dataset. It uses a dynamic time warping algorithm to calculate vibration waveform similarity and an image feature matching algorithm to detect surface deformation areas on the slide plate. These evaluations complement each other to more accurately detect slide plate anomalies. The early warning and judgment module calculates the vibration waveform feature offset and the area growth rate of the image deformation region, and sets corresponding thresholds for early warning and judgment, which further improves the accuracy of detection and reduces the possibility of false alarms and missed alarms.

[0023] The present invention is further configured as follows: the data grouping module includes: Label setting unit: takes the train running direction and load type as the first-level label, where the running direction includes up and down, and the load type includes empty and loaded; takes the train passing sequence and track section as the second-level label, where the track section includes straight section, curved section, and switch section; Weight allocation unit: Assign weights to different tag combinations based on historical desoldering event statistics.

[0024] In this technical solution, the label setting unit groups multidimensional data using train direction, load type, transit sequence, and track section as labels. Train direction and load type serve as primary labels, reflecting the different directions and magnitudes of the train's forces on the slide plate. Under different directions and loads, the forces acting on the slide plate vary significantly, leading to different wear and deformation patterns. Track section and train transit sequence serve as secondary labels. Different track sections result in different lateral and longitudinal forces on the slide plate. Transit sequence, on the other hand, considers the temporal patterns of train operation and provides a comprehensive and detailed description of the slide plate's operating environment and usage, making data classification more scientific and reasonable. This hierarchical structure of primary and secondary labels not only highlights the primary influencing factors of direction and load type, but also considers the secondary factors of transit sequence and track section. This makes data classification clearer and more organized, facilitates subsequent analysis and processing of data with different label combinations, and improves the system's maintainability and scalability.

[0025] The beneficial effects of the present invention are as follows: (1) it realizes desoldering displacement detection and risk warning based on multimodal data fusion and dynamic correlation analysis, thereby avoiding train operation accidents caused by slide bed failure; (2) it synchronously collects multidimensional data such as slide bed vibration frequency, vibration amplitude, slide bed force and slide bed image through multiple sensors, which can comprehensively reflect the working status of the slide bed from multiple angles, make up for the limitations that may exist in a single data source, and improve the accuracy and reliability of slide bed desoldering displacement detection. It groups the multidimensional data according to the train running direction, load type, passing sequence and track section position to generate a basic data set containing spatiotemporal correlation, which can better understand the working characteristics of the slide bed under different operating conditions, provide a more accurate basis for subsequent detection and warning, establish a mapping relationship between vibration signals and image features, and calculate the vibration waveform similarity according to the dynamic time warping algorithm. , according to the image feature matching algorithm, the deformation area of the slide bed surface is detected, and the abnormal situation of the slide bed can be detected from the two levels of vibration signal and image feature at the same time, which improves the comprehensiveness and accuracy of the detection. The vibration waveform feature offset and the image deformation area area growth rate are calculated by intra-group comparison, and preset thresholds and benchmark values are set. When the corresponding thresholds are exceeded, the primary warning is triggered independently, which can improve the accuracy and timeliness of the warning and avoid false alarms and missed alarms. When the two detection results form a spatial position correlation, it is determined to be a desoldering abnormal event. The spatial position correlation judgment can further improve the reliability of the detection results, eliminate the misjudgment caused by some accidental factors, and can timely discover the desoldering displacement problem of the slide bed and issue an early warning, providing railway maintenance personnel with sufficient time for repair and processing, thereby effectively avoiding train operation accidents caused by slide bed failures and improving the safety of railway operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1Schematic diagram of the flow of the method for detecting the desoldering and displacement of the railway track slide bed plate of the present invention; Figure 2 This is a schematic diagram of the structure of a railway track slide bed desoldering and displacement detection system; Figure 3 This is a schematic structural diagram of the railway track bed cover in the present invention; Figure 4 A schematic diagram of the structure of the slide bed plate in the present invention when it is desoldered and displaced; Figure 5 It is a structural schematic diagram of a slide bed plate with cracks in the present invention. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] like Figure 1 、 Figures 3 to 5 As shown in the embodiment 1 of the present invention, a method for detecting the desoldering and displacement of a railway track slide bed plate includes the following steps: a slide plate data acquisition step, synchronously collecting multi-dimensional data of the slide plate vibration frequency, vibration amplitude, slide plate force and slide plate image through multiple sensors; The basic data set generation step is to group the multidimensional data according to the train running direction, load type, passing time sequence and track section location to generate a basic data set containing spatiotemporal correlation; Data matrix construction and anomaly detection steps: obtain vibration signals, synchronously collect image sequences, establish a mapping relationship between vibration signals and image features, construct a multidimensional data matrix containing the slide plate vibration frequency, vibration amplitude, slide plate force and slide plate image, perform time synchronization comparison between the current detection group data and the historical normal state data set, calculate the vibration waveform similarity based on the dynamic time warping algorithm, and detect the deformation area on the slide plate surface based on the image feature matching algorithm; The desoldering abnormality judgment step calculates the vibration waveform feature offset and the image deformation area area growth rate through intra-group comparison. When the vibration waveform feature offset exceeds the preset threshold or the image deformation area area growth rate exceeds the baseline value, the primary warning is triggered independently. When the two detection results form a spatial position correlation, it is determined to be a desoldering abnormality event.

[0029] In this embodiment, multiple sensors simultaneously collect multidimensional data, including slide plate vibration frequency, vibration amplitude, slide plate force, and slide plate image data, enabling comprehensive analysis of the slide plate's operating status from multiple perspectives. This overcomes the potential limitations of a single data source and improves the accuracy and reliability of slide plate desoldering and displacement detection. Multidimensional data is grouped by train direction, load type, transit time sequence, and track section location to generate a basic dataset with temporal and spatial correlations. This allows for a better understanding of the slide plate's operating characteristics under different operating conditions, providing a more accurate basis for subsequent detection and early warning. A mapping relationship is established between vibration signals and image features, and vibration waveform similarity is calculated using a dynamic time warping algorithm. Deformation areas on the slide plate's surface are detected using an image feature matching algorithm. This allows for simultaneous detection of slide plate anomalies from both the vibration signal and image features, improving the comprehensiveness and accuracy of detection. Vibration waveform feature offsets and image deformation area growth rates are calculated through intra-group comparisons. Preset thresholds and baseline values are set, and when these thresholds are exceeded, primary early warnings are triggered independently. This improves the accuracy and timeliness of early warnings and avoids false alarms and missed alerts. When the two test results form a spatial correlation, a desoldering anomaly is identified. This spatial correlation further improves the reliability of the test results and eliminates misjudgments caused by accidental factors. This allows for timely detection of desoldering and displacement issues on the slideway plate and issues an early warning, providing railway maintenance personnel with ample time to carry out repairs and address the issue. This effectively avoids train accidents caused by slideway plate failures and improves railway safety.

[0030] It can be understood that the multiple sensors are connected to the time synchronization module so that the data collected by the multiple sensors are synchronized in time.

[0031] It is understandable that sensor data of different dimensions are normalized to have the same dimension to facilitate subsequent data analysis and comparison. For example, the maximum and minimum normalization method is used to linearly map the data to the interval [0, 1].

[0032] As can be understood, since data collection times from different sensors may vary slightly, a linear interpolation algorithm is used to unify the data from these sensors onto the same timeline based on the sensor data's timestamps. The collected data is then spatially aligned with the track mileage coordinate system. The track location corresponding to each data point is determined based on the train's transit time sequence and track section location labels. For example, data collected by vibration sensors, force sensors, and image sensors are associated with the track mileage coordinate system to ensure that each data point has clear geographic location information. The collected multidimensional data is then correlated based on train direction, load type, transit time sequence, and track section location. The temporally aligned and correlated multidimensional data sets are integrated to generate a basic dataset that incorporates temporal and spatial correlations. This basic dataset includes multidimensional data on each data set's train direction, load type, transit time sequence, track section location, and the corresponding vibration frequency, vibration amplitude, slide plate force, and slide plate image.

[0033] As you can understand, vibration frequency and amplitude directly reflect the dynamic response of the slide bed under conditions such as train traffic. Multiple sensors monitor the physical changes in the slide bed as it is impacted by the train in real time, and the collected data is a concrete manifestation of the vibration signal. For example, when a train wheel passes over the slide bed, the slide bed vibrates, and the vibration sensors capture the corresponding vibration frequency and amplitude information, forming part of the dynamic response signal.

[0034] Calculating the vibration waveform characteristic offset includes: Preprocess the current detected vibration waveform data and the historical vibration waveform data under normal conditions, use a low-pass filter to remove high-frequency noise interference, and use the minimum-maximum normalization method to map the data to the [0, 1] interval to unify the data scale; The time domain and frequency domain features of the two waveform data are extracted separately: in the time domain, the waveform features such as mean, variance, and peak are calculated; in the frequency domain, the time domain signal is converted into a frequency domain signal through fast Fourier transform, and features such as main frequency, bandwidth, and spectrum energy are extracted; The extracted features are combined into a feature vector, and the Euclidean distance formula is used to calculate the distance between the current vibration waveform feature vector and the historical normal vibration waveform feature vector, which is the vibration waveform feature offset.

[0035] The calculation formula for the area growth rate of the image deformation region is: ; Among them, the growth rate of the area of the deformed area of the image is usually used to describe the degree of change of the area of the deformed area on the surface of the slide bed with time or the number of detections; if deformation already exists during the first detection, the normal image detected after the most recent maintenance is used as the benchmark; when deformation occurs multiple times, the undeformed area in the benchmark image is used as a reference to avoid cumulative errors.

[0036] In one embodiment of the present invention, it also includes: a desoldering feature classification step, determining the abnormal starting point based on the vibration frequency, vibration amplitude, force of the slide bed plate and the slide bed plate image, dividing the desoldering position into areas based on the train formation information when the abnormal event occurs, identifying the desoldering features, and classifying the desoldering features.

[0037] In this technical solution, the starting point of the abnormality is determined by integrating multi-dimensional data such as the vibration frequency, vibration amplitude, force on the slide plate, and the slide plate image. Compared with a single data source, this can more comprehensively and accurately capture the starting moment of the slide plate desoldering and displacement abnormality, improving the accuracy and timeliness of abnormality detection, helping to identify potential risks at the early stage of the problem and avoid further deterioration of the abnormality. The desoldering location is divided into regions based on the train formation information at the time of the abnormal event, taking into account the impact of the train operation status and load on the slide plate, making the regional division more scientific and targeted. Under different formations and loads, the force conditions of the slide plate are different, and the area and characteristics of the desoldering may also vary. This can better reflect the actual situation and provide accurate regional positioning for subsequent repairs and maintenance.

[0038] It can be understood that the train formation information includes the number and type of carriages of the train, which can be obtained according to the train operation plan.

[0039] Determining the abnormal starting point according to the vibration frequency, vibration amplitude, force on the slide plate and the slide plate image includes: For the vibration frequency and amplitude data of the slide bed, a sliding average filter method is used for noise reduction. The sliding window size is set, and the average value of the data within the window is calculated as the filtered data. The first-order difference of the filtered data is calculated to obtain the rate of change of the data. When the vibration frequency change rate at a certain moment is greater than the preset frequency change threshold, and the vibration amplitude change rate is greater than the preset amplitude change threshold, the moment is recorded as a suspicious point of vibration data anomaly. For the slide plate force data, we first use fast Fourier transform to convert the time domain data into frequency domain data, analyze the energy distribution in the frequency domain, and extract the energy value corresponding to the main frequency component. When the deviation between the energy value of the main frequency component and the corresponding energy value under the historical normal state exceeds the preset energy deviation threshold, we record that moment as a suspicious abnormal point in the force data. For the slide bed image, the current image and the historical normal image are grayscaled, and the image edge contours are extracted using the Canny edge detection algorithm. The Hausdorff distance between the edge contours of the current image and the historical normal image is calculated. The Hausdorff distance reflects the maximum mismatch between the two point sets. When the Hausdorff distance is greater than the preset contour difference threshold, the image acquisition moment is recorded as a suspicious image abnormality point. The abnormal suspicious points obtained from the vibration data, force data and image data are time-aligned. If there are abnormal suspicious points of at least two data types at a certain moment, and there are no abnormal suspicious points within a certain time interval before this moment, then this moment is determined as the abnormal starting point; if there is no moment that meets the conditions at the same time, then the earliest abnormal suspicious point in each data type is selected, and the earliest moment among these points is used as the abnormal starting point.

[0040] The step of dividing the desoldering position into regions, identifying desoldering features, and classifying the desoldering features includes: First, based on the abnormal starting point determined in step S4, an initial detection area is delineated in the track mileage coordinate system, centered on that point and combined with the location information of the suspected desoldering area in the slide bed image. The slide bed image is processed using an edge detection algorithm (Canny edge detection algorithm). By calculating the gradient amplitude and direction of the pixel points in the image, the edges with significant grayscale changes in the image are found, outlining the desoldering location, further refining the desoldering area boundary, and completing the region division. When identifying desoldering features, a gray-level co-occurrence matrix analysis is performed on the image of the desoldering area to calculate texture feature parameters such as contrast, correlation, energy, and entropy to describe the texture characteristics of the desoldering area. At the same time, the shape of the desoldering area is quantified by calculating geometric shape features such as the area, perimeter, and circularity of the desoldering area. In addition, combined with the vibration frequency, vibration amplitude, and slide plate force data, statistical features such as the mean, variance, and peak value of the data in the period before and after the abnormal starting point are extracted to supplement the desoldering features. Based on the extracted texture features, geometric shape features and vibration force statistical features, the support vector machine classification algorithm is used to classify the desoldering features. A large amount of sample data of known desoldering types is collected in advance, the corresponding features are extracted and labeled, and a training data set is constructed. The training data set is used to train the SVM model, and the optimal classification hyperplane is determined. The desoldering feature vectors to be classified are input into the pre-trained SVM model. According to the model output results, the desoldering features are divided into different types, such as micro-crack type desoldering, local debonding type desoldering, and large-area separation type desoldering.

[0041] In one embodiment of the present invention, in the basic data set generation step, the acquisition through time series includes: Install visual sensors on the train. When the train passes through a tunnel for the first time, it constructs a 3D map of the tunnel using visual SLAM technology and records the location information of key feature points. When the train passes through the tunnel, the visual sensor captures real-time image information inside the tunnel. Using the visual SLAM algorithm, feature points are extracted from the real-time image and matched with feature points in the pre-built map. Based on the real-time train position determined by the visual SLAM algorithm and combined with train speed information, the time when the train wheels pass through the slide bed is predicted. With the predicted passing time as the center, the same period of time is extended forward and backward to form a fixed time window; Taking the time when the train first arrives at the slide bed detection area as the benchmark, calculate the time difference between each subsequent data point and the benchmark time. According to the train speed information, multiply the time difference by the inverse of the speed to obtain the time code value. According to the track section division, assign unique spatial identification codes to different sections.

[0042] In this technical solution, traditional positioning methods fail in areas such as tunnels where GPS signals are not covered. By installing visual sensors and utilizing visual SLAM technology, a three-dimensional map can be constructed in the tunnel and the train position can be located in real time, effectively solving the positioning problem in the tunnel. Visual SLAM technology can provide high-precision positioning information by extracting and matching feature points in the image. Compared with traditional positioning methods, visual SLAM is not affected by factors such as electromagnetic interference and multipath effects, and the positioning results are more stable and reliable. By combining train speed information and visual SLAM positioning results, the time when the train wheels pass through the slide bed can be accurately predicted, and a fixed time window can be formed with this as the center. At the same time, based on the time when the train first arrives at the slide bed detection area, the time difference is calculated and encoded. Combined with the spatial identification code of the track section, a precise correlation between time and space is achieved, providing a reliable foundation for subsequent data analysis.

[0043] It can be understood that visual SLAM technology is a technology that uses visual sensors to perceive the surrounding environment in real time, build an environmental map, and determine its own position in the map.

[0044] In the basic data set generation step, grouping the multidimensional data by train running direction, load type, passing time sequence, and track section location includes: The train running direction and load type are used as the first-level labels, wherein the running direction includes up and down, and the load type includes empty and loaded; The train passing time sequence and track section are used as secondary labels, and the track section includes straight segments, curved segments and switch areas; According to the historical desoldering event statistics, weights are assigned to different tag combinations.

[0045] In this technical solution, a multi-level data grouping system is constructed by using the train's running direction and load type as the first-level labels, and the train's passing sequence and track section as the second-level labels. This allows for more detailed data division and classification of the complex railway operating environment and train status information. The selected labels cover the key factors that affect the desoldering and displacement of the slide plate. The train's running direction affects the contact force and friction between the wheelset and the slide plate; the load type directly determines the pressure the train exerts on the slide plate; the train's passing sequence reflects the train's operating status within a specific time period; and the track sections have different geometric features and stress conditions. Historical desoldering events reflect the actual situation of slide plate failures under different operating conditions. By assigning weights to different label combinations, those label combinations that are more likely to cause desoldering events can be highlighted.

[0046] It can be understood that the total weight ≥ 80% of the rated load is considered heavy load; the curvature radius of the straight section is ≥ 3000m, the curvature radius of the curved section is < 3000m, and the turnout area is divided into the transition section between the point rail and the basic rail.

[0047] In the data matrix construction and anomaly detection steps, establishing a mapping relationship between vibration signals and image features includes: The collected vibration signal of the slide plate is processed in segments according to a certain signal duration; According to the sliding window technology, the vibration signal is processed with a certain step size to continuously extract time domain features; Perform fast Fourier transform on each vibration signal to convert the time domain signal into a frequency domain signal; The texture features of the slide bed image are extracted based on the local binary pattern algorithm; Extract key point features of the slide bed image based on the scale-invariant feature transformation algorithm; Based on the deep learning method, a convolutional neural network model is used to extract features from the image; A regression model between vibration signal features and image features is established based on the support vector regression algorithm, with the time domain features and frequency domain features of the vibration signal as input and the texture features and key point features of the image as output.

[0048] In this technical solution, the vibration signal is segmented into segments of fixed duration and processed using a sliding window technique with a fixed step size to continuously extract time-domain features. This technique captures the time-series variation of the vibration signal, such as amplitude fluctuations and periodicity, providing rich time-domain information for subsequent analysis and helping to fully understand the dynamic response of the slide plate. The time-domain signal is converted into a frequency-domain signal using a fast Fourier transform, revealing the distribution of the vibration signal across different frequency components. A local binary pattern algorithm is used to extract texture features from the slide plate image. These texture features reflect microstructural variations on the slide plate surface, such as surface roughness and scratches. These features are closely related to issues such as wear, desoldering, and displacement, providing important visual information for detection. A scale-invariant feature transform algorithm is used to extract key point features from the image. Key point features are invariant to transformations such as scaling, rotation, and translation, enabling accurate identification of important feature points in the slide plate image, such as bolt locations and edge contours. This provides a stable image feature foundation for subsequent correlation analysis with vibration signal features. A convolutional neural network model is used to extract features from images. Convolutional neural networks have powerful feature learning capabilities and can automatically learn deeper, more abstract features from images, including potential features of slideway desoldering and displacement, further enriching the representation of image features. A support vector regression algorithm is used to establish a regression model between vibration signal features and image features. This model uses the time and frequency domain features of the vibration signal as input and the texture and key point features of the image as output. This model can explore the inherent correlation between the vibration signal and image features. By continuously optimizing the model parameters, the accuracy and stability of the mapping relationship are improved, providing a more scientific basis for slideway desoldering and displacement detection.

[0049] Specifically, the spectral characteristics of the frequency domain signal are analyzed, and the frequency domain characteristics of the main frequency, bandwidth, and spectral energy are extracted. The main frequency represents the frequency component with the most concentrated energy in the vibration signal, the bandwidth reflects the frequency distribution range of the signal, and the spectral energy reflects the energy distribution of the signal at each frequency.

[0050] It is understandable that the time-frequency analysis method of short-time Fourier transform or wavelet transform can also be used to perform time-frequency joint analysis on the vibration signal to obtain the joint characteristics of the signal in time and frequency, so as to more comprehensively describe the dynamic characteristics of the vibration signal.

[0051] It can be understood that the local binary pattern algorithm is the LBP algorithm. The LBP algorithm can effectively describe the texture information of the local area of the image. By calculating the gray value relationship between each pixel and its neighboring pixels, an LBP feature vector is generated to characterize the texture changes on the surface of the slide bed.

[0052] It can be understood that the scale-invariant feature transform algorithm is the SIFT algorithm. The SIFT algorithm has scale invariance and rotation invariance, can detect key points of images at different scales, and calculate its feature descriptor to describe the key feature points on the surface of the slide bed and the features of the surrounding areas.

[0053] It can be understood that the convolutional neural network model is a CNN model. A CNN model suitable for the image features of the slide bed is constructed and trained through a large amount of image data so that the model can automatically learn the high-level feature representation of the image, such as edge, shape, and structure.

[0054] It can be understood that the support vector regression algorithm is an SVR algorithm, which can find the optimal regression hyperplane in the high-dimensional feature space and model the nonlinear relationship between the vibration signal features and the image features.

[0055] In one embodiment of the present invention, in the data matrix construction and anomaly detection step, the construction of a multidimensional data matrix containing the vibration frequency, vibration amplitude, force on the slide plate, and image of the slide plate includes: For the synchronously collected vibration frequency, vibration amplitude, and slide plate force data, adaptive median filtering is used to remove impulse noise, and wavelet threshold denoising is used to eliminate high-frequency interference. Different normalized weights are assigned to the vibration frequency, vibration amplitude, and slide plate force. The data are linearly mapped to the [0, 1] interval using the maximum and minimum normalization method to form a fused vibration-force feature vector. The features of the slide bed image are extracted by a lightweight convolutional neural network, and the extracted feature vector is normalized by the Z-Score method to make the feature mean 0 and variance 1. Taking the time when the train first arrives at the slide plate detection area as the benchmark, calculate the time difference between each subsequent data point and the benchmark time. Based on the train speed information, multiply the time difference by the inverse of the speed to obtain the time code value. Based on the track section division, unique spatial identification codes are assigned to different sections. A four-dimensional data matrix is constructed, in which the four dimensions correspond to time, space, vibration-force characteristics, and image characteristics, respectively. According to the spatiotemporal coding rules, the preprocessed vibration-force feature vectors and image feature vectors are filled into the corresponding positions of the matrix, and the missing data points are supplemented by the weighted average interpolation method based on spatiotemporal neighboring data.

[0056] In this technical solution, adaptive median filtering is used to remove impulse noise, and wavelet threshold denoising eliminates high-frequency interference, significantly improving the quality of vibration frequency, vibration amplitude, and slide plate force data. Impulse noise and high-frequency interference can affect the authenticity and reliability of data. Denoising can make the data smoother and more accurate, laying a good foundation for subsequent feature extraction and data analysis. Different normalization weights are assigned to the vibration frequency, vibration amplitude, and slide plate force, and the data are linearly mapped to the [0, 1] interval using the maximum-minimum normalization method, eliminating the influence of different dimensions and orders of magnitude on data analysis. The denoised and normalized vibration frequency, vibration amplitude, and slide plate force data are fused into a vibration-force feature vector, effectively integrating different types of features. Vibration and force information reflects the operating state of the slide plate from different perspectives. The fused feature vector provides a more comprehensive description of the slide plate's dynamic characteristics, providing richer information for subsequent fusion with image features. Slide plate image features are extracted using a lightweight convolutional neural network, and Z-score normalization is used to normalize the feature distribution, stabilizing the feature mean to 0 and the variance to 1. Using the time when the train first arrives at the slide plate inspection area as a benchmark, the time difference between each subsequent data point and the benchmark is calculated. Based on the train's speed information, the time difference is multiplied by the inverse of the speed to obtain a time code value. This considers the influence of train speed on time, making the time information more accurately reflect the train's operating status on the track and enhancing the correlation between the data and time. Based on the track segment division, each section is assigned a unique spatial identifier, clarifying the spatial location of the data. By combining the time code and spatial identifier, a data structure with spatiotemporal correlation is constructed, facilitating the analysis of slide plate state changes at different temporal and spatial locations, providing strong support for fault location and prediction. A data matrix encompassing four dimensions—temporal, spatial, vibration-stress, and image features—is constructed, integrating multi-source data. This comprehensively describes the slide plate's status, encompassing temporal, spatial, and physical characteristics. This provides richer information for slide plate desoldering and displacement detection, improving detection integrity. Missing data points are supplemented using a weighted average interpolation method based on spatiotemporal neighboring data, ensuring data continuity and integrity.

[0057] It can be understood that adaptive median filtering is a nonlinear filtering technology used to remove impulse noise in images or signals. Adaptive median filtering can dynamically adjust the window size according to the noise density in the local window, thereby more effectively retaining the detailed information of the image or signal while removing noise.

[0058] As you can understand, wavelet threshold denoising is a signal denoising method based on the wavelet transform. First, the signal is subjected to a wavelet transform to obtain wavelet coefficients of different scales. Then, a threshold is set based on the statistical properties of the wavelet coefficients (such as the noise level). Wavelet coefficients below the threshold are set to zero. Finally, the denoised signal is reconstructed through an inverse wavelet transform. This effectively removes high-frequency noise while preserving the signal's low-frequency components.

[0059] It can be understood that a lightweight convolutional neural network refers to a convolutional neural network that reduces network parameters and computational complexity by optimizing the network structure and adopting lightweight modules. While maintaining a high accuracy, the lightweight convolutional neural network has a faster inference speed and lower resource consumption.

[0060] As you can understand, the Z-Score normalization method is a data normalization technique used to convert data into a standard normal distribution with a mean of 0 and a variance of 1. For each data point, the difference between its mean and the dataset is calculated and then divided by the standard deviation of the dataset. This can eliminate dimensional differences between the data, making different features comparable, and also helps improve the convergence speed and stability of the model.

[0061] It can be understood that the four-dimensional data matrix is M(t, s, f, i), where: t is the time dimension, which is the time when the train first enters the detection area As a benchmark, the time code value calculation formula is: ; in, For the current moment and The time difference is , v is the real-time speed of the train; S is the spatial dimension, based on the track mileage coordinate system, with the center of the slide plate as the reference point, and the track section division rules are: Straight line segment: Every 50 meters is a coding unit, numbered L001, L002, ...; Curve segment: divided into segments according to the curvature radius (e.g. <3000m is a segment), numbered C001, C002, ...; Turnout area: With the tip of the switch rail as the origin, the coding units are 10m extending to both sides, numbered S001, S002, ...; f is the vibration-force characteristic dimension, including the vibration frequency , vibration amplitude , the normalized eigenvector of the force F ,in + + =1 is the weight coefficient, which is set according to the sensor accuracy; i is the image feature dimension, and the 128-dimensional feature vector extracted by lightweight CNN is stored after Z-Score normalization.

[0062] As can be understood, the sensor synchronization mechanism adopts GPS clock synchronization + hardware triggering: when the train wheels trigger the track sensor, a hardware trigger signal is sent to all sensors to ensure that the start time of data collection is synchronized.

[0063] It can be understood that spatiotemporal correlation refers to the dynamic correspondence between multidimensional data in time series (train passage sequence, detection cycle) and spatial position (track mileage, section type), while also associating train operation status parameters (speed, load, and formation) to form a "time-space-state" trinity data correlation model.

[0064] In the data matrix construction and anomaly detection steps, the calculation of vibration waveform similarity based on the dynamic time warping algorithm includes: According to the DTW algorithm, a distance matrix is constructed. The matrix elements represent the distance between two vibration waveforms at different time points. When calculating the cumulative distance matrix, the path slope is restricted to a certain range. When calculating the path cost, if the vibration amplitude change between adjacent time points exceeds a preset threshold, the cost of this path point is additionally weighted to obtain the cumulative distance reflecting the similarity of the two vibration waveforms.

[0065] In this technical solution, a distance matrix is constructed based on the dynamic time warping algorithm, with the matrix elements accurately representing the distance between two vibration waveforms at different time points. Distance calculations based on waveform decomposition at specific time points fully consider the waveform's local characteristics along the time axis, providing a detailed and accurate data foundation for subsequent similarity calculations and avoiding the potential bias associated with similarity assessments based solely on the overall waveform profile. When calculating the cumulative distance matrix, the path slope is constrained to a certain range, ensuring a more reasonable match between the two vibration waveforms along the time axis and preventing mismatches caused by excessive distortion of the time axis. When calculating the path cost, additional weighting is applied to path points where the vibration amplitude change between adjacent time points exceeds a preset threshold. When a slide plate desoldering or shifting fault occurs, the vibration waveform amplitude often exhibits abnormal changes. By applying additional weighting to key change points, the algorithm focuses on these fault-related features, effectively capturing the slide plate's fault characteristics and improving fault detection sensitivity. By processing the path slope and path cost described above, the DTW algorithm can more accurately identify patterns in the vibration waveform associated with slide plate desoldering or shifting faults.

[0066] Preferably, the limiting slope is between 0.5 and 2.

[0067] In the data matrix construction and anomaly detection steps, detecting the deformation area on the slide bed surface according to the image feature matching algorithm includes: The key point features of the current detection image and the historical normal state image are extracted respectively according to the scale-invariant feature transformation algorithm; By calculating the Euclidean distance between the two sets of image feature descriptors, we filter out feature point pairs with a matching degree higher than a preset threshold, generate an initial matching point set, and eliminate mismatched points in the initial matching point set using a random sampling consistency algorithm. According to the coordinate information of the matching point pairs, the geometric transformation parameters between the current image and the historical normal image are calculated through the affine transformation model, and the current image is mapped to the coordinate system of the historical normal image so that the two have the same geometric reference; The grayscale value difference between the corresponding position of the current image and the historical normal image after mapping is compared pixel by pixel, and a grayscale difference threshold is set. When the proportion of pixels in a certain area whose grayscale value difference exceeds the threshold reaches a preset ratio, the area is determined to be a deformation area of the slide bed surface.

[0068] In this technical solution, a scale-invariant feature transformation algorithm is used to extract key point features of the current detection image and the historical normal state image respectively. By calculating the Euclidean distance between the two sets of image feature descriptors to screen feature point pairs with high matching, and using a random sampling consistency algorithm to exclude mismatched points, the accuracy of feature matching is further improved and detection errors caused by mismatching are reduced. After mapping the current image to the coordinate system of the historical normal image, the grayscale value differences of the corresponding positions are compared pixel by pixel. The grayscale value differences can intuitively reflect the changes in the image surface. By setting a reasonable grayscale difference threshold and a preset ratio of the number of pixels, it is possible to accurately determine whether there is a deformation area on the slide bed surface. It has high sensitivity and accuracy and can effectively detect small deformations on the slide bed surface. Based on the coordinate information of the matching point pairs, the geometric transformation parameters between the current image and the historical normal image are calculated through the affine transformation model, and the current image is mapped to the coordinate system of the historical normal image, achieving precise alignment of the two images. When comparing the grayscale value differences, they can accurately correspond to the same position, improving the accuracy of fault location. Even if the image has a certain degree of geometric deformation such as translation, rotation and scaling, the affine transformation model can correct it, ensuring the accuracy of the detection results.

[0069] As you can understand, Euclidean distance refers to the straight-line distance between two points. In image feature matching, Euclidean distance is often used to calculate the similarity or distance between two feature vectors. The smaller the Euclidean distance, the more similar the two feature vectors are.

[0070] As you can understand, the random sampling consensus algorithm is the RANSAC algorithm. It is an algorithm used to estimate mathematical model parameters from a dataset containing outliers. It randomly selects a portion of points in the dataset as inliers. It then fits the model based on the inliers and calculates the distance from other points to the model. If the distance is less than a certain threshold, these points are also considered inliers and the model is refitted. This process is repeated until the optimal model parameters are found. The RANSAC algorithm can effectively handle datasets containing a large number of outliers and obtain robust model estimation results.

[0071] As you can understand, geometric transformation parameters are parameters used to describe the geometric transformation relationship between images. Common geometric transformations include translation, rotation, scaling, and affine transformations. In image matching and stitching, geometric transformation parameters must be calculated to align images taken from different perspectives or at different times to the same coordinate system. These parameters are typically calculated based on the matching results of feature points in the images, using the least squares method or other optimization algorithms.

[0072] like Figure 2 As shown in the second embodiment of the present invention, a railway track slide bed plate desoldering and displacement detection system includes: Data acquisition module: used to synchronously collect multi-dimensional data of the slide bed vibration frequency, vibration amplitude, slide bed force and slide bed image through multiple sensors; Data grouping module: groups the multidimensional data according to train running direction, load type, passing time sequence and track section position to generate a basic data set containing spatiotemporal correlation; Relationship building and comparison module: This module acquires vibration signals, synchronously collects image sequences, establishes a mapping relationship between vibration signals and image features, and constructs a multidimensional data matrix containing the slide plate's vibration frequency, vibration amplitude, slide plate force, and slide plate image. It also performs a time-synchronized comparison of the current detection group data with the historical normal state data set, calculates the vibration waveform similarity using a dynamic time warping algorithm, and detects the deformation area on the slide plate surface using an image feature matching algorithm. Early warning and judgment module: Through intra-group comparison, the vibration waveform feature offset and the image deformation area growth rate are calculated. When the vibration waveform feature offset exceeds the preset threshold or the image deformation area growth rate exceeds the baseline value, a primary early warning is triggered independently. When the two detection results form a spatial correlation, it is determined to be a desoldering abnormality event; Abnormal analysis module: Determine the abnormal starting point based on the vibration frequency, vibration amplitude, force and image of the slide bed plate. Based on the train formation information when the abnormal event occurs, divide the desoldering location into regions, identify the desoldering features, and classify the desoldering features.

[0073] In this technical solution, the data acquisition module uses multiple sensors to synchronously collect multidimensional data, including the slide plate's vibration frequency, amplitude, forces acting on the slide plate, and images. This data covers various aspects of the slide plate's mechanical motion and appearance. The data grouping module groups the data by train direction, load type, transit time, and track section location, generating a basic dataset with temporal and spatial correlations, making the data more organized and analyzable. The relationship building and comparison module further establishes a mapping relationship between vibration signals and image features, constructing a multidimensional data matrix. This multi-dimensional, multi-level integration approach comprehensively reflects the slide plate's operating status, significantly improving the comprehensiveness of the inspection. The relationship building and comparison module performs a time-synchronized comparison of the current inspection data set with a historical normal state dataset. It uses a dynamic time warping algorithm to calculate vibration waveform similarity and an image feature matching algorithm to detect surface deformation areas on the slide plate. These evaluations complement each other to more accurately detect slide plate anomalies. The early warning and judgment module calculates the vibration waveform feature offset and the area growth rate of the image deformation region, and sets corresponding thresholds for early warning and judgment, which further improves the accuracy of detection and reduces the possibility of false alarms and missed alarms.

[0074] Preferably, the data grouping module includes: Label setting unit: takes the train running direction and load type as the first-level label, where the running direction includes up and down, and the load type includes empty and loaded; takes the train passing sequence and track section as the second-level label, where the track section includes straight section, curved section, and switch section; Weight allocation unit: Assign weights to different tag combinations based on historical desoldering event statistics.

[0075] In this technical solution, the label setting unit groups multidimensional data using train direction, load type, transit sequence, and track section as labels. Train direction and load type serve as primary labels, reflecting the different directions and magnitudes of the train's forces on the slide plate. Under different directions and loads, the forces acting on the slide plate vary significantly, leading to different wear and deformation patterns. Track section and train transit sequence serve as secondary labels. Different track sections result in different lateral and longitudinal forces on the slide plate. Transit sequence, on the other hand, considers the temporal patterns of train operation and provides a comprehensive and detailed description of the slide plate's operating environment and usage, making data classification more scientific and reasonable. This hierarchical structure of primary and secondary labels not only highlights the primary influencing factors of direction and load type, but also considers the secondary factors of transit sequence and track section. This makes data classification clearer and more organized, facilitates subsequent analysis and processing of data with different label combinations, and improves the system's maintainability and scalability.

[0076] The specific description of the present invention in the above embodiments is only used to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Technical engineers in this field may make some non-essential improvements and adjustments to the present invention based on the contents of the above invention, which fall within the scope of protection of the present invention.

Claims

1. A method for detecting the desoldering and displacement of a railway track slide bed, characterized in that: The following steps are involved: a slide plate data acquisition step, synchronously collecting multi-dimensional data of the slide plate vibration frequency, vibration amplitude, slide plate force and slide plate image through multiple sensors; The basic data set generation step is to group the multidimensional data according to the train running direction, load type, passing time sequence and track section location to generate a basic data set containing spatiotemporal correlation; Data matrix construction and anomaly detection steps: obtain vibration signals, synchronously collect image sequences, establish a mapping relationship between vibration signals and image features, construct a multidimensional data matrix containing the slide plate vibration frequency, vibration amplitude, slide plate force and slide plate image, perform time synchronization comparison between the current detection group data and the historical normal state data set, calculate the vibration waveform similarity based on the dynamic time warping algorithm, and detect the deformation area on the slide plate surface based on the image feature matching algorithm; The desoldering abnormality judgment step calculates the vibration waveform feature offset and the image deformation area area growth rate through intra-group comparison. When the vibration waveform feature offset exceeds the preset threshold or the image deformation area area growth rate exceeds the baseline value, the primary warning is triggered independently. When the two detection results form a spatial position correlation, it is determined to be a desoldering abnormality event.

2. A railway track slide plate desoldering and displacement detection method according to claim 1, characterized in that: Also includes: The desoldering feature classification step determines the abnormal starting point based on the vibration frequency, vibration amplitude, force of the slide bed and the slide bed image, divides the desoldering location into regions based on the train formation information when the abnormal event occurs, identifies the desoldering features, and classifies the desoldering features.

3. A railway track slide plate desoldering and displacement detection method according to claim 2, characterized in that: In the basic data set generation step, the time sequence includes: Install visual sensors on the train. When the train passes through a tunnel for the first time, it uses visual SLAM technology to build a 3D map of the tunnel and record the location information of key feature points. When the train passes through the tunnel, the visual sensor captures real-time image information inside the tunnel. Using the visual SLAM algorithm, feature points are extracted from the real-time image and matched with feature points in the pre-built map. Based on the real-time train position determined by the visual SLAM algorithm and combined with train speed information, the time when the train wheels pass through the slide bed is predicted. With the predicted passing time as the center, the same period of time is extended forward and backward to form a fixed time window; Taking the time when the train first arrives at the slide bed detection area as the benchmark, calculate the time difference between each subsequent data point and the benchmark time. According to the train speed information, multiply the time difference by the inverse of the speed to obtain the time code value. According to the track section division, assign unique spatial identification codes to different sections.

4. A railway track slide plate desoldering and displacement detection method according to claim 3, characterized in that: In the basic data set generation step, grouping the multidimensional data by train running direction, load type, passing time sequence, and track section location includes: The train running direction and load type are used as the first-level labels, wherein the running direction includes up and down, and the load type includes empty and loaded; The train passing time sequence and track section are used as secondary labels, and the track section includes straight segments, curved segments and switch areas; According to the historical desoldering event statistics, weights are assigned to different tag combinations.

5. A railway track slide bed plate desoldering and displacement detection method according to claim 4, characterized in that: In the data matrix construction and anomaly detection steps, establishing a mapping relationship between vibration signals and image features includes: The collected vibration signal of the slide plate is processed in segments according to a certain signal duration; According to the sliding window technology, the vibration signal is processed with a certain step size to continuously extract time domain features; Perform fast Fourier transform on each vibration signal to convert the time domain signal into a frequency domain signal; The texture features of the slide bed image are extracted based on the local binary pattern algorithm; Extract key point features of the slide bed image based on the scale-invariant feature transformation algorithm; Based on the deep learning method, a convolutional neural network model is used to extract features from the image; A regression model between vibration signal features and image features is established based on the support vector regression algorithm, with the time domain features and frequency domain features of the vibration signal as input and the texture features and key point features of the image as output.

6. A railway track slide plate desoldering and displacement detection method according to claim 5, characterized in that: In the data matrix construction and anomaly detection step, the construction of a multidimensional data matrix including the vibration frequency, vibration amplitude, force on the slide plate, and image of the slide plate includes: For the synchronously collected vibration frequency, vibration amplitude, and slide plate force data, adaptive median filtering is used to remove impulse noise, and wavelet threshold denoising is used to eliminate high-frequency interference. Different normalized weights are assigned to the vibration frequency, vibration amplitude, and slide plate force. The data are linearly mapped to the [0, 1] interval using the maximum and minimum normalization method to form a fused vibration-force feature vector. The features of the slide bed image are extracted by a lightweight convolutional neural network, and the extracted feature vector is normalized by the Z-Score method to make the feature mean 0 and variance 1. Taking the time when the train first arrives at the slide plate detection area as the benchmark, calculate the time difference between each subsequent data point and the benchmark time. Based on the train speed information, multiply the time difference by the inverse of the speed to obtain the time code value. Based on the track section division, unique spatial identification codes are assigned to different sections. A four-dimensional data matrix is constructed, in which the four dimensions correspond to time, space, vibration-force characteristics, and image characteristics, respectively. According to the spatiotemporal coding rules, the preprocessed vibration-force feature vectors and image feature vectors are filled into the corresponding positions of the matrix, and the missing data points are supplemented by the weighted average interpolation method based on spatiotemporal neighboring data.

7. A railway track slide plate desoldering and displacement detection method according to claim 6, characterized in that: In the data matrix construction and anomaly detection steps, the calculation of vibration waveform similarity based on the dynamic time warping algorithm includes: According to the DTW algorithm, a distance matrix is constructed. The matrix elements represent the distance between two vibration waveforms at different time points. When calculating the cumulative distance matrix, the path slope is restricted to a certain range. When calculating the path cost, if the vibration amplitude change between adjacent time points exceeds a preset threshold, the cost of this path point is additionally weighted to obtain the cumulative distance reflecting the similarity of the two vibration waveforms.

8. A method for detecting desoldering and displacement of a railway track slide plate according to claim 5, 6 or 7, characterized in that: In the data matrix construction and anomaly detection steps, detecting the deformation area on the slide bed surface according to the image feature matching algorithm includes: The key point features of the current detection image and the historical normal state image are extracted respectively according to the scale-invariant feature transformation algorithm; By calculating the Euclidean distance between the two sets of image feature descriptors, we filter out feature point pairs with a matching degree higher than a preset threshold, generate an initial matching point set, and eliminate mismatched points in the initial matching point set using a random sampling consistency algorithm. According to the coordinate information of the matching point pairs, the geometric transformation parameters between the current image and the historical normal image are calculated through the affine transformation model, and the current image is mapped to the coordinate system of the historical normal image so that the two have the same geometric reference; The grayscale value difference between the corresponding position of the current image and the historical normal image after mapping is compared pixel by pixel, and a grayscale difference threshold is set. When the proportion of pixels in a certain area whose grayscale value difference exceeds the threshold reaches a preset ratio, the area is determined to be a deformation area of the slide bed surface.

9. A railway track slide bed plate desoldering and displacement detection system, used to implement the railway track slide bed plate desoldering and displacement detection method according to claims 1-8, characterized in that: include: Data acquisition module: used to synchronously collect multi-dimensional data of the slide bed vibration frequency, vibration amplitude, slide bed force and slide bed image through multiple sensors; Data grouping module: groups the multidimensional data according to train running direction, load type, passing time sequence and track section position to generate a basic data set containing spatiotemporal correlation; Relationship building and comparison module: This module acquires vibration signals, synchronously collects image sequences, establishes a mapping relationship between vibration signals and image features, and constructs a multidimensional data matrix containing the slide plate's vibration frequency, vibration amplitude, slide plate force, and slide plate image. It also performs a time-synchronized comparison of the current detection group data with the historical normal state data set, calculates the vibration waveform similarity using a dynamic time warping algorithm, and detects the deformation area on the slide plate surface using an image feature matching algorithm. Early warning and judgment module: Through intra-group comparison, the vibration waveform feature offset and the image deformation area growth rate are calculated. When the vibration waveform feature offset exceeds the preset threshold or the image deformation area growth rate exceeds the baseline value, a primary early warning is triggered independently. When the two detection results form a spatial correlation, it is determined to be a desoldering abnormality event; Abnormal analysis module: Determine the abnormal starting point based on the vibration frequency, vibration amplitude, force and image of the slide bed plate. Based on the train formation information when the abnormal event occurs, divide the desoldering location into regions, identify the desoldering features, and classify the desoldering features.

10. The railway track slide plate desoldering and displacement detection system according to claim 9, characterized in that: The data grouping module includes: Label setting unit: takes the train running direction and load type as the first-level label, where the running direction includes up and down, and the load type includes empty and loaded; takes the train passing sequence and track section as the second-level label, where the track section includes straight section, curved section, and switch section; Weight allocation unit: Assign weights to different tag combinations based on historical desoldering event statistics.

Images