A cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning

Abstract

The present invention relates to the technical field of high-speed train fault diagnosis, and discloses a cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning. First, multi-channel sensor monitoring data is obtained from the high-speed train monitoring system; a D-TernP network model with a multi-level export structure is constructed, including a backbone feature extraction network and multiple export sub-networks; the multi-channel sensor monitoring data is input into the D-TernP network model to obtain the prediction results of each export sub-network; the multi-channel sensor monitoring data to be detected is input into the trained D-TernP network model for reasoning and judgment, and each export sub-network calculates the entropy value of the prediction result and determines the confidence level. If the confidence level is sufficient, the prediction result is directly output, otherwise it is uploaded to the cloud model for more in-depth analysis. The present invention can reduce the delay between the on-board system and the cloud, and at the same time complete some tasks on the on-board system, which can reduce the amount of computing on the cloud while reducing bandwidth consumption.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed train fault diagnosis, and specifically to a cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning. Background Art

[0002] High-speed trains have become a vital mode of intercity transportation in modern society, where passenger safety and system reliability are paramount. During high-speed operation, various subsystems, such as the bogie system, traction system, and wheelset system, may malfunction due to various factors, including component aging, humidity, temperature, and environmental factors. Therefore, achieving real-time monitoring and intelligent fault diagnosis of key train components has become a key research topic in the rail transit field.

[0003] With the continuous development of onboard detection and control systems, modern high-speed trains typically include hundreds or even thousands of sensors to monitor the status of various components of the running train, including vibrating wheelsets, bearings, and bogies. The suspension system, due to its direct impact on vibration isolation and attitude control, is a key monitoring target. However, due to the multi-dynamic coupling complexity of its structure, its fault characteristics are affected by multiple factors such as speed and track conditions, making diagnosis more difficult. While the large amount of monitoring data provides a rich information foundation for fault diagnosis, it also presents challenges in data analysis and processing. To identify fault relationships under complex dynamic conditions using this real-time monitoring data, artificial intelligence methods, particularly deep learning, have been introduced into train fault diagnosis systems. While these methods have achieved significant improvements in accuracy and feasibility compared to traditional detection methods, the increasing computational complexity and model size pose new challenges to the computational and storage resources of onboard detection systems.

[0004] Due to the computational limitations of onboard systems, traditional approaches often upload monitored data to the cloud for processing and return the results to the onboard system. This is because cloud computing provides high-performance computing power, making it easier to handle computationally intensive tasks and more efficient in processing large amounts of monitoring data. However, trains, as high-speed mobile units, face bandwidth constraints, and the ever-increasing speeds of trains pose even greater challenges to the design of communication systems. This requires efficient communication between the onboard system and the cloud. Directly uploading all monitoring data to the cloud results in significant bandwidth consumption and latency. High-speed operation reduces available bandwidth, making it impossible to support the real-time upload of all sensor data. Furthermore, the communication system must reserve bandwidth for the transmission of critical commands, such as train control, further compressing the space available for uploading monitoring data. Furthermore, wireless communication between trains and external systems is susceptible to environmental fluctuations (e.g., in tunnels). When network connectivity is interrupted, fault diagnosis systems that rely solely on cloud computing will not function properly. Therefore, effectively addressing the latency-sensitive train fault diagnosis problem in the collaboration between onboard systems and the cloud remains a significant research challenge. Summary of the Invention

[0005] To address the above issues, the present invention aims to provide a cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning. By optimizing resource allocation in cloud-edge computing, this method can reduce latency between the onboard system and the cloud. Furthermore, by completing some tasks on the onboard system, it can reduce both the computational load on the cloud and bandwidth consumption. The technical solution is as follows:

[0006] Step 1: Obtain multi-channel sensor monitoring data from the high-speed train monitoring system;

[0007] Step 2: Build a D-TernP (Dynamic TernP) network model with a multi-level export structure, including a backbone feature extraction network and multiple export sub-networks. The backbone feature extraction network is used to extract deep semantic information from the input signal and is composed of a cascade of lightweight residual TernP modules adapted for edge device operation. Each export sub-network is connected to different depths of the backbone feature extraction network to implement branch prediction.

[0008] Step 3: Train and fit the D-TernP network model based on historical monitoring data samples:

[0009] Input the multi-channel sensor monitoring data into the D-TernP network model to obtain the prediction results of each export sub-network;

[0010] Calculate the cross entropy loss of each export sub-network separately, dynamically set the weight coefficient of each export sub-network loss, and construct a weighted total loss function;

[0011] Count the sample exit ratios of each exit sub-network, calculate the corresponding reward items, combine the weighted total loss function with the reward items to obtain the final total loss, and minimize the final total loss to optimize the network parameters;

[0012] Step 4: Input the multi-channel sensor monitoring data to be detected into the trained D-TernP network model for inference and judgment:

[0013] Each exit sub-network calculates the entropy value of the prediction result in the order of priority from the edge to the center. If the entropy value is lower than the corresponding threshold, it means that the D-TernP network model meets the confidence condition for the exit sub-network result, and the prediction result is directly output without subsequent forward propagation;

[0014] If all edge exit sub-networks fail to meet the confidence condition, the multi-channel sensor monitoring data to be tested will be uploaded to the cloud model for more in-depth analysis, and the final prediction results will be output.

[0015] The beneficial effects of the present invention are:

[0016] 1) This paper proposes a novel dynamic reasoning mechanism to achieve efficient task allocation. Most fault samples can be successfully processed at the edge device, thereby significantly reducing communication overhead while maintaining high diagnostic accuracy.

[0017] 2) This paper designs the TernP module, a lightweight structure that combines partial convolution and position normalization techniques. It effectively reduces model complexity while improving recognition accuracy, and has obvious advantages over traditional methods.

[0018] 3) This paper introduces an adaptive training strategy that combines dynamic weight adjustment with a reward mechanism to ensure stable model convergence during training and achieve optimal sample allocation between different branches. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the structural diagram of the TernP module.

[0020] Figure 2 This is the overall framework diagram of the cloud-edge collaborative fault diagnosis system. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] This invention provides a cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning. This method determines the execution steps based on the recognition difficulty of each sample, enabling rapid local processing on edge devices and uploading uncertain samples to the cloud for processing, significantly reducing bandwidth consumption and communication latency. First, multi-channel sensor signals are input into a D-TernP network model, which consists of three efficient processing modules. This network extracts multi-level features of the signals, reducing computational complexity to meet the real-time processing requirements of edge devices. Then, based on sample complexity, a dynamic reasoning mechanism is employed, implementing dynamic exits through a multi-exit structure. This ensures rapid response to easily identifiable samples on the edge device, while complex samples are uploaded to the cloud for diagnosis by a more complex model. This mechanism dynamically adjusts the training weights and reward strategies for different exits, promoting the adaptive optimization of the model's exit strategy during training and improving overall reasoning efficiency and accuracy. Finally, the system, combined with a cloud-edge collaborative architecture design, reduces data transmission bandwidth while ensuring the real-time and accuracy of fault diagnosis, making it suitable for large-scale high-speed train fault detection scenarios.

[0023] Specifically, the following steps are included:

[0024] Step 1: Obtain multi-channel sensor monitoring data from the high-speed train monitoring system.

[0025] Step 2: Build a D-TernP network model with a multi-level export structure, including: a backbone feature extraction network: used to extract deep semantic information of the input signal, composed of a cascade of lightweight residual structure TernP modules, adapted for edge device operation; multiple export sub-networks: located at different depths of the backbone network, each export sub-network can output a set of prediction results for staged reasoning.

[0026] The present invention adopts a combination of partial convolution and position normalization to form a TernP module, which can be described as follows:

[0027] For an input data tensor , B is the number of data samples for a single input, S is the total number of channels, L is the signal sample length, and partial convolution is used in the front The convolution operation is performed on the channel, and the subsequent channels remain unchanged to obtain the intermediate features. :

[0028] ;

[0029] Then the features are normalized in position, that is, the results of each channel at each time step are normalized in position. The formula is as follows:

[0030] ;

[0031] ;

[0032] in, and Represent the mean and standard deviation of all channels of the b-th sample and the l-th position respectively, is the component at the bth sample, sth channel, and lth position in the input data tensor X.

[0033] Then, the complete operation structure of the TernP module can be expressed as follows:

[0034] ;

[0035] in, is the weight matrix used to learn the feature combination between channels, is the bias term. and Respectively represent the mean and standard deviation calculated for all S channels at each position l, and is a learnable parameter.

[0036] Step 3: Train the D-TernP network model.

[0037] Multi-channel sensor monitoring data is input into the D-TernP network model to obtain the prediction results of each outlet sub-network. For each outlet sub-network, the cross entropy loss function is used to evaluate the difference between the predicted label and the true label. The formula is as follows:

[0038] ;

[0039] Where C represents the total number of categories, Represents the network parameters from the input to the output sub-network, y c is the encoding of the true label, is the predicted probability of the current export sub-network for class c calculated by the Softmax function. The formula is as follows:

[0040] ;

[0041] in, Indicates that network parameters are taken into account , the input sample x passes through the output of the nth exit sub-network, this formula is for the output Softmax normalization is performed to finally normalize the output into a probability distribution. The weighted total loss function formula is as follows:

[0042] ;

[0043] in, represents the weight of the nth export sub-network in the total loss, Represents the output prediction result. By controlling the weight coefficients of different export sub-networks, the tendency of controlling cloud-edge allocation during the detection process can be adjusted.

[0044] In order to achieve automated parameter design, this solution adopts a loss function that combines dynamic weighting and exit reward mechanism. The formula is as follows:

[0045] ;

[0046] ;

[0047] Where N represents the total number of export sub-networks, represents the cross entropy loss of the n-th export sub-network, Represents the corresponding dynamic weight coefficient, which can be adaptively adjusted during the training process to make the model focus on the branch with larger errors. t and T represent the current training step and the total training step, respectively. The second term of the loss function formula is the reward term, which encourages the model to make predictions as early as possible in the shallow exit sub-network. n represents the proportion of samples exiting the nth exit sub-network, is the reward adjustment coefficient of each exit sub-network, which can control the degree of attention the model pays to the classification accuracy on different exit sub-networks. is the adjustment weight of the overall reward term. Based on this formula, this design can automatically optimize the learning objectives and sample distribution of different export sub-networks, thereby improving the performance of the entire model in multi-layer reasoning.

[0048] Step 4: Input the multi-channel sensor data to be tested into the trained D-TernP network model for inference judgment. Its inference mechanism is based on a dynamic exit strategy, that is, through the confidence of the exit sub-network, it is determined whether the sample has been pre-classified in the exit sub-network. In order to evaluate the confidence, the present invention uses entropy as an indicator:

[0049] ;

[0050] The entropy value of the prediction result is calculated for each export sub-network. If the entropy value is lower than the corresponding threshold, it means that the model has sufficient confidence in the result of the export sub-network, and the prediction result is directly output; if the entropy value is higher than the threshold, it means that there is cognitive uncertainty in the model, and all edge export sub-networks do not meet the confidence condition, then the data is uploaded to the cloud model for more in-depth analysis, and the final prediction result is output.

[0051] Figure 1The structural diagram of the TernP module is shown. This module integrates three key technologies: partial convolution, position normalization (PONO), and point-wise convolution. In each TernP module, the input signal first undergoes a partial convolution operation, where convolution calculations are performed only in key regions of the sequence, thereby significantly reducing redundant calculations while ensuring feature extraction capabilities. Subsequently, position normalization is used to normalize each time step, effectively alleviating the uneven distribution of multi-channel signals in position-sensitive tasks and enhancing the model's ability to capture key timing features. After several stacking operations, point-wise convolution is used to compress and fuse information between channels, further improving the compactness and discriminability of feature extraction. The entire TernP module can be deployed independently on edge devices to complete lightweight diagnostic tasks, or it can serve as a basic component of deep cloud networks to facilitate the goal of end-cloud collaborative reasoning.

[0052] Figure 2 The overall framework of the cloud-edge collaborative fault diagnosis system is demonstrated, which has three important components: the device edge model, the vehicle gateway edge model, and the cloud model.

[0053] The device edge sub-model is used in each high-speed train carriage to directly process the multi-channel sensor data collected by that carriage. This model utilizes a lightweight TernP modular structure, enabling preliminary diagnosis of most fault samples within resource-constrained conditions. For samples with high-confidence predictions, the system completes the judgment at this layer and directly outputs the results.

[0054] When the output results of the device edge sub-model do not meet the output requirements, the relevant feature information will be transmitted to the vehicle gateway sub-model in the train system. The model will fuse and further analyze the feature information from different carriages to achieve secondary diagnosis of more complex fault modes.

[0055] If the above two-level models still cannot make a reliable judgment, the system will transmit the sample to the cloud sub-model of the remote monitoring center for processing.

[0056] To further improve communication efficiency, the framework adopts a feature-level transmission strategy between the edge and cloud, uploading abstracted feature representations only when necessary. This avoids large-scale transmission of raw data and significantly reduces bandwidth consumption. Furthermore, the system introduces a dynamic exit mechanism that dynamically determines whether to send samples to higher-level processing based on the predicted entropy output of the model, thereby enabling adaptive resource allocation and control of inference depth.

[0057] All experiments in this example were conducted on a workstation equipped with an Intel(R) Xeon(R) W-2245 3.90GHz CPU and an NVIDIA RTX 4090 GPU. Model training was accelerated using the GPU (Graphics Processing Unit), while inference time measurement and computational performance evaluation were all performed on the CPU (Central Processing Unit). All methods were limited to using two threads to ensure a fair comparison and simulate the application environment of actual edge devices, which often lack GPU acceleration and have limited computing resources.

[0058] This example uses multibody dynamics simulation software to construct a dynamic model of the coupled bogie frame and carbody of a high-speed train to validate and evaluate the effectiveness of the proposed method. The model comprehensively considers the structural parameters of real vehicles and measured track spectra as external excitations to simulate various operating conditions. The research focuses on a frame monitoring system consisting of eight triaxial accelerometers (four sensors installed on each bogie frame). Each sensor collects vibration acceleration data in the lateral, longitudinal, and vertical directions, generating a 24-channel high-frequency monitoring signal with a sampling frequency of 243 Hz. The simulated operating conditions include the dynamic behavior of the train under normal operating conditions, as well as abnormal conditions covering 30 typical fault scenarios, including: one air spring failure per bogie, two lateral damper failures per bogie, four anti-snaking damper failures per bogie, four vertical damper failures per bogie, and four coil spring failures per bogie. All operating conditions simulate train operation at 200 km / h. The dataset is divided into training, validation, and test sets.

[0059] The dynamic exit mechanism allows for manual setting of the entropy threshold for each egress subnetwork based on specific application requirements. In this experiment, the normalized entropy thresholds for the device edge egress subnetwork and the vehicle gateway egress subnetwork were set to 0.2 and 0.5, respectively, to determine the confidence level of the model's predictions. To evaluate the effectiveness of this approach, the experimental results were compared with several existing cutting-edge methods. ResNet21 and BranchyNet20 were selected as the primary baseline models, replacing two-dimensional convolutions with one-dimensional convolutions while maintaining the original architecture to adapt them to multi-channel sensor signal processing tasks. To accurately evaluate the performance of the proposed approach, the experimental results are shown in Table 1. As can be seen, compared with the baseline approach, this approach significantly reduces model complexity while achieving higher accuracy. The introduction of the TernP module alone significantly improves the ResNet architecture, while the integration of PONO further enhances the model's ability to process multi-channel sensor signals with diverse characteristics.

[0060] Table 1 Performance comparison of different methods in fault diagnosis tasks

[0061] .

[0062] The real-time processing capability evaluation shown in Table 2 was performed independently on CPU hardware, aiming to simulate the resource-constrained environment of actual edge devices, where GPU acceleration is generally not possible. Under such deployment conditions, our method shows stable advantages under different batch sizes. For the single-sample processing scenario that is critical in real-time monitoring, our method significantly reduces the inference time compared to the baseline methods. The performance comparison between different method variants also reveals a relationship: although the computational overhead introduced by PONO when processing large batches of samples is negligible, it significantly improves feature accuracy in scenarios with only a single sample input, especially for tasks that require high-confidence output.

[0063] Table 2 Analysis of inference time under different batch sizes

[0064] .

[0065] Table 3's performance analysis of each branch demonstrates the efficiency of this multi-layered design. The device edge branch successfully processed the vast majority of samples with very high accuracy, while the vehicle gateway branch effectively handled the majority of the remaining samples. Only a small number of samples, typically those involving complex or ambiguous failure modes, required data upload to the cloud for processing. The confidence level determines whether to forward a task from the edge device to the cloud, not the final diagnosis result. The cloud environment supports more complex analysis methods.

[0066] Table 3 Performance analysis of each branch of the Dynamic TernP framework

[0067] .

[0068] This comprehensive series of evaluation results demonstrates that the proposed method effectively addresses key challenges in cloud-edge collaborative fault diagnosis. High accuracy and efficient sample distribution at the edge significantly reduce communication overhead, while the step-by-step processing strategy ensures reliable fault diagnosis under a wide range of operating conditions. The system demonstrates stable performance across various speeds and fault types, validating its feasibility and practicality in real-world high-speed train applications.

[0069] In summary, this paper proposes the D-TernP framework, an efficient cloud-edge collaborative fault diagnosis method for high-speed train scenarios that combines lightweight TernP modules with a dynamic reasoning mechanism. The main contributions of this paper include the following three aspects:

[0070] 1) This paper proposes a novel dynamic reasoning mechanism to achieve efficient task allocation. Most fault samples can be successfully processed at the edge device, thereby significantly reducing communication overhead while maintaining high diagnostic accuracy.

[0071] 2) This paper designs the TernP module, a lightweight structure that combines partial convolution and position normalization techniques. It effectively reduces model complexity while improving recognition accuracy, and has obvious advantages over traditional methods.

[0072] 3) This paper introduces an adaptive training strategy that combines dynamic weight adjustment with a reward mechanism to ensure stable model convergence during training and achieve optimal sample allocation between different branches.

[0073] Experimental results demonstrate that this method significantly outperforms existing baseline models in terms of accuracy while significantly reducing model complexity. In tests on edge devices, the device edge branch was able to process the majority of samples with extremely high accuracy, while the vehicle gateway branch effectively handled more complex scenarios. In terms of computational efficiency, this method also significantly optimized inference time on edge devices, making it well-suited for practical real-time monitoring needs. Furthermore, the system demonstrated consistent performance across various batch sizes and operating conditions, validating its stability in real-world deployments.

Claims

1. A cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning, characterized by: The following steps are involved: Step 1: Obtain multi-channel sensor monitoring data from the high-speed train monitoring system; Step 2: Construct a D-TernP network model with a multi-level export structure, including a backbone feature extraction network and multiple export sub-networks. The backbone feature extraction network is used to extract deep semantic information of the input signal and is composed of a cascade of lightweight residual structure TernP modules to adapt to edge device operation. Each export sub-network is connected to different depths of the backbone feature extraction network to implement branch prediction. Step 3: Train and fit the D-TernP network model based on historical monitoring data samples: Input the multi-channel sensor monitoring data into the D-TernP network model to obtain the prediction results of each export sub-network; Calculate the cross entropy loss of each export sub-network separately, dynamically set the weight coefficient of each export sub-network loss, and construct a weighted total loss function; Count the sample exit ratios of each exit sub-network, calculate the corresponding reward items, combine the weighted total loss function with the reward items to obtain the final total loss, and minimize the final total loss to optimize the network parameters; Step 4: Input the multi-channel sensor monitoring data to be detected into the trained D-TernP network model for inference and judgment: Each exit sub-network calculates the entropy value of the prediction result in the order of priority from the edge to the center. If the entropy value is lower than the corresponding threshold, it means that the D-TernP network model meets the confidence condition for the exit sub-network result, and the prediction result is directly output; If all edge exit sub-networks fail to meet the confidence condition, the multi-channel sensor monitoring data to be tested will be uploaded to the cloud model for more in-depth analysis and the final prediction results will be output; In step 2, the TernP module is represented as follows: For the input data tensor , B is the number of data samples for a single input, S is the total number of channels, L is the signal sample length, and partial convolution is used in the front The convolution operation is performed on the channel, and the subsequent channels remain unchanged to obtain the intermediate features. : ； in, and Represents the front of the input data tensor Channel and Channel data; Conv represents convolution operation; Then the features are normalized in position, that is, the results of each channel at each time step are normalized in position. The formula is as follows: ； ； in, and Represent the mean and standard deviation of all channels of the b-th sample and the l-th position respectively; is the component of the bth sample, sth channel, and lth position in the input data tensor X; Then, the complete operation structure of the TernP module is expressed as follows: ； in, is the weight matrix used to learn the feature combination between channels, is the bias term; and Respectively represent the mean and standard deviation calculated for all S channels at each position l, and is a learnable parameter; X is the input data tensor; Y is the output of the TernP module.

2. The cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning according to claim 1 is characterized in that: In step 3, for each outlet sub-network, the cross entropy function is used to evaluate the gap between the predicted label and the true label , the formula is as follows: ； Where C represents the total number of categories, Represents the network parameters from the input to the output sub-network, y c is the encoding of the true label, is the predicted probability of the current export sub-network for class c calculated by the Softmax function. The formula is as follows: ； in, Indicates that network parameters are taken into account , the output of the input sample x after passing through the nth exit sub-network; exp represents the exponential function; Represents the output value corresponding to class c; The weighted total loss function The formula is as follows: ； in, represents the weight of the nth export sub-network in the total loss, Represents the output prediction result; N is the number of export sub-networks; is the loss function; The final total loss To adopt a loss function that combines dynamic weighting and exit reward mechanism, the formula is as follows; ； ； in, represents the cross entropy loss of the n-th export sub-network, Represents the corresponding dynamic weight coefficient, t and T represent the current training step and the total training step respectively; the second term of the loss function formula is the reward term, which encourages the model to make predictions as early as possible in the shallow exit sub-network, P n represents the proportion of samples exiting the nth exit sub-network, is the reward adjustment coefficient of each exit sub-network; is the adjustment weight of the overall reward item.

3. The cloud-edge collaborative high-speed train fault diagnosis method based on dynamic neural reasoning according to claim 2 is characterized in that: In step 4, the entropy value of the prediction result is calculated as the prediction exit threshold, as shown in the following formula: ； in, is the entropy value of the prediction result.

Images