An Unsupervised Self-Learning Column Inspection Algorithm

Abstract

This invention discloses an unsupervised self-learning train inspection algorithm, belonging to the field of railway vehicle fault detection technology. It utilizes the periodicity of vehicle operation to construct a three-dimensional baseline of time, route, and vehicle, distinguishing between fluctuations in operating conditions and genuine faults. A periodically prior-guided distributed decoding network is employed, using difference constraints to limit the reconstructed output to a normal distribution, avoiding missed detections or misjudgments. An out-of-distribution perception module is set up, integrating latent density, reconstruction difference probability, and multi-branch consistency to actively identify unknown samples, triggering incremental learning based on prototype replay and elastic weight consolidation, achieving unlabeled continuous accumulation of new faults without forgetting old knowledge. Image, vibration, temperature, and semantic metadata are fused, and cross-modal attention is used to enhance robustness. A lightweight network is deployed at edge nodes for online inference; abnormal and unknown samples are reported to the data center, and the baseline is periodically reconstructed and the model updated before being redeployed, forming a self-learning closed loop.

Description

Technical Field

[0001] This invention relates to the field of railway vehicle fault detection technology, and in particular to an unsupervised self-learning train inspection algorithm. Background Technology

[0002] Currently, railway vehicle fault trackside image detection systems (such as TFDS) are widely deployed, and automated detection methods based on deep learning are gradually being applied. Existing technologies mainly fall into three categories: The first category is unsupervised anomaly detection based on reconstruction paradigms, using autoencoders or generators trained only on normal images to locate anomalies through reconstruction differences. However, this type of method has a fundamental flaw—the model tends to "perfectly reconstruct" anomalous regions into normal shapes, leading to the missed detection of minor faults such as small cracks or loose bolts; conversely, if the anomaly is too significant, the entire image may be misjudged as an anomaly, making accurate localization impossible. The second category adds artificial noise (such as Gaussian white noise) to the feature layer to simulate anomalies, thereby training the model's sensitivity to anomalies. However, the distribution of artificial noise differs significantly from that of real faults, making it difficult to cover the complex and ever-changing fault patterns at railway sites. The third category is methods based on meta-learning or few-shot learning, but these still require a certain number of labeled samples, and frequent retraining or manual intervention is needed when new faults emerge, lacking continuous self-learning capabilities. The aforementioned technologies share the following core problems: they fail to utilize the inherent operational periodicity of railway vehicles (the same vehicle operates under similar conditions when running repeatedly on the same route), making it impossible to distinguish between normal fluctuations in weather, load, speed, etc., and real faults; they lack the ability to judge their own cognitive boundaries, meaning the model "does not know what it does not know," and cannot actively trigger incremental learning; the self-learning process often falls into catastrophic forgetting, that is, forgetting old knowledge after learning new fault types; and detection relies on single-modal data, resulting in insufficient robustness when faced with image blurring, changes in lighting, or sensor noise. Summary of the Invention

[0003] The main objective of this invention is to provide an unsupervised self-learning column inspection algorithm that can effectively solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] An unsupervised self-learning column inspection algorithm includes the following steps:

[0006] Step S1: Obtain railway vehicle operation and route information, construct a three-dimensional normal baseline based on the periodicity of vehicle operation, the three-dimensional normal baseline includes time dimension baseline, route dimension baseline and vehicle dimension baseline, and automatically classify historical detection data through clustering method to filter normal operation status for baseline modeling;

[0007] Step S2: Construct an anomaly detection network with distributed decoding. The network adopts a dual-path encoder-decoder structure with physical information enhancement. It takes running intersection coding and time-mileage information as conditional inputs. The encoder end is equipped with a feature decoupling module. The decoder receives periodic prior guidance to generate a reconstruction output within the normal distribution. It controls the difference between the reconstruction result and the input through a difference constraint loss function. During inference, it generates anomaly scores based on reconstruction differences and dynamically calculates anomaly thresholds.

[0008] Step S3: Construct an out-of-distribution sensing module. By integrating latent spatial density estimation, reconstructing the probability of difference distribution, and multi-branch consistency difference, it determines whether the sample exceeds the scope of learned knowledge. When it is determined to be an out-of-distribution unknown sample, it triggers active incremental learning. The active incremental learning adopts a continuous learning framework based on prototype playback and uses experience playback and parameter regularization strategies to update the model.

[0009] Step S4: Fuse multi-source heterogeneous data, including image data, vibration or acoustic data, temperature data and semantic metadata, perform feature alignment and fusion through a cross-modal attention mechanism, and inject semantic metadata as conditional input into the detection network;

[0010] Step S5: Deploy the detection network on the trackside edge computing node for online inference, outputting three types of results: normal, abnormal, or unknown outside the distribution. Abnormal results and out-of-distribution samples are reported to the data center. The data center periodically performs the incremental model update in step S2 and the baseline reconstruction in step S1, generates a new model version and distributes it to the edge node, forming a self-learning closed loop.

[0011] Preferably, in step S1, the time dimension baseline is used to capture the gradual aging trend of components over time, the route dimension baseline uses the consistency of operating patterns within the same route to filter random disturbances, and the vehicle dimension baseline establishes a healthy reference distribution based on group data of similar components of the same model.

[0012] Preferably, in step S2, the difference constraint loss function includes reconstruction loss, difference upper limit constraint loss, feature distribution alignment loss, and periodic consistency loss, wherein the periodic consistency loss is used to ensure that the reconstruction result maintains the same running path semantics as the input.

[0013] Preferably, in step S3, the potential spatial density estimation adopts a Gaussian mixture model or normalized flow modeling, and the tail behavior of the differential distribution is reconstructed using extreme value theory modeling; the prototype playback buffer stores the core embedding of normal samples, the prototype of the identified fault type, and the temporary embedding of out-of-distribution samples. When the out-of-distribution samples accumulate to a preset number and are confirmed as a new fault type by manual review, their prototypes are stably written into the buffer.

[0014] Preferably, in step S4, when a certain type of sensor data is missing, the network completes the information through cross-modal inference.

[0015] Preferably, in step S5, the detection network deployed on the edge nodes is a lightweight model after pruning and quantization; when the data center performs incremental updates, it adopts a strategy combining elastic weight consolidation and knowledge distillation to prevent catastrophic forgetting.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] This invention utilizes the periodicity of railway vehicle operation to construct a three-dimensional normal baseline of time-routes-vehicles, decoupling normal fluctuations in operating conditions from actual faults, thereby reducing false alarms caused by changes in operating conditions. It employs a periodically prior-guided distributed decoding network, using difference constraints to control the reconstruction output within a normal distribution, avoiding the over-matching of normal patterns by traditional reconstruction paradigms that may miss minor faults or fail to locate anomalies. An out-of-distribution sensing module is introduced, integrating latent spatial density, reconstruction difference probability, and multi-branch consistency to identify unknown samples and triggering incremental learning based on prototype playback and elastic weight consolidation. This allows the model to continuously accumulate new fault knowledge without forgetting old knowledge, even without annotation. Finally, it fuses image, vibration, temperature, and semantic metadata, utilizing cross-modal attention enhancement features to improve detection robustness when a single sensor fails. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the algorithm flow of the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] like Figure 1 As shown, during the preparation phase, the system obtains the target vehicle's route information from the railway transportation management information system or train operation monitoring device, including the vehicle's unique number, the train number and route code, the order of stations and planned arrival and departure times, and cumulative mileage, among other semantic metadata. Based on this information, an operational identity is established for each vehicle. Simultaneously, multiple batches of detection data generated by the vehicle's historical trackside detection system (such as a freight car fault trackside image detection system) are collected, forming a dataset indexed by vehicle and arranged chronologically.

[0021] Leveraging the inherent operational cyclicality of railway vehicles—that is, when the same vehicle repeatedly runs on the same route, the detection points it passes through, its operating speed range, load conditions, and ambient lighting are highly consistent—the system constructs a three-dimensional normal baseline. The first dimension is the time baseline: for multiple batches of data collected from the same vehicle at the same detection point, time-series models (such as exponentially weighted moving averages or Gaussian process regression) are used to characterize the long-term evolution trend of each component's features, thereby distinguishing between slow wear and aging processes and sudden fault states. The second dimension is the route baseline: statistical modeling is performed on the detection data of the same vehicle on the same route section, using the consistency of operating patterns to filter out random disturbances caused by speed fluctuations, load changes, or slight weather differences, avoiding misjudging these normal operating condition changes as abnormalities. The third dimension is the vehicle baseline: a reference benchmark is established based on health data of the same vehicle type and component group (e.g., the characteristic distribution of vehicles with recent no-fault repair reports), used to cross-validate whether individual vehicles deviate from the group's normal range. Before constructing the baseline, the system uses a multi-stage clustering method (first using density-based clustering to remove outliers, and then using partitioning clustering to subdivide the remaining data) to automatically classify the unlabeled historical data, filter out the data in "normal operation status", exclude samples caused by environmental interference or suspected failures, and ensure that the baseline is supported only by healthy data.

[0022] In the construction and training phase of the anomaly detection network, this invention employs a physically-enhanced dual-path encoder-decoder network, referred to as the distributed intra-decoder network. This network comprises two parallel encoder branches: the path network takes the running route encoding as conditional input, mapping normal samples belonging to the same route or vehicle to a unified embedding space and imposing contrastive learning constraints to bring similar samples closer together; the time network takes timestamps and cumulative mileage as conditional input, encoding the long-term temporal features of the same component to capture natural aging trends. At the encoder's end, a mutual information minimization decorrelation module is set up. By imposing decorrelation constraints on each feature channel (e.g., using covariance regularization or the Hilbert-Schmidt independence criterion), different feature dimensions express mutually independent fault modes, avoiding false detections or missed detections caused by feature coupling between multiple faults.

[0023] The core design of the decoder differs from the traditional "perfect reconstruction" paradigm. Traditional methods aim to reconstruct the input image or signal as accurately as possible. When the input contains anomalies, the model tends to reconstruct the anomalous parts as normal, leading to missed detections. If the anomalies are too significant, the reconstruction error of the entire image becomes too large, resulting in global misjudgments. To address this issue, the decoder of this invention receives the periodic prior established in step S1 as a guiding signal. Its goal is not to indiscriminately reconstruct the input, but to map the latent feature vectors of the input onto a normal distribution manifold, generating a reconstructed output that is semantically consistent with the input but has corrected for anomaly deviations. To achieve this goal, a combined loss function is used during training: reconstruction loss ensures basic reconstruction capability; difference upper bound constraint penalizes the difference between the reconstructed result and the input exceeding a preset upper bound, preventing the model from being overly sensitive to fluctuations under normal operating conditions; feature distribution alignment loss promotes the latent vectors to follow a standard normal distribution to facilitate generalization; periodic consistency loss ensures that the reconstructed samples and the original samples maintain consistency in intersection semantic features, meaning the reconstruction result cannot lose key contextual information such as the running intersection and time position. Through this difference constraint mechanism, the network learns to strike a balance between "preserving semantics" and "correcting anomalies," enabling minor faults to accumulate into detectable differences while preventing normal fluctuations in operating conditions from being excessively amplified.

[0024] During inference, for any input sample, the network outputs its in-distribution reconstruction result and calculates residuals pixel-by-pixel or feature-by-feature to obtain a difference map. Combining residual neighborhood consistency (i.e., the continuity and abruptness of differences within local regions) and multi-scale feature distances, a component-level anomaly heatmap is generated. Anomaly thresholds are determined dynamically: based on the difference distribution of historical normal data at each detection point, a kernel density estimation is used to fit the probability density function, and then, combined with a small amount of prior knowledge (such as manually set safety margins), an adaptive threshold for each detection point and each component type is calculated. This threshold is updated periodically as operational data accumulates.

[0025] Step S3 implements out-of-distribution sensing and proactive incremental learning. At the encoder end, the system constructs an out-of-distribution scorer, which integrates three signals: First, it models the density of the latent vector output by the encoder in the embedding space using a Gaussian mixture model or normalized flow; the lower the density, the more likely the sample is to exceed the learned distribution. Second, it uses extreme value theory to analyze the tail behavior of the decoder's reconstruction of differences within the normal distribution; rare high difference values ​​suggest possible out-of-distribution scenarios. Third, it calculates the consistency between outputs at different scales or between different network branches; the greater the difference between branches, the more likely the sample belongs to an unknown mode. When the comprehensive score exceeds a preset threshold, the model determines the current sample as an "out-of-distribution unknown sample," meaning that the sample contains structures that the model cannot currently interpret—it could be a new fault type that has not yet been learned, or it could be an extreme condition or sensor anomaly. At this point, the model does not directly assign a fault category but triggers the proactive incremental learning mechanism.

[0026] Incremental learning employs a continuous learning framework based on prototype replay. The system maintains a fixed-capacity prototype replay buffer, storing three types of prototypes: core embeddings of normal samples (centroids representing health status), confirmed fault type prototypes (one or a set of embedding prototypes for each known fault), and temporary embeddings of out-of-distribution samples. When out-of-distribution samples accumulate to a preset number (e.g., 20 to 50 similar samples) and are confirmed as new fault types through manual inspection or remote verification, the prototypes for that sample set are stably written into the buffer. During incremental training, the system extracts historical prototypes from the buffer and mixes them with newly collected samples to form training batches. Elastic Weight Consolidation (EWC) is used to apply secondary constraints to parameters in the model that are important for the old task, and knowledge distillation loss is used to maintain the stability between the new and old model outputs. This approach learns the ability to distinguish new fault types while avoiding catastrophic forgetting. The incremental learning process is executed asynchronously in the background, and the updated model weights are validated and then distributed to each edge detection node via the model management service.

[0027] Step S4 involves the fusion of multimodal data. At the data acquisition front end, the system simultaneously acquires data from multiple sources: visible light or infrared images (covering key components such as the undercarriage, bogies, wheelsets, couplers, braking systems, and pantographs); wheel-rail acoustic signals collected by a trackside microphone array (used to detect wheel flats and bearing failures); temperature distribution of bearings and gearboxes acquired by an infrared thermal imager; and semantic metadata including vehicle ID, train number, route, mileage, speed, load, date and time, and weather conditions. Features for each modality are extracted using a dedicated encoder—convolutional neural networks for images and temporal convolutional networks or Transformers for acoustic time-series signals. At the fusion layer, a cross-modal attention mechanism is introduced: for example, anomalies in vibration signals cause attention weights to be allocated more to the corresponding component area in the image; temperature anomalies enhance the model's sensitivity to structural anomalies in that component. The fused multimodal features are input to the main detection network, while semantic metadata is injected as conditional input into multiple layers of the network, enabling the detection benchmark to be standardized and aligned for "the same vehicle, the same operating condition." When data from a certain type of sensor (such as infrared temperature) is temporarily missing, the network uses features from other modalities to fill in the missing information through cross-modal inference. For example, it can infer the possible thermal state based on image texture and acoustic features, thereby ensuring the fault tolerance of the system.

[0028] Finally, step S5 achieves online deployment and a self-learning closed loop. Lightweight network models, pruned and quantized for compression, are deployed on trackside edge computing nodes to meet millisecond-level real-time inference requirements. Edge nodes perform online detection on data collected as each train passes, outputting three types of results: normal (no anomalies, direct passage); anomaly (outputting the anomaly type and visual location, triggering an alarm and providing manual verification); and unknown outside the distribution (the model proactively indicates "beyond existing knowledge," triggering data labeling and uploading). Anomaly results and unknown outside the distribution sample data are reported by the edge nodes to the data center. The data center aggregates reported data from multiple detection points along the entire line, periodically (e.g., weekly or monthly) performing incremental model updates (step S2) and baseline reconstruction (step S1) to generate new model versions. After the new model is evaluated on an offline validation set and confirmed to have no performance regression through small-scale A / B testing, it is distributed hierarchically to each edge node via the model management service, forming a complete self-learning closed loop of "collection—detection—reporting—update—distribution." For new fault types confirmed by incremental learning, their prototypes will be included in the fault knowledge graph, and subsequent similar faults can be directly detected without further manual intervention.

[0029] Through the above implementation methods, this invention fully utilizes mechanisms such as the periodicity of railway vehicle operation, distributed decoding, distributed perception and continuous learning, and multimodal fusion to achieve a highly robust train inspection algorithm that requires no manual annotation, can adapt to changes in operating conditions, can proactively recognize unknown faults, and can continuously accumulate knowledge.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An unsupervised self-learning column inspection algorithm, characterized in that: Includes the following steps: Step S1: Obtain railway vehicle operation and route information, construct a three-dimensional normal baseline based on the periodicity of vehicle operation, the three-dimensional normal baseline includes time dimension baseline, route dimension baseline and vehicle dimension baseline, and automatically classify historical detection data through clustering method to filter normal operation status for baseline modeling; Step S2: Construct an anomaly detection network with distributed decoding. The network adopts a dual-path encoder-decoder structure with physical information enhancement. It takes running intersection coding and time-mileage information as conditional inputs. The encoder end is equipped with a feature decoupling module. The decoder receives periodic prior guidance to generate a reconstruction output within the normal distribution. It controls the difference between the reconstruction result and the input through a difference constraint loss function. During inference, it generates anomaly scores based on reconstruction differences and dynamically calculates anomaly thresholds. Step S3: Construct an out-of-distribution sensing module. By integrating latent spatial density estimation, reconstructing the probability of difference distribution, and multi-branch consistency difference, it determines whether the sample exceeds the scope of learned knowledge. When it is determined to be an out-of-distribution unknown sample, it triggers active incremental learning. The active incremental learning adopts a continuous learning framework based on prototype playback and uses experience playback and parameter regularization strategies to update the model. Step S4: Fuse multi-source heterogeneous data, including image data, vibration or acoustic data, temperature data and semantic metadata, perform feature alignment and fusion through a cross-modal attention mechanism, and inject semantic metadata as conditional input into the detection network; Step S5: Deploy the detection network on the trackside edge computing node for online inference, outputting three types of results: normal, abnormal, or unknown outside the distribution. Abnormal results and out-of-distribution samples are reported to the data center. The data center periodically performs the incremental model update in step S2 and the baseline reconstruction in step S1, generates a new model version and distributes it to the edge node, forming a self-learning closed loop.

2. The unsupervised self-learning column inspection algorithm according to claim 1, characterized in that: In step S1, the time dimension baseline is used to capture the gradual aging trend of components over time, the route dimension baseline uses the consistency of operating patterns within the same route to filter random disturbances, and the vehicle dimension baseline establishes a healthy reference distribution based on group data of the same type of components of the same model.

3. The unsupervised self-learning column inspection algorithm according to claim 1, characterized in that: In step S2, the difference constraint loss function includes reconstruction loss, difference upper limit constraint loss, feature distribution alignment loss, and periodic consistency loss, wherein the periodic consistency loss is used to ensure that the reconstruction result maintains the same operational path semantics as the input.

4. The unsupervised self-learning column inspection algorithm according to claim 1, characterized in that: In step S3, the potential spatial density estimation adopts Gaussian mixture model or normalized flow modeling, and the tail behavior of the differential distribution is reconstructed using extreme value theory modeling. The prototype playback buffer stores the core embedding of normal samples, the prototype of the identified fault type, and the temporary embedding of out-of-distribution samples. When the out-of-distribution samples accumulate to a preset number and are confirmed as a new fault type by manual review, their prototypes are stably written into the buffer.

5. The unsupervised self-learning column inspection algorithm according to claim 1, characterized in that: In step S4, when data from a certain type of sensor is missing, the network completes the information through cross-modal inference.

6. The unsupervised self-learning column inspection algorithm according to claim 1, characterized in that: In step S5, the detection network deployed at the edge nodes is a lightweight model after pruning and quantization; when the data center performs incremental updates, it adopts a strategy that combines elastic weight consolidation and knowledge distillation to prevent catastrophic forgetting.

Images