Method for real-time detection of bullet train faults on parking line

Abstract

The invention provides a method for detecting motor train faults in real time on a parking line. The method comprises the following steps: decoding attribute information of a to-be-detected motor train unit; detecting pressure data on the parking line; determining a vehicle structure template matched with the attribute information of the to-be-detected motor train in a vehicle structure template library, and determining the current position of the to-be-detected motor train unit on the parking line based on the pressure data to generate a motor train dragging instruction; analyzing the motor train dragging instruction to obtain a matched vehicle structure template and a current position, and determining a target position to which the to-be-detected motor train unit needs to be moved when the to-be-detected motor train unit is overhauled on the parking line based on the matched vehicle structure template so as to drag the to-be-detected motor train unit to move from the current position to the target position. A feedback signal indicating that the to-be-detected motor train unit is in a ready state is generated and sent to the control module; and based on triggering of the feedback model, generating a troubleshooting instruction to drive a troubleshooting combination installed on the train storage line to carry out fault detection on the to-be-overhauled motor train unit.

Description

Technical Field

[0001] The invention relates to the technical field of railway freight transportation, and in particular to a method for real-time detection of motor vehicle faults on a storage line. Background Art

[0002] In the current field of EMU maintenance and overhaul, with the rapid development of high-speed railway technology, the operational safety and efficiency of EMUs have become crucial factors. Traditional EMU fault detection methods usually rely on manual inspections or static inspections, which have many limitations. For example, manual inspections are not only time-consuming and laborious, but also limited by the experience and skill level of the inspectors, making it difficult to ensure the comprehensiveness and accuracy of the inspections; static inspections need to be carried out when the EMU is stopped, which seriously affects the operational efficiency and availability of the EMU.

[0003] In the existing technology, although there are some EMU information management systems based on RFID (radio frequency identification) technology, most of these systems are only used for EMU identification and tracking, and fail to be closely integrated with the maintenance of EMUs. At the same time, although the pressure sensors on the storage line can detect pressure data, they often fail to effectively use these data to achieve precise positioning and drag control of the EMU. Therefore, the existing EMU fault detection methods have obvious deficiencies in intelligence, automation, and maintenance efficiency.

[0004] In addition, due to its particularity and technical requirements, the location of EMU maintenance sites is usually limited, and due to the high construction costs, capacity expansion becomes relatively difficult. Therefore, EMU maintenance work often needs to arrange a schedule in advance to ensure the smooth progress of the maintenance work. At present, maintenance robots are installed in the maintenance depot, and they move back and forth along the vehicle tracks to inspect the bottom of the EMU. These robots use image recognition technology to determine the components currently being inspected, but due to technical limitations, they have difficulty in accurately measuring the dimensions of the components. This means that if there is a slight deformation or damage to the component, the traditional maintenance robot may not be able to detect these potential faults in time, thus affecting the safe operation of the EMU. Summary of the invention

[0005] In order to solve the above technical problems, the present application provides a method for real-time detection of EMU faults on a storage line, so as to at least solve or alleviate the above problems existing in the prior art.

[0006] A method for real-time detection of motor vehicle faults on a storage line, comprising: Park the EMU to be inspected on the storage line; Activate the train-mounted tag set on the train-mounted train to be detected based on the RFID reader / writer to decode the attribute information of the train-mounted train to be detected and send it to the control module; Detecting pressure data on the parking line based on a pressure sensor disposed on the parking line and sending the pressure data to the control module; The control module determines the vehicle structure template that matches the attribute information of the motor vehicle to be detected in the vehicle structure template library, and determines the current position of the motor vehicle group to be detected on the storage line based on the pressure data, so as to generate a motor vehicle towing instruction based on the matched vehicle structure template and the current position and send it to the motor vehicle towing device; The EMU towing device parses the EMU towing instruction to obtain the matched vehicle structure template and the current position, and determines the target position to which the EMU to be detected is to be moved when the EMU to be detected is inspected on the storage line based on the matched vehicle structure template, so as to drag the EMU to be detected from the current position to the target position, and generates a feedback signal to the control module that the EMU to be detected is in a ready state; The control module generates a fault repair instruction based on the triggering of the feedback model to drive the fault repair combination installed on the storage line to perform fault detection on the EMU to be repaired according to the customized repair strategy under the drive of the mechanical arm, and repair the repairable faults.

[0007] A method for real-time detection of motor vehicle faults on a storage line, comprising: Receive the decoded EMU attribute information sent from the RFID reader / writer, which is obtained after the RFID reader / writer activates the EMU onboard tag; Receiving pressure data on the parking line detected and sent by the pressure sensor; According to the received EMU attribute information, the corresponding vehicle structure template is matched in the vehicle structure template library, and the current position of the EMU on the storage line is determined in combination with the pressure data. According to the matched vehicle structure template and the current position, the EMU towing instruction is generated and sent to the EMU towing device; receiving a ready signal fed back by the EMU towing device after towing the EMU from the current position to the target position, where the target position is determined by the EMU towing device after parsing the EMU towing instruction; After receiving the ready signal, a fault inspection and repair instruction is generated to drive the fault inspection and repair combination installed on the storage line to perform fault detection and repair on the EMU under the drive of the mechanical arm.

[0008] In the present application, a local screenshot of a typical freight car fault and a complete fault train monitoring image are obtained as the original railway freight car fault image; the original railway freight car fault image is scored to obtain an image quality score value, and based on the image quality score value, the original railway freight car fault image is cleaned to obtain a valid railway freight car fault image; the valid railway freight car fault image is image balanced to obtain a balanced fault image; the valid railway freight car fault image is classified to obtain a fault classification label and a fault level; the balanced fault image is fully labeled with components and partially labeled with fault locations to obtain fault labeling data; based on the balanced fault image and the fault classification label set, the fault level set, and the fault labeling data, a railway freight car fault image intelligent recognition data set is constructed, which meets the training requirements of the deep learning algorithm, effectively promotes the development of TFDS railway freight car fault image intelligent recognition technology, and lays a good technical foundation for ensuring the safe operation of railway freight cars. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flow chart of a method for real-time detection of motor vehicle faults on a storage line according to an embodiment of the present application. DETAILED DESCRIPTION

[0010] Figure 1 FIG. 1 is a flow chart of a method for real-time detection of motor vehicle faults on a storage line according to an embodiment of the present application. Figure 1 As shown, the device for real-time detection of motor vehicle faults on a storage line comprises: Park the EMU to be inspected on the storage line; Activate the train-mounted tag set on the train-mounted train to be detected based on the RFID reader / writer to decode the attribute information of the train-mounted train to be detected and send it to the control module; Detecting pressure data on the parking line based on a pressure sensor disposed on the parking line and sending the pressure data to the control module; The control module determines the vehicle structure template that matches the attribute information of the motor vehicle to be detected in the vehicle structure template library, and determines the current position of the motor vehicle group to be detected on the storage line based on the pressure data, so as to generate a motor vehicle towing instruction based on the matched vehicle structure template and the current position and send it to the motor vehicle towing device; The EMU towing device parses the EMU towing instruction to obtain the matched vehicle structure template and the current position, and determines the target position to which the EMU to be detected is to be moved when the EMU to be detected is inspected on the storage line based on the matched vehicle structure template, so as to drag the EMU to be detected from the current position to the target position, and generates a feedback signal to the control module that the EMU to be detected is in a ready state; The control module generates a fault repair instruction based on the triggering of the feedback model to drive the fault repair combination installed on the storage line to perform fault detection on the EMU to be repaired according to the customized repair strategy under the drive of the mechanical arm, and repair the repairable faults.

[0011] Optionally, the fault inspection and repair combination includes a fault detection module and a fault repair module, the mechanical arm includes a first mechanical arm and a second mechanical arm, the fault detection module is arranged on the first mechanical arm, and the fault repair module is arranged on the second mechanical arm, and the control module generates a fault inspection and repair instruction based on the triggering of the feedback model to drive the fault inspection and repair combination installed on the storage line to perform fault detection on the EMU to be inspected according to a customized inspection and repair strategy under the drive of the mechanical arm, and repair the repairable fault, including: The control module generates a fault detection instruction based on the triggering of the feedback model to drive the fault detection module installed on the first mechanical arm to perform fault detection on the EMU to be repaired according to a customized detection strategy; When a fault is detected, the control module generates a fault repair instruction to drive the fault repair module installed on the second mechanical arm to repair the repairable fault according to a customized repair strategy.

[0012] Optionally, the fault detection module includes: an image detection component, an ultrasonic detection component, a waveform detection component, and at least one of a temperature detection component; The image detection component is used to perform the following steps to determine whether there are component failures at the bottom and side of the body of the EMU to be inspected: Capturing images of the bottom and / or side of the vehicle body of the EMU to be inspected, extracting features from the captured images, and determining the morphological changes of the bottom and side of the vehicle body accordingly; Calibrate the distance between the bottom and / or side of the vehicle body and the image detection assembly to determine the thickness change of the bottom and side of the vehicle body according to the calibrated distance; According to the morphological change and the thickness change, combined with the vehicle structure template of the EMU to be tested, it is determined whether key EMU components at the bottom and / or side of the body of the EMU to be tested are faulty.

[0013] Optionally, the waveform detection component is used to perform the following steps to determine whether there are any faulty components on the EMU to be detected: Obtain the jitter data of the train passing detected by the train jitter sensor on the train to be detected, Performing waveform analysis on the vehicle passing vibration data to extract waveform features therein; According to the waveform characteristics and in combination with the vehicle structure template of the EMU to be tested, it is determined whether there are any faulty components on the EMU to be tested.

[0014] Optionally, the temperature detection component is used to perform the following steps to determine whether there are any faulty components on the EMU to be detected; Performing non-contact temperature collection on set detection points on the body of the EMU to be detected; According to the collected temperature, starting from the set detection point and in combination with the vehicle structure template of the EMU to be detected, it is determined whether there are any faulty parts on the EMU to be detected.

[0015] Preferably, the specific implementation details of the above steps are as follows: (1) Feedback signal triggering and command generation Continuously monitor the feedback signal input of the control module. When a specific feedback signal is triggered (which can be set to signal strength, frequency and other parameters meet the set threshold conditions), the following operations are performed: According to the preset logic, the control module generates fault detection instructions. This process can be regarded as a simple conditional judgment algorithm. For example, if the feedback signal is , the trigger threshold is ,when When a fault detection instruction is generated, the operation is executed.

[0016] (2) Fault detection module performs detection After the fault detection command is sent to the fault detection module installed on the first robot arm, the module works according to the customized detection strategy, which involves the following detection components and related algorithms: After the image is collected, the convolutional neural network (CNN) algorithm is used to extract features. In CNN, features are extracted by performing convolution operations between the convolution kernel of the convolution layer and the image matrix. Suppose the input image is , the convolution kernel is , the feature map after convolution operation It can be expressed as: (* indicates convolution operation). Through multi-layer convolution, pooling and other operations, the feature representation of the image is obtained, which is used to determine the changes in the vehicle body shape.

[0017] The distance calibration adopts the triangulation principle and related algorithm. Assume that the image detection component and the vehicle body have two points , The distances are , , the thickness change is calculated using trigonometric formulas based on known geometric relationships and measured angles. For example, in a simple triangle model, if the angle is known, and one side length , can be based on ( is the required thickness related quantity, is the distance) to calculate thickness related information.

[0018] When combining the vehicle structure template to identify faults, a pattern recognition algorithm (such as support vector machine SVM) is used. The extracted morphology and thickness change features are used as input vectors. , through the decision function of SVM ( is the weight vector, is bias) to determine whether it is a fault. When certain conditions are met (such as greater than 0 or less than 0, according to training settings), it is judged as a fault.

[0019] For the ultrasonic detection component: After transmitting and receiving the ultrasonic signal, the fast Fourier transform (FFT) algorithm is used to convert the time domain signal into a frequency domain signal. Assume that the time domain signal is , the frequency domain signal after FFT transformation for: By analyzing the frequency domain signal characteristics (such as frequency components, amplitude, etc.) and comparing them with the ultrasonic signal characteristics under normal conditions, internal defects of components can be determined.

[0020] For the waveform detection component: After obtaining the jitter data of the passing vehicle, the wavelet transform algorithm is used to extract the waveform features. Assume that the jitter data is , its wavelet transform for: ( is the scale parameter, is the translation parameter, is a wavelet function). The extracted features are compared with the normal waveform features corresponding to the vehicle structure template to determine whether there are faulty parts.

[0021] For the temperature detection component: After the temperature is collected, the temperature data is analyzed based on the algorithm. Suppose the collected temperature value is ( Indicates different detection points), by setting the temperature threshold The temperature field distribution model constructed in combination with the vehicle structure template is used to determine the conditional formula, such as when ( is the set temperature change threshold), it is determined that there may be a faulty component near the detection point.

[0022] (3) Fault diagnosis and repair instruction generation The fault detection module feeds back the detection results to the control module, and the processor makes fault judgment based on these results. If a fault is detected, the following operations are performed: Using the rule-based judgment algorithm, let the fault feature vector be , according to the preset fault rule set , determine whether the fault condition is met. For example, if Some eigenvalues ​​of If the conditions of a specific fault type are met, it is judged as a fault of that type.

[0023] When a fault is determined, the control module generates a fault repair instruction.

[0024] (4) The fault repair module performs repair The fault repair instruction is sent to the fault repair module installed on the second robot arm, and the module works according to the customized repair strategy: For cleaning components: Use a path planning algorithm (such as the A* algorithm) to plan the cleaning path. In the A* algorithm, the heuristic function of the calculation node is used And the actual cost function , get the node's evaluation function , select the optimal path for cleaning. At the same time, combined with the pressure value fed back by the pressure sensor , use a closed-loop control algorithm (such as PID control algorithm) to adjust the cleaning parameters. Suppose the output of the PID controller is for: ( , , are proportional, integral and differential coefficients, is pressure deviation), and adjusts parameters such as cleaning fluid flow, cleaning tool speed, etc. in real time.

[0025] For heating components: When heating electrical connection components, etc., use the PID control algorithm to accurately control the temperature. Set the target temperature to The actual temperature is , temperature deviation , adjust the heating power through the PID control algorithm , so that the actual temperature approaches the target temperature.

[0026] In the PID control algorithm of the heating component, in order to accurately control the temperature, it is necessary to coordinate the three links of proportion, integration, and differentiation to ensure that the actual temperature is stable and quickly approaches the target temperature. The specific algorithm for accurately controlling the temperature using the PID control algorithm is as follows: 1. Proportional link: The proportional link responds quickly based on the current temperature deviation and adjusts the heating power. Set the target temperature to The actual temperature is , the temperature deviation is The proportionality factor is set to , whose output It is proportional to the temperature deviation, and the formula is In the initial stage of heating, Larger, It can rapidly increase the heating power and quickly narrow the gap with the target temperature. Too large may cause temperature overshoot.

[0027] 2. Integral link: The integral link is used to eliminate steady-state errors and integrate the temperature deviation over time. The integral coefficient is , whose output The formula is As heating progresses, even if the proportional link brings the temperature close to the target value, there may still be a small deviation. The integral link will continue to accumulate this deviation, and its output will increase over time, thereby increasing the heating power and reducing the steady-state error. If it is too large, the system response will become slow or even unstable.

[0028] 3. Differential link: The differential link adjusts the heating power in advance according to the rate of change of temperature deviation to suppress overshoot. The differential coefficient is , whose output The formula is When the temperature approaches the target value, The rate of change is negative, Reduce heating power to prevent overheating. If it is too large, the system will be sensitive to noise, causing heating power fluctuations.

[0029] 4. Comprehensive output: Add the outputs of the three links to obtain the final signal for controlling the heating power , the formula is In practical applications, the thermal characteristics and heating requirements of electrical connection components are continuously optimized. , , , to achieve precise temperature control, ensure heating effect while avoiding overheating and damage to components.

[0030] 5. Intelligent adaptive adjustment: Introduce fuzzy control algorithm to realize intelligent adaptive adjustment of PID parameters. and its rate of change , dynamically adjusted through fuzzy rules , , For example, when and When both are larger, increase To heat up quickly; when Smaller but When it is larger, reduce , increase Prevent overshoot. Through this intelligent adjustment, PID control can maintain good temperature control performance under different working conditions.

[0031] For laser obstacle removal components: a path planning algorithm based on a geometric model is used to plan the laser processing path. According to the geometric shape and fault location of the component, a geometric model is constructed to calculate the optimal motion trajectory of the laser beam to ensure accurate processing of stubborn attachments, cracks, etc.

[0032] Preferably, the specific implementation process of planning the laser processing path using a path planning algorithm based on a geometric model is as follows: The following will combine specific mathematical formulas to elaborate on the path planning process of the laser obstacle removal component, from building a geometric model to calculating the optimal motion trajectory.

[0033] 1. Build the geometric model: First, the parts are digitally modeled. The surface of the parts is represented by parametric surfaces, such as the common Bezier surface. The mathematical expression of the secondary Bezier surface is:

[0034] in, is a parameter, and is the Bernstein polynomial, These control vertices are used to accurately describe the geometry of the component.

[0035] For the fault location, its coordinates in the Cartesian coordinate system are: Project the fault location onto the constructed component geometry model and find its corresponding parameter value on the parametric surface , which can be achieved by solving the system of equations:

[0036] in, It is a parametric surface The x, y, z coordinates are expressed in Cartesian coordinates.

[0037] The A algorithm is used to find the optimal trajectory of the laser beam. The core of the A algorithm is to select the optimal path through the evaluation function f(n), which is defined as:

[0038] in, represents the nodes on the path, From the starting point to the node The actual cost of It is a slave node Estimated cost to the target point (fault location).

[0039] For the movement of the laser beam on the surface of the component, It can be defined as the distance between adjacent nodes. Assume that the nodes and The corresponding parameter values ​​on the parametric surface are and , then the distance between them It can be calculated by the following formula:

[0040] - The Euclidean distance can be used as the estimated cost, that is, from the current node To the corresponding point of the fault location on the parametric surface The Euclidean distance of:

[0041] in, Is the current node Parameter value on a parametric surface.

[0042] During the search process of the A* algorithm, we continuously update and ,choose The node with the smallest value is used as the next expansion node until the optimal path from the starting position of the laser beam to the fault position is found. This path is the optimal movement trajectory of the laser beam on the surface of the component, ensuring that stubborn attachments, cracks, etc. can be accurately processed.

[0043] For mechanical obstacle removal components: When disassembling, replacing or repairing parts, use the inverse kinematics algorithm to control the movement of the robot arm. Assume that the desired position of the end of the robot arm is , the posture is , solve the angles of each joint through the inverse kinematics equation ( represents the joint number), so that the robot arm reaches the specified position and posture to complete the repair task.

[0044] From constructing the geometric model to calculating the optimal motion trajectory, the following steps are taken to implement path planning for the laser obstacle removal component: 1. Using parametric surfaces (such as Bezier surfaces) to build component geometry models is an innovation. Traditional geometric modeling methods may have difficulty accurately describing complex component shapes, while parametric surfaces can flexibly adjust the surface shape by controlling vertices and Bernstein polynomials to more accurately reflect the actual geometric features of components. This modeling method provides a more accurate basis for subsequent fault location and path planning, and is an improvement and breakthrough in traditional geometric modeling technology.

[0045] 2. Applying the A* algorithm to the path planning of laser obstacle removal components is an innovative attempt. The A* algorithm is widely used in fields such as robot path planning, but its application in the specific scenario of laser obstacle removal combines the component geometry model and fault location information through custom and Function, making it more in line with the actual needs of laser obstacle removal. The integrated application of this cross-domain algorithm provides new ideas and methods for path planning of laser obstacle removal technology, and improves the intelligence and automation level of laser obstacle removal.

[0046] To this end, the above-mentioned construction of the geometric model to calculate the optimal motion trajectory and the specific implementation details of the path planning of the laser obstacle removal component have the following technical benefits: 1. Accurate positioning and processing: By constructing a geometric model in the form of a parametric surface, such as a Bezier surface, the complex geometric shapes of parts can be accurately described. By projecting the fault location onto the model, the corresponding position of the fault on the surface of the part can be accurately found, providing accurate positioning for subsequent laser processing. For example, when processing tiny cracks in engine parts, accurate positioning ensures that the laser beam only acts on the crack site, avoiding damage to the surrounding normal area, and improving the accuracy and safety of obstacle removal.

[0047] 2. Efficient path planning: A* algorithm is used to calculate the optimal trajectory of the laser beam, and the evaluation function Consider both the actual cost and the estimated cost. In practical applications, reflects the actual distance cost of the laser beam moving, The search direction is guided toward the target fault location, which greatly reduces the search space and improves the efficiency of path planning. Compared with the traditional blind search algorithm, the A* algorithm can quickly find the best path from the laser starting position to the fault location, saving processing time and improving overall work efficiency. It is especially suitable for industrial production scenarios with strict requirements on processing time.

[0048] 3. Strong adaptability: This path planning method has good adaptability to parts of different shapes and types. Whether it is a part with a regular shape or a part with a complex surface, path planning can be performed by building a suitable geometric model.

[0049] To this end, the technical benefits brought by the above specific steps in the application scenario of EMU fault maintenance are as follows: 1. In fault detection, when multiple detection components (image detection, ultrasonic detection, waveform detection, temperature detection) point to the same component, confidence evaluation and fusion are performed through formulas. Single confidence evaluation formula Different detection principles were considered ( ), different components under the same principle ( ) and the parts to be tested ( ) factors, through the weight value , Comprehensively evaluate the reliability of the test results of each test component for a specific component. Fusion confidence evaluation formula, such as as well as , the single confidence evaluations of multiple detection components are integrated. This multi-source data fusion and confidence evaluation method is an innovation that breaks the limitations of a single detection method and can judge component failures more comprehensively and accurately.

[0050] 2. Detection Algorithm For the image detection component: Convolutional neural network (CNN) is used for image feature extraction through convolution operation It can automatically learn the key features in the image. Compared with the traditional method of manually designing features, it is more efficient and accurate in extracting the complex changes in the shape of the train body. The triangulation principle-related algorithm is used for distance calibration, and the thickness change is calculated based on the trigonometric formula, which provides a quantitative basis for accurately judging the shape of the component. The decision function of the support vector machine (SVM) Used for fault identification, it can accurately divide fault and non-fault categories in high-dimensional feature space.

[0051] For ultrasonic detection components: using the fast Fourier transform (FFT) algorithm Converting ultrasonic time domain signals into frequency domain signals and analyzing the internal characteristics of components from a frequency domain perspective can more clearly identify potential defects than simple time domain analysis.

[0052] For the waveform detection component: wavelet transform algorithm By processing the jitter data of passing vehicles, it is possible to analyze waveform characteristics at different scales and positions, adapt to complex and changeable jitter signals, and more accurately capture waveform anomalies caused by faults.

[0053] For temperature detection components: Based on the temperature data analysis algorithm, combined with the set temperature threshold and temperature change threshold And the temperature field distribution model constructed by the vehicle structure template, through the conditional judgment formula such as To judge the fault, the temperature data is combined with the vehicle structure to realize the temperature-structure correlation analysis of the fault.

[0054] 3. Repair module control algorithm For cleaning components: A* algorithm is used for cleaning path planning, through the evaluation function Find the optimal cleaning path to improve cleaning efficiency and coverage. PID control algorithm Adjust cleaning parameters in real time based on pressure sensor feedback to ensure the stability and consistency of the cleaning effect.

[0055] For heating components: PID control algorithm accurately controls the heating temperature to and Deviation For input, adjust the heating power , ensuring that the heating treatment can achieve the repair purpose without causing overheating damage to the components.

[0056] For laser obstacle removal components: The path planning algorithm based on the geometric model plans the laser processing path according to the geometric shape of the components and the fault location, so as to achieve accurate processing of stubborn attachments and cracks and improve the repair accuracy.

[0057] For mechanical obstacle removal components: the inverse kinematics algorithm is based on the desired position of the end of the robotic arm and posture Solving for joint angles , to achieve precise motion control of the robotic arm and ensure the accuracy of parts disassembly, replacement or repair operations.

[0058] Optionally, if fault detection is performed based on the image detection component, ultrasonic detection component, waveform detection component, and temperature detection component and points to the same component, the method further includes: performing individual confidence evaluations on the faulty components detected by the image detection component, ultrasonic detection component, waveform detection component, and temperature detection component, and fusing all individual confidence evaluations to obtain a fused confidence evaluation, so as to classify and confirm the fault of the component based on the fused confidence evaluation and the individual confidence evaluation.

[0059] Optionally, confidence evaluation is performed on the faulty components detected by the image detection component, the ultrasonic detection component, the waveform detection component, and the temperature detection component, respectively, including: Based on the following formula, the single confidence evaluation of the image detection component, the ultrasonic detection component, the waveform detection component, and the temperature detection component is determined:

[0060] in, is the mapping function, is the weight value, , , n=1,2,3,4, ; All confidence evaluations are integrated to obtain a fused confidence evaluation, including: based on the following formula, all confidence evaluations are integrated to obtain a fused confidence evaluation:

[0061] in, is the weight value, and the weight value is , determined based on prior knowledge; or, ,in, is a constant, is the weight value, which is determined based on prior knowledge.

[0062] Optionally, the fault repair module includes: at least one of a cleaning component, a heating component, a laser obstacle removal component, and a mechanical obstacle removal component; The cleaning component removes surface attachments, dirt, and grease from the following faulty parts: Wheels, axles, brake system components: remove oil, rust, and foreign matter from wheel treads, axle surfaces, brake discs, and brake calipers; Suspension system components: clean dust and oil on air springs, shock absorbers, suspension arms and other components; Vehicle side parts: clean the stains on doors, windows and side wall panels; Fasteners: remove rust and oil stains on fasteners such as bolts, nuts, and rivets; The heating assembly performs heating treatment on the following faulty parts: Electrical connection parts: Heat the wiring terminals and cable joints with poor contact to soften the insulation layer or solder, re-weld or tighten them to restore the electrical connection; Braking system components: locally heat the worn brake pads and repair or replace them after softening; Other metal parts: heating the parts that need welding repair, including broken suspension arms and bogie frames, and restoring their integrity through welding; Preferably, for the above-mentioned cleaning component and heating component, the specific implementation process of the above-mentioned solution is as follows: 1. Cleaning components 1. Cleaning process based on machine vision and path planning Machine vision detection: Use the OpenCV library to build an image recognition system based on a convolutional neural network (CNN). Use TensorFlow or PyTorch as a deep learning framework to build a simple CNN model. Assume that the LeNet-5 model structure is used, which contains two convolutional layers, two pooling layers, and three fully connected layers. Input layer: Receives a size of Color image (the image size can be adjusted according to actual conditions). Convolutional layer 1: Use indivual The convolution kernel of is , filled with , the activation function uses the ReLU function, and the output feature map size is The calculation formula is: ,in It is Tier Line The output of the column, It is The convolution kernel weights of the layer, It is The input of the layer, is the bias. Pooling layer 1: using The maximum pooling with a step size of , the output feature map size is . Convolutional layer 2: uses 16 The convolution kernel of is , filled with , the activation function is ReLU, and the output feature map size is . Pooling layer 2: also uses The maximum pooling with a step size of , the output feature map size is . Fully connected layer 1: Flatten the output of pooling layer 2 into a one-dimensional vector, connected to $120$ neurons, with ReLU as the activation function. Fully connected layer 2: Connected to 84 neurons, with ReLU as the activation function. Output layer: Connected to Neurons ( is the number of stain categories, such as oil, rust, foreign matter, etc.), and the Softmax function is used for classification, and the probability of each category is output. Path planning: A* algorithm is used to plan the cleaning path. First, a two-dimensional array is defined to represent the cleaning area. Each element in the array represents a position. A value of 0 indicates that it is passable, and a value of 1 indicates an obstacle (such as the non-washable part of a component).

[0063] Define the node class, including location information ( , coordinates), parent node pointer, value (the actual cost from the starting point to the node), value (the estimated cost from this node to the target node) and value( ).

[0064] Initialize a priority queue to store the nodes to be expanded, according to The values ​​are sorted from smallest to largest.

[0065] Add the starting point to the priority queue and set the starting point The value is 0, The value can be calculated by Manhattan distance (such as ).

[0066] Loop from the priority queue The node with the smallest value is expanded: If the expanded node is the target node (i.e., the location where the stain is detected), the path is found and the path is obtained by backtracking through the parent node pointer.

[0067] Otherwise, check the four adjacent nodes above, below, left and right of the node (if they are in the cleaning area and are not obstacles): Calculate the neighboring nodes value( , assuming each move costs 1) and value.

[0068] If the adjacent node is not in the priority queue, add it to the priority queue and set its parent node to the current node.

[0069] If the adjacent node is already in the priority queue and the newly calculated The value is smaller than the original value, update its value, value, value and parent node.

[0070] 2. Cleaning parameter adjustment based on fuzzy control The input variables are defined as the severity of the stain and the cleaning time, and the output variables are the cleaning fluid flow rate and the cleaning tool rotation speed.

[0071] Fuzzify the input and output variables: The severity of stains is divided into three fuzzy subsets: "low", "medium" and "high", and their membership functions are defined respectively. For example, for "low", the Gaussian membership function is used ,in is the center value, is the standard deviation and can be adjusted according to actual conditions; “medium” and “high” are similarly defined.

[0072] The cleaning time is divided into three fuzzy subsets: “short”, “medium” and “long”, and the corresponding membership functions are also defined.

[0073] The cleaning fluid flow rate is divided into three fuzzy subsets of “small”, “medium” and “large”, and the cleaning tool speed is divided into three fuzzy subsets of “slow”, “medium” and “fast”, and their respective membership functions are defined.

[0074] A fuzzy rule base is established based on experience, for example: if the severity of the stain is "high" and the cleaning time is "short", then the flow rate of the cleaning liquid is "large" and the speed of the cleaning tool is "fast".

[0075] Rules are expressed in "if-then" form, such as :if (the severity of the stain is high) and (the cleaning time is short) then (the cleaning liquid flow is large) and (the cleaning tool speed is fast).

[0076] Fuzzy reasoning is performed using the Mamdani reasoning method. For each rule, the membership of the output variable is calculated based on the membership of the input variable, and then synthesized by taking the maximum value.

[0077] Finally, the centroid method is used to perform defuzzification and calculate the precise control value of the cleaning fluid flow rate and the cleaning tool speed. Assume that the output variable The membership function is , then the exact value after defuzzification is .

[0078] (ii) Heating components 1. Accurate temperature control based on model predictive control Establish a heating model: Take the heating of electrical connection components as an example, consider its heat transfer process, and establish a lumped parameter thermal model. Assume that the component consists of multiple thermal resistors. and heat capacity According to the law of conservation of energy, for each node ,have ,in Is a node The temperature, Is with the node Connected nodes The temperature, Is a node and The thermal resistance between Is a node Internal heat source (such as heating power).

[0079] Model Predictive Control (MPC): Defining the Prediction Horizon and control time domain . Predicting the future based on the heating model Temperature at each time step ( , is the current moment). Define the objective function ,in is the target temperature, yes Time has come The heating power change at each moment, is the weight coefficient used to balance the temperature tracking error and the heating power change. Use an optimization algorithm (such as a quadratic programming algorithm) to solve the objective function and obtain the heating power control value at the current moment. In practical applications, prediction and optimization are performed again after each time step to achieve rolling optimization control.

[0080] 2. Heating strategy optimization based on deep learning Data collection and preprocessing: Collect heating repair data of parts with different fault types and different materials, including heating time, heating power, temperature change, repair results, etc. Normalize the data, for example, for temperature data , using the formula Normalize, where and are the minimum and maximum temperatures in the data set, respectively.

[0081] Model construction and training: Long short-term memory network (LSTM) is used to build a heating strategy prediction model. Taking Keras library as an example, a model including LSTM layer and fully connected layer is constructed: Input layer: Receives normalized historical heating data sequence, such as past The heating power, temperature and other data of each time step are input in the form of , where features is the number of features. LSTM layer: set a certain number of neurons (such as 64) and return a sequence (i.e., the output at each time step). Fully connected layer: connects to 1 neuron and outputs the predicted control parameters such as heating power or temperature. Compile the model, select a suitable loss function (such as mean squared error loss function) and optimizer (such as Adam optimizer). Train the model using the collected data, set the number of training rounds (such as 100 rounds) and batch size (such as 32).

[0082] In the actual heating repair process, the current component information (such as material, fault type) and historical heating data are input into the trained model. The model outputs the optimized heating strategy, such as heating power curve, heating time, etc., to guide the work of the heating component.

[0083] Furthermore, in combination with the above working method, the specific structure of the cleaning component and the heating component is described as follows.

[0084] 1. Cleaning components 1. The cleaning component is mainly composed of a robotic arm, a cleaning nozzle, a cleaning brush and other parts. The robotic arm is responsible for driving the cleaning nozzle and the cleaning brush to the position of the parts that need to be cleaned, and its movement trajectory is determined by the A* algorithm path planning mentioned above. The robotic arm has multiple joint degrees of freedom and can flexibly adjust the angle and position to meet the cleaning needs of complex-shaped parts such as wheels, axles, and brake system components. For example, when cleaning the wheel tread, the robotic arm can accurately fit the cleaning brush to the tread according to the path planning and perform all-round brushing.

[0085] 2. The nozzle is used to spray the cleaning liquid, and the flow rate of the cleaning liquid is controlled by a parameter adjustment system based on fuzzy control. The nozzle can be a high-pressure atomizing nozzle, which can atomize the cleaning liquid into tiny particles, increase the contact area between the cleaning liquid and the stains, and enhance the cleaning effect. For the axle surface with serious oil pollution, the flow rate of the cleaning liquid is increased after judgment by the fuzzy control algorithm. The high-pressure atomizing nozzle can quickly and evenly spray the cleaning liquid onto the axle surface, so that the oil pollution is quickly emulsified and decomposed.

[0086] 3. The cleaning brush is made of different materials and shapes according to the different cleaning objects and stain types. For rust and foreign matter on metal parts such as axles and brake discs, a hard wire brush can be used; for dust and oil on rubber or plastic parts such as air springs and shock absorbers, a soft nylon brush can be used to avoid damage to the parts. The speed of the cleaning brush is also adjusted by the fuzzy control algorithm according to the severity of the stain and the cleaning time.

[0087] (ii) Heating components 1. For electrical connection parts, the heating assembly uses electromagnetic induction heating coils as the heating source. Electromagnetic induction heating uses an alternating magnetic field to generate an induced current in a metal part, which in turn generates heat and achieves rapid heating. The design and layout of the heating coils are optimized according to the shape and size of the electrical connection parts to ensure uniform heating. For example, for a terminal block, the heating coil is wrapped around the terminal, and the heating power and speed are precisely controlled by adjusting the frequency and amplitude of the alternating current to quickly soften the insulation layer or solder.

[0088] 2. Use high-precision thermocouples or thermistors as temperature sensors to monitor the temperature of the heating components in real time. These sensors convert temperature signals into electrical signals and feed them back to the control system based on model predictive control (MPC) or deep learning models. When the brake pad is locally heated, multiple temperature sensors are distributed on the surface of the brake pad to collect temperature data in real time and provide accurate temperature information to the control system to achieve precise temperature control.

[0089] 3. It consists of a microcontroller (such as STM32 series), a power drive module, etc. The microcontroller is responsible for running the MPC algorithm or deep learning model, calculating the heating power control value based on the information fed back by the temperature sensor and the preset target temperature, and controlling the operation of the heating source through the power drive module. The power drive module can adjust the current and voltage of the heating source according to the instructions of the microcontroller to achieve precise adjustment of the heating power.

[0090] The laser obstacle removal component performs laser processing on the following faulty parts: Wheels, axles, brake system components: remove stubborn attachments on wheel treads, axle surfaces, brake discs, and tiny fragments on crack edges; Suspension system components: Precisely cut and remove foreign matter and cracks on suspension arms, shock absorbers and other components; Body side parts: laser treatment of stubborn stains and scratches on doors, windows and side panels; The mechanical obstacle removal assembly is used to mechanically process the following faulty parts: Wheels and axles: Cut the cracks and broken parts on the wheels and replace them with new wheels or axle parts; Suspension system components: dismantle, replace or repair broken or deformed suspension arms, shock absorbers and other components; Bogie parts: repair cracks and deformations on the bogie frame, crossbeams and longitudinal beams, and replace damaged parts; Fasteners: Remove and replace loose or broken bolts, nuts, rivets and other fasteners to ensure the fastening effect.

[0091] Specifically, the process of the laser obstacle removal component and the mechanical obstacle removal component implementing the above processing is as follows: (I) Innovative implementation principle of laser obstacle removal components 1. Laser processing flow based on machine vision and path planning Machine vision inspection: Use the OpenCV library combined with Python language to build an image recognition system based on convolutional neural network (CNN). Taking the TensorFlow framework as an example, a simple CNN model is constructed to detect the fault location and type of parts. Input layer: The receiving size is Color image (can be adjusted according to actual situation). Convolutional layer 1: use $32$ The convolution kernel of is , filled with , the activation function uses the ReLU function. The calculation formula is ,in It is Tier Line The output of the column, It is The convolution kernel weights of the layer, It is The input of the layer, is the bias. After this convolutional layer, the output feature map size is . Pooling layer 1: using The maximum pooling with a step size of , the output feature map size becomes . Subsequently, convolutional layers and pooling layers are added in sequence to deepen the network structure and extract more advanced features. For example, add another convolutional layer 2 and use 64 The convolution kernel of is , filled with , the activation function is ReLU, and the output feature map size is ; Pooling layer 2 also uses The maximum pooling with a step size of , the output feature map size is Fully connected layer: Flatten the output of the pooling layer into a one-dimensional vector, connect it to multiple neurons, and finally connect it to the output layer. The output layer sets the number of neurons according to the number of fault types, uses the Softmax function for classification, and outputs the probability of each fault type, thereby determining the fault location and type on the component, such as the location of stubborn attachments on the wheel tread, the location of cracks on the suspension arm, etc.

[0092] Path planning: Use Dijkstra algorithm to plan the laser processing path. Divide the area to be processed into grids, with each grid as a node. For each node, define the distance between it and the adjacent node (which can be set according to the actual cost of laser movement, such as the cost of straight-line movement is 1, and the cost of non-straight-line movement is appropriately increased). Initialize a distance array , used to store the minimum distance from the starting point to each node, the initial value is set to infinity, and the distance of the starting point is set to 0. Initialize a priority queue , used to store the nodes to be expanded, sorted from small to large distance. Add the starting point to the priority queue . Loop through the priority queue Take out the node with the smallest distance : If the node If it is the target node (i.e. the node where the fault is located), the path is found. By recording the predecessor node of each node, the path is obtained by backtracking. Otherwise, check the node Neighboring nodes :Calculate the number of nodes from the starting point To Node Distance ,in Is a node To Node If , then update , and the node The predecessor node of , the node Join the priority queue .

[0093] 2. Laser parameter adjustment based on adaptive control Sensor data acquisition: Use power sensor to monitor laser power, spot sensor to monitor laser spot size, and position sensor to monitor laser head position. Assume laser power is , the spot size is , the laser head position is .

[0094] Adopts adaptive proportional-integral-derivative (PID) control algorithm. The output of traditional PID control for ,in is the deviation (e.g. the difference between the expected laser power and the actual laser power), , , are proportional, integral, and differential coefficients. In adaptive PID, these coefficients are adjusted in real time according to sensor data. For example, when it is detected that the hardness of stubborn attachments on the wheel tread is large, fuzzy logic reasoning (similar to fuzzy control in cleaning components, a fuzzy rule base is established, and the fuzzy input is fuzzified according to factors such as attachment hardness and thickness, and the coefficient adjustment value is obtained by reasoning) is used to increase the value, so that the laser power increases rapidly to effectively remove stubborn attachments; at the same time, according to the spot size and position information, adjust the laser focusing parameters to ensure that the laser energy is concentrated on the fault location.

[0095] 2. Mechanical obstacle removal components 1. Robotic arm control based on kinematic model and optimization algorithm Kinematic model establishment: Taking the 6-DOF robot as an example, the D-H parameter method is used to establish the kinematic model. For each joint, four D-H parameters are defined: connecting rod length , connecting rod torsion angle , joint offset and joint angle . Through the homogeneous transformation matrix Represents the transformation relationship between two adjacent joint coordinate systems, . The pose of the end effector of the robot arm It can be obtained by multiplying the homogeneous transformation matrices of each joint: ,in .

[0096] - Optimization algorithm solution: When mechanical processing is required on wheels, suspension system components, etc., the optimization algorithm is used to solve the angles of each joint of the robot arm according to the target position and posture (such as determining the target position and posture of the cutting tool when cutting wheel cracks). The gradient descent method is used to define the objective function , such as the sum of squared errors between the actual position and posture of the end effector of the robot and the target position and posture. Calculate the gradient of the objective function for each joint angle , and then update the joint angles by iteration: ,in is the learning rate, is the number of iterations. Through continuous iterations, the end effector of the robot arm reaches the target position and posture, achieving precise mechanical processing of parts.

[0097] 2. Mechanical processing based on force feedback and intelligent decision-making Force feedback system: A force sensor is installed on the end effector of the robot arm to monitor the force during mechanical processing in real time. For example, when disassembling suspension system components, the force sensor can sense the magnitude and direction of the disassembly force. Assume that the force vector measured by the force sensor is .

[0098] Intelligent decision-making algorithm: Based on the force feedback data and preset processing rules, a finite state machine (FSM) is used for intelligent decision-making. Different states are defined, such as "preparing for disassembly", "disassembling", "encountering resistance", "disassembly completed", etc. When in the "disassembling" state, if the force detected by the force sensor exceeds the preset threshold, it enters the "encountering resistance" state. In the "encountering resistance" state, by analyzing the direction and magnitude of the force, combined with the structural information of the parts (such as the tightening torque of the bolts, the connection method of the parts, etc.), the cause of the resistance (such as rusted bolts, stuck parts, etc.) is determined, and then corresponding measures are taken, such as increasing the torque, changing the disassembly angle, etc., to achieve intelligent control of the mechanical obstacle removal process, ensure the smooth progress of mechanical processing and the safe disassembly and replacement of parts.

[0099] Optionally, an auxiliary positioning mark is provided at a first predetermined position of the parking lane, an auxiliary positioning device is provided at a second predetermined position of the parking lane, the first predetermined position and the second predetermined position form a facing relationship, and the method further includes: When the train group to be detected moves from the current position to the target position, the auxiliary positioning mark is detected by the auxiliary positioning device to generate a detection result, and the detection result is sent to the control module to perform auxiliary positioning on the train group to be detected from the current position to the target position.

[0100] Optionally, the auxiliary positioning mark includes at least one of an image mark, an ultrasonic material mark, and an infrared material mark, and the auxiliary positioning equipment includes at least one of an image recognition device, an ultrasonic transmitting device, and an infrared transmitting device.

[0101] Preferably, the creative implementation details of the above-mentioned technical processing steps based on auxiliary positioning marks are as follows: 1. Auxiliary positioning based on image marking and image recognition device 1. Image mark design: Image marks are designed with unique patterns and features, such as a QR code pattern or a pattern with a specific shape combination. This design makes the mark easy to identify and distinguish against a complex background. To ensure the accuracy and reliability of the mark, the pattern can be encoded, such as embedding location information about the positioning point, parking lane number, etc. in the QR code.

[0102] 2. Image recognition algorithm: Use OpenCV library and Python language to implement image recognition algorithm based on feature point matching. Taking SIFT (Scale Invariant Feature Transform) algorithm as an example, the specific steps are as follows: Scale space extreme value detection: Construct a Gaussian pyramid and perform Gaussian blur processing on the input image at different scales to obtain a series of images at different scales. Suppose the original image is , the image after Gaussian blur for: .

[0103] in is the scale parameter, and * represents the convolution operation. The DOG (Difference of Gaussian) image is obtained by subtracting adjacent scale images, and the extreme points in the DOG image are detected as potential feature points.

[0104] Feature point positioning and direction assignment: Accurately locate potential feature points and remove unstable edge response points. Calculate the gradient direction in the neighborhood of the feature point and use the main gradient direction as the direction of the feature point. Suppose the gradient of the image at point (x, y) is: .

[0105] Gradient Amplitude and direction for: , .

[0106] Feature descriptor generation: With the feature point as the center, sample in its neighborhood according to certain rules to generate a 128-dimensional feature descriptor. The feature descriptor contains the gradient information of the image around the feature point and is invariant to scale, rotation, and illumination.

[0107] Feature point matching: feature point matching is performed on the train image to be detected and the pre-stored image tag template. The Euclidean distance is used as the similarity metric to find the point in the image to be detected that is closest to the feature point descriptor in the template image as the matching point. Suppose the feature point descriptor of the template image is , the descriptor of the feature point of the image to be detected is , Euclidean distance for: .

[0108] in is the dimension of the feature descriptor. By matching a sufficient number of feature points, the RANSAC (Random Sampling Consensus) algorithm is used to remove mismatched points and obtain accurate matching point pairs.

[0109] 3. Positioning calculation: Based on the matching point pairs, use the perspective transformation matrix to calculate the position and posture of the train to be detected relative to the image marker. , matching points in the image to be detected , the perspective transformation matrix satisfy:

[0110] The perspective transformation matrix can be solved by at least 4 pairs of matching points According to the perspective transformation matrix and the known position of the image marker in the parking line, the position coordinates of the train set to be inspected in the parking line are calculated.

[0111] 2. Auxiliary positioning based on ultrasonic material marking and ultrasonic transmitting device 1. Ultrasonic material marking: A material with unique ultrasonic reflection characteristics is set as a marker at the first predetermined position of the parking line. This material can efficiently reflect ultrasonic signals, and its reflected signal has identifiable characteristics, such as specific reflection frequency or phase characteristics.

[0112] 2. Ultrasonic ranging principle: The ultrasonic transmitter transmits an ultrasonic pulse signal to the ultrasonic material marker and then receives the reflected signal. and the time difference between transmitting and receiving signals , calculate the distance between the ultrasonic transmitter and the ultrasonic material marker : .

[0113] In order to improve the distance measurement accuracy, multiple measurements can be taken to obtain the average value. The time difference between each measurement is , then the average distance for: .

[0114] 3. Positioning algorithm: If multiple ultrasonic material markers are set at different positions on the storage line, and the relative position relationship between these markers is known. The distance to multiple markers is measured by the ultrasonic transmitter, and the position of the EMU to be detected is determined by the triangulation positioning algorithm. Assuming that three ultrasonic material markers are known , , The coordinates of , , , Ultrasonic transmitter and marker , , The distances are , , . According to the equation of a circle: , , .

[0115] Solving the above equations together, we can get the coordinates of the ultrasonic transmitter (i.e. the train to be tested) in the plane of the storage line: .

[0116] (III) Auxiliary positioning based on infrared material marking and infrared emission device 1. Infrared material marking characteristics: Select materials with specific infrared emission or reflection characteristics as markers, such as materials with high emissivity or high reflectivity at specific wavelengths. Such markers can be clearly distinguished from the background in infrared images.

[0117] 2. Infrared image acquisition and processing: The infrared emitting device emits infrared light to the parking area, and the infrared camera is used to collect images containing infrared material marks. The collected infrared images are pre-processed, such as grayscale, filtering, etc., to enhance the contrast and clarity of the image. Suppose the infrared image is , the grayscale image It can be obtained by weighted average method: .

[0118] in , , are the red, green, and blue channel values ​​of the color infrared image, respectively. Then, the threshold segmentation algorithm is used to segment the infrared material mark from the background. Set the threshold to , the binary image after segmentation for: .

[0119] 3. Positioning calculation: Perform morphological operations on the segmented binary image, such as corrosion and expansion, to remove noise and fill the holes inside the mark. Use the contour detection algorithm to find the contour of the infrared material mark, and calculate the geometric center of the contour as the position of the mark. Let the point on the contour be , contour center coordinates for: , .

[0120] in is the number of points on the contour. Based on the known position of the infrared material marker in the parking line and its position in the image, combined with the camera imaging model (such as the pinhole camera model), the actual position of the train to be detected in the parking line is calculated. In the pinhole camera model, the point on the image plane With the point in the world coordinate system The relationship is:

[0121] in , is the focal length of the camera, are the coordinates of the image center, is the rotation matrix, is the translation vector. The position of the train to be detected can be calculated by using the known camera parameters and the position of the infrared material marker in the world coordinate system.

[0122] The innovations and technical benefits of the above specific technical steps are described as follows: 1. Image tagging and image recognition device Image tags with unique patterns and coded information, such as QR codes or specific shape combination patterns, are used, and information such as the location of the positioning point and the parking line number is embedded in them. This breaks through the limitation of traditional simple tags that are only used for identification, allowing the tags to carry more positioning-related information and provide a basis for subsequent precise positioning.

[0123] Using SIFT algorithm for image recognition, from scale space extreme value detection, feature point location and direction assignment, feature descriptor generation to feature point matching, each step is based on rigorous mathematical principles. For example, by constructing Gaussian pyramid and DOG image to detect extreme value points ( Used to generate Gaussian blurred images), so that the algorithm has scale invariance; by calculating the gradient direction to assign feature point directions ( The gradient amplitude and direction are calculated by formulas such as , which have rotation invariance; by generating a 128-dimensional feature descriptor and using Euclidean distance matching ( ), improving matching accuracy. These features enable the algorithm to accurately identify markers in complex environments, which is a significant innovation compared to traditional image recognition algorithms.

[0124] Using the perspective transformation matrix ( ) calculates the position and posture of the train to be detected relative to the image marker, solves the matrix through at least 4 pairs of matching points, and realizes the conversion from image coordinates to actual position coordinates, which provides an effective method for accurate positioning calculation.

[0125] The SIFT algorithm's scale, rotation, and illumination invariance, as well as the precise calculation of perspective transformation, enable accurate identification of image markers and calculation of the position of the EMU in a variety of complex environments, providing high-precision positioning for repairing train faults, ensuring that maintenance equipment can accurately reach the fault location, and improving maintenance efficiency. For example, accurate positioning can still be achieved when there are stains on the surface of the train or uneven lighting. The coded information embedded in the image marker can not only be used for positioning, but also provide relevant information such as the storage line, which makes it convenient for maintenance personnel to quickly understand the location of the train and related site information, and better plan the maintenance process.

[0126] 2. Based on ultrasonic material marking and ultrasonic transmitting device Materials with unique ultrasonic reflection properties (such as specific reflection frequency or phase characteristics) are used as markers. Compared with ordinary materials, they can reflect ultrasonic signals more efficiently and their reflected signals have identifiable characteristics, providing a better signal basis for accurate ranging.

[0127] In order to improve the ranging accuracy, the time difference between ultrasonic emission and reception signals is measured multiple times. , and using the formula Calculate the average distance, reduce single measurement errors, and improve ranging stability and accuracy.

[0128] The ultrasonic ranging is combined with the triangulation positioning algorithm to measure the distance to multiple ultrasonic material markers with known positions ( Calculate the distance), and use the circle equation to solve the train position ( The method realizes the transformation from one-dimensional distance measurement to two-dimensional position determination by solving the same equations, which innovates the traditional ultrasonic positioning method.

[0129] Ultrasonic signal propagation is not affected by factors such as light and dust, and has strong anti-interference ability in the complex environment of the parking line, ensuring the reliability of positioning. For example, it can still work normally at night or in a dusty environment.

[0130] The application of averaging multiple measurements and triangulation positioning algorithms has effectively improved the positioning accuracy, and can accurately determine the position of the EMU in the plane of the storage line, providing accurate position information for maintenance equipment.

[0131] 3. Based on infrared material marking and infrared emission device Materials with high emissivity or high reflectivity at specific wavelengths are selected as markers so that the markers can be clearly distinguished from the background in infrared images, providing a good foundation for subsequent image processing and positioning.

[0132] Complex infrared image processing flow: from grayscale conversion after image acquisition ( ), filtering to threshold segmentation ( ), then to morphological operations and contour detection, a series of image processing techniques are used to accurately extract marker position information.

[0133] Camera model combined with positioning: The marker position information obtained by infrared image processing is combined with the pinhole camera model ( ) calculates the actual position and realizes accurate conversion from image position to actual space position.

[0134] Infrared positioning technology has unique advantages at night or in temperature-sensitive scenarios. It can accurately identify markers and locate in these special environments, expanding the scope of application of the positioning system. For example, it can work effectively during nighttime maintenance or when detecting faults in heating components of trains. Through image processing flow and camera model calculation, the actual position of the EMU can be accurately obtained, providing a reliable location basis for train fault maintenance and improving the accuracy and efficiency of maintenance.

[0135] The present application also provides a real-time detection system for motor vehicle faults on a storage line, which includes: RFID reader: used to activate the onboard tag of the EMU to be tested, decode the attribute information of the EMU to be tested, and send it to the control module.

[0136] Pressure sensor: installed on the parking line, used to detect the pressure data on the parking line and send the data to the control module.

[0137] Control module: used to determine the vehicle structure template that matches the attribute information of the EMU to be tested in the vehicle structure template library, determine the current position of the EMU to be tested on the storage line based on the pressure data, and generate an EMU towing instruction and send it to the EMU towing device according to the matched vehicle structure template and the current position; it is also used to generate a fault inspection and maintenance instruction after receiving the feedback signal sent by the EMU towing device that the EMU to be tested is in a ready state, drive the fault inspection and maintenance combination installed on the storage line to perform fault detection on the EMU to be inspected according to the customized maintenance strategy under the drive of the mechanical arm, and repair the repairable faults.

[0138] EMU towing device: used to parse the EMU towing instruction, obtain the matched vehicle structure template and current position, determine the target position to which the EMU to be tested is to be moved when it is inspected on the storage line based on the matched vehicle structure template, drag the EMU to be tested from the current position to the target position, and generate a feedback signal to the control module that the EMU to be tested is in a ready state.

[0139] Fault inspection and repair combination and robotic arm: The fault inspection and repair combination is installed on the storage line. Driven by the robotic arm, it detects faults on the EMU to be inspected and repairs repairable faults according to the fault inspection and repair instructions generated by the control module and the customized inspection and repair strategy.

[0140] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for real-time detection of motor vehicle faults on a storage line, characterized in that: include: Park the EMU to be inspected on the storage line; Activate the train-mounted tag set on the train-mounted train to be detected based on the RFID reader / writer to decode the attribute information of the train-mounted train to be detected and send it to the control module; Detecting pressure data on the parking line based on a pressure sensor disposed on the parking line and sending the pressure data to the control module; The control module determines the vehicle structure template that matches the attribute information of the motor vehicle to be detected in the vehicle structure template library, and determines the current position of the motor vehicle group to be detected on the storage line based on the pressure data, so as to generate a motor vehicle towing instruction based on the matched vehicle structure template and the current position and send it to the motor vehicle towing device; The EMU towing device parses the EMU towing instruction to obtain the matched vehicle structure template and the current position, and determines the target position to which the EMU to be detected is to be moved when the EMU to be detected is inspected on the storage line based on the matched vehicle structure template, so as to drag the EMU to be detected from the current position to the target position, and generates a feedback signal to the control module that the EMU to be detected is in a ready state; The control module generates a fault repair instruction based on the triggering of the feedback model to drive the fault repair combination installed on the storage line to perform fault detection on the EMU to be repaired according to the customized repair strategy under the drive of the mechanical arm, and repair the repairable faults.

2. A method for real-time detection of motor vehicle faults on a storage line according to claim 1, characterized in that: The fault inspection and repair combination includes a fault detection module and a fault repair module, the mechanical arm includes a first mechanical arm and a second mechanical arm, the fault detection module is arranged on the first mechanical arm, the fault repair module is arranged on the second mechanical arm, the control module generates a fault inspection and repair instruction based on the triggering of the feedback model, so as to drive the fault inspection and repair combination installed on the storage line to perform fault detection on the EMU to be inspected according to the customized inspection and repair strategy under the drive of the mechanical arm, and repair the repairable fault, including: The control module generates a fault detection instruction based on the triggering of the feedback model to drive the fault detection module installed on the first mechanical arm to perform fault detection on the EMU to be repaired according to a customized detection strategy; When a fault is detected, the control module generates a fault repair instruction to drive the fault repair module installed on the second mechanical arm to repair the repairable fault according to a customized repair strategy.

3. A method for real-time detection of motor vehicle faults on a storage line according to claim 2, characterized in that: The fault detection module includes: an image detection component, an ultrasonic detection component, a waveform detection component, and at least one of a temperature detection component; The image detection component is used to perform the following steps to determine whether there are component failures at the bottom and side of the body of the EMU to be inspected: Capturing images of the bottom and / or side of the vehicle body of the EMU to be inspected, extracting features from the captured images, and determining the morphological changes of the bottom and side of the vehicle body accordingly; Calibrate the distance between the bottom and / or side of the vehicle body and the image detection assembly to determine the thickness change of the bottom and side of the vehicle body according to the calibrated distance; According to the morphological change and the thickness change, combined with the vehicle structure template of the EMU to be tested, it is determined whether key EMU components at the bottom and / or side of the body of the EMU to be tested are faulty.

4. A method for real-time detection of motor vehicle faults on a storage line according to claim 2, characterized in that: The waveform detection component is used to perform the following steps to determine whether there are any faulty components on the EMU to be detected: Obtain the jitter data of the train passing detected by the train jitter sensor on the train to be detected, Performing waveform analysis on the vehicle passing vibration data to extract waveform features therein; According to the waveform characteristics and in combination with the vehicle structure template of the EMU to be tested, it is determined whether there are any faulty components on the EMU to be tested.

5. The method for real-time detection of motor vehicle faults on a storage line according to claim 2, characterized in that: The temperature detection component is used to perform the following steps to determine whether there are any faulty components on the EMU to be detected; Performing non-contact temperature collection on set detection points on the body of the EMU to be detected; According to the collected temperature, starting from the set detection point and in combination with the vehicle structure template of the EMU to be detected, it is determined whether there are any faulty parts on the EMU to be detected.

6. A method for real-time detection of motor vehicle faults on a storage line according to claim 5, characterized in that: If the fault detection based on the image detection component, ultrasonic detection component, waveform detection component, and temperature detection component points to the same component, then the method further includes: performing individual confidence evaluations on the faulty components detected by the image detection component, ultrasonic detection component, waveform detection component, and temperature detection component, and fusing all individual confidence evaluations to obtain a fused confidence evaluation, so as to classify and confirm the fault of the component based on the fused confidence evaluation and the individual confidence evaluation.

7. A method for real-time detection of motor vehicle faults on a storage line according to claim 6, characterized in that: The confidence evaluation is performed on the faulty parts detected by the image detection component, the ultrasonic detection component, the waveform detection component, and the temperature detection component, respectively, including: Based on the following formula, the single confidence evaluation of the image detection component, the ultrasonic detection component, the waveform detection component, and the temperature detection component is determined: ,in, is the mapping function, is the weight value, , , n=1,2,3,4, ; All confidence evaluations are integrated to obtain a fused confidence evaluation, including: based on the following formula, all confidence evaluations are integrated to obtain a fused confidence evaluation: ,in, is the weight value, and the weight value is , determined based on prior knowledge; or, ,in, is a constant, is the weight value, which is determined based on prior knowledge.

8. A method for real-time detection of motor vehicle faults on a storage line according to claim 7, characterized in that: The fault repair module includes: at least one of a cleaning component, a heating component, a laser obstacle removal component, and a mechanical obstacle removal component; The cleaning component removes surface attachments, dirt, and grease from the following faulty parts: Wheels, axles, brake system components: remove oil, rust, and foreign matter from wheel treads, axle surfaces, brake discs, and brake calipers; Suspension system components: clean dust and oil on air springs, shock absorbers, suspension arms and other components; Vehicle side parts: clean the stains on doors, windows and side wall panels; Fasteners: remove rust and oil stains on fasteners such as bolts, nuts, and rivets; The heating assembly performs heating treatment on the following faulty parts: Electrical connection parts: Heat the wiring terminals and cable joints with poor contact to soften the insulation layer or solder, re-weld or tighten them to restore the electrical connection; Braking system components: locally heat the worn brake pads and repair or replace them after softening; Other metal parts: heating the parts that need welding repair, including broken suspension arms and bogie frames, and restoring their integrity through welding; The laser obstacle removal component performs laser processing on the following faulty parts: Wheels, axles, brake system components: remove stubborn attachments on wheel treads, axle surfaces, brake discs, and tiny fragments on crack edges; Suspension system components: Precisely cut and remove foreign matter and cracks on suspension arms, shock absorbers and other components; Body side parts: laser treatment of stubborn stains and scratches on doors, windows and side panels; The mechanical obstacle removal assembly is used to mechanically process the following faulty parts: Wheels and axles: Cut the cracks and broken parts on the wheels and replace them with new wheels or axle parts; Suspension system components: dismantle, replace or repair broken or deformed suspension arms, shock absorbers and other components; Bogie parts: repair cracks and deformations on the bogie frame, crossbeams and longitudinal beams, and replace damaged parts; Fasteners: Remove and replace loose or broken bolts, nuts, rivets and other fasteners to ensure the fastening effect.

9. A method for real-time detection of motor vehicle faults on a storage line according to claim 8, characterized in that: An auxiliary positioning mark is provided at a first predetermined position of the parking lane, an auxiliary positioning device is provided at a second predetermined position of the parking lane, the first predetermined position and the second predetermined position are in a facing relationship, and the method further includes: When the train group to be detected moves from the current position to the target position, the auxiliary positioning mark is detected by the auxiliary positioning device to generate a detection result, and the detection result is sent to the control module to perform auxiliary positioning on the train group to be detected from the current position to the target position.

10. A method for real-time detection of motor vehicle faults on a storage line according to claim 9, characterized in that: The auxiliary positioning mark includes at least one of an image mark, an ultrasonic material mark, and an infrared material mark, and the auxiliary positioning equipment includes at least one of an image recognition device, an ultrasonic transmitting device, and an infrared transmitting device.

Images