Rail transit fault analysis system and method based on image recognition

Abstract

The invention discloses a rail transit fault analysis system and method based on image recognition, and the system comprises a data collection and processing module which is used for collecting and preprocessing an original image and vibration data of rail transit equipment; the data storage module is used for storing the original image and vibration data, the preprocessed image and vibration data and historical detection records; extracting features of defect images and vibration data, and establishing a defect feature database; and the algorithm analysis library is used for constructing a deep learning network model capable of identifying defects by utilizing the defect feature database, carrying out fault identification on the preprocessed image and vibration data, storing an identification result, and simultaneously pushing a high-priority defect alarm to a maintenance work order system. According to the invention, defect detection is carried out on the image and vibration data of the rail transit equipment through image recognition and a deep learning algorithm, and intelligent analysis and efficient maintenance of faults are realized in combination with a data storage and alarm pushing mechanism.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of rail transit fault identification, and particularly relates to a rail transit fault analysis system and method based on image identification. BACKGROUND

[0002] In recent years, electrified railways, especially high-speed railways, have developed rapidly, and the operating mileage has continued to expand. The workload of the overhead contact line maintenance unit is increasing, and the difficulty of equipment maintenance and management has also increased significantly. The maintenance standards and hidden defect processing requirements of high-speed railways are much higher than those of ordinary speed railways. In the past, for the hidden faults of the overhead contact line, periodic manual inspection and portable overhead contact line inspection devices were usually used for detection, but this method has been difficult to meet the actual needs of the operation of overhead contact line equipment of high-speed railways.

[0003] Manual inspection is usually carried out at night during the window period, and the detection personnel need to work continuously for 4 hours. The detection effect is affected by many human factors, such as the fatigue state and working meticulousness of the personnel. According to the introduction of the contact line maintenance personnel on the front line, a group of detection workers can maintain about 100 contact line support in a night window period. For hundreds of kilometers of power supply lines, such operation efficiency is obviously low. In addition, manual inspection also has a high safety risk.

[0004] Although the portable overhead contact line inspection equipment has higher efficiency than manual inspection, it still has many limitations. When a single 5 million pixel camera is used to shoot overhead contact line equipment, the resolution is low, and only obvious defects such as bird nests and foreign matters on the overhead contact line frame can be identified. Some inspection devices even cannot clearly see the dropper. In addition, the imaging effect of the inspection equipment in areas such as bridges, tunnels and stations is also limited by the light conditions. In the tunnel, it is difficult to image because a fill light cannot be carried; in the station and bridge interior, it is also difficult to obtain clear images due to the sharp light contrast and small dynamic range. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a rail transit fault analysis system and method based on image identification, which detects defects in the images and vibration data of rail transit equipment through image recognition and deep learning algorithms, and realizes intelligent analysis and efficient maintenance of faults by combining data storage and alarm pushing mechanisms.

[0006] To achieve the above purpose, the present application provides the following technical solutions:

[0007] A rail transit fault analysis system based on image identification, comprising:

[0008] a data acquisition and processing module, configured to acquire original images and vibration data of rail transit equipment and perform preprocessing;

[0009] a data storage module, configured to store original image and vibration data, preprocessed image and vibration data, and historical detection records, extract features of image and vibration data of defects based on the historical detection records, and establish a defect feature database;

[0010] a fault analysis module, internally provided with an algorithm analysis library, configured to train a deep learning network model by using the defect feature database, and obtain the deep learning network model capable of identifying defects of the rail transit equipment;

[0011] the deep learning network model capable of identifying defects of the rail transit equipment is used to perform fault identification on the preprocessed image and vibration data, and the result of the fault identification is stored in the data storage module and a high-priority fault alarm is pushed to a maintenance work order system.

[0012] Further, the data collection and processing module comprises:

[0013] a data collection module, configured to collect original image and vibration data of the rail transit equipment;

[0014] a preprocessing module, configured to clean, format convert, and time sequence synchronize the original image and vibration data, and obtain preprocessed image and vibration data;

[0015] a cache and distribution module, configured to divide the preprocessed image and vibration data into real-time paths and archive paths; the real-time paths are used to be pushed to Kafka topics for real-time consumption by the algorithm analysis library; and the archive paths are stored in the data storage module for offline analysis and long-term storage.

[0016] Further, the original image of the rail transit equipment comprises foreign matters on the track and deformation conditions of train components.

[0017] Further, the preprocessing process comprises grayscale, noise reduction, or normalization processing.

[0018] Further, the data storage module comprises a relational table, a space-time database, a file storage, an index service, and an API gateway module.

[0019] Further, the algorithm analysis library comprises:

[0020] an algorithm container, configured to encapsulate deep learning algorithms by using a Docker container, so as to realize running environment isolation and rapid deployment;

[0021] a task scheduling module, configured to dynamically allocate CPU / GPU resources, so as to ensure efficient running of the algorithm container;

[0022] The intermediate result caching module is used to cache the surface feature vectors of rail transit equipment using the in-memory database Redis, in order to accelerate the inference process of subsequent models.

[0023] Furthermore, a deep learning network model capable of identifying defects in rail transit equipment is used to perform fault identification on the preprocessed images and vibration data, including:

[0024] Using a YOLOv7 defect detection model based on convolutional neural networks, defect features in images are identified;

[0025] We use an LSTM vibration analysis model based on recurrent neural networks to process and analyze time series data of vibrations from rail transit equipment.

[0026] Furthermore, the results of the fault identification include: the defect type, confidence level, and location coordinates of key components in the rail transit equipment; wherein key components include: insulators, poles, pole numbers, locator electrical connections, U-shaped clamps, bases, wire clamps, and screws.

[0027] The present invention also provides a method for applying an image recognition-based rail transit fault analysis system, comprising the following steps:

[0028] S1. Acquire raw images and vibration data of rail transit equipment and perform preprocessing;

[0029] S2. Store the original images and vibration data, the preprocessed images and vibration data, and historical inspection records; and based on the historical inspection records, compare the images and vibration data with defects with the normal images and vibration data to establish a defect database.

[0030] S3. Use the defect database to train the deep learning network model to obtain a deep learning network model that can identify defects in rail transit equipment.

[0031] A deep learning network model capable of identifying defects in rail transit equipment is used to perform fault identification on preprocessed images and vibration data, and the identification results are stored in the data storage module. At the same time, a high-priority defect alarm is pushed to the maintenance work order system.

[0032] According to the specific embodiments provided by the present application, the following technical effects are disclosed: the present application reduces the workload and time cost of manual detection and improves the detection efficiency by automatically collecting and preprocessing the original image and vibration data; the potential faults of the equipment can be more accurately detected by using the deep learning network model to identify the preprocessed image and vibration data, reducing the possibility of false detection and missed detection; the historical detection records can be stored and managed through the establishment of the defect database, facilitating subsequent analysis and decision-making. At the same time, the system can push high-priority defect alarms to the maintenance work order system, realizing timely processing and maintenance of faults; by discovering and processing equipment faults in a timely manner, safety accidents caused by equipment faults can be effectively avoided, improving the safety of rail transit; by analyzing historical data, the rules and trends of equipment faults can be found, providing decision support for equipment maintenance and updating. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0034] The image recognition-based rail transit fault analysis system and method of the present application will be further described below in combination with the drawings.

[0035] Figure 1 is the system structure diagram of the rail transit fault analysis provided by the present application;

[0036] Figure 2 is the system architecture diagram of the rail transit fault analysis provided by the present application;

[0037] Figure 3 The detection demonstration diagram of the rail transit fault analysis system provided by the present application. DETAILED DESCRIPTION

[0038] The specific embodiments of the present application will be further described in detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0039] In order to better understand the purpose, structure and function of the present application, the present application will be further described in detail below in combination with the drawings.

[0040] Embodiment 1

[0041] As shown in Figure 1 , the present application provides an image recognition-based rail transit fault analysis system, which comprises:

[0042] a data collection and processing module configured to collect and preprocess raw images and vibration data of the rail transit equipment, wherein the raw images of the rail transit equipment include foreign matters on the track and deformation conditions of train components.

[0043] In this embodiment, the data collection process is specifically as follows: multiple high-definition cameras are installed on the subway platforms and along the track, for example, at both ends of each platform and at key positions of the track, such as curves and switches. These cameras need to have high resolution and support multiple protocol access (HTTP / MQTT / Kafka) such as Figure 2 as shown in the figure, which can clearly capture details of the track and train components. For example, an infrared camera that can take clear images in a relatively dark tunnel can monitor the track conditions for 24 hours without interruption and provide high-quality raw data for image recognition.

[0044] The data preprocessing process is specifically as follows: edge computing devices such as NVIDIA Jetson series modules are deployed near the cameras. These devices can preliminarily process and analyze the images collected by the cameras locally, reducing the computing pressure on the cloud server and the data transmission delay. For example, it can detect whether there are obvious abnormal objects or areas in the image in real time, such as foreign matters on the track and deformation of train components, and quickly upload the images that may have faults to the cloud for further analysis.

[0045] The preprocessing process of the raw images and vibration data collected by the rail transit equipment includes grayscale, noise reduction, or normalization processing.

[0046] In this embodiment, the parameters of the cameras, such as focal length, aperture, and exposure time, are set to ensure the best image quality. At the same time, the collected images are preprocessed, including grayscale, noise reduction, and normalization operations. For example, in an environment with large changes in light, the camera parameters are automatically adjusted to keep the image clear and stable, and noise reduction algorithms are used to remove noise points in the image, improving the clarity and recognizability of the image.

[0047] The data collection and processing module includes:

[0048] a data collection module configured to collect raw images and vibration data of the rail transit equipment;

[0049] a preprocessing module configured to clean, format convert (such as RAW image to standard PNG / JSON), and time synchronize the raw images and vibration data, to obtain preprocessed images and vibration data;

[0050] The cache and distribution module is used for temporarily storing high-frequency data through a distributed cache (such as Redis) after preprocessing of the image and vibration data, and is divided into a real-time path and an archive path; the real-time path is used for pushing to a Kafka topic for real-time consumption by an algorithm analysis library; and the archive path is stored in a data storage module for offline analysis and long-term storage.

[0051] The data storage module is an application layer, and is used for storing original image and vibration data, preprocessed image and vibration data, and historical detection records; based on the historical detection records, a large number of picture samples are analyzed, normal and defect samples are compared, a defect feature database is established; and a message queue (such as RabbitMQ / Kafka) is used to push a preprocessed image and vibration data packet to the algorithm analysis library to trigger an algorithm analysis task.

[0052] The data storage module includes a relational table, a space-time database, a file storage, an index service, and an API gateway module.

[0053] In this embodiment, the relational table is used for storing defect attributes (type, position, severity level); the space-time database is used for recording GPS coordinates and detection time of the defect, and supports geographic fence query; the file storage is used for storing a storage path of associated original data (such as image slices and vibration waveforms); the index service is used for establishing a B+ tree or an inverted index for high-frequency query fields (such as line numbers and defect types); and the API gateway is used for providing a query interface (such as GraphQL) to an operation and maintenance system or a visualization platform.

[0054] The fault analysis module is built-in with an algorithm analysis library, is connected with the data storage module, subscribes to tasks in a message queue, pulls data packets and starts a processing pipeline, is used for training a deep learning network model by using the defect feature database, and obtains a deep learning network model capable of identifying defects of rail transit equipment.

[0055] The deep learning network model capable of identifying defects of rail transit equipment is used for fault identification on preprocessed image and vibration data, and the result of fault identification is stored in the data storage module, and a high-priority fault alarm is pushed to a repair work order system.

[0056] In this embodiment, a deep learning algorithm such as a convolutional neural network (CNN) is used. A large amount of rail transit fault image data is used to train the model in advance, so that the model learns the characteristics of different fault types. For example, for the wear and tear fault of the train wheel, the model can accurately identify whether the wheel is worn and the degree of wear by learning the image features of the worn part. In actual application, the real-time collected image is input into the trained model, and the model can quickly output the fault detection result. The detection result (such as "wave abrasion defect, confidence 92%") is encapsulated as JSON and written into the data storage module through the gRPC interface.

[0057] According to the fault detection result, the running data and historical fault records of rail transit are combined to comprehensively diagnose and analyze the fault. If a foreign object is detected on the track, the software will immediately determine the size and position of the foreign object and the degree of influence on train operation, and timely issue a warning information to inform the relevant staff to take appropriate measures, such as suspending train operation, organizing personnel to clean up foreign objects, etc. At the same time, the relevant information of the fault is recorded to provide data support for subsequent fault analysis and maintenance.

[0058] The pre-processed image and vibration data are subjected to fault recognition using a deep learning network model capable of identifying rail transit equipment defects, including:

[0059] A YOLOv7 defect detection model based on convolutional neural network is used to identify the defect features in the image.

[0060] An LSTM vibration analysis model based on recurrent neural network is used to process and analyze the time series data of rail transit equipment vibration.

[0061] The algorithm analysis library includes:

[0062] An algorithm container is used to encapsulate deep learning algorithms through a Docker container to realize running environment isolation and rapid deployment.

[0063] A task scheduling module is used to dynamically allocate CPU / GPU resources to ensure efficient operation of the algorithm container.

[0064] An intermediate result caching module is used to cache the surface feature vectors of rail transit equipment using a memory database Redis to speed up the inference process of subsequent models.

[0065] The training process of the YOLOv7 defect detection model in this embodiment is as follows:

[0066] I. Data preparation and preprocessing

[0067] (1) Data Collection: High-definition industrial cameras are used to collect image data of rail transit equipment (such as tracks, wheels, and overhead lines), covering different lighting conditions, equipment poses, and defect types (cracks, wear, deformation, etc.).

[0068] (2) Data Annotation:

[0069] Labeling is done using Labelme or CVAT tools, generating YOLO format label files (.txt), each line containing class ID, normalized center point coordinates, and width and height. Example label format: 0 0.50.50.20.3 (class 0, center point (0.5, 0.5), width and height ratio 0.2 and 0.3).

[0070] (3) Data Augmentation:

[0071] Mosaic augmentation: randomly crop and paste 4 images to enrich small target context.

[0072] Geometric transformation: random rotation (-15° to 15°), translation (±10%), and scaling (0.8 to 1.2 times).

[0073] Color disturbance: adjust brightness, contrast, and saturation (±20%) to simulate complex lighting environments.

[0074] II. Model Selection and Configuration

[0075] Network structure:

[0076] Backbone: ELAN-W (Extended Efficient Layer Aggregation Network-Wide) is used to extract multi-scale features through multi-branch convolution.

[0077] Neck: PAFPN (Path Aggregation Feature Pyramid Network) is used to fuse shallow positioning information and deep semantic information.

[0078] Head: Decoupled Head is introduced to separate classification and regression tasks, improving detection accuracy.

[0079] Configuration file adjustment:

[0080] Modify the anchors parameter in yolov7.yaml to adjust the prior box according to the defect size (e.g., use [[10, 13], [16, 30], [33, 23]] for small targets).

[0081] The depth_multiple and width_multiple are set to control the depth and width of the network, balancing speed and accuracy.

[0082] III. Training Process

[0083] (1) Hyperparameter Settings:

[0084] Optimizer: AdamW with weight decay 0.01.

[0085] Learning Rate: Initial learning rate 0.001, decay to 0.0001 using cosine annealing.

[0086] Batch Size: Adjusted according to GPU memory (e.g., 64 for 8 RTX 3090s).

[0087] (2) Loss Function:

[0088] Classification Loss: BCEWithLogitsLoss, Regression Loss: CIoU Loss, balancing position and shape errors.

[0089] (3) Training Strategy:

[0090] Warm-up Training: Use a small learning rate (0.0001) for the first 500 iterations to stabilize the model.

[0091] Mixed Precision Training: Use FP16 to accelerate training and reduce memory usage.

[0092] IV. Evaluation and Optimization

[0093] (1) Evaluation Indicators:

[0094] mAP@0.5: Average precision at IoU threshold 0.5.

[0095] Recall@0.5: Recall rate at IoU threshold 0.5.

[0096] (2) Optimization Strategy:

[0097] Hard Example Mining: Re-label missed samples and add them to the training set.

[0098] Model Distillation: Use YOLOv7-X as the teacher model, distill to YOLOv7-tiny student model, and improve inference speed.

[0099] V. Deployment and Application

[0100] Model Export: Convert to ONNX or TensorRT format, inference speed improved by 2-3 times.

[0101] Real-time detection: integrated into edge computing devices (such as Jetson AGX Xavier), achieving 1080P image 30FPS real-time detection.

[0102] The process of training the LSTM vibration analysis model in this embodiment is as follows:

[0103] I. Data preparation and preprocessing

[0104] (1) Data collection: collect vibration signals of rail transit equipment (such as bearings, gears) through acceleration sensors, sampling frequency ≥ 10kHz.

[0105] (2) Data cleaning:

[0106] De-noising: use wavelet threshold de-noising (such as Sym8 wavelet, soft threshold processing).

[0107] Standardization: Z-Score standardization of signals to eliminate dimensional effects.

[0108] (3) Feature engineering:

[0109] Time domain features: extract mean, variance, peak factor, kurtosis, etc.

[0110] Frequency domain features: extract spectral energy, dominant frequency, and frequency band energy ratio through FFT.

[0111] Time-frequency features: use STFT to generate time-frequency spectrograms and convert them into grayscale images as CNN inputs.

[0112] II. Model design

[0113] (1) Network structure:

[0114] Input layer: accepts vibration signal sequences of length T (such as T = 1024).

[0115] LSTM layer: bidirectional LSTM with hidden layer dimension 128 to capture forward and backward time sequence dependencies.

[0116] Attention mechanism: introduce self-attention mechanism to weight key time steps.

[0117] Fully connected layer: outputs fault type probabilities (such as normal, wear, and fracture).

[0118] (2) Hyperparameter settings:

[0119] Learning rate: 0.001, using Adam optimizer.

[0120] Batch size: adjust according to data length (such as 128).

[0121] Dropout: Set dropout of 0.5 after LSTM layer to prevent overfitting.

[0122] III. Training Process

[0123] (1) Loss Function: Cross-entropy loss function with class weight balancing (e.g., set class_weight when there are fewer fault samples).

[0124] (2) Training Strategy:

[0125] Early Stopping: Stop training when the validation set loss does not decrease for 5 consecutive rounds.

[0126] Learning Rate Decay: Decay the learning rate to 0.9 times every 10 rounds.

[0127] IV. Evaluation and Optimization

[0128] (1) Evaluation Metrics:

[0129] Accuracy: The proportion of correctly classified samples.

[0130] F1-Score: Balance precision and recall.

[0131] (2) Optimization Strategy:

[0132] Feature Selection: Use mutual information method to select key features and reduce redundancy.

[0133] Model Fusion: Integrate multiple LSTM models (e.g., different time steps) for voting decision.

[0134] V. Deployment and Application

[0135] (1) Real-time Analysis: Deploy the model to edge computing nodes to realize real-time classification and early warning of vibration signals.

[0136] (2) Visualization Interface: Develop a web interface to display vibration signal time-frequency spectrum, fault probability, and historical trends.

[0137] In this embodiment, the application is also equipped with a high-performance server cluster for rapid processing and analysis of image data that may have faults uploaded to the cloud. For example, Dell PowerEdge servers are used to run complex image recognition algorithms. When the server can perform in-depth analysis on the images uploaded by the edge computing device, it can accurately determine the type and location of the fault in the image data by comparing it with the pre-set data storage module

[0138] When a record containing latitude, longitude, image ID, confidence "tunnel crack" is written in the algorithm analysis library, the data storage module automatically associates the previous detection data of the location, calculates the crack propagation rate, and if the rate exceeds the threshold, notifies the work order system through Webhook to generate an emergency maintenance task.

[0139] As Figure 3 shown, the present application provides a detection result of demonstration, including: defect type, confidence and location coordinates of key components in rail transit equipment; wherein the key components include: insulator, pole, pole number, positioner electrical connection, U-shaped clamp, base, wire clamp and screw. The detection result of the insulator in the deep learning network model in the present application is shown in Table 1; the detection result of the wire clamp is shown in Table 2; and the detection result of the equipotential line is shown in Table 3.

[0140] Table 1 Insulator detection

[0141]

[0142] In summary, the overall accuracy rate of insulator detection is: (96%+100%+100%+100%+95%) ÷5=98.2%.

[0143] Table 2 Wire clamp detection

[0144]

[0145] In summary, the overall accuracy rate of wire clamp detection is: (98%+97%+100%+100%+93%) ÷5=97.6%.

[0146] Table 3 Equipotential line detection

[0147]

[0148]

[0149] In summary, the overall accuracy rate of equipotential line detection is: (100%+95%+96%+100%+94%) ÷5=97%.

[0150] The rail transit fault analysis system based on image recognition in the present application also has the following advantages:

[0151] 1. High precision and multi-dimensional detection capability

[0152] Multi-source signal fusion: the system integrates multi-source signals such as structure light, infrared light, ultraviolet light from line array camera, laser radar and optical fiber sensor, and combines deep learning algorithm, which can detect parameters such as catenary wear, pantograph temperature, hard point and burning alone at the same time, and the recognition accuracy reaches the international advanced level 1210.

[0153] Three-dimensional visual modeling: Using two-dimensional / three-dimensional visual joint measurement technology, generate three-dimensional models of tracks and vehicle components, covering visible defects such as rail peeling, loose bolts, and track bed cracks.

[0154] 2. Real-time and automated processing

[0155] Millisecond response: The vehicle 360° appearance image fault detection system collects images at high speed through a line array camera, and analyzes them in real time with edge computing devices, shortening the fault identification time to milliseconds, supporting the detection of trains in operation without stopping.

[0156] Closed-loop alarm mechanism: After detecting an anomaly, the system automatically triggers an alarm signal and pushes it to the maintenance work order system.

[0157] 3. Multi-dimensional data integration and intelligent decision-making

[0158] Heterogeneous data fusion: The system supports the integration and analysis of multi-source heterogeneous data such as catenary parameters, vibration signals, and image data, builds a fault correlation model, and improves the efficiency of alarm troubleshooting.

[0159] Predictive maintenance: Based on the LSTM algorithm, analyze time series data such as turnout current, predict the probability of failure, and improve performance by 8% compared to traditional methods.

[0160] 4. Cost and safety optimization

[0161] Low cost: The pantograph detection patent reduces the need for manual inspection, reducing maintenance costs by 40%.

[0162] Safety: Markers avoid misjudgment by maintenance personnel through marking, combined with real-time alarm mechanisms, to prevent secondary failures such as loosening and breaking.

[0163] Example 2

[0164] The application also provides a method for applying the image recognition-based rail transit fault analysis system of Example 1, comprising the following steps:

[0165] S1, collect the original images and vibration data of the rail transit equipment and perform preprocessing;

[0166] S2, store the original images and vibration data, preprocessed images and vibration data, and historical detection records; and based on the historical detection records, compare the images and vibration data with defects with normal images and vibration data, and establish a defect database;

[0167] S3, train the deep learning network model using the defect database to obtain a deep learning network model capable of identifying defects in rail transit equipment;

[0168] The preprocessed image and vibration data are subjected to fault recognition by using a deep learning network model capable of identifying defects of rail transit equipment, and the recognition result is stored in the data storage module, and a high-priority defect alarm is pushed to a maintenance work order system.

[0169] The foregoing description of the disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rail transit fault analysis system based on image recognition, characterized in that, include: The data acquisition and processing module is used to acquire raw images and vibration data of rail transit equipment and perform preprocessing. The data storage module is used to store raw images and vibration data, preprocessed images and vibration data, and historical inspection records; and based on the historical inspection records, it extracts the features of the defect images and vibration data to establish a defect feature database. The fault analysis module has a built-in algorithm analysis library, which is used to train a deep learning network model using a defect feature database to obtain a deep learning network model that can identify defects in rail transit equipment. Using a deep learning network model capable of identifying defects in rail transit equipment, fault identification is performed on preprocessed images and vibration data, and the results of the fault identification are stored in the data storage module. At the same time, high-priority fault alarms are pushed to the maintenance work order system.

2. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, The data acquisition and processing module includes: The data acquisition module is used to acquire raw images and vibration data of rail transit equipment; The preprocessing module is used to clean, convert the format and synchronize the timing of the original images and vibration data to obtain preprocessed images and vibration data. The caching and distribution module is used to divide the preprocessed image and vibration data into real-time paths and archived paths. The real-time path is used to push the data to a Kafka topic for real-time consumption by the algorithm analysis library. The archived path is stored in the data storage module for offline analysis and long-term storage.

3. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, The original images of the rail transit equipment include: the size and location of foreign objects on the track, and the deformation of train components.

4. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, The preprocessing process includes: grayscale conversion, noise reduction, or normalization.

5. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, The data storage module includes relational tables, a spatiotemporal database, file storage, an indexing service, and an API gateway module.

6. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, The algorithm analysis library includes: Algorithm containers are used to encapsulate deep learning algorithms using Docker containers, enabling runtime environment isolation and rapid deployment. The task scheduling module is used to dynamically allocate CPU / GPU resources to ensure the efficient operation of the algorithm container; The intermediate result caching module is used to cache the surface feature vectors of rail transit equipment using the in-memory database Redis, in order to accelerate the inference process of subsequent models.

7. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, Fault identification is performed on preprocessed images and vibration data using a deep learning network model capable of identifying defects in rail transit equipment, including: Using a YOLOv7 defect detection model based on convolutional neural networks, defect features in images are identified; We use an LSTM vibration analysis model based on recurrent neural networks to process and analyze time series data of vibrations from rail transit equipment.

8. The rail transit fault analysis system based on image recognition according to claim 1, characterized in that, The results of the fault identification include: the defect type, confidence level, and location coordinates of key components in rail transit equipment; among which key components include: insulators, poles, pole numbers, locator electrical connections, U-shaped clamps, bases, wire clamps, and screws.

9. A rail transit fault analysis method based on image recognition, applied to the rail transit fault analysis system based on image recognition as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. Acquire raw images and vibration data of rail transit equipment and perform preprocessing; S2. Store the original images and vibration data, the preprocessed images and vibration data, and historical inspection records; and based on the historical inspection records, compare the images and vibration data with defects with the normal images and vibration data to establish a defect database. S3. Use the defect database to train the deep learning network model to obtain a deep learning network model that can identify defects in rail transit equipment. A deep learning network model capable of identifying defects in rail transit equipment is used to perform fault identification on preprocessed images and vibration data, and the identification results are stored in the data storage module. At the same time, a high-priority defect alarm is pushed to the maintenance work order system.

Images