Subway relay residual life prediction method based on physical information neural network

Abstract

The present application relates to the field of relay health management and residual life prediction of urban rail vehicles, in particular to a subway relay residual life prediction method based on physical information neural network, comprising collecting subway relay multi-source time series monitoring data and preprocessing; establishing a degradation physical mechanism model including contact wear dynamics, arc ablation aging and mechanical fatigue attenuation and setting constraint conditions; constructing a physical information neural network of fully connected deep neural network architecture, embedding the degradation physical mechanism model and the constraint conditions into the loss function, which is composed of data fitting loss, physical residual loss and boundary constraint loss; training the network using standardized data set; inputting real-time data into the trained model to output degradation health degree and residual life prediction value. The present application combines the advantages of physical mechanism and data-driven, realizes high-precision residual life prediction under small sample variable working conditions, and supports predictive maintenance of urban rail vehicles.

Description

Technical Field

[0001] This invention relates to the field of urban rail vehicle relay health management and remaining life prediction technology, specifically a method for predicting the remaining life of metro relays based on physical information neural networks. Background Technology

[0002] Subway relays are core components of the electrical systems of urban rail vehicles, widely distributed in train control systems, signaling systems, traction systems, and auxiliary power supply systems. Their health directly affects driving safety and the stability of each system's operation. During long-term service, relays are frequently subjected to multiple factors such as opening and closing actions, arc erosion, mechanical contact wear, and environmental temperature and humidity disturbances, exhibiting obvious progressive degradation characteristics and easily leading to faults such as contact adhesion, poor contact, and malfunctions.

[0003] Currently, the existing technologies for predicting the remaining lifespan and managing the health of subway relays mainly include the following solutions: (a) Traditional empirical model scheme This scheme is based on the rated parameters of the relay's mechanical and electrical lifespan, using a fixed threshold of the number of operations to determine the end of its lifespan, and relying solely on the factory-specified values ​​to roughly divide the maintenance cycle. When the relay's cumulative number of operations reaches the preset threshold, maintenance or replacement is scheduled. The disadvantages of this scheme are: the fixed threshold method is too crude, failing to consider differences in actual operating conditions, environmental disturbances, and degradation differences between individual relays, resulting in large prediction errors. This can easily lead to premature replacement (over-maintenance) causing resource waste, or premature replacement (under-maintenance) causing malfunctions and downtime. (II) Pure Data-Driven Machine Learning / Deep Learning Solutions This scheme employs deep learning methods such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Backpropagation (BP) neural networks. It trains the model based on time-series data from historical operational monitoring and predicts remaining lifespan by fitting degradation trends. However, this scheme has drawbacks: the purely data-driven model relies on a large number of failure sample data, lacks physical mechanism constraints, and suffers a sharp decline in extrapolation and generalization ability when training samples are insufficient or operating conditions change, resulting in significantly reduced prediction accuracy and a susceptibility to overfitting. In actual subway relay operation and maintenance scenarios, relays have long lifespans and few failure samples, making this scheme difficult to implement effectively. (III) Traditional Mechanism Modeling Scheme This scheme establishes mathematical degradation equations based on arc physics and contact wear dynamics, and deduces the life decay law through numerical solution. The disadvantages of this scheme are: the pure physical mechanism model simplifies a large number of boundary conditions for the convenience of solution, making it difficult to fully characterize the complex nonlinear degradation process of relays, and the model parameters are difficult to tune and have weak adaptability, making it difficult to meet the needs of remaining life prediction under variable operating conditions.

[0004] In summary, existing methods cannot simultaneously leverage the advantages of physical mechanism constraints and data feature mining, making it difficult to achieve high-precision remaining life prediction of metro relays under small sample and variable operating conditions. This makes it unsuitable for supporting the practical needs of lean predictive operation and maintenance in urban rail transit. Therefore, a metro relay remaining life prediction method based on physical information neural networks is proposed to address these issues. Summary of the Invention

[0005] To address the technical problems of difficulty in mining degradation characteristics of subway relays under varying operating conditions and small sample sizes, low accuracy in predicting remaining life, and inability to intelligently formulate maintenance plans, this invention provides a subway relay remaining life prediction method based on physical information neural networks to solve the above problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the remaining life of subway relays based on physical information neural networks includes the following steps: Step S1: Multi-source operational data acquisition and preprocessing Multi-source timing monitoring data is collected during the operation of the subway relay. The multi-source timing monitoring data includes, but is not limited to: opening and closing time, contact resistance, circuit voltage, operating current, operating frequency, ambient temperature and humidity, and arc duration. The above data can be obtained through the existing sensor system and data acquisition device of the subway vehicle.

[0007] The following preprocessing operations were performed sequentially on the collected multi-source time-series monitoring data: (1) Outlier removal: The Raida criterion (3σ criterion) or interquartile range method is used to identify and remove data that exceed the normal range, so as to avoid the interference of abnormal data on model training.

[0008] (2) Missing value imputation: Linear interpolation or cubic spline interpolation is used to fill missing data points to ensure the continuity and integrity of time series data.

[0009] (3) Temporal normalization: Min-Max Normalization or Z-score normalization is used to map data of different dimensions to a unified dimension range, thereby eliminating the impact of dimension differences on model training.

[0010] (4) Sliding window sample construction: Using a fixed time window length, the data segments are truncated along the time sequence to form standardized input samples containing historical sequences and current observations, and a training dataset suitable for supervised learning is constructed.

[0011] Step S2: Modeling the physical mechanism of relay degradation The physical mechanisms of degradation of subway relays during long-term service are investigated, and partial differential equations and constraints characterizing the degradation process are established. The degradation physical mechanisms include the following three aspects: (1) Dynamic modeling of contact wear The contact wear dynamics equation is established based on the Archard classic wear model. The Archard wear volume equation is:

[0012] in, For contact wear volume, The wear coefficient of the contact material (based on the silver-based contact material, the value range is...). ), The contact pressure at which the contacts close. The total travel distance between the contact surfaces is the relative sliding distance. The Brinell hardness is the value of the contact material.

[0013] Relay operating frequency Using the independent variable, it is transformed into a differential degradation equation for wear thickness in a single action:

[0014] in, The wear thickness of the contact point during a single operation. The number of times the relay operates. The wear coefficient of the contact material is given, and the value range is based on the silver-based contact material. , The contact pressure at which the contacts close. The Brinell hardness of the contact material. This refers to the effective contact area of ​​the contact point.

[0015] The boundary conditions are , The critical wear thickness for contact failure, reaching The contact point is determined to be worn and unusable.

[0016] (2) Arc ablation aging modeling An arc erosion aging equation is established based on the correlation between arc energy and contact material loss. The formula for calculating contact weight loss caused by arc erosion during a single opening and closing operation is as follows:

[0017] in, Weight loss due to single arc ablation. The energy coefficient for converting arc heat energy into material vaporization loss is used; the value for silver contacts is 0.15~0.4. The total energy of the electric arc in a single action. The latent heat of vaporization of the contact metal material.

[0018] The total energy of the electric arc is measured by the collected circuit voltage. Operating current and duration of electric arc Perform time-series integration:

[0019] The continuous-time degeneracy differential equation is:

[0020] The boundary conditions include: the remaining effective mass of the contact. , The critical mass for arc ablation failure; arc time , and It was determined by statistical analysis of data collected from multiple sources on site.

[0021] (3) Modeling of mechanical fatigue attenuation A mechanical fatigue attenuation model is established based on the SN fatigue curve and Miner's linear cumulative damage theory.

[0022] SN stress-life empirical equation:

[0023] in, For alternating mechanical stress during contact opening and closing, This represents the fatigue limit number of actions under the current stress level. , This is the inherent fatigue constant of the contact material, determined by accelerated aging tests. Miner's linear fatigue cumulative damage formula:

[0024] in, For the first The actual operating frequency of the relay under various operating conditions To determine the critical number of fatigue failures under the corresponding stress level, mechanical fatigue failure is defined as when the cumulative damage D ≥ 1. The alternating mechanical stress Corrected for ambient temperature:

[0025] in, is the coefficient of linear expansion of the material. To measure the ambient temperature, The reference temperature is used as the basis for the damage constraint range. .

[0026] The three degradation physical mechanisms mentioned above together constitute the physical constraint framework of the partial differential equations for relay degradation evolution, defining the life decay boundary, operating condition influence constraints, and physical quantity change range, providing embedded physical rules for the subsequent construction of physical information neural networks.

[0027] Step S3: Construction and Training of Physical Information Neural Network (PINN) A Physics-Informed Neural Network (PINN) based on a fully connected deep neural network is constructed, and the degeneracy evolution partial differential equation and constraints established in step S2 are embedded as physical constraints into the network loss function.

[0028] The fully connected deep neural network comprises at least one input layer, multiple hidden layers, and one output layer. The number of nodes in the input layer corresponds to the dimension of the input features (including normalized multi-source monitoring data). The hidden layers adopt a fully connected structure, and the activation function is either the hyperbolic tangent function or an adaptive activation function. The output layer outputs the predicted value of the relay's degradation health or remaining service life.

[0029] The loss function consists of the following three parts: (1) Data fitting loss The mean squared error loss function is used to measure the difference between the network output and the measured data labels.

[0030] in, The number of training samples. This represents the predicted remaining lifetime output by the network. This corresponds to the actual remaining lifetime tag value.

[0031] (2) Physical mechanism residual loss The residuals of the metric network output at each sampling point of the degenerate evolution partial differential equation:

[0032] in, The number of sampling points is a physical constraint. This is a unified residual expression of the degradation evolution partial differential equation established in step S2 (including a comprehensive expression of the contact wear differential equation, the arc ablation continuous degradation equation, and the mechanical fatigue cumulative damage equation). and The time-domain and state-domain coordinates of the sampling points. and The partial derivatives of the network output with respect to time and state variables are obtained using automatic differentiation techniques.

[0033] (3) Boundary condition constraint loss :Measure the degree of deviation of the network output from the physical constraint boundary:

[0034] in, The number of sampling points is determined by boundary constraints. Boundary condition operators The output value of the network at the boundary sampling points. For a given boundary value (such as the critical wear thickness for contact failure) Critical mass of arc ablation failure wait).

[0035] The total loss function is a weighted sum of the three losses mentioned above:

[0036] in and These are the weighting coefficients for the physical mechanism residual loss and the boundary condition constraint loss, respectively, used to balance the impact of data-driven and physical constraints on network training.

[0037] The physical information neural network is trained using the standardized model input dataset obtained in step S1. During training, the dataset is divided into a training set, a validation set, and a test set according to a preset ratio.

[0038] The network parameters are optimized using the Adaptive Moment Estimation (Adam) optimizer, with an initial learning rate set to... During training, the learning rate is dynamically adjusted based on the validation set error using a decay strategy. An early stopping mechanism is introduced, terminating training when the validation set error no longer decreases within a preset number of rounds (e.g., 50 rounds) to prevent overfitting. Network parameters are simultaneously optimized through backpropagation, ensuring that the network output both fits the measured data and conforms to the physical degradation law of relays, ultimately yielding a trained physical information neural network model.

[0039] Step S4: Remaining life prediction The real-time collected relay operation data (after the same data preprocessing as in step one) is input into the trained physical information neural network model, and the model outputs the current degradation health and the predicted value of Remaining Useful Life (RUL).

[0040] The degradation health level is a dimensionless index between 0 and 1, used to characterize the current health status of the relay; the remaining service life prediction value is the remaining service time of the relay from the current time to the predicted failure time (expressed as the number of operations or the number of operating days).

[0041] Step S5, Health Status Classification The health status of relays is divided into four levels according to the proportion of remaining life to rated total life: healthy (remaining life > 80%), sub-healthy (50% to 80%), early warning (20% to 50%), and fault (≤ 20%), so as to realize the visualization assessment and display of status and provide intuitive basis for operation and maintenance decision-making.

[0042] Step S6: Intelligent maintenance linkage and full lifecycle traceability Based on the remaining life prediction results and health status classification, differentiated maintenance plans and spare parts replacement reminders are automatically generated; a full life cycle data archive of relays is established from commissioning, operation, degradation, maintenance to scrapping, so as to achieve full traceability of prediction results, operation and maintenance records and status changes.

[0043] This invention combines the advantages of both physical mechanisms and data-driven approaches. By embedding the degradation physical partial differential equations of relay contact wear dynamics, arc erosion aging, and mechanical fatigue decay into the neural network loss function as built-in constraints, the model can both fit measured data and strictly adhere to physical degradation laws during training. Compared to purely data-driven models, this invention maintains high prediction accuracy even with small samples and varying operating conditions, effectively solving the technical problems of weak generalization ability and overfitting in traditional models.

[0044] This invention significantly reduces the reliance on a large number of failure sample data for model training by introducing physical mechanism constraints. Even in actual subway relay operation and maintenance scenarios, where relay lifespan is long and failure samples are scarce, this invention still works effectively, demonstrating extremely low implementation barriers and practical engineering value.

[0045] This invention can accurately predict the failure time of relays and rationally plan maintenance windows, transforming urban rail vehicle operation and maintenance from the traditional regular inspection and post-fault repair mode to a predictive maintenance mode. This effectively avoids resource waste caused by over-maintenance and operational losses caused by sudden failures and outages. Furthermore, it can predict potential faults such as relay contact sticking and poor contact in advance, providing maintenance personnel with sufficient early warning time, effectively avoiding signal system malfunctions and traffic safety hazards, and significantly improving the operational reliability of subway signal and electrical control systems.

[0046] This invention reduces manpower input for ineffective maintenance, optimizes spare parts inventory management strategies, and lowers downtime losses and maintenance material costs, resulting in significant economic benefits. Furthermore, this invention has the potential to be extended to batch online monitoring and lifespan prediction of subway relays across the entire network, providing core technological support for the full lifecycle management of urban rail vehicles. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] Example 1: Please refer to Figure 1 The method for predicting the remaining life of subway relays based on physical information neural networks, as shown, includes the following steps: Step S1: Multi-source operational data acquisition and preprocessing This embodiment takes a certain type of subway DC relay as the research object. This relay has a rated voltage of DC 110V and a rated current of 10A, and is widely used in the control and signal circuits of subway trains. Through various types of sensors arranged on the relay and a subway vehicle data acquisition system, the following multi-source timing monitoring data are collected: (1) Opening and closing time: The time from when the coil is energized to when the main contact closes, and the time from when the coil is de-energized to when the main contact opens, are collected by the auxiliary contact signal. The sampling frequency is 1kHz.

[0050] (2) Contact resistance: The contact resistance of the relay contacts in the closed state was measured using the four-terminal method (Kelvin connection) with a sampling frequency of 10Hz.

[0051] (3) Circuit voltage: The circuit voltage u(t) of the main contacts of the relay is collected by a voltage sensor at a sampling frequency of 10kHz.

[0052] (4) Operating current: The operating current i(t) of the main contacts of the relay is collected by a current sensor at a sampling frequency of 10kHz.

[0053] (5) Frequency of operation: The frequency of operation per unit time is recorded by accumulating the number of relay opening and closing operations N; (6) Ambient temperature and humidity: A temperature and humidity sensor is used to collect the ambient temperature T and relative humidity at the relay installation location, with a sampling frequency of 0.1Hz.

[0054] (7) Duration of electric arc The duration of abnormal fluctuations in circuit voltage and current during the contact disconnection process is recorded using a high-speed data acquisition card.

[0055] The data acquisition cycle covers the entire process from the relay being put back into operation to its degradation and failure, with a total acquisition time of no less than 6 months.

[0056] Preprocessing of the collected multi-source time-series monitoring data: (1) Outlier removal: The Raida criterion is adopted. (Criteria), for each monitored variable, calculate its mean. and standard deviation It will fall Data points outside the specified range are identified as outliers and removed.

[0057] (2) Missing value imputation: For the gaps generated after removing outliers and the missing data points in the original collection, cubic spline interpolation is used to imput them to ensure the continuity and integrity of the time series data.

[0058] (3) Time series normalization: The minimum-maximum normalization method is used to map all monitored variables to the interval [0,1]. The normalization formula is as follows:

[0059] in, The original data values, and These are the minimum and maximum values ​​of the variable across the entire dataset, respectively.

[0060] (4) Sliding window sample construction: Set a fixed time window length L=100 (corresponding to the monitoring data of 100 consecutive actions of the relay), sliding step size S=10, and slide along the time sequence to extract data segments. Each sample contains the time sequence of all monitored variables within the window, and the label is the actual value of the remaining lifespan corresponding to the end time of the window (determined through full life cycle test data) to form a standardized supervised learning dataset.

[0061] Step S2: Modeling the physical mechanism of relay degradation This paper analyzes three physical mechanisms of degradation in subway relays and establishes partial differential equations and constraints for degradation evolution. (a) Contact Wear Dynamics Based on the Archard wear model, a differential degradation equation for wear thickness per single operation is established with the relay operation frequency N as the independent variable:

[0062] In this embodiment, the contact material is silver cadmium oxide (AgCdO), and the material wear coefficient K is taken as... Contact closure contact positive pressure The value is 0.5N, and the Brinell hardness H of the contact material is [value missing]. Effective contact area of ​​the contact point Values .

[0063] The wear thickness of a single operation was calculated. .

[0064] Boundary conditions: Critical wear thickness at contact failure The value is 0.1 mm, when the cumulative wear thickness is... Cumulatively reached When this happens, the contact point is determined to be worn out and has failed.

[0065] (ii) Arc erosion aging Based on the correlation between arc energy and contact material loss, the formula for calculating the weight loss of the contact point due to arc erosion during a single opening and closing operation is as follows:

[0066] in, Weight loss due to single arc ablation. The energy coefficient for converting arc heat energy into material vaporization loss is used; the value for silver contacts is 0.15~0.4. The total energy of the electric arc in a single action. The latent heat of vaporization of the contact metal material.

[0067] In this embodiment, the circuit voltage u(t), operating current i(t), and arc duration are collected. Perform time-series integration to solve for the total energy of the electric arc. Under typical operating conditions, the energy of a single arc breaking is approximately 0.5 J to 2.0 J.

[0068] The energy coefficient of electric arc heat energy converted into material vaporization loss The value is 0.25, representing the latent heat of vaporization of the silver contact metal material. Values .

[0069] The continuous-time degeneracy differential equation is:

[0070] Boundary conditions: Remaining effective mass of the contact Critical mass of contact failure The value is 60% of the initial mass; arc duration It is determined by the statistical distribution of data collected from multiple sources on site.

[0071] (iii) Mechanical fatigue attenuation Based on SN fatigue curves and Miner's linear cumulative damage theory: SN stress-life empirical equation:

[0072] in, For alternating mechanical stress during contact opening and closing, This represents the fatigue limit number of actions under the current stress level. , This is the inherent fatigue constant of the contact material, calibrated by accelerated aging tests.

[0073] In this embodiment, the inherent fatigue constant of the contact material is calibrated through accelerated aging tests. The value is 3.0. Values .

[0074] Miner's linear fatigue cumulative damage formula:

[0075] in, Let k be the actual operating frequency of the relay under the k-th operating condition. To determine the critical number of fatigue failures under the corresponding stress level, mechanical fatigue failure is defined as when the cumulative damage D≥1. Based on the measured working conditions, the operating frequency of the relays under each stress level was statistically analyzed, and the cumulative damage D was calculated accordingly. Mechanical fatigue failure was determined when the cumulative damage D ≥ 1.

[0076] Alternating mechanical stress Corrected for ambient temperature:

[0077] in, Here, T represents the coefficient of linear expansion of the material, and T represents the measured ambient temperature. The reference temperature is used as the base temperature, and the damage constraint range is... .

[0078] In this embodiment, the reference temperature The value is taken as 20℃, and the coefficient of linear expansion of the material is... Values Reference alternating mechanical stress The value is 300 MPa, and the damage constraint range is... .

[0079] The partial differential equations and constraints of the above three degradation mechanisms together form the physical constraint framework for relay degradation evolution, which is then embedded in the subsequent PINN model.

[0080] Step S3: Construction and Training of Physical Information Neural Network (PINN) The PINN model is built using Python as the programming language and based on the TensorFlow or PyTorch deep learning framework.

[0081] The network structure is as follows: Input layer: The number of nodes is the dimension of the input features. In this embodiment, each sample contains a time series of 7 types of monitoring variables. The number of nodes in the input layer is 7×L (L is the sliding window length, with a value of 100), that is, 700 input nodes.

[0082] Hidden layers: Six fully connected hidden layers are used, with 128 nodes in each layer. The hyperbolic tangent function (tanh) is selected as the activation function.

[0083] Output layer: 1 node, outputs the predicted remaining service life of the relay (in terms of the remaining number of operations).

[0084] Automatic Differentiation Module: Utilizes the Automatic Differentiation feature of deep learning frameworks to automatically calculate the network output during training. First and higher-order partial derivatives with respect to the input variables (time t and state variable x) and Used for physical mechanism residual loss The calculation.

[0085] The loss function is designed as follows: (1) Data fitting loss : Using the mean squared error loss function:

[0086] (2) Physical mechanism residual loss The residuals of the three degenerate physical equations:

[0087] in, , and These are the predicted values ​​of wear thickness, remaining mass, and cumulative damage output by the network, respectively.

[0088] (3) Boundary condition constraint loss :

[0089] In this context, the subscript "+" indicates the positive part, and the subscript "-" indicates the negative part.

[0090] The total loss function is: In this embodiment, the weighting coefficient The value is 0.1. The value is 0.01.

[0091] Model training: The standardized dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The Adam optimizer was used to optimize the network parameters, with an initial learning rate set to... The batch size is set to 64. A learning rate decay strategy is introduced during training—if the validation set error does not decrease for 10 consecutive epochs, the learning rate decays to 0.8 times the current value. An early stopping mechanism is introduced—training terminates when the validation set error does not decrease for 50 consecutive epochs. The maximum number of iterations is set to 5000 epochs.

[0092] Step S4: Remaining life prediction The real-time collected relay operation data (after the same data preprocessing as in step S1) is input into the trained PINN model. The model outputs the current degradation health status (a dimensionless index between 0 and 1) and the predicted value of the remaining service life (expressed as the number of remaining actions).

[0093] In this embodiment, the mean absolute percentage error (MAPE) between the model's prediction results on the test set and the actual remaining lifetime value is controlled within 8%, and the prediction accuracy is improved by about 20% compared with the traditional LSTM model.

[0094] Step S5, Health Status Classification The health status of relays is classified into four levels based on the proportion of remaining life to rated total life:

[0095] Step S6: Intelligent maintenance linkage and full lifecycle traceability Based on remaining life expectancy predictions and health status classifications, the system automatically generates differentiated maintenance plans and spare parts replacement reminders. Health status: Perform maintenance according to the standard maintenance cycle and record normal operation data.

[0096] Sub-health status: Shorten the monitoring period (e.g., from monthly to weekly) and increase the number of key parameter testing items.

[0097] Warning status: A maintenance work order is generated, and a spare parts procurement reminder is automatically pushed. It is recommended to arrange replacement within 2 weeks.

[0098] Fault Status: An emergency replacement work order is generated, a fault warning notification is pushed, and immediate replacement is recommended.

[0099] Simultaneously, a full lifecycle data archive for relays is established, encompassing commissioning, operation, degradation, maintenance, and scrapping. This archive includes: commissioning date, cumulative number of operations, maintenance records, health status assessment results at each stage, remaining lifespan prediction curve, actual replacement date, and replacement reason. This enables full traceability of prediction results, maintenance records, and status changes, providing data support for subsequent optimization of lifespan prediction models and adjustments to maintenance strategies for the same model of relays.

[0100] Example 2: Based on Example 1 of the present invention, those skilled in the art can also use the following alternative solutions to achieve the objective of the present invention: (a) Alternative solutions for model network structure In Embodiment 1 above, a fully connected deep neural network is used as the basic architecture of PINN. As an alternative, the basic fully connected PINN can be replaced with LSTM-PINN (Long Short-Term Memory Physical Information Neural Network) or GRU-PINN (Gated Recurrent Unit Physical Information Neural Network). LSTM-PINN and GRU-PINN add temporal memory units to the original physical constraint embedding, enabling them to better capture long-range temporal dependence features in multi-source monitoring data of relays, and are particularly suitable for relay life prediction scenarios where the degradation process has a long-term memory effect. The training methods, loss function construction, and physical constraint embedding methods of the two alternative network structures are basically the same as in the aforementioned embodiments, the only difference being that the basic network architecture is replaced by LSTM or GRU layers instead of fully connected layers.

[0101] (II) Physical Mechanism Extension Alternatives The above embodiment 1 considers three degradation physical mechanisms: contact wear dynamics, arc erosion aging, and mechanical fatigue attenuation. As an alternative, vibration and noise degradation mechanism constraints or insulation resistance degradation mechanism constraints can be added, depending on the specific type of relay and application scenario. For example, for high-power vehicle relays, which are subjected to continuous train operation vibration during service, vibration accumulation damage can be included as a fourth type of degradation mechanism in the physical constraint framework; for signal-specific relays, insulation resistance degradation is one of their main degradation modes, and the insulation resistance degradation equation can be included as a component of the physical constraints. The above extended scheme only requires adding corresponding physical mechanism residual loss terms and boundary condition constraint loss terms to the loss function; the rest of the implementation is the same as the aforementioned embodiment.

[0102] (III) Alternative solutions for data collection In Embodiment 1 above, relay operation data is collected via a wired method. As an alternative, an acquisition architecture of "edge sensing nodes + wireless LoRa / 5G IoT" can replace the traditional wired acquisition method. Specifically, an integrated edge sensing node is deployed at the relay installation location. This node has multiple types of sensors built-in (voltage / current / temperature / humidity, etc.) and performs preliminary edge-end data preprocessing (outlier removal, data compression, etc.). The data is then transmitted to a central server or cloud platform via a LoRa wireless communication module (suitable for long-distance, low-power scenarios) or a 5G wireless communication module (suitable for high-bandwidth, low-latency scenarios), enabling unmanned real-time data transmission on-site, reducing wiring costs, and improving the deployment flexibility of the data acquisition system.

[0103] (iv) Predicted Output Alternatives In the above embodiment 1, the model output is the degradation health and remaining service life prediction values. As an alternative, the following output functions can be extended based on the remaining service life prediction: (1) Fault type identification - by clustering or classifying the features of the network intermediate layer, the main degradation mode of the relay (wear-dominated, ablation-dominated or fatigue-dominated) can be identified to provide a basis for fault diagnosis; (2) Degradation trend visualization curve fitting - the degradation trajectory of the relay from commissioning to the current moment can be displayed graphically, and the complete degradation curve can be extrapolated to the failure threshold to enrich the information dimension of operation and maintenance decision-making.

[0104] (v) Deployment Architecture Alternatives The trained PINN model can be deployed in the following two architectures: (1) Local server deployment - The model is deployed on the local server of the metro depot or line control center. The real-time monitoring data of the relays of each line are collected to the central server through the local area network for centralized inference calculation. This is suitable for single line or station-level operation and maintenance management mode; (2) Cloud-edge collaborative edge computing deployment - The lightweight PINN model is deployed on the train edge computing gateway or trackside edge node. Each edge node independently completes the life prediction inference of the local relays. Only the prediction results and key alarm information are uploaded to the cloud operation and maintenance management platform for clustered monitoring and analysis. This is suitable for clustered operation and maintenance management scenarios of multiple lines and multiple stations. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining life of subway relays based on physical information neural networks, characterized in that, Includes the following steps: Step S1: Multi-source operational data acquisition and preprocessing Multi-source time-series monitoring data during the operation of subway relays are collected. The collected multi-source time-series monitoring data are then subjected to outlier removal, missing value imputation, time-series normalization, and sliding window sample construction to form a standardized model input dataset. Step S2: Modeling the physical mechanism of relay degradation The physical mechanism of degradation of subway relays was investigated, and partial differential equations and constraints characterizing the degradation process of relays were established. Step S3: Construction and Training of Physical Information Neural Network A fully connected deep neural network is constructed. The degradation evolution partial differential equation and constraint conditions established in step S2 are used as physical constraints and embedded into the network loss function. The fully connected deep neural network is trained using the standardized model input dataset obtained in step S1. The network parameters are optimized synchronously through backpropagation so that the network output fits the measured data and conforms to the physical degradation law of relays, thus obtaining the trained physical information neural network model. Step S4: Remaining life prediction The real-time collected relay operation data is input into the trained physical information neural network model, which outputs the current degradation health level and the predicted value of the remaining service life.

2. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step S1, the multi-source timing monitoring data includes opening and closing time, contact resistance, circuit voltage, operating current, operating frequency, ambient temperature and humidity, and arc duration. The outlier removal method uses the Raida criterion or interquartile range method to identify and remove data that exceeds the normal range; The missing value imputation uses linear interpolation or cubic spline interpolation to fill in the missing data; The time series normalization uses minimum-maximum normalization or Z-score standardization to map the data to a uniform dimension range. The sliding window sample construction uses a fixed time window length to slide and extract time series data, forming a standardized input sample that includes historical sequences and current observations.

3. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step S2, the degradation physical mechanism includes contact wear dynamics, arc ablation aging and mechanical fatigue decay, and the constraint conditions include life decay boundary, operating condition influence constraint and physical quantity change range. The degradation physical mechanism of the contact wear dynamics is established based on the Archard wear model, with the relay operating frequency as the reference. Using the independent variable, establish the differential degradation equation for the wear thickness in a single action: in, The wear thickness of the contact point during a single operation. The number of times the relay operates. The wear coefficient of the contact material is given, and the value range is based on the silver-based contact material. , The contact pressure at which the contacts close. The Brinell hardness of the contact material. The effective contact area of ​​the contact point; the constraint boundary conditions are as follows: , The critical wear thickness for contact failure, reaching The contact point is determined to be worn and unusable.

4. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step two, the degradation physical mechanism of arc ablation aging is established based on the correlation between arc energy and contact material loss. The formula for calculating the weight loss of the contact during a single opening and closing arc ablation is as follows: in, Weight loss due to single arc ablation. The energy coefficient for converting arc heat energy into material vaporization loss is used; the value for silver contacts is 0.15~0.

4. The total energy of the electric arc in a single action. The latent heat of vaporization of the contact metal material; The total energy of the electric arc is measured by the collected circuit voltage. Operating current and duration of electric arc Perform time-series integration: The continuous-time degeneracy differential equation is: The boundary conditions include: the remaining effective mass of the contact. , The critical mass for arc ablation failure; arc time , and It was determined by statistical analysis of data collected from multiple sources on site.

5. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step S2, the degradation physical mechanism of mechanical fatigue attenuation is established based on the SN fatigue curve and Miner's linear cumulative damage theory, including: SN stress-life empirical equation: in, For alternating mechanical stress during contact opening and closing, This represents the fatigue limit number of actions under the current stress level. , This is the inherent fatigue constant of the contact material, determined by accelerated aging tests. Miner's linear fatigue cumulative damage formula: in, Let k be the actual operating frequency of the relay under the k-th operating condition. To determine the critical number of fatigue failures under the corresponding stress level, mechanical fatigue failure is defined as when the cumulative damage D ≥ 1. The alternating mechanical stress Corrected for ambient temperature: in, is the coefficient of linear expansion of the material. To measure the ambient temperature, The reference temperature is used as the baseline; the damage constraint range is... .

6. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step S3, the loss function consists of three parts: data fitting loss, physical mechanism residual loss, and boundary condition constraint loss. The mathematical expression of the loss function is: in, For data fitting loss, the mean squared error loss function is used to measure the difference between the network output and the measured data labels; The physical mechanism residual loss is used to measure the residual of the network output at each sampling point of the degenerate evolution partial differential equation. Boundary condition constraint loss is used to measure the degree of deviation of the network output at the constraint boundary. and These are the weighting coefficients for physical mechanism residual loss and boundary condition constraint loss, respectively.

7. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 6, characterized in that: The data fitting loss The expression is: in, The number of training samples. This represents the predicted remaining lifetime output by the network. This corresponds to the actual remaining lifetime tag value; The physical mechanism of residual loss The expression is: in, The number of sampling points is a physical constraint. This provides a unified residual expression for degenerate evolution partial differential equations. and The time-domain and state-domain coordinates of the sampling points. and The partial derivatives of the network output with respect to time and state variables are obtained using automatic differentiation techniques. The boundary condition constraint loss The expression is: in, The number of sampling points is determined by boundary constraints. For boundary condition operators, The output value of the network at the boundary sampling points. For the given boundary values.

8. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step S3, the training process of the fully connected deep neural network includes: The standardized model input dataset is divided into training set, validation set and test set according to a preset ratio; An adaptive moment estimation optimizer is used to optimize network parameters, with an initial learning rate set to... During training, the learning rate is dynamically adjusted based on the validation set error using a learning rate decay strategy. An early stopping mechanism is introduced during training. Training is terminated when the validation set error no longer decreases within a preset number of consecutive rounds to prevent overfitting.

9. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 1, characterized in that: In step S4, after the current degradation health level and the predicted remaining useful life are output, It also includes a health status classification step: the relay health status is divided into four levels: healthy, sub-healthy, warning, and fault, according to the proportion of remaining life to rated total life, so as to realize the visual assessment and display of status.

10. The method for predicting the remaining life of subway relays based on physical information neural networks according to claim 9, characterized in that: After the health status classification is completed, the process also includes intelligent maintenance linkage and full life cycle traceability steps: Based on the remaining life expectancy prediction results and health status classification, the system automatically generates differentiated maintenance plans and spare parts replacement reminders. Establish a full lifecycle data archive for relays, from commissioning, operation, degradation, maintenance to scrapping, to achieve full traceability of prediction results, operation and maintenance records and status changes.

Images