A train operation state analysis method and system

Abstract

This invention discloses a method and system for analyzing train operation status, including: A1. Multimodal data collaborative sampling based on immune sensing cell units; A2. Construction and updating of dynamic causal graphs for closed-loop systems; A3. Generation and optimization of diagnostic models based on federated optimization; and A4. Early warning of minor faults based on digital twin stress fields. Train status analysis transforms from a discrete, passive, and retrospective auxiliary maintenance tool into a continuous, proactive, and a priori autonomous health management system. It can not only detect faults earlier and more accurately, but also interpret faults, predict their development, and intelligently optimize its own monitoring strategies.

Description

Technical Field

[0001] This invention belongs to the field of dynamic monitoring vehicle technology, specifically, it relates to a method and system for analyzing train operation status. Background Technology

[0002] With the rapid expansion of China's high-speed railway network and the continuous increase in train operating speed, the analysis of EMU operating status and fault early warning have become core issues in ensuring transportation safety and improving operation and maintenance efficiency. Currently, this field mainly relies on a series of technologies based on a combination of data-driven and mechanistic models, but there are still several deep-seated and systemic technical bottlenecks that restrict the in-depth development of status analysis towards real-time, precision, and intelligence.

[0003] First, at the data sensing level, existing technologies generally employ monitoring methods with fixed frequencies and thresholds. Sensor networks collect data at a constant rate (e.g., 1 kHz). This method generates massive amounts of redundant data during long-term stable train operation, resulting in a huge waste of storage and transmission resources. Furthermore, during critical periods such as transient shocks and the onset of anomalies, insufficient sampling rates or bandwidth limitations lead to the loss of crucial fault characteristic information. This uneven data collection pattern leaves the backend analysis system facing the dilemma of both "information deserts" and "information floods," failing to provide a high-quality data foundation for precise diagnosis.

[0004] Secondly, at the level of condition diagnosis, existing methods often exhibit the shortcomings of "model silos" and "shallow cognition." On the one hand, although algorithms such as deep learning have made progress in fault classification, they are essentially "black boxes," only able to map symptoms to fault codes, unable to explain the root cause, mechanism, and propagation path of the fault, making it difficult for maintenance personnel to make accurate maintenance decisions. On the other hand, for complex systems with strong coupling and closed-loop control, such as traction drives, the dynamic causal relationships between internal variables and the fault propagation mechanism are still unclear. Static fault trees or expert rule bases are difficult to adapt to system parameter drift and complex operating condition changes, resulting in persistently high false alarm and false negative rates.

[0005] Furthermore, at the system architecture level, existing solutions lack the ability for collaborative evolution and swarm intelligence. Each train set acts as an independent data silo, and its local diagnostic model can only learn from its own limited historical data, resulting in weak generalization ability for new and unknown faults. Although federated learning frameworks have been proposed to break down data silos, existing parameter averaging aggregation methods are prone to "negative transfer" in train groups with heterogeneous data distribution (different lines, train conditions), leading to a decline in global model performance and failing to balance generality and personalized needs.

[0006] Finally, in terms of early warning and foresight, current technologies lag significantly behind the occurrence of physical damage. Current systems primarily rely on threshold exceedances of macroscopic physical quantities such as vibration and temperature for alarms. These signals typically indicate that macroscopic damage (such as crack propagation or severe wear) has already occurred in the component materials. For microscopic damage such as early microcracks in critical bogie structural components or initial pitting corrosion in bearings, existing monitoring systems cannot detect it when there are no obvious abnormalities in macroscopic signals. This results in the loss of the optimal window for fault prevention, leaving predictive maintenance only in the theoretical stage.

[0007] In summary, current EMU operation status analysis technology faces four interconnected and progressively worsening core technical problems: inefficient data acquisition, superficial diagnostic understanding, isolated model evolution, and severely lagging early warning systems. These problems collectively lead to insufficient intelligence in existing systems, high operation and maintenance costs, and an inability to fundamentally achieve early prevention and precise control of faults. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for analyzing train operation status.

[0009] To achieve the aforementioned objectives, the technical solution adopted by this invention includes: a train operation status analysis method, comprising the following steps: A1. Multimodal data collaborative sampling based on immune-sensing cell units: A1.1 Bionic Design: Each physical sensor (such as a vibration accelerometer) is equipped with a microprocessor unit, forming an immune sensing cell unit (ISC). The ISC not only collects data but also incorporates a lightweight anomaly scorer. A1.2 Co-sampling algorithm: The base sampling rate f_base is kept at 1kHz; The anomaly score S(t) is dynamically calculated based on the signal local entropy H(t) and the mutual information I(t) with adjacent sensors: S(t) = α * (1 - H(t) / H_max) + β * I(t), where α and β are weights, and H_max is the maximum entropy reference value; Dynamic adjustment: When S(t) > θ_high (high threshold), the sampling rate of the immune sensing cell unit and its three physically associated neighboring immune sensing cell units is increased to f_high = 10kHz, and full-band acquisition is enabled; when S(t) < θ_low, the sampling rate is reduced to f_low = 100Hz, and only the characteristic frequency band is acquired; θ_high is defined as the 99.9th percentile of the historical S(t) value of the ISC in a healthy state; θ_low is defined as the 50th percentile (median). During the startup phase, conservative initial values ​​based on physical model simulation or expert experience are used, and these values ​​are dynamically updated based on health data during operation. Coordination mechanism: Through local communication, if an ISC triggers upsampling, it will notify all ISCs in its functional coupling group (such as all vibration and displacement sensors on the bogie) to synchronously increase the sampling rate. A2. Construction and updating of dynamic cause-effect graphs for closed-loop systems: A2.1 Initial Graph Construction: Based on the system schematic diagram (all measurable or estimable key physical variables are abstracted as graph nodes. The variable set V = {v1, v2, ..., vn}, for example: v1: traction motor current, v2: motor speed, v3: output torque, v4: gearbox bearing temperature, v5: inverter DC side voltage, etc.), construct a directed graph G_0 with physical variables (such as current, speed, torque) as nodes and known physical laws (such as motor equations, gear ratios) as edges; A2.2 Online Learning and Updates: Under normal system operation, data is continuously collected, and a strategy is selected: for variable pairs with known accurate physical models (such as motors), Kalman filtering residual cross-correlation analysis is enforced. For variable pairs with unknown or complex models, if the data exhibits significant linearity, Granger causality testing with time delay is used; if nonlinearity is significant, transitive entropy is used. Bayesian fusion: the weight w of each edge X→Y in the causal graph G_t is considered its existence probability P(edge). The Kalman filtering analysis result based on the physical model is used as the prior probability P_prior(edge), and the causal strength calculated by the data-driven method (Granger or transitive entropy) is normalized and used as the likelihood probability P_likelihood(edge). The posterior probability is updated according to the Bayesian formula: P(edge)_new = (P_likelihood(edge) * P_prior(edge)) / Σ. This process ensures that known physical knowledge continuously "anchors" the direction of data-driven discovery, preventing the emergence of false causality contrary to physics. A2.3 Fault Propagation Simulation: An influence propagation model based on a directed weighted graph is adopted. When the outlier A_X (z-score) of node X exceeds the threshold, the "influence flow" received by each of its downstream neighbor nodes Y is calculated as follows: I_{X→Y} = A_X * w_{X→Y} * α_{X→Y}. Where w_{X→Y} is the causal edge weight, and α_{X→Y} is the physical attenuation factor (determined based on energy or information transmission efficiency, such as a transmission efficiency of 0.97). The potential anomaly level P_Y of node Y is updated to: P_Y = max( A_Y, Σ I_{X→Y} ). If P_Y exceeds its own anomaly threshold, then Y is activated and continues to propagate downstream; The algorithm outputs all activated nodes and their P-values, forming a hypothesis about the fault propagation path. Maintenance personnel can use this to determine the root cause of the fault (the initially activated node with the highest P-value) and potentially affected secondary components. A3. Generation and optimization of diagnostic models based on federated optimization: A3.1 Local Model Training: Each EMU edge node maintains a local model library containing N different diagnostic sub-models (such as CNN, LSTM, GNN). When training with local data, a meta-learning update rule is introduced: the parameter update direction Δθ_local of the local model is determined not only by the local gradient ∇L_local, but also by a personalized correction vector v_personal: Δθ_local = η * (▽L_local + λ * v_personal), where Δθ_local: the update vector of the local model parameters; η: the global learning rate; ▽L_local: the gradient of the local loss function; λ: the personalized balancing coefficient; v_personal: the personalized correction feature vector, which is a feature vector issued by the cloud platform specifically for the operating environment of the train crew. A3.3, Federated Aggregation and Optimization: The cloud platform does not directly aggregate model parameters, but instead runs a population of models; Each edge node uses a homomorphic encryption algorithm (such as the CKKS scheme) to encrypt the average gradient of all batches of data in this training cycle (i.e., the update amount Δθ_local of the local model parameters), obtaining ciphertext C_i. This ciphertext, along with the fault identification rate P_i and the false alarm rate F_i of unknown faults obtained from the local evaluation in this round, is uploaded to the cloud platform for optimization using the Non-Dominated Ranking Genetic Algorithm (NSGA-II). The two optimization objectives are: maximizing the global fault identification rate and minimizing the model's alarm rate for unknown faults to avoid false alarms. Each round of optimization generates a new generation of model population. The Pareto optimal solution set models are encrypted and distributed back to the edge nodes to replace the old models. The cloud platform uses its private key to generate a digital signature for the distributed model files. Upon receiving the model, the edge nodes verify the signature using the pre-configured cloud platform public key, ensuring the model's integrity and provenance. Simultaneously, the model file itself is encrypted using a one-time pad symmetric key, which is then encrypted using the edge nodes' asymmetric public key before being transmitted together. A4. Early warning of minor faults based on digital twin stress fields: A4.1 High-fidelity twin construction: A finite element model integrating geometry, materials, and load boundary conditions is created for key components (such as bogies) and run on a cloud platform. A4.2 Dimensionality Reduction and Edge Computing: Dimensionality Reduction Modeling: On the cloud platform, a reduced-order model (ROM) is generated through parametric modeling and offline simulation. The input to this ROM is a few key load parameters L(t) = [L1, L2, ... Lk], which are derived from real-time data in the train control network (such as MVB or Ethernet), such as traction motor torque commands, train speed, track curvature information, axle load signals, etc. The output of the ROM is the theoretical stress vector σ_theoretical_keypoints(t) of the key monitoring points. This ROM model and the necessary load-stress mapping matrix are then distributed to the edge nodes of the corresponding components. A4.3 Difference Detection and Early Warning: The ROM model is distributed to the edge nodes. The edge nodes calculate σ_theoretical_keypoints(t) in real time and compare it only with the measured stress σ_measured_keypoints(t) of the corresponding key points measured in the ISC network to calculate the residuals; Reconstructing early warning indicators: Based on the residual vector of key monitoring points r(t) = σ_measured_keypoints(t) -σ_theoretical_keypoints(t), Residual abnormal amplitude: F_magnitude(t) = || r(t) ||_2 (2-norm of the vector), F_magnitude(t): the comprehensive amplitude intensity of the stress residual at time t; Fault mode similarity: The calculation method of Sim(t) remains the same, but the comparison object becomes r(t) and the standard fault mode vector. Sim(t): the shape similarity between the residual vector at time t and the standard fault mode. Joint early warning logic: When the conditions F_magnitude(t) > F_threshold(t) and Sim(t) > S_threshold are simultaneously satisfied for M consecutive periods, an early warning is triggered, where M = ceil( τ_min / T_s ) sampling periods, where τ_min is the minimum physical time (e.g., 100 milliseconds) required by the system to confirm a persistent fault phenomenon, T_s is the current actual sampling period (dynamically changing with S(t), F_threshold(t) is the dynamic amplitude early warning threshold at time t, and S_threshold is the threshold for determining pattern similarity.

[0010] The joint early warning logic constructs an early fault detection mechanism with extremely high robustness and extremely low false alarm rate by introducing dynamic thresholds, pattern matching, and adaptive continuous decision-making. It comprehensively judges from three physical dimensions—amplitude, space, and time—achieving accurate capture of early, small but potentially damaging damage. This ensures that the physical timescale of the early warning judgment remains stable regardless of changes in the sampling rate, triggering early warnings. The dynamic model of F_threshold(t) remains applicable, but its computational basis changes from "field integral value" to "historical statistics of the residual vector norm."

[0011] A train operation status analysis system, comprising: The sensing unit, integrated with the physical sensor, is responsible for basic signal conditioning and fixed-rate sampling, realizing the basic sampling in A1.2; Cluster head processors, deployed on top of multiple sensing units (such as all sensors of a bogie) that are physically adjacent or functionally related, perform anomaly scoring calculation (S(t)), dynamic sampling rate decision-making, and intra-cluster coordination. The cluster head processor controls the sampling rate of each sensing unit via a high-speed bus (such as SPI). A global adaptive optimization layer is implemented, with a Health Management Core (HMC) module designed at the top level of the system. This module receives all diagnostic and early warning results and dynamically adjusts its settings. 1. For A1: When a component is suspected of having an early failure (from A4) or the cause-effect graph shows that it is in a critical propagation path (from A2), HMC can instruct to increase the base sampling rate f_base of the relevant ISC cluster or decrease its anomaly scoring threshold θ_high to achieve "diagnostic guidance perception".

[0012] 2. For A3: Newly identified fault cases will be used as labels to trigger the federated learning training rounds of the relevant models first; Cloud Platform: The cloud platform's strategy optimization engine receives information from all edge nodes, generates next-generation detection model parameters and diagnostic rules through a multi-objective game algorithm, and then encrypts and distributes them to each edge node.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of a train operation status analysis method.

[0014] Compared with the prior art, the advantages of the present invention include: (1) The train operation status analysis method and system provided by the present invention solves the deployment problem of intelligent collaboration of massive sensors through the hierarchical hardware architecture of "sensing unit (SU) + cluster head processor (CHP)"; it clarifies the exclusive real-time domain of TSN, isolates management traffic, and ensures the reliability and low latency of key data transmission. (2) The present invention provides a train operation status analysis method and system, which uses a reduced-order model (ROM) to replace the real-time calculation of the full field finite element on the cloud platform and pushes the key calculations down to the edge, thus solving the performance bottleneck of stress field analysis; (3) The train operation status analysis method and system provided by the present invention supplements the model signature verification mechanism and designs a health management core (HMC) to realize global adaptive optimization from diagnosis to perception, making the system a truly intelligent and adaptive organic whole. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a train operation status analysis method and system according to the present invention. Detailed Implementation

[0017] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The technical solution, its implementation process, and principles will be further explained below with reference to the accompanying drawings and specific implementation examples in the embodiments of this application.

[0018] It should be noted that the embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, the present invention covers any substitutions, modifications, equivalent methods and solutions made on the spirit, principles and scope of the present invention as defined by the claims. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this application, the terms "first," "second," "third," and similar terms do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "one" or "a" do not indicate a quantity limitation, but rather indicate the presence of at least one. Terms such as "include" or "contain" mean that the elements or objects preceding "include" cover the elements or objects listed following "include" or their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0020] In the description of this application, the terms center, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. In addition, when using positional terms such as sides, outer side, top and bottom, etc., it should be understood that they are used only for the convenience of understanding and description, taking into account that the structure may be oriented to other positions.

[0021] In the description of this application, unless otherwise expressly specified and limited, the technical or scientific terms used shall have the ordinary meaning understood by a person with ordinary skills in the art to which this application pertains. Terms such as installation, connection, and linking shall be interpreted broadly, for example, they may refer to fixed connection, detachable connection, mating connection, or integral connection. For a person skilled in the art, the specific meaning of the above terms in this application may be understood according to the specific circumstances.

[0022] The present invention aims to introduce and explain the structural composition of a train operation status analysis method and system, as well as the coordination relationship between the various components. Unless otherwise specified, the dimensions, materials, and manufacturing processes of the various components suitable for the train operation status analysis method and system in the present invention can be selected according to specific circumstances, and no special limitations or explanations are made here.

[0023] Furthermore, to provide the public with a better understanding of the present invention, certain specific details are described in detail in the following description of the invention. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0024] Example 1 Please see Figure 1 A method for analyzing train operation status includes the following steps: A1. Multimodal data collaborative sampling based on immune-sensing cell units: A1.1 Bionic Design: Each physical sensor (such as a vibration accelerometer) is equipped with a microprocessor unit, forming an immune sensing cell unit (ISC). The ISC not only collects data but also incorporates a lightweight anomaly scorer. A1.2 Co-sampling algorithm: The base sampling rate f_base is kept at 1kHz; The anomaly score S(t) is dynamically calculated based on the signal local entropy H(t) and the mutual information I(t) with adjacent sensors: S(t) = α * (1 - H(t) / H_max) + β * I(t), where α and β are weights, and H_max is the maximum entropy reference value; Dynamic adjustment: When S(t) > θ_high (high threshold), the sampling rate of the immune sensing cell unit and its three physically associated neighboring immune sensing cell units is increased to f_high = 10kHz, and full-band acquisition is enabled; when S(t) < θ_low, the sampling rate is reduced to f_low = 100Hz, and only the characteristic frequency band is acquired; θ_high is defined as the 99.9th percentile of the historical S(t) value of the ISC in a healthy state; θ_low is defined as the 50th percentile (median). During the startup phase, conservative initial values ​​based on physical model simulation or expert experience are used, and these values ​​are dynamically updated based on health data during operation. Coordination mechanism: Through local communication, if an ISC triggers upsampling, it will notify all ISCs in its functional coupling group (such as all vibration and displacement sensors on the bogie) to synchronously increase the sampling rate. A2. Construction and updating of dynamic cause-effect graphs for closed-loop systems: A2.1 Initial Graph Construction: Based on the system schematic diagram (all measurable or estimable key physical variables are abstracted as graph nodes. The variable set V = {v1, v2, ..., vn}, for example: v1: traction motor current, v2: motor speed, v3: output torque, v4: gearbox bearing temperature, v5: inverter DC side voltage, etc.), construct a directed graph G_0 with physical variables (such as current, speed, torque) as nodes and known physical laws (such as motor equations, gear ratios) as edges; A2.2 Online Learning and Updates: Under normal system operation, data is continuously collected, and a strategy is selected: for variable pairs with known accurate physical models (such as motors), Kalman filtering residual cross-correlation analysis is enforced. For variable pairs with unknown or complex models, if the data exhibits significant linearity, Granger causality testing with time delay is used; if nonlinearity is significant, transitive entropy is used. Bayesian fusion: the weight w of each edge X→Y in the causal graph G_t is considered its existence probability P(edge). The Kalman filtering analysis result based on the physical model is used as the prior probability P_prior(edge), and the causal strength calculated by the data-driven method (Granger or transitive entropy) is normalized and used as the likelihood probability P_likelihood(edge). The posterior probability is updated according to the Bayesian formula: P(edge)_new = (P_likelihood(edge) * P_prior(edge)) / Σ. This process ensures that known physical knowledge continuously "anchors" the direction of data-driven discovery, preventing the emergence of false causality contrary to physics. A2.3 Fault Propagation Simulation: An influence propagation model based on a directed weighted graph is adopted. When the outlier A_X (z-score) of node X exceeds the threshold, the "influence flow" received by each of its downstream neighbor nodes Y is calculated as follows: I_{X→Y} = A_X * w_{X→Y} * α_{X→Y}. Where w_{X→Y} is the causal edge weight, and α_{X→Y} is the physical attenuation factor (determined based on energy or information transmission efficiency, such as a transmission efficiency of 0.97). The potential anomaly level P_Y of node Y is updated to: P_Y = max( A_Y, Σ I_{X→Y} ). If P_Y exceeds its own anomaly threshold, then Y is activated and continues to propagate downstream; The algorithm outputs all activated nodes and their P-values, forming a hypothesis about the fault propagation path. Maintenance personnel can use this to determine the root cause of the fault (the initially activated node with the highest P-value) and potentially affected secondary components. A3. Generation and optimization of diagnostic models based on federated optimization: A3.1 Local Model Training: Each EMU edge node maintains a local model library containing N different diagnostic sub-models (such as CNN, LSTM, GNN). When training with local data, a meta-learning update rule is introduced: the parameter update direction Δθ_local of the local model is determined not only by the local gradient ▽L_local, but also by a personalized correction vector v_personal: Δθ_local = η * (▽L_local + λ * v_personal), where Δθ_local: the update vector of the local model parameters; η: the global learning rate; ▽L_local: the gradient of the local loss function; λ: the personalized balance coefficient; v_personal: the personalized correction feature vector, which is a feature vector issued by the cloud platform for the operating environment of the train crew; the essence of this update rule is to vectorize the experience driven by local data (▽L_local) and the personalized prior knowledge extracted from the cloud (v_personal) to jointly guide the evolution direction of the model; A3.3, Federated Aggregation and Optimization: The cloud platform does not directly aggregate model parameters, but instead runs a population of models; Each edge node uses a homomorphic encryption algorithm (such as the CKKS scheme) to encrypt the average gradient of all batches of data in this training cycle (i.e., the update amount Δθ_local of the local model parameters), obtaining ciphertext C_i. This ciphertext, along with the fault identification rate P_i and the false alarm rate F_i of unknown faults obtained from the local evaluation in this round, is uploaded to the cloud platform for optimization using the Non-Dominated Ranking Genetic Algorithm (NSGA-II). The two optimization objectives are: maximizing the global fault identification rate and minimizing the model's alarm rate for unknown faults to avoid false alarms. Each round of optimization generates a new generation of model population. The Pareto optimal solution set models are encrypted and distributed back to the edge nodes to replace the old models. The cloud platform uses its private key to generate a digital signature for the distributed model files. Upon receiving the model, the edge nodes verify the signature using the pre-configured cloud platform public key, ensuring the model's integrity and provenance. Simultaneously, the model file itself is encrypted using a one-time pad symmetric key, which is then encrypted using the edge nodes' asymmetric public key before being transmitted together. A4. Early warning of minor faults based on digital twin stress fields: A4.1 High-fidelity twin construction: A finite element model integrating geometry, materials, and load boundary conditions is created for key components (such as bogies) and run on a cloud platform. A4.2 Dimensionality Reduction and Edge Computing: Dimensionality Reduction Modeling: On the cloud platform, a reduced-order model (ROM) is generated through parametric modeling and offline simulation. The input to this ROM is a few key load parameters L(t) = [L1, L2, ... Lk], which are derived from real-time data in the train control network (such as MVB or Ethernet), such as traction motor torque commands, train speed, track curvature information, axle load signals, etc. The output of the ROM is the theoretical stress vector σ_theoretical_keypoints(t) of the key monitoring points. This ROM model and the necessary load-stress mapping matrix are then distributed to the edge nodes of the corresponding components. A4.3 Difference Detection and Early Warning: The ROM model is distributed to the edge nodes. The edge nodes calculate σ_theoretical_keypoints(t) in real time and compare it only with the measured stress σ_measured_keypoints(t) of the corresponding key points measured in the ISC network to calculate the residuals; Reconstructing early warning indicators: Based on the residual vector of key monitoring points r(t) = σ_measured_keypoints(t) -σ_theoretical_keypoints(t), Residual abnormal amplitude: F_magnitude(t) = || r(t) ||_2 (2-norm of the vector), F_magnitude(t): the comprehensive amplitude intensity of the stress residual at time t; Fault mode similarity: The calculation method of Sim(t) remains the same, but the comparison object becomes r(t) and the standard fault mode vector. Sim(t): the shape similarity between the residual vector at time t and the standard fault mode. Joint early warning logic: An early warning is triggered when the conditions F_magnitude(t) > F_threshold(t) and Sim(t) > S_threshold are simultaneously satisfied for M consecutive periods, where M = ceil( τ_min / T_s ) sampling periods, where τ_min is the minimum physical time (e.g., 100 milliseconds) required by the system to confirm a persistent fault phenomenon, and T_s is the current actual sampling period (dynamically changing with S(t)). F_threshold(t) is the dynamic amplitude early warning threshold at time t; S_threshold: the threshold for determining pattern similarity.

[0025] The joint early warning logic constructs an early fault detection mechanism with extremely high robustness and extremely low false alarm rate by introducing dynamic thresholds, pattern matching, and adaptive continuous decision-making. It comprehensively judges from three physical dimensions—amplitude, space, and time—achieving accurate capture of early, small but potentially damaging damage. This ensures that the physical timescale of the early warning judgment remains stable regardless of changes in the sampling rate, triggering early warnings. The dynamic model of F_threshold(t) remains applicable, but its computational basis has changed from "field integral value" to "historical statistics of the residual vector norm."

[0026] A train operation status analysis system, comprising: The sensing unit, integrated with the physical sensor, is responsible for basic signal conditioning and fixed-rate sampling, realizing the basic sampling in A1.2; Cluster head processors, deployed on top of multiple sensing units (such as all sensors of a bogie) that are physically adjacent or functionally related, perform anomaly scoring calculation (S(t)), dynamic sampling rate decision-making, and intra-cluster coordination. The cluster head processor controls the sampling rate of each sensing unit via a high-speed bus (such as SPI). A global adaptive optimization layer is implemented, with a Health Management Core (HMC) module designed at the top level of the system. This module receives all diagnostic and early warning results and dynamically adjusts its settings. 1. For A1: When a component is suspected of having an early failure (from A4) or the cause-effect graph shows that it is in a critical propagation path (from A2), HMC can instruct to increase the base sampling rate f_base of the relevant ISC cluster or decrease its anomaly scoring threshold θ_high to achieve "diagnostic guidance perception".

[0027] 2. For A3: Newly identified fault cases will be used as labels to trigger the federated learning training rounds of the relevant models first; Cloud Platform: The cloud platform's strategy optimization engine receives information from all edge nodes, generates next-generation detection model parameters and diagnostic rules through a multi-objective game algorithm, and then encrypts and distributes them to each edge node.

[0028] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of a train operation status analysis method.

[0029] Understandably, step A1 addresses the uneven data distribution in traditional technologies: fixed sampling leads to data redundancy during normal periods and loss of features during abnormal periods. It achieves "event-driven" adaptive sensing, reducing redundant data traffic by over 60% at the source while ensuring 100% capture of complete high-frequency waveforms of abnormal events, significantly improving data value density and transmission and storage efficiency. Step A2 addresses the problem of traditional diagnostic techniques that "know what but not why": black-box models only report faults but cannot pinpoint root causes or predict propagation. An interpretable and evolvable system knowledge graph is constructed, upgrading diagnosis from a simple "symptom-fault" mapping to reasoning about fault propagation paths and root causes, providing direct evidence for precise maintenance and significantly shortening fault diagnosis time. Step A3 addresses the limitations of traditional models, such as "partiality" and "isolation": poor generalization of single models and data silos between vehicles prevent the formation of collective intelligence. It achieves privacy-preserving collective intelligence evolution through "data remaining stationary while the model moves," enabling the diagnostic model to continuously absorb experience from the entire railway network while maintaining optimal performance for specific trains and lines through global evolution and personalized adaptation. This significantly improves the comprehensive recognition rate of both known and unknown faults. Step A4 addresses the "after-the-fact" nature of traditional early warning technologies: traditional vibration and temperature monitoring cannot detect early micro-damage in materials. It achieves quantitative comparison and early warning based on "physical field consistency," detecting micro-damage tens to hundreds of hours before visible cracks or functional failures appear in components. This transforms the maintenance model from "repair after failure" to "intervention before damage," fundamentally preventing major safety accidents.

[0030] Working principle: The raw signal is sampled by the sensing unit and intelligently filtered by the cluster head processor, and then uploaded to various analysis engines in the edge layer via the TSN real-time bus.

[0031] The analysis results (causal reasoning, stress residuals, and fault classification) are aggregated into the core module of health management for comprehensive decision-making.

[0032] Edge nodes upload encrypted model gradients and performance metrics to the cloud platform via 5G / satellite.

[0033] The cloud platform generates a better model through federated evolution and combines individual features to generate a personalized vector v_personal. After encryption and signing, the vector is sent to each train edge node to complete the model update and personalized adaptation.

[0034] Based on the diagnostic and early warning results, the core health management module makes optimization decisions and issues adjustment instructions to the cluster head processor (such as increasing the base sampling rate f_base in a specific area).

[0035] The cluster head processor executes instructions to change the sampling behavior of the sensing units, thereby enabling key monitoring of high-risk points and forming an intelligent closed loop of "the more suspicious, the more attention; the more attention, the clearer."

[0036] It should be understood that the above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. It should not be considered that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several simple deductions or substitutions can be made without departing from the concept of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for analyzing train operation status, characterized in that: Includes the following steps: A1. Multimodal data collaborative sampling based on immune sensing cell units; A2. Construction and updating of dynamic cause-effect graphs for closed-loop systems; A3. Generation and optimization of diagnostic models based on federated optimization; A4. Early warning of minor faults based on digital twin stress field.

2. The train operation status analysis method according to claim 1, characterized in that: Step A1 specifically includes: A1.1 Bionic Design: Each physical sensor is equipped with a microprocessor unit, which constitutes an immune sensing cell unit. The immune sensing cell unit is used to collect data and also includes a built-in lightweight abnormality scorer. A1.2, Cooperative Sampling Algorithm: The anomaly score S(t) is dynamically calculated based on the signal local entropy H(t) and the mutual information I(t) with adjacent sensors: S(t) = α * (1 - H(t) / H_max) + β * I(t), where α and β are weights, and H_max is the maximum entropy reference value; Dynamic adjustment: When S(t) > θ_high, θ_high is a high threshold, the sampling rate of the immune sensing cell unit and its three physically associated neighboring immune sensing cell units is increased to f_high = 10kHz, and full-band acquisition is enabled; when S(t) < θ_low, θ_low is the 50th percentile, the sampling rate is reduced to f_low = 100Hz, and only the characteristic frequency band is acquired; Coordination mechanism: Through local communication, if an immune sensing cell unit triggers upsampling, it will notify all immune sensing cell units in its functional coupling group to synchronously increase the sampling rate.

3. The train operation status analysis method according to claim 1, characterized in that: Step A2 specifically includes: A2.1 Initial Graph Construction: Construct a directed graph G_0 with physical variables as nodes and known physical laws as edges; A2.2 Online Learning and Updates: Under normal system operation, data is continuously collected, and the following strategies are selected: For variable pairs with known accurate physical models, Kalman filtering and residual cross-correlation analysis are enforced; for variables with unknown or complex models, if the data have obvious linear characteristics, Granger causality test with time delay is used; if the nonlinear characteristics are obvious, transfer entropy is used. Bayesian fusion: The weight w of each edge X→Y in the causal graph G_t is regarded as its existence probability P(edge), the Kalman filter analysis result based on the physical model is used as the prior probability P_prior(edge), and the causal strength calculated by the data-driven method is normalized and used as the likelihood probability P_likelihood(edge). The posterior probability is updated according to the Bayesian formula: P(edge)_new = (P_likelihood(edge) * P_prior(edge)) / Σ.

4. The train operation status analysis method according to claim 3, characterized in that: Also includes: A2.3 Fault Propagation Simulation: An influence propagation model based on a directed weighted graph is adopted. When the outlier value A_X (z-score) of node X exceeds the threshold, the influence flow received by each of its downstream neighbor nodes Y is calculated as follows: I_{X→Y} = A_X * w_{X→Y} * α_{X→Y}, where w_{X→Y} is the causal edge weight and α_{X→Y} is the physical attenuation factor. The potential anomaly level P_Y of node Y is updated to: P_Y = max( A_Y, Σ I_{X→Y} ). If P_Y exceeds its own anomaly threshold, then Y is activated and continues to propagate downstream. The algorithm outputs all activated nodes and their P values, forming a hypothesis about the fault propagation path.

5. The train operation status analysis method according to claim 1, characterized in that: Step A3 specifically includes: A3.1 Local Model Training: Each EMU edge node maintains a local model library containing N different diagnostic sub-models; When training with local model data, a meta-learning update rule is introduced: Δθ_local = η * (▽L_local + λ* v_personal), where Δθ_local is the update vector of local model parameters; η is the global learning rate; ▽L_local is the gradient of the local loss function; λ is the personalized balancing coefficient; and v_personal is the personalized correction feature vector, which is a feature vector issued by the cloud platform specifically for the operating environment of the train crew. A3.3, Federated Aggregation and Optimization: Each edge node uses a homomorphic encryption algorithm to encrypt the average gradient of all batches of data in this training cycle to obtain ciphertext C_i, which, together with the fault identification rate P_i and the unknown fault false alarm rate F_i evaluated locally in this round, is uploaded to the cloud platform and optimized using the non-dominated sorting genetic algorithm NSGA-II. Each round of optimization generates a new generation of model population. The Pareto optimal solution set models are encrypted and distributed back to the edge nodes to replace the old models. The cloud platform uses its private key to generate a digital signature for the distributed model file. After receiving the model, the edge node uses the pre-set cloud platform public key to verify the signature, ensuring the integrity and trustworthiness of the model's origin. At the same time, the model file itself is encrypted using a one-time pad symmetric key, which is then encrypted with the edge node's asymmetric public key and transmitted together.

6. The train operation status analysis method according to claim 1, characterized in that: Step A4 specifically includes: A4.1 High-fidelity twin construction: For key components, a finite element model integrating geometry, materials, and load boundary conditions is established and run on a cloud platform. A4.2 Dimensionality Reduction and Edge Computing: Dimensionality reduction modeling: On the cloud platform, a reduced-order model is generated through parametric modeling and offline simulation; A4.3 Difference Detection and Early Warning: The reduced-order model is distributed to the edge nodes, and the edge nodes calculate σ_theoretical_keypoints(t) in real time and compare it only with the measured stress σ_measured_keypoints(t) of the corresponding key points measured by the ISC network to calculate the residual. Reconstructing early warning indicators: Based on the residual vector of key monitoring points r(t) = σ_measured_keypoints(t) - σ_theoretical_keypoints(t), Residual abnormal amplitude: F_magnitude(t) = || r(t) ||_2 , F_magnitude(t): the comprehensive amplitude intensity of stress residual at time t; Fault mode similarity: The calculation method of Sim(t) remains the same, but the comparison object becomes r(t) and the standard fault mode vector. Sim(t): the shape similarity between the residual vector at time t and the standard fault mode. Joint early warning logic: When the conditions F_magnitude(t) > F_threshold(t) and Sim(t) > S_threshold are simultaneously satisfied for M consecutive periods, an early warning is triggered, where M = ceil( τ_min / T_s ) sampling periods, where τ_min is the minimum physical time required by the system to confirm a persistent fault phenomenon, T_s is the current actual sampling period, F_threshold(t) is the dynamic amplitude early warning threshold at time t, and S_threshold is the threshold for determining pattern similarity.

7. The train operation status analysis method according to claim 6, characterized in that: The input to the reduced-order model is a few key load parameters L(t) = [L1, L2, ... Lk], and the output of the reduced-order model is the theoretical stress vector σ_theoretical_keypoints(t) of the key monitoring points. The reduced-order model and the necessary load-stress mapping matrix are then sent to the edge nodes of the corresponding components.

8. The train operation status analysis method according to claim 7, characterized in that: The M periods: M = ceil( τ_min / T_s ) sampling periods, where τ_min is the minimum physical time required for the system to confirm a persistent fault phenomenon, and T_s is the current actual sampling period, ensuring that the physical time scale for early warning judgment is stable regardless of changes in the sampling rate, thus triggering early warning.

9. A train operation status analysis system, used to implement the steps of the method as described in any one of claims 1-8, characterized in that: include: The sensing unit, integrated with the physical sensor, is responsible for basic signal conditioning and fixed-rate sampling, realizing the basic sampling in step A1.2; Cluster head processors, deployed on multiple sensing units that are physically adjacent or functionally related, can realize anomaly score calculation S(t), dynamic sampling rate decision-making, and intra-cluster collaboration. A global adaptive optimization layer, located at the top level of the system, contains a core health management module that receives all diagnostic and early warning results and dynamically adjusts the system. For step A1: When a component is suspected of having an early failure or the cause-effect graph shows that it is in a critical propagation path, the instruction is to increase the base sampling rate f_base of the relevant immune sensing cell unit or decrease its abnormal score threshold θ_high to achieve diagnostic guidance sensing. For step A3: Use newly confirmed fault cases as labels to trigger the federated learning training rounds of the relevant models first; Cloud Platform: The cloud platform's strategy optimization engine receives information from all edge nodes, generates next-generation detection model parameters and diagnostic rules through a multi-objective game algorithm, and then encrypts and distributes them to each edge node.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it performs the steps of the method as described in any one of claims 1-8.

Images