An intelligent diagnosis data service platform based on rail transit

Abstract

The application discloses a kind of based on rail transit intelligent diagnosis data service platform, comprising: constructing rail transit intelligent diagnosis data service system, decouples historical operation data into working condition latent space code and fault latent space code based on variational autoencoder;In working condition latent space, unseen working condition space is generated by convex hull sampling, and in fault latent space, unseen fault mode space is generated by extrapolation based on Gaussian process regression;Two kinds of latent space codes are combined and decoded to generate a generalization test sample set;Batch inference is carried out on the diagnostic model on the sample set, and accuracy, confidence distribution and stability index are obtained;Combined with performance threshold, the generalization failure boundary is identified, and a structured verification report is generated for model deployment access decision.The reliability of the model under unseen scenarios can be quantitatively evaluated before model deployment, and continuous generalization capability determination basis is provided during operation.

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, and in particular to a rail transit intelligent diagnostic data service platform. Background Technology

[0002] With the expansion of urban rail transit networks and the continuous increase in operational density, health monitoring and intelligent diagnosis of rail transit vehicles and related signaling equipment have become key technologies for ensuring operational safety and reducing maintenance costs. In recent years, digital twin technology, machine learning methods, and data service architectures have been gradually introduced into the field of rail transit diagnosis, resulting in a number of data-driven intelligent diagnostic systems.

[0003] For example, existing technologies utilize adaptive antlion strategies and virtual modeling techniques to construct digital twin health analysis models. These models are trained on vehicle lifecycle data and combined with Bayesian judgment principles to output health diagnosis results. Such methods can achieve real-time assessment of vehicle health status to a certain extent. Another approach uses rough set theory to reduce the attributes of a fault diagnosis decision table, and uses the reduced attribute combinations as input to a neural network. This neural network is then trained to probabilistically identify faulty areas, thereby diagnosing signal equipment faults. Furthermore, other technologies establish multi-dimensional analysis processes, including energy consumption diagnosis, energy efficiency diagnosis, energy quality diagnosis, and environmental quality diagnosis, to systematically determine anomalies in the energy consumption of rail transit networks.

[0004] However, in the actual deployment and application of the above-mentioned existing technologies, there is a key technical problem that has not yet been effectively solved: the generalization ability of the diagnostic model under unseen operating conditions and unseen fault modes lacks a systematic verification mechanism.

[0005] Specifically, rail transit systems exhibit significant operational complexity and failure mode uncertainty. On the one hand, vehicle operating conditions are influenced by a combination of factors, including track conditions, load variations, ambient temperature, and scheduling strategies, inevitably leading to unseen combinations of operating conditions not present in the training data during actual operation. On the other hand, with equipment aging, changes in the external environment, or the emergence of new faults, the system may also exhibit unseen failure modes, meaning that the manifestations and characteristic distributions of these faults differ fundamentally from historical training samples.

[0006] Current technologies for model training and validation typically employ cross-validation or hold-out validation using training and test sets partitioned from historical data. This implicitly assumes that the test and training data follow the same distribution. However, this assumption often fails in actual rail transit operations. When a model is deployed to a new line, a new model type, a new season, or a new operating period, even if the model demonstrates high accuracy on the historical test set, its diagnostic performance under real-world, unseen conditions may drastically decline, manifesting as increased false alarm rates, increased false negative rates, or distorted diagnostic confidence.

[0007] More seriously, existing technologies lack proactive verification methods for model generalization capabilities. Before deploying a model, operations personnel cannot quantitatively assess the extent to which the model can maintain reliable diagnosis when faced with unseen operating conditions. During model operation, when the distribution of input data deviates from that of training data, the system cannot automatically identify the deviation and trigger a re-verification mechanism for generalization capabilities. This leads to a common dilemma in practical applications: high testing accuracy but low field reliability, severely hindering the engineering implementation and large-scale promotion of intelligent diagnostic data services.

[0008] Therefore, how to construct a mechanism that can systematically verify the generalization ability of diagnostic models under unseen operating conditions and unseen fault modes, and provide quantifiable generalization performance evaluation before model deployment and during operation, has become a specific technical problem that urgently needs to be solved in the field of intelligent diagnostic data services for rail transit. Summary of the Invention

[0009] To address the lack of a systematic verification mechanism for the generalization ability of existing diagnostic models under unseen operating conditions and unseen fault modes, this invention provides a rail transit intelligent diagnostic data service platform.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention discloses a rail transit intelligent diagnostic data service platform, comprising: The working condition feature decoupling module is configured to acquire the historical operation dataset of rail transit vehicles. The historical operation dataset includes multiple working condition labels and multiple fault mode labels. The variational autoencoder is used to decouple each operation sample in the historical operation dataset into working condition latent space coding and fault latent space coding. The unseen working condition space construction module is configured to receive the working condition latent space encoding, perform spatial interpolation operation based on convex hull sampling within the working condition feature distribution boundary spanned by the working condition latent space encoding, and generate several sets of synthetic working condition latent space encodings that do not overlap with the working condition labels of any running sample in the historical running dataset, as the unseen working condition space. The unseen fault mode space construction module is configured to receive the fault latent space code, perform an extrapolation sampling operation based on Gaussian process regression on the fault latent space code, and generate several sets of synthetic fault latent space codes that exceed the boundary range of the fault latent space codes corresponding to all fault mode labels in the historical running dataset, as the unseen fault mode space. The generalization test sample generation module is configured to perform a vector concatenation operation between each synthetic working condition latent space code in the unseen working condition space and each synthetic fault latent space code in the unseen fault mode space to generate a generalization test sample set. The diagnostic model performance evaluation module is configured to deploy the diagnostic model to be validated on the generalization test sample set for batch diagnostic inference, calculate the diagnostic accuracy, diagnostic confidence distribution and diagnostic result stability index of the diagnostic model on the generalization test sample set, and output multi-dimensional generalization performance quantification results. The generalization boundary identification module is configured to receive the multidimensional generalization performance quantization result, back locate the critical sample region in the generalization test sample set whose diagnostic performance is lower than the generalization performance acceptance threshold based on a preset generalization performance acceptance threshold, map the critical sample region back to the working condition feature distribution space and the fault feature distribution space, and output the generalization failure boundary description information of the diagnostic model to be verified. The diagnostic service orchestration module is configured to encapsulate the multidimensional generalization performance quantification results and the generalization failure boundary description information into a generalization capability verification report, and push the generalization capability verification report to the operation and maintenance decision-making end, so that the operation and maintenance decision-making end can perform deployment admission judgment before the diagnostic model to be verified is deployed online.

[0011] Furthermore, the operating condition feature decoupling module includes: a historical data receiving unit, configured to access the rail transit vehicle monitoring data service interface to obtain the historical operating dataset. Each operating sample in the historical operating dataset is pre-labeled with a corresponding operating condition label and a corresponding fault mode label. The operating condition label is used to identify the combination category of line conditions, load status and ambient temperature during vehicle operation, and the fault mode label is used to identify the type of fault that occurred in the vehicle component. The parallel encoder unit is configured to receive the historical running dataset, use a first encoder to map each running sample to the working condition latent space code, and use a second encoder to map the same running sample to the fault latent space code. The first encoder and the second encoder share the underlying feature extraction network but are decoupled at the high-level network branches. The mutual exclusion constraint unit is configured to calculate the mutual information between the working condition latent space code and the fault latent space code, and backpropagate to update the network parameters of the first encoder and the second encoder with minimizing the mutual information as a constraint condition, thereby forcing the working condition latent space code and the fault latent space code to be orthogonal to each other in the feature dimension. The reconstruction verification unit is configured to concatenate the working condition latent space encoding and the fault latent space encoding execution vector and input them into the decoder. The decoder outputs the reconstructed running sample and uses the reconstruction error between the reconstructed running sample and the running sample as part of the training loss to feed back to the parallel encoder unit and the decoder. The encoding output unit is configured to receive all running samples in the historical running dataset after the parallel encoder unit has converged during training, output the working condition latent space code and the fault latent space code corresponding to each running sample, and transmit the feature distribution boundary formed by all the working condition latent space codes to the unseen working condition space construction module, and transmit the feature distribution boundary formed by all the fault latent space codes to the unseen fault mode space construction module.

[0012] Furthermore, the unseen working condition space construction module includes: a convex hull boundary calculation unit, configured to receive all the working condition latent space codes output by the working condition feature decoupling module, calculate the convex hull boundary of the point set formed by all the working condition latent space codes in the working condition latent space dimension where the working condition latent space codes are located, and use the convex hull boundary as the geometric description of the known working condition distribution boundary. The spatial sampling strategy unit is configured to establish a uniform sampling grid within the spatial region enclosed by the convex hull boundary, traverse each candidate sampling point in the uniform sampling grid, calculate the Euclidean distance between each candidate sampling point and each known point in the working condition latent space encoding, and select the candidate sampling points whose Euclidean distances to all known points are greater than a preset distance threshold as boundary interpolation points. The synthetic coding generation unit is configured to obtain the boundary interpolation points, output each boundary interpolation point as a set of synthetic working condition latent space codes, and generate a virtual working condition label corresponding to each synthetic working condition latent space code. The virtual working condition label is used to identify that the working condition corresponding to the synthetic working condition latent space code does not have the same type of working condition label in the historical working dataset. The unseen working condition space output unit is configured to associate and store all the synthetic working condition latent space codes with all the virtual working condition labels to form the unseen working condition space, and transmit the unseen working condition space to the generalization test sample generation module.

[0013] Furthermore, the unseen fault mode space construction module includes a Gaussian process modeling unit, configured to receive all the fault latent space codes output by the working condition feature decoupling module, use the fault latent space dimension where the fault latent space code is located as the input space, and use the fault mode label corresponding to each fault latent space code as the observation value to construct a Gaussian process regression model. The Gaussian process regression model is used to describe the nonlinear mapping relationship between the fault latent space code and the fault mode label, as well as the prediction uncertainty at the location of the fault latent space code. The extrapolation sampling unit is configured to obtain the feature distribution boundary of the fault latent space code in the fault latent space dimension, generate several sets of candidate fault latent space codes in the extended region extending outward from the feature distribution boundary, input each set of candidate fault latent space codes into the Gaussian process regression model, calculate the prediction mean and prediction variance corresponding to each set of candidate fault latent space codes, and select the candidate fault latent space codes whose prediction variance is greater than a preset variance threshold as extrapolation sampling points. The virtual fault label generation unit is configured to acquire the extrapolated sampling points, output each extrapolated sampling point as a group of synthetic fault latent space codes, and generate a corresponding virtual fault mode label for each group of synthetic fault latent space codes. The virtual fault mode label is used to identify that the fault mode corresponding to the synthetic fault latent space code does not have the same type of fault mode label in the historical running dataset, and the virtual fault mode label is different from all the fault mode labels in the historical running dataset. The unseen fault mode space output unit is configured to associate and store all the synthetic fault latent space codes with all the virtual fault mode labels to form the unseen fault mode space, and transmit the unseen fault mode space to the generalization test sample generation module.

[0014] Furthermore, the generalization test sample generation module includes: an encoding splicing unit, configured to receive all the synthetic working condition latent space codes output by the unseen working condition space construction module and all the synthetic fault latent space codes output by the unseen fault mode space construction module, perform a Cartesian product combination operation on each group of synthetic working condition latent space codes and each group of synthetic fault latent space codes, and sequentially splice the synthetic working condition latent space codes and the synthetic fault latent space codes in each combination generated by the Cartesian product combination operation in the vector dimension to generate a spliced ​​latent space code with a working condition-fault joint latent space expression; The decoding and reconstruction unit is configured to input each of the concatenated latent space codes into the decoder included in the working condition feature decoupling module, and the decoder maps the concatenated latent space codes back to the original data space to generate a generalized test sample corresponding to the concatenated latent space code. The generalized test sample is consistent with the running sample in the historical running dataset in terms of data form. The sample labeling unit is configured to obtain the virtual operating condition label of the synthetic operating condition latent space encoding and the virtual fault mode label of the synthetic fault latent space encoding corresponding to each generalization test sample, and to jointly label the virtual operating condition label and the virtual fault mode label as the real generalization attribute labeling information of the generalization test sample. The test sample set output unit is configured to associate and store all the generalized test samples with the real generalized attribute annotation information corresponding to all the generalized test samples to form the generalized test sample set, and transmit the generalized test sample set to the diagnostic model performance evaluation module.

[0015] Furthermore, the diagnostic model performance evaluation module includes: a diagnostic model deployment unit, configured to receive the generalized test sample set output by the generalized test sample generation module, obtain the model file of the diagnostic model to be verified, instantiate the diagnostic model to be verified and load it into the inference runtime environment, and establish a data reading channel between the inference runtime environment and the generalized test sample set. The batch inference execution unit is configured to sequentially read each generalized test sample in the generalized test sample set through the data reading channel, input each generalized test sample into the diagnostic model to be verified, call the inference interface of the diagnostic model to be verified to perform diagnostic classification operation on each generalized test sample, and output the diagnostic classification result and diagnostic confidence value corresponding to each generalized test sample. The accuracy calculation unit is configured to obtain the diagnostic classification result and the real generalization attribute annotation information corresponding to each generalization test sample, compare the diagnostic classification result with the virtual fault mode label in the real generalization attribute annotation information, and count the proportion of the number of samples that match the comparison in the generalization test sample set to the total number of samples as the diagnostic accuracy. The confidence distribution calculation unit is configured to obtain the diagnostic confidence value corresponding to each generalization test sample, divide all the diagnostic confidence values ​​into several confidence levels according to the value range, statistically analyze the distribution of the number of generalization test samples in each confidence level and the distribution of the number of samples that match the diagnostic classification result with the real generalization attribute labeling information, and output a confidence-accuracy joint distribution histogram. The stability index calculation unit is configured to perform several repeated sampling operations on the generalization test sample set. Each repeated sampling operation randomly selects a fixed number of samples from the generalization test sample set to form a repeated sampling subset. The unit calculates the subset diagnostic accuracy corresponding to each repeated sampling subset and calculates the variance of the diagnostic accuracy of all subsets as the stability index of the diagnostic result. The generalization performance quantification result output unit is configured to encapsulate the diagnostic accuracy, the confidence-accuracy joint distribution histogram, and the diagnostic result stability index in a structured manner to form the multidimensional generalization performance quantification result, and transmit the multidimensional generalization performance quantification result to the generalization boundary identification module and the diagnostic service orchestration module.

[0016] Furthermore, the generalization boundary identification module includes: a performance threshold determination unit, configured to receive the multidimensional generalization performance quantification result output by the diagnostic model performance evaluation module, extract the diagnostic accuracy and the diagnostic result stability index from the multidimensional generalization performance quantification result, obtain a preset generalization performance acceptance threshold, the generalization performance acceptance threshold including a lower accuracy threshold and a higher stability threshold, and mark the generalization test samples in the generalization test sample set whose diagnostic accuracy is lower than the lower accuracy threshold or whose diagnostic result stability index is higher than the higher stability threshold as critical samples; The critical sample localization unit is configured to acquire the real generalized attribute annotation information corresponding to each critical sample, extract the virtual operating condition label and the virtual fault mode label corresponding to each critical sample from the real generalized attribute annotation information, establish a co-occurrence frequency matrix of the virtual operating condition label and the virtual fault mode label in the critical sample set, and mark the combination of the virtual operating condition label and the virtual fault mode label with a frequency exceeding a preset frequency threshold in the co-occurrence frequency matrix as a failure label combination; The working condition space boundary mapping unit is configured to receive all the synthetic working condition latent space codes and their corresponding virtual working condition labels output by the unseen working condition space construction module, filter out all the synthetic working condition latent space codes of the virtual working condition labels that belong to the virtual working condition labels in the failure label combination, calculate the minimum convex hull boundary of all the filtered synthetic working condition latent space codes in the working condition latent space dimension, and use the minimum convex hull boundary as the working condition generalization failure boundary. The fault space boundary mapping unit is configured to receive all the synthetic fault latent space codes and their corresponding virtual fault mode labels output by the unseen fault mode space construction module, filter out all the synthetic fault latent space codes of the virtual fault mode labels belonging to the virtual fault mode labels in the failure label combination, calculate the minimum convex hull boundary of all the filtered synthetic fault latent space codes in the fault latent space dimension, and use the minimum convex hull boundary as the fault generalization failure boundary. The boundary description information generation unit is configured to obtain the generalized failure boundary of the operating condition and the generalized failure boundary of the fault, convert the operating condition feature value range corresponding to the generalized failure boundary of the operating condition into rail transit physical operating condition description text, convert the fault feature value range corresponding to the generalized failure boundary of the fault into rail transit fault mode description text, perform a structured combination of the rail transit physical operating condition description text and the rail transit fault mode description text to generate the generalized failure boundary description information, and transmit the generalized failure boundary description information to the diagnostic service orchestration module.

[0017] 8. A rail transit intelligent diagnostic data service platform according to claim 1, characterized in that the diagnostic service orchestration module includes: a report encapsulation unit, configured to receive the multidimensional generalization performance quantification result output by the diagnostic model performance evaluation module and the generalization failure boundary description information output by the generalization boundary identification module, and to structurally fill the diagnostic accuracy, the confidence-accuracy joint distribution histogram, and the diagnostic result stability index in the multidimensional generalization performance quantification result with the rail transit physical condition description text and the rail transit fault mode description text in the generalization failure boundary description information according to a preset report template to generate a generalization capability verification report; The report push unit is configured to obtain the generalization capability verification report, establish an encrypted communication link with the operation and maintenance decision-making end, transmit the generalization capability verification report to the operation and maintenance decision-making end in structured data format through the encrypted communication link, and send a report reception confirmation instruction to the operation and maintenance decision-making end after the transmission is completed. The admission determination execution unit is configured to receive a deployment admission instruction or a deployment rejection instruction from the operation and maintenance decision-making end based on the generalization capability verification report. When the deployment admission instruction is received, the unit generates a deployment license credential for the diagnostic model to be verified and outputs the deployment license credential to the model deployment executor. When the deployment rejection instruction is received, the unit encapsulates the generalization failure boundary description information in the generalization capability verification report into a model improvement suggestion and feeds it back to the model training end for the model training end to perform training sample amplification or model structure optimization operations.

[0018] The technological advancements achieved by this invention compared to existing technologies are as follows: This invention achieves a systematic evaluation of the performance of diagnostic models under unseen operating conditions and unseen fault modes by constructing a generalization verification mechanism that combines decoupling of operating condition features, generation of unseen operating condition space, and extrapolation of unseen fault modes. Utilizing the decoupled operating condition latent space and fault latent space, representative generalization test samples are generated both within and outside the distribution, respectively. This ensures that the test data no longer depends on the same distribution assumption, fundamentally overcoming the problem of insufficient coverage of real-world complex operating conditions by traditional verification methods.

[0019] Meanwhile, by constructing a unified generalization test sample set and introducing multi-dimensional performance quantification indicators, the diagnostic accuracy, confidence consistency, and result stability of the model are comprehensively evaluated, realizing a quantifiable expression of generalization ability. Furthermore, by combining a generalization boundary identification method, the performance degradation region is mapped back to the operating condition and fault feature space, forming a physically meaningful generalized failure boundary description, transforming the model's generalization ability from an abstract indicator into an interpretable spatial range.

[0020] Based on the above mechanism, this invention can quantitatively assess the reliability of a model in unseen scenarios before deployment and provide a continuous basis for judging its generalization ability during operation, effectively avoiding the application risk of high testing accuracy but low actual reliability, thereby significantly improving the engineering availability and safety assurance capabilities of the intelligent diagnostic system for rail transit. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0022] In the attached diagram: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0023] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0024] like Figure 1 As shown, this invention discloses a rail transit intelligent diagnostic data service platform, comprising: The working condition feature decoupling module is configured to acquire the historical operation dataset of rail transit vehicles. The historical operation dataset includes multiple working condition labels and multiple fault mode labels. The variational autoencoder is used to decouple each operation sample in the historical operation dataset into working condition latent space coding and fault latent space coding. The unseen working condition space construction module is configured to receive the working condition latent space encoding, perform spatial interpolation operation based on convex hull sampling within the working condition feature distribution boundary spanned by the working condition latent space encoding, and generate several sets of synthetic working condition latent space encodings that do not overlap with the working condition labels of any running sample in the historical running dataset, as the unseen working condition space. The unseen fault mode space construction module is configured to receive the fault latent space code, perform an extrapolation sampling operation based on Gaussian process regression on the fault latent space code, and generate several sets of synthetic fault latent space codes that exceed the boundary range of the fault latent space codes corresponding to all fault mode labels in the historical running dataset, as the unseen fault mode space. The generalization test sample generation module is configured to perform a vector concatenation operation between each synthetic working condition latent space code in the unseen working condition space and each synthetic fault latent space code in the unseen fault mode space to generate a generalization test sample set. The diagnostic model performance evaluation module is configured to deploy the diagnostic model to be validated on the generalization test sample set for batch diagnostic inference, calculate the diagnostic accuracy, diagnostic confidence distribution and diagnostic result stability index of the diagnostic model on the generalization test sample set, and output multi-dimensional generalization performance quantification results. The generalization boundary identification module is configured to receive the multidimensional generalization performance quantization result, back locate the critical sample region in the generalization test sample set whose diagnostic performance is lower than the generalization performance acceptance threshold based on a preset generalization performance acceptance threshold, map the critical sample region back to the working condition feature distribution space and the fault feature distribution space, and output the generalization failure boundary description information of the diagnostic model to be verified. The diagnostic service orchestration module is configured to encapsulate the multidimensional generalization performance quantification results and the generalization failure boundary description information into a generalization capability verification report, and push the generalization capability verification report to the operation and maintenance decision-making end, so that the operation and maintenance decision-making end can perform deployment admission judgment before the diagnostic model to be verified is deployed online.

[0025] Specifically, the implementation of the operating condition feature decoupling module involves constructing a variational autoencoder framework with structural constraints, which decomposes historical operating data into mutually independent operating condition representations and fault representations in the latent representation layer, thereby providing a physically interpretable low-dimensional representation foundation for subsequent construction of the unseen space.

[0026] The system accesses the rail transit vehicle monitoring data service interface through the historical data receiving unit to obtain historical operational datasets containing multi-source observation information. Each operational sample in this dataset corresponds to a unique operating condition label and a fault mode label. The operating condition label describes the characteristics of the operating environment and combination of operating conditions, while the fault mode label characterizes the equipment status category. Before entering the model, the data undergoes unified feature alignment and numerical normalization processing to ensure comparability of features with different dimensions within the same numerical domain, while preserving the original physical meaning.

[0027] In the feature encoding stage, the parallel encoder unit adopts a structure with a shared bottom-level network and decoupled high-level branches. The bottom-level feature extraction network performs a nonlinear mapping on the input running samples to extract general representation features; based on this, it splits into two independent encoding branches to construct the first encoder and the second encoder, respectively. The first encoder outputs the latent space code of the operating condition, and the second encoder outputs the latent space code of the fault. Both adopt a variational inference mechanism to map the input to latent distribution parameters, namely the mean vector and the variance vector, and achieve random sampling through reparameterization, so that the latent variables have a continuously differentiable structure. This expression characterizes how the latent code is generated, where the mean and variance are output by the encoder network, and random variables are used to introduce distribution modeling capabilities, enabling the model to learn the sample distribution rather than a single-point mapping.

[0028] To decouple operating condition features from fault features, the mutual exclusion constraint unit introduces a constraint mechanism based on minimizing mutual information. By constructing a mutual information estimation function between the operating condition latent space encoding and the fault latent space encoding, this function is added as a regularization term to the overall loss function. During training, the degree of information coupling between the two is continuously suppressed, making the two types of latent variables statistically independent. This process can be expressed as minimizing the objective: ,in This represents the implicit space encoding of the operating conditions. This represents fault latent space encoding. The reduction of mutual information means that operating condition information will not be leaked into the fault representation, and vice versa, thus achieving semantic separation at the feature level.

[0029] While decoupling constraints, the reconstruction verification unit introduces a decoder to perform a reverse mapping of the concatenated latent space encoding. The decoder receives the concatenated vector of the operating condition latent space encoding and the fault latent space encoding, reconstructs the original operating sample, and constrains the integrity of the encoding space representation through reconstruction error. The reconstruction loss is typically in the form of mean squared error, ensuring that the output sample maintains consistency with the original sample in statistical features. This mechanism guarantees that the decoupling process does not lead to the loss of key information, allowing the latent space to still possess sufficient expressive power.

[0030] The overall training objective consists of three parts: variational distribution constraints, reconstruction error terms, and mutual information constraints. Through multi-objective joint optimization, the model achieves structural decoupling while maintaining its generative capabilities. After training convergence, the encoding output unit performs forward encoding on all historical running samples, obtaining corresponding latent space encoding sets for operating conditions and fault modes, each forming its own feature distribution. These latent space encodings geometrically constitute two independent latent distribution domains, representing the observed operating condition range and the observed fault mode range, respectively. This result not only provides boundary criteria for the subsequent construction of unseen operating condition and fault mode spaces but also ensures the controllability and interpretability of the combination of the two spaces through the decoupling structure, thereby avoiding interference from the aliasing of different factors on the generalization analysis.

[0031] Specifically, the unseen working condition space construction module is based on the working condition latent space encoding output by the working condition feature decoupling module. It performs geometric modeling and structural expansion on the observed working condition distribution. Its goal is to systematically generate potential working condition expressions that are inside the known distribution but not covered by historical samples without destroying the physical consistency of the original distribution, thereby constructing a representative unseen working condition space.

[0032] In the convex hull boundary calculation unit, the entire latent space encoding set of all working conditions output by the working condition feature decoupling module is first received. This set can be regarded as a discrete point set located in a high-dimensional Euclidean space, with each encoding corresponding to an observed working condition combination. To characterize the overall distribution range of this point set in the latent space, a convex hull construction operation is performed on this point set, and the minimum convex set of all points is defined as the working condition feature distribution boundary. This boundary can be formally represented as the set of convex combinations of all sample points: in Indicates the first Hidden space encoding for each working condition. This represents the corresponding weight coefficient. This expression demonstrates that any point inside the convex hull can be generated by a linearly weighted combination of existing samples, thus ensuring that the generated space does not exceed the physically accessible range of the actual working condition distribution. Through this geometric description, the known working condition region in the latent space is transformed into a closed and continuous polyhedral structure.

[0033] In the spatial sampling strategy unit, a uniform sampling grid is constructed within the convex hull boundary as the constrained region. The sampling grid is constructed based on the equally spaced division of the value range of each dimension, discretizing the latent space into a set of regular grid points. For each candidate sampling point, the Euclidean distance between it and all known working condition latent space codes is calculated to evaluate the similarity between the point and existing samples. The Euclidean distance is defined as follows: ,in Indicates candidate sampling points, This represents known coded points. By setting a uniform distance threshold, candidate sampling points are filtered, retaining only those points whose distance to all known coded points is greater than the threshold. This filtering mechanism ensures that the retained points are located in sparse regions of known samples in the latent space, i.e., regions not fully covered by historical data. This avoids generating redundant samples that highly overlap with existing operating conditions, while ensuring that the generated samples are still within the reasonable operating condition range.

[0034] In the synthetic coding generation unit, the selected boundary interpolation points are directly used as the output of new latent space codes for work conditions. These codes are distributed among known work conditions, but under distance constraints, they maintain significant differences from any single known sample, thus corresponding to new work condition combinations. For each synthetic latent space code, a corresponding virtual work condition label is constructed. This label is not a simple copy of an existing label, but is distinguished based on its relative position in the latent space, explicitly indicating that the work condition corresponding to this code does not have a similar label in the historical running dataset, ensuring its unseen attribute from a semantic perspective. In the unseen work condition space output unit, all synthetic latent space codes and their corresponding virtual work condition labels are associated and stored one-to-one, forming a structured unseen work condition space dataset. This dataset maintains the same representation as the original latent space codes for work conditions, while semantically expanding the coverage of work condition distribution. Since all generated samples are constrained by convex hull boundaries, their distribution will not deviate from the real physical range of work conditions. At the same time, the distance filtering mechanism ensures that they are sufficiently different from existing samples, giving this space good representativeness and discriminability.

[0035] This module is tightly coupled with the operating condition feature decoupling module. The operating condition feature decoupling module decouples operating condition features from fault features through mutual information constraints, ensuring that the operating condition latent space encoding only represents the operating environment and operating condition factors, excluding fault interference information. This guarantees that the convex hull boundary depicts a pure operating condition distribution structure. Based on this, the unseen operating condition space construction module geometrically expands this distribution structure and fills sparse regions, ensuring that the unseen operating condition space maintains both physical consistency and structural diversity, providing a high-quality input foundation for subsequent generalization test sample generation.

[0036] Specifically, the unseen fault mode space construction module, based on the fault latent space encoding obtained by the operating condition feature decoupling module, performs uncertainty-driven extrapolation and expansion of the observed fault distribution. This involves using a probabilistic regression model to characterize the known structures in the latent space and identifying potential novel fault mode regions outside the distribution boundaries, thereby constructing a forward-looking unseen fault mode space. This process complements the unseen operating condition space construction module, which generates unseen operating conditions within the known distribution, while the unseen fault mode space construction module expands unknown faults outside the known distribution.

[0037] In the Gaussian process modeling unit, the complete set of fault latent space codes output by the operating condition feature decoupling module is received. Since the operating condition feature decoupling module decouples operating condition factors from fault factors through mutual information constraints, the current fault latent space codes only reflect fault-related features, allowing them to be considered a pure representation of the fault mode. Based on this, a Gaussian process regression model is constructed using the fault latent space codes as input variables and the corresponding fault mode labels as observed outputs. This model characterizes the correlation between any two points in the input space by defining a covariance function, thereby forming the ability to predict the distribution of unknown points. Its basic form can be expressed as: in Represents the mean function, The kernel function is used to measure the similarity between latent space codes for different faults. By learning from historical data, this model can not only fit the distribution of known fault modes, but also output the predicted mean and predicted variance at any position in the latent space, where the predicted variance reflects the degree of uncertainty at that position.

[0038] In the extrapolation sampling unit, the distribution boundary of the point set composed of all fault latent space codes is first estimated. Its extension range can be determined using a minimum envelope region or a distance-based boundary description method. Based on this, an extended region is constructed outside the boundary, and candidate fault latent space codes are generated within this region. Each candidate code is input into a Gaussian process regression model to obtain the corresponding prediction mean and prediction variance. The variance term of the prediction result is expressed as follows: in Indicates candidate coding points, Represents the covariance matrix of the training samples. This represents the covariance vector between candidate points and training samples. This variance term describes the cognitive uncertainty of candidate points relative to known data. The larger the value, the further the region is from the known sample distribution, and the more likely it corresponds to an unknown failure mode.

[0039] By setting a uniform variance threshold, all candidate codes are screened, and only codes with a prediction variance higher than the threshold are retained as extrapolation sampling points. The essence of this screening mechanism lies in using the uncertainty response of the probabilistic model to actively locate knowledge gaps in the latent space, avoiding the generation of samples that are merely minor perturbations of existing fault modes, while ensuring that the generated samples statistically represent potential novel faults.

[0040] In the virtual fault label generation unit, the selected extrapolated sampling points are used as the output of the synthetic fault latent space encoding. Since these codes are located outside the known distribution boundaries, their corresponding fault features do not have a direct mapping relationship in historical data. Therefore, an independent virtual fault mode label is generated for each code. This label does not overlap with the historical fault label set and is used to identify that the fault mode represented by the code belongs to an unknown category. The label generation process combines the differences in latent space location for differentiation, so that different extrapolation regions correspond to different virtual labels, thereby maintaining structural separability in subsequent analysis. In the unseen fault mode space output unit, all synthetic fault latent space codes and corresponding virtual fault mode labels are associated and stored to form the unseen fault mode space. Structurally, this space represents an extension of the original fault distribution, and semantically, it represents a potential set of new faults. Since all generated points come from high uncertainty regions, this space has strong exploratory and forward-looking characteristics.

[0041] This module forms a hierarchical collaborative relationship with the operating condition feature decoupling module and the unseen operating condition space construction module. The operating condition feature decoupling module provides a decoupled fault latent space, ensuring that Gaussian process modeling is not affected by operating condition factors and guaranteeing the singularity of the modeling object. The unseen operating condition space construction module performs distributed internal expansion along the operating condition dimension, while the unseen fault mode space construction module performs distributed external expansion along the fault dimension. The two are subsequently combined to form a complete generalization test space covering unseen operating conditions, unseen faults, and combinations thereof. Through this structured expansion approach, the unseen fault mode space not only possesses statistical rationality but also can be physically mapped to potential unobserved equipment anomaly patterns, providing crucial support for evaluating the generalization capability of diagnostic models.

[0042] Specifically, in the generalization test sample generation module, based on the unseen operating condition space constructed by the unseen operating condition space construction module and the unseen fault mode space constructed by the unseen fault mode space construction module, the module systematically combines and reverse-maps the two types of decoupled latent spaces. Its goal is to generate standardized data samples that simultaneously contain features of unseen operating conditions and unseen faults, thereby forming a unified test carrier for evaluating the generalization ability of diagnostic models. This module essentially completes the structured reconstruction from the latent space to the original data space while maintaining the traceability of semantic labels.

[0043] In the encoding and splicing unit, all synthetic operating condition latent space encoding sets and all synthetic fault latent space encoding sets are received. Since the operating condition feature decoupling module has achieved statistical independence of the two types of latent variables through mutual information constraints, the operating condition latent space and the fault latent space are semantically independent, allowing them to be freely combined without introducing information conflicts. Based on this, a Cartesian product combination operation is performed on the two encoding sets, establishing a one-to-one pairing relationship between each set of operating condition codes and each set of fault codes. The combination result can be expressed as: in This indicates that the set of codes in the working condition space is not found. This represents the set of codes in the unseen fault mode space. This represents the combined set of codes. For any combination pair, the corresponding latent space codes for operating conditions and faults are sequentially concatenated along the vector dimension to form a unified joint latent space representation: The splicing result retains independent subspaces for the two types of information in its structure, while forming a complete representation of the operating state as a whole, which is equivalent to a potential expression form that includes combinations of new operating conditions and new faults.

[0044] In the decoding and reconstruction unit, each concatenated latent space code is input into the decoder trained in the condition feature decoupling module. This decoder, having learned the nonlinear mapping from the joint latent space to the original data space during training, can thus reconstruct the concatenated code into physically meaningful running samples. The decoding process can be abstractly represented as follows: in This represents the decoder mapping function. This represents the generalized test sample obtained from the reconstruction. Since the training objective of the decoder includes the reconstruction error constraint, this mapping maintains the same data structure and feature distribution as the original running samples in a statistical sense, so that the generated samples are isomorphic to the samples in the historical running dataset in terms of data form, including consistency in feature dimension, numerical range and physical semantics.

[0045] In the sample annotation unit, each generated generalized test sample is annotated with attributes. The annotation information comes from its corresponding latent space composition, namely the virtual operating condition label encoded by the operating condition latent space and the virtual fault mode label encoded by the fault latent space. By jointly annotating the two types of labels, the true generalized attribute annotation information of the sample is formed. This annotation method differs from traditional single-label classification; it is essentially a combined label structure that reflects both unseen operating condition attributes and unseen fault attributes, thus providing an accurate reference for subsequent multi-dimensional performance evaluation.

[0046] In the test sample set output unit, all generated generalized test samples and their corresponding real generalized attribute annotation information are uniformly organized and stored to form a complete generalized test sample set. Structurally, this sample set covers a combined region of unseen operating condition space and unseen fault mode space, and semantically contains a large number of state combinations not found in historical data, exhibiting significant out-of-distribution characteristics. This module forms a closed-loop logical relationship with the operating condition feature decoupling module and the unseen fault mode space construction module. The operating condition feature decoupling module provides decoupled latent representations, allowing different factors to be manipulated independently; the unseen operating condition space construction module generates in-distribution unseen samples in the operating condition dimension, and the unseen fault mode space construction module generates out-of-distribution unknown samples in the fault dimension; while the generalized test sample generation module, through combination and decoding, uniformly maps these two types of extensions into observable data forms. Through this process, the originally abstract latent space extensions are transformed into standardized test data that can be directly input into the diagnostic model, transforming the generalization ability assessment from theoretical analysis into an executable sample-level verification process, significantly enhancing the system's coverage and assessment accuracy for complex unseen scenarios.

[0047] Specifically, in the diagnostic model performance evaluation module, based on the generalization test sample set generated by the generalization test sample generation module, the diagnostic model performance evaluation module performs systematic reasoning evaluation on the diagnostic model to be validated. Its purpose is to characterize the performance of the model in the combination of unseen working conditions and unseen faults through multi-dimensional statistical indicators, thereby expanding the generalization ability from a single accuracy rate to a comprehensive measurement system that includes confidence and stability.

[0048] In the diagnostic model deployment unit, the generalization test sample set output by the generalization test sample generation module is received, and the diagnostic model file to be verified is obtained. Through instantiation, the model is loaded into a unified inference runtime environment, and a data reading channel is established between the sample set and the inference environment. This channel organizes the data flow in batch processing, enabling the model to receive samples generated by the generalization test sample generation module under the same input specifications. Because the generalization test samples maintain consistency with historical runtime data in data structure, the model can directly execute inference without structural adjustments, thus ensuring the comparability of evaluation results.

[0049] In the batch inference execution unit, generalization test samples are loaded batch by batch through the data reading channel. Each sample is input into the diagnostic model, and its inference interface is called to output the diagnostic classification result and the corresponding diagnostic confidence score. The diagnostic confidence score is usually derived from the probability distribution of the model's output layer, which can be expressed as: in The model represents the categories The response value, Indicates that the sample belongs to the category The predicted probability. This probability value reflects the model's confidence in its classification decision and is an important basis for measuring the reliability of generalization.

[0050] In the accuracy calculation unit, the diagnostic classification result of each generalization test sample is compared with its true generalization attribute annotation information. Since the generalization test sample generation module has assigned a virtual fault mode label to each sample, this label serves as an evaluation benchmark to determine whether the prediction result is correct. The overall diagnostic accuracy is obtained by statistically analyzing the proportion of samples whose prediction results match the virtual fault mode labels out of the total number of samples, expressed as follows: in Represents the total number of samples. Indicates the predicted label, The true label represents the model's classification, and the indicator function is used to determine whether the classifications are consistent. This metric reflects the model's overall discriminative ability on unseen distributions.

[0051] In the confidence distribution calculation unit, the diagnostic confidence values ​​of all samples are divided into intervals, mapping continuous probability values ​​to several discrete confidence level intervals. Within each confidence interval, the number of samples and the number of correctly predicted samples are counted, thus constructing a joint distribution relationship between confidence and accuracy. This distribution is used to analyze the consistency between the model's confidence output and the actual accuracy, revealing whether the model is overconfident or underconfident. If a high confidence interval corresponds to a low accuracy, it indicates that the model has a confidence distortion problem in unseen scenarios.

[0052] In the stability index calculation unit, the volatility of model performance is evaluated through a repeated sampling mechanism. Specifically, a fixed-size subset is randomly selected from the generalization test sample set, and this process is repeated independently multiple times, calculating the diagnostic accuracy for each subset. The stability index is obtained by statistically analyzing the variance of the accuracies across all subsets, and its expression is as follows: in Indicates the number of samples. Indicates the first Accuracy of subsampled subset This represents the average accuracy. This metric reflects the model's consistent performance across different sample subsets; a larger variance indicates that the model is more sensitive to input distribution perturbations and has poorer generalization stability.

[0053] In the generalization performance quantification output unit, diagnostic accuracy, confidence level, joint distribution of accuracy, and stability indicators are uniformly and structurally encapsulated to form a multidimensional generalization performance quantification result. This result not only describes the model's average performance on unseen samples but also reveals its decision confidence mechanism and performance fluctuation characteristics, characterizing the model's generalization ability from multiple dimensions. This module, together with the operating condition feature decoupling module and the generalization test sample generation module, forms a complete evaluation chain. The operating condition feature decoupling module constructs a controllable latent space through decoupling; the unseen operating condition space construction module and the unseen fault mode space construction module expand the operating condition and fault distributions, respectively; the generalization test sample generation module maps the expanded results to observable samples; and the diagnostic model performance evaluation module performs real inference evaluation on these samples. Through this continuous process, the model no longer relies on identically distributed test data for validation but is tested on unseen distributions constructed by the system, thereby achieving a quantitative characterization of generalization ability and providing a data foundation for subsequent generalization boundary identification.

[0054] Specifically, in the generalization boundary identification module, based on the multidimensional generalization performance quantification results output by the diagnostic model performance evaluation module, the module reverse-maps the statistically significant performance degradation phenomenon to the latent space structure, thereby transforming performance failure into a spatial boundary and achieving an interpretable characterization of the generalization limit of the diagnostic model. The essence of this module is to establish a mapping relationship from sample-level performance indicators to the distribution region of latent features, so that model failure is no longer an abstract conclusion, but a spatial region description with clear operating conditions and fault semantics.

[0055] In the performance threshold determination unit, the multidimensional generalization performance quantification results output by the diagnostic model performance evaluation module are received, and diagnostic accuracy and diagnostic result stability indicators are extracted from them. Samples are then screened based on a preset generalization performance acceptance threshold system. This threshold system includes a lower accuracy threshold and an upper stability threshold, which respectively constrain the model's discriminative ability and result volatility. For any generalization test sample, if its corresponding prediction result exhibits low accuracy or high volatility in a locally statistical sense, it is determined that it does not meet the generalization performance requirements. This determination process can be abstractly represented as follows: in Represents the critical sample set. and These represent the accuracy threshold and the stability threshold, respectively. Through this screening, all samples falling within the model performance failure range are uniformly marked as critical samples, providing a basis for subsequent spatial localization.

[0056] In the critical sample localization unit, the true generalization attribute annotation information of each critical sample is extracted, including the corresponding virtual operating condition label and virtual fault mode label. By statistically analyzing the co-occurrence relationships of these labels in the critical sample set, a co-occurrence frequency matrix between operating condition labels and fault labels is constructed. Each element of this matrix represents the number of times a combination of a certain operating condition label and a certain fault label appears in the critical sample, expressed in the following form: in and They represent the first Each sample is assigned a working condition label and a fault label. By setting a frequency threshold, frequently occurring label combinations are selected and defined as failure label combinations. This process achieves aggregation from single-sample failures to pattern-level failures, making subsequent boundary identification statistically stable.

[0057] In the working condition space boundary mapping unit, all synthetic working condition latent space codes and their corresponding virtual working condition labels output by the unseen working condition space construction module are received. Based on the failure label combinations, all coded points belonging to the failure working condition label set are selected, and minimum convex hull calculation is performed on these points in the working condition latent space. This convex hull serves as the generalized failure boundary in the working condition dimension, and its mathematical form still satisfies the convex combination constraint: in This represents the set of codes belonging to the failure condition label. This boundary characterizes the range of operating condition combinations within which the model begins to exhibit systematic performance degradation.

[0058] In the fault space boundary mapping unit, operations symmetrical to those in the operating condition space are performed. It receives all synthetic fault latent space codes and their corresponding virtual fault mode labels output by the unseen fault mode space construction module, filters out coding points belonging to the failure label combinations, and constructs a minimum convex hull in the fault latent space. This convex hull describes the range of unknown fault modes in which the model cannot maintain stable discrimination capability, thus forming a generalized failure boundary in the fault dimension. In the boundary description information generation unit, the two latent space boundaries are semantically mapped. Specifically, the boundary regions in the operating condition latent space are converted into interpretable physical operating condition intervals through decoding mapping or feature back-inference methods, such as a combination of line condition ranges, load level intervals, and ambient temperature intervals; simultaneously, the boundary regions in the fault latent space are converted into fault feature mode descriptions, such as abnormal vibration feature intervals or electrical feature offset ranges. By structurally combining these two types of descriptions, complete generalized failure boundary description information is generated.

[0059] This descriptive information is no longer confined to an abstract space but transformed into an engineering semantic expression that can be understood by operations and maintenance personnel. It clearly indicates under what operating conditions and what combination of failure modes the model faces the risk of performance failure. The overall module and the operating condition feature decoupling module form a progressive relationship to the diagnostic model performance evaluation module. The operating condition feature decoupling module provides the decoupling space, the unseen operating condition space construction module and the unseen failure mode space construction module construct the extended distribution, the generalization test sample generation module generates test samples, the diagnostic model performance evaluation module outputs performance results, and the generalization boundary identification module completes the inversion from performance results to spatial boundaries, realizing the transformation from data-driven evaluation to structured cognition, providing a direct basis for subsequent model optimization and deployment decisions.

[0060] Specifically, in the diagnostic service orchestration module, based on the multidimensional generalization performance quantification results formed by the diagnostic model performance evaluation module and the generalization failure boundary description information formed by the generalization boundary identification module, the diagnostic service orchestration module completes the structured transformation from technical evaluation results to engineering decision-making basis. It integrates the model generalization capability from scattered indicators into a unified report and embeds it into the operation and maintenance decision-making process through standardized service interfaces to realize a controllable access mechanism before the model goes online.

[0061] The report encapsulation unit receives the multidimensional generalization performance quantification results output by the diagnostic model performance evaluation module and the generalization failure boundary description information output by the generalization boundary identification module. The multidimensional generalization performance quantification results include diagnostic accuracy, confidence and accuracy joint distribution, and stability indicators. These indicators characterize the model's capabilities from three dimensions: overall performance, decision credibility, and result volatility. The generalization failure boundary description information provides the operating condition range and failure mode range corresponding to performance degradation. Through a preset report template, the above information is uniformly structured and filled in, aligning data from different sources within the same semantic framework. The report adopts a hierarchical structure: the upper layer is an overall performance summary, the middle layer is statistical distribution and stability analysis, and the lower layer is a spatial description of the generalization failure boundary, thus forming a complete information chain from result to cause.

[0062] During report generation, key indicators are standardized to ensure comparability across different dimensions. A normalization function can be used to unify the scaling of each indicator. in This represents the original index value. This indicates the normalized result. Through this processing, different performance indicators in the report can be comprehensively analyzed under a unified evaluation framework, which is beneficial for the operation and maintenance side to make an overall judgment.

[0063] In the report delivery unit, the generated generalization capability verification report is encapsulated in a structured data format, and an encrypted communication link is established between the report and the operations and maintenance decision-making end. The report maintains data integrity and consistency during transmission, enabling the operations and maintenance system to directly parse the report content and perform automated processing. After transmission is complete, a confirmation command is sent to the operations and maintenance decision-making end to ensure that the report has been successfully received and entered into the decision-making process. This process transforms model evaluation results from offline analysis into an online service capability that can be embedded into business systems.

[0064] The access determination execution unit receives decision instructions generated by the operations and maintenance decision-making end based on the report content. The operations and maintenance decision-making process can essentially be abstracted as a decision function based on multiple indicator constraints: in Indicates whether the deployment conditions are met. and These represent the accuracy and stability thresholds, respectively. When both the overall model performance and stability meet the preset requirements, a deployment acceptance instruction is output; otherwise, a deployment rejection instruction is output.

[0065] Upon receiving a deployment access instruction, the system generates a model deployment license credential and transmits it to the model deployment executor, enabling the model to enter the actual operating environment. This credential serves as the unique authorization identifier for model deployment, ensuring controllable management of the model deployment process. Upon receiving a deployment rejection instruction, the system does not directly terminate the process. Instead, it extracts and reorganizes the generalization failure boundary description information from the generalization capability verification report to form targeted model improvement suggestions. These suggestions are fed back to the model training end in the form of operating condition ranges and failure mode ranges, indicating in which data regions the current model has capability gaps. Based on this, the model training end performs training sample augmentation or model structure optimization, ensuring that subsequent training processes have clear goal orientation rather than blind adjustments.

[0066] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A data service platform for intelligent diagnostics of rail transit, characterized in that, include: The working condition feature decoupling module is configured to acquire the historical operation dataset of rail transit vehicles. The historical operation dataset includes multiple working condition labels and multiple fault mode labels. The variational autoencoder is used to decouple each operation sample in the historical operation dataset into working condition latent space coding and fault latent space coding. The unseen working condition space construction module is configured to receive the working condition latent space encoding, perform spatial interpolation operation based on convex hull sampling within the working condition feature distribution boundary spanned by the working condition latent space encoding, and generate several sets of synthetic working condition latent space encodings that do not overlap with the working condition labels of any running sample in the historical running dataset, as the unseen working condition space. The unseen fault mode space construction module is configured to receive the fault latent space code, perform an extrapolation sampling operation based on Gaussian process regression on the fault latent space code, and generate several sets of synthetic fault latent space codes that exceed the boundary range of the fault latent space codes corresponding to all fault mode labels in the historical running dataset, as the unseen fault mode space. The generalization test sample generation module is configured to perform a vector concatenation operation between each synthetic working condition latent space code in the unseen working condition space and each synthetic fault latent space code in the unseen fault mode space to generate a generalization test sample set. The diagnostic model performance evaluation module is configured to deploy the diagnostic model to be validated on the generalization test sample set for batch diagnostic inference, calculate the diagnostic accuracy, diagnostic confidence distribution and diagnostic result stability index of the diagnostic model on the generalization test sample set, and output multi-dimensional generalization performance quantification results. The generalization boundary identification module is configured to receive the multidimensional generalization performance quantization result, back locate the critical sample region in the generalization test sample set whose diagnostic performance is lower than the generalization performance acceptance threshold based on a preset generalization performance acceptance threshold, map the critical sample region back to the working condition feature distribution space and the fault feature distribution space, and output the generalization failure boundary description information of the diagnostic model to be verified. The diagnostic service orchestration module is configured to encapsulate the multidimensional generalization performance quantification results and the generalization failure boundary description information into a generalization capability verification report, and push the generalization capability verification report to the operation and maintenance decision-making end, so that the operation and maintenance decision-making end can perform deployment admission judgment before the diagnostic model to be verified is deployed online.

2. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The operating condition feature decoupling module includes: a historical data receiving unit, configured to access the rail transit vehicle monitoring data service interface to obtain the historical operating dataset. Each operating sample in the historical operating dataset is pre-labeled with a corresponding operating condition label and a corresponding fault mode label. The operating condition label is used to identify the combination category of line conditions, load status and ambient temperature during vehicle operation, and the fault mode label is used to identify the type of fault that occurs in the vehicle components. The parallel encoder unit is configured to receive the historical running dataset, use a first encoder to map each running sample to the working condition latent space code, and use a second encoder to map the same running sample to the fault latent space code. The first encoder and the second encoder share the underlying feature extraction network but are decoupled at the high-level network branches. The mutual exclusion constraint unit is configured to calculate the mutual information between the working condition latent space code and the fault latent space code, and backpropagate to update the network parameters of the first encoder and the second encoder with minimizing the mutual information as a constraint condition, thereby forcing the working condition latent space code and the fault latent space code to be orthogonal to each other in the feature dimension. The reconstruction verification unit is configured to concatenate the working condition latent space encoding and the fault latent space encoding execution vector and input them into the decoder. The decoder outputs the reconstructed running sample and uses the reconstruction error between the reconstructed running sample and the running sample as part of the training loss to feed back to the parallel encoder unit and the decoder. The encoding output unit is configured to receive all running samples in the historical running dataset after the parallel encoder unit has converged during training, output the working condition latent space code and the fault latent space code corresponding to each running sample, and transmit the feature distribution boundary formed by all the working condition latent space codes to the unseen working condition space construction module, and transmit the feature distribution boundary formed by all the fault latent space codes to the unseen fault mode space construction module.

3. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The unseen working condition space construction module includes: a convex hull boundary calculation unit, configured to receive all the working condition latent space codes output by the working condition feature decoupling module, calculate the convex hull boundary of the point set formed by all the working condition latent space codes in the working condition latent space dimension where the working condition latent space codes are located, and use the convex hull boundary as the geometric description of the known working condition distribution boundary. The spatial sampling strategy unit is configured to establish a uniform sampling grid within the spatial region enclosed by the convex hull boundary, traverse each candidate sampling point in the uniform sampling grid, calculate the Euclidean distance between each candidate sampling point and each known point in the working condition latent space encoding, and select the candidate sampling points whose Euclidean distances to all known points are greater than a preset distance threshold as boundary interpolation points. The synthetic coding generation unit is configured to obtain the boundary interpolation points, output each boundary interpolation point as a set of synthetic working condition latent space codes, and generate a virtual working condition label corresponding to each synthetic working condition latent space code. The virtual working condition label is used to identify that the working condition corresponding to the synthetic working condition latent space code does not have the same type of working condition label in the historical working dataset. The unseen working condition space output unit is configured to associate and store all the synthetic working condition latent space codes with all the virtual working condition labels to form the unseen working condition space, and transmit the unseen working condition space to the generalization test sample generation module.

4. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The unseen fault mode space construction module includes a Gaussian process modeling unit, configured to receive all the fault latent space codes output by the working condition feature decoupling module, take the fault latent space dimension where the fault latent space code is located as the input space, and take the fault mode label corresponding to each fault latent space code as the observation value to construct a Gaussian process regression model. The Gaussian process regression model is used to describe the nonlinear mapping relationship between the fault latent space code and the fault mode label and the prediction uncertainty at the location of the fault latent space code. The extrapolation sampling unit is configured to obtain the feature distribution boundary of the fault latent space code in the fault latent space dimension, generate several sets of candidate fault latent space codes in the extended region extending outward from the feature distribution boundary, input each set of candidate fault latent space codes into the Gaussian process regression model, calculate the prediction mean and prediction variance corresponding to each set of candidate fault latent space codes, and select the candidate fault latent space codes whose prediction variance is greater than a preset variance threshold as extrapolation sampling points. The virtual fault label generation unit is configured to acquire the extrapolated sampling points, output each extrapolated sampling point as a group of synthetic fault latent space codes, and generate a corresponding virtual fault mode label for each group of synthetic fault latent space codes. The virtual fault mode label is used to identify that the fault mode corresponding to the synthetic fault latent space code does not have the same type of fault mode label in the historical running dataset, and the virtual fault mode label is different from all the fault mode labels in the historical running dataset. The unseen fault mode space output unit is configured to associate and store all the synthetic fault latent space codes with all the virtual fault mode labels to form the unseen fault mode space, and transmit the unseen fault mode space to the generalization test sample generation module.

5. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The generalization test sample generation module includes: an encoding splicing unit, configured to receive all the synthetic working condition latent space codes output by the unseen working condition space construction module and all the synthetic fault latent space codes output by the unseen fault mode space construction module, perform a Cartesian product combination operation on each group of synthetic working condition latent space codes and each group of synthetic fault latent space codes, and sequentially splice the synthetic working condition latent space codes and the synthetic fault latent space codes in each combination generated by the Cartesian product combination operation in the vector dimension to generate a spliced ​​latent space code with a working condition-fault joint latent space expression; The decoding and reconstruction unit is configured to input each of the concatenated latent space codes into the decoder included in the working condition feature decoupling module, and the decoder maps the concatenated latent space codes back to the original data space to generate a generalized test sample corresponding to the concatenated latent space code. The generalized test sample is consistent with the running sample in the historical running dataset in terms of data form. The sample labeling unit is configured to obtain the virtual operating condition label of the synthetic operating condition latent space encoding and the virtual fault mode label of the synthetic fault latent space encoding corresponding to each generalization test sample, and to jointly label the virtual operating condition label and the virtual fault mode label as the real generalization attribute labeling information of the generalization test sample. The test sample set output unit is configured to associate and store all the generalized test samples with the real generalized attribute annotation information corresponding to all the generalized test samples to form the generalized test sample set, and transmit the generalized test sample set to the diagnostic model performance evaluation module.

6. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The diagnostic model performance evaluation module includes: a diagnostic model deployment unit, configured to receive the generalized test sample set output by the generalized test sample generation module, obtain the model file of the diagnostic model to be verified, instantiate the diagnostic model to be verified and load it into the inference runtime environment, and establish a data reading channel between the inference runtime environment and the generalized test sample set. The batch inference execution unit is configured to sequentially read each generalized test sample in the generalized test sample set through the data reading channel, input each generalized test sample into the diagnostic model to be verified, call the inference interface of the diagnostic model to be verified to perform diagnostic classification operation on each generalized test sample, and output the diagnostic classification result and diagnostic confidence value corresponding to each generalized test sample. The accuracy calculation unit is configured to obtain the diagnostic classification result and the real generalization attribute annotation information corresponding to each generalization test sample, compare the diagnostic classification result with the virtual fault mode label in the real generalization attribute annotation information, and count the proportion of the number of samples that match the comparison in the generalization test sample set to the total number of samples as the diagnostic accuracy. The confidence distribution calculation unit is configured to obtain the diagnostic confidence value corresponding to each generalization test sample, divide all the diagnostic confidence values ​​into several confidence levels according to the value range, statistically analyze the distribution of the number of generalization test samples in each confidence level and the distribution of the number of samples that match the diagnostic classification result with the real generalization attribute labeling information, and output a confidence-accuracy joint distribution histogram. The stability index calculation unit is configured to perform several repeated sampling operations on the generalization test sample set. Each repeated sampling operation randomly selects a fixed number of samples from the generalization test sample set to form a repeated sampling subset. The unit calculates the subset diagnostic accuracy corresponding to each repeated sampling subset and calculates the variance of the diagnostic accuracy of all subsets as the stability index of the diagnostic result. The generalization performance quantification result output unit is configured to encapsulate the diagnostic accuracy, the confidence-accuracy joint distribution histogram, and the diagnostic result stability index in a structured manner to form the multidimensional generalization performance quantification result, and transmit the multidimensional generalization performance quantification result to the generalization boundary identification module and the diagnostic service orchestration module.

7. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The generalization boundary identification module includes: a performance threshold determination unit, configured to receive the multidimensional generalization performance quantification result output by the diagnostic model performance evaluation module, extract the diagnostic accuracy and the diagnostic result stability index from the multidimensional generalization performance quantification result, obtain a preset generalization performance acceptance threshold, the generalization performance acceptance threshold including a lower accuracy threshold and a higher stability threshold, and mark the generalization test samples in the generalization test sample set whose diagnostic accuracy is lower than the lower accuracy threshold or whose diagnostic result stability index is higher than the higher stability threshold as critical samples; The critical sample localization unit is configured to acquire the real generalized attribute annotation information corresponding to each critical sample, extract the virtual operating condition label and the virtual fault mode label corresponding to each critical sample from the real generalized attribute annotation information, establish a co-occurrence frequency matrix of the virtual operating condition label and the virtual fault mode label in the critical sample set, and mark the combination of the virtual operating condition label and the virtual fault mode label with a frequency exceeding a preset frequency threshold in the co-occurrence frequency matrix as a failure label combination; The working condition space boundary mapping unit is configured to receive all the synthetic working condition latent space codes and their corresponding virtual working condition labels output by the unseen working condition space construction module, filter out all the synthetic working condition latent space codes of the virtual working condition labels that belong to the virtual working condition labels in the failure label combination, calculate the minimum convex hull boundary of all the filtered synthetic working condition latent space codes in the working condition latent space dimension, and use the minimum convex hull boundary as the working condition generalization failure boundary. The fault space boundary mapping unit is configured to receive all the synthetic fault latent space codes and their corresponding virtual fault mode labels output by the unseen fault mode space construction module, filter out all the synthetic fault latent space codes of the virtual fault mode labels belonging to the virtual fault mode labels in the failure label combination, calculate the minimum convex hull boundary of all the filtered synthetic fault latent space codes in the fault latent space dimension, and use the minimum convex hull boundary as the fault generalization failure boundary. The boundary description information generation unit is configured to obtain the generalized failure boundary of the operating condition and the generalized failure boundary of the fault, convert the operating condition feature value range corresponding to the generalized failure boundary of the operating condition into rail transit physical operating condition description text, convert the fault feature value range corresponding to the generalized failure boundary of the fault into rail transit fault mode description text, perform a structured combination of the rail transit physical operating condition description text and the rail transit fault mode description text to generate the generalized failure boundary description information, and transmit the generalized failure boundary description information to the diagnostic service orchestration module.

8. The intelligent diagnostic data service platform for rail transit according to claim 1, characterized in that, The diagnostic service orchestration module includes a report encapsulation unit, configured to receive the multidimensional generalization performance quantification result output by the diagnostic model performance evaluation module and the generalization failure boundary description information output by the generalization boundary identification module, and to structurally fill the diagnostic accuracy, the confidence-accuracy joint distribution histogram, and the diagnostic result stability index in the multidimensional generalization performance quantification result with the rail transit physical condition description text and the rail transit fault mode description text in the generalization failure boundary description information according to a preset report template to generate a generalization capability verification report; The report push unit is configured to obtain the generalization capability verification report, establish an encrypted communication link with the operation and maintenance decision-making end, transmit the generalization capability verification report to the operation and maintenance decision-making end in structured data format through the encrypted communication link, and send a report reception confirmation instruction to the operation and maintenance decision-making end after the transmission is completed. The admission determination execution unit is configured to receive a deployment admission instruction or a deployment rejection instruction from the operation and maintenance decision-making end based on the generalization capability verification report. When the deployment admission instruction is received, the unit generates a deployment license credential for the diagnostic model to be verified and outputs the deployment license credential to the model deployment executor. When the deployment rejection instruction is received, the unit encapsulates the generalization failure boundary description information in the generalization capability verification report into a model improvement suggestion and feeds it back to the model training end for the model training end to perform training sample amplification or model structure optimization operations.

Images