PHM system, method, medium and product of heavy haul railway group signal system

Abstract

The invention relates to the technical field of train control system monitoring, in particular to a PHM system and method of a heavy haul railway group signal system, a medium and a product, and the system comprises a data collection module which is configured to collect operation data of a target heavy haul railway group signal system; the machine learning module is configured to extract data features of the operation data based on the CNN network and the RNN network, and determine the operation state of the heavy haul railway group signal system; and determining the working strategy of the heavy haul railway group signal system based on the MADDPG. According to the PHM scheme of the heavy haul railway group signal system, a deep learning method in the field of artificial intelligence is closely combined with fault prediction and health management requirements of the heavy haul railway group signal system, signal system operation state monitoring is achieved based on the CNN-RNN network, and meanwhile working strategies of the heavy haul railway group signal system are determined based on the MADDPG.

Description

Technical Field

[0001] The present disclosure relates to the technical field of train control system monitoring, and particularly to a PHM system, method, medium and product for a heavy-haul railway group signal system. Background Art

[0002] In recent years, the heavy-haul railway transportation has developed rapidly. As an innovative means to improve the train operation efficiency, the train group technology has been applied in multiple lines and various scenarios. However, the group technology involves numerous devices and requires multi-disciplinary collaboration and multi-worker cooperation, which undoubtedly increases the difficulty of learning and experimentation. At present, for the fault prediction and health management of the heavy-haul railway signal system, traditional methods mainly rely on expert experience and simple mathematical models, and judge by threshold for the operating parameters of the signal system. When the parameters exceed the set threshold, it is considered that the device may have a fault. For example, by monitoring parameters such as the voltage and current of the switch machine, when the voltage or current exceeds the normal range, a fault warning is issued.

[0003] However, these traditional methods have obvious limitations. On the one hand, their fault prediction mainly relies on preset thresholds, and it is difficult to adapt to the complex and changeable operating environment and working conditions of the signal system, and it is easy to have misjudgment or missed judgment. On the other hand, traditional methods cannot comprehensively and accurately evaluate the health state of the signal system, cannot detect potential fault hazards in advance, and it is difficult to meet the requirements of high reliability and safety of the heavy-haul railway for the signal system. In addition, some existing simulation technologies often only consider the ground signal system when simulating the heavy-haul railway group signal system, and cannot realize the simulation of vehicle-ground signal communication and vehicle-vehicle signal communication, let alone handle the interaction of factors such as the track environment included in group train operation, and cannot truly reflect the actual operation of the signal system, resulting in low accuracy and reliability of the simulation results and unable to provide effective support for fault prediction and health management. Summary of the Invention

[0004] The present disclosure aims to at least partly solve one of the technical problems in the above technologies, and for this purpose, a PHM system for a heavy-haul railway group signal system is proposed, including: A data acquisition module, configured to: Collect the operation data of the target heavy-haul railway group signal system; A machine learning module, configured to: Extract the data features of the operation data based on the CNN network and the RNN network, and determine the operation state of the heavy-haul railway group signal system; and Determine the working strategy of the heavy-haul railway group signal system based on MADDPG.

[0005] Further, the machine learning module extracts data features of the operation data based on a CNN network and an RNN network, and the executed method includes: Extracting first data features of first operation data based on the CNN network; wherein, the first operation data is image data; the image data includes: appearance images of signal system devices, display state images of signal lamps, and surveillance videos; Extracting second data features of second operation data based on the RNN network; wherein, the second operation data is time series data; Fusing the first data features and the second data features to obtain target data features.

[0006] Further, the machine learning module is further configured to: Determine the operation state of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method.

[0007] Further, the machine learning module determines the operation state of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method, and the executed method includes: Selecting several types of monitoring data and respectively calculating the operation state grades of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method; Fusing the obtained several operation state grades to obtain a target operation state grade.

[0008] Further, the monitoring data includes: Equipment remaining life data, equipment actual monitoring data, and equipment failure history data.

[0009] Further, the machine learning module fuses the obtained several operation state grades to obtain a target operation state grade, and the executed method includes: Based on the D-S evidence fusion method and / or the grey relational analysis method, fusing the obtained several operation state grades to obtain a target operation state grade.

[0010] Further, the PHM system of the heavy-haul railway group signal system further includes: A simulation module, configured to: Perform simulation on the heavy-haul railway group signal system to provide a working environment for the training optimization and performance evaluation of the machine learning module.

[0011] The present disclosure also proposes a PHM method for a heavy-haul railway group signal system, including: Collecting operation data of a target heavy-haul railway group signal system; Extracting data features of the operation data based on a CNN network and an RNN network, and determining the operation state of the heavy-haul railway group signal system; and Determine the working strategy of the heavy-haul railway group signal system based on MADDPG.

[0012] Furthermore, extract the data features of the operation data based on the CNN network and the RNN network, including: Extract the first data feature of the first operation data based on the CNN network; wherein, the first operation data is image data; the image data includes: the appearance image of the signal system equipment, the display status image of the signal lamp, and the monitoring video. Extract the second data feature of the second operation data based on the RNN network; wherein, the second operation data is time-series data. Fuse the first data feature and the second data feature to obtain the target data feature.

[0013] Furthermore, the PHM method further includes: Determine the operation status of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method.

[0014] Furthermore, determine the operation status of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method, including: Based on the fuzzy comprehensive evaluation method, select several types of monitoring data and calculate the operation status level of the heavy-haul railway group signal system respectively. Fuse the obtained several operation status levels to obtain the target operation status level.

[0015] Furthermore, fuse the obtained several operation status levels to obtain the target operation status level, including: Based on the D-S evidence fusion method and / or the grey relational analysis method, fuse the obtained several operation status levels to obtain the target operation status level.

[0016] The present disclosure also proposes a computer-readable storage medium, in which a computer program or instruction is stored. When the computer program or instruction is executed by a processor, it is at least used to implement the above method.

[0017] The present disclosure also proposes a computer program product, which is stored in a computer-readable storage medium. When the computer program product is executed by a processor, it is at least used to implement the above method.

[0018] Compared with the prior art, the beneficial effects of the present disclosure are as follows: The PHM solution of the heavy-haul railway group signal system given by the present disclosure closely combines the deep learning method in the field of artificial intelligence with the fault prediction and health management requirements of the heavy-haul railway group signal system. It realizes the operation status monitoring of the signal system based on the CNN-RNN network, determines the working strategy of the heavy-haul railway group signal system based on MADDPG, and achieves efficient and accurate management of the heavy-haul railway group signal system by means of simulation. It gives full play to the powerful capabilities of deep learning in data analysis and pattern recognition, deeply analyzes the operation data of the heavy-haul railway group signal system, can accurately predict potential faults in advance, and evaluates the health status of the system in real time and dynamically. At the same time, the simulation technology is used to repeatedly verify and continuously optimize the fault prediction and health management strategies. By applying this technology, the potential fault hazards hidden in the signal system can be captured more accurately, greatly improving the safety and reliability of the operation of the heavy-haul railway signal system and laying a solid foundation for the safe and efficient operation of the heavy-haul railway.

[0019] Other features and advantages of the present disclosure will be described in the following specification, and some of them will be obvious from the specification, or can be understood by implementing the present disclosure. The objectives and other advantages of the present disclosure can be realized and obtained by the structures specifically pointed out in the written specification and the drawings. The technical solutions of the present disclosure will be further described below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings are used to provide a further understanding of the present disclosure, and constitute a part of the specification. They are used together with the embodiments of the present disclosure to explain the present disclosure, but do not constitute a limitation to the present disclosure. In the drawings: Figure 1 Schematic diagram of the PHM system of the heavy-haul railway group signal system given in the embodiment; Figure 2 Flowchart of the intelligent agent reinforcement learning given in the embodiment; Figure 3 Schematic diagram of the MADDPG algorithm given in the embodiment; Figure 4 Flowchart of the calculation of the health level of the switch machine given in the embodiment; Figure 5 Schematic diagram of the evaluation of the effect of turnout fault diagnosis given in the embodiment; Figure 6 Structural diagram of the adaptive enhanced convolutional neural network given in the embodiment; Figure 7 Schematic diagram of the computer-readable storage medium given in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The present disclosure will be described below in conjunction with the accompanying drawings. The preferred embodiments described herein are only used to illustrate and explain the present disclosure, and are not used to limit the present disclosure.

[0022] The present disclosure focuses on solving the short - board problems of the existing heavy - haul railway group signal system in the aspect of fault prediction and health management technology, and innovatively proposes a PHM (Prognostics and Health Management) technology for the heavy - haul railway group signal system based on artificial intelligence deep learning.

[0023] Figure 1 The PHM system of the heavy - haul railway group signal system given by the present disclosure includes: A data acquisition module, configured to: Collect the operation data of the target heavy - haul railway group signal system; A machine learning module, configured to: Extract the data features of the operation data based on the CNN network and the RNN network, and determine the operation state of the heavy - haul railway group signal system; and Determine the working strategy of the heavy - haul railway group signal system based on MADDPG.

[0024] Further, the method executed by the machine learning module for extracting the data features of the operation data based on the CNN network and the RNN network includes: Extract the first data features of the first operation data based on the CNN network; wherein, the first operation data is image data; the image data includes: the appearance image of the signal system equipment, the display state image of the signal lamp, and the monitoring video; Extract the second data features of the second operation data based on the RNN network; wherein, the second operation data is time - series data; Fuse the first data features and the second data features to obtain the target data features.

[0025] According to some embodiments of the present disclosure, the working principle of the machine learning module is as follows: 1. Use a convolutional neural network (CNN) to extract the operation data features of the signal system. CNN has a strong feature extraction ability and is used to process image and video data in the heavy - haul railway signal system, such as the appearance image of track equipment, the display state image of the signal lamp, etc. Through learning a large amount of image data, CNN can automatically extract key features and identify abnormal conditions of the equipment, such as damaged signal lamp bulbs, worn track components, etc.

[0026] 2. Using recurrent neural networks (RNNs) and their variants (LSTMs, GRUs) to extract the characteristics of the signal system operation data, RNNs are suitable for processing sequential data. In the heavy-haul railway signal system, train operation status data, signal transmission data, etc. all have sequential characteristics. As improved versions of RNNs, LSTMs and GRUs can effectively solve the problem of gradient disappearance in traditional RNNs, capture long-term dependencies in data more accurately, and are used to predict the state changes of the signal system during train operation, such as signal transmission delays, communication interruptions, etc.

[0027] According to some embodiments of the present disclosure, deep reinforcement learning continuously learns the optimal policy through the interaction between the agent and the environment. In the heavy-haul railway group signal system, the adjustment of the signal control policy can be regarded as the behavior of the agent, and indicators such as the operating efficiency and safety of the system are used as reward signals. Through the deep reinforcement learning algorithm, the agent can learn the optimal signal control policy under different working conditions, improving the transportation efficiency and safety of heavy-haul railways.

[0028] In this embodiment, the deep Q-network (DQN), multi-agent deep deterministic policy gradient (MADDPG), and proximal policy optimization (PPO) are adopted to determine the signal system control policy. At the same time, by integrating deep reinforcement learning (DRL) with traditional control theories (such as PID control, LQR control), the stability of traditional methods can be used as a benchmark policy, and the risk actions that may occur in the early exploration of DRL can be filtered through a reference mechanism, thereby improving the safety of actual deployment. For the group collaborative control problem, the MADDPG algorithm realizes the comprehensive optimization goal by designing a multi-dimensional reward function (such as weighted vehicle delay, length, driving speed, etc.), which is more adaptable to scenarios than traditional single-index optimization methods. For single-point control, DQN or PPO can be selected to balance efficiency and complexity.

[0029] In this embodiment, the train group operation algorithm based on multi-agent deep deterministic policy gradient (MADDPG) constructs a multi-agent dynamic interaction model, takes each train as an independent agent, integrates real-time perception data such as position, speed, acceleration, distance to the preceding vehicle, and signal status, and uses the Actor-Critic deep neural network architecture for centralized training to generate deterministic policies such as speed control and track switching, and realizes local real-time decision-making through distributed execution; the reward function integrates the goals of safety (safety distance constraint penalty), operating efficiency (time / energy consumption optimization), and fairness (balanced waiting time), combines PID control and reinforcement learning to achieve dynamic collision avoidance and signal collaborative optimization, and finally verifies its potential for conflict avoidance and improvement of operating efficiency compared with traditional fixed-block systems in complex road network scenarios in simulation platforms such as SUMO / Gazebo.

[0030] Further, the method for determining the working strategy of the heavy-haul railway group signal system based on MADDPG includes: Regarding the trains in the heavy-haul railway group signal system as agents; wherein, the train status information is used as the network input of the corresponding policy network of the agent, and the train control strategy is used as the network output of the corresponding policy network of the agent; Construct a reward function according to the operation requirements of the heavy-haul railway group signal system; Optimize the agent based on the reward function; Determine the working strategy of the heavy-haul railway group signal system according to the optimized agent.

[0031] Further, constructing the reward function according to the operation requirements of the heavy-haul railway group signal system includes: constructing the reward function based on the operation safety requirements of the signal system, the operation efficiency requirements of the signal system, and the train operation fairness requirements.

[0032] Further, the reward function at least includes a penalty term based on the train spacing, wherein the expression of the penalty term includes: R safe =-k·exp[(d min -d actual ) / d min wherein, R safe represents the safety distance penalty term; d min represents the minimum safety distance; d actual represents the current distance between the train and the preceding train; k represents the adjustment coefficient.

[0033] According to some embodiments of the present disclosure, when multiple trains operate on a complex railway, by designing a (MADDPG) dynamic interactive multi-agent environment model and performing centralized training to achieve collaborative decision-making, the speed, interval, and path are mainly adjusted to avoid conflicts and optimize the operation efficiency.

[0034] 1. Multi-agent system modeling Each train is an independent agent and has the ability of state perception (the sensor transmits back the state information). The state information includes the train position, speed, acceleration, distance from the preceding train, signal status, etc. The adjustable parameters include: adjusting the speed, switching the track, and requesting signals.

[0035] 2. Deep neural network Output a deterministic policy (such as a speed control instruction). The input is the global state (such as the positions of all trains) or the local state (such as the information of adjacent trains). Distributed execution is adopted, that is, during operation, each train makes independent decisions only relying on local observations, and stores the historical state-action-reward-next state (SARS’) samples.

[0036] ​3. Design Principles of Reward Function When designing the reward function, at least the following three points should be considered: (1) Safety, avoiding rear-end collisions or crashes (penalize if the safety distance is violated); In this embodiment, a safety distance penalty term is introduced. When the distance from the preceding train is less than the dynamic threshold, an exponentially increasing negative reward is triggered. The calculation formula is as follows: R safe =-k·exp[(d min -d actual ) / d min ) where R safe represents the safety distance penalty term; d min represents the minimum safety distance; d actual represents the current distance between the train and the preceding train; k represents the adjustment coefficient.

[0037] (2) Efficiency, minimizing time and energy consumption; In this embodiment, a two-dimensional reward mechanism is adopted. The time efficiency reward is inversely proportional to the deviation from the planned timetable, and the energy consumption efficiency reward is negatively correlated with the traction power loss.

[0038] (3) Fairness, balancing the waiting time of each train and avoiding local congestion.

[0039] In this embodiment, through the equalized waiting time algorithm, additional compensation rewards are given to the trains in the areas with long-term detention to avoid local congestion.

[0040] Furthermore, each train acts as an independent agent, and through sensors, it can perceive its own state (position, speed, acceleration) and environmental information (distance from the preceding train, signal light status, track topology) in real time. The Actor network (i.e., the policy network) of each agent generates deterministic actions (such as speed adjustment, track switching instructions) based on local observations. The Critic network (i.e., the evaluation network) evaluates the action value using global state information (positions of all trains, signal system data) during the centralized training phase. And, in this embodiment, a hybrid control mechanism fusion technology is adopted, combining PID control with reinforcement learning strategies. The PID controller processes basic speed tracking (such as speed limit constraints in sections), and the MADDPG policy layer superimposes dynamic optimization instructions (such as sudden obstacle avoidance, signal collaborative response), discretizing the continuous action space into executable track switching and acceleration level instructions. As Figure 2 shown, reinforcement learning enables the agent to continuously interact with the environment, and then continuously learn and optimize the strategy during the process of obtaining reward rewards, aiming to obtain the maximum cumulative reward. This interaction process is repeated continuously, and finally, the agent reaches the optimal strategy through the collected data.

[0041] Different from supervised learning and unsupervised learning, in the process of learning and interaction, the agent in reinforcement learning has no standard answers or pre-defined labels, but tries to take different actions to learn how to obtain the maximum reward in a given scenario.

[0042] Furthermore, the evaluation network corresponding to the agent is constructed based on the train operation state and the heavy-haul railway group signal system state.

[0043] According to some embodiments of the present disclosure, the agent's Critic network additionally accesses the signal system state (such as the occupancy of the block section and the locked state of the turnout), inputs the signal equipment state into the neural network, and realizes the two-way game optimization between the train group and the signal equipment.

[0044] According to some embodiments of the present disclosure, the SUMO / Gazebo is used to train and deploy the train working strategy based on the MADDPG algorithm, and the process is as follows: 1. Simulation environment construction Based on SUMO / Gazebo, a three-dimensional road network model is constructed and embedded in the "partially observable Markov game" framework: set the dynamic parameters of heterogeneous train types (high-speed rail / passenger train / freight train); simulate emergency failure scenarios (such as track intrusion and signal machine abnormality).

[0045] 2. Policy iteration optimization The MADDPG algorithm process: Each agent has a set of Actor and Critic, which output the action ai according to the observed state si to maximize the overall reward of the agent; the Critic is only used in the centralized training stage. In the training stage, the Critic evaluates and analyzes the action according to the action output by the Actor and feedbacks it to the Actor to realize the optimization of the Actor.

[0046] As Figure 3 shown, the MADDPG algorithm is composed of multiple agents respectively implementing a deep deterministic policy gradient algorithm. After all agents i obtain the current observation data si from the environment, their online policy network decides the action ai according to si and outputs it as the action of agent i. After multiple agents make decisions, they form an action set a. The environment is affected by the action set a, updates the current state to s′, and feedbacks the reward r = {r1, r2,..., ri}. Then, the obtained quadruple (si, ai, ri, si′) is stored in the experience replay pool for the next training need. The experience replay pool stores interaction data (state-action-reward tuples) of the order of 10^6, and the soft update coefficient τ of the target network is set to 0.01 to balance the convergence speed and stability.

[0047] 3. Online decision-making mechanism During the deployment phase, each train only retains the lightweight Actor network (with a parameter quantity <1MB) and achieves millisecond-level response through edge computing nodes. The observation information compression technology is adopted to input the original state data into the network after PCA dimensionality reduction, taking into account both computational efficiency and decision-making accuracy.

[0048] Further, the machine learning module is further configured to: Determine the operating state of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method.

[0049] Further, the machine learning module determines the operating state of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method, and the executed method includes: Based on the fuzzy comprehensive evaluation method, select several types of monitoring data and calculate the operating state levels of the heavy-haul railway group signal system respectively; Fuse the obtained several operating state levels to obtain the target operating state level.

[0050] Further, the monitoring data includes: Equipment remaining life data, equipment actual monitoring data, and equipment failure history data.

[0051] Further, the machine learning module fuses the obtained several operating state levels to obtain the target operating state level, and the executed method includes: Based on the D-S evidence fusion method and / or the grey relational analysis method, fuse the obtained several operating state levels to obtain the target operating state level.

[0052] The following is an illustration of this solution in combination with the fault prediction and health management of the switch machine: The switch machine is the actuator of the turnout control system, used for the conversion and locking of the turnout, and the supervision of the position and state of the turnout. At present, in China, the decisive factors for the maintenance and repair of turnouts are the practical experience of the staff themselves and the professional knowledge they master. Such fault analysis and fault location methods have many problems, such as low efficiency, large workload, inability to ensure timeliness and accuracy, etc. At present, the electric hydraulic ZYJ7 type is often used to complete the conversion and locking of the turnout position in high-speed railway and speed-up turnout sections. According to on-site research and understanding. The ZYJ7 type speed-up turnout is selected as the research object.

[0053] The microcomputer monitoring system monitors the action current of the switch with ZYJ-7 type electro-hydraulic switch machine. Signal workers can analyze the state of the switch conversion process based on the switch action current curve and power curve. However, the fault characteristics in the power curve are not obvious, and sometimes it is necessary to refer to the switch action current curve for further judgment. Therefore, it is more intuitive to use the action current curve to judge switch faults. The data to be used in this paper are all from the switch fault characteristic information provided by the on-site microcomputer monitoring system, and thus the research work on intelligent diagnosis of switch faults has been carried out. According to on-site research and practical work experience, the fault types of the research object are determined.

[0054] During the normal conversion process of the switch, the current curve of the speed-up switch of the ZYJ7 type electro-hydraulic switch machine can be roughly divided into three stages: the starting stage, the action stage, and the slow release stage.

[0055] In the starting stage of the switch, at the moment of starting, the motor needs to overcome a large resistance to drive the transmission device to operate. Therefore, the current formed in its starting circuit is relatively large, the current value rises suddenly and forms a peak, completing the unlocking process of the switch; in the action stage of the switch, the three-phase action current of the ZYJ7 type speed-up switch is generally between 4A and 5A. At this time, if the switch working state is normal, the switch machine steadily pushes the switch rail to move until it is in close contact with the stock rail and locks the switch rail. After locking, the switch control circuit cuts off the power supply, and the current curve drops rapidly. The action time of this stage is close to about 8s; in the slow release stage of the 1DQJ of the switch, there should be a "small step" composed of two-phase current curves in this stage. The formation of the "small step" is generated by the loop connected after locking. This loop connects the two-phase power supply terminals of the switch machine. Coupled with the slow release characteristic of the 1DQJ, the microcomputer monitoring will continue to collect the current curve of the switch. If the indication circuit is normal, the current of this "small step" is about between 0.5A and 0.6A.

[0056] Through on-site actual research and analysis, according to the microcomputer monitoring current curve of the switch operation of the ZYJ7 type electro-hydraulic switch machine, the typical fault modes can be roughly divided into four categories and several types. For its four typical fault types, namely typical idling fault, abnormal state reverse operation, abnormal resistance, and stuck notch, the fault causes are explained based on the composition and working principle of the switch.

[0057] ① Idling fault: Switch jamming is a fault with a high incidence rate. The fault cause may be mechanical jamming or internal jamming of the switch machine, such as too tight contact for non-locking, non-locking of the switch rail, imbalance of the inner indicating rod, etc. At this time, the switch often cannot be unlocked or locked, and the switch cannot complete the normal conversion action. The action time of the curve in the action area is extended and the current value does not drop back until the phase-breaking protector cuts off the switch action circuit and the current drops to zero, and no small step appears in the slow release area.

[0058] ② Abnormal state reverse operation: Abnormal state reverse operation refers to the situation where the reverse operation time of the switch is too short. This kind of fault generally occurs because the switch jams and idles during the previous conversion, that is, the switch was not in place completely before. If the switch jams during locking, resulting in the switch rail not being completely converted in place, then the time of the next reverse operation will be shortened. For the jamming situation during the unlocking process, there may not be an obvious change in the power curve in terms of time during the next reverse operation, but it may lead to a too high peak or abnormal peak state in the unlocking area due to difficult unlocking, such as the appearance of two peak values.

[0059] ③ Abnormal resistance fault: During the switch conversion process, the switch machine is relied on to push the switch rail to move. In this process, the switch machine needs to overcome various frictional resistances to do work. Therefore, the change of the external resistance will also be reflected in the conversion section of the power curve. The reasons for the abnormal increase in resistance can be roughly divided into the following types: the slide plate is not smooth, the switch rail moves on a rough or damaged slide plate, the geometric dimensions of the internal components of the switch are not square, etc. Through observation, it is found that abnormal resistance generally occurs between 1.5 s and 3 s. If there is an abnormal bulge in the curve during this time period, it indicates that there is resistance during the conversion or locking process, and in severe cases, it will lead to an idling fault.

[0060] ④ Switch stuck notch fault: The switch stuck notch fault is also called the slow release area missing fault, which is caused by the slow release characteristic of 1DQJ. If there is a situation where small steps are missing in the current curve, the circuit loop is not connected. The possible reasons include that the switch is not in place, resulting in the relevant indication relay not being energized, the diode being burned out and causing an open circuit, the protection resistance being open circuited, and the quick-acting switch contact being damaged. In this case, it is possible to check whether the positions of the detection rods in the switch machine and the detection rods of the locking detector are correct, whether the locking block is jammed, and whether the quick-acting switch group is jammed.

[0061] According to some embodiments of the present disclosure, the machine learning module determines the operating state of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method, the D-S evidence fusion method, and the grey relational analysis method. Figure 4 The figure is a schematic diagram for monitoring the operating state of the switch machine in the heavy-haul railway group signal system based on this method. The process includes: According to the algorithm and scoring rules, the health state of the switch machine is divided into 5 levels: "healthy", "sub-healthy", "abnormal", "morbid", and "faulty". Obtain the remaining life data of the switch machine components, the actual monitoring data, and the historical fault data. Calculate the time deterioration degree of the switch machine components, and further calculate the membership degree of the health state level based on the switch machine components based on the triangular fuzzy function according to the calculation results; based on the same method, calculate the membership degree of the health state level based on the monitoring data, and the membership degree of the health state level based on the historical fault data of the switch machine. Fuse the membership degrees of the three obtained health state levels according to the D-S evidence method to obtain the membership degree of the component health state, and further obtain the component evaluation membership degree vector; Based on grey relational analysis, synthesize the evaluation membership degree of the switch machine, and then obtain its health level.

[0062] According to some other embodiments of the present disclosure, evaluate the effect of turnout fault diagnosis based on an improved convolutional neural network, and the steps are as Figure 5 shown: First, collect the original samples of the current curve of the ZYJ7 type turnout, and extract the features of the current curve through Feature Extraction Method 1 and Feature Extraction Method 2 respectively, that is, extract the feature vectors according to the characteristic data of each state stage of the turnout action current curve, divide the original data of the turnout action current curve into N time sections according to time and extract the eigenvalue parameters in each time section to obtain two feature subsets, and input them into the adaptive boosting convolutional neural network and the original convolutional neural network for diagnosis respectively, and compare the diagnosis results of the two.

[0063] The network structure of the adaptive boosting convolutional neural network is as Figure 6 shown. Among them, Region 1 is the process of forward feature extraction and target classification, Region 2 is the process of backpropagation of residuals, and Region 3 is the adaptive boosting module. The input data obtains the classification result after the forward process. Each classification value corresponds to a unique category, and the category corresponding to the maximum value is recognized as the category to which the input belongs, while the category to which the input data actually belongs is called the true value category; the classification true value is the training supervision data, which stores the true classification result corresponding to the input data, where the value corresponding to the true value category is 1, and the values corresponding to the other categories are 0; the classification error is generated from the classification result and the classification true value. The CNN uses the target error function E(ω,b) to measure the learning effect of each parameter in the hidden layer on the input data. By adjusting the parameters of the hidden layer in the iterative process to reduce the output of the error function, the classification result y j is as close as possible to the classification true value y' j . When the error output of two adjacent times is not greater than the established threshold, it is considered to reach convergence and the learning is completed.

[0064] In summary, in this embodiment, the effect of the PHM system is evaluated by comparing different convolutional neural networks and different input sample data.

[0065] Furthermore, the PHM system of the heavy-haul railway group signal system further includes: A simulation module, configured to: Perform simulation on the heavy-haul railway group signal system to provide a working environment for the training optimization and performance evaluation of the machine learning module.

[0066] According to some embodiments of the present disclosure, constructing a virtual environment in the simulation of a heavy-haul railway group signal system involves multiple aspects such as track line, signal equipment simulation, and train operation simulation, and it is necessary to combine modeling techniques, dynamics algorithms, and digital twin platforms.

[0067] I. Simulation of Track Lines and Signal Equipment 1. 3D Modeling and Scene Construction Track line modeling: Use CAD or BIM tools (such as AutoCAD, Revit) to create a track geometry model, including parameters such as straight lines, curves, and gradients, and achieve modular integration through the OpenUSD format for easy reuse in multiple scenarios.

[0068] Signal equipment modeling: Perform 3D modeling on equipment such as signal lights, turnouts, and communication base stations, and design interaction logic in combination with the signal system logic (such as interlocking control, train automatic protection ATP). For example, simulate the relationship between signal switching and train response through discrete event simulation.

[0069] 2. System Dynamics and Signal Logic Simulation Dynamics model: Adopt a multi-body dynamics model to simulate the physical interaction between the track and the train, such as wheel-rail contact force, coupler force, etc., and combine the PhysX engine to achieve high-precision mechanical calculations.

[0070] Signal system simulation: Based on the state machine model, simulate the working process of signal equipment, such as the logic of stopping at a red light and passing at a green light, and integrate the real-time data synchronization function to ensure that the virtual signals are consistent with the actual equipment.

[0071] II. Simulation of Train Group Operation 1. Construction of Train Dynamics Model Multi-particle model: Decompose the train into multiple particles (such as carriages, bogies), establish a dynamics equation including traction force, braking force, and air resistance, and simulate the changes in acceleration and speed through numerical methods (such as Euler's method).

[0072] Motor and braking model: Integrate the characteristic curve of the traction motor (such as torque-speed relationship) and braking algorithms (such as electro-pneumatic combined braking) to dynamically adjust the train operation state.

[0073] 2. Integration of Group Driving Algorithm Operation control algorithm: Adopt PID control or model predictive control (MPC) to adjust the train speed and achieve functions such as automatic start and stop, and precise alignment. For example, preview control strategies under different working conditions in the digital twin platform.

[0074] Scene simulation: Test the emergency response ability of the train by injecting random events (such as sudden obstacles, weather changes), and use AI algorithms to optimize the scheduling strategy.

[0075] Based on the same technical concept, the present disclosure also proposes a PHM method for an overloaded railway group signal system, and the method includes: Using a data acquisition module, perform the following method: Collect the operation data of the target overloaded railway group signal system; Using a machine learning module, perform the following method: Extract the data features of the operation data based on the CNN network and the RNN network, and determine the operation state of the overloaded railway group signal system; and Determine the working strategy of the overloaded railway group signal system based on MADDPG; Using a simulation module, perform the following method: Perform a simulation on the overloaded railway group signal system to provide a working environment for the training optimization and performance evaluation of the machine learning module.

[0076] As Figure 7 shown, the present disclosure also proposes a computer-readable storage medium, in which a computer program or instruction is stored, and when the computer program or instruction is executed by a processor, it is at least used to implement the above method.

[0077] The present disclosure also proposes a computer program product, which is stored in a computer-readable storage medium, and when the computer program product is executed by a processor, it is at least used to implement the above method.

[0078] Obviously, those of ordinary skill in the art can make various changes and modifications to the present disclosure without departing from the spirit and scope of the present disclosure. Thus, if these modifications and variations of the present disclosure fall within the scope of the claims of the present disclosure and their equivalent technologies, the present disclosure also intends to include these changes and modifications.

Claims

1. A PHM system for a heavy-haul railway group signal system, characterized in that: include: The data acquisition module is configured as follows: Collect the operation data of the target heavy-haul railway group signal system; The machine learning module is configured to: Extracting data features of the operation data based on the CNN network and the RNN network to determine the operation status of the heavy-haul railway group signal system; and Determine the working strategy of heavy-haul railway group signal system based on MADDPG.

2. The PHM system of the heavy-haul railway group signal system according to claim 1, characterized in that: The machine learning module extracts data features of the operating data based on the CNN network and the RNN network, and the execution method includes: Extracting first data features of first operation data based on the CNN network; wherein the first operation data is image data; the image data includes: an appearance image of a signal system device, an image of a display state of a signal machine, and a monitoring video; Extracting a second data feature of second operating data based on the RNN network; wherein the second operating data is time series data; The first data feature and the second data feature are fused to obtain a target data feature.

3. The PHM system of the heavy-haul railway group signal system according to claim 1, characterized in that: The machine learning module is further configured to: Determine the operating status of heavy-haul railway group signal system based on fuzzy comprehensive evaluation method.

4. The PHM system of the heavy-haul railway group signal system according to claim 3, characterized in that: The machine learning module determines the operating status of the heavy-haul railway group signal system based on the fuzzy comprehensive evaluation method, and the execution method includes: Based on the fuzzy comprehensive evaluation method, several types of monitoring data are selected to respectively calculate the operating status level of the heavy-haul railway group signal system; The several operating status levels obtained are fused to obtain the target operating status level.

5. The PHM system of the heavy-haul railway group signal system according to claim 4, characterized in that: The monitoring data include: Equipment residual life data, equipment actual monitoring data and equipment failure history data.

6. The PHM system of the heavy-haul railway group signal system according to claim 4, characterized in that: The machine learning module fuses the several operating status levels obtained to obtain a target operating status level, and the execution method includes: The target operating status level is obtained by fusing several operating status levels obtained based on the DS evidence fusion method and / or the grey correlation analysis method.

7. The PHM system of the heavy-haul railway group signal system according to claim 1, characterized in that: Also includes: The simulation module is configured as: The heavy-load railway group signal system is simulated to provide a working environment for the training optimization and performance evaluation of the machine learning module.

8. A PHM method for a heavy-haul railway group signal system, characterized in that: include: Collect the operation data of the target heavy-haul railway group signal system; Extracting data features of the operating data based on the CNN network and the RNN network to determine the operating status of the heavy-haul railway group signal system; and Determine the working strategy of heavy-haul railway group signal system based on MADDPG.

9. The PHM method for a heavy-haul railway group signal system according to claim 8, characterized in that: The data features of the operating data are extracted based on the CNN network and the RNN network, including: Extracting first data features of first operation data based on the CNN network; wherein the first operation data is image data; the image data includes: an appearance image of a signal system device, an image of a display state of a signal machine, and a monitoring video; Extracting a second data feature of second operating data based on the RNN network; wherein the second operating data is time series data; The first data feature and the second data feature are fused to obtain a target data feature.

10. The PHM method for a heavy-haul railway group signal system according to claim 8, characterized in that: The PHM method further comprises: Determine the operating status of heavy-haul railway group signal system based on fuzzy comprehensive evaluation method.

11. The PHM method for a heavy-haul railway group signal system according to claim 10, characterized in that: The operating status of the heavy-haul railway group signal system is determined based on the fuzzy comprehensive evaluation method, including: Based on the fuzzy comprehensive evaluation method, several types of monitoring data are selected to respectively calculate the operating status level of the heavy-haul railway group signal system; The several operating status levels obtained are fused to obtain the target operating status level.

12. The PHM method for a heavy-haul railway group signal system according to claim 11, characterized in that: The target operating status level is obtained by integrating several operating status levels, including: The target operating status level is obtained by fusing several operating status levels obtained based on the DS evidence fusion method and / or the grey correlation analysis method.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program or instruction, and when the computer program or instruction is executed by the processor, it is used to implement at least the method of claim 8.

14. A computer program product, the computer program product being stored in a computer-readable storage medium, characterized in that: When the computer program product is executed by a processor, it is used to at least implement the method of claim 8.

Images