Search

CN-122020484-A - Multi-path section rail fault collaborative diagnosis method and system based on federal learning

CN122020484ACN 122020484 ACN122020484 ACN 122020484ACN-122020484-A

Abstract

The invention provides a multi-path section railway fault collaborative diagnosis method and system based on federal learning in the technical field of railway track safety monitoring, and the method comprises the steps of S1, initializing a global fault diagnosis module and a global output module by a server, sending initial global diagnosis parameters and initial global output parameters to a client, S2, initializing a local fault diagnosis module and a personalized output module by the client to construct a local fault diagnosis model, S3, training the local fault diagnosis model based on a fault characteristic data set, S4, aggregating local update parameters by the server to update the global fault diagnosis module to obtain optimized global diagnosis parameters and send the client, and repeating training until preset conditions are reached, and S5, carrying out fault diagnosis by the client through the local fault diagnosis model. The invention has the advantages of effectively breaking the core bottleneck of data island, insufficient cross-domain adaptation, contradiction between safety and real-time performance and weak anti-interference capability.

Inventors

  • ZHANG PEI
  • YANG XIAOYU
  • JI PENGXIAO
  • MENG QINGBO
  • ZHANG GUILIN
  • DUAN QIANG

Assignees

  • 郑州铁路职业技术学院

Dates

Publication Date
20260512
Application Date
20260224

Claims (10)

  1. 1. A multi-path section rail fault collaborative diagnosis method based on federal learning is characterized by comprising the following steps: Step S1, road section clients deployed on different road section rails collect heterogeneous monitoring data of the road section rails, which at least comprise rail vibration signals, deformation data and environmental parameters, and carry out local encryption storage; Step S2, preprocessing and labeling the heterogeneous monitoring data by each road section client to construct a local fault characteristic data set; Step S3, initializing a global fault diagnosis module and a global output module by the central server, and encrypting and transmitting initial global diagnosis parameters of the global fault diagnosis module and initial global output parameters of the global output module to clients of each road section; step S4, each road section client initializes a local fault diagnosis module based on the initial global diagnosis parameters, initializes a personalized output module based on the initial global output parameters, and constructs a local fault diagnosis model based on the local fault diagnosis module and the personalized output module; Step S5, each road section client performs local training on the local fault diagnosis model based on the fault characteristic data set, calculates quality evaluation indexes of the round of training data after the training is completed, and encrypts and uploads local update parameters and the quality evaluation indexes of the local fault diagnosis module to a central server; Step S6, a central server adopts a weighted aggregation strategy with Bayesian robustness, calculates aggregation weights based on the quality evaluation indexes uploaded by the road section clients, carries out parameter aggregation on local update parameters based on the aggregation weights so as to update the global fault diagnosis module to obtain optimized global diagnosis parameters, encrypts and transmits the optimized global diagnosis parameters to the road section clients, and repeatedly executes the steps of local training and parameter aggregation until the global fault diagnosis module reaches a preset convergence condition or training round, and transmits the latest optimized global diagnosis parameters to the road section clients so as to update the local fault diagnosis module; and S7, carrying out fault identification and classification on the local real-time heterogeneous monitoring data by each road section client through the latest local fault diagnosis model, outputting a real-time fault diagnosis result, and realizing the collaborative diagnosis of the multi-path section railway faults.
  2. 2. The method for collaborative diagnosis of multi-path segment rail faults based on federal learning according to claim 1, wherein the step S1 is specifically as follows: Road section clients deployed on different road section rails collect heterogeneous monitoring data of the road section rails, which at least comprise rail vibration signals, deformation data and environmental parameters, through a sensor array; the rail vibration signal at least comprises a rail vertical vibration acceleration, a rail transverse vibration acceleration, a sleeper vibration amplitude and a vibration frequency, the deformation data at least comprises a rail vertical displacement, a rail transverse displacement, a rail gauge deviation and a rail bending degree, and the environment parameters at least comprise an environment temperature, an environment humidity, a precipitation amount, a wind power level and a dust concentration; And the road section client stores the collected heterogeneous monitoring data into a database which is locally encrypted by adopting an AES-256 encryption algorithm.
  3. 3. The multi-path segment rail fault collaborative diagnosis method based on federal learning according to claim 2, wherein the database is provided with a role grading authorization mechanism, corresponding operation authorities are only allocated for authorized operation staff, illegal access is prevented by combining multi-factor authentication, and operation logs of all data access, data modification and data export are recorded at the same time and kept for at least 90 days; And before updating the database, carrying out integrity check on the data fingerprint based on the previous blockchain storage, periodically carrying out cloud encryption backup on the data stored in the database, and continuously clearing invalid data in the database based on the storage space of the road client and a preset clearing rule.
  4. 4. The method for collaborative diagnosis of multi-path segment rail faults based on federal learning according to claim 1, wherein the step S2 is specifically as follows: Each road section client performs preprocessing including outlier processing, noise filtering, data alignment and synchronization, normalization processing and characteristic engineering on the heterogeneous monitoring data to obtain standardized data; marking the fault type and the fault level of each piece of standardized data according to railway industry standards to obtain corresponding marking information; The SMOTE oversampling technology or the random undersampling technology is adopted, the sample proportion of a fault sample and a normal sample in each standardized data is balanced based on the marking information, and a local fault characteristic data set is further constructed; Identifying and removing abnormal data points generated by instantaneous failure or external interference of a sensor in the heterogeneous monitoring data by adopting a statistical method based on a quartile range or 3 sigma rule; The noise filtering is specifically implemented by filtering high-frequency noise and power frequency interference from track vibration signals in the heterogeneous monitoring data by adopting a Butterworth band-pass filter or a wavelet threshold denoising method, retaining effective frequency band signals related to rail fault characteristics, smoothing the deformation data by adopting a sliding average filtering method, and identifying and processing physically impossible abnormal jump through a preset instantaneous change rate threshold by adopting a sliding window median filtering method and introducing self-adaptive filtering and multi-sensor data verification based on physical constraint to the environment parameters and cross-verifying by utilizing correlation among parameters; the data alignment and synchronization is specifically that based on a unified first time stamp, time alignment is carried out on the heterogeneous monitoring data from different sampling frequency sensors, the data are unified to the same time sequence through an interpolation resampling method, and time sequence relevance among the data is ensured; The normalization processing specifically comprises the steps of performing Z-score normalization or Min-Max normalization on the heterogeneous monitoring data, and eliminating dimension influence; The characteristic engineering specifically comprises the steps of extracting time domain characteristics, frequency domain characteristics and time-frequency domain characteristics from the heterogeneous monitoring data.
  5. 5. The method for collaborative diagnosis of multi-path segment rail faults based on federal learning according to claim 1, wherein the step S3 is specifically: the central server initializes a global fault diagnosis module and a global output module; The global fault diagnosis module is constructed based on the heterogeneous data adapting unit, the multi-scale feature extracting unit, the road crossing section feature fusion unit and the feature calibration unit, and the global output module is constructed based on the feature enhancement unit, the multi-task prediction unit and the probability calibration unit; Acquiring initial global diagnosis parameters of the global fault diagnosis module, wherein the initial global output parameters of the global output module; Obtaining a current second timestamp, calculating hash values of the initial global diagnosis parameter, the initial global output parameter and the second timestamp through an SM3 algorithm, encrypting the initial global diagnosis parameter, the initial global output parameter, the second timestamp and the hash values into an encryption parameter packet through a preset SM4 original key, retrieving public keys of clients of all road sections through an auxiliary server, respectively encrypting the SM4 original key through the public keys to obtain corresponding encryption keys, and transmitting the encryption parameter packet and the encryption keys to the clients of the corresponding road sections through an SSL encryption transmission link.
  6. 6. The multi-path segment rail fault collaborative diagnosis method based on federal learning according to claim 5, wherein the heterogeneous data adapting unit is constructed based on vibration signal branches, deformation data branches, environmental parameter branches and fusion subunits; The vibration signal branch is used for capturing time-frequency local characteristics of a non-stationary track vibration signal by utilizing a wavelet basis function through the wavelet convolution layer; the deformation data branches are used for capturing the spatial correlation characteristics of the steel rail deformation from the deformation data through a spatial convolution layer by utilizing a 3 multiplied by 3 convolution kernel; the environment parameter branches are used for capturing time sequence variation trend characteristics of environment parameters through a time sequence convolution layer by utilizing a1 xk convolution kernel; the fusion subunit is used for dynamically calculating the attention weights of the time-frequency local features, the spatial correlation features and the time-varying trend features through a multi-head attention weighting layer, and carrying out weighted fusion to obtain unified embedded features; The multi-scale feature extraction unit is constructed based on a multi-scale cavity convolution block, a channel attention subunit and a space attention subunit; The multi-scale cavity convolution block is used for capturing local features and global features from the unified embedded features through 4 parallel cavity convolution layers; the channel attention subunit is used for carrying out channel dimension extrusion-excitation on the output of the multi-scale cavity convolution block through the extrusion-excitation module, enhancing a fault sensitive channel and inhibiting a redundant channel; the space attention subunit is used for introducing position information for the output of the channel attention subunit through the coordinate attention module, focusing the spatial position related features of the occurrence of faults, improving the spatial discriminant of the features and outputting multi-scale fusion features; The cross-road segment feature fusion unit is constructed based on a graph construction subunit, a graph convolution layer and a knowledge distillation subunit; the graph construction subunit is used for extracting a normalized adjacency matrix from the multi-scale fusion features through a dynamic federation graph generator; the graph roll stacking layer is used for aggregating the characteristics of adjacent or similar road sections in the normalized adjacent matrix through weighted graph convolution and outputting a common fault characteristic matrix; The knowledge distillation subunit is used for taking the strong road section characteristics with rich fault samples and high characteristic quality in the common fault characteristic matrix as teacher signals, distilling the strong road section characteristics into weak road section characteristics with rare fault samples so as to improve the characteristic expression capacity of the weak road section and output the common characteristics of the road sections; the characteristic calibration unit is constructed based on a characteristic credibility evaluation subunit, a self-adaptive filtering subunit and a characteristic smoothing subunit; the feature credibility evaluation subunit is used for calculating a credibility score matrix of the cross-road segment commonality feature of the rail of each road section based on each multi-scale fusion feature and a global feature distribution statistic by a Markov distance calculator, wherein the global feature distribution statistic is a mean vector mu and a covariance matrix sigma based on the multi-scale fusion feature statistics of the rail of all road sections; The self-adaptive filtering subunit is used for setting a dynamic soft threshold value based on the credibility score matrix through a soft threshold value filtering layer, and carrying out weight suppression on the cross-road segment commonality characteristic to obtain a filtering commonality characteristic matrix; The characteristic smoothing subunit is used for carrying out Gaussian smoothing on the filtering commonality characteristic matrix through a Gaussian kernel smoothing layer so as to reduce characteristic fluctuation, improve stability of commonality characteristics and output calibration commonality fault characteristics; the feature enhancement unit is constructed based on a residual connection block and a task attention subunit; The residual error connecting block is used for carrying out characteristic enhancement on the calibration common fault characteristic through a ResNet bottleneck structure to obtain a residual error enhancement common fault characteristic; the task attention subunit is used for dynamically adjusting the feature enhancement weight of the residual enhancement common fault feature based on the task difference of fault type prediction and fault grade prediction through the dynamic attention layer, improving the suitability of the feature and the task, and obtaining the task enhancement common feature; The multi-task prediction unit is constructed based on a shared feature layer, a fault type prediction branch, a fault grade prediction branch and a task interaction subunit; The shared feature layer is used for extracting a dual-task shared basic feature universal for a fault type prediction task and a fault grade prediction task from the task enhanced commonality features through a 2-layer full-connection layer; The fault type prediction branch is used for identifying fault type probability distribution defined by railway industry standards from the dual-task sharing basic characteristics through 3-layer MLP and Softmax activation; The fault level prediction branch is used for identifying fault level probability distribution from the dual-task sharing basic characteristics through 3-layer MLP and Sigmoid activation; the task interaction subunit is used for calculating an interaction loss value through an interaction loss function, and updating network parameters of a shared feature layer, a fault type prediction branch and a fault grade prediction branch based on the interaction loss value in a back propagation mode; the probability calibration unit is constructed based on the distribution adaptation subunit and the probability correction subunit; The distribution adaptation subunit is used for carrying out temperature scaling on fault type probability distribution and fault grade probability distribution; The probability correction subunit is used for carrying out nonlinear correction on the fault type probability distribution and the fault grade probability distribution after temperature scaling through the Beta calibration layer so as to output a fault diagnosis result carrying the fault type and the fault grade.
  7. 7. The method for collaborative diagnosis of multi-path segment rail faults based on federal learning according to claim 1, wherein the step S4 is specifically: Each road section client receives an encryption parameter packet and an encryption key issued by a central server, calls a locally stored private key, decrypts the received encryption key to obtain an SM4 original key, and decrypts the received encryption parameter packet through the SM4 original key to obtain an initial global diagnosis parameter, an initial global output parameter, a second timestamp and a hash value; After carrying out integrity verification on the initial global diagnosis parameter, the initial global output parameter and the second timestamp based on the hash value, carrying out aging verification based on the second timestamp, and completing parameter receiving if verification passes; And initializing a local fault diagnosis module which is the same as the network architecture of the global fault diagnosis module by each road section client based on the initial global diagnosis parameters, initializing a personalized output module which is the same as the architecture of the global output module based on the initial global output parameters, and constructing a local fault diagnosis model based on the local fault diagnosis module and the personalized output module.
  8. 8. The method for collaborative diagnosis of multi-path segment rail faults based on federal learning according to claim 1, wherein the step S5 is specifically: The client side of each road section preliminarily divides the fault characteristic data set into a training set and a verification set according to the proportion of 7:3 by taking 30 days as a basic time window, and ensures that the training set covers samples of all fault types; Calculating fault distribution entropy of the verification set, and dynamically adjusting the proportion of the training set to the verification set in the basic time window if the fault distribution entropy is smaller than a preset entropy threshold; Performing fault scene driven online data enhancement on the training set and the verification set, wherein the online data enhancement specifically comprises the following steps: aiming at a sample of the rail vibration signal, applying preset amplitude scaling, time stretching and Gaussian noise superposition based on a physical simulation rule so as to simulate fault signal changes under different train loads and speeds; Aiming at the samples of deformation data, combining with the history distribution of local environmental parameters, generating deformation derivative samples under different environmental couplings through linear transformation; Aiming at the samples of the environmental parameters, generating virtual samples with the synergistic effect of the multiple environmental factors through permutation and combination, and compensating the defect of scarcity of the local composite scene samples; The training set after the online data enhancement carries out local training on the local fault diagnosis model, and the verification set after the online data enhancement carries out verification on the trained local fault diagnosis model; After training, calculating the quality evaluation index of the training data of the current round, and uploading the local updating parameters and the quality evaluation index of the local fault diagnosis module to a central server through an SSL (secure socket layer) encryption transmission link after adopting a homomorphic encryption algorithm to encrypt the local updating parameters and the quality evaluation index of the local fault diagnosis module.
  9. 9. The method for collaborative diagnosis of multi-path segment rail faults based on federal learning of claim 1, wherein in the step S6, the weighted aggregation strategy with Bayesian robustness is specifically that a central server calculates the aggregation weight of each local update parameter based on the quality evaluation index uploaded by each road segment client, eliminates extreme abnormal parameters in the local update parameters by adopting a pruning average algorithm, and performs weighted summation on effective parameters in the local update parameters based on the aggregation weight.
  10. 10. A multi-path section rail fault collaborative diagnosis system based on federal learning is characterized by comprising the following modules: The heterogeneous monitoring data acquisition module is used for being deployed at road section clients of different road section rails, acquiring heterogeneous monitoring data of the road section rails, which at least comprise rail vibration signals, deformation data and environmental parameters, and carrying out local encryption storage; the fault characteristic data set construction module is used for preprocessing and marking the heterogeneous monitoring data by each road section client to construct a local fault characteristic data set; the server model initializing module is used for initializing a global fault diagnosis module and a global output module by the central server and encrypting and transmitting initial global diagnosis parameters of the global fault diagnosis module and initial global output parameters of the global output module to clients of each road section; the local fault diagnosis model initialization module is used for initializing a local fault diagnosis module based on the initial global diagnosis parameters by each road section client, initializing a personalized output module based on the initial global output parameters, and constructing a local fault diagnosis model based on the local fault diagnosis module and the personalized output module; The local training module is used for carrying out local training on the local fault diagnosis model by each road section client based on the fault characteristic data set, calculating the quality evaluation index of the round of training data after the training is finished, and encrypting and uploading the local updating parameters and the quality evaluation index of the local fault diagnosis module to the central server; The parameter aggregation module is used for calculating an aggregation weight based on the quality evaluation index uploaded by each road section client by adopting a weighted aggregation strategy with Bayesian robustness by the central server, carrying out parameter aggregation on each local update parameter based on each aggregation weight so as to update the global fault diagnosis module to obtain an optimized global diagnosis parameter, encrypting and issuing the optimized global diagnosis parameter to each road section client, and repeatedly executing the steps of local training and parameter aggregation until the global fault diagnosis module reaches a preset convergence condition or training round, and encrypting and issuing the latest optimized global diagnosis parameter to the road section client to update the local fault diagnosis module; the rail fault diagnosis module is used for carrying out fault identification and classification on the local real-time heterogeneous monitoring data through the latest local fault diagnosis model by each road section client, outputting a real-time fault diagnosis result and realizing the collaborative diagnosis of the multi-path section rail faults.

Description

Multi-path section rail fault collaborative diagnosis method and system based on federal learning Technical Field The invention relates to the technical field of railway track safety monitoring, in particular to a multi-path section railway track fault collaborative diagnosis method and system based on federal learning. Background Along with the transformation of the railway industry from a large-scale 'design building' stage to 'design building and operation maintenance and the repeated' stage, the railway mileage and the running speed are continuously improved, the heavy load transportation capacity is continuously increased, various faults such as irregularity in height, fine deformation and the like are easily generated under the effects of repeated rolling of wheels, alternating temperature change and complex geological environment, if the faults cannot be accurately diagnosed in time, the minor hidden trouble can gradually accumulate and evolve into serious safety accidents, and the running safety and the life and property safety of passengers are directly threatened. According to industry prediction, the railway maintenance equipment scale reaches 301.02 hundred million yuan by 2025, the annual composite growth rate reaches 11.64%, and the track fault diagnosis technology is used as a core support of a railway operation and maintenance system, so that the demands on accuracy, instantaneity and synergy are increasingly urgent. In order to solve the difficult problem of rail fault diagnosis, the prior art has gradually developed from traditional manual detection and periodical rail inspection vehicle detection to the online diagnosis direction of data driving and artificial intelligence assistance. For example, a track condition monitoring method based on machine vision is provided in the prior art, the primary identification of track faults is realized through vision acquisition and analysis, meanwhile, a diagnosis method based on data driving and expert knowledge fusion and an online diagnosis method based on knowledge transfer learning are also gradually applied, and the sensitivity and model instantaneity of small sample fault diagnosis are optimized through constructing the mapping relation between fault features and diagnosis results and combining the technologies such as reinforcement learning, average influence degree analysis and the like, and millimeter-level fault feature capturing and grading early warning can be realized by part of methods, so that the diagnosis efficiency and accuracy are improved to a certain extent. In addition, the global first set of track intelligent diagnosis analysis system is put into operation, the data scattering bottleneck is broken through a dynamic planning flexible matching algorithm, the ordered association and the rapid screening of multi-period data are realized, and the track maintenance is promoted to be upgraded to a mode of 'real-time monitoring-intelligent early warning-accurate treatment'. However, the existing rail fault diagnosis technology still has a plurality of bottlenecks under the multi-path collaborative diagnosis scene, and is difficult to meet the safety operation and maintenance requirements of a large-scale railway network: First, the problem of data islanding is prominent, and the collaborative diagnosis basis is weak. Railway track data are generally stored in different railway offices, working sections and operation and maintenance units in a scattered mode, sensitive information such as track vibration signals, deformation data, environmental parameters and the like is covered, and the sensitive information is limited by data privacy protection regulations, industry safety specifications and department management boundaries, and the data are forbidden to be transmitted outwards or collected together in a cross-unit mode, so that mutually isolated data islands are formed. The existing diagnosis method is mostly based on a single road section or a single data source training model, and cannot integrate the complementary value of heterogeneous data of multiple road sections, so that the generalization capability of the model is limited, and rail fault diagnosis requirements under different regions and different working conditions (such as climate difference, geological conditions and traffic level) are difficult to adapt. Secondly, the cross-domain suitability is insufficient, and the diagnosis precision is limited by the data distribution difference. The existing migration learning method can reduce the model deviation of a source domain (such as a rail inspection vehicle environment) and a target domain (such as a practical operation environment) to a certain extent, but aiming at a multi-road-section scene, the rail materials, the abrasion degree, the operation age and the monitoring equipment types of each road section are different, so that the fault data distribution heterogeneity is obvious, the single migration model is difficu