Deep learning-based rail transit equipment thermal fault early warning method and system

Abstract

The invention relates to the technical field of rail transit safety monitoring, in particular to a rail transit equipment thermal fault early warning method and system based on deep learning. The core of the method is that multi-mode sensing data is formed by synchronously collecting multi-source data such as infrared temperature, visible light images and vibration of rail transit equipment; afterwards, real-time reasoning is carried out on the edge side by utilizing a mixed deep learning model, the model analyzes an image through a convolutional neural network branch to realize target recognition and infrared emissivity dynamic calibration, and conjoint analysis is carried out on calibrated temperature and vibration time sequence data through a long-short-term memory network branch to learn a change mode of the temperature and vibration time sequence data; and finally, the model generates an early warning signal by judging whether the real-time data mode deviates from the inherent normal working condition mode or not, so that early and accurate early warning independent of a fixed threshold is realized. The system correspondingly comprises a multi-source sensing module, an edge intelligent processing module and a communication early warning module. According to the invention, the accuracy, real-time performance and foresight of early warning are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of rail transit safety monitoring technology, specifically to a method and system for early warning of thermal faults in rail transit equipment based on deep learning. Background Technology

[0002] As the main artery of urban and intercity transportation, the safe operation of rail transit is of paramount importance. Under long-term high-load operation, critical mechanical components of trains, such as wheel bearings, gearboxes, and braking systems, will experience temperature rise due to friction and fatigue. Overheating is a precursor to major safety accidents such as axle shedding, melting, and even fires. Therefore, continuous and reliable temperature monitoring of these critical components is one of the core technical means to ensure the safe operation of rail transit. Currently, the industry widely uses online monitoring systems based on non-contact infrared temperature measurement. These systems typically consist of infrared temperature sensors installed beside the track or on the bottom of the train, a simple signal processing unit, and a data communication module. Their basic working principle is to collect the infrared radiation energy of the target object through infrared sensors, convert it into a temperature reading, and upload the temperature data to the back-end monitoring center via wired or wireless means. The monitoring center software presets a fixed temperature alarm threshold; when the reported temperature data exceeds this threshold, the system triggers an alarm to alert maintenance personnel.

[0003] However, this traditional technical solution has several inherent and unresolved drawbacks. First, its core alarm logic relies on a single, fixed temperature threshold, resulting in extremely low intelligence. In the complex actual rail transit operating environment, factors such as ambient temperature, weather conditions, train load, and operating speed significantly affect the surface temperature of components, leading to high false alarm and false negative rates for the fixed threshold solution. For example, in hot summer weather, the normal temperature of a component may be close to the threshold, while in winter or when operating without load, even if there are early potential problems, the temperature may be below the threshold and thus be missed. Second, the accuracy of non-contact infrared temperature measurement heavily depends on the emissivity of the surface of the object being measured. Rail transit components are made of different materials such as metal, rubber, and composite materials, with vastly different emissivity that varies due to surface oxidation and contamination. Traditional systems lack an effective emissivity calibration mechanism, resulting in significant errors in the basic temperature measurement data itself. Furthermore, existing systems have limited functionality, only performing the basic "sensing-transmission" task, placing all data analysis burden in the background, and failing to achieve real-time intelligent diagnosis and early warning at the edge. More importantly, single temperature parameters are insufficient to comprehensively reflect the health status of equipment and cannot identify early mechanical faults that have not yet caused significant temperature rises, such as early pitting of bearings or minor wear of gears. These faults often first manifest as abnormal vibration characteristics. Therefore, there is an urgent need in this field for a new type of thermal fault monitoring technology for rail transit equipment that can integrate multi-dimensional information, possess edge intelligence analysis capabilities, and achieve accurate early warning, in order to overcome the many limitations of existing technologies.

[0004] Therefore, the existing technology still needs further development. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for early warning of thermal faults in rail transit equipment based on deep learning, so as to solve the problems existing in the prior art.

[0006] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a method for early warning of thermal faults in rail transit equipment based on deep learning, comprising: S1: Collect multimodal sensing data of rail transit equipment through multi-source sensors. The multimodal sensing data includes at least infrared temperature data, visible light image data, and vibration data. S2: The multimodal sensing data is input into a deep learning model deployed at the edge for real-time inference to identify the device status; wherein, the deep learning model is used to extract the correlation features in the multimodal sensing data and determine whether to generate a warning signal based on the correlation features; S3: When the deep learning model generates a warning signal, a warning action is triggered.

[0007] Specifically, in step S1, the multi-source sensor includes an infrared temperature sensor, a visible light camera, and a vibration sensor, and the multimodal sensing data is a time-series data set composed of the infrared temperature data, visible light image data, and vibration data.

[0008] Specifically, the deep learning model in step S2 is a hybrid neural network model, which includes a convolutional neural network (CNN) branch for processing the visible light image data and a recurrent neural network (RNN) branch for processing the time-series data in the time-series data set.

[0009] Specifically, the CNN branch is configured to perform target recognition on the visible light image data and dynamically calibrate the emissivity of the infrared temperature sensor based on the identified target material information.

[0010] Specifically, the RNN branch is a Long Short-Term Memory (LSTM) network, which is configured to perform joint time-series analysis on calibrated infrared temperature data and vibration data to learn their variation patterns.

[0011] Specifically, step S2, "determining whether to generate a warning signal based on the associated features," includes: The current data change pattern obtained from the joint time series analysis is compared with the normal operating condition pattern stored in the deep learning model; when the current data change pattern deviates from the normal operating condition pattern, an early warning signal is generated, wherein the triggering of the early warning signal does not depend on a fixed temperature threshold.

[0012] Specifically, the method further includes a model update step S4: Based on historical multimodal sensing data that has been confirmed to be in a normal state, the parameters of the deep learning model on the edge side are fine-tuned to adapt to the operating characteristics of the specific monitored device.

[0013] Specifically, the model update step S4 is executed locally on the edge side and uses a federated learning framework to aggregate encrypted model parameters with the central server without uploading the original data.

[0014] According to a second aspect of the present invention, a deep learning-based early warning system for thermal faults in rail transit equipment is provided, comprising: A multi-source sensing module, which includes an infrared temperature sensor, a visible light camera, and a vibration sensor, is used to collect multimodal sensing data; The edge intelligence processing module is communicatively connected to the multi-source sensing module. It has a built-in AI computing unit for loading and running deep learning models to perform real-time inference on the multimodal sensing data. The communication early warning module is connected to the edge intelligent processing module and is used to perform early warning actions after receiving instructions from the edge intelligent processing module.

[0015] Specifically, the edge intelligence processing module is a system-on-a-chip (SoC) integrating a neural network processing unit (NPU), and the multi-source sensing module, edge intelligence processing module, and communication early warning module are integrated into a protective housing suitable for rail transit environments.

[0016] Beneficial effects: The deep learning-based thermal fault early warning method and system for rail transit equipment provided by this invention has produced significant and beneficial technical effects compared with existing technologies. The primary benefit lies in the qualitative leap in early warning capabilities achieved by introducing multimodal perception and edge-side deep learning models, moving from "threshold judgment" to "pattern recognition." The system comprehensively utilizes infrared temperature, visible light images, and vibration data, making the analysis dimensions more comprehensive. In particular, the use of visible light images to automatically identify the measured target and dynamically calibrate the infrared emissivity through a convolutional neural network fundamentally solves the industry problem of measurement inaccuracies caused by uncertain emissivity in traditional infrared thermometry, greatly improving the reliability of basic data. Furthermore, the joint analysis of temperature and vibration time-series data using a long short-term memory network can keenly capture subtle, cross-modal abnormal correlation patterns between the two, thus issuing early warnings when faults are still in their nascent stage and have not yet caused a drastic temperature rise, achieving true early fault prediction and significantly advancing the accident prevention process.

[0017] Secondly, this invention, through its edge intelligent computing architecture, delivers a dual improvement in real-time response and system reliability. Deploying complex deep learning models on edge terminals allows data acquisition, analysis, and decision-making to be completed instantly locally, eliminating the network latency and bandwidth pressure associated with uploading data to the cloud. This meets the stringent millisecond-level real-time requirements of rail transit safety monitoring. Even in extreme cases where the network connection to the backend monitoring center is interrupted, the edge terminal can still operate independently and execute local audible and visual alarms, greatly enhancing the robustness and survivability of the entire monitoring system. Furthermore, this architecture reduces reliance on backend system computation, thus helping to lower overall operation and maintenance costs.

[0018] Finally, by introducing online learning and federated learning mechanisms, this invention endows the system with unique adaptive and collaborative evolutionary capabilities. The model can be fine-tuned locally using normal data verified on-site, continuously adapting to the personalized operating characteristics of specific monitored devices, thus becoming more accurate with use. Through the federated learning framework, multiple terminals can jointly optimize a global model without sharing original sensitive data, achieving knowledge sharing and collaborative progress under data privacy protection, giving the entire monitoring network a continuously optimizing vitality. In summary, this invention not only achieves breakthroughs in monitoring accuracy and foresight but also brings comprehensive and significant benefits in terms of system intelligence, real-time performance, reliability, and adaptability. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the deep learning-based early warning method for thermal faults in rail transit equipment provided in a specific embodiment of the present invention. Figure 2 This is a schematic diagram of the system composition of a deep learning-based early warning system for thermal faults in rail transit equipment, provided in a specific embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0021] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0022] Please see Figure 1 This invention provides a deep learning-based method for early warning of thermal faults in rail transit equipment, comprising: S1: Collect multimodal sensing data of rail transit equipment through multi-source sensors. The multimodal sensing data includes at least infrared temperature data, visible light image data, and vibration data. It should be further explained that, in this embodiment, the hardware foundation of the method is an edge computing terminal integrating multiple source sensors. This terminal is installed beside the track or under the train, directly facing the critical equipment to be monitored, such as wheel bearings. The specific implementation process of step S1 is as follows: After the system is powered on, the main controller synchronously triggers the infrared temperature sensor, visible light camera, and vibration sensor to collect data. The infrared temperature sensor preferably uses an MLX90640 infrared thermal imager chip with a resolution of 32x24 pixels and a frame rate of 10Hz (i.e., acquiring 10 frames of temperature matrix data per second). This frame rate selection is to balance the real-time nature of the data with the processing power of the edge computing unit, capturing rapid temperature changes without placing an excessive burden on the processor. The visible light camera preferably uses an OV2640 2-megapixel CMOS sensor, acquiring RGB images at a rate of 10 frames per second, with an image resolution of 800x600 pixels. This resolution is sufficient to clearly distinguish the surface details of components such as bearings, while also providing a moderate amount of data. The vibration sensor preferably uses the ADXL345 triaxial digital accelerometer, with a range set to ±16g and an output data rate (ODR) configured to 1000Hz to meet the requirements for capturing high-frequency impact vibrations. To ensure data synchronization, the main controller reads data from all sensors every 100 milliseconds (i.e., 10Hz) and packages the infrared temperature matrix (32x24), visible light image (800x600), and 100 vibration data points (100 points per axis, 300 points in total across all three axes) acquired at that moment into a single data packet, forming a synchronized multimodal data sample.

[0023] Step S2: Input the multimodal sensing data into a deep learning model deployed at the edge for real-time inference to identify the device status; wherein, the deep learning model is used to extract the correlation features in the multimodal sensing data and determine whether to generate a warning signal based on the correlation features; It should be further explained that step S2 is the core intelligent analysis stage. A deep learning model deployed in an edge AI computing unit (such as the Huawei Atlas 200DK acceleration module) continuously receives the aforementioned data packets. The model first preprocesses the data and then performs inference. The model's output is a risk score between 0 and 1. The closer the risk score is to 1, the higher the risk of abnormal device status. A warning threshold of 0.85 is set. When the risk score output by the model is greater than or equal to 0.85, a warning signal is generated. The choice of 0.85 as the threshold is based on a trade-off between the model's performance on the validation set: a threshold that is too low will lead to more false alarms, while a threshold that is too high will miss real early faults. 0.85 can ensure a high detection rate while keeping the false alarm rate within an acceptable range (e.g., <5%).

[0024] Step S3: When the deep learning model generates a warning signal, a warning action is triggered; It should be further explained that the warning action in step S3 is tiered. When a warning signal is generated, the edge terminal immediately sends an alarm message to the host monitoring center via its RS485 interface, containing the device ID, timestamp, risk score, and main abnormal characteristics (such as high temperature or abnormal vibration spectrum). Simultaneously, the terminal's local red LED indicator flashes three times per second, and the buzzer sounds continuously, achieving on-site audible and visual alarm. This tiered response mechanism ensures the reliable transmission of warning information.

[0025] Specifically, in step S1, the multi-source sensor includes an infrared temperature sensor, a visible light camera, and a vibration sensor, and the multimodal sensing data is a time-series data set composed of the infrared temperature data, visible light image data, and vibration data.

[0026] It should be further explained that, in this embodiment, the time-series data set is constructed as follows: the system maintains a circular buffer of size 30 in memory to store data within the last 3 seconds (10 frames / second because the sampling rate is 10Hz). A continuous data packet (3 seconds = 30 frames). Therefore, the time-series data set is actually a tensor of dimension (30,...), where 30 is the time step. For infrared data, the temperature matrix in each data packet is flattened into a tensor of length 768 (32...). The infrared data from 30 consecutive time steps (24=768) forms a 30x768 matrix. Visible light images undergo scaling and cropping preprocessing. The center of the 800x600 image is cropped to a 224x224 region and scaled to 64x64 pixels to reduce computation. It is then flattened into a 4096-dimensional vector, and the 30 frames form a 30x4096 matrix. Vibration data undergoes Fast Fourier Transform (FFT) on 100 points along each axis to extract the top 10 frequency components with the largest amplitudes and their magnitudes, forming 20 features per axis (10 frequency points + 10 amplitude points), for a total of 60 features across the three axes. These 30 time steps form a 30x60 matrix. These three matrices are concatenated along the channel dimension and used as the input to the model. The 3-second time window was chosen because the fault characteristics of rail transit bearings (such as the generation of hot spots and changes in vibration modes) can usually be seen within seconds. This window can effectively capture the dynamic process without introducing too much delay.

[0027] Specifically, the deep learning model in step S2 is a hybrid neural network model, which includes a convolutional neural network (CNN) branch for processing the visible light image data and a recurrent neural network (RNN) branch for processing the time-series data in the time-series data set.

[0028] It should be further explained that, in this embodiment, the specific architecture of the hybrid neural network model is as follows. The model adopts a two-branch structure, and finally performs feature fusion and decision-making.

[0029] 1. CNN Branch: The input is a single 64x64x3 RGB image. The first layer is a convolutional layer using 32 7x7 convolutional kernels with a stride of 2 and padding of 3, followed by a ReLU activation function and a 2x2 max pooling layer. The second convolutional layer uses 64 5x5 convolutional kernels with a stride of 1, followed by ReLU and pooling. The third convolutional layer uses 128 3x3 convolutional kernels. The feature map is then flattened and connected to a fully connected layer with 256 neurons, outputting a 256-dimensional image feature vector. .

[0030] 2. RNN Branch: The input consists of infrared vectors (30x768) and vibration feature vectors (30x60) for 30 consecutive time steps. First, these are concatenated into a 30x828 matrix along the feature dimension. This matrix is ​​then fed into a two-layer bidirectional LSTM network. The hidden state dimension of each LSTM layer is set to 128, therefore the final output dimension of the bidirectional LSTM layer is 30x256 (because bidirectional is 2). 128 = 256). The output of the last time step (dimension 256) is taken as the temporal feature vector. .

[0031] 3. Feature fusion and classification: This involves fusing and classifying the features output by the CNN branches. and the output of the RNN branch The features are concatenated to obtain a 512-dimensional fused feature vector. This vector is then input into a three-layer fully connected classification network: the first layer has a 512-dimensional input and a 128-dimensional output, followed by ReLU and Dropout (dropout rate 0.5); the second layer has a 128-dimensional input and a 64-dimensional output, followed by ReLU; and the third layer has a 64-dimensional input and a 1-dimensional output, followed by a Sigmoid activation function to obtain the final risk score.

[0032] Specifically, the CNN branch is configured to perform target recognition on the visible light image data and dynamically calibrate the emissivity of the infrared temperature sensor based on the identified target material information.

[0033] It should be further explained that, in this embodiment, the specific steps for the CNN branch to implement dynamic reactivity calibration are as follows: a. Object Detection and Material Classification: In the CNN branch, we actually embed a lightweight semantic segmentation model (such as a simplified version of U-Net). The output of its last convolutional layer is not a feature vector, but a 64x64 semantic segmentation map, where each pixel is classified as "background," "metal bearing," "rubber seal," etc. By calculating the area ratio of the main material category in the segmentation map, the dominant material in the current image is determined. ; b. Emissivity Mapping: The system internally stores an emissivity lookup table (LUT), for example: metal bearing: 0.90, rubber seal: 0.94, composite material: 0.88; c. Temperature calibration: Raw temperature values ​​read from the infrared sensor. This is calculated based on a default emissivity (e.g., 0.95). The actual emissivity needs to be determined based on the identified material emissivity. and default emission rate A correction is performed. The correction formula is based on a simplified approximation of Planck's radiation law, and is as follows: in, It is a calibrated estimate of the true temperature, which physically means a temperature that is closer to the actual surface temperature of the object. It is the temperature value read directly from the infrared sensor. This is the default emissivity set by the factory for infrared sensors, typically 0.95. This is an estimate of the true emissivity of the target object's material obtained through image recognition. The fourth root in the formula is used because radiant energy is proportional to the fourth power of temperature. This calibration step can effectively reduce the temperature measurement error from ±5℃ to within ±2℃.

[0034] Specifically, the RNN branch is a Long Short-Term Memory (LSTM) network, which is configured to perform joint time-series analysis on calibrated infrared temperature data and vibration data to learn their variation patterns.

[0035] It should be further explained that, in this embodiment, the training process of the LSTM network for joint temporal analysis is as follows: First, training data needs to be prepared. A large number of temporal data samples covering various states such as normal, slightly abnormal, and severely abnormal are collected from the historical database, with each sample lasting 3 seconds (30 frames). Each sample is labeled by experts, with normal samples labeled 0 and faulty samples labeled 1. The goal of model training is to enable the network to learn to map the input temporal patterns to the corresponding risk labels. The following hyperparameter settings are used during training: the optimizer uses Adam, with its initial learning rate set to 0.001, which is a commonly used value that has been verified by a large number of experiments to achieve stable convergence. The batch size is set to 32 to achieve a balance between training speed and memory usage. The number of training epochs is set to 100, and early stopping is used. Training stops when the validation set loss no longer decreases for 10 consecutive epochs to prevent overfitting. The loss function used is binary cross-entropy. During training, the forget gate, input gate, and output gate in the LSTM network work together to learn which historical information needs to be retained, which needs to be forgotten, and how to combine the current input with the historical state. In this way, the network can automatically learn, for example, cross-modal, time-related anomalous patterns such as "while the 200Hz component in the vibration spectrum continuously increases, the outer ring temperature of the bearing rises by more than 2°C within 3 seconds," and its risk score should be significantly improved.

[0036] Specifically, step S2, "determining whether to generate a warning signal based on the associated features," includes: comparing the current data change pattern obtained from the joint time series analysis with the normal operating condition pattern stored in the deep learning model; when the current data change pattern deviates from the normal operating condition pattern, a warning signal is generated, wherein the triggering of the warning signal does not depend on a fixed temperature threshold.

[0037] It should be further explained that in this embodiment, the "comparison" is achieved by calculating the reconstruction error. Specifically, based on the above classification model, we additionally train an LSTM autoencoder to specifically learn the data patterns under normal operating conditions.

[0038] 1. Autoencoder training: An LSTM autoencoder is trained using only a large amount of normal-state time-series data (data with label 0). The goal of this autoencoder is to reconstruct the normal input sequence as accurately as possible. The encoder compresses the input sequence (e.g., 30x828) into a low-dimensional latent vector, and the decoder then reconstructs the original sequence from this latent vector.

[0039] 2. Reconstruction error calculation: During the inference phase, for each new real-time sequence... The pre-trained autoencoder is used to reconstruct the result. Then, the root mean square error (RMSE) of both is calculated as the reconstruction error. The calculation formula is as follows: in, It is the total number of all feature points in the sequence (for a 30x828 sequence, N=30). 828 = 24840). It is the value of the i-th feature point in the real-time sequence. It is the value of the i-th feature point in the sequence reconstructed by the autoencoder. The physical meaning of is the degree of deviation between real-time data and the normal mode baseline; the larger the value, the more serious the deviation.

[0040] 3. Dynamic threshold early warning: We set a dynamic reconstruction error threshold. This threshold is not a fixed value, but rather is calculated during system initialization by collecting the reconstruction errors of all samples during a period of normal operation (e.g., 24 hours), and then adding three times the standard deviation to the average of these errors as the initial threshold. When the real-time calculated... Greater than When a situation is identified as a serious deviation from the normal pattern, a high-risk score from the classification model is used to trigger an early warning. This method also has the capability to detect unknown types of faults.

[0041] Specifically, the method further includes a model update step S4: based on historical multimodal sensing data that has been confirmed to be in a normal state, the parameters of the deep learning model on the edge side are fine-tuned to adapt to the operating characteristics of the specific monitored device.

[0042] It should be further explained that, in this embodiment, the specific online learning process of step S4 is as follows: The system sets up a buffer with a capacity of 1000 samples to store "normal" data samples that have been manually confirmed by maintenance personnel on the host computer software. When the buffer is full, the system will start the fine-tuning process during network idle periods (such as 2:00 AM to 4:00 AM). Fine-tuning adopts a transfer learning strategy: keeping the parameters of the backbone feature extraction network of the model (i.e., the first few layers of the CNN and the LSTM network) unchanged (frozen), and only retraining the last few fully connected layers used for classification. The learning rate for fine-tuning is set to a smaller value, such as 0.0001, using mini-batch gradient descent with a batch size of 16, and training for 2 epochs is sufficient. After fine-tuning, the updated model is validated using the data in the buffer. If the accuracy improves and the false positive rate decreases, the old model is replaced with the new model; otherwise, it is rolled back. This mechanism enables the system to adapt to the unique "personality" of specific devices, such as the inherent vibration characteristics caused by minor installation deviations.

[0043] Specifically, the model update step S4 is executed locally on the edge side and uses a federated learning framework to aggregate encrypted model parameters with the central server without uploading the original data.

[0044] It should be further explained that, in this embodiment, the implementation details of federated learning are as follows: After each edge terminal completes the model fine-tuning locally, it will update the model parameter increments (i.e., weight changes) generated by this update. Encryption is performed using the AES-256 algorithm. The encrypted result... Uploaded to a central server. After the server collects uploads from at least 10 endpoints, it performs a federated averaging (Fed Avg) algorithm: in, These are the updated global model parameters. These are the global model parameters before the update. This refers to the number of terminals participating in this round of aggregation. This is the increment of model parameters obtained by encryption and decryption from the k-th terminal. After aggregation, the server will generate the new global model. The encrypted data is distributed to all participating terminals. Each terminal receives the data and updates its local model. This process fully protects the security and privacy of sensitive data at each site.

[0045] Please see Figure 2 The present invention provides another embodiment, which provides a deep learning-based early warning system for thermal faults in rail transit equipment. The deep learning-based early warning system for thermal faults in rail transit equipment includes: The multi-source sensing module 100 includes an infrared temperature sensor, a visible light camera, and a vibration sensor, used to collect multimodal sensing data; The edge intelligence processing module 200 is communicatively connected to the multi-source sensing module. It has a built-in AI computing unit for loading and running deep learning models to perform real-time inference on the multimodal sensing data. The communication early warning module 300 is communicatively connected to the edge intelligent processing module and is used to execute early warning actions after receiving instructions from the edge intelligent processing module.

[0046] It should be further explained that, in this embodiment, the hardware configuration of the system is as follows: all modules are integrated into a cast aluminum housing that meets the IP67 protection standard. The sensors in the multi-source sensing module 100 are connected to the core board via FPC cables. The core of the edge intelligent processing module 200 is a Rockchip RK3568 chip, which integrates four ARM Cortex-A55 cores and an NPU (Neural Processing Unit) with a computing power of 0.8 TOPS, sufficient to smoothly run the hybrid deep learning model. The module is equipped with 4GB LPDDR4 memory and 32GB eMMC storage. The communication warning module 300 includes an RS485 interface chip (model MAX3485), a Wi-Fi / Bluetooth dual-mode module (model AP6255) for wireless debugging and data transmission, a high-brightness tri-color LED (red / yellow / green), and a passive piezoelectric buzzer. The system is powered by an industrial-grade 24VDC power supply line, which is converted to 5V and 3.3V by an internal power management chip for use by each module. The mounting brackets of the housing are designed with shock-absorbing rubber pads to cope with the strong vibrations in the rail transit environment.

[0047] Specifically, the edge intelligent processing module 200 is a system-on-a-chip (SoC) integrating a neural network processing unit (NPU), and the multi-source sensing module 100, the edge intelligent processing module 200, and the communication early warning module 300 are integrated into a protective housing suitable for rail transit environments.

[0048] Further explanation is needed regarding the key details of hardware integration in this embodiment. The NPU of the SoC (RK3568) supports INT8 quantization inference, which significantly improves model running speed and reduces power consumption. Before deploying the model, the trained FP32 model needs to be quantized into an INT8 model using the manufacturer's conversion tool. This process introduces a slight loss of accuracy, but the inference speed can be increased by 3-5 times, fully meeting real-time requirements. The protective housing is made of 6061 aluminum alloy with an anodized surface for corrosion resistance. A silicone sealing ring is used between the housing and the cover to ensure dust and water resistance. All external interfaces (power, RS485) use waterproof aviation connectors. Thermally conductive silicone pads are attached to the inner wall of the housing to conduct the main heat generated by the SoC to the metal housing for heat dissipation. The entire system is designed to operate within a temperature range of -40℃ to +85℃, fully meeting the outdoor environmental requirements of most rail transit lines in China.

[0049] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the deep learning-based thermal fault early warning method for rail transit equipment. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0050] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0051] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0052] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0053] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for early warning of thermal faults in rail transit equipment based on deep learning, characterized in that, Includes the following steps: S1: Collect multimodal sensing data of rail transit equipment through multi-source sensors. The multimodal sensing data includes at least infrared temperature data, visible light image data, and vibration data. S2: The multimodal sensing data is input into a deep learning model deployed at the edge for real-time inference to identify the device status; wherein, the deep learning model is used to extract the correlation features in the multimodal sensing data and determine whether to generate a warning signal based on the correlation features; S3: When the deep learning model generates a warning signal, a warning action is triggered.

2. The method according to claim 1, characterized in that, In step S1, the multi-source sensor includes an infrared temperature sensor, a visible light camera, and a vibration sensor, and the multimodal sensing data is a time-series data set composed of the infrared temperature data, visible light image data, and vibration data.

3. The method according to claim 2, characterized in that, The deep learning model in step S2 is a hybrid neural network model, which includes a convolutional neural network (CNN) branch for processing the visible light image data and a recurrent neural network (RNN) branch for processing the time-series data in the time-series data set.

4. The method according to claim 3, characterized in that, The CNN branch is configured to perform target recognition on the visible light image data and dynamically calibrate the emissivity of the infrared temperature sensor based on the identified target material information.

5. The method according to claim 4, characterized in that, The RNN branch is a Long Short-Term Memory (LSTM) network, which is configured to perform joint time-series analysis on calibrated infrared temperature data and vibration data to learn their variation patterns.

6. The method according to claim 5, characterized in that, Step S2, "determining whether to generate a warning signal based on the associated features," specifically includes: The current data change pattern obtained from the joint time series analysis is compared with the normal operating condition pattern stored in the deep learning model; when the current data change pattern deviates from the normal operating condition pattern, an early warning signal is generated, wherein the triggering of the early warning signal does not depend on a fixed temperature threshold.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes a model update step S4: Based on historical multimodal sensing data that has been confirmed to be in a normal state, the parameters of the deep learning model on the edge side are fine-tuned to adapt to the operating characteristics of the specific monitored device.

8. The method according to claim 7, characterized in that, The model update step S4 is executed locally on the edge side and uses a federated learning framework to aggregate encrypted model parameters with the central server without uploading the original data.

9. A deep learning-based early warning system for thermal faults in rail transit equipment, used to implement the method described in any one of claims 1 to 8, characterized in that, The system includes: A multi-source sensing module, which includes an infrared temperature sensor, a visible light camera, and a vibration sensor, is used to collect multimodal sensing data; The edge intelligence processing module is communicatively connected to the multi-source sensing module. It has a built-in AI computing unit for loading and running deep learning models to perform real-time inference on the multimodal sensing data. The communication early warning module is connected to the edge intelligent processing module and is used to perform early warning actions after receiving instructions from the edge intelligent processing module.

10. The system according to claim 9, characterized in that, The edge intelligent processing module is a system-on-a-chip (SoC) with an integrated neural network processing unit (NPU). The multi-source sensing module, the edge intelligent processing module, and the communication early warning module are integrated into a protective housing suitable for rail transit environments.

Images