CN-122020315-A - Rail online fault diagnosis method and system based on knowledge transfer learning

CN122020315ACN 122020315 ACN122020315 ACN 122020315ACN-122020315-A

Abstract

The invention provides a rail online fault diagnosis method and a rail online fault diagnosis system based on knowledge transfer learning in the technical field of intelligent monitoring and artificial intelligence crossing of railway infrastructure, wherein the method comprises the following steps that S1, a server acquires a teacher diagnosis model by utilizing a source domain data set training model frame; the method comprises the steps of obtaining a student diagnosis model by training a teacher diagnosis model by adopting a mixed training strategy based on a target domain data set, collecting real-time multi-mode monitoring data from a target domain line by an edge computing device, compensating to obtain correction data, dynamically selecting the most important feature subset from the correction data to form a simplified feature set, inputting the simplified feature set into the student diagnosis model to obtain a student fault diagnosis result and confidence level thereof, and introducing a D-S evidence theory to generate a student fault diagnosis report. The method has the advantages that the accuracy, the instantaneity, the cross-domain adaptability and the overall safety of the system of the railway track fault diagnosis are greatly improved.

Inventors

ZHANG PEI
LI ZHENGHUI
ZHANG NA
WANG XIYAN
ZHU WENYU
ZOU YANG

Assignees

郑州铁路职业技术学院

Dates

Publication Date: 20260512
Application Date: 20260224

Claims (10)

1. A rail online fault diagnosis method based on knowledge transfer learning is characterized by comprising the following steps: step S1, acquiring source domain multi-mode monitoring data comprising rail geometric data, rail images, rail loads, vibration acceleration, vibration noise and environmental temperature from a plurality of source domain lines of a rail, meanwhile, acquiring target domain multi-mode monitoring data with the same mode from a target domain line, encrypting the source domain multi-mode monitoring data and the target domain multi-mode monitoring data into an encryption historical data packet, and uploading the encryption historical data packet to a server; s2, after decrypting and checking the received encryption historical data packet, the server obtains the source domain multi-mode monitoring data and the target domain multi-mode monitoring data and stores the source domain multi-mode monitoring data and the target domain multi-mode monitoring data in a security database; S3, the server carries out preprocessing, fault labeling and data enhancement on the source domain multi-mode monitoring data, builds a source domain data set, and builds a target domain data set based on each target domain multi-mode monitoring data; Step S4, a model framework based on a lightweight confidence rule base is built by a server, a teacher diagnosis model is obtained by training the source domain data set, and meanwhile, a lightweight student diagnosis model is obtained by training the teacher diagnosis model by adopting a mixed training strategy combining knowledge distillation and domain countermeasure self-adaptive transfer learning based on the target domain data set; s5, the edge computing equipment collects real-time multi-mode monitoring data from the target domain line, compensates the real-time monitoring data to obtain correction data, and dynamically selects the most important feature subset from the correction data according to the current service capability index of the edge computing equipment and the global feature priori list issued by the server to form a simplified feature set; S6, inputting the simplified feature set into a deployed student diagnosis model by the edge computing equipment to obtain a student fault diagnosis result and the confidence coefficient thereof, introducing a D-S evidence theory, and fusing the confidence coefficient, the historical diagnosis consistency and the vision auxiliary verification information from the image sensor to generate a student fault diagnosis report carrying a reliability level; Step S7, the edge computing equipment judges whether the reliability level is higher than a preset level threshold, if so, the student fault diagnosis report is digitally signed and directly sent to an operation and maintenance center through a secure channel, otherwise, the original real-time multi-mode monitoring data is uploaded to a server through the secure channel, the teacher diagnosis model is requested to carry out secure reasoning in a TEE trusted execution environment, a teacher fault diagnosis report is generated, and the teacher fault diagnosis report is sent to the operation and maintenance center through the secure channel after being digitally signed; And S8, the operation and maintenance center marks the student fault diagnosis report or the teacher diagnosis report based on the input field verification result to generate an increment sample set, the increment sample set is sent to the server through a safety channel, and the server performs increment learning and model updating by using the increment sample set.
2. The method for diagnosing the rail on-line fault based on the knowledge transfer learning of claim 1, wherein the step S1 is specifically as follows: Acquiring source domain multi-mode monitoring data comprising rail geometric data, rail images, rail loads, vibration accelerations, vibration noise signals and ambient temperatures from a plurality of source domain routes of a rail, wherein the rail geometric data at least comprises rail height, rail direction, rail gauge and levelness; Meanwhile, target domain multi-mode monitoring data with the same mode are collected from the target domain line; calculating unique hash fingerprints for the source domain multi-mode monitoring data and the target domain multi-mode monitoring data of each data batch by using an SHA-256 algorithm, binding the hash fingerprints and a data acquisition timestamp, broadcasting and recording the hash fingerprints and the data acquisition timestamp on a blockchain in a transaction mode, and finishing the certification; Encrypting source domain multi-mode monitoring data and target domain multi-mode monitoring data of each data batch into an encryption historical data packet by adopting an end-to-end encryption technology based on a national encryption algorithm, and uploading each encryption historical data packet to a server through a secure channel; The end-to-end encryption technology based on the cryptographic algorithm specifically comprises the following steps: Converting the source domain multi-mode monitoring data and the target domain multi-mode monitoring data into binary data blocks, distributing a unique batch ID for the source domain multi-mode monitoring data and the target domain multi-mode monitoring data of each data batch, and acquiring data mode type descriptions of the source domain multi-mode monitoring data and the target domain multi-mode monitoring data based on data acquisition time stamps of the source domain multi-mode monitoring data and the target domain multi-mode monitoring data; Constructing a data enhancement block based on the binary data block and metadata, wherein the metadata comprises a batch ID, a time range and a data modality type description; Negotiating with a server through an SM2 algorithm to generate a master key, taking the master key and a batch ID as input, and generating a batch subkey through an SM3 algorithm, carrying out hash calculation on the batch ID through the SM3 algorithm to obtain an initial vector; the batch subkey is called through an SM4 algorithm, and the data enhancement block is encrypted based on the initial vector and a CBC mode to obtain an encrypted data block; Dividing the encrypted data block into a plurality of data sub-blocks by taking 128 bytes as a unit, applying cyclic left shift operation to the data sub-blocks based on the initial vector to obtain confusion sub-blocks, and splicing the confusion sub-blocks to obtain a confusion data block; and packaging the confusion data block and the metadata into encryption historical data.
3. The method for diagnosing the rail on-line fault based on the knowledge transfer learning of claim 1, wherein the step S2 is specifically as follows: the server receives the encryption historical data packet in real time, analyzes the encryption historical data packet to obtain an confusion data block and metadata comprising a batch ID, a time range and a data mode type description; Taking the negotiated master key and the batch ID as input, and generating batch subkeys by using an SM3 algorithm, carrying out hash calculation on the batch ID by using the SM3 algorithm to obtain an initial vector; dividing the confusion data block into a plurality of confusion sub-blocks by taking 128 bytes as a unit, performing cyclic right shift operation on the confusion sub-blocks based on the initial vector to obtain data sub-blocks, and splicing the data sub-blocks to obtain an encrypted data block; Invoking the batch subkey through an SM4 algorithm, decrypting an encrypted data block based on the initial vector and a CBC mode to obtain a data enhancement block, analyzing the data enhancement block to obtain a binary data block and metadata, performing consistency check on the metadata obtained by two analyses, performing inverse conversion on the binary data block to obtain source domain multi-mode monitoring data and target domain multi-mode monitoring data, and performing integrity check on the source domain multi-mode monitoring data and the target domain multi-mode monitoring data through hash fingerprints of a blockchain memory card; And storing the source domain multi-mode monitoring data and the target domain multi-mode monitoring data which pass the verification in a security database deployed in a TEE trusted execution environment.
4. The method for diagnosing the rail on-line fault based on the knowledge transfer learning of claim 1, wherein the step S3 is specifically as follows: The server performs preprocessing at least comprising data cleaning and outlier processing, data alignment and fusion, signal noise reduction and filtering, and data normalization and standardization on the source domain multi-mode monitoring data, performs labeling at least comprising a fault type label, a fault severity grade label, a fault position label, data quality and a confidence coefficient label on the preprocessed source domain multi-mode monitoring data, and performs data enhancement on a small sample with severe faults aiming at the fault severity grade label by adopting a condition constraint countermeasure generation network to construct a source domain data set; based on the source domain data set, carrying out multi-round feature screening by utilizing average influence degree analysis and maximum correlation minimum redundancy criterion, calculating and generating a global feature priori list arranged in descending order of feature importance, and transmitting the global feature priori list to edge computing equipment; The condition constraint countermeasure generation network specifically comprises the steps of introducing constraint conditions based on a physical mechanism of track faults into a loss function of a generator of the countermeasure generation network, limiting vibration frequency bands and amplitude ranges of generated samples to be consistent with known fault mechanisms, and ensuring effectiveness and safety of the generated samples.
5. The method for diagnosing the rail online faults based on knowledge transfer learning of claim 1, wherein in the step S4, the construction method of the lightweight confidence rule base is characterized in that a rule front part merging and confidence degree distribution sparsification technology is adopted, redundant confidence rules with low contribution degree are removed, K neighbor search strategies are introduced in a confidence rule activation stage, and only the confidence rule most relevant to the current input is calculated; the mixed training strategy is specifically that soft labels and middle layer features output by the teacher diagnosis model are distilled, and a gradient inversion layer is introduced to minimize feature distribution differences of a source domain and a target domain.
6. The method for on-line fault diagnosis of rail based on knowledge transfer learning of claim 1, wherein the step S5 is specifically: the edge computing equipment collects real-time multi-mode monitoring data from a target domain line of a rail, carries out zero drift and sensitivity drift estimation on the real-time multi-mode monitoring data through a Kalman filter to obtain sensor drift amount, and carries out real-time compensation on the real-time multi-mode monitoring data based on the sensor drift amount to obtain correction data; according to the service capability index which comprises network bandwidth, calculation resource allowance and diagnosis response time limit requirement, dynamically selecting the most important feature subset from the correction data based on the global feature priori list issued by the server to form a simplified feature set; and when the sensor drift amount exceeds a preset drift threshold value, selecting the feature least sensitive to drift from the correction data as a feature subset based on the global feature priori list.
7. The method for on-line fault diagnosis of rail based on knowledge transfer learning of claim 1, wherein the step S6 is specifically: the edge computing equipment inputs the simplified feature set into a deployed student diagnosis model to obtain a preliminary student fault diagnosis result and a corresponding confidence coefficient thereof; then, constructing a multisource information fusion framework based on a D-S evidence theory, and specifically executing the following steps: A. defining a set formed by all possible results of fault diagnosis as an identification frame; B. constructing a basic probability distribution function, namely constructing the basic probability distribution function for the following three independent evidence sources respectively: b1, model output evidence, namely constructing basic probability distribution about various faults in the identification frame based on the confidence coefficient output by the student diagnosis model; based on a preset time window, searching a locally stored historical diagnosis record, calculating the consistency degree of the current student fault diagnosis result and the historical fault diagnosis result on the fault type and the fault severity level, and quantifying the consistency degree into basic probability distribution, wherein the consistency degree is obtained by calculating sequence similarity or counting the continuous occurrence times of the same fault conclusion; Triggering and receiving a rail image from an image sensor, performing visual analysis on the rail image through a light-weight image analysis model, converting an analysis result into a support degree for specific fault propositions, and constructing corresponding basic probability distribution; C. The evidence fusion and decision-making are carried out by applying a synthesis rule in a D-S evidence theory to fuse three basic probability distributions to obtain a joint distribution function, calculating a credibility function and a plausibility function aiming at each fault proposition according to the joint distribution function, generating a final student fault diagnosis result according to a preset decision rule, and calculating the credibility level of the student fault diagnosis result; D. And generating a report, namely integrating the final student fault diagnosis result and the reliability level thereof, and generating a student fault diagnosis report carrying the reliability level.
8. The method for online fault diagnosis of rail based on knowledge transfer learning of claim 1, wherein in step S7, when the reliability level is not higher than a level threshold, a redundancy check mechanism is automatically triggered and the low reliability event is recorded in a chain as a basis for model performance evaluation and system maintenance, and the redundancy check mechanism specifically comprises: The time dimension verification comprises the steps of calling historical data of multi-mode monitoring data in the same historical period or a preset time interval of the current diagnosis time point, carrying out trend comparison on the historical data and fault characteristics extracted from a simplified characteristic set according to the student fault diagnosis report, and if the fault characteristics continuously appear or are in a deterioration trend in the historical data, improving the credibility level; When the service capability index of the edge computing equipment meets the preset triggering condition, one or more standby lightweight diagnosis models with different structures are called in parallel to infer the same simplified feature set, and the voting method or the weighted average method is adopted to fuse the output of the models so as to achieve the final student fault diagnosis report; Before transmission, the student fault diagnosis report, the teacher fault diagnosis report and the real-time multi-mode monitoring data are encrypted by adopting an end-to-end encryption technology based on a national encryption algorithm.
9. The method for on-line fault diagnosis of rail based on knowledge transfer learning of claim 1, wherein the step S8 is specifically: the operation and maintenance center marks the student fault diagnosis report or the teacher diagnosis report based on the input field verification result to generate an increment sample set, encrypts the increment sample set by adopting an end-to-end encryption technology based on a national encryption algorithm, and then sends the increment sample set to a server through a secure channel; the server performs incremental learning and model updating by using the incremental sample set, specifically: And then, the student diagnosis model carries out knowledge distillation learning on the updated teacher diagnosis model through the replay strategy, and after updating and verifying the teacher diagnosis model and the student diagnosis model, the student diagnosis model of a new version is safely issued to edge computing equipment for model iteration.
10. The rail online fault diagnosis system based on knowledge transfer learning is characterized by comprising the following modules: the system comprises a historical data acquisition and uploading module, a server, a source domain multi-mode monitoring module, a target domain monitoring module and a storage module, wherein the historical data acquisition and uploading module is used for acquiring source domain multi-mode monitoring data comprising rail geometric data, rail images, rail loads, vibration acceleration, vibration noise and environmental temperature from a plurality of source domain lines of a rail; The historical data decryption storage module is used for obtaining the source domain multi-mode monitoring data and the target domain multi-mode monitoring data and storing the source domain multi-mode monitoring data and the target domain multi-mode monitoring data in a security database after the server decrypts and verifies the received encrypted historical data packet; The system comprises a data set construction module, a global feature priori list, a feature importance ranking module, a data set analysis module and a data set analysis module, wherein the data set construction module is used for preprocessing, fault marking and data enhancement on the source domain multi-mode monitoring data by a server, constructing a source domain data set, and constructing a target domain data set based on each target domain multi-mode monitoring data; the system comprises a model training and deploying module, a training and deploying module and a training module, wherein the model training and deploying module is used for constructing a model framework based on a lightweight confidence rule base by a server, utilizing the source domain data set to train to obtain a teacher diagnosis model, and simultaneously, adopting a mixed training strategy combining knowledge distillation and domain countermeasure self-adaptive transfer learning to train the teacher diagnosis model to obtain a lightweight student diagnosis model based on the target domain data set; The feature dynamic selection module is used for acquiring real-time multi-mode monitoring data from a target domain line by the edge computing equipment, compensating in real time to obtain correction data, and dynamically selecting the most important feature subset from the correction data according to the current service capability index of the edge computing equipment and the global feature priori list issued by the server to form a simplified feature set; the student fault diagnosis report generation module is used for inputting the simplified feature set into a deployed student diagnosis model by the edge computing equipment to obtain a student fault diagnosis result and the confidence coefficient thereof, introducing a D-S evidence theory, fusing the confidence coefficient, the historical diagnosis consistency and the vision auxiliary verification information from the image sensor, and generating a student fault diagnosis report carrying a reliability level; The system comprises a fault diagnosis report sending module, a fault diagnosis report receiving module, a fault diagnosis report sending module and a fault diagnosis report receiving module, wherein the fault diagnosis report sending module is used for judging whether the reliability level is higher than a preset level threshold value or not by edge computing equipment, if so, the student fault diagnosis report is digitally signed and directly sent to an operation and maintenance center through a safety channel, otherwise, the original real-time multi-mode monitoring data is uploaded to a server through the safety channel, the teacher diagnosis model is requested to carry out safety reasoning in a TEE trusted execution environment, a teacher fault diagnosis report is generated, and the teacher fault diagnosis report is sent to the operation and maintenance center through the safety channel after being digitally signed; The incremental learning module is used for marking the student fault diagnosis report or the teacher diagnosis report based on the input field verification result by the operation and maintenance center to generate an incremental sample set, and sending the incremental sample set to the server through the safety channel, wherein the server performs incremental learning and model updating by using the incremental sample set.

Description

Rail online fault diagnosis method and system based on knowledge transfer learning Technical Field The invention relates to the technical field of intelligent monitoring and artificial intelligence crossing of railway infrastructure, in particular to a rail online fault diagnosis method and system based on knowledge transfer learning. Background The health status of a railway track (rail) as a core infrastructure of a railway transportation system is directly related to driving safety and operation efficiency. Along with the continuous extension of railway operation mileage, continuous lifting of train operation speed and gradual increase of carrying load, dynamic load and environmental stress born by a track structure are increasingly complex and severe, so that risks of various faults such as geometrical irregularity, gauge deviation, part damage and the like of the track are obviously increased. Therefore, the method develops a real-time, accurate and reliable online fault diagnosis technology, and has important significance for realizing safety early warning and promoting the transformation of a railway maintenance mode from periodic 'planning maintenance' to predictive 'state maintenance'. In recent years, intelligent diagnosis methods using data driving as a core are becoming a mainstream research direction in the field of rail fault diagnosis. The method relies on the multi-mode monitoring data such as vibration, images, acoustics and the like acquired by the track detection vehicle, the vehicle-mounted sensing device and the fixed monitoring network, and realizes the automatic identification and evaluation of the track state by constructing a machine learning model or a deep learning model. Although the related method shows good prospects, in practical engineering application and system deployment, a plurality of key challenges are still faced, so that the reliability of diagnosis performance is restricted, and potential threats are formed for the overall safety of a railway system: 1. Small samples and data imbalance problem the performance of the data driven model relies on adequate and balanced training samples. However, in actual railway operation, the probability of occurrence of serious faults (such as level III and above irregularity) is extremely low, resulting in serious starvation of the corresponding samples. The extremely unbalanced sample makes the model difficult to fully learn the identification characteristics of key faults, and high-risk faults are easy to miss detection, so that potential safety hazards are formed. 2. The difficulty in real-time computing and feature extraction is that the online diagnostic system needs to have efficient computing power to meet the real-time response requirements. The existing method is beneficial to complete information if the high-dimensional multi-source features are adopted in full quantity, but introduces significant calculation load, is difficult to adapt to real-time diagnosis requirements under high-speed operation, and can cause key information loss and influence accurate characterization of complex faults or compound faults if the feature screening is carried out by excessively relying on manual experience. 3. The model has insufficient cross-domain adaptability, namely when an optimized diagnosis model is trained in a specific line or experimental environment (source domain) and deployed to different lines, climates or operation conditions (target domain), the performance is obviously reduced due to the differences of data distribution, noise background, load spectrum and the like, namely the problem of domain deviation is solved. The model has poor universality, and the quick popularization and application among different lines are restricted. 4. System level security and data credibility challenges existing research focuses on model algorithm performance itself, and lacks overall consideration from the point of view of information physical system (CPS) security. The method is characterized in that the monitoring data can face stealing, tampering or poisoning attacks in transmission, storage and sharing, so that the model is misjudged, the data quality and the labeling reliability are affected by noise, sensor drift and artificial subjectivity, the model learning upper limit is limited, in addition, the system generally lacks long-term evolution and self-adaptive capability, and is difficult to adapt to slow change (concept drift) of data distribution caused by natural degradation of the track state. In summary, when dealing with the first three challenges, the existing methods are often only studied in isolation from the viewpoint of model performance optimization, and lack a systematic integration perspective. More importantly, these challenges are interwoven with the system-level security and trust issues described at point 4, which together highlight the limitations of current technology paths-namely the failure to incorporate model