Fireproof flexible connection state monitoring system for motor train unit based on embedded sensor network

Abstract

The present application relates to the technical field of EMU equipment state monitoring, and discloses a fireproof soft connection state monitoring system for EMU based on embedded sensor network. The system comprises an embedded distributed sensor network, an edge computing node and an adaptive control scheduler. The embedded distributed sensor network collects acoustic emission signals and infrared thermal imaging data of the fireproof soft connection of the EMU. The edge computing node is provided with a multi-physical field coupling digital twin body of a fusion material aging constitutive model, receives the above data and generates continuous prediction values of insulation degradation gradient, fireproof layer failure probability and mechanical fatigue damage degree. The adaptive control scheduler adopts an adaptive federated edge learning algorithm to aggregate model update parameters of the same type of soft connection in multiple carriages. The present application is used for realizing non-invasive online monitoring and early warning of the whole life cycle of the fireproof soft connection of the EMU, avoiding data transmission packet loss and leakage in the strong electromagnetic environment of the EMU, and reducing the overall power consumption of the sensor network.

Description

Technical Field

[0001] This invention relates to the field of EMU equipment condition monitoring technology, and discloses a fireproof soft connection condition monitoring system for EMUs based on embedded sensor networks. Background Technology

[0002] Existing fire-resistant flexible connection monitoring solutions for high-speed trains typically employ single-point temperature sensing technology for anomaly detection. In practice, thermistors or thermocouples are fixedly installed at the flexible connection terminals or on the outer surface of the fireproof layer, transmitting temperature data to the onboard central control unit via wired or wireless means. The central control unit has a preset temperature alarm threshold; when the real-time collected temperature data exceeds this threshold, the system triggers an alarm mechanism. At the data transmission level, existing solutions typically use a fixed sampling frequency and transmission power to centrally upload the raw temperature sensing data to the central control unit for centralized processing and judgment.

[0003] The problem with the existing technical solutions is that the monitoring mechanism based on a single-point fixed temperature threshold cannot provide early warning of the latent degradation state of fireproof flexible connections. Since the degradation of fireproof flexible connections is a gradual process involving the coupled evolution of multiple physical fields such as temperature, stress, and partial discharge, the surface temperature has not yet exceeded the fixed alarm threshold in the early stages of insulation aging or mechanical fatigue. Single-point temperature sensing cannot capture the continuous changes in the internal insulation degradation gradient and mechanical fatigue damage. The existing fixed threshold mechanism can only trigger a post-mortem alarm in the late stage when the temperature rises sharply, thus losing the time window for proactive intervention in the early stages of degradation. Summary of the Invention

[0004] The purpose of this invention is to provide a solution that can effectively address the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is: a fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks, comprising: An embedded distributed sensor network is used to collect acoustic emission signals and infrared thermal imaging data from the fireproof flexible connection of the EMU. Edge computing nodes are connected to embedded distributed sensor networks and have built-in multi-physics field coupled digital twins that integrate material aging constitutive models. They receive acoustic emission signals and infrared thermal imaging data to generate continuous prediction values ​​of insulation degradation gradient, fireproof layer failure probability and mechanical fatigue damage. The adaptive control scheduler communicates with the embedded distributed sensor network and the edge computing node respectively. It uses an adaptive federated edge learning algorithm to aggregate the model update parameters of the same type of soft connection in multiple carriages, and dynamically adjusts the monitoring priority and early warning threshold of the embedded distributed sensor network based on continuous prediction values ​​to form a closed-loop control.

[0006] Preferably, when adjusting the state of the embedded distributed sensor network, the adaptive control scheduler adopts a dynamic time slot allocation method that integrates non-orthogonal multiple access power domain multiplexing and Markov decision process to map the EMU operating conditions into a Markov state space. The joint reward function is to maximize the network lifetime and minimize the data acquisition delay. The transmit power and access time slot ratio of each sensor node are output through a deep deterministic policy gradient network.

[0007] Preferably, in the process of constructing a multi-physics coupled digital twin, a flow field order reduction projection method based on intrinsic orthogonal decomposition is used to process the partial differential equations in the material aging constitutive model, mapping the high-dimensional temperature field and stress field to a low-dimensional subspace, and combining the simulated annealing algorithm to optimize the boundary conditions in the low-dimensional subspace, so as to realize the real-time parameter evolution and deduction of the digital twin on the edge computing node.

[0008] Preferably, when processing acoustic emission signals and infrared thermal imaging data, a dual-modal feature alignment method that integrates micro-Doppler feature extraction and visual Transformer mechanism is adopted. The acoustic emission signal is converted into a time-frequency map and the instantaneous frequency trajectory of micro-Doppler is extracted as a query sequence. The patches of infrared thermal imaging data are embedded as key-value pairs. The adaptive fusion of acoustic and thermal heterogeneous data in the latent feature space is achieved through a cross-attention mechanism.

[0009] Preferably, when aggregating model update parameters, the adaptive federated edge learning algorithm introduces a clonal selection mechanism based on an artificial immune system to screen local models. It treats global model parameters as antigens and local model parameters uploaded by edge computing nodes in each carriage as antibodies. By calculating affinity and concentration suppression strategies, it removes inferior local parameters containing abnormal operating condition noise and aggregates only the model gradients corresponding to high-affinity antibodies.

[0010] Preferably, when calculating the joint reward function, the deep deterministic policy gradient network introduces non-cooperative game theory to construct a power competition model among sensor nodes. The signal-to-interference-plus-noise ratio and remaining energy of each node are used as utility functions. By finding the Nash equilibrium point, the action space output of the deep deterministic policy gradient network is constrained, so that the sensor nodes spontaneously converge to the optimal power allocation strategy when competing for shared wireless channel resources.

[0011] Preferably, after optimizing the boundary conditions in the low-dimensional subspace, the ensemble Kalman filter method in meteorological data assimilation is introduced to correct the error in the real-time parameter evolution of the digital twin. The boundary conditions output by the simulated annealing algorithm are used as the background field, and the real-time acquired acoustic emission signals and infrared thermal imaging data are used as the observation field. Covariance expansion and localization processing are used to suppress the filtering divergence in the nonlinear constitutive model derivation process.

[0012] Preferably, after achieving adaptive fusion of acoustic and thermal heterogeneous data in the latent feature space, word vector mapping and graph neural network methods from natural language processing are introduced to construct a fault feature semantic graph. The fused feature vectors are regarded as word sequences, and the contextual co-occurrence relationship between features is learned through the Skip-gram model. A dynamic topology graph is constructed with feature vectors as nodes and co-occurrence similarity as edge weights. Neighbor node information is aggregated through graph convolutional networks to quantify the degree of correlation of partial discharge.

[0013] Preferably, when aggregating the model gradient corresponding to high affinity antibodies, the Shapley value mechanism design theory in microeconomics is introduced to evaluate the marginal contribution of each carriage edge computing node to the global model accuracy. The differences in the operating conditions of the EMU are regarded as a coalition game environment. By calculating the average marginal revenue of local model parameters in all possible combinations of carriage nodes, the weight coefficients of each node gradient in the global aggregation formula are adaptively allocated.

[0014] Preferably, when adaptively allocating the weight coefficients of each node gradient in the global aggregation formula, active disturbance rejection control technology from cybernetics is introduced to respond to the dynamic fluctuations of the weight coefficients. The expected weight value calculated by Shapley value is used as the tracking reference signal, and the weight mutation caused by the electromagnetic interference of the EMU is regarded as the total disturbance. The disturbance is estimated and compensated in real time using an extended state observer, and the trajectory of weight coefficient change is smoothed by a nonlinear state error feedback law to maintain the stability of adaptive federated edge learning.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves continuous prediction of insulation degradation gradient, fireproof layer failure probability, and mechanical fatigue damage levels for fireproof flexible connections by embedding a multiphysics-coupled digital twin with a fused material aging constitutive model within edge computing nodes and processing acoustic emission signals and infrared thermal imaging data using a dual-modal feature alignment method. This transforms the original reactive surface temperature alarm into an advanced warning based on the evolution of multiphysics parameters, enabling the monitoring system to acquire the continuous changes in the degradation gradient as early as the initial stages of insulation degradation or mechanical fatigue within the fireproof flexible connection. This solves the technical problem of existing technologies being unable to provide advanced warnings of latent degradation states.

[0016] 2. This invention employs an adaptive federated edge learning algorithm to aggregate model update parameters for soft connections of the same type across multiple carriages. By performing local model filtering and gradient aggregation at the edge, it avoids the centralized uploading of raw sensor data under strong electromagnetic interference conditions of high-speed trains, reducing the risk of data loss and leakage. The adaptive control scheduler dynamically adjusts the monitoring priority and early warning threshold of the embedded distributed sensor network based on continuous prediction values. Combined with a dynamic time slot allocation method based on Markov decision processes, it balances the monitoring accuracy and overall network power consumption under complex operating conditions. In the digital twin derivation process, a reduced-order projection method based on intrinsic orthogonal decomposition and ensemble Kalman filter error correction are introduced to reduce the solution overhead of high-dimensional partial differential equations at edge computing nodes and suppress filter divergence in the nonlinear constitutive model derivation process. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall workflow of the present invention. Figure 2 This is a flowchart of the dual-modal feature alignment and fault semantic graph construction process of the present invention; Figure 3 This is a flowchart of the digital twin order reduction and error correction process of the present invention; Figure 4 This is a flowchart of the local model selection and weight allocation process of the present invention; Figure 5 This is a flowchart of the dynamic time slot allocation and power control of the present invention; Figure 6 This is a flowchart of the smoothing weight coefficient for the active disturbance rejection control of the present invention. Detailed Implementation

[0018] Example 1: Please refer to Figure 1 This embodiment is a basic implementation of the fireproof soft connection status monitoring system for EMUs based on embedded sensor networks. The system includes an embedded distributed sensor network, an edge computing node, and an adaptive control scheduler, which are sequentially coupled in communication. The adaptive control scheduler establishes bidirectional communication connections with the embedded distributed sensor network and the edge computing node, forming a closed-loop control link.

[0019] The embedded distributed sensor network is deployed at each fireproof flexible connection installation location in each carriage of the EMU. Each fireproof flexible connection is configured with a set of sensor node units. Each set of sensor node units includes an acoustic emission sensor module, an infrared thermal imaging module, an embedded processing unit, a wireless communication module, and a power supply module. The acoustic emission sensing module employs a piezoelectric ceramic sensor with a resonant frequency of 100kHz-1MHz, coupled and installed at the connection between the conductive terminal and the flexible conductor of the fireproof flexible connector. It is configured to collect elastic stress waves generated during insulation degradation, partial discharge, and mechanical fatigue crack propagation within the fireproof flexible connector, i.e., acoustic emission signals. The infrared thermal imaging module uses an uncooled focal plane infrared detector with a spectral response range of 8-14μm and a field of view covering the entire fireproof insulation layer surface of the fireproof flexible connector. It is configured to collect temperature field distribution data on the surface of the fireproof flexible connector, i.e., infrared thermal imaging data. The embedded processing unit uses a microcontroller with an ARM Cortex-M7 core, electrically connected to both the acoustic emission sensing module and the infrared thermal imaging module. It is configured to preprocess the collected acoustic emission signals and infrared thermal imaging data, including signal filtering, data frame encapsulation, and timestamp synchronization. The wireless communication module supports 5G-NR vehicle-to-ground communication and WiFi. The dual-mode communication unit of the 6 wireless local area network is electrically connected to the embedded processing unit and is configured to upload the pre-processed sensor data to the edge computing node and receive control commands issued by the adaptive control scheduler; the power supply module adopts a DC-DC isolated power supply unit powered by the on-board auxiliary power supply of the EMU to provide stable power supply for the above modules.

[0020] The edge computing nodes are deployed in the onboard electrical cabinets of each carriage of the EMU (Electric Multiple Unit). They utilize embedded industrial computers equipped with NPU (Neural Processing Unit) acceleration units and are connected to the sensor nodes of the embedded distributed sensor network within the same carriage via a wireless local area network. One edge computing node is configured for each carriage, and the edge computing nodes in each carriage are connected via the EMU's Ethernet backbone network. Each edge computing node incorporates a multi-physics coupled digital twin that integrates a material aging constitutive model. This material aging constitutive model is a nonlinear aging constitutive equation constructed based on the material properties of a fire-resistant silicone rubber fire-resistant insulating layer and a copper flexible conductor. Specifically, it includes an insulation thermal aging constitutive model, a mechanical fatigue constitutive model, and a partial discharge evolution constitutive model. The insulation thermal aging constitutive model uses a nonlinear aging equation modified from the Arrhenius model, expressed as:

[0021] In the formula, This refers to the remaining lifespan of the insulating material at a given temperature. The initial life constant of the material. The activation energy of insulating materials, Let be the ideal gas constant. The absolute temperature of the insulating material. The discharge damage coefficient is... This represents the cumulative discharge amount of partial discharge.

[0022] The mechanical fatigue constitutive model adopts a fatigue damage model modified by the Miner linear cumulative damage criterion, and its expression is:

[0023] In the formula, Mechanical fatigue damage degree, This represents the actual number of cycles under the stress amplitude of level n. The fatigue limit cycle number of the material under the first stress amplitude is obtained through the material's SN curve.

[0024] The multiphysics coupled digital twin is a multiphysics numerical model that couples temperature, stress, and partial discharge fields. Its governing equations include the heat conduction equation, the elasticity equilibrium equation, and the Poisson equation, covering all nonlinear terms of the material aging constitutive model. The edge computing node receives acoustic emission signals and infrared thermal imaging data uploaded from the embedded distributed sensor network, inputs the data into the multiphysics coupled digital twin, and outputs continuous predicted values ​​of insulation degradation gradient, fireproof layer failure probability, and mechanical fatigue damage degree along the thickness direction of the fireproof flexible connection through real-time parameter evolution and deduction of the digital twin. The unit of insulation degradation gradient is % / mm, and the fireproof layer failure probability and mechanical fatigue damage degree are both continuous values ​​in the range of 0-1.

[0025] The adaptive control scheduler, deployed in the train's central control unit, employs an industrial control computer equipped with a multi-core processor and FPGA acceleration unit. It communicates with edge computing nodes in each carriage via the train's Ethernet backbone and with sensor nodes in an embedded distributed sensor network via wireless communication links. The adaptive control scheduler is configured to aggregate model update parameters of similar soft connections across multiple carriages using an adaptive federated edge learning algorithm. This algorithm utilizes a federated averaging framework. Each carriage's edge computing node acts as a local client, updating its local model based on its own sensor data, generating model update parameters (gradient data), and uploading this gradient data to the adaptive control scheduler. The adaptive control scheduler, acting as a global server, receives the model update parameters uploaded from each carriage's edge computing node, performs global aggregation, generates updated global model parameters, and distributes these parameters to the edge computing nodes in each carriage, completing one round of federated learning iteration.

[0026] The adaptive control scheduler is simultaneously configured to receive continuous predicted values ​​of insulation degradation gradient, fireproof layer failure probability, and mechanical fatigue damage degree output by edge computing nodes. Based on the continuous predicted values, monitoring priorities are divided, and fireproof soft connections are classified into three priorities: high risk, medium risk, and low risk. For the sensor node units corresponding to high-priority fireproof soft connections, their sampling frequency is increased, their transmission power is increased, and their warning threshold is lowered. For the sensor node units corresponding to low-priority fireproof soft connections, their sampling frequency is reduced, their transmission power is decreased, and their warning threshold is increased. This achieves dynamic scheduling of monitoring resources and forms a closed-loop control link of data acquisition - status prediction - scheduling control - data acquisition.

[0027] As a preferred embodiment, please refer to Figure 5 When adjusting the state of the embedded distributed sensor network, the adaptive control scheduler adopts a dynamic time slot allocation method that integrates non-orthogonal multiple access power domain multiplexing and Markov decision process to map the EMU operating conditions into Markov state space. The joint reward function is to maximize the network lifetime and minimize the data acquisition delay. The transmit power and access time slot ratio of each sensor node are output through a deep deterministic policy gradient network.

[0028] The five-tuple of the Markov decision process is defined as follows: ,in: state space Each state Based on the current operating speed of the EMU Contact network voltage Carriage ambient temperature Remaining energy of each sensing node Current data queue length Composition, in which This is the number of the sensor node. The current moment; Action space Each action , which is the transmit power of each sensor node. Access time slot ratio ,in , ,and , This represents the minimum transmit power of the sensing node. This represents the maximum transmit power of the sensor node; State transition probability : , for the current state Next action Then, transition to the next state. The probability is determined by a state transition matrix constructed based on historical data of the EMU's operating conditions. Discount factor : This is used to weigh the immediate rewards against the future rewards; Joint reward function The optimization objective is to maximize network lifetime and minimize data acquisition latency. The expression is:

[0029] In the formula, and Let be the weighting coefficient, satisfying , For network lifecycle rewards, This is a reward for delayed data collection.

[0030] The expression for the network lifecycle reward item is:

[0031] In the formula, For the first The initial total energy of each sensing node, This represents the total number of sensor nodes. To perform the action The next The remaining energy of each sensing node.

[0032] The expression for the data acquisition delay reward is:

[0033] In the formula, For the first Maximum data queue length for each sensor node To perform the action The next The length of the data queue for each sensor node.

[0034] The deep deterministic policy gradient network includes an actor network and a critic network, wherein the actor network is configured to adjust according to the current state. Output deterministic actions , For the actor network, the network parameters are set; for the critic network, the configuration is based on the current state. With action Output action value , These are the network parameters for the critic network. Both the actor network and the critic network use fully connected neural networks, with input layer dimensions and state space... The dimensions are consistent; the output layer dimension of the actor network is consistent with the action space. The dimensions are consistent, and the output layer dimension of the critic network is 1.

[0035] When calculating the joint reward function, the deep deterministic policy gradient network introduces non-cooperative game theory to construct a power competition model among sensor nodes. The signal-to-interference-plus-noise ratio and remaining energy of each node are used as utility functions. By finding the Nash equilibrium point, the action space output of the deep deterministic policy gradient network is constrained, so that the sensor nodes spontaneously converge to the optimal power allocation strategy when competing for shared wireless channel resources.

[0036] In the non-cooperative game, each sensor node is considered a participant, and the strategy space of the game is the transmission power of each sensor node. utility function The expression is:

[0037] In the formula, To exclude the first The set of transmit powers of all other sensor nodes besides the one sensor node. The signal-to-interference-plus-noise ratio (SINR) weighting coefficient. Energy consumption weighting coefficient For the first The signal-to-interference-plus-noise ratio of each sensing node.

[0038] The expression for the signal-to-interference-plus-noise ratio is:

[0039] In the formula, For the first Channel gain between sensor nodes and edge computing nodes The Gaussian white noise power of the channel.

[0040] The Nash equilibrium point satisfies: for any sensing node Any selectable transmit power All have ,in The first Nash equilibrium point The transmit power of each sensor node is controlled. The transmit power range corresponding to the Nash equilibrium point is used as the constraint boundary of the action space of the deep deterministic policy gradient network, ensuring that the transmit power output by the actor network is within the Nash equilibrium range. Simultaneously, non-orthogonal multiple access power domain multiplexing technology is employed to allocate different transmit power levels to multiple sensor nodes within the same time slot. At the receiving end, serial interference cancellation technology is used to separate the signals of different sensor nodes, enabling simultaneous access by multiple nodes.

[0041] As a preferred embodiment, please refer to Figure 3 In the process of constructing the multiphysics coupled digital twin, the flow field order reduction projection method based on intrinsic orthogonal decomposition is used to process the partial differential equations in the material aging constitutive model, mapping the high-dimensional temperature field and stress field to the low-dimensional subspace, and combining the simulated annealing algorithm to optimize the boundary conditions in the low-dimensional subspace, so as to realize the real-time parameter evolution and deduction of the digital twin on the edge computing node.

[0042] The governing equations of the multiphysics coupled digital twin are a set of partial differential equations involving the coupled temperature field, stress field, and partial discharge field, in the general form of:

[0043] In the formula, It is a high-dimensional physical field variable vector, including temperature field variables. Stress field variables Partial discharge field variables , For spatial coordinates, Using time as the coordinate, For differential operators, The source term vector contains the nonlinear terms of the material aging constitutive model.

[0044] The specific implementation process of the flow field order reduction projection method based on intrinsic orthogonal decomposition is as follows: A multiphysics sample dataset was collected, and snapshot matrices of the temperature field, stress field, and partial discharge field under different boundary conditions and aging stages were obtained through finite element simulation. ,in For the number of snapshots, For the first The physical field variable vector corresponding to each snapshot For solving physical fields; Calculate the covariance matrix of the snapshot matrix ; For covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and the corresponding feature vectors The eigenvectors are the intrinsic orthogonal decomposition basis functions; Before selection The eigenorthogonal decomposition basis functions corresponding to the largest eigenvalues ​​are used to construct a reduced-order basis matrix. ,in The selection satisfies the cumulative energy contribution rate. ; high-dimensional physical field variable vectors Projecting onto a lower-dimensional subspace yields a lower-dimensional vector of state variables. The projection relationship is as follows:

[0045] Substituting the above projection relations into the original partial differential equations, we obtain the reduced-order governing equations in the low-dimensional subspace through Galerkin projection:

[0046] Original The high-dimensional partial differential equation system is reduced to the following order: A system of ordinary differential equations in dimension.

[0047] The specific implementation process of the simulated annealing algorithm for boundary condition optimization in low-dimensional subspace is as follows: The boundary conditions of the reduced-order control equations are used as optimization variables, and the goal is to minimize the mean square error between the derivation results of the digital twin and the measured sensing data. The objective function is:

[0048] In the formula, For boundary condition vectors, Based on boundary conditions The predicted physical field values ​​obtained through the derivation of the reduced-order model These are the measured values ​​of the physical field obtained by reconstructing from infrared thermal imaging data and acoustic emission signals.

[0049] The optimization process of the simulated annealing algorithm is as follows: Initial annealing temperature Initial boundary conditions Current optimal solution The current optimal objective function value ; Under the current boundary conditions Generate new boundary conditions within the neighborhood. Calculate the difference of the objective function ; like Then accept the new boundary conditions. ;like Then, based on probability Accepting new boundary conditions ,in This is the current annealing temperature; Decrease the annealing temperature according to the annealing rate. ,in This is the annealing coefficient; Repeat the above iterative process until the annealing temperature drops to the termination temperature. Alternatively, if the objective function value converges to a preset threshold, the optimal boundary condition is output. .

[0050] After optimizing the boundary conditions in the low-dimensional subspace, the ensemble Kalman filter method in meteorological data assimilation is introduced to correct the error in the real-time parameter evolution of the digital twin. The boundary conditions output by the simulated annealing algorithm are used as the background field, and the real-time acquired acoustic emission signals and infrared thermal imaging data are used as the observation field. Covariance expansion and localization processing are used to suppress the filtering divergence in the nonlinear constitutive model derivation process.

[0051] The specific implementation process of the ensemble Kalman filtering method is as follows: Construct the state vector of the ensemble Kalman filter, the state vector For the low-dimensional state variables of the reduced-order model It also includes key parameters of the material aging constitutive model; Initialize the state set and generate Initial state samples ,in The initial state sample is generated based on the prior distribution of the initial boundary conditions and material parameters, to determine the number of members in the set. Prediction steps: Input each state sample into the reduced-order multiphysics coupled digital twin, complete one-step time extrapolation, and obtain the predicted state set. And calculate the mean of the forecast state. Covariance matrix with background field ; Update steps: Using the optimal boundary conditions output by the simulated annealing algorithm as the background field, and the real-time acquired acoustic emission signals and infrared thermal imaging data as the observation field, an observation vector is constructed. Observation operator Configured as a linear transformation matrix to map the state vector to the observation space, resulting in an observation-prediction set. ; Calculate the Kalman gain matrix ,in To observe the noise covariance matrix; The analysis state set is obtained by updating each forecast state sample. The updated formula is:

[0052] In the formula, The observed noise samples follow a mean of 0 and a covariance matrix of... Gaussian distribution; For the background field covariance matrix Covariance dilation and localization: Covariance dilation is achieved by multiplying by a dilation coefficient. Localization processing attenuates the covariance between state variables whose distance exceeds a preset threshold by introducing a local correlation function. The state variables and material parameters of the digital twin are updated based on the analysis of the state set to complete the error correction.

[0053] As a preferred embodiment, please refer to Figure 2 When processing acoustic emission signals and infrared thermal imaging data, a dual-modal feature alignment method that integrates micro-Doppler feature extraction and visual Transformer mechanism is adopted. The acoustic emission signal is converted into a time-frequency map and the instantaneous frequency trajectory of micro-Doppler is extracted as a query sequence. The patches of infrared thermal imaging data are embedded as key-value pairs. The adaptive fusion of acoustic and thermal heterogeneous data in the latent feature space is achieved through cross-attention mechanism.

[0054] The specific implementation process of the dual-modal feature alignment method is as follows: Preprocessing and micro-Doppler feature extraction of acoustic emission signals: Acquired acoustic emission time-domain signals Perform continuous wavelet transform to obtain the time-frequency diagram. The expression for the continuous wavelet transform is:

[0055] In the formula, As a scaling factor, and frequency Inversely proportional, For Morlet mother wavelet function, It is the complex conjugate of the mother wavelet function.

[0056] Based on time-frequency graph The instantaneous frequency trajectory of the micro-Doppler signal is extracted. The effective signal region in the time-frequency graph is extracted using adaptive threshold segmentation. A peak detection algorithm is then used to extract the instantaneous frequency peak at each moment, and the resulting segments are connected to obtain the micro-Doppler instantaneous frequency trajectory. The micro-Doppler instantaneous frequency trajectory is divided into blocks according to a fixed time window, and the feature vector corresponding to each block is used as the query sequence. ,in For sequence length, For feature dimensions.

[0057] Preprocessing and tile embedding of infrared thermal imaging data: The collected infrared thermal imaging data Normalization is performed using the following formula:

[0058] In the formula, The minimum pixel value of the infrared thermal imaging data. This represents the maximum pixel value of the infrared thermal imaging data.

[0059] The normalized infrared thermal imaging data is divided into patches of varying sizes. The non-overlapping tiles are flattened into one-dimensional vectors and embedded through a linear projection layer to obtain tile embedding vectors. Position encoding is added to each tile embedding vector to finally obtain a sequence of key-value pairs. , ,in The number of tiles, As a feature dimension, and the query sequence The feature dimensions are consistent.

[0060] Feature adaptive fusion is achieved through a cross-attention mechanism. The calculation expression for cross-attention is as follows:

[0061] In the formula, The scaling factor is used to output the fused multimodal feature vector. .

[0062] After achieving adaptive fusion of acoustic and thermal heterogeneous data in the latent feature space, word vector mapping and graph neural network methods from natural language processing are introduced to construct a fault feature semantic graph. The fused feature vectors are regarded as word sequences. The contextual co-occurrence relationship between features is learned through the Skip-gram model. A dynamic topology graph is constructed with feature vectors as nodes and co-occurrence similarity as edge weights. Neighbor node information is aggregated through graph convolutional networks to quantify the degree of correlation of partial discharge.

[0063] The construction and processing flow of the fault feature semantic map is as follows: The fused multimodal feature vector Treating the words as a sequence, with each feature vector corresponding to a word, the Skip-gram model is used to learn the contextual co-occurrence relationships between features. The optimization objective of the Skip-gram model is to maximize the probability of occurrence of context words given a center word. The log-likelihood function is:

[0064] In the formula, For sequence length, For the context window size, With the core vocabulary, For contextual vocabulary, Given the central vocabulary Contextual vocabulary The probability of occurrence is calculated using the softmax function.

[0065] The word embedding vector corresponding to each feature vector is obtained by training the Skip-gram model. A dynamic topology graph is constructed using the word embedding vector corresponding to each feature vector as a node and the co-occurrence similarity between features as the edge weight. ,in For a set of nodes, Let be a set of edges, nodes With nodes Border weight between The cosine similarity between two feature vectors is expressed as:

[0066] In the formula, For nodes The corresponding word embedding vector, For nodes The corresponding word embedding vector.

[0067] A graph convolutional network is used to perform feature aggregation on the above dynamic topology graph. The inter-layer propagation formula of the graph convolutional network is as follows:

[0068] In the formula, For the first The node feature matrix of the layer, For the initial word embedding vector matrix, The normalized adjacency matrix, For the first The trainable weight matrix of the layer, It is the ReLU nonlinear activation function.

[0069] By aggregating neighbor node information through a graph convolutional network, an aggregated feature vector for each node is obtained. The cosine similarity between the aggregated feature vector and the preset partial discharge fault feature template is calculated, and the degree of correlation of partial discharge is quantified based on the similarity value.

[0070] As a preferred embodiment, please refer to Figure 4 The adaptive federated edge learning algorithm introduces a clonal selection mechanism based on an artificial immune system to screen local models when aggregating model update parameters. It treats global model parameters as antigens and local model parameters uploaded by edge computing nodes in each carriage as antibodies. It eliminates inferior local parameters containing abnormal operating condition noise by calculating affinity and concentration suppression strategies, and only aggregates the model gradients corresponding to high affinity antibodies.

[0071] The specific implementation process of the clone selection mechanism is as follows: Treat global model parameters as antigens The local model parameters uploaded by the edge computing nodes of each carriage are regarded as antibodies. ,in The node number is used to calculate the edge of the carriage. , This represents the total number of carriages. Calculate each antibody With antigen Affinity between Affinity is characterized by a negative mean squared error, expressed as:

[0072] Calculate each antibody concentration Concentration is used to characterize antibody similarity, and the expression is:

[0073] In the formula, Antibody With antibodies The similarity function between them, when hour, ,otherwise , This is the similarity threshold.

[0074] Antibody selection probability was calculated using a concentration inhibition strategy. The expression is:

[0075] In the formula, and Let be the weighting coefficient, satisfying , .

[0076] Based on selection probability The antibodies are sorted, and high-affinity antibodies with a selection probability higher than a preset threshold are selected. Inferior antibodies with a selection probability lower than the preset threshold are removed. Only the local model gradients corresponding to high-affinity antibodies are retained for subsequent global aggregation.

[0077] When aggregating the model gradients corresponding to high-affinity antibodies, the Shapley value mechanism design theory from microeconomics is introduced to evaluate the marginal contribution of each carriage edge computing node to the global model accuracy. The differences in the operating conditions of the EMU are regarded as a coalition game environment. By calculating the average marginal revenue of local model parameters in all possible combinations of carriage nodes, the weight coefficients of each node gradient in the global aggregation formula are adaptively allocated.

[0078] The specific implementation process for the Shapley value weighting allocation is as follows: Treat all edge computing nodes of the train carriages as a set of participants in a coalition game. ,in Given the total number of carriages, any subset For a consortium, the model accuracy corresponding to different consortia is equal to the consortium's profit. , For only adopting alliance The accuracy of the global model on the test dataset is obtained by aggregating the local model gradients of the nodes within the model. Calculate the first Shapley value of each carriage edge node The expression is:

[0079] In the formula, For the alliance The number of nodes in For the first One node joined the alliance After that, the marginal increase in alliance revenue.

[0080] The weighting coefficients of each node's gradient in the global aggregation formula are calculated based on the Shapley value. The normalization formula is:

[0081] In the formula, ,and .

[0082] The local model gradients corresponding to high-affinity antibodies are weighted and aggregated based on weight coefficients. The global model gradient update formula is as follows:

[0083] In the formula, For the aggregated gradient of the global model, For the first The local model gradient uploaded by the edge computing node of each carriage.

[0084] Please refer to Figure 6 When adaptively allocating the weight coefficients of each node gradient in the global aggregation formula, active disturbance rejection control (ADRC) technology from control theory is introduced to respond to the dynamic fluctuations of the weight coefficients. The expected weight value calculated by Shapley value is used as the tracking reference signal. The weight mutation caused by electromagnetic interference of the EMU is regarded as the total disturbance. The disturbance is estimated and compensated in real time using an extended state observer. The trajectory of weight coefficient change is smoothed by a nonlinear state error feedback law to maintain the stability of adaptive federated edge learning.

[0085] The specific implementation process of the active disturbance rejection control technology is as follows: Construct an active disturbance rejection controller, which comprises three parts: a tracking differentiator, an extended state observer, and a nonlinear state error feedback law. The expected weight value is calculated using the Shapley value. As a tracking reference signal, the weight abrupt changes caused by electromagnetic interference of the EMU and the weight fluctuations caused by changes in operating conditions are regarded as the total disturbance. Tracking differentiator for reference signal Perform tracking, arrange the transition process, and output the tracking signal. With differential signal Its discrete form is:

[0086]

[0087] In the formula, The sampling step size, For velocity factor, For the filter factor, This is the fastest control synthesis function.

[0088] The extended state observer estimates the total disturbance of the system in real time, and applies the actual weighting coefficients. As the system output, its discrete form is:

[0089]

[0090]

[0091] In the formula, For tracking signals of weighting coefficients, This is the estimated total disturbance of the system in the extended state. , For observer gain, To control the gain, To control the quantity.

[0092] The nonlinear state error feedback law outputs a control quantity based on the tracking error and disturbance estimate, smoothing the trajectory of the weighting coefficient changes. Its expression is:

[0093]

[0094]

[0095] In the formula, For proportional gain, It is a non-linear power function. It is a nonlinear factor. is the length of the linear segment interval.

[0096] The control quantity output by the active disturbance rejection controller As a correction to the weighting coefficients, the weighting coefficients are... Real-time corrections are performed to obtain smoothed weight coefficients. .

[0097] As a preferred embodiment, the complete workflow of the fireproof flexible connection status monitoring system for high-speed trains based on embedded sensor networks is as follows: During the system initialization phase, the adaptive control scheduler initializes the global multiphysics coupled digital twin model parameters and distributes them to the edge computing nodes of each carriage. The edge computing nodes of each carriage construct a local multiphysics coupled digital twin based on the global model parameters and complete the intrinsic orthogonal decomposition and order reduction processing. Each sensor node unit of the embedded distributed sensor network completes initialization and completes time synchronization according to the default sampling frequency, transmission power and time slot ratio.

[0098] During the data acquisition phase, each sensor node unit of the embedded distributed sensor network collects the acoustic emission signal and infrared thermal imaging data of the corresponding fireproof soft connection. After preprocessing and timestamp synchronization, the data is uploaded to the edge computing node of the corresponding carriage.

[0099] In the dual-modal feature fusion stage, the edge computing nodes adopt a dual-modal feature alignment method that combines micro-Doppler feature extraction and visual Transformer mechanism to extract features and adaptively fuse acoustic emission signals and infrared thermal imaging data, thereby obtaining the fused multimodal feature vector.

[0100] During the state prediction phase, the edge computing nodes input the fused multimodal feature vectors into the built-in multiphysics coupled digital twin. The real-time parameter evolution is deduced through a reduced-order model, and error correction is completed by combining ensemble Kalman filtering. The corresponding insulation degradation gradient of the fireproof soft connection, the failure probability of the fireproof layer, and the continuous predicted values ​​of mechanical fatigue damage are output and uploaded to the adaptive control scheduler.

[0101] During the local model training phase, the edge computing nodes of each carriage complete the iterative update of the local model based on the sensor data and state prediction results collected in the carriage, generate local model update parameters, i.e. gradient data, and upload them to the adaptive control scheduler.

[0102] In the local model screening and global aggregation stage, the adaptive control scheduler uses a clonal selection mechanism based on an artificial immune system to screen the uploaded local model parameters and remove inferior local parameters; it calculates the weight coefficients of each node based on the Shapley value mechanism and completes the smooth correction of the weight coefficients through active disturbance rejection control; it performs weighted aggregation of the gradients of high-affinity local models based on the corrected weight coefficients, updates the global model parameters, and distributes the updated global model parameters to the edge computing nodes of each carriage.

[0103] In the adaptive closed-loop control phase, the adaptive control scheduler, based on the received state prediction values, divides the monitoring priorities of each fire protection soft connection, adopts a dynamic time slot allocation method that integrates non-orthogonal multiple access and Markov decision process, outputs the transmit power and access time slot ratio of each sensor node through a deep deterministic policy gradient network, dynamically adjusts the sampling frequency, monitoring priority and early warning threshold of each sensor node, and sends adjustment instructions to each sensor node unit of the embedded distributed sensor network to complete one round of closed-loop control.

[0104] The system cyclically executes the entire process from data acquisition to closed-loop control, realizing real-time monitoring, early warning, and adaptive closed-loop control of the fire-resistant soft connection status of the EMU.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fireproof flexible connection status monitoring system for high-speed trains based on embedded sensor networks, characterized in that, include: An embedded distributed sensor network is used to collect acoustic emission signals and infrared thermal imaging data from the fireproof flexible connection of the EMU. Edge computing nodes are connected to embedded distributed sensor networks and have built-in multi-physics field coupled digital twins that integrate material aging constitutive models. They receive acoustic emission signals and infrared thermal imaging data to generate continuous prediction values ​​of insulation degradation gradient, fireproof layer failure probability and mechanical fatigue damage. The adaptive control scheduler communicates with the embedded distributed sensor network and the edge computing node respectively. It uses an adaptive federated edge learning algorithm to aggregate the model update parameters of the same type of soft connection in multiple carriages, and dynamically adjusts the monitoring priority and early warning threshold of the embedded distributed sensor network based on continuous prediction values ​​to form a closed-loop control.

2. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 1, characterized in that, When adjusting the state of the embedded distributed sensor network, the adaptive control scheduler adopts a dynamic time slot allocation method that integrates non-orthogonal multiple access power domain multiplexing and Markov decision process. It maps the EMU operating conditions into a Markov state space and uses the maximization of network lifetime and minimization of data acquisition latency as the joint reward function. The deep deterministic policy gradient network outputs the transmit power and access time slot ratio of each sensor node.

3. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 1, characterized in that, In the process of constructing a multiphysics coupled digital twin, a flow field order reduction projection method based on intrinsic orthogonal decomposition is used to process the partial differential equations in the material aging constitutive model, mapping the high-dimensional temperature field and stress field to a low-dimensional subspace, and combining the simulated annealing algorithm to optimize the boundary conditions in the low-dimensional subspace.

4. The fireproof flexible connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 1, characterized in that, When processing acoustic emission signals and infrared thermal imaging data, a dual-modal feature alignment method that integrates micro-Doppler feature extraction and visual Transformer mechanism is adopted. The acoustic emission signal is converted into a time-frequency map and the instantaneous frequency trajectory of micro-Doppler is extracted as a query sequence. The patches of infrared thermal imaging data are embedded as key-value pairs. The adaptive fusion of acoustic and thermal heterogeneous data in the latent feature space is achieved through a cross-attention mechanism.

5. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 1, characterized in that, When aggregating model update parameters, the adaptive federated edge learning algorithm introduces a clonal selection mechanism based on an artificial immune system to screen local models. It treats global model parameters as antigens and local model parameters uploaded by edge computing nodes in each carriage as antibodies. By calculating affinity and concentration suppression strategies, it eliminates inferior local parameters containing abnormal operating condition noise and only aggregates the model gradients corresponding to high-affinity antibodies.

6. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 2, characterized in that, When calculating the joint reward function, the deep deterministic policy gradient network introduces non-cooperative game theory to construct a power competition model among sensor nodes. The signal-to-interference-plus-noise ratio and remaining energy of each node are used as utility functions. By finding the Nash equilibrium point, the action space output of the deep deterministic policy gradient network is constrained, so that the sensor nodes spontaneously converge to the optimal power allocation strategy when competing for shared wireless channel resources.

7. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 3, characterized in that, After optimizing the boundary conditions in the low-dimensional subspace, the ensemble Kalman filter method in meteorological data assimilation is introduced to correct the error in the real-time parameter evolution of the digital twin. The boundary conditions output by the simulated annealing algorithm are used as the background field, and the real-time acquired acoustic emission signals and infrared thermal imaging data are used as the observation field. Covariance expansion and localization processing are used to suppress the filtering divergence in the nonlinear constitutive model derivation process.

8. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 4, characterized in that, After achieving adaptive fusion of acoustic and thermal heterogeneous data in the latent feature space, word vector mapping and graph neural network methods from natural language processing are introduced to construct a fault feature semantic graph. The fused feature vectors are regarded as word sequences. The contextual co-occurrence relationship between features is learned through the Skip-gram model. A dynamic topology graph is constructed with feature vectors as nodes and co-occurrence similarity as edge weights. Neighbor node information is aggregated through graph convolutional networks to quantify the degree of correlation of partial discharge.

9. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 5, characterized in that, When aggregating the model gradients corresponding to high-affinity antibodies, the Shapley value mechanism design theory from microeconomics is introduced to evaluate the marginal contribution of each carriage edge computing node to the global model accuracy. The differences in the operating conditions of the EMU are regarded as a coalition game environment. By calculating the average marginal revenue of local model parameters in all possible combinations of carriage nodes, the weight coefficients of each node gradient in the global aggregation formula are adaptively allocated.

10. The fireproof soft connection status monitoring system for high-speed trains based on embedded sensor networks according to claim 9, characterized in that, When adaptively allocating the weight coefficients of each node gradient in the global aggregation formula, active disturbance rejection control (ADRC) technology from cybernetics is introduced to respond to the dynamic fluctuations of the weight coefficients. The expected weight value calculated by Shapley value is used as the tracking reference signal, and the weight mutation caused by the electromagnetic interference of the EMU is regarded as the total disturbance. The disturbance is estimated and compensated in real time using an extended state observer, and the trajectory of weight coefficient change is smoothed by a nonlinear state error feedback law to maintain the stability of adaptive federated edge learning.

Images