Locomotive bearing fault detection method and system based on neural network and time sequence attention mechanism

Abstract

The application relates to the technical field of locomotive bearing fault detection, and discloses a locomotive bearing fault detection method and system based on a neural network and a time sequence attention mechanism; the system comprises a sensor assembly, an AD acquisition module, a data processing CPU, a locomotive stability detection module, a double-path redundant network communication module, a CAN bus module, a 5G network card module and a double-path power supply redundancy module; vibration signals, temperature signals, acceleration signals and stability information of a locomotive to be detected are collected in real time; the vibration signals collected in real time are subjected to data preprocessing, and are divided into time sequence segments with a fixed length according to fault characteristic frequencies; after the time sequence segments after preprocessing are subjected to feature-level fusion with multi-source signals, the time sequence segments are input into a convolutional neural network for feature extraction; and the feature sequences extracted by the convolutional neural network are input into a long short-term memory network for time sequence modeling. The application can more accurately perform bearing fault diagnosis and adapt to variable working conditions and multiple fault types.

Description

Technical Field

[0001] This invention relates to the field of locomotive bearing fault detection technology, specifically to a locomotive bearing fault detection method and system based on neural networks and temporal attention mechanisms. Background Technology

[0002] Locomotive bogie bearings operate in a complex environment, bearing enormous mechanical loads during railway transportation. Due to factors such as high-speed travel, temperature fluctuations, and vibrations, the bearings' working environment is exceptionally complex. Over time, bearings may experience various faults such as fatigue, wear, and cracks. These faults may initially manifest as weak vibration signals, making them difficult to detect in a timely manner using traditional fault diagnosis methods.

[0003] Traditional vibration monitoring methods are relatively limited, typically relying on experience-based rules or manually extracted features (such as time-domain and frequency-domain features) to determine fault conditions. However, these methods often suffer from the following problems when dealing with complex and variable vibration signals from locomotive bogie bearings: Feature extraction is difficult: In actual operation, the vibration signal of locomotive bearings is often affected by noise interference and there are many different types of faults. Manual feature extraction cannot cover all situations and it is difficult to extract the key features that are truly helpful for fault diagnosis.

[0004] Missing temporal characteristics: Bearing fault signals are temporal and cumulative. Traditional methods have difficulty effectively capturing the temporal dependence and long-term trend of vibration signals, which makes it difficult to detect many minor or early faults.

[0005] Low diagnostic accuracy: Traditional vibration data fault diagnosis methods usually rely on simple statistical analysis and pattern recognition, which cannot handle complex multidimensional data, resulting in low classification accuracy. In particular, when faced with multiple fault types, misjudgment is prone to occur.

[0006] In recent years, with the rise and rapid development of deep learning technology, methods based on deep learning models such as convolutional neural networks (CNN), long short-term memory networks (LSTM), and attention mechanisms have made significant progress in the fields of vibration signal analysis and fault detection. These methods can automatically learn to extract features from raw signals and are particularly adept at handling time-series data and complex nonlinear problems.

[0007] CNN: Convolutional layers can effectively extract local features from signals, and are especially suitable for processing vibration signals with spatial relationships, such as frequency components and amplitude changes. They can automatically extract key features from signals without the need for manual feature design.

[0008] LSTM: As a recurrent neural network (RNN) capable of capturing long-term dependencies, LSTM has advantages in processing time series data. It can effectively capture the dynamic changes in vibration signals and help diagnose periodic or cumulative faults.

[0009] Attention Mechanism: The temporal attention mechanism can assign different weights to different parts of the input signal according to the needs of the current task, thereby focusing on the features most useful for fault diagnosis. This is of great significance for feature combinations at multiple time steps in vibration signals, and can avoid the problem of insufficient local feature extraction in traditional methods.

[0010] Traditional methods for analyzing locomotive bearing vibration monitoring data typically rely on classical signal processing techniques and analysis methods based on frequency domain, time domain, and statistical characteristics. Below are some common traditional data analysis methods: Fast Fourier Transform (FFT) Analysis: The Fast Fourier Transform (FFT) is a standard method for converting vibration signals from the time domain to the frequency domain. By performing FFT analysis on the vibration signal, the spectrum of the vibration signal can be obtained, and then the various frequency components contained in the vibration signal can be identified. In vibration monitoring of locomotive bearings, FFT is often used to detect typical frequencies of bearing faults, such as rolling element frequencies, inner and outer ring frequencies, and friction frequencies. By analyzing the amplitude of different frequency components, it can be determined whether the equipment has a fault. For example, when a fault occurs, the amplitude of a specific frequency component will increase significantly, indicating that a fault may have occurred. FFT analysis can intuitively display the frequency components in the vibration signal, helping to detect common bearing faults, such as rolling element and inner / outer ring damage. However, FFT only provides frequency domain information and cannot effectively capture dynamic changes in time series. It is also sensitive to noise and requires an accurate fault frequency model.

[0011] Time-domain analysis: Time-domain analysis methods determine the presence of faults by directly analyzing the changes in vibration signals over time. Commonly used time-domain features include signal amplitude, mean, standard deviation, peak value, and root mean square (RMS) value. In bearing fault monitoring, time-domain analysis can be used to assess the changing trends and fluctuations of vibration signals. For example, by calculating the RMS value of the vibration signal, it can be determined whether the signal exceeds a preset threshold, thus indicating a potential equipment fault. Common time-domain analysis methods include RMS, peak value analysis, and pulse frequency analysis. Time-domain analysis is simple and intuitive with low computational complexity. However, it is less effective than frequency-domain analysis for complex vibration modes and cannot handle high-frequency components in the signal.

[0012] Envelope analysis is a time-domain signal analysis method that extracts the envelope of a signal to analyze the instantaneous amplitude changes of the vibration signal, thereby determining the likelihood of a fault. Envelope analysis is primarily used to identify early faults in components such as rolling bearings. It is widely applied in rolling bearing fault detection because rolling bearing faults typically manifest as periodic impact signals, and the envelope effectively extracts the periodic characteristics of these signals. Envelope analysis is effective for early detection of rolling bearing faults and can effectively handle complex nonlinear vibration signals. However, envelope analysis requires multiple signal processing steps, resulting in high computational complexity, and it relies heavily on appropriate parameter settings.

[0013] Waveform Feature Analysis (Time-Frequency Analysis): Time-frequency analysis methods, by simultaneously considering the time and frequency domain characteristics of a signal, can accurately analyze the non-stationarity of a signal (e.g., the change of signal frequency over time). Common time-frequency analysis methods include Short-Time Fourier Transform (STFT) and Wavelet Transform (WT). In bearing fault detection, time-frequency analysis helps reveal complex changes in vibration signals, especially when the fault signal exhibits non-stationarity and abrupt changes. The Short-Time Fourier Transform (STFT) can be used to analyze the spectral changes of a signal within a short time window, while the Wavelet Transform can extract detailed features of the vibration signal at multiple scales. While time-frequency analysis can consider both the time and frequency domain characteristics of a signal, it is suitable for processing complex vibration signals. However, time-frequency analysis involves significant computation and requires careful selection of the time window and scale.

[0014] Frequency domain statistical analysis: Frequency domain statistical analysis methods diagnose signals by calculating the statistical characteristics of vibration signals in the frequency domain (such as power spectral density and spectral amplitude). By assessing the energy in specific frequency bands, the likelihood of a fault can be identified. This method is widely used to analyze the operating status of locomotive bearings. For example, certain bearing faults can cause an increase in energy within a specific frequency range; by analyzing the power spectrum of these frequency bands, the presence of a fault can be determined. Frequency domain statistical methods are intuitive and reliable, and suitable for analyzing frequency changes related to specific fault types. However, this method may not be effective in capturing non-periodic faults and complex fault modes.

[0015] Time-Frequency Transform (TFA): A TFA is a representation of a vibration signal after a time-frequency transformation, allowing for analysis in both the time and frequency domains. Common TFAs include the Wigner-Ville distribution and the Hilbert-Huang transform (HHT). For complex nonlinear and non-stationary signals, TFAs can reveal multi-scale variations in the signal and help detect potential fault modes in locomotive bearings.

[0016] Currently, the existing methods for diagnosing locomotive bogie bearing faults face the following technical bottlenecks: Complexity of vibration signals: Vibration signals of bogie bearings usually contain multiple frequency components and noise interference, and the changes in signals in the early stage of a fault are weak. Traditional vibration fault detection methods often find it difficult to automatically extract effective features from complex vibration signals when processing locomotive bogie bearing vibration data, especially when facing vibration signals under noise interference and complex working conditions, making it difficult to guarantee high-precision fault detection.

[0017] Timing dependence is difficult to capture: Bearing failures often have certain timing variation characteristics. Fault signals of locomotive bogie bearings usually have strong timing dependence. Traditional methods are difficult to effectively capture the changing trends and periodic characteristics over a long period of time. Therefore, there is a great challenge in identifying early bearing failures and weak failures.

[0018] The problem of classifying multiple types of faults: In practical applications, locomotive bogie bearings may experience a variety of different types of faults (such as wear, cracks, missing rolling elements, etc.). These faults may exhibit similar characteristics in vibration signals, which increases the difficulty of fault classification. Traditional classification methods have low accuracy when faced with multiple fault types and cannot identify and distinguish different fault types in a timely and accurate manner.

[0019] Therefore, a locomotive bearing fault detection method and system based on neural networks and temporal attention mechanism are proposed to solve the above-mentioned problems. Summary of the Invention

[0020] To address the shortcomings of existing technologies, this invention provides a locomotive bearing fault detection method and system based on neural networks and temporal attention mechanisms. By combining CNN for local feature extraction, LSTM for capturing temporal dependencies, and an attention mechanism to enhance focus on key fault features, this invention also introduces multi-source signal fusion, adaptive gating under varying operating conditions, edge-end quantization deployment, and redundant communication design. This improves the accuracy of vibration data analysis, enables more accurate bearing fault diagnosis, adapts to various fault types and complex signal patterns, significantly improves the accuracy, efficiency, and real-time performance of fault detection, and solves problems such as difficulty in vibration data feature extraction, insufficient capture of temporal dependencies, and insufficient accuracy in fault classification.

[0021] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A locomotive bearing fault detection method based on neural networks and temporal attention mechanism, comprising the following processes: Real-time acquisition of vibration signals, temperature signals, acceleration signals, and stability information of the locomotive under test; The real-time acquired vibration signals are preprocessed, including AD conversion, noise reduction, and normalization. Based on the fault characteristic frequency of the locomotive bearing, the vibration signals are divided into several time series segments of fixed length. The length of each time series segment contains at least one complete fault impact cycle. The preprocessed vibration data time series segments are fused with synchronously acquired temperature signals, acceleration signals, and stationarity information at the feature level to form a multi-dimensional input feature matrix. A multi-dimensional input feature matrix is ​​input into a convolutional neural network for feature extraction. The convolutional neural network uses multi-scale convolutional kernels to extract high-frequency impact features and low-frequency degradation features in parallel. The feature sequences extracted by the convolutional neural network are input into the long short-term memory network for time-series modeling. The gating mechanism of the long short-term memory network dynamically adjusts the forgetting gate parameters according to the locomotive speed signal to adapt to the time-series dependency changes under varying operating conditions. A temporal attention mechanism is applied to the temporal modeling results output by the Long Short-Term Memory Network to perform feature weighting, resulting in a weighted feature vector of the vibration signal. The temporal attention mechanism assigns weight values ​​to each time step to pinpoint the precise moment when the fault impact occurs. The weighted feature vector of the vibration signal is fused with the statistical features of the temperature signal and acceleration signal at the decision level to obtain a comprehensive fault judgment result, and the bearing condition of the locomotive under test is judged for multiple categories based on the result. The fault diagnosis results are transmitted simultaneously to the locomotive host and the cloud server via redundant network modules.

[0022] The objective of this invention can also be achieved through the following technical solutions.

[0023] A locomotive bearing fault detection system based on neural network and temporal attention mechanism includes a data processing CPU, sensor components, AD acquisition module, locomotive stability detection module, dual-redundant network communication module, CAN bus module, 5G network card module, and dual-redundant power supply module. The sensor assembly is used to collect vibration and temperature signals of the locomotive under test in real time. The AD acquisition module is used to convert the vibration and temperature signals acquired by the sensor components into digital signals and upload them to the data processing CPU. The locomotive stability detection module is used to detect the locomotive's acceleration signal and stability information in real time and upload them to the data processing CPU. The data processing CPU has a built-in fusion model of a convolutional neural network, a long short-term memory network, and a time-series attention mechanism, which is trained offline. The model is deployed at the edge after quantization and compression to achieve millisecond-level inference latency. The data processing CPU is used to preprocess the real-time acquired vibration signals and divide them into several fixed-length time series segments. The convolutional neural network is used to extract features from each vibration data time series segment, and the long short-term memory network is used to perform time-series modeling on the extracted feature sequences. The time-series attention mechanism is applied to the time-series modeling results to perform feature weighting to obtain a weighted feature vector of the vibration signal. Based on the weighted feature vector, the condition of the locomotive bearing under test is judged for multiple types of faults. Both the dual-redundant network communication module and the CAN bus module are used to upload the locomotive bearing fault diagnosis results obtained by the data processing CPU to the locomotive host, and the dual-redundant network communication module automatically switches to the backup link when the main link is interrupted. The 5G network card module is used to upload the diagnostic results of the data processing CPU and the locomotive status information to the cloud server in real time. The dual-power redundant module is used to supply power to the data processing CPU, sensor components, AD acquisition module, locomotive stability detection module, dual-redundant network communication module, CAN bus module, and 5G network card module.

[0024] The beneficial effects of this invention are: (1) The hardware system of this invention uses Rockchip RK3588 as the core for data acquisition and processing, combined with high-precision sensors (including ADS7691, composite piezoelectric ceramic sensor, PT100 sensor, MEMS sensor and gyroscope) and multiple data transmission methods to meet the high requirements of locomotive bearing vibration monitoring. These sensors can not only provide accurate vibration signals, but also judge the dynamic state and stability of the vehicle through multi-dimensional data fusion, further improving the accuracy of fault diagnosis.

[0025] (2) This invention employs a fire-resistant, bendable cable design to connect equipment and sensors, ensuring stable signal transmission in harsh environments. This invention utilizes a redundant design, including dual-path redundant network communication modules and dual-path redundant power supply modules, ensuring the system can continue to operate normally in the event of a fault. This invention employs multiple data communication interfaces (5G, Modbus TCP / TRDP, CAN bus, etc.) and their redundant network design to ensure the stability and reliability of the system in different scenarios.

[0026] (3) This invention combines CNN for local feature extraction, LSTM to capture temporal dependencies, and Attention mechanism to enhance attention to key fault features. It also introduces multi-source signal fusion, variable working condition adaptive gating, edge quantization deployment and redundant communication design, thereby improving the analysis accuracy of vibration data, enabling more accurate bearing fault diagnosis, adapting to various fault types and complex signal patterns, significantly improving the accuracy, efficiency and real-time performance of fault detection, and solving problems such as difficulty in vibration data feature extraction, insufficient capture of temporal dependencies and insufficient accuracy of fault classification.

[0027] (4) This invention employs deep learning techniques (CNN, LSTM, Attention mechanism) to automatically extract and learn complex features of signals. These techniques can effectively capture the nonlinear and time-varying characteristics in vibration signals. In particular, the LSTM layer can identify long-term dependencies and adapt to the non-stationarity of signals, while the Attention mechanism helps the model focus on the signal parts that are crucial for fault judgment, thereby better dealing with complex faults.

[0028] (5) This invention achieves automatic feature learning and optimization through deep learning. CNN can automatically extract fault-related features from the original vibration data, while LSTM and Attention layers can automatically adjust the focus of attention according to the dynamic changes of the data itself, thereby avoiding the difficulties of manual feature selection and threshold setting.

[0029] (6) This invention effectively models the time dependence of signals by introducing an LSTM network, which can capture the dynamic changes of vibration signals in time series, such as periodic faults and sudden faults. In addition, the Attention mechanism can make the model focus on the abnormal features of those key time periods by weighting the outputs of different time steps, which helps to identify and locate faults more accurately.

[0030] (7) Although deep learning models usually require a lot of computing resources, with the development of modern hardware (such as GPUs), the CNN+LSTM+Attention method can achieve high real-time performance and computational efficiency while maintaining high accuracy through parallel computing and model optimization. Especially in the real-time detection of equipment failures, deep learning models can be trained end-to-end, respond quickly to real-time signals, and provide fast and accurate diagnostic results.

[0031] (8) In its specific implementation, this invention adopts an offline training and edge quantization deployment method, which makes the inference latency of the model on the RK3588 platform less than 10 milliseconds, meeting the requirements of real-time locomotive monitoring. Through experimental comparison, the method achieved an average accuracy of 97.6% and a recall of 96.8% on a real-world dataset containing five fault types, including bearing wear, missing rolling elements, and inner and outer ring cracks. This is significantly better than the traditional FFT (accuracy 71.2%), temporal feature method (accuracy 68.5%), and single CNN model (accuracy 85.3%).

[0032] The locomotive bearing fault detection method and system based on neural networks and temporal attention mechanism has the advantages of high fault detection accuracy, strong adaptability to changing working conditions, good real-time performance, high system reliability, wide versatility, and convenient deployment and maintenance. It can realize the status monitoring and fault diagnosis of locomotive bearings throughout their entire life cycle, providing reliable technical support for the intelligent operation and maintenance and safe operation of railway locomotives.

[0033] Based on the above technical solution, the present invention can be further improved as follows.

[0034] Furthermore, the AD acquisition module uses an ADS7691 chip, and its input end is connected to the sensor assembly via a fire-resistant, bendable cable. The output end of the AD acquisition module is connected to a data processing CPU, which is also bidirectionally connected to the locomotive stability detection module, the dual-redundant network communication module, the CAN bus module, and the 5G network card module. The dual-redundant power supply module consists of two power supply modules, one of which is used as the main power supply module and the other as a backup power supply module. Both power supply modules can supply power to the above modules, with a power consumption of less than 8W.

[0035] The beneficial effects of adopting the above-mentioned further solutions are that a standardized unidirectional signal transmission link for hardware modules is realized, and the fireproof and bendable cable effectively improves the anti-interference and physical toughness of signal transmission, making it suitable for the harsh operating environment of locomotives with high temperature and vibration. The dual-power supply redundancy design enables uninterrupted power supply, avoids data acquisition interruption and diagnostic result loss due to power failure, ensures stable operation of the system around the clock, and the low power consumption design is suitable for locomotive on-board power supply scenarios.

[0036] Furthermore, the sensor assembly consists of a composite piezoelectric ceramic sensor and a PT100 temperature sensor. The composite piezoelectric ceramic sensor is installed in the vertical and horizontal directions of the bearing housing of the locomotive under test, and is used to collect dual-channel vibration signals in real time, which are then input to different channels of the convolutional neural network to extract radial and axial fault features. The PT100 temperature sensor is installed on the outer ring of the bearing and the gearbox of the locomotive under test, and is used to collect the temperature signals of the environment and components of the locomotive under test in real time, with a temperature measurement accuracy of ±0.5℃.

[0037] The beneficial effects of adopting the above-mentioned further solutions are that dual-channel vibration signal acquisition can simultaneously capture the radial and axial vibration characteristics of the bearing, making up for the deficiency that single-channel acquisition cannot fully reflect the bearing's operating status, and allowing the convolutional neural network to extract more complete fault features; the PT100 high-precision temperature sensor can capture the weak temperature changes in the early stage of bearing failure, forming a multi-physical quantity complementarity with the vibration signal, greatly improving the sensitivity of early fault detection, and providing dual judgment basis for fault diagnosis.

[0038] Furthermore, the locomotive stability detection module is installed on the floor inside the locomotive under test and consists of a MEMS six-axis accelerometer and a gyroscope. The MEMS six-axis accelerometer is used to detect the lateral, longitudinal, and vertical acceleration signals of the locomotive in real time, and the gyroscope is used to detect the rate of change of the locomotive's angular velocity in real time to characterize stability.

[0039] The beneficial effects of adopting the above-mentioned further solution are that the MEMS six-axis accelerometer, in conjunction with the gyroscope, can accurately distinguish between the inherent vibration caused by bearing failure and the environmental vibration caused by locomotive road conditions and driving posture, effectively eliminating the influence of environmental interference on the misjudgment of fault diagnosis; the module is installed on the floor of the carriage, and is less directly affected by bearing vibration, and the collected stability information can truly reflect the locomotive's operating conditions, providing reliable operating condition data support for the adaptive adjustment of the LSTM forget gate speed, and improving the model's adaptability to changing operating conditions.

[0040] Furthermore, the dual-redundant network communication module consists of two RJ45 Ethernet interfaces, which support ModbusTCP and TRDP protocols respectively, and are used to realize the interaction between the data processing CPU and the locomotive host, and to provide real-time feedback of processing results to the locomotive host; the CAN bus module adopts the CANFD protocol with a transmission rate of 500kbps, and is used for lightweight reporting of low-data-volume fault results.

[0041] The advantages of adopting the above-mentioned further solutions are that the dual network interface primary and backup redundancy design ensures the continuity of communication links, and quickly switches to the backup link when the primary link fails, ensuring that the fault diagnosis results are uploaded to the locomotive host in real time; the dual protocol support can be adapted to the vehicle network standards of different locomotive models, greatly improving the system's versatility and compatibility; the lightweight design of the CANFD protocol reduces the vehicle network bandwidth occupation while meeting the fault result reporting requirements, and forms a high- and low-speed communication complement with the redundant Ethernet.

[0042] Furthermore, the training dataset of the fusion model of the data processing CPU contains measured data of six states of locomotive bearings, with a total duration of over 500 hours and a sampling rate of 25.6kHz. After 100 rounds of offline training, the model achieved an average accuracy of 97.6% and a recall of 96.8% on the measured dataset.

[0043] The beneficial effects of adopting the above-mentioned further solutions are that the large sample and high sampling rate of the actual test dataset allow the model to fully learn the characteristic patterns of different fault types of locomotive bearings, thereby improving the model's fault identification and generalization capabilities; after the model is fully trained offline, it can directly achieve high-precision fault diagnosis when deployed at the edge, without the need for secondary on-site training, thus reducing the on-site deployment and maintenance costs of the system. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the hardware module structure of the present invention; Figure 2 This is a schematic diagram of the overall process of the locomotive bearing fault detection method of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0046] In the embodiments, by Figure 1-2 This invention presents a method and system for detecting locomotive bearing faults based on neural networks and temporal attention mechanisms. The invention includes a data processing CPU, sensor components, an AD acquisition module, a locomotive stability detection module, a dual-redundant network communication module, a CAN bus module, a 5G network card module, and a dual-redundant power supply module.

[0047] The sensor assembly described in this invention is used to acquire vibration and temperature signals of the locomotive under test in real time. Preferably, the sensor assembly can be composed of a composite piezoelectric ceramic sensor and a temperature sensor. The composite piezoelectric ceramic sensor is installed in the vertical and horizontal directions of the bearing housing of the locomotive under test, and is used to acquire dual-channel vibration signals in real time. These signals are then input to different channels of a convolutional neural network to extract radial and axial fault features, adapting to high-frequency vibration monitoring. The temperature sensor can be a PT100 sensor, installed in the bearings or gearboxes of the locomotive under test, or other locations requiring temperature measurement, to monitor temperature changes and acquire environmental and component temperature signals of the locomotive under test in real time.

[0048] In this invention, the input terminal of the AD acquisition module is connected to the sensor assembly, and the output terminal of the AD acquisition module is connected to the data processing CPU. The AD acquisition module is used to perform AD conversion on the vibration and temperature signals acquired by the sensor assembly, converting them from analog signals to digital signals, and then uploading them to the data processing CPU.

[0049] Preferably, the AD acquisition module can be an ADS7691, which has high sampling accuracy and low power consumption, and can provide accurate digital signals for subsequent processing. The AD acquisition module and the sensor components are connected by a fire-resistant, bendable cable to ensure long-term stable operation of the system in harsh environments. The cable has strong anti-interference capabilities and physical toughness, adapting to the challenges of high temperature, humidity, and vibration in industrial environments, thus extending the system's service life.

[0050] The locomotive stability detection module described in this invention is used to detect the locomotive's acceleration signal and stability information in real time and upload them to the data processing CPU. The locomotive stability detection module is installed on the floor inside the locomotive under test and can be composed of a MEMS six-axis accelerometer and a gyroscope. The MEMS six-axis accelerometer is used to detect the locomotive's acceleration signal in real time, and the gyroscope is used to detect the locomotive's stability in real time.

[0051] In this invention, the data processing CPU is connected to the AD acquisition module, the locomotive stability detection module, the dual-redundant network communication module, the CAN bus module, the 5G network card module, and the dual-redundant power supply module. The data processing CPU not only receives vibration and temperature signals after AD conversion in real time, as well as acceleration signals and stability information acquired by the locomotive stability detection module, but also incorporates a fusion model of an offline-trained convolutional neural network, a long short-term memory network, and a time-series attention mechanism. During the training phase, this model used measured locomotive bearing vibration data including normal conditions and five typical fault types, with a total data duration exceeding 500 hours and a sampling rate of 25.6kHz. After training, the model is quantized and compressed before being deployed at the edge, achieving millisecond-level inference latency. Specifically, the data processing CPU preprocesses the real-time acquired vibration signals, dividing them into several fixed-length time series segments. A convolutional neural network is used to extract features from each vibration data time series segment, and a long short-term memory network is used to perform time-series modeling on the extracted feature sequences. A time-series attention mechanism is applied to weight the features based on the time-series modeling results, obtaining a weighted feature vector of the vibration signal. Based on this weighted feature vector, multi-category fault judgments are made regarding the bearing status of the locomotive under test. The bearing fault diagnosis results and locomotive status information are fed back to remote maintenance personnel via 5G or a wired network.

[0052] Preferably, the data processing CPU can be Rockchip RK3588, which serves as the core of data acquisition and processing. It has powerful computing and data processing capabilities. The RK3588 not only supports real-time data processing but also AI model inference and optimization. It can perform complex fault detection tasks at the edge, providing strong support for real-time diagnosis.

[0053] This invention's system supports wired networks and interacts with the locomotive network using ModbusTCP or TRDP protocols. Both the dual-redundant network communication module and the CAN bus module can be used to upload the locomotive bearing fault diagnosis results obtained by the data processing CPU to the locomotive host. The dual-redundant network communication module enables interaction between the data processing CPU and the locomotive host, providing real-time feedback of processing results to the locomotive host and ensuring seamless connection between the locomotive host and this invention's system. The dual-redundant network communication module consists of two network interfaces, ensuring that data transmission can continue through the backup link even if one network link fails, thereby improving system reliability and security. For reporting low-volume results, this invention can use a CAN bus module, which simplifies data transmission and improves system compatibility, especially advantageous in scenarios requiring low-latency and low-bandwidth transmission.

[0054] The system of this invention supports 5G data reporting. The 5G network card module can transmit the real-time collected locomotive status information (i.e., vibration signal, temperature signal, acceleration signal, and stability information) and the diagnostic results of the data processing CPU to the cloud server in real time through the high-speed mobile network. This information is then fed back to remote operation and maintenance personnel for centralized processing and remote monitoring. Operation and maintenance personnel can monitor the equipment status at any time, perform system debugging and remote maintenance, and meet the needs of large-scale data analysis.

[0055] Considering power redundancy and power fluctuation tolerance, this invention features a dual-power redundancy module. This module powers the data processing CPU, sensor components, AD acquisition module, locomotive stability detection module, dual-redundant network communication module, CAN bus module, and 5G network card module. The dual-power redundancy module consists of two power modules: one as the primary power module and the other as a backup power module. Both modules can power the aforementioned modules. In the event of a power failure in the primary power module, the backup power module can maintain normal operation of the equipment, ensuring the continuity of data acquisition and processing and preventing data loss or equipment downtime due to power issues.

[0056] Based on the principle of the locomotive bearing fault detection system described above, this invention also proposes a locomotive bearing fault detection method based on neural networks and temporal attention mechanisms, such as... Figure 2 As shown, the specific process includes the following: S1 Data Acquisition Vibration sensors, temperature sensors, MEMS six-axis accelerometers, and gyroscopes are installed on key parts of the locomotive to collect vibration signals, temperature signals, acceleration signals, and stability information of the locomotive under test in real time, while also acquiring the locomotive speed signal and uploading it to the data processing CPU.

[0057] S2 Data Preprocessing The real-time acquired vibration signals undergo preprocessing operations such as AD conversion, denoising, and normalization. Based on the fault characteristic frequencies of the locomotive bearings, the vibration signals are divided into several fixed-length time series segments (each segment representing a sample). The length of each time series segment must contain at least one complete fault impact cycle. The preprocessed vibration signals are one-dimensional time series segments (100 sampling points in length), with an input dimension of (batch_size, 100, 1), where batch_size represents the number of samples input for each training iteration (to be modified based on training progress), 100 is the length of a single sample, and 1 represents a single-channel vibration signal. Normalization scales the signal values ​​to the [0, 1] range to avoid differences in numerical range affecting model convergence.

[0058] S3 Multi-Source Feature Fusion The preprocessed vibration data time series segments are fused with synchronously acquired temperature signals, acceleration signals, and stationarity information at the feature level to form a multi-dimensional input feature matrix. Statistical features such as mean and slope are extracted from the temperature signals; time-domain features such as root mean square and peak value are extracted from the acceleration signals; and the rate of change of angular velocity output from the gyroscope is extracted from the stationarity information. These features are then concatenated with the time-domain and frequency-domain features of the vibration signals to form a multi-dimensional input feature matrix with dimensions (batch_size, 100, 8).

[0059] S4 Feature Extraction The multidimensional input feature matrix is ​​fed into a Convolutional Neural Network (CNN) for feature extraction. Specifically, convolution and pooling operations are performed on each time series segment of data. The convolution operation extracts local features, such as frequency components and amplitude changes, from the time series by sliding the convolution kernel, resulting in a series of feature sequences. Pooling operations (such as max pooling) further reduce the dimensionality of the features, preserving the main features.

[0060] In this embodiment, the CNN consists of three convolutional layers, each using 64 convolutional kernels. The first layer has a kernel size of 3×1 and is used to extract high-frequency short-term features (such as impact signals and transient vibrations). The second layer has a kernel size of 4×1 and is used to capture periodic features with a medium time span (such as periodic wear of bearing rolling elements). The third layer has a kernel size of 5×1 and is used to capture low-frequency long-term trends (such as slow degradation caused by temperature changes). The convolution operation is represented by the following formula: (1) In the formula, This represents the convolution output at time t. Indicates the convolution kernel weights; σ represents the value of the input signal at time t+k; b represents the bias term, which is added to the output of the convolution operation to further adjust the numerical range of the feature map; σ represents the activation function, which introduces nonlinear characteristics and enhances the model's ability to fit complex fault modes (such as nonlinear friction and resonance); K represents the convolution kernel size.

[0061] Pooling downsamples the convolution result, which can be expressed by the following formula: (2) In the formula, The pooled output retains the maximum value in the local region to enhance salient features; , All input values ​​are within the pooling window, i.e., the output of the convolution. , Reducing the feature dimension (from 100→50→25) decreases the computational cost of subsequent LSTM; enhancing translation invariance improves robustness to small signal shifts.

[0062] S5 time series modeling The feature sequences extracted by the convolutional neural network are input into a Long Short-Term Memory (LSTM) network for temporal modeling. LSTM can effectively capture the temporal dependencies in vibration data and update the hidden states and memory units through a gating mechanism. In this embodiment, the LSTM adopts a two-layer bidirectional structure with a hidden layer dimension of 128.

[0063] The LSTM state update formula is as follows: Forgotten Gate: (3) The forgetting gate determines which information needs to be retrieved from the cell state. Forgotten in the middle, This is the result of the cell state update function. In the formula, It is a value between 0 and 1, representing the degree of retention (0 means complete forgetting, 1 means complete retention). The weight matrix of the forget gate (learnable parameters); Forget gate bias term (learnable parameter); It is the hidden state of the previous time step, obtained through the hidden state update function; It is the current input, i.e., the output after pooling. ; This is the Sigmoid function, which maps the output range to the interval between 0 and 1. In this step, the forget gate parameters are dynamically adjusted based on the locomotive's speed signal. When the speed changes, the forget gate bias term... Scaling the model proportionally allows it to adapt to changes in time-dependent conditions under varying operating conditions.

[0064] Input Gate: (4) The input gate determines which new information needs to be added to the cell state. In the formula, It controls the proportion of information to be updated, and the Sigmoid function is used to map the output range to the interval between 0 and 1. It is the weight matrix of the input gate (learnable parameters); It is the bias term of the input gate (a learnable parameter); It is the hidden state at the current time step, obtained through the hidden state update function. Update unit: (5) The update unit can generate new candidate information for the current time step, ready to be added to the cell state. In the formula, These are candidate values ​​obtained through the hyperbolic tangent function (tanh), ranging from [-1, 1], representing information that may contribute to the cell state when the current input is combined with the historical state; It is the weight matrix (learnable parameters) of the update unit; It is the bias term (learnable parameter) of the update unit.

[0065] Output gate: (6) The output gate determines how much cell state information should be included in the hidden state (output) at the current time step. In the formula, It controls the output ratio, and the Sigmoid function is used to map the output range to the interval between 0 and 1. , These are the weight matrix and bias terms corresponding to the gating unit.

[0066] Cell status update: (7) Cell state updates can update the cell state, i.e., the long-term memory portion, by multiplying the result of the forgetting gate and the cell state update at the previous time point, and then adding the result of the input gate and the update unit. In the formula, , It represents the cell state at the current moment compared to the previous moment.

[0067] Hidden state update: (8) Hidden state updates compute the hidden state at the current time step, which serves as the input for the next time step. In the formula, The output of the LSTM is passed to the next time step, where the cell state is nonlinearly transformed by the tanh function, and then the output intensity is controlled by the output gate.

[0068] S6 Feature Weighting A temporal attention mechanism is applied to the temporal modeling results output by a Long Short-Term Memory (LSTM) network to weight features, assigning different weights to features at each time step to highlight key features related to the fault, ultimately obtaining a weighted feature vector of the vibration signal. This embodiment employs an additive attention mechanism.

[0069] Attention score: (9) In the formula, The attention score at the i-th time step; The weight matrix is ​​a learnable matrix; For bias terms; This is the hyperbolic tangent function, used for nonlinear mapping; Let be the hidden state update value of LSTM at the i-th time step, i.e., in formula (8) .

[0070] Attention weights: (10) In the formula, Let be the attention weight for the i-th time step, which is normalized using Softmax so that the sum of the weights for all time steps is 1; T is the total length of the time series. Time steps with larger attention weight values ​​correspond to the precise moment when the fault impact occurs.

[0071] Weighted output: (11) In the formula, The weighted context vector is a fixed-length vector that integrates features from all time steps; T is the total number of time steps (e.g., the total length of the time series output by the LSTM). The LSTM output for each time step... According to its corresponding attention weight Perform a weighted summation to obtain the final context vector. This serves as input for subsequent classification or decoding tasks.

[0072] S7 Decision-Level Fusion and Fault Classification The weighted feature vector of the vibration signal is fused with the statistical features of the temperature and acceleration signals at the decision-level to obtain a comprehensive fault judgment result. Specifically, the context vector output by the Attention mechanism is... The input is fed into a fully connected layer to obtain the fault category probability distribution of the vibration signal. Simultaneously, statistical features are extracted from the temperature signal (mean, rate of change) and the acceleration signal (root mean square, peak value), and then input into independent shallow neural networks (two fully connected layers) to obtain the fault category probability distributions of the temperature and acceleration signals, respectively. Finally, the results from the three sources are fused using a weighted voting method to obtain a comprehensive fault judgment.

[0073] Fully connected layers and the Softmax function are used to calculate the failure probability. (12) In the formula, The score for the j-th type of fault; This represents the total number of bearing failure categories (such as normal operation, bearing wear, missing rolling elements, cracks in inner and outer rings, etc.). The sum of scores for all fault categories. The probability of the j-th type of fault. The probability of each fault category of the bearing is calculated by formula (12), and the category with the highest probability value is taken as the fault category of the locomotive bearing in the final diagnosis.

[0074] S8 Results Reporting The fault diagnosis results are uploaded to the locomotive host via a dual-redundant network communication module or a CAN bus module, and simultaneously uploaded to the cloud server via a 5G network card module for remote maintenance personnel to monitor.

[0075] The invention was validated on a real-world locomotive bearing dataset using the aforementioned system and method. The dataset contains five states: normal, bearing wear, missing rolling elements, inner ring crack, and outer ring crack, totaling 15,000 samples. Experimental results show that the method achieves an average accuracy of 97.6%, a recall of 96.8%, and an F1 score of 97.2%, with an average inference time of 8.2 milliseconds on the RK3588 platform. Comparison with traditional methods and single-model approaches is shown in the table below. Working principle: Step 1: The composite piezoelectric ceramic sensor (bearing housing vertical / horizontal direction) of the sensor assembly acquires dual-channel vibration signals, and the temperature sensor (PT100) acquires the temperature signals of the bearing and the surrounding environment. The above analog signals are converted into digital signals and amplified by the AD acquisition module (ADS7691) and then uploaded to the data processing CPU. At the same time, the MEMS six-axis accelerometer and gyroscope of the locomotive stability detection module acquire the locomotive acceleration signals and stability information, and directly upload them to the data processing CPU. The dual-power redundant module provides uninterrupted power supply to all hardware modules to ensure that the acquisition process is not interrupted. Step 2: Data Processing. The CPU sequentially performs preprocessing, feature-level fusion, feature extraction, temporal modeling, feature weighting, decision-level fusion, and fault classification operations on the received multi-source signals. First, the vibration signal is denoised and normalized, and divided into fixed-length time series segments according to the fault characteristic frequency. Then, the vibration sequence is fused with temperature, acceleration, and stationarity information at the feature level to form a multi-dimensional input feature matrix. High-frequency impact and low-frequency degradation features are extracted through a multi-scale convolutional neural network. The feature sequence is input into a speed-adaptive LSTM network for temporal modeling to capture the temporal dependencies of the signal. Then, the modeling results are weighted through a temporal attention mechanism to locate the fault impact moment and obtain a vibration weighted feature vector. Finally, the vibration weighted feature vector is fused with the statistical features of temperature and acceleration at the decision level, and a comprehensive fault judgment result is obtained through weighted voting, achieving accurate classification of six states: normal bearing condition, wear, missing rolling elements, inner ring crack, outer ring crack, and cage damage. Step 3: The data processing CPU uploads the fault diagnosis results to the locomotive host in real time via a dual-redundant network communication module or a CAN bus module, providing fault warnings and decision-making basis for the locomotive's local control system. At the same time, it uploads the fault results, raw data, and locomotive operating information to the cloud server via a 5G network card module, enabling real-time monitoring, data storage, and analysis by remote maintenance personnel. If the local network link fails, the dual-redundant network communication module automatically switches to the backup link. If the 5G network is interrupted, the data processing CPU will cache the latest diagnostic results for the past hour and automatically retransmit them after the network is restored, ensuring the integrity and continuity of the data.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting locomotive bearing faults based on neural networks and temporal attention mechanisms, characterized in that: Includes the following processes: Real-time acquisition of vibration signals, temperature signals, acceleration signals, and stability information of the locomotive under test; The real-time acquired vibration signals are preprocessed, including AD conversion, noise reduction, and normalization. Based on the fault characteristic frequency of the locomotive bearing, the vibration signals are divided into several time series segments of fixed length. The length of each time series segment contains at least one complete fault impact cycle. The preprocessed vibration data time series segments are fused with synchronously acquired temperature signals, acceleration signals, and stationarity information at the feature level to form a multi-dimensional input feature matrix. A multi-dimensional input feature matrix is ​​input into a convolutional neural network for feature extraction. The convolutional neural network uses multi-scale convolutional kernels to extract high-frequency impact features and low-frequency degradation features in parallel. The feature sequences extracted by the convolutional neural network are input into the long short-term memory network for time-series modeling. The gating mechanism of the long short-term memory network dynamically adjusts the forgetting gate parameters according to the locomotive speed signal to adapt to the time-series dependency changes under varying operating conditions. A temporal attention mechanism is applied to the temporal modeling results output by the Long Short-Term Memory Network to perform feature weighting, resulting in a weighted feature vector of the vibration signal. The temporal attention mechanism assigns weight values ​​to each time step to pinpoint the precise moment when the fault impact occurs. The weighted feature vector of the vibration signal is fused with the statistical features of the temperature signal and acceleration signal at the decision level to obtain a comprehensive fault judgment result, and the bearing condition of the locomotive under test is judged for multiple categories based on the result. The fault diagnosis results are transmitted simultaneously to the locomotive host and the cloud server via redundant network modules.

2. A locomotive bearing fault detection system based on neural networks and temporal attention mechanism, characterized in that: It includes a data processing CPU, sensor components, AD acquisition module, locomotive stability detection module, dual-redundant network communication module, CAN bus module, 5G network card module, and dual-redundant power supply module; The sensor assembly is used to collect vibration and temperature signals of the locomotive under test in real time. The AD acquisition module is used to convert the vibration and temperature signals acquired by the sensor components into digital signals and upload them to the data processing CPU. The locomotive stability detection module is used to detect the locomotive's acceleration signal and stability information in real time and upload them to the data processing CPU. The data processing CPU has a built-in fusion model of a convolutional neural network, a long short-term memory network, and a time-series attention mechanism, which is trained offline. The model is deployed at the edge after quantization and compression to achieve millisecond-level inference latency. The data processing CPU is used to preprocess the real-time acquired vibration signals and divide them into several fixed-length time series segments. The convolutional neural network is used to extract features from each vibration data time series segment, and the long short-term memory network is used to perform time-series modeling on the extracted feature sequences. The time-series attention mechanism is applied to the time-series modeling results to perform feature weighting to obtain a weighted feature vector of the vibration signal. Based on the weighted feature vector, the condition of the locomotive bearing under test is judged for multiple types of faults. Both the dual-redundant network communication module and the CAN bus module are used to upload the locomotive bearing fault diagnosis results obtained by the data processing CPU to the locomotive host, and the dual-redundant network communication module automatically switches to the backup link when the main link is interrupted. The 5G network card module is used to upload the diagnostic results of the data processing CPU and the locomotive status information to the cloud server in real time. The dual-power redundant module is used to supply power to the data processing CPU, sensor components, AD acquisition module, locomotive stability detection module, dual-redundant network communication module, CAN bus module, and 5G network card module.

3. The locomotive bearing fault detection system based on neural network and temporal attention mechanism according to claim 2, characterized in that: The input end of the AD acquisition module is connected to the sensor assembly, and the output end of the AD acquisition module is connected to the data processing CPU. The data processing CPU is also connected to the locomotive stability detection module, the dual-channel redundant network communication module, the CAN bus module, and the 5G network card module. The dual-channel redundant power supply module consists of two power supply modules, one of which is used as the main power supply module and the other as the backup power supply module. Both power supply modules can supply power to the above modules.

4. The locomotive bearing fault detection system based on neural network and temporal attention mechanism according to claim 2, characterized in that: The sensor assembly consists of a composite piezoelectric ceramic sensor and a temperature sensor. The composite piezoelectric ceramic sensor is installed in the vertical and horizontal directions of the bearing seat of the locomotive under test to collect dual-channel vibration signals in real time, which are then input into different channels of a convolutional neural network to extract radial and axial fault features. The temperature sensor is installed on the locomotive under test to collect environmental and component temperature signals of the locomotive under test in real time.

5. The locomotive bearing fault detection system based on neural network and temporal attention mechanism according to claim 2, characterized in that: The locomotive stability detection module is installed on the floor inside the locomotive under test. It consists of a MEMS six-axis accelerometer and a gyroscope. The MEMS six-axis accelerometer is used to detect the locomotive's acceleration signal in real time, and the gyroscope is used to detect the locomotive's stability in real time.

6. The locomotive bearing fault detection system based on neural network and temporal attention mechanism according to claim 2, characterized in that: The dual-redundant network communication module consists of two network interfaces, which are used to realize the interaction between the data processing CPU and the locomotive host, and to provide real-time feedback of the processing results to the locomotive host.

Images