Rail transit dynamic safety early warning system based on AI

Abstract

The invention discloses an AI-based rail transit dynamic safety early warning system, which comprises a sensing layer for deploying an intelligent sensor network with dynamically adjustable spacing, integrating a steel rail health monitoring suite to carry out microdefect detection on a steel rail, capturing vibration deformation characteristics under high-speed operation by a dynamic monitoring array, acquiring external environment data by an environment sensing module, and sending the external environment data to an early warning layer; preprocessing the collected data by utilizing an edge computing node; according to the cognitive layer, a line adaptive layer performs adaptive fusion on multi-line features through three-stage training of'basic training-meta training-fine tuning 'in combination with a dynamic weight generation mechanism, meanwhile, a causal cognitive module constructs a three-stage variable causal graph, a hybrid inference engine combines a CLIPS symbol inference engine, a Bayesian probability network and a D-S evidence theory, and the line adaptive layer performs multi-line feature fusion on the basis of the CLIPS symbol inference engine. Logic deduction and uncertainty quantification of fault attribution are realized; according to the decision-making layer, a dynamic threshold generator adjusts an early warning boundary in real time based on an LSTM prediction model, and an emergency plan engine recommends a maintenance strategy in combination with a digital twinborn simulation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit safety technology, and particularly relates to an AI-based rail transit dynamic safety early warning system. BACKGROUND

[0002] Under the background of the continuous expansion of the global rail transit network, the continuous improvement of the operation speed and the increasing pressure of safe operation and maintenance, the traditional railway safety guarantee mode relying on periodic manual inspection, fixed sensor array and post-fault processing has gradually shown the limitations of response lag, many blind areas and weak prediction ability. The breakthrough development of artificial intelligence technology, especially the mature application of deep learning algorithm in image recognition, sequence prediction and anomaly detection, combined with the real-time data collection capability of multi-source sensors realized by Internet of Things technology and the efficient data processing efficiency brought by the collaborative architecture of edge computing and cloud computing, provides technical support for building a new generation of rail transit dynamic safety early warning system.

[0003] The existing system may only analyze the correlation of variables through correlation coefficient or regression model, the essence of which is to capture the statistical correlation between variables, rather than causal relationship. In the scene of rail transit safety early warning, this limitation will directly lead to the lack of causal reasoning, which will affect the accuracy of early warning, the effectiveness of maintenance strategy and the long-term reliability of the system. Therefore, the AI-based rail transit dynamic safety early warning system is proposed. SUMMARY

[0004] The purpose of the present application is to solve the problems in the prior art, and an AI-based rail transit dynamic safety early warning system is proposed.

[0005] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0006] The AI-based rail transit dynamic safety early warning system comprises:

[0007] The perception layer: the intelligent sensor network with dynamically adjustable deployment interval is deployed, the rail health monitoring kit is integrated to detect the microscopic defects of the rail, the dynamic monitoring array captures the vibration and deformation characteristics under high-speed operation, the environment perception module obtains external environment data, and the edge computing node is used for pre-processing the collected data;

[0008] Cognition layer: including line adaptive layer and causal cognition module, the line adaptive layer combines dynamic weight generation mechanism for adaptive fusion of multi-line features through three-stage training of "basic training → meta-training → fine-tuning", the basic training extracts microscopic metallographic features of the rail through the material encoder, the curvature encoder models the topological relationship of the track curvature radius, and the speed encoder captures the high-speed dynamic response, the meta-training learns the parameter initialization of cross-line generalization, the fine-tuning adapts to the new line while retaining the old knowledge, and the dynamic weight generator generates Softmax weight based on the real-time line feature vector through the MLP network, and at the same time, the causal cognition module constructs a three-level variable causal graph, and the mixed reasoning engine combines the CLIPS symbolic reasoning engine, Bayesian probability network and D-S evidence theory to realize the logical deduction and uncertainty quantification of fault attribution.

[0009] Decision layer: the dynamic threshold generator adjusts the early warning boundary in real time based on the LSTM prediction model, and the emergency plan engine recommends the maintenance strategy combined with the digital twin simulation result.

[0010] The above technical scheme further comprises:

[0011] Further, the ultrasonic guided wave sensor is used to detect the microscopic cracks inside the rail, the eddy current probe is used to detect the surface cracks of the rail, the fiber Bragg grating strain gauge is used to measure the stress and strain of the rail, the six-degree-of-freedom accelerometer is used to measure the acceleration and three rotational angular velocities in three directions, the laser displacement sensor is used to measure the vertical / lateral deformation of the track, the millimeter wave radar is used to detect the obstacles in front of the track, the microclimate station is used to measure the rainfall, temperature and humidity, and the laser radar is used to measure the axle load in combination with the strain gauge.

[0012] Further, the edge computing node uses Kalman filtering for data filtering, and the state equation is expressed as x k =Ax k+1 +Bu k +w k , x k is a state vector, A is a state transition matrix, and w k is process noise.

[0013] Further, the line adaptive layer is trained through the specific steps of "basic training → meta-training → fine-tuning" three-stage training.

[0014] Basic training: feature extractor pre-training: initialize the general feature extraction ability of the material / curvature / speed encoder, the material encoder is pre-trained based on ImageNet, the last 1 layer is fine-tuned, and the output is f m , the curvature encoder pre-training models the topological relationship, and the output is f c , the speed encoder pre-training captures the time series pattern, and the output is f v , and the MLP parameters of the dynamic weight generator are randomly initialized.

[0015] Meta-training: Cross-path generalization initialization:

[0016] Inner loop: For each line T, update the parameters using the support set data: Where θ includes encoder parameters and MLP parameters, and the loss L sup The cross-entropy loss for the early warning task;

[0017] Outer loop: Update initial parameters with query set data: At this point, the MLP parameters are optimized along with the encoder parameters;

[0018] Fine-tuning: New route adaptation and knowledge retention:

[0019] Calculate the Fisher information matrix: Based on the five lines of meta-training data, calculate the sensitivity of parameters to the old task. Where, θ * These are the parameters after meta-training. This is the meta-training data;

[0020] Fine-tuning loss: L total =L new +λ∑ i F i (θ-θ * ) 2 L new The cross-entropy loss of the new route is represented by λ, which controls the strength of old knowledge retention.

[0021] Furthermore, the specific steps of the adaptive fusion of multiple line features by the line adaptive layer combined with the dynamic weight generation mechanism are as follows:

[0022] Real-time line feature vector definition: Input line feature vector L = [M, C, V], where:

[0023] M∈{0,1} represents the material type code;

[0024] C = 1 / R represents the reciprocal of the radius of curvature;

[0025] V∈[0,400] represents the real-time speed of the train;

[0026] Normalize the line feature vector:

[0027] Material type code C: Normalized to

[0028] The reciprocal of the radius of curvature V: normalized to

[0029] Train real-time speed M: No normalization required;

[0030] Normalized feature vector:

[0031] MLP structure and weight generation:

[0032] MLP architecture: two-layer fully connected network, hidden layer activation function is ReLU, output layer activation function is Softmax;

[0033] Forward propagation formula:

[0034] Hidden layer: where W1 is the weight matrix from the input layer to the hidden layer, and b1 is the bias vector;

[0035] Output layer: o = W2h + b z ∈R 3 where W2 is the weight matrix from the hidden layer to the output layer, and b z is the bias vector;

[0036] Softmax normalization:

[0037] Output weight W = [w m , w c , w v ], w m + w c + w v = 1;

[0038] Feature fusion and early warning output: the fusion feature is the weighted sum of the material / curvature / speed encoder outputs: f fusion = w m f m + w c f c + w v f v , where f m is the material encoder output, f c is the curvature encoder output, and f v is the speed encoder output, the fusion feature is input into the classifier, and the output is the safety warning level.

[0039] Further, the specific steps of the structural causal modeling are:

[0040] Variable classification and definition: root variables represent external inputs or controllable variables, directly affecting system state, intermediate variables represent intermediate state of the system derived from root variables through physical mechanisms or empirical models, and result variables represent system output or failure indicators directly caused by intermediate variables;

[0041] Causal relationship modeling: explicit causal relationships between variables based on directed acyclic graphs, thus establishing a structured knowledge framework to locate fault impact paths;

[0042] CLIPS symbolic reasoning engine: through expert rules, pattern matching is performed on observation data to quickly screen possible fault causes to realize deterministic logic;

[0043] Bayesian probability network: based on a probabilistic graphical model that is causal Figure One , the conditional dependence relationship between variables is quantified using a conditional probability table, and the posterior probability of each fault cause is calculated using Bayes' theorem to handle random uncertainty;

[0044] D-S evidence theory: define the trust degree of each fault cause through basic probability assignment, and use Dempster combination rule to fuse multi-source heterogeneous evidence from CLIPS rules, Bayesian network inference results and real-time sensor data. The fusion formula is Where m(A) is the fused basic probability assignment, K=∑ B∩C＝0 m1(B)·m2(C) is the conflict coefficient, is the total trust degree of proposition A, is the degree of not denying A.

[0045] Further, the dynamic threshold generator adjusts the specific steps of the early warning boundary based on the LSTM prediction model in real time:

[0046] Step 1: Data preprocessing and feature engineering: collect historical sensor data, normalize or standardize, eliminate dimensional differences, extract time series features, and construct input vectors;

[0047] Step 2: LSTM prediction model construction and training: build an LSTM network, including an input layer, a hidden layer, a fully connected layer, and an output layer:

[0048] LSTM unit recursion: h(t) = LSTM(X(t), h(t-1));

[0049] The fully connected layer outputs the predicted value:

[0050] Where h(t) is the hidden state, W σ , b σ is the output layer weight and bias;

[0051] The training target is to minimize the mean square error of the predicted value and the actual value y(t+1): Use the historical data set to train the model and update the network parameters through backpropagation;

[0052] Step 3: Dynamic threshold calculation: based on LSTM prediction value and the current working condition data, the dynamic threshold Threshold(t+1) is calculated, the dynamic threshold includes a basic threshold and a working condition adjustment term, the basic threshold is the LSTM prediction value reflecting the expected value of the future state, the working condition adjustment term is an additional margin based on the current working condition, and the dynamic threshold Threshold(t+1) calculation formula is: Wherein, alpha, beta, gamma are weight coefficients;

[0053] Step 4: Real-time early warning judgment: input real-time sensor data into the trained LSTM model to obtain the prediction value at the future time Combine the current working condition to calculate the dynamic threshold Threshold(t+1), if the real-time monitoring value y(t+1)>Threshold(t+1), the early warning is triggered; otherwise, keep normal state.

[0054] Further, the specific steps of the decision layer to generate cost-optimal / schedule-optimal / comprehensive balanced alternative solutions are:

[0055] Step 1: Digital twin model construction and real-time synchronization: based on the physical parameters of the rail transit system, a high-precision digital twin model is constructed;

[0056] Step 2: Maintenance strategy generation and simulation: generate multiple maintenance strategies for the current fault cause, and simulate the execution in the digital twin model, and record the simulation results of each strategy;

[0057] Step 3: Strategy evaluation and recommendation: based on the simulation results, calculate the comprehensive benefits of each strategy, and select the strategy with the highest comprehensive benefits to recommend to the operation and maintenance personnel, and the comprehensive benefit formula is: Wherein, S i is the comprehensive benefit, B is the maintenance budget, Delta sigma i is the stress reduction ratio of strategy i, C i is the cost of strategy i.

[0058] The present application has the following beneficial effects:

[0059] In the present application, by defining root variables, intermediate variables and result variables, combined with expert rules and data driving, the system can clearly show the cause and effect chain of fault occurrence, rather than only outputting correlation results. When the system warns, the cause and effect diagram can be traced back to the root variable or the intermediate variable, rather than only outputting a fuzzy conclusion. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is the system block diagram of the AI-based rail transit dynamic safety early warning system proposed by the present application;

[0061] Figure 2 System block diagram of the line adaptation layer proposed by the present application;

[0062] Figure 3 System block diagram of the causal cognition module proposed by the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0064] Please refer to Figures 1-3 The present application is an AI-based rail transit dynamic safety early warning system, which comprises:

[0065] The perception layer: the intelligent sensor network with dynamically adjustable deployment interval (5-50 meters) is deployed, the steel rail health monitoring kit (including ultrasonic guided wave sensor, eddy current probe, fiber Bragg strain gauge) is integrated to detect microscopic defects of the steel rail, the dynamic monitoring array (six-degree-of-freedom accelerometer, laser displacement sensor) captures the vibration deformation characteristics under high-speed operation, the environmental perception module (millimeter wave radar, micro weather station, laser radar) acquires external environmental data such as rainfall and axle load, and the edge computing node (NVIDIA Jetson AGX Orin) is used to pre-process the collected data;

[0066] Cognition layer: including line adaptive layer and causal cognition module, the line adaptive layer combines dynamic weight generation mechanism for adaptive fusion of multi-line features through three-stage training of "basic training → meta-training → fine-tuning", the basic training extracts microstructure features of materials such as steel rail U75V / U71Mn through material encoder (ResNet50 architecture), the curvature encoder (GAT graph neural network) models the topological relationship of track curvature radius (500-10000 meters), and the speed encoder (Transformer time module) captures the high-speed dynamic response of 350km / h, the meta-training learns the parameter initialization of cross-line generalization, the fine-tuning adapts to new lines while retaining old knowledge, the dynamic weight generator generates Softmax weight based on real-time line feature vector (material type, curvature radius, train speed) through MLP network to realize real-time adjustment of encoder weight, and at the same time, the causal cognition module constructs a three-level variable causal graph (including 28 root variables (rainfall, axle load, fastener torque, etc.), 17 intermediate variables (track irregularity index, rail stress amplitude, etc.), and 9 result variables (crack length, settlement rate, etc.)), and the mixed reasoning engine combines CLIPS symbolic reasoning engine (300+ expert rules), Bayesian probability network (15 conditional probability tables) and D-S evidence theory to realize logical deduction and uncertainty quantification of fault attribution.

[0067] Decision layer: the dynamic threshold generator adjusts the early warning boundary in real time based on the LSTM prediction model, and the emergency plan engine recommends maintenance strategies combined with digital twin simulation results.

[0068] In one embodiment, ultrasonic guided wave sensors are used to detect internal micro-cracks (such as transverse cracks and core damage) in the rail, eddy current probes are used to detect surface cracks (such as tread scratches) in the rail, fiber Bragg grating strain gauges are used to measure rail stress and strain, six-degree-of-freedom accelerometers are used to measure acceleration and three rotational angular velocities in three directions, laser displacement sensors are used to measure track vertical / lateral deformation, millimeter wave radar is used to detect obstacles (such as falling rocks and foreign objects) in front of the track, a micro-meteorological station is used to measure rainfall, temperature and humidity, and a laser radar is used in combination with a strain gauge to measure axle load.

[0069] In one embodiment, the edge computing node uses Kalman filtering for data filtering, and the state equation is expressed as x k = Ax k+1 + Bu k + w k , x k is a state vector, A is a state transition matrix, and w k is process noise.

[0070] In one embodiment, the line adaptive layer uses the following specific steps for three-stage training of "basic training → meta-training → fine-tuning":

[0071] Base training: Feature extractor pre-training: initialize general feature extraction capability of material / curvature / speed encoders, material encoder (ResNet50) is pre-trained based on ImageNet, fine-tune the last 1 layer, output f m , curvature encoder (GAT) pre-train to model topological relationship, output f c , speed encoder (Transformer) pre-train to capture temporal patterns, output f v , the MLP parameters of the dynamic weight generator are randomly initialized;

[0072] Meta-training (MAML): cross-line generalization initialization:

[0073] Inner loop: for each line T, update the parameters with support set data (100 samples): where θ contains encoder parameters (ResNet50 / GAT / Transformer) and MLP parameters, and loss L sup is the cross-entropy loss of the early warning task;

[0074] Outer loop: update the initial parameters with query set data (50 samples): At this time, the MLP parameters are optimized together with the encoder parameters, learning "how to generate weights according to line features";

[0075] Fine-tuning (EWC): new line adaptation and knowledge preservation:

[0076] Calculate the Fisher information matrix: based on the meta-training data of 5 lines, calculate the sensitivity of the parameters to the old task: where θ * is the parameter after meta-training, is the meta-training data;

[0077] Fine-tuning loss: L total = L new + λ∑ i F i (θ-θ * ) 2 , where L new is the cross-entropy loss of the new line, λ controls the strength of old knowledge preservation, and the MLP parameters are further adapted to the new line features in this stage.

[0078] In one embodiment, the specific steps of the line adaptive layer combining the dynamic weight generation mechanism for adaptive fusion of multi-line features are as follows:

[0079] Real-time line feature vector definition: input line feature vector L = [M, C, V], where:

[0080] M ∈ {0, 1} represents material type code (0 = U75V, 1 = U71Mn);

[0081] C = 1 / R represents the reciprocal of the radius of curvature (R ∈ [500, 10000] meters, C ∈ [0.0001, 0.002]);

[0082] V ∈ [0, 400] represents the real-time speed of the train (km / h);

[0083] Normalize the line feature vector:

[0084] Material type code C: normalized to

[0085] Reciprocal of the radius of curvature V: normalized to

[0086] Real-time speed of the train M: no need for normalization (0 / 1 direct input);

[0087] Normalized feature vector:

[0088] MLP structure and weight generation:

[0089] MLP architecture: two-layer fully connected network, hidden layer activation function is ReLU, output layer activation function is Softmax;

[0090] Forward propagation formula:

[0091] Hidden layer: where W1 is the weight matrix from the input layer to the hidden layer, and b1 is the bias vector;

[0092] Output layer: o = W2h + b z ∈ R 3 , where W2 is the weight matrix from the hidden layer to the output layer, and b z is the bias vector;

[0093] Softmax normalization:

[0094] Output weight W = [w m , w c , w v ], w m + w c + w v = 1;

[0095] Feature fusion and early warning output: the fused feature is the weighted sum of the material / curvature / speed encoder outputs: f fusion = w m fm +w c f c +w v f v where f m is the material encoder (ResNet50) output, f c is the curvature encoder (GAT) output, and f v is the speed encoder (Transformer) output, which is fused into a classifier (e.g., fully connected layer + Softmax) to output the safety warning level (0 / 1 / 2).

[0096] In one embodiment, the specific steps of the structural causal modeling are:

[0097] Variable classification and definition: root variables (e.g., rainfall, axle load, fastener torque, etc.) represent external inputs or controllable variables that directly affect the system state, intermediate variables (e.g., track irregularity index, rail stress amplitude, etc.) represent the intermediate state of the system derived from root variables through physical mechanisms or empirical models, and result variables (e.g., crack length, settlement rate, etc.) represent system outputs or failure indicators directly caused by intermediate variables;

[0098] Causal relationship modeling: based on directed acyclic graph (DAG), the causal relationship between variables is determined (if variable X has a direct impact on Y, a directed edge X→Y is added, such as rainfall indirectly causing the track irregularity index to rise by affecting the degree of roadbed compaction, axle load directly causing rail stress amplitude to exceed the limit by increasing wheel-rail contact force, and the rationality of the edge is verified through expert knowledge (such as railway engineering specifications) and historical data (such as sensor records) to exclude spurious correlations, such as "temperature" and "crack length" have no direct causal relationship, but may indirectly affect the intermediate variable "rail stress amplitude" through "thermal expansion and contraction"), thus establishing a structured knowledge framework to locate the fault impact path;

[0099] CLIPS symbolic reasoning engine: through expert rules, the observed data (such as crack length exceeding the threshold) is pattern-matched to quickly screen possible fault causes (such as "rail stress amplitude exceeding the limit" and "track irregularity index being too high") to realize deterministic logical reasoning;

[0100] Bayesian probability network: based on a probabilistic graphical model related to causal Figure One relationship, the conditional dependency relationship between variables is quantified using a conditional probability table (CPT), and the posterior probability of each fault cause is calculated through Bayes' theorem to handle random uncertainty;

[0101] E-S evidence theory: define the belief degree of each fault cause through basic probability assignment (BPA), fuse the multi-source heterogeneous evidence from CLIPS rules, Bayesian network inference results and real-time sensor data (such as stress sensor values) using Dempster combination rule, and the fusion formula is expressed as wherein m(A) is the basic probability assignment after fusion, K = ∑ B∩C＝0 m1(B)·m2(C) is the conflict coefficient, is the total belief degree to proposition A, is the degree of not denying A.

[0102] Example:

[0103] Evidence 1 (CLIPS rule): m1(stress over limit) = 0.7, m1(other) = 0.3;

[0104] Evidence 2 (Bayesian network): m2(stress over limit) = 0.6, m2(unevenness too high) = 0.4;

[0105] After fusion:

[0106] K = m1(stress over limit)·m2(unevenness too high) + m1(other)·m2(stress over limit) + m1(other)·m2(unevenness too high) = 0.58

[0107] m(stress over limit) = (0.7×0.6 + 0.3×0.6) / (1-0.58) ≈ 1.428.

[0108] In one embodiment, the dynamic threshold generator adjusts the specific steps of the early warning boundary based on the LSTM prediction model in real time:

[0109] Step 1: Data preprocessing and feature engineering: collect historical sensor data, perform normalization (Min-Max normalization) or standardization (Z-score standardization), eliminate dimension differences, extract time series features (such as sliding window mean, change rate), construct input vector X(t) = [I(t-k), σ(t-k), R(t-k), W(t-k), V(t-k)], track irregularity index I(t), rail stress amplitude σ(t), rainfall R(t), axle load W(t), train speed V(t), wherein k is the time window length (such as the past 1 hour of data).

[0110] Step 2: LSTM prediction model construction and training: build an LSTM network, including an input layer, a hidden layer, a fully connected layer and an output layer:

[0111] LSTM unit recursion: h(t) = LSTM(X(t), h(t-1));

[0112] Fully connected layer output prediction value:

[0113] where h(t) is the hidden state, W σ , b σ are the output layer weights and bias;

[0114] The training objective is to minimize the mean squared error between the prediction value and the actual value y(t+1): Use the historical data set (such as the past 1 month data) to train the model, update the network parameters by back propagation (Adam optimizer);

[0115] Step 3: Dynamic threshold calculation: based on LSTM prediction value and current working condition data, calculate the dynamic threshold Threshold(t+1), the dynamic threshold includes the basic threshold and the working condition adjustment term, the basic threshold is the LSTM prediction value reflecting the expected value of the future state, the working condition adjustment term is an additional margin based on the current working condition, to avoid false negatives caused by prediction errors under variable working conditions, the dynamic threshold Threshold(t+1) calculation formula is: where α, β, γ are weight coefficients (obtained by historical data regression fitting);

[0116] Step 4: Real-time early warning judgment: input real-time sensor data into the trained LSTM model to get the prediction value at the future time Combine the current working condition to calculate the dynamic threshold Threshold(t+1), if the real-time monitoring value y(t+1) > Threshold(t+1), trigger the alarm; otherwise, keep normal state.

[0117] In one embodiment, the decision layer generates the specific steps of the cost optimal / period optimal / comprehensive balanced alternative solution:

[0118] Step 1: Digital twin model construction and real-time synchronization: based on the physical parameters of the rail transit system (such as track structure (rail type, fastener spacing), train characteristics (axle load, suspension system), environmental factors (temperature, rainfall)) to construct a high-precision digital twin model;

[0119] Step 2: Maintenance strategy generation and simulation: for the current fault reason (such as "rail stress out of limits"), generate multiple maintenance strategies and simulate execution in the digital twin model, record the simulation results of each strategy;

[0120] Strategy 1 (adjust fastener torque): adjust fastener torque from T0=50 N·m to T1=60 N·m, after simulation, rail stress amplitude is reduced by Δσ=15%, maintenance cost C1=5000 yuan;

[0121] Strategy 2 (limit speed operation): limit train speed from V0=120 km / h to V1=100 km / h, after simulation, rail stress amplitude is reduced by Δσ=20%, but it affects operation efficiency (equivalent cost C2=2000 yuan (such as passenger delay cost));

[0122] Strategy 3 (replace rail): replace with high-strength rail, after simulation, rail stress amplitude is reduced by Δσ=50%, maintenance cost C3=50000 yuan.

[0123] Step 3: strategy evaluation and recommendation: based on the simulation results, calculate the comprehensive benefit of each strategy (such as benefit=stress reduction ratio-cost / budget), select the strategy with the highest comprehensive benefit to recommend to the operation and maintenance personnel. The comprehensive benefit formula is: Where S i is the comprehensive benefit, B is the maintenance budget, Δσ i is the stress reduction ratio of strategy i (such as 15%, 20%, 50%), C i is the cost of strategy i.

[0124] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based dynamic safety early warning system for rail transit, characterized in that: include: Perception layer: Deploys a smart sensor network with dynamically adjustable spacing, integrates a rail health monitoring kit to detect microscopic defects in the rail, a dynamic monitoring array to capture vibration and deformation characteristics under high-speed operation, an environmental perception module to acquire external environmental data, and uses edge computing nodes to preprocess the collected data; The cognitive layer includes a line adaptation layer and a causal cognition module. The line adaptation layer undergoes three stages of training: "basic training → meta-training → fine-tuning". It combines a dynamic weight generation mechanism to adaptively fuse features from multiple lines. Basic training extracts the microscopic metallographic features of the rails through a material encoder, models the topological relationship of the track curvature radius through a curvature encoder, and captures the high-speed dynamic response through a speed encoder. Meta-training learns parameter initialization for cross-line generalization. Fine-tuning retains old knowledge when adapting to new lines. The dynamic weight generator generates Softmax weights based on real-time line feature vectors through an MLP network. Meanwhile, the causal cognition module constructs a three-level variable causal graph. The hybrid inference engine combines the CLIPS symbolic inference engine, Bayesian probabilistic network, and DS evidence theory to achieve logical deduction and uncertainty quantification of fault attribution. Decision-making level: The dynamic threshold generator adjusts the early warning boundary in real time based on the LSTM prediction model, and the emergency plan engine recommends maintenance strategies based on the results of digital twin simulation.

2. The AI-based dynamic safety early warning system for rail transit according to claim 1, characterized in that, Ultrasonic guided wave sensors are used to detect microscopic cracks inside the rails, eddy current probes are used to detect cracks on the rail surfaces, fiber optic strain gauges are used to measure rail stress and strain, six-degree-of-freedom accelerometers are used to measure acceleration in three directions and rotational angular velocities in three directions, laser displacement sensors are used to measure vertical / lateral deformation of the track, millimeter-wave radar is used to detect obstacles in front of the track, micro-weather stations are used to measure rainfall, temperature, and humidity, and lidar combined with strain gauges is used to measure axle load.

3. The AI-based dynamic safety early warning system for rail transit according to claim 1, characterized in that, The edge computing node uses Kalman filtering for data filtering, and its state equation is represented as x. k =Ax k+1 +Bu k +w k x k Let A be the state vector, and let w be the state transition matrix. k This is process noise.

4. The AI-based dynamic safety early warning system for rail transit according to claim 1, characterized in that, The adaptive layer of the line is trained through a three-stage process: "basic training → meta-training → fine-tuning". Basic training Feature Extractor Pre-training: Initializes the general feature extraction capabilities of the material / curvature / velocity encoder. The material encoder is pre-trained based on ImageNet, with fine-tuning of the last layer, outputting f. m The curvature encoder is pre-trained to model topological relationships and outputs f. c The speed encoder is pre-trained to capture temporal patterns and outputs f. v The MLP parameters of the dynamic weight generator are randomly initialized; Meta-training: Cross-path generalization initialization: Inner loop: For each line T, update the parameters using the support set data: Where θ includes encoder parameters and MLP parameters, and the loss L sup The cross-entropy loss for the early warning task; Outer loop: Update initial parameters with query set data: At this point, the MLP parameters are optimized along with the encoder parameters; Fine-tuning: New route adaptation and knowledge retention: Calculate the Fisher information matrix: Based on the five lines of meta-training data, calculate the sensitivity of parameters to the old task. Where, θ * These are the parameters after meta-training. This is the meta-training data; Fine-tuning loss: L total =L new +λ∑ i F i (θ-θ * ) 2 L new The cross-entropy loss of the new route is represented by λ, which controls the strength of old knowledge retention.

5. The AI-based dynamic safety early warning system for rail transit according to claim 4, characterized in that, The specific steps of the adaptive fusion of multiple line features by the line adaptive layer combined with the dynamic weight generation mechanism are as follows: Real-time line feature vector definition: Input line feature vector L = [M, C, V], where: M∈{0,1} represents the material type code; C = 1 / R represents the reciprocal of the radius of curvature; V∈[0,400] represents the real-time speed of the train; Normalize the line feature vector: Material type code C: Normalized to The reciprocal of the radius of curvature V: normalized to Train real-time speed M: No normalization required; Normalized eigenvectors: MLP Structure and Weight Generation: MLP architecture: Two fully connected layers, with ReLU activation function in the hidden layer and Softmax activation function in the output layer; Forward propagation formula: Hidden layer: Where W1 is the weight matrix from the input layer to the hidden layer, and b1 is the bias vector; Output layer: o = W2h + b z ∈R 3 Where W2 is the weight matrix from the hidden layer to the output layer, b z It is the bias vector; Softmax normalization: Output weight W = [w m ,w c ,w v ], w m +w c +w v =1; Feature fusion and early warning output: The fused features are a weighted sum of the outputs from the material / curvature / velocity encoders: f fusion =w m f m +w c f c +w v f v , where f m For material encoder output, f c For the curvature encoder output, f v The speed encoder output is fused with the feature input classifier to output a safety warning level.

6. The AI-based dynamic safety early warning system for rail transit according to claim 5, characterized in that, The specific steps of the structural causal modeling are as follows: Variable classification and definition: Root variables represent external inputs or controllable variables that directly affect the system state; intermediate variables are derived from root variables through physical mechanisms or empirical models and reflect the intermediate state of the system; and result variables represent system outputs or fault indicators directly caused by intermediate variables. Causal relationship modeling: Based on directed acyclic graphs, the causal relationships between variables are clarified, thereby establishing a structured knowledge framework to locate the path of failure impact; CLIPS Symbolic Reasoning Engine: It uses expert rules to pattern match observed data, quickly filters possible causes of failure, and achieves deterministic logic. Bayesian probabilistic networks: Based on a probabilistic graphical model consistent with causal graphs, they use conditional probability tables to quantify the conditional dependencies between variables and calculate the posterior probability of each fault cause using Bayes' theorem to handle random uncertainty. Dempster's evidence theory defines the confidence level of each fault cause by assigning basic probability values. It then uses Dempster's combination rule to fuse multi-source heterogeneous evidence from CLIPS rules, Bayesian network inference results, and real-time sensor data. The fusion formula is expressed as follows: Where m(A) is the base probability assignment after fusion, and K = ∑ B∩C＝0 m1(B)·m2(C) is the conflict coefficient, Bel The total confidence level for proposition A, To the degree to which A is not denied.

7. The AI-based dynamic safety early warning system for rail transit according to claim 1, characterized in that, The specific steps of the dynamic threshold generator in adjusting the early warning boundary in real time based on the LSTM prediction model are as follows: Step 1: Data preprocessing and feature engineering: Collect historical sensor data, normalize or standardize it, eliminate differences in units, extract time series features, and construct input vectors; Step 2: LSTM Prediction Model Construction and Training: Construct an LSTM network, including an input layer, hidden layers, fully connected layers, and an output layer. LSTM cell recursion: h(t) = LSTM(X(t), h(t-1)); Fully connected layer outputs predicted values: Where h(t) is the hidden state, W σ b σ For output layer weights and biases; The training objective is to minimize the predicted value. Mean square error of the actual value y(t+1): The model is trained using a historical dataset, and the network parameters are updated via backpropagation. Step 3: Dynamic threshold calculation: based on LSTM prediction values Based on the current operating data, calculate the dynamic threshold Threshold(t+1). The dynamic threshold includes a base threshold and an operating condition adjustment term. The base threshold is the LSTM prediction value. The expected value reflects the future state. The operating condition adjustment term is an additional margin based on the current operating condition. The dynamic threshold Threshold(t+1) is calculated using the following formula: Where α, β, and γ are weighting coefficients; Step 4: Real-time early warning judgment: Input real-time sensor data into the trained LSTM model to obtain the predicted value for future time moments. The dynamic threshold Threshold(t+1) is calculated based on the current operating conditions. If the real-time monitoring value y(t+1) > Threshold(t+1), an early warning is triggered; otherwise, the normal state is maintained.

8. The AI-based dynamic safety early warning system for rail transit according to claim 1, characterized in that, The specific steps by which the decision-making level generates alternative solutions that are optimal in cost, optimal in schedule, and balanced overall are as follows: Step 1: Digital Twin Model Construction and Real-time Synchronization: Construct a high-precision digital twin model based on the physical parameters of the rail transit system; Step 2: Maintenance Strategy Generation and Simulation: Based on the current cause of the failure, generate multiple maintenance strategies, simulate their execution in the digital twin model, and record the simulation results of each strategy; Step 3: Strategy Evaluation and Recommendation: Based on the simulation results, calculate the overall benefit of each strategy, and select the strategy with the highest overall benefit to recommend to the operations and maintenance personnel. The formula for the overall benefit is: Among them, S i For overall benefits, B represents the maintenance budget, and Δσ i C represents the stress reduction ratio of strategy i. i Let be the cost of strategy i.

Images