Railway sleeper intelligent detection method and system based on multi-source data fusion

Abstract

The invention provides an intelligent railway sleeper detection method and system based on multi-source data fusion, and relates to the technical field of railway track detection, and the method comprises the steps: obtaining multi-source original data of a railway track; performing data preprocessing according to the multi-source original data, and eliminating sensor inherent noise and environmental interference through denoising operation to obtain standardized data; performing feature extraction according to the standardized data to obtain a feature vector set of sleeper arrangement; performing data fusion according to the feature vector set to obtain a comprehensive state score of the sleeper; performing anomaly detection according to the comprehensive state score to obtain an anomaly identification result; and performing prediction analysis according to an abnormal identification result, analyzing a historical data trend through time sequence modeling to predict future change of the sleeper state, and generating a maintenance suggestion based on a prediction result. The automation level, the result accuracy and the operation and maintenance efficiency of the detection process are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of railway track inspection technology, and more specifically, to an intelligent inspection method and system for railway sleepers using multi-source data fusion. Background Technology

[0002] In the field of railway track inspection and maintenance technology, sleepers, as the core load-bearing component of the track structure, directly determine the geometric stability, smoothness, and train operation safety of the track through their arrangement and health condition. With the continuous development of sensing technology, data acquisition and processing technology, railway inspection has gradually evolved from relying on manual foot inspections and handheld tool measurements in the early days to adopting various online monitoring systems, such as infrared axle temperature detection systems, track dynamic detection systems, and acoustic detection systems, realizing the automated acquisition of sleeper and track component status parameters. However, existing technologies still have significant limitations in dealing with the complex issue of sleeper arrangement: on the one hand, traditional manual inspection methods are inefficient and costly, and the results are easily affected by personnel experience and subjective state, making it difficult to achieve routine monitoring of large-scale road networks; on the other hand, even with the use of automated detection equipment, existing methods are mostly limited to independent analysis of data from single-type sensors and simple threshold alarms, lacking effective fusion and collaborative analysis of multi-source heterogeneous data, resulting in a single dimension in the characterization of sleeper condition, failing to comprehensively and accurately reflect its true physical state and evolution trend, and especially making it difficult to detect complex anomalies such as loosening, cracks, and settlement at an early stage.

[0003] Based on the shortcomings of the existing technologies, there is an urgent need for a smart detection method and system for railway sleepers that integrates multi-source data. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent detection of railway sleepers based on multi-source data fusion, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides a method for intelligent detection of railway sleepers based on multi-source data fusion, including:

[0006] Acquire multi-source raw data of railway tracks, including infrared thermometry data, dynamic mechanical data, acoustic spectrum data, image data, and train operation status data;

[0007] Data preprocessing is performed based on the multi-source raw data. Noise reduction is performed to eliminate inherent sensor noise and environmental interference. Combined with normalization, data of different dimensions are scaled to the same numerical range to obtain standardized data.

[0008] Feature extraction is performed based on the standardized data. By extracting the geometric, dynamic, and thermal features of the sleeper arrangement from multi-source data, a feature vector set of the sleeper arrangement is obtained.

[0009] Data fusion is performed based on the set of feature vectors. Through a comprehensive fusion process that integrates multi-source features and combines prior information, a comprehensive status score for the sleeper is obtained.

[0010] Anomaly detection is performed based on the comprehensive status score. Real-time data is compared with historical normal patterns through pattern recognition, and abnormal sleeper arrangement is identified based on threshold comparison to obtain anomaly identification results.

[0011] Based on the anomaly identification results, predictive analysis is performed, and historical data trends are analyzed through time series modeling to predict future changes in sleeper condition. Maintenance recommendations are then generated based on the prediction results.

[0012] Secondly, this application also provides a multi-source data fusion-based intelligent detection system for railway sleepers, comprising:

[0013] The acquisition module is used to acquire multi-source raw data of railway tracks, including infrared temperature measurement data, dynamic mechanical data, acoustic spectrum data, image data, and train operation status data.

[0014] The processing module is used to perform data preprocessing based on the multi-source raw data, eliminate inherent noise of the sensor and environmental interference through denoising operation, and combine normalization processing to uniformly scale data of different dimensions to the same numerical range to obtain standardized data.

[0015] The extraction module is used to extract features based on the standardized data. By extracting the geometric, dynamic and thermal features of the sleeper arrangement from the multi-source data, a feature vector set of the sleeper arrangement is obtained.

[0016] The fusion module is used to perform data fusion based on the feature vector set. Through a comprehensive fusion process that integrates multi-source features and combines prior information, a comprehensive status score of the sleeper is obtained.

[0017] The detection module is used to detect anomalies based on the comprehensive status score, compare real-time data with historical normal patterns through pattern recognition, and identify abnormal sleeper arrangement based on threshold comparison to obtain anomaly identification results.

[0018] The output module is used to perform predictive analysis based on the anomaly identification results, predict future changes in sleeper condition by analyzing historical data trends through time series modeling, and generate maintenance suggestions based on the prediction results.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention acquires multi-source raw data of railway tracks and performs continuous intelligent processing, including data preprocessing, feature extraction, data fusion, anomaly detection, and predictive analysis, thereby achieving comprehensive perception, accurate diagnosis, and proactive maintenance of sleeper arrangement status. This significantly improves the automation level, accuracy of results, and efficiency of operation and maintenance in the detection process. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the intelligent detection method for railway sleepers based on multi-source data fusion as described in this embodiment of the invention.

[0023] Figure 2 This is a schematic diagram of the intelligent railway sleeper detection system based on multi-source data fusion as described in this embodiment of the invention.

[0024] Figure 3 This is a schematic diagram of a multi-sensor collaborative detection system for railway tracks.

[0025] Figure 4 This is a schematic diagram of the overall system architecture.

[0026] The markings in the diagram are: 1. Roadbed; 2. Ballast; 3. Sleeper; 4. Rail; 5. Image acquisition device; 6. Infrared temperature measurement acquisition device; 7. Mechanical sensor array acquisition device; 8. Train status detection on-board detection equipment; 9. Acoustic acquisition device; 901. Acquisition module; 902. Processing module; 903. Extraction module; 904. Fusion module; 905. Detection module; 906. Output module. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] Example 1:

[0030] This embodiment provides a method for intelligent detection of railway sleepers based on multi-source data fusion.

[0031] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0032] Step S100: Obtain multi-source raw data of railway track, including infrared temperature measurement data, dynamic mechanical data, acoustic spectrum data, image data, and train operation status data;

[0033] Specifically, multi-source raw data is processed through methods such as... Figure 3 The railway track multi-sensor collaborative detection system shown integrates various sensor technologies to achieve comprehensive monitoring of the railway track and its surrounding environment. These sensors include image acquisition devices, infrared temperature measurement devices, mechanical sensor arrays, onboard detection equipment, and acoustic acquisition devices. They work together to improve the accuracy and efficiency of detection, ensuring the safety and reliability of railway transportation. Figure 3 As shown in the figure, the roadbed 1 is located at the bottom of the figure and is the basic support structure for the railway track. It is usually made of earth and stone materials and is used to support the entire track system. Figure 3The ballast 2, located above the roadbed, is a layer of crushed stone or crushed stone concrete used to distribute the load generated by passing trains, reduce pressure on the roadbed, and provide drainage. The sleepers 3, located above the ballast, are crucial components for fixing the rails and maintaining track geometry; they are typically made of wood, concrete, or steel. The rails 4 are the core components of the railway track, guiding the train's direction, bearing its weight, and providing the path for the train. Video cameras (image acquisition devices 5), installed on the train, have four imaging capabilities: visible light, fog penetration, short-wave infrared, and long-wave infrared, used to capture image information of the track and its surrounding environment. Temperature infrared sensors (infrared temperature measurement acquisition devices 6) are used to monitor track temperature, detecting the temperature distribution on the track surface using infrared technology to prevent track deformation or damage caused by temperature changes. A mechanical sensor array acquisition device 7, installed on the track, is used to detect the mechanical effects exerted on the track by passing trains, including pressure and vibration, to assess the track's load-bearing capacity and condition. The onboard train condition monitoring equipment 8 is installed on the train to monitor its operating status in real time, including parameters such as speed and acceleration, to ensure safe operation. The acoustic acquisition device 9 captures acoustic signals from the track and its surrounding environment, detecting abnormal sounds in the track, such as cracks or loosening, through acoustic analysis to identify potential faults in advance. This data acquisition is based on actual railway inspection scenarios, and sensor deployment considers key points on the line to ensure comprehensive coverage of multi-dimensional information on the physical condition of the sleepers, providing raw input for subsequent processing.

[0034] Step S200: Perform data preprocessing based on multi-source raw data, eliminate sensor inherent noise and environmental interference through denoising operation, and combine normalization processing to uniformly scale data of different dimensions to the same numerical range to obtain standardized data.

[0035] It is understandable that the data preprocessing process in step S200 eliminates inherent sensor noise (such as thermal noise or electromagnetic interference) and environmental influences (such as weather factors) by performing denoising operations on the multi-source raw data, and combines normalization processing to uniformly scale data of different dimensions (such as temperature units and vibration units) to the same numerical range, thereby obtaining standardized data and improving data consistency and comparability.

[0036] Step S300: Perform feature extraction based on standardized data. By extracting the geometric, dynamic, and thermal features of sleeper arrangement from multi-source data, a feature vector set of sleeper arrangement is obtained.

[0037] It should be noted that the feature extraction in step S300 derives the key features of sleeper arrangement from standardized data, including extracting geometric features from image data, extracting dynamic features from dynamic mechanical data, and extracting thermal features from infrared thermometry data. These features are then integrated to form a set of feature vectors to comprehensively characterize the multidimensional physical state of the sleepers.

[0038] Step S400: Perform data fusion based on the feature vector set. Through a comprehensive fusion process that integrates multi-source features and combines prior information, obtain the comprehensive status score of the sleeper.

[0039] The data fusion in step S400 is a comprehensive process that integrates multi-source features and combines prior information to collaboratively process heterogeneous data in the feature vector set, generating a comprehensive condition score for the sleeper and enhancing the robustness and accuracy of the condition assessment.

[0040] Step S500: Anomaly detection is performed based on the comprehensive status score. Real-time data is compared with historical normal patterns through pattern recognition, and abnormal sleeper arrangement is identified based on threshold comparison to obtain anomaly identification results.

[0041] The anomaly detection in step S500 is based on a comprehensive status score. It uses pattern recognition technology to compare real-time data with historical normal patterns and identifies abnormal sleeper arrangement through a threshold comparison mechanism, outputs anomaly identification results, and achieves accurate anomaly location.

[0042] Step S600: Perform predictive analysis based on the anomaly identification results, predict future changes in sleeper condition by analyzing historical data trends through time series modeling, and generate maintenance recommendations based on the prediction results.

[0043] Step S600's predictive analysis, based on anomaly identification results, applies time series modeling to analyze historical data trends, predicts future changes in sleeper condition, and generates maintenance recommendations based on the prediction results, supporting forward-looking operation and maintenance decisions. The overall system architecture of the above method is as follows: Figure 4 As shown.

[0044] Further, step S200 includes steps S210 to S230.

[0045] Step S210: Perform noise elimination processing on the multi-source raw data. By applying the sensor-specific noise model, eliminate the track vibration interference and thermal noise unique to sleeper detection to obtain preliminary clean data.

[0046] Step S220: Based on the preliminary cleaning data, perform effective data extraction processing, and remove invalid data points by screening criteria based on the spatial distribution pattern of sleepers to obtain the filtered data;

[0047] Step S230: Perform dimension unification processing on the filtered data, and convert the multi-source data into a dimensionless form based on the data distribution characteristics to obtain standardized data.

[0048] In the data processing workflow, preprocessing of raw data is involved to improve the accuracy and efficiency of subsequent analysis. Denoising refers to reducing random errors or interference in the data through algorithms, thereby extracting a purer signal. Screening involves selecting useful information from a large amount of data and eliminating irrelevant or redundant data. Labeling involves classifying or marking data to facilitate machine learning and pattern recognition. Image enhancement aims to improve the visual effects of images, such as contrast and brightness, making features in the image more prominent. Normalization scaling involves scaling the data to a specific range, such as 0 to 1, to eliminate the influence of different units or magnitudes and ensure data consistency and comparability. In steps S210 to S230 of this invention, these preprocessing methods are specifically applied to railway sleeper inspection scenarios, forming a complete data processing chain: Step S210 performs noise removal processing on multi-source raw data, eliminating track vibration interference and thermal noise unique to sleeper inspection by applying a sensor-specific noise model, thus achieving denoising and obtaining preliminary clean data; Step S220 performs effective data extraction processing on the preliminary clean data, removing invalid data points based on screening criteria based on the spatial distribution patterns of sleepers, such as using the periodic characteristics of sleeper arrangement to identify and exclude abnormal collected values ​​caused by train obstruction or environmental transients, thus achieving screening and obtaining filtered data; Step S230 performs dimensional unification processing on the filtered data, converting multi-source data into a dimensionless form based on data distribution characteristics, for example, scaling heterogeneous data such as temperature and vibration to a uniform range through Min-Max normalization, thus achieving normalization and obtaining standardized data. This series of processes closely integrates with the actual needs of railway sleeper inspection, providing reliable input for subsequent feature extraction by progressively purifying, screening, and standardizing the raw data.

[0049] Further, step S300 includes steps S310 to S330.

[0050] Step S310: Perform multi-source feature extraction based on standardized data. Extract sleeper surface temperature distribution features from infrared thermometry data, vibration frequency features from dynamic mechanical data, sleeper spacing and angle features from image data, and integrate acoustic spectrum data and train status data features to obtain a preliminary feature pool.

[0051] Step S320: Based on the preliminary feature pool, feature selection is performed, and based on the importance analysis of historical data and the verification of the physical meaning of sleepers, redundant features are removed to obtain a set of key features.

[0052] Step S330: Perform matrix construction processing based on the key feature set. By organizing the feature vectors according to the sleeper number and introducing a spatiotemporal expansion mechanism, a feature vector set of sleeper arrangement is obtained.

[0053] In steps S310 to S330, the feature extraction process unfolds based on the logical system for constructing the key feature matrix of sleeper arrangement. First, step S310 performs multi-source feature extraction based on standardized data. This involves extracting sleeper surface temperature distribution features (such as average temperature and gradient distribution) from infrared thermography data, vibration frequency features (such as dominant frequency and peak amplitude) from dynamic mechanical data, and sleeper spacing and angle features from image data. It also integrates acoustic spectrum data (such as average sound pressure level) and train status data features (such as dynamic response delay) to form a preliminary feature pool covering the multidimensional physical state of the sleepers. Specifically, initial features are extracted from the five types of sensor data shown in the table below to form the preliminary feature pool:

[0054] Table 1 Initial Feature Extraction Table

[0055]

[0056] Subsequently, step S320 performs feature selection based on the preliminary feature pool. This involves importance analysis based on historical data, such as evaluating feature contribution through machine learning models and verifying the physical meaning of sleepers (retaining key features like sleeper spacing and dominant vibration frequency), while eliminating redundant features (e.g., eliminating information overlap through correlation analysis) to obtain a set of key features, ensuring the simplicity and representativeness of the feature set. Preferably, during feature selection and importance assessment, multiple machine learning models, such as Random Forest, XGBoost / LightGBM, LASSO regression, and mutual information methods, are first trained based on historical detection data (including normal and abnormal annotations). These models are then used to calculate Gini importance, average impurity reduction, gain importance, and coverage importance. The contribution of each feature to the state classification is evaluated using indicators such as sex, coefficient shrinkage, and nonlinear correlation strength. Based on this, key features with clear physical significance are verified and confirmed by combining the experience of railway engineering experts. Features such as sleeper spacing (directly affecting track geometry), dominant vibration frequency (reflecting the coupling state between sleeper and track bed), surface temperature gradient (early warning of local stress concentration), and crack propagation rate (characterizing structural damage evolution) must be retained. Finally, through methods such as Pearson correlation coefficient analysis (with a threshold greater than 0.85), principal component analysis to evaluate feature independence, and variance expansion factor to detect multicollinearity, highly correlated or redundant features with overlapping information are systematically eliminated, completing the entire process optimization of feature selection.

[0057] In step S330, the features retained after screening are organized into a key feature matrix arranged in sleeper configuration according to the following structure:

[0058] ;

[0059] in, This represents the key feature matrix of sleeper arrangement. The numbering indicates the number of the sleeper (e.g., from sleeper #001 to sleeper #n), with each row corresponding to one sleeper; This indicates the number of key features retained after filtering, with each column corresponding to one feature. Indicates the first The root sleeper in the first The values ​​for each feature (already standardized). Furthermore, the table below provides examples of the physical definitions, data sources, and calculation methods for typical features in the matrix (this can be expanded as needed):

[0060] Table 2. Detailed list of key features of sleeper arrangement

[0061]

[0062] After constructing a two-dimensional feature matrix, a spatiotemporal expansion mechanism is introduced to stack the feature matrices of multiple time segments along the time dimension, forming a three-dimensional feature data volume. ,in Indicates the first The feature matrix at each detection time. Using timestamps, this structure supports time-series prediction tasks in feature trend analysis (such as displacement accumulation over time), state evolution modeling, and predictive maintenance. Simultaneously, all features undergo standardization before being stored (e.g., Z-score standardization or Min-Max normalization) to eliminate the influence of dimensions, and are stored in shards using Parquet or a time-series database for fast querying and incremental updates. Combined with version management (e.g., generating matrix version identifiers for each round of detection) and dynamic update mechanisms (e.g., generating matrix slices according to the detection cycle and using a sliding window to retain recent data), the final result is a set of feature vectors for sleeper arrangement, providing a structured foundation for subsequent data fusion.

[0063] Further, step S400 includes steps S410 to S430.

[0064] Step S410: Perform feature weight allocation processing based on the feature vector set. Differentiated weights are assigned to each feature through a weight allocation mechanism based on the historical detection accuracy of the sensor, resulting in a weighted feature set.

[0065] Step S420: Perform probability fusion based on the weighted feature set, update the sleeper state confidence based on the combination of prior probability and real-time data likelihood, and obtain a preliminary state score;

[0066] Step S430: Based on the preliminary status score, a comprehensive evaluation process is carried out. Through multi-dimensional score fusion and threshold comparison mechanism, the comprehensive status score of the sleeper is obtained.

[0067] Specifically, in the data fusion process from steps S410 to S430, multi-source data is integrated based on the feature vector set. First, step S410 performs feature weight allocation processing, assigning differentiated weights based on the historical detection accuracy of sensors (e.g., the weighted average fusion principle, dynamically adjusting weights based on the historical accuracy of infrared, dynamic mechanics, and other sensors) to obtain a weighted feature set. Then, step S420 performs probabilistic fusion, updating the sleeper state confidence based on prior probabilities (e.g., historical statistical probabilities of sleeper abnormal states) and real-time data likelihood (e.g., the matching degree between current features and abnormal conditions) (reflecting the Bayesian inference fusion principle) to obtain a preliminary state score. Finally, step S430 performs comprehensive evaluation processing, integrating multi-dimensional scores (e.g., geometric, dynamic, and thermal feature scores) and performing threshold comparisons (e.g., the voting mechanism in decision-level fusion) to generate a comprehensive sleeper state score. This process comprehensively utilizes weighted averaging, Bayesian inference, and decision-level fusion methods, progressively refining from the feature level to the decision level to ensure the robustness and accuracy of the state assessment.

[0068] Further, step S500 includes steps S510 to S530.

[0069] Step S510: Establish a multi-dimensional benchmark pattern library based on historical normal data. By analyzing the stable correlation patterns of sleeper temperature distribution, vibration characteristics, acoustic features and train operation status under normal conditions, a benchmark pattern library containing spatiotemporal features is constructed.

[0070] Step S520: Perform multi-source feature collaborative analysis based on the comprehensive state score, extract the current multi-dimensional feature sequence through a time-series sliding window, and perform feature layer similarity matching with the benchmark pattern library to obtain the multi-dimensional deviation evaluation result.

[0071] Step S530: Based on the multi-dimensional deviation assessment results, perform comprehensive anomaly judgment, identify specific anomaly types and their spatial distribution through a dynamic threshold comparison mechanism combined with an anomaly type feature library, and obtain anomaly identification results.

[0072] In the anomaly detection process, step S510 first establishes a multi-dimensional benchmark pattern library based on historical normal data. By analyzing the stable correlation patterns between sleeper temperature distribution characteristics, vibration characteristic parameters, acoustic characteristic indicators, and train operation status data under normal conditions, a benchmark pattern library containing spatiotemporal features is constructed to provide a reliable reference benchmark for anomaly identification. Subsequently, step S520 performs multi-source feature collaborative analysis based on the comprehensive state score, extracts the current multi-dimensional feature sequence using a time-series sliding window, and performs feature layer similarity matching between these real-time data and the benchmark pattern library to obtain multi-dimensional deviation evaluation results, accurately quantifying the anomaly tendency of sleeper status. Finally, step S530 performs comprehensive anomaly judgment based on the multi-dimensional deviation evaluation results. Through a dynamic threshold comparison mechanism combined with predefined anomaly feature patterns such as loosening and settlement in the anomaly type feature library, specific anomaly types and their spatial distribution are identified, ultimately obtaining anomaly identification results and completing the closed-loop detection process from data to decision.

[0073] Further, step S600 includes steps S610 to S630.

[0074] Step S610: Establish a trend prediction model based on the anomaly identification results and historical state data. Construct a sleeper state change trend model by analyzing the temporal evolution law of anomaly characteristics, and obtain the state evolution prediction function.

[0075] Step S620: Perform predictive analysis based on the state evolution prediction function, and predict the trajectory of sleeper state parameter changes within a specific future period through a sliding time window mechanism to obtain the state change prediction result;

[0076] Step S630: Generate maintenance strategies based on the state change prediction results. By matching historical maintenance records with predicted state thresholds, generate specific maintenance measures suggestions to obtain maintenance recommendations.

[0077] Specifically, step S610 first establishes a trend prediction model based on the anomaly identification results and historical state data. By deeply analyzing the temporal evolution of anomaly characteristics, a mathematical model capable of accurately describing the changing trend of sleeper state is constructed, thus obtaining a state evolution prediction function with predictive capabilities. Subsequently, step S620 performs multi-step predictive analysis based on this state evolution prediction function, using a sliding time window mechanism to dynamically predict the trajectory of sleeper state parameter changes within a specific future period, obtaining a state change prediction result with temporal continuity. Finally, step S630 intelligently generates maintenance strategies based on the state change prediction results. By associating effective maintenance schemes from the historical maintenance record database and accurately matching them with the predicted state threshold, targeted and specific maintenance measures are automatically generated, ultimately outputting maintenance suggestions that can guide on-site operations, achieving a complete closed loop from predictive analysis to maintenance decision-making.

[0078] Example 2:

[0079] This embodiment is a preferred implementation, specifically demonstrating the application scenario of multi-source data fusion in the detection of railway sleeper loosening anomalies. By integrating data from three types of sensors—infrared thermometry, dynamic mechanical sensors, and acoustic sensors—it intelligently determines whether a particular sleeper has exhibited loosening anomalies. In this embodiment, the infrared thermometry sensor detects an abnormal increase in the sleeper surface temperature, possibly due to increased friction, and outputs a confidence level for the temperature anomaly. The dynamic mechanical sensor detected a decrease in the dominant vibration frequency and an abnormal amplitude, and output the vibration anomaly confidence level. The acoustic sensor detected an increase in impact sound events and output an acoustic anomaly confidence level. These data provide the foundation for subsequent fusion analysis.

[0080] In the application of the weighted average fusion method, weights are dynamically assigned based on the detection accuracy of various sensors for loosening anomalies in historical data, among which infrared weights are... (Because temperature has a weak indication of loosening), dynamic mechanical weighting (Because vibration is sensitive to loosening), acoustic weighting (Acoustic events are susceptible to environmental interference). Weighted average fusion results. The calculation formula is:

[0081] ;

[0082] Substituting the numerical values, we get The threshold was set at 0.70, because... If so, it is determined that the sleeper is abnormally loose.

[0083] In the application of Bayesian inference fusion methods, it is assumed that prior probabilities are based on historical statistics. The likelihoods of each sensor under the condition of sleeper loosening are as follows: , , By simplifying the modeling, the posterior probability The calculation formula is:

[0084] ;

[0085] Substituting the numerical values ​​into the posterior probability calculation, which is significantly higher than the threshold, also supports the loosening determination.

[0086] During the decision output and maintenance recommendation phase, the fusion results trigger a system alarm, and the sleeper is highlighted in red on the management platform. By combining the decreasing trend of the dominant vibration frequency and the crack propagation rate in the feature matrix, it is predicted that the loosening may develop into structural damage within 30 days. Therefore, maintenance recommendations are generated: tightening work should be carried out within 14 days, and vibration parameters should be closely monitored in the following 3 inspection cycles. This embodiment demonstrates the integrated application of multi-source data fusion analysis in a multi-dimensional coordinate system (including spatial dimensions X, Y, Z, time dimension, and feature dimension), achieving a comprehensive assessment of the track condition by forming a multi-dimensional fused data matrix.

[0087] Example 3:

[0088] This embodiment demonstrates the application of multidimensional fused data matrix in railway sleeper condition evolution analysis and predictive maintenance decision-making. Continuous condition monitoring of a certain section (a total of 100 sleepers) is required. The goal is to identify potential deterioration trends and formulate maintenance plans in advance by constructing and analyzing the multidimensional fused data matrix, thereby achieving closed-loop management from data acquisition to decision support.

[0089] The multidimensional fusion data matrix, serving as the output carrier of the aforementioned feature extraction and data fusion steps, structurally integrates the feature vector set and comprehensive status score in the spatiotemporal dimensions, providing a unified multidimensional data analysis foundation for subsequent anomaly detection and predictive analysis. The construction of the multidimensional fusion data matrix encompasses five key dimensions, including the spatial dimension (X, Y, Z) defining the precise track mileage of each sleeper. Horizontal position With elevation Time dimension ( ) Record detection timestamps at 5-minute intervals, and feature dimensions ( The integrated and fused set of state features, such as health scores (0–100 points), Temperature Anomaly Index Vibration intensity Acoustic anomaly count Cumulative displacement .lattice The mathematical representation of each data point is defined as a six-tuple:

[0090] ;

[0091] Where the eigenvector for:

[0092] ;

[0093] Full dot matrix This forms a unified framework that covers the multi-dimensional states of railway sleepers.

[0094] The real-time update mechanism of the dot matrix ensures the timeliness and accuracy of data, including the time-series sliding update, where a new time slice is added every 5 minutes. The feature vector is updated according to the latest sensor data fusion result. The spatial coordinate is updated by image matching when the sleeper is displaced due to maintenance. The data archiving strategy retains the dot matrix data of the last 30 days in high-performance memory, and the earlier data is compressed and stored in the historical database to support long-term trend analysis.

[0095] The specific application of the dot matrix in analysis and decision-making is as follows: First, through deterioration trend clustering analysis, all sleepers are clustered in the feature-time subspace. It is found that 5 sleepers (numbered #34, #35, #41, #42, #47) show an average daily decrease of 1.2 points, a 15% increase, and a cumulative amount exceeding the threshold in 10 consecutive time slices. The clustering result is marked as a potential settlement group. Subsequently, spatio-temporal association pattern mining is carried out and analyzed in the three-dimensional subspace of space-time-vibration intensity. It is found that the sleepers in this group are continuously distributed in space, and the vibration intensity shows a propagation trend along the line direction over time (increasing from #34 to #47), indicating that there may be stress transfer caused by local ballast softening. Based on the time-series analysis of the dot matrix, the system calls the built-in prediction model, inputs the feature sequence of the last 30 time slices of this group, and outputs a prediction that within the next 7 days it will drop below 70 points (alarm threshold), and generates decision suggestions such as jointly detecting and tamping the sleepers in this group and the underlying ballast within 3 days. In the visualization and intervention feedback link, the sleepers in the group are displayed in the three-dimensional view of the management platform with a gradient color - flashing warning, and an Automatic Local Settlement Early Warning Report is automatically generated with a trend curve and a spatial heat map. After the maintenance is completed, the railway maintenance staff mark the processing results through the platform, and accordingly update the dot matrix feature baseline to form a decision-making closed loop. This embodiment demonstrates the comprehensive application of the multi-dimensional data dot matrix in anomaly detection, status assessment, and predictive maintenance, ensuring the safe and reliable operation of railway transportation.

[0096] Example 4:

[0097] This embodiment specifically shows how to automatically generate specific executable maintenance suggestions based on the detected deterioration trend of sleepers, and fully presents the whole process from data input to the output of maintenance suggestions. The system detects the following fusion analysis results (from the multi-dimensional data dot matrix) for sleepers numbered #201 to #205 within 7 consecutive days: the health score decreases from 92 points to 76 points; the cumulative lateral displacement increases by 8 mm; the vibration intensity increases by 22%; the acoustic anomaly count More than 5 times a day; infrared images showed that the local temperature was 4°C higher than the surrounding area. Based on this data, the system determined that there was a combined anomaly in this section: "loose sleepers accompanied by local track bed settlement".

[0098] The maintenance suggestion generation model is built based on historical maintenance records. The system has accumulated all maintenance work orders and their corresponding sleeper inspection data from the past three years, forming a structured record that includes anomaly type (loosening, cracks, settlement, wear, etc.), detection feature values ​​(feature sequences before the anomaly occurred), measures taken ("tightening bolts", "replacing sleepers", "compacting the track bed", etc.), and maintenance effects (the degree of recovery of health score within 30 days after maintenance). The system is trained using a multi-label classification model (RandomForest + One-vs-Rest), with the input feature sequence ( , , , , The system outputs suggested measures labels (multiple possibilities), with the training objective being to maximize positive matching of historical maintenance effects. Simultaneously, the system establishes a maintenance strategy rule base, for example, using the Apriori algorithm to mine rules from historical work orders: IF Cumulative Displacement > 5mm AND Vibration Intensity Increase > 20% THEN Suggested Measures IN {"Supplementary Compaction of Track Bed", "Tightening of Sleeper Bolts"}.

[0099] In the suggestion generation and natural language output stage, based on the current detection data, the model outputs the following through inference: the primary suggestion is to further compact the track bed (confidence level 87%); the secondary suggestion is to retighten the sleeper bolts (confidence level 72%); and the supplementary suggestion is to re-measure the vibration parameters within 7 days. The Natural Language Generation (NLG) module dynamically generates readable suggestions based on preset templates, automatically filling in key fields to form a complete maintenance suggestion report. The report includes the anomaly type (loose sleepers accompanied by local track bed settlement), detection basis (the cumulative increase in lateral displacement by 8mm in the past 7 days, the increase in vibration intensity by 22%, frequent acoustic anomalies, and local temperature differences), specific recommended measures (compact the track bed in this section within 3 days, with a compaction width of not less than 0.5m on each side of the sleeper; simultaneously check and tighten the sleeper bolts, with a torque standard of not less than 300 N·m; re-measure the vibration intensity and displacement parameters through the on-board detection system within 24 hours after the operation is completed), expected results (based on similar historical cases, the above measures can improve the health score to above 85 points within 7 days), and risk warnings (if the treatment is delayed, it may develop into structural damage to the sleepers within 30 days).

[0100] The suggestion is automatically inserted into the "Maintenance Suggestion" section of the management platform, and a PDF version of the "Sleeper Maintenance Work Order" is generated, complete with trend curves and work instruction diagrams. It also supports voice synthesis for on-site inspection terminals. Through a decision-making closed-loop and model optimization mechanism, after maintenance personnel implement the suggestion, the platform provides feedback on actual maintenance measures and post-maintenance inspection data. This maintenance record is added to the historical database as a new sample. The suggestion generation model is periodically retrained, and a reinforcement learning mechanism is used to score the effectiveness of the suggestions, dynamically adjusting model weights to achieve continuous system optimization.

[0101] Example 5:

[0102] like Figure 2 As shown, this embodiment provides a multi-source data fusion-based intelligent railway sleeper detection system, the system comprising:

[0103] The acquisition module 901 is used to acquire multi-source raw data of railway tracks, including infrared temperature measurement data, dynamic mechanical data, acoustic spectrum data, image data, and train operation status data.

[0104] The processing module 902 is used to perform data preprocessing based on multi-source raw data, eliminate inherent sensor noise and environmental interference through denoising operations, and combine normalization processing to uniformly scale data of different dimensions to the same numerical range to obtain standardized data.

[0105] The extraction module 903 is used to extract features based on standardized data. By extracting the geometric, dynamic and thermal features of the sleeper arrangement from multi-source data, a set of feature vectors for the sleeper arrangement is obtained.

[0106] The fusion module 904 is used to perform data fusion based on the feature vector set. Through a comprehensive fusion process that integrates multi-source features and combines prior information, a comprehensive status score of the sleeper is obtained.

[0107] The detection module 905 is used to detect anomalies based on the comprehensive status score. It compares real-time data with historical normal patterns through pattern recognition and identifies abnormal sleeper arrangement based on threshold comparison to obtain anomaly identification results.

[0108] The output module 906 is used to perform predictive analysis based on the anomaly identification results, predict future changes in sleeper condition by analyzing historical data trends through time series modeling, and generate maintenance suggestions based on the prediction results.

[0109] In one specific embodiment of this application, the processing module 902 includes:

[0110] The first processing unit is used to perform noise elimination processing on multi-source raw data. By applying a sensor-specific noise model, it eliminates track vibration interference and thermal noise unique to sleeper detection and obtains preliminary clean data.

[0111] The second processing unit is used to extract effective data based on the preliminary cleaning data, and remove invalid data points by screening criteria based on the spatial distribution pattern of sleepers to obtain the filtered data.

[0112] The third processing unit is used to perform dimension unification processing on the filtered data, and convert multi-source data into dimensionless form based on the data distribution characteristics to obtain standardized data.

[0113] In one specific embodiment of this application, the extraction module 903 includes:

[0114] The first extraction unit is used to perform multi-source feature extraction based on standardized data. It extracts sleeper surface temperature distribution features from infrared thermometry data, vibration frequency features from dynamic mechanical data, sleeper spacing and angle features from image data, and integrates acoustic spectrum data and train status data features to obtain a preliminary feature pool.

[0115] The second extraction unit is used to perform feature filtering based on the preliminary feature pool, analyze the importance of historical data and verify the physical significance of sleepers, and remove redundant features to obtain a set of key features.

[0116] The third extraction unit is used to perform matrix construction processing based on the key feature set. By organizing the feature vectors according to the sleeper number and introducing a spatiotemporal expansion mechanism, the feature vector set of sleeper arrangement is obtained.

[0117] In one specific embodiment of this application, the fusion module 904 includes:

[0118] The first fusion unit is used to perform feature weight allocation processing based on the feature vector set. It assigns differentiated weights to each feature through a weight allocation mechanism based on the historical detection accuracy of the sensor, thereby obtaining a weighted feature set.

[0119] The second fusion unit is used to perform probability fusion based on the weighted feature set, and update the sleeper state confidence based on the combination of prior probability and real-time data likelihood to obtain a preliminary state score.

[0120] The third fusion unit is used to perform comprehensive evaluation based on the preliminary status score. Through multi-dimensional score fusion and threshold comparison mechanism, the comprehensive status score of the sleeper is obtained.

[0121] In one specific embodiment of this application, the detection module 905 includes:

[0122] The first detection unit is used to establish a multi-dimensional benchmark pattern library based on historical normal data. By analyzing the stable correlation patterns of sleeper temperature distribution, vibration characteristics, acoustic features and train operation status under normal conditions, a benchmark pattern library containing spatiotemporal features is constructed.

[0123] The second detection unit is used to perform multi-source feature collaborative analysis based on the comprehensive state score. It extracts the current multi-dimensional feature sequence through a time-series sliding window and performs feature layer similarity matching with the benchmark pattern library to obtain the multi-dimensional deviation evaluation result.

[0124] The third detection unit is used to make a comprehensive judgment on anomalies based on the multi-dimensional deviation evaluation results. By combining a dynamic threshold comparison mechanism with an anomaly type feature library, it identifies specific anomaly types and their spatial distribution, and obtains anomaly identification results.

[0125] In one specific embodiment of this application, the output module 906 includes:

[0126] The first output unit is used to establish a trend prediction model based on the anomaly identification results and historical state data. It constructs a sleeper state change trend model by analyzing the temporal evolution law of anomaly characteristics and obtains the state evolution prediction function.

[0127] The second output unit is used to perform predictive analysis based on the state evolution prediction function, and predict the trajectory of sleeper state parameter changes within a specific future period through a sliding time window mechanism to obtain the state change prediction result.

[0128] The third output unit is used to generate maintenance strategies based on the state change prediction results. By matching historical maintenance records with the predicted state thresholds, specific maintenance measures are suggested, and maintenance recommendations are obtained.

[0129] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent detection of railway sleepers using multi-source data fusion, characterized in that, include: Acquire multi-source raw data of railway tracks, including infrared thermometry data, dynamic mechanical data, acoustic spectrum data, image data, and train operation status data; Data preprocessing is performed based on the multi-source raw data. Noise reduction is performed to eliminate inherent sensor noise and environmental interference. Combined with normalization, data of different dimensions are scaled to the same numerical range to obtain standardized data. Feature extraction is performed based on the standardized data. By extracting the geometric, dynamic, and thermal features of the sleeper arrangement from multi-source data, a feature vector set of the sleeper arrangement is obtained. Data fusion is performed based on the set of feature vectors. Through a comprehensive fusion process that integrates multi-source features and combines prior information, a comprehensive status score for the sleeper is obtained. Anomaly detection is performed based on the comprehensive status score. Real-time data is compared with historical normal patterns through pattern recognition, and abnormal sleeper arrangement is identified based on threshold comparison to obtain anomaly identification results. Based on the anomaly identification results, predictive analysis is performed, and historical data trends are analyzed through time series modeling to predict future changes in sleeper condition. Maintenance recommendations are then generated based on the prediction results.

2. The intelligent detection method for railway sleepers based on multi-source data fusion according to claim 1, characterized in that, Data preprocessing is performed on the multi-source raw data. Noise reduction is used to eliminate inherent sensor noise and environmental interference. Normalization is then applied to scale data of different dimensions to the same numerical range, resulting in standardized data, including: Noise removal processing is performed on the multi-source raw data. By applying a sensor-specific noise model, the track vibration interference and thermal noise unique to sleeper detection are eliminated to obtain preliminary clean data. Based on the preliminary cleaning data, effective data extraction and processing are performed, and invalid data points are removed by screening criteria based on the spatial distribution pattern of sleepers to obtain the filtered data; The filtered data is then subjected to dimensionless processing, and the multi-source data is converted into a dimensionless form based on the data distribution characteristics to obtain standardized data.

3. The intelligent detection method for railway sleepers based on multi-source data fusion according to claim 1, characterized in that, Feature extraction is performed based on the standardized data. By extracting the geometric, dynamic, and thermal features of the sleeper arrangement from multi-source data, a feature vector set of sleeper arrangement is obtained, including: Multi-source feature extraction is performed based on the standardized data. By extracting sleeper surface temperature distribution features from infrared thermometry data, vibration frequency features from dynamic mechanical data, sleeper spacing and angle features from image data, and integrating acoustic spectrum data and train status data features, a preliminary feature pool is obtained. Feature filtering is performed based on the preliminary feature pool, and the importance of historical data and the physical significance of sleepers are verified. Redundant features are then removed to obtain a set of key features. The matrix is ​​constructed based on the set of key features. By organizing the feature vectors according to the sleeper number and introducing a spatiotemporal expansion mechanism, the set of feature vectors for sleeper arrangement is obtained.

4. The intelligent detection method for railway sleepers based on multi-source data fusion according to claim 1, characterized in that, Data fusion is performed based on the aforementioned feature vector set. Through a comprehensive fusion process combining multi-source feature integration and prior information, a comprehensive sleeper status score is obtained, including: Based on the feature vector set, feature weight allocation processing is performed. Differentiated weights are assigned to each feature through a weight allocation mechanism based on the historical detection accuracy of the sensor, resulting in a weighted feature set. Probability fusion is performed based on the weighted feature set, and the sleeper state confidence is updated based on the combination of prior probability and real-time data likelihood to obtain a preliminary state score. Based on the preliminary status score, a comprehensive evaluation process is performed, and a comprehensive status score for the sleeper is obtained through multi-dimensional score fusion and threshold comparison mechanism.

5. The intelligent detection method for railway sleepers based on multi-source data fusion according to claim 1, characterized in that, Anomaly detection is performed based on the comprehensive status score. Real-time data is compared with historical normal patterns using pattern recognition, and sleeper arrangement anomalies are identified based on threshold comparison, resulting in anomaly identification results, including: A multi-dimensional benchmark pattern library is established based on historical normal data. By analyzing the stable correlation patterns of sleeper temperature distribution, vibration characteristics, acoustic features and train operation status under normal conditions, a benchmark pattern library containing spatiotemporal features is constructed. Based on the comprehensive state score, multi-source feature collaborative analysis is performed. The current multi-dimensional feature sequence is extracted through a time-series sliding window and matched with the benchmark pattern library to obtain the multi-dimensional deviation evaluation result. Based on the multi-dimensional deviation evaluation results, anomalies are comprehensively determined. By combining a dynamic threshold comparison mechanism with an anomaly type feature library, specific anomaly types and their spatial distributions are identified, resulting in anomaly identification results.

6. A multi-source data fusion intelligent detection system for railway sleepers, characterized in that, include: The acquisition module is used to acquire multi-source raw data of railway tracks, including infrared temperature measurement data, dynamic mechanical data, acoustic spectrum data, image data, and train operation status data. The processing module is used to perform data preprocessing based on the multi-source raw data, eliminate inherent noise of the sensor and environmental interference through denoising operation, and combine normalization processing to uniformly scale data of different dimensions to the same numerical range to obtain standardized data. The extraction module is used to extract features based on the standardized data. By extracting the geometric, dynamic and thermal features of the sleeper arrangement from the multi-source data, a feature vector set of the sleeper arrangement is obtained. The fusion module is used to perform data fusion based on the feature vector set. Through a comprehensive fusion process that integrates multi-source features and combines prior information, a comprehensive status score of the sleeper is obtained. The detection module is used to detect anomalies based on the comprehensive status score, compare real-time data with historical normal patterns through pattern recognition, and identify abnormal sleeper arrangement based on threshold comparison to obtain anomaly identification results. The output module is used to perform predictive analysis based on the anomaly identification results, predict future changes in sleeper condition by analyzing historical data trends through time series modeling, and generate maintenance suggestions based on the prediction results.

7. The intelligent railway sleeper detection system based on multi-source data fusion according to claim 6, characterized in that, The processing module includes: The first processing unit is used to perform noise elimination processing on the multi-source raw data, and eliminate the track vibration interference and thermal noise unique to sleeper detection by applying a sensor-specific noise model to obtain preliminary clean data. The second processing unit is used to perform effective data extraction processing based on the preliminary cleaning data, and to remove invalid data points by screening criteria based on the spatial distribution pattern of sleepers to obtain the filtered data. The third processing unit is used to perform dimension unification processing on the filtered data, and convert multi-source data into dimensionless form based on the data distribution characteristics to obtain standardized data.

8. The intelligent railway sleeper detection system based on multi-source data fusion according to claim 6, characterized in that, The extraction module includes: The first extraction unit is used to perform multi-source feature extraction based on the standardized data. It extracts the surface temperature distribution features of the sleeper from infrared thermometry data, the vibration frequency features from dynamic mechanical data, and the sleeper spacing and angle features from image data. It also integrates the acoustic spectrum data and train status data features to obtain a preliminary feature pool. The second extraction unit is used to perform feature filtering based on the preliminary feature pool, analyze the importance of historical data and verify the physical meaning of sleepers, and remove redundant features to obtain a set of key features. The third extraction unit is used to perform matrix construction processing based on the key feature set. By organizing the feature vectors according to the sleeper number and introducing a spatiotemporal expansion mechanism, a feature vector set of sleeper arrangement is obtained.

9. The intelligent railway sleeper detection system based on multi-source data fusion according to claim 6, characterized in that, The fusion module includes: The first fusion unit is used to perform feature weight allocation processing based on the feature vector set, and to assign differentiated weights to each feature through a weight allocation mechanism based on the historical detection accuracy of the sensor, so as to obtain a weighted feature set; The second fusion unit is used to perform probability fusion based on the weighted feature set, update the sleeper state confidence based on the combination of prior probability and real-time data likelihood, and obtain a preliminary state score. The third fusion unit is used to perform comprehensive evaluation processing based on the preliminary status score, and obtain the comprehensive status score of the sleeper through multi-dimensional score fusion and threshold comparison mechanism.

10. The intelligent railway sleeper detection system based on multi-source data fusion according to claim 6, characterized in that, The detection module includes: The first detection unit is used to establish a multi-dimensional benchmark pattern library based on historical normal data. By analyzing the stable correlation patterns of sleeper temperature distribution, vibration characteristics, acoustic features and train operation status under normal conditions, a benchmark pattern library containing spatiotemporal features is constructed. The second detection unit is used to perform multi-source feature collaborative analysis based on the comprehensive state score, extract the current multi-dimensional feature sequence through a time-series sliding window, and perform feature layer similarity matching with the benchmark pattern library to obtain the multi-dimensional deviation evaluation result. The third detection unit is used to make a comprehensive judgment on anomalies based on the multi-dimensional deviation evaluation results. By combining the dynamic threshold comparison mechanism with the anomaly type feature library, it identifies specific anomaly types and their spatial distribution to obtain anomaly identification results.

Images