A kind of urban rail vehicle running gear system level intelligent fault diagnosis system of synergic optimization computing networking

Abstract

The application discloses a kind of urban rail vehicle running gear system level intelligent fault diagnosis systems of synergic optimization computing networking, belong to urban rail vehicle running gear fault diagnosis technical field.The synergic optimization computing networking urban rail vehicle running gear system level intelligent fault diagnosis system proposed in the application builds the diagnosis framework of edge real-time fault preliminary screening-cloud depth analysis-edge cloud dynamic cooperation, completes the reasonable division and decoupling collaborative deployment of edge and cloud diagnosis subsystem.Secondly, by the data slicing algorithm and downsampling algorithm proposed and combined with edge diagnosis result intelligent screening upload data, on the basis of effectively relieving communication link pressure, the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] This invention relates to a system-level intelligent fault diagnosis system for the running gear of urban rail vehicles with collaborative optimization computing networking, belonging to the field of fault diagnosis technology for the running gear of urban rail vehicles. Background Technology

[0002] With the rapid development of urban rail transit systems and the large number of vehicles put into operation, the operational safety of urban rail vehicles has become increasingly prominent, becoming a focus of industry attention. The running gear (also known as the bogie), as a core component of urban rail vehicles, is constantly exposed to complex and changing operating conditions, making it prone to performance degradation and even component failure, posing a significant threat to the safe operation of trains. Accurate and timely identification of the emergence and evolution of faults during equipment operation is of great significance for ensuring the safe operation of the system and reducing or avoiding major catastrophic accidents. However, currently, the urban rail vehicle fault monitoring systems deployed in various regions mainly employ two approaches: 1. The "single-machine acquisition + offline analysis" model. This approach deploys data acquisition equipment on-site to collect signals such as vibration and temperature, recording a large amount of raw data as files stored locally. Subsequently, data is periodically copied manually or read and analyzed by an independent offline computing program to achieve fault diagnosis and condition assessment. 2. The "cloud-based centralized diagnosis" model. This solution uploads all raw waveform data acquired by the on-site acquisition equipment directly to a remote cloud server or data center via the network, where the cloud platform performs unified data storage, processing, modeling, and diagnostic analysis.

[0003] Existing technical solutions have obvious shortcomings and defects in practical applications:

[0004] First, the core drawback of the "single-machine data acquisition + offline analysis" solution lies in its insufficient real-time performance and scalability. Because this solution relies on offline analysis, it cannot complete processing and diagnosis simultaneously with data acquisition, resulting in the inability to detect and issue timely alarms for abnormal changes in device status in real time, making it difficult to detect early faults promptly. Simultaneously, as the number of monitoring points and sampling channels increases, the pressure on local storage surges, and the computational load of offline processing also increases accordingly, limiting the overall scalability of the system and making it difficult to meet the needs of large-scale monitoring scenarios.

[0005] Secondly, the core drawback of the "cloud-based centralized diagnostics" solution lies in its high data transmission cost and excessive reliance on the network. Under high sampling rates and multi-channel conditions, the amount of raw waveform data is enormous. Uploading it completely to the cloud consumes a significant amount of network bandwidth, resulting in high transmission costs and significant data return delays. This model heavily relies on a continuous, stable, and high-speed network connection. In industrial sites with poor network conditions or at the network edge, the reliability and real-time performance of data transmission are difficult to guarantee, directly impacting the timeliness of fault diagnosis and the reliability of system operation. Summary of the Invention

[0006] This invention provides a data processing method for the running gear of urban rail vehicles, and further, based on the data processing method for the running gear of urban rail vehicles, constructs a system-level intelligent fault diagnosis system for the running gear of urban rail vehicles with a collaborative optimization computing network. This system processes the massive monitoring data continuously generated by the running gear of urban rail vehicles by coupling and coordinating edge computing and cloud computing.

[0007] The technical solution of this invention is:

[0008] According to a first aspect of the present invention, a data processing method for the running gear of an urban rail vehicle is provided, comprising:

[0009] The data slicing step is implemented using a data slicing algorithm, which includes:

[0010] Data from multiple sensors deployed on the running gear of urban rail vehicles is obtained; an adaptive physical threshold is constructed based on the key sampling channel signals in the sensor data; wherein the sensors include a first sensor and a second sensor; the first sensor is used to collect key sampling channel signals;

[0011] Using the adaptive physical threshold, the falling edge feature time representing the mechanical rotation reference point in the key sampling channel signal is extracted to construct a phase feature index sequence;

[0012] Based on the extracted phase feature index sequence, a slicing window is constructed using an equal-phase periodic locking slicing strategy;

[0013] The data from each sensor in the synchronously acquired second sensor is sliced ​​according to the slicing window to obtain the slice data of the corresponding sensor.

[0014] The downsampling step is implemented using a downsampling algorithm, which includes:

[0015] For the current slice of data extracted by the data slicing algorithm, set For any target index, the target feature dimension is... Calculate the corresponding virtual floating-point coordinates ;

[0016] The calculated floating-point coordinates are used to obtain the resampled feature values.

[0017] Furthermore, the adaptive physical threshold is defined as:

[0018] ;

[0019] In the formula, An adaptive physical threshold for the key sampling channel signal; This represents the maximum signal amplitude of the key sampling channel signal. This represents the minimum signal amplitude of the key sampling channel signal.

[0020] Furthermore, the step of using the adaptive physical threshold to extract the falling edge feature time representing the mechanical rotation reference point in the key sampling channel signal to construct a phase feature index sequence specifically involves: using the adaptive physical threshold to perform binarization logic determination on the signal located in the key sampling channel, and extracting the falling edge feature time representing the mechanical rotation reference point. That is, meeting the following conditions Constructing phase feature index sequences :

[0021] ;

[0022] In the formula, For the key sampling channel signal, the first The index position of each falling edge feature; For the key sampling channel signal in The signal amplitude at the index; For the key sampling channel signal in The signal amplitude at the next sampling point at the index.

[0023] Furthermore, the step of constructing a slicing window based on the extracted phase feature index sequence using an equal-phase periodic locking slicing strategy is specifically as follows: for the phase feature index sequence... Index sequence by phase feature Randomly select the first one Index position of each falling edge feature Starting from the point, lock the containing The interval of a complete mechanical cycle is used as the slicing window, and the starting index of the slice is... With Termination Index They respectively satisfy:

[0024] ;

[0025] In the formula, The first one extracted from the key sampling channel signal The index position of each falling edge feature; For the key sampling channel signal, the first The index position of each falling edge feature; .

[0026] Furthermore, the virtual floating-point coordinates The specific expression is:

[0027] ;

[0028] In the formula, For target index The corresponding virtual floating-point coordinates; This represents the total number of actual sampling points in the current slice data.

[0029] Furthermore, the step of obtaining the resampled feature values ​​using the calculated floating-point coordinates specifically involves: Let The integer index is the result of rounding down. Then the first Each resampled feature value for:

[0030] ;

[0031] In the formula, , This represents the signal amplitude of two adjacent discrete sampling points in the current slice of data.

[0032] According to a second aspect of the present invention, a system-level intelligent fault diagnosis system for the running gear of an urban rail vehicle is provided, which is used in the running gear of an urban rail vehicle. Sensors are installed on the running gear of the urban rail vehicle. The fault diagnosis system includes an edge system and a cloud system. The sensors establish communication with the edge system, and the edge system establishes communication with the cloud system. The running gear of the urban rail vehicle mainly includes an axle box, a traction motor, and a gearbox. A photoelectric sensor installed on the traction motor serves as a first sensor, and other sensors installed on the axle box, the traction motor, and the gearbox serve as second sensors.

[0033] The data parsing module is used to receive and parse sensor input signals.

[0034] The first message queue management module is used to realize communication between various modules of the edge system;

[0035] A data storage module, which is used to store data locally;

[0036] The waveform preprocessing module is used to resample the specified sensor data in the buffer of the waveform processing module according to the sensor selection command issued by the edge visualization interface, thereby reducing the amount of data pushed to the edge visualization interface in the edge system.

[0037] The rapid diagnosis module includes a built-in rapid diagnosis data receiving buffer, a downsampling data buffer, a data slicing unit, a downsampling unit, a diagnostic inference unit, and a judgment unit. The rapid diagnosis data receiving buffer caches data distributed by the first message queue management module based on the rapid diagnosis module start command issued via the visual interface. The downsampling data buffer caches downsampling data processed by the downsampling unit. The data slicing unit incorporates the data slicing algorithm from the aforementioned data processing method. The downsampling unit incorporates the downsampling algorithm from the aforementioned data processing method. The diagnostic inference unit performs fault diagnosis on the downsampling data to obtain edge diagnosis results. The judgment unit performs differentiated processing based on the edge diagnosis results to determine edge-distributed data.

[0038] An edge communication module, which uses the MQTT protocol to achieve transmission and interaction with the cloud system, and realizes the transmission of edge-distributed data to the cloud system;

[0039] An edge visualization interface is used to issue sampling control commands, sensor selection commands, rapid diagnostic module start commands, and for visualization display.

[0040] The first API service module is used for information transmission between the edge visualization interface and other modules of the edge system.

[0041] The cloud system includes:

[0042] The cloud communication module uses the MQTT protocol to communicate with the edge system, receiving edge distribution data transmitted by the edge system. It also pushes the edge diagnostic results from the received edge distribution data directly to the cloud visualization interface through the second API service module, and distributes the downsampled data from the edge distribution data to the data management module and the fine diagnosis module through the second message queue management module.

[0043] The second message queue management module is responsible for communication between various modules of the cloud system;

[0044] The data management module is used to manage downsampled data in the edge distribution data uploaded by the edge system and the cloud diagnostic results of the cloud system, which are distributed by the second message queue management module.

[0045] The fine-grained diagnostic module is used to identify faults in the downsampled data uploaded by the edge system and distributed by the second message queue management module in order to obtain cloud-based diagnostic results.

[0046] A cloud-based visualization interface is used to display cloud-based diagnostic results, edge diagnostic results, and perform statistical analysis.

[0047] The second API service module is used for information transmission between the cloud-based visual interface and other modules of the cloud system.

[0048] Furthermore, the cloud system also includes a system monitoring module, which is used to manage the entire fault diagnosis system based on the cloud diagnostic results.

[0049] The beneficial effects of this invention are as follows: Based on edge computing and cloud computing collaborative technologies, this invention proposes a system-level intelligent fault diagnosis system for the running gear of urban rail vehicles with a collaboratively optimized computing network. Based on an in-depth analysis of the dual requirements of real-time response and accurate judgment in running gear fault diagnosis, a diagnostic framework of edge real-time fault screening, cloud-based in-depth analysis, and edge-cloud dynamic collaboration is established, completing the reasonable division and decoupled collaborative deployment of edge and cloud diagnostic subsystems. Secondly, by proposing data slicing and downsampling algorithms and combining them with edge diagnostic results to intelligently filter uploaded data, the diagnostic accuracy is improved while effectively alleviating communication link pressure. As can be seen from the above, the data processing method of this invention significantly improves the robustness and survivability of the system in complex industrial environments. The system-level intelligent fault diagnosis system for the running gear of urban rail vehicles constructed based on this collaboratively optimized computing network provides solid technical support for accurate and efficient intelligent fault diagnosis of urban rail vehicle running gear. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the architecture of the intelligent fault diagnosis system for the running gear of urban rail vehicles based on the collaborative optimization computing network of the present invention.

[0051] Figure 2 This is a system workflow diagram provided according to an embodiment of the present invention.

[0052] Figure 3 This is a schematic diagram illustrating the principle of edge data slicing in an embodiment of the present invention.

[0053] Figure 4 This is a comparison chart of edge downsampling results in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0055] like Figure 1 As shown, a collaborative optimization computing network-based intelligent fault diagnosis system for the running gear of urban rail vehicles is provided. Sensors are installed on the running gear. The fault diagnosis system includes an edge system and a cloud system. The sensors communicate with the edge system, and the edge system communicates with the cloud system. Specifically, the running gear and the sensors installed on it serve as terminal devices. The running gear mainly includes three components: an axle box, a traction motor, and a gearbox. Sensors are installed on the corresponding components (accelerometer and current sensor are installed on the axle box; acceleration sensor, current sensor, and photoelectric sensor are installed on the traction motor; acceleration sensor and current sensor are installed on the gearbox).

[0056] Edge systems include:

[0057] The data parsing module is used to receive and parse sensor input signals.

[0058] The first message queue management module is used to realize communication between various modules of the edge system. Specifically, the first message queue management module distributes the parsed data to the buffer areas in the data storage module, the rapid diagnosis module, and the waveform processing module of the edge system. Among them, the data storage module is used for data backup, the rapid diagnosis module caches data for a set threshold duration according to the start command, and the waveform processing module caches data from specified sensors according to the sensor selection command.

[0059] The data storage module is used to store data locally; specifically, it is used to store the cache data in the data storage module locally according to the sensor number and the component in a time window segmentation manner.

[0060] The waveform preprocessing module is used to resample the specified sensor data in the buffer of the waveform processing module according to the sensor selection command issued by the edge visualization interface, thereby reducing the amount of data pushed to the edge visualization interface in the edge system.

[0061] The rapid diagnosis module includes a built-in rapid diagnosis data receiving buffer, a downsampling data buffer, a data slicing unit, a downsampling unit, a diagnostic inference unit, and a judgment unit. The rapid diagnosis data receiving buffer caches data distributed by the first message queue management module based on the rapid diagnosis module start command issued via the visual interface. The downsampling data buffer caches downsampling data processed by the downsampling unit. The data slicing unit incorporates a data slicing algorithm. The downsampling unit incorporates a downsampling algorithm. The diagnostic inference unit performs fault diagnosis on the downsampling data to obtain edge diagnosis results. The judgment unit performs differentiated processing based on the edge diagnosis results to determine edge-distributed data.

[0062] An edge communication module, which uses the MQTT protocol to transmit and interact with the cloud system, and transmits edge distribution data to the cloud system; wherein, the edge distribution data includes edge diagnostic results, or includes edge diagnostic results and downsampled data.

[0063] An edge visualization interface is used to issue sampling control commands, sensor selection commands, rapid diagnostic module start commands, and for visualization display.

[0064] The first API service module is used for information transmission between the edge visualization interface and other modules of the edge system. Specifically: it establishes information transmission between the data parsing module, waveform preprocessing module, rapid diagnosis module, and edge visualization interface; the edge visualization interface issues sampling control commands, and the data parsing module receives the commands through the first API service module for parsing; the waveform preprocessing module issues sensor selection commands, and the waveform preprocessing module receives the commands through the first API service module for resampling, and pushes the resampling data to the edge visualization interface in real time for waveform monitoring; the edge visualization interface issues a rapid diagnosis module start command, and the rapid diagnosis module receives the commands through the first API service module for data caching, and simultaneously pushes the edge diagnosis results to the edge visualization interface in real time through the first API service module.

[0065] The cloud system includes:

[0066] The cloud communication module interacts with the edge system via the MQTT protocol to receive edge distribution data transmitted by the edge system; it also pushes the edge diagnostic results in the received edge distribution data directly to the cloud visualization interface through the second API service module, and distributes the downsampled data in the edge distribution data to the data management module and the fine diagnosis module through the second message queue management module.

[0067] The second message queue management module is responsible for communication between various modules of the cloud system; specifically, the second message queue management module distributes the downsampled data received by the cloud communication module to the data management module and the fine diagnosis module.

[0068] The data management module is used to manage downsampled data in the edge distribution data uploaded by the edge system and the cloud diagnostic results of the cloud system distributed by the second message queue management module; the data management module is also used for storing and querying data.

[0069] The fine-grained diagnostic module is used to identify faults in the downsampled data uploaded by the edge system and distributed by the second message queue management module in order to obtain cloud-based diagnostic results.

[0070] A cloud-based visualization interface is used to display cloud-based diagnostic results, edge diagnostic results, and perform statistical analysis.

[0071] The second API service module is used for information exchange between the cloud-based visualization interface and other modules of the cloud system. Specifically, it establishes information exchange between the cloud communication module, the fine-grained diagnosis module, the data management module, and the cloud-based visualization interface. The cloud communication module directly pushes the acquired edge diagnosis results to the cloud-based visualization interface through the second API service module; the fine-grained diagnosis module feeds back the generated cloud diagnosis results to the cloud-based visualization interface through the second API service module; and the cloud-based visualization interface uses the second API service module to perform statistical analysis (such as querying and exporting) of the edge diagnosis results and cloud diagnosis results.

[0072] The cloud system also includes a system monitoring module, which is used to automatically manage the entire fault diagnosis system based on the cloud diagnostic results to ensure the normal operation of the system.

[0073] Furthermore, both the edge system and the cloud system deploy pre-trained convolutional neural network models based on the PyTorch architecture. The edge system's rapid diagnosis module deploys a lightweight neural network model for real-time processing of massive amounts of raw data (such as CNN models like 1D-CNN and WD-CNN), aiming to quickly filter and identify potential fault data segments (i.e., downsampled data). The cloud system's fine-grained diagnosis module deploys a complex neural network model to perform deep feature extraction on the fault data segments uploaded by the edge system to determine the specific fault type. To ensure the stability of collaborative diagnosis, data slicing and downsampling units are added before the lightweight neural network model's inference and diagnosis in the edge system. This ensures that the data features input to the lightweight neural network model and the subsequent complex neural network model in the cloud maintain a high degree of consistency in physical cycles and data dimensions. This effectively shields the signal distribution from the effects of vehicle speed fluctuations and operating condition interference, avoiding a decrease in recognition accuracy due to unstable input segments. After model training, the `torch.save` command serializes and saves the object containing the complete network structure, training weights, embedded preprocessing modules, and necessary configuration information as a `.pth` file. Therefore, during the inference phase, the edge system or cloud system only needs to load a single model file corresponding to the component. It can then directly input a fixed-length vibration data segment, and the model will automatically complete the entire process from preprocessing to feature inference and output diagnostic results. This ultimately achieves efficient and reliable interaction throughout the entire process of "rapid initial screening at the edge and accurate diagnosis in the cloud." Applying the above technical solution, it can be seen that this invention uses an edge-cloud collaborative architecture to achieve efficient deployment and dynamic inference of the intelligent fault diagnosis model for the running gear of urban rail vehicles.

[0074] Furthermore, the complex neural network model adopts a model based on an improved CNN model: such as the existing publicly available MMD-CNN model, IMACNN-BiGRU model (MMD-CNN uses a ResNet-18 backbone network and MMD loss function to dynamically align the feature distributions of the source and target domains to achieve cross-speed fault classification. The IMACNN-BiGRU model integrates multi-channel CNN and BiGRU layers to capture multi-scale spatial-temporal features to adapt to complex working condition analysis), or a nonlinear spatial pyramid network model guided by a triple attention mechanism. The nonlinear spatial pyramid network model guided by a triple attention mechanism includes a nonlinear spatial pyramid pooling layer, a triple attention mechanism module, and a nonlinear enhancement classification head based on the KA principle, connected in sequence; wherein, the nonlinear spatial pyramid pooling layer (SPP, a type of...) The pooling layer technique used in CNNs enhances robustness by extracting local and global features in parallel through multi-scale spatial partitioning. The triple attention mechanism module models the interaction between channels and spatial dimensions after pooling at each scale to suppress noise interference and highlight fault discrimination features. Meanwhile, the nonlinear enhancement classification head, based on the KA principle, adaptively fits complex features through the superposition of learnable nonlinear functions to construct accurate classification boundaries, thereby achieving high-precision identification and decision-making for bearing faults. (The nonlinear enhancement module based on the KA principle replaces the traditional fully connected layer. Unlike fully connected layers, which are limited by fixed activation functions, the KA module uses the superposition of learnable nonlinear functions to approximate complex features. This adaptive fitting ability allows it to construct accurate classification boundaries even when faced with unclear fault features due to noise, significantly improving the model's noise robustness.)

[0075] like Figure 2 As shown, the specific workflow of a collaborative optimization computing network-based intelligent fault diagnosis system for the running gear of urban rail vehicles is as follows:

[0076] After the intelligent fault diagnosis system is started, the edge visualization interface of the edge system can issue system start and sampling control commands (the sampling control commands include sampling frequency and sampling channel). Each sensor corresponds to an independent sampling channel, and at least one photoelectric sensor in the traction motor and one acceleration sensor in any component are selected as sampling channels. Among them, the photoelectric sensor is the key sampling channel. After receiving the command, the edge system data parsing module begins to receive and parse the data collected by the sensor (i.e., perform A / D conversion on the received data), and distributes the parsed data to the buffer in the edge system's data storage module, fast diagnosis module, and waveform processing module through the first message queue management module. The data storage module stores the cached data locally based on sensor numbers and time-window segmentation (for example, if three accelerometers and one current sensor are installed on the axle box, numbered 1, 2, 3, and 4, the data from sensors numbered 1, 2, 3, and 4 are segmented into 1024 data points and stored in a file named after the axle box; the same applies to others). The waveform preprocessing module resamples the specified sensor data in the waveform preprocessing module's cache (using linear interpolation) according to the sensor selection command issued by the edge visualization interface and pushes it to the edge visualization interface in real time for waveform monitoring. The rapid diagnosis module has a built-in rapid diagnosis data receiving cache and a downsampling data cache. The former caches data distributed by the first message queue management module, and the latter caches downsampling data processed by the downsampling algorithm. Furthermore, the rapid diagnosis module adopts an on-demand startup mode, remaining in standby mode normally and only activated after receiving a startup command through the first API service module. After activation, data distributed by the first message queue management module begins to accumulate in the fast diagnostic data receiving buffer. When a set threshold is reached, receiving new data stops, the current buffer content is copied for subsequent processing, and the buffer is cleared for use in the next round. Subsequently, the data slicing algorithm built into the data slicing unit automatically extracts signal segments containing the complete working cycle based on the falling edge event of the key sampling channel to obtain sliced ​​data. The downsampling algorithm then processes the sliced ​​data according to the standard length to eliminate sampling fluctuations and period differences, reducing the amount of data while ensuring consistency and reducing the communication pressure of transmission. This downsampled data will then enter the downsampled data buffer.Downsampling data is input into the respective pre-trained lightweight neural network models of each component for real-time rapid inference and diagnosis. (If the sampling control command issued by the edge visualization interface involves acceleration and current sensors from one component, then one pre-trained lightweight neural network model is set; if the sampling control command issued by the edge visualization interface involves acceleration and current sensors from two components, then two pre-trained lightweight neural network models are set; if the sampling control command issued by the edge visualization interface involves acceleration and current sensors from three components, then three pre-trained lightweight neural network models are set. For example, if the photoelectric sensor in the traction motor, one acceleration sensor in the axle box, and one acceleration sensor in the gearbox are selected as sampling channels, then two pre-trained lightweight neural network models are set. Taking the axle box pre-trained lightweight neural network model diagnosis as an example, the edge diagnosis results are divided into three categories: all components are healthy (label 1), itself is healthy and any other component is faulty (label 2), itself is faulty and all other components are healthy (label 3); the same applies to others.) The edge diagnosis results are pushed to the edge visualization interface in real time. The rapid diagnosis module in the edge system performs differentiated processing based on the results: if the edge diagnosis results of all models correspond to label 1, then only the edge diagnosis results are fed back to the cloud system through the edge communication module; if the edge diagnosis results of any model correspond to label 2 or label 3, then the edge diagnosis results are encapsulated with the downsampled data input from all models through the edge communication module, and uploaded to the cloud system via the MQTT protocol for refined diagnosis, greatly reducing the amount of data uploaded and alleviating transmission pressure. Throughout the entire process of data slicing, downsampling, inference diagnosis, and result reporting, the system blocks new data from being written to the rapid diagnosis data receiving buffer to ensure processing timing and data consistency; after a single diagnosis process is completed, the system immediately unblocks, and the module automatically starts a new round of data accumulation and diagnosis. The above mechanism realizes efficient collaboration and orderly data flow between real-time edge processing and deep cloud analysis.

[0077] The cloud communication module establishes a stable connection with the edge communication module via a broker based on pre-configured server addresses and port information, ensuring reliable information reception. After receiving messages uploaded from the edge system, the cloud communication module categorizes and processes them: the acquired edge diagnostic results are directly pushed to the cloud visualization interface via the second API service module; downsampled data marked as faulty is distributed to the data management module and the fine-grained diagnostic module via the second message queue management module for subsequent in-depth processing. The data management module, acting as the cloud's data hub, continuously subscribes to downsampled data marked as faulty from the message queue and persistently stores it in a standardized format in the cloud database within the data management module. This establishes a traceable and queryable fault data archive, while providing unified recording and archiving support for all cloud-generated diagnostic results. The fine-grained diagnostic module verifies the downsampled data labeled as fault states uploaded by the edge system based on a pre-trained cloud-based complex neural network model. It not only reviews the edge diagnostic results but also further identifies the specific fault type and confidence level as the cloud diagnostic result. (For example, if two pre-trained lightweight neural network models are set (corresponding to the axle box and gearbox), and the edge system's fast diagnostic module determines that either model's edge diagnostic result corresponds to label 2 or label 3, then the edge communication module encapsulates the edge diagnostic result with the downsampled data of the original input of the two pre-trained lightweight neural network models and transmits it to the cloud system. The pre-trained cloud-based complex neural network model in the cloud system then diagnoses and identifies the specific fault type and confidence level.) Simultaneously, the generated cloud diagnostic results are also fed back to the cloud visualization interface through the second API service module. If the cloud diagnostic result indicates that the equipment status is normal, the system continues to monitor operation. If a fault is confirmed and its severity reaches a preset standard (confidence level reaches a preset value), the system monitoring module will automatically trigger an early warning or issue a shutdown command, thereby achieving controllable management of the fault diagnosis system's operating status.

[0078] The data slicing algorithm is as follows:

[0079] 1) To adapt to dynamic changes in sensor signal amplitude and avoid errors from manually set thresholds, adaptive physical threshold calculation is performed:

[0080] Based on the maximum signal amplitude of the key sampling channel signal segment located in the rapid diagnostic data receive buffer. Minimum signal amplitude Calculate the adaptive physical threshold The adaptive physical threshold (midpoint threshold) is defined as follows:

[0081] ;

[0082] In the formula, An adaptive physical threshold for the key sampling channel signal; This represents the maximum signal amplitude of the key sampling channel signal. This represents the minimum signal amplitude of the critical sampling channel signal. In the above, the critical sampling channel signal segment corresponds to the data signal segment collected by the photoelectric sensor mounted on the traction motor rotating shaft, stored in the rapid diagnostic data receiving buffer.

[0083] 2) Falling edge feature time localization: Using the adaptive physical threshold, the continuous key sampling channel signals located in the fast diagnostic data receiving buffer are binarized and logically determined to extract the falling edge feature time representing the mechanical rotation reference point. That is, meeting the following conditions Constructing phase feature index sequences :

[0084] ;

[0085] In the formula, For the key sampling channel signal, the first The index position of each falling edge feature (i.e., the instantaneous sampling index when the signal jumps from a high level to a low level across the threshold); For the key sampling channel signal in The signal amplitude at the index; For the key sampling channel signal in The signal amplitude at the next sampling point at the index.

[0086] 3) Based on the extracted phase feature index sequence Instead of the traditional fixed-time slicing, an equal-phase periodic locking slicing strategy is adopted: for phase feature index sequences Index sequence by phase feature Randomly select the first one Index position of each falling edge feature Starting from the point, lock the containing One complete mechanical cycle (in this embodiment) The interval is used as the slicing window, and the starting falling edge of the slice is indexed. index with the terminating (ending) falling edge They respectively satisfy:

[0087] ;

[0088] In the formula, The first one extracted from the key sampling channel signal The index position of each falling edge feature; For the key sampling channel signal, the first The index position of each falling edge feature; .

[0089] 4) Based on the slicing window, slice the data from each sensor (excluding the photoelectric sensor) that was acquired simultaneously to obtain the slice data for the corresponding sensor. In practical applications, the sample points of the slice data are indexed starting from 0.

[0090] This step ensures that each extracted data segment physically corresponds precisely to a fixed angle of rotation of the motor shaft, thus achieving physical alignment of rotating machinery fault characteristics at the data source. The principle and method of data slicing have also been validated in the BJTU-RAO bogie dataset, such as... Figure 3 As shown, starting from a falling edge of the key sampling channel signal (key phase signal), and using four falling edges as the endpoint, the synchronously acquired data slices are divided into four complete cycles of data segments.

[0091] As can be seen from the above technical solution, the data slicing algorithm proposed in this invention is a method of equiphase periodic locking slicing based on the midpoint of the physical amplitude of the key sampling channel signal. This method analyzes the pulse waveform of the key sampling channel signal in real time (collected by a photoelectric sensor installed on the rotating shaft of the traction motor), and adaptively determines an adaptive physical threshold using the current physical amplitude range of the signal to accurately extract the rotation period features, thereby achieving synchronous angular interception of vibration data. This process not only significantly reduces the amount of original data but also transforms the non-stationary time-domain signal into a stationary angular-domain signal, ensuring that each sample strictly corresponds to a fixed mechanical rotation period (containing complete fault information), rather than a fixed time length. This processing method not only effectively eliminates the influence of speed fluctuations and ensures the consistency of the relative position and period number of the fault impact signal in the sample vector, but also solves the problem of characteristic frequency drift and "spectral ambiguity" that easily occurs under variable speed conditions through physical alignment of feature dimensions.

[0092] The downsampling algorithm is as follows:

[0093] 1) The current slice of data extracted by the data slicing algorithm ,set up For the target feature dimension (in this embodiment) In order to transform high-density slice data Compress and map to a standard space for any target index Calculate the corresponding virtual floating-point coordinates The specific expression is:

[0094] ;

[0095] In the formula, For target index The corresponding virtual floating-point coordinates; For the current slice data The actual total number of sampling points (i.e., data length, which varies with the instantaneous speed of the motor). This step ensures accurate alignment of the signal's beginning and end phases during compression / stretching.

[0096] 2) Using the calculated floating-point coordinates Obtain the resampled feature values (i.e., downsampled data): Let The integer index is the result of rounding down. Weight of the decimal part (0) ), then the first Each resampled feature value for:

[0097] ;

[0098] In the formula, Mapped coordinates The offset weight; Mapped coordinates The integer part; , This represents the signal amplitude of two adjacent discrete sampling points in the current slice of data.

[0099] The downsampling method of this invention aims to solve the problem of inconsistent sample data length caused by motor speed fluctuations during the equal-angle slicing of data slices, and to eliminate the dependence of the input dimension of deep learning models on physical rotation speed. Since data slicing locks a fixed mechanical rotation period, the number of original sampling points (physical time length) corresponding to this period changes dynamically at different rotation speeds. To adapt to the requirements of convolutional neural networks for fixed input tensor dimensions, downsampling maps the variable-length original physical slices to a unified feature space. This not only achieves forced alignment of data dimensions but also implicitly completes the normalization from the time domain to the feature domain, allowing the model to focus on the fault characteristics of the waveform itself without needing to pay attention to the time scale differences of the signal. In other words, after processing through data slicing and downsampling, the system can reduce the original variable-length data segments, which could potentially reach tens of thousands of points, to a more consistent length. Unified transformation into feature vectors with fixed standards This not only provides standardized input for subsequent neural networks but also serves as an efficient compression method, significantly reducing the bandwidth required to upload data to the cloud while ensuring the integrity of fault characteristics. Figure 4The image shows a comparison of sliced ​​data at different speeds after downsampling. The left side shows data segments obtained by slicing signals collected by a sensor in the running gear of an urban rail vehicle under low, medium, and high speed conditions. It can be seen that the data lengths obtained by slicing at equal angles differ at different speeds. To ensure that data from different speeds can be successfully incorporated into the model inference, the data slices are downsampled to obtain the following... Figure 4 The data segments on the right are of the same length. This means that the downsampled data segments all have the same length and can be directly fed into the pre-trained lightweight neural network model in the rapid diagnosis module for inference and diagnosis. The model's output inference results will be pushed to the edge visualization interface, and the system will determine whether to upload the downsampled data segment to the cloud system for detailed diagnosis based on the diagnostic results.

[0100] To validate this study, data spanning various speed conditions (low, medium, and high speeds) and four fault states (normal, fault 1, fault 2, and fault 3) were selected from the BJTU-RAO bogie public dataset. Three preprocessing methods (A: fixed 4096-point isochronous slices; B: data slicing + cubic spline interpolation; C: data slicing + downsampling (this invention)) were trained and tested using the same sample size and speed-fault state distribution ratio. Table 1 shows the classification accuracy and single-sample preprocessing time of the 1D-CNN and WD-CNN models after preprocessing using each method. The results show that in the 1D-CNN model, group B had the highest classification accuracy (Accuracy=1.000) but the longest single-sample preprocessing time (approximately 0.696 s). Group C had comparable classification accuracy to group B (Accuracy=0.994), but significantly reduced single-sample preprocessing time (approximately 0.084 s). Group A had the fastest preprocessing speed (approximately 0.003 s). While achieving the highest accuracy (accuracy = 0.764) in the WD-CNN model, group A had the lowest classification accuracy and the shortest preprocessing time per sample (approximately 0.002 s). Group B had better classification accuracy than group A (accuracy = 0.861), but the longest preprocessing time per sample (approximately 0.631 s). Group C had the highest classification accuracy, with preprocessing speed between groups A and B (approximately 0.083 s). In summary, method A consistently had the lowest classification accuracy, indicating a significant performance loss under varying rotational speeds. While method B had higher accuracy, its excessively long preprocessing time compromised its real-time performance. Method C, while maintaining high accuracy, significantly reduced computational overhead, achieving a better accuracy-efficiency trade-off and making it suitable for real-time preprocessing scenarios in edge systems.

[0101] Table 1

[0102]

[0103] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A data processing method for the running gear of urban rail vehicles, characterized in that, include: The data slicing step is implemented using a data slicing algorithm, which includes: Data from multiple sensors deployed on the running gear of urban rail vehicles is obtained; an adaptive physical threshold is constructed based on the key sampling channel signals in the sensor data; wherein the sensors include a first sensor and a second sensor; the first sensor is used to collect key sampling channel signals; Using the adaptive physical threshold, the falling edge feature time representing the mechanical rotation reference point in the key sampling channel signal is extracted to construct a phase feature index sequence; Based on the extracted phase feature index sequence, a slicing window is constructed using an equal-phase periodic locking slicing strategy; The data from each sensor in the synchronously acquired second sensor is sliced ​​according to the slicing window to obtain the slice data of the corresponding sensor. The downsampling step is implemented using a downsampling algorithm, which includes: For the current slice of data extracted by the data slicing algorithm, set For any target index, the target feature dimension is... Calculate the corresponding virtual floating-point coordinates ; The calculated floating-point coordinates are used to obtain the resampled feature values.

2. The data processing method for the running gear of urban rail vehicles according to claim 1, characterized in that, The adaptive physical threshold is defined as: ； In the formula, An adaptive physical threshold for the key sampling channel signal; This represents the maximum signal amplitude of the key sampling channel signal. This represents the minimum signal amplitude of the key sampling channel signal.

3. The data processing method for the running gear of urban rail vehicles according to claim 1, characterized in that, The step of using the adaptive physical threshold to extract the falling edge feature time representing the mechanical rotation reference point in the key sampling channel signal and constructing a phase feature index sequence specifically involves: using the adaptive physical threshold to perform binarization logic determination on the signal located in the key sampling channel, and extracting the falling edge feature time representing the mechanical rotation reference point. That is, meeting the following conditions Constructing phase feature index sequences : ； In the formula, For the key sampling channel signal, the first The index position of each falling edge feature; For the key sampling channel signal in The signal amplitude at the index; For the key sampling channel signal in The signal amplitude at the next sampling point at the index.

4. The data processing method for the running gear of urban rail vehicles according to claim 1, characterized in that, The extracted phase feature index sequence is used to construct a slicing window using an equal-phase periodic locking slicing strategy. Specifically, for the phase feature index sequence... Index sequence by phase feature Randomly select the first one Index position of each falling edge feature Starting from the point, lock the containing The interval of a complete mechanical cycle is used as the slicing window, and the starting index of the slice is... With Termination Index They respectively satisfy: ； In the formula, The first one extracted from the key sampling channel signal The index position of each falling edge feature; For the key sampling channel signal, the first The index position of each falling edge feature; .

5. The data processing method for the running gear of urban rail vehicles according to claim 1, characterized in that, The virtual floating-point coordinates The specific expression is: ； In the formula, For target index The corresponding virtual floating-point coordinates; This represents the total number of actual sampling points in the current slice data.

6. The data processing method for the running gear of urban rail vehicles according to claim 1, characterized in that, The process of obtaining the resampled feature values ​​using the calculated floating-point coordinates specifically involves: Let The integer index is the result of rounding down. Then the first Each resampled feature value for: ； In the formula, , This represents the signal amplitude of two adjacent discrete sampling points in the current slice of data.

7. A system-level intelligent fault diagnosis system for the running gear of urban rail vehicles with collaborative optimization computing networking, characterized in that, This system is used for the running gear of urban rail vehicles, and sensors are installed on the running gear. The fault diagnosis system includes an edge system and a cloud system. The sensors establish communication with the edge system, and the edge system establishes communication with the cloud system. The running gear of the urban rail vehicle mainly includes an axle box, a traction motor, and a gearbox. The photoelectric sensor installed on the traction motor serves as the first sensor, and other sensors installed on the axle box, traction motor, and gearbox serve as the second sensor. The data parsing module is used to receive and parse sensor input signals. The first message queue management module is used to realize communication between various modules of the edge system; A data storage module, which is used to store data locally; The waveform preprocessing module is used to resample the specified sensor data in the buffer of the waveform processing module according to the sensor selection command issued by the edge visualization interface, thereby reducing the amount of data pushed to the edge visualization interface in the edge system. The rapid diagnosis module includes a built-in rapid diagnosis data receiving buffer, a downsampling data buffer, a data slicing unit, a downsampling unit, a diagnostic inference unit, and a judgment unit. The rapid diagnosis data receiving buffer caches data distributed by the first message queue management module based on the rapid diagnosis module start command issued by the visual interface. The downsampling data buffer caches downsampling data processed by the downsampling unit. The data slicing unit incorporates the data slicing algorithm from the data processing method of claim 1. The downsampling unit incorporates the downsampling algorithm from the data processing method of claim 1. The diagnostic inference unit performs fault diagnosis on the downsampling data to obtain edge diagnosis results. The judgment unit performs differentiated processing based on the edge diagnosis results to determine edge-distributed data. An edge communication module, which uses the MQTT protocol to achieve transmission and interaction with the cloud system, and realizes the transmission of edge-distributed data to the cloud system; An edge visualization interface is used to issue sampling control commands, sensor selection commands, rapid diagnostic module start commands, and for visualization display. The first API service module is used for information transmission between the edge visualization interface and other modules of the edge system. The cloud system includes: The cloud communication module uses the MQTT protocol to communicate with the edge system, receiving edge distribution data transmitted by the edge system. It also pushes the edge diagnostic results from the received edge distribution data directly to the cloud visualization interface through the second API service module, and distributes the downsampled data from the edge distribution data to the data management module and the fine diagnosis module through the second message queue management module. The second message queue management module is responsible for communication between various modules of the cloud system; The data management module is used to manage downsampled data in the edge distribution data uploaded by the edge system and the cloud diagnostic results of the cloud system, which are distributed by the second message queue management module. The fine-grained diagnostic module is used to identify faults in the downsampled data uploaded by the edge system and distributed by the second message queue management module in order to obtain cloud-based diagnostic results. A cloud-based visualization interface is used to display cloud-based diagnostic results, edge diagnostic results, and perform statistical analysis. The second API service module is used for information transmission between the cloud-based visual interface and other modules of the cloud system.

8. The intelligent fault diagnosis system for the running gear of urban rail vehicles with collaborative optimization computing networking as described in claim 7, characterized in that, The cloud system also includes a system monitoring module, which is used to manage the entire fault diagnosis system based on the cloud diagnostic results.

Images