Method and system for monitoring quality of electric service vehicle-mounted equipment based on LMD system

Abstract

The invention discloses a method and system for monitoring the quality of electric service vehicle-mounted equipment based on an LMD system. The method comprises the following steps: S1, collecting data through the LMD system and transmitting the data to a ground system; s2, cleaning and generating multi-dimensional time sequence data; s3, analyzing the operation state by using an LSTM-Transform model, and generating an equipment health index; s4, generating a personalized monitoring rule through Bayesian optimization; s5, identifying an abnormal state through the graph neural network, marking early warning levels in a classified manner, and triggering an alarm; s6, dynamically planning and generating a resource scheduling plan; s7, a scheduling plan and alarm information are displayed, tasks are distributed, and maintenance feedback is uploaded; and S8, comparing the maintenance feedback with the health index, optimizing model parameters and updating rules. Through the intelligent monitoring and scheduling optimization technology, accurate monitoring of the operation state of the electric service vehicle-mounted equipment, fault prediction and efficient resource scheduling are achieved, and the safety and the intelligent level of equipment management are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent management of signal equipment, and particularly to a method and system for monitoring the quality of on-vehicle signal equipment based on the LMD system. Background Art

[0002] In recent years, with the rapid development of the railway transportation industry, the running speed of railway trains has been continuously increased, the mileage of train routes has been gradually extended, and the train operation density has increased significantly, posing higher requirements for the operation safety and management efficiency of railway signal equipment. As an important part of ensuring the safe operation of trains, on-vehicle signal equipment has become a key technology in railway locomotive operation. However, there are still many problems in the existing technologies for the status monitoring, maintenance management, and fault warning of on-vehicle signal equipment, making it difficult to meet the requirements of intelligent and efficient management in modern railway transportation.

[0003] In the existing management mode of on-vehicle signal equipment, data collection and transmission mostly rely on manual or semi-automatic methods. For example, the operation data of the LKJ system equipment is usually collected manually or through a local terminal, and then dumped and analyzed. Due to the lack of real-time performance and efficiency of this method, it is impossible to dynamically monitor and warn the status during the operation of the equipment. Especially for locomotive equipment scattered on different lines, the existing monitoring means are difficult to timely grasp the technical status and operation trend of the equipment, and cannot effectively respond to sudden failures and abnormal situations. In addition, traditional status monitoring methods usually adopt fixed-threshold alarm rules, and such static rules are difficult to adapt to the dynamic changes of the equipment operation environment and load, which may lead to missed alarms or false alarms, further increasing the potential safety hazards.

[0004] At the same time, there is generally a lack of the ability to integrate and deeply mine multi-modal data in the existing technologies. A large amount of multi-dimensional data is generated during the operation of on-vehicle signal equipment, including sensor information such as voltage, current, temperature, and vibration, as well as GPS location information and historical operation data. However, these data are often stored in isolation, lacking a unified platform for integrated analysis, resulting in the inability to effectively mine the correlation information between the data. This data island phenomenon not only affects the accurate assessment of the equipment status but also restricts the intelligent process of fault prediction and maintenance management.

[0005] In terms of resource scheduling and fault response, the existing systems also have a relatively low level of intelligence. Currently, the allocation and scheduling of maintenance resources mostly rely on manual decision-making or empirical rules, making it difficult to perform efficient optimization scheduling based on real-time alarm information, equipment priorities, and resource locations. This mode not only increases the scheduling time and operating costs but also may lead to the risk of fault expansion due to untimely responses. In addition, due to the lack of a perfect GIS visualization platform, there are still problems such as non-intuitive information transmission and inaccurate task allocation in the display and execution of scheduling plans, further affecting the efficiency of fault handling.

[0006] Another significant defect of the existing technologies is the insufficient utilization of historical data. During the operation of the on-vehicle equipment of the signal department, a large amount of historical fault data and maintenance records have been accumulated, and these data contain rich patterns and rules. However, the existing systems lack effective models and algorithms for deep learning and analysis of these historical data and cannot make full use of these data to support equipment health management and rule optimization. In addition, the lack of the ability to conduct linkage analysis of historical data and real-time data also makes it difficult for the existing systems to achieve dynamic optimization and closed-loop management of equipment status.

[0007] Therefore, how to provide a method and system for monitoring the quality of on-vehicle equipment of the signal department based on the LMD system is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a method and system for monitoring the quality of on-vehicle equipment of the signal department based on the LMD system. The present invention fully combines multi-dimensional equipment operation data analysis, personalized rule optimization, fault status classification, and resource scheduling optimization technologies, and details the real-time monitoring of equipment operation status, the generation and prediction of health indexes, intelligent resource scheduling, and the closed-loop optimization process of task feedback, having the advantages of high monitoring accuracy, fast fault response, efficient resource allocation, and intelligent system operation.

[0009] The method for monitoring the quality of on-vehicle equipment of the signal department based on the LMD system according to the embodiments of the present invention includes the following steps:

[0010] S1. Collect equipment data through the LMD system on the locomotive, and combine with the GPS and Beidou systems to transmit the data to the ground system through the MTUP platform of the railway intranet to generate multi-dimensional equipment operation data;

[0011] S2. Preprocess the equipment data transmitted through the MTUP platform to generate multi-dimensional time series data;

[0012] S3. Combine the multi-dimensional equipment operation data and multi-dimensional time series data, and use the LSTM-Transformer hybrid model to analyze the equipment operation status to generate an equipment health index;

[0013] S4. Combine multi-dimensional device operation data, multi-dimensional time series data, and device health index, and use the Bayesian optimization algorithm to dynamically adjust the monitoring rule parameters and alarm logic to generate personalized monitoring rules;

[0014] S5. Use the personalized monitoring rules to identify the abnormal state of the device through the graph neural network, classify and label the abnormal state of the device according to the warning level, and trigger alarm information;

[0015] S6. Combine device data and alarm information, analyze task priorities, resource distribution, and device locations based on the dynamic programming model, and generate a resource scheduling plan;

[0016] S7. Use the GIS module of the LMD system to display the resource scheduling plan and alarm information in real time, dispatch tasks to maintenance personnel, and upload feedback after maintenance is completed;

[0017] S8. Compare the maintenance feedback information with the device health index and alarm information, optimize the parameters of the LSTM-Transformer hybrid model, and update the device status history record and monitoring rules.

[0018] Optionally, the specific steps of S3 include:

[0019] S31. Combine the multi-dimensional device operation data X = {x 1 , x 2 ,..., x t} and the multi-dimensional time series data Z = {z 1 , z 2 ,..., z t} into the input data I, where I = {I 1 , I 2 ,..., I t}, representing the comprehensive data combining device physical parameters and time series characteristics;

[0020] S32. Input the input data I into the LSTM module for time-dependent modeling, and the hidden state is H LSTM = {h 1 , h 2 ,..., h t}, specifically:

[0021] h t = f(h t-1 , I t ; θ LSTM );

[0022] where θ LSTM are the training parameters of the LSTM, h t is the time series hidden feature at time t, and h t-1is the time series implicit feature at time t-1, and f is the recursive function of the LSTM;

[0023] S33. Input H LSTM into the Transformer module and calculate the global feature through the multi-head self-attention mechanism:

[0024] H Trans = Concat(head 1 , head 2 ,..., head n )·W O ;

[0025] where:

[0026]

[0027] where H Trans is the global feature, Concat is to splice the output features of n attention heads, W O is the weight matrix of the output layer, head i is the output feature of the i-th attention head, head n is the output feature of the n-th attention head, W Q,i , W K,i , W V,i is the weight matrix of the i-th attention head, is the feature dimension normalization factor, and Softmax is the normalization function;

[0028] S34. Map the global feature H Trans output by the Transformer to the device health index HI:

[0029] HI = σ(W 1 ·H Trans + b 1 );

[0030] where σ is the activation function, W 1 and b 1 are the linear mapping parameters respectively;

[0031] S35. Combine the exponential smoothing and weighted sliding window algorithms to smooth the device health index HI and enhance the trend:

[0032]

[0033] where HI smooth is the processed device health index, HI t is the health index at the current time t, HI t-iis the health index at the i-th moment in the past, α is the exponential smoothing factor, and w i is the weighting coefficient of historical data, and N is the size of the sliding window;

[0034] S36. Output the processed device health index HI smooth and predict the potential failure risk and time window in combination with the historical trend of device operation.

[0035] Optionally, the S4 specifically includes:

[0036] S41. Input the multi-dimensional device operation data X = {x 1 , x 2 ,..., x t}, the device health index HI smooth , and the historical failure mode data H = {H 1 , H 2 ,..., H N} accumulated by the LMD system through long-term monitoring into the rule optimization module as the basic input for monitoring rule optimization;

[0037] S42. Use the Bayesian optimization algorithm to extract key failure modes from the historical data H and calculate the posterior distribution of the monitoring rule parameter θ:

[0038]

[0039] where P(θ∣H, HI smooth ) is the joint probability of the historical failure mode data and the health index, θ is the set of monitoring rule parameters, P(θ) is the prior distribution of the rule parameters, and Θ is the search space of the rule parameters;

[0040] S43. Optimize the rule parameters by the maximum a posteriori probability method:

[0041] θ opt = argmax θ P(H, HI smooth ∣θ)·P(θ);

[0042] where θ opt is the optimized monitoring rule parameter, and argmax θ is the parameter value corresponding to the maximum a posteriori probability of θ;

[0043] S44. Combine the device health index HI smooth and the optimized monitoring rule parameter θ opt , and dynamically adjust the rule parameters through the exponential weight mechanism and linear correction to generate the real-time optimized rule parameter θ real-time :

[0044]

[0045] Among them, μ is the exponential adjustment factor, ∈ is the linear adjustment factor, e is the base of the natural logarithm, and HI t is the real-time health index of the device at the current moment;

[0046] S45. According to the rule parameter θ optimized in real time real-time , optimize the alarm logic through non-linear combination

[0047]

[0048] Among them, is the current alarm logic, is the updated alarm logic, θ prev is the rule parameter in the previous cycle, δ is the smoothing parameter to avoid the denominator being zero, ρ is the exponential amplification factor, τ is the alarm logic update factor, and ln is the natural logarithm;

[0049] S46. Update the rule parameter θ optimized in real time real-time and the alarm logic to the rule library to generate personalized monitoring rules.

[0050] Optionally, the specific content of S5 includes:

[0051] S51. Input the rule parameter θ optimized in real time real-time and the alarm logic into the status classification module as the basic basis for abnormal status recognition;

[0052] S52. Combine the device health index HI smooth and the multi-dimensional device operation data X = {x 1 , x 2 ,..., x t} to construct a device operation feature graph G = (V, E);

[0053] S53. Input the device operation feature graph G and the rule parameter θ optimized in real time real-time into the graph neural network, extract the deep relationship between features through the message passing mechanism, and adjust the feature embedding weights in combination with the rule parameters to generate the device operation status feature embedding H = {h 1 , h 2 ,..., h m}:

[0054]

[0055] Among them, is the feature embedding of vertex v i at the k-th layer, is the vertex vi The feature embedding at the (k + 1)-th layer, where N(i) is the set of neighbors of vertex v i and deg(i) is the degree of vertex v i , deg(j) is the degree of vertex v j , W (k) is the weight matrix of the k-th layer of the graph neural network, θ is the activation function, and j is the neighbor node of vertex v i , w ij is the dynamic weight of the relationship between vertex v i and v j ;

[0056] S54. Using the device operation status feature embedding H, combined with the device health index HI smooth and the rule parameter θ real-time optimized in real time, generate the status classification result S through a non-linear classification model with attention weights:

[0057] s i = sigmoid(W s ·ω(∑ j∈N(i) α ij ·(h j + θ real-time ·h i )) + HI smooth ·δ + b s );

[0058] where s i is the classification result of node i, sigmoid maps the classification result to a probability value, W s is the weight matrix of the classification layer, ω is the non-linear activation function, α ij is the attention weight of neighbor node j to central node i, h i and h j are the embedding features of node i and neighbor node j, δ is the influence factor of the device health index, and b s is the bias term of the classification layer;

[0059] S55. According to the status classification result S, combined with the device health index HI smooth , the rule parameter θ real-time optimized in real time, and the device operation status feature embedding H, calculate the alarm level A of the abnormal event:

[0060]

[0061] where λ 1 is the health index weight factor, λ 2 is the weighting factor of the status classification result, λ 3is the rule parameter weight factor, K is the total number of abnormal states detected, and s k is the k-th classification result, max(HI smooth ,H is the equipment health index HI smooth and the larger value of the equipment operation status feature embedding H;

[0062] S56. Mark the alarm level A of the abnormal event, classify it into multiple alarm categories according to preset classification rules, and trigger corresponding alarm information.

[0063] Optionally, the S6 specifically includes:

[0064] S61. Combine the alarm level A of the abnormal event and the real-time geographic location L of the device i , construct the objective function of the dynamic programming model:

[0065]

[0066] Among them, C is the total cost of dispatch, N is the total number of alarm devices to be dispatched, and D i is the geographical distance from the resource center to the alarm device i, T i The time required for device i to respond, γ 1 ,γ 2 ,γ 3 is the weight factor, Minimize is to minimize the objective function C;

[0067] S62, through the real-time location of the alarm device i (x i ,y i ,z i ) and resource center location (x r ,y r ,z r ), combined with the path complexity weight w d and the vertical distance weight ω p , calculate the geographical distance D i :

[0068]

[0069] S63. Dynamically estimate the response time of the alarm device, taking into account geographic distance, road congestion, resource movement speed and historical response data:

[0070]

[0071] Among them, v r is the average speed of resource scheduling, C i is the real-time road congestion factor, ω a is the impact weight of the historical correction term, mean(T history) is the mean of historical response times, std(T history ) is the standard deviation of historical response times, is a small positive number to prevent the denominator from being zero, T i is the response time of alarm device i;

[0072] S64. Dynamically calculate the priority weight by combining the alarm level, geographical location, and historical event frequency of abnormal events:

[0073]

[0074] Among them, P i is the priority weight of device i, exp is the exponential function, α 1 is the attenuation factor of geographical distance, f i is the historical event frequency of device i, f j is the historical event frequency of device j, N is the total number of all alarm devices, A i is the alarm level of device i, A j is the alarm level of device j, D i is the geographical distance from device i to the resource center, D j is the geographical distance from device j to the resource center;

[0075] S65. Generate a resource scheduling plan using a dynamic programming model based on the priority weight P i , the response time T of the alarm device i and the real-time geographical location of the device:

[0076]

[0077] Among them, C total is the total cost of the scheduling plan, min is the optimization objective, γ 1 is the weight of the distance factor, γ 2 is the weight of the response time factor, γ 3 is the weight of the priority factor;

[0078] S66. Output the resource scheduling plan, including task priority, resource allocation, path, and completion time.

[0079] The system for monitoring the quality of on-board signaling equipment based on the LMD system according to the embodiment of the present invention includes the following modules:

[0080] Data acquisition and transmission module, used to collect multi-dimensional data from device sensors and obtain the real-time geographical location information of the device by combining the vehicle-mounted GPS and Beidou systems;

[0081] Data processing and analysis module, used to clean, fill in missing values, align time, normalize, and extract features from the collected data;

[0082] A rule generation and optimization module, which is used to dynamically adjust monitoring rules and alarm logic through the Bayesian optimization method;

[0083] A status monitoring and alarm module, which is used to monitor the running status in real time through a classification model, identify abnormalities and generate alarm information;

[0084] A resource scheduling and path optimization module, which is used to synthesize the alarm level of the device, the real-time geographical location and the response time, and generate a resource scheduling plan by using a dynamic programming model;

[0085] A GIS visualization and task execution module, which is used to display the positions of alarm devices, scheduling paths and task statuses in real time through the GIS module, and dispatch tasks to maintenance personnel;

[0086] A historical data management and model optimization module, which is used to store the historical operation data of the device, alarm records and maintenance feedback, and continuously optimize the performance of the system model through offline training and deployment.

[0087] The beneficial effects of the present invention are as follows:

[0088] By combining multi-dimensional device operation data, real-time monitoring technology and optimization algorithms, the present invention effectively solves the problems existing in the prior art, such as non-real-time device status monitoring, serious data island phenomenon, low scheduling response efficiency, etc., and realizes the intelligent management and optimized scheduling of on-board signal equipment. By using the LMD system to collect multi-dimensional status data and combining the health index generation and fault prediction model, the running status of the device can be grasped in real time and potential faults can be predicted, so as to give early warnings and reduce the safety risks brought by sudden faults. The personalized rule optimization algorithm further improves the adaptability of the alarm rules and effectively reduces the false alarms and missed alarms caused by traditional static rules. The resource scheduling module comprehensively considers multiple factors such as the alarm level, device location and historical frequency, and uses dynamic programming to generate the optimal scheduling path, greatly improving the accuracy and efficiency of scheduling. Through the GIS visualization display and task feedback function, the present invention not only realizes the visualization of task allocation, but also forms a closed-loop optimization process from device status monitoring to task execution, ensuring the high efficiency and reliability of device management. In addition, the present invention makes full use of historical operation data, combines offline model training and real-time optimization, improves the adaptability of the system in multiple scenarios, and further promotes the improvement of the intelligent operation and maintenance level of railway signal equipment, having significant advantages in safety, accuracy and economy. Description of the Drawings

[0089] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0090] Figure 1 Flow chart of the method and system for monitoring the quality of on-board signaling equipment based on the LMD system proposed by the present invention;

[0091] Figure 2 Schematic structural diagram of the system for monitoring the quality of on-board signaling equipment based on the LMD system proposed by the present invention. Specific embodiments

[0092] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0093] Reference Figure 1 , the method for monitoring the quality of on-board signaling equipment based on the LMD system includes the following steps:

[0094] S1. Collect equipment data through the LMD system on the locomotive, combine with the GPS and Beidou systems, and transmit the data to the ground system through the MTUP platform on the railway intranet to generate multi-dimensional equipment operation data;

[0095] S2. Preprocess the equipment data transmitted through the MTUP platform to generate multi-dimensional time series data;

[0096] S3. Combine the multi-dimensional equipment operation data and multi-dimensional time series data, and use the LSTM-Transformer hybrid model to analyze the equipment operation status to generate an equipment health index;

[0097] S4. Combine the multi-dimensional equipment operation data, multi-dimensional time series data and equipment health index, and use the Bayesian optimization algorithm to dynamically adjust the monitoring rule parameters and alarm logic to generate personalized monitoring rules;

[0098] S5. Use the personalized monitoring rules to identify the abnormal status of the equipment through the graph neural network, classify and label the abnormal status of the equipment according to the warning level, and trigger alarm information;

[0099] S6. Combine the equipment data and alarm information, and analyze the task priority, resource distribution and equipment location based on the dynamic programming model to generate a resource scheduling plan;

[0100] S7. Use the GIS module of the LMD system to display the resource scheduling plan and alarm information in real time, dispatch tasks to maintenance personnel, and upload feedback after maintenance;

[0101] S8. Compare the maintenance feedback information with the equipment health index and alarm information, optimize the parameters of the LSTM-Transformer hybrid model, and update the equipment status history record and monitoring rules.

[0102] In this embodiment, step S3 specifically includes:

[0103] S31. Merge the multi-dimensional device operation data X = {x 1 , x 2 ,..., x t} and the multi-dimensional time series data Z = {z 1 , z 2 ,..., z t} into the input data I, where I = {I 1 , I 2 ,..., I t}, representing the comprehensive data combining device physical parameters and time series characteristics;

[0104] S32. Input the input data I into the LSTM module for time-dependence modeling, with the hidden state H LSTM = {h 1 , h 2 ,..., h t}, specifically:

[0105] h t = f(h t-1 , I t ; θ LSTM );

[0106] where θ LSTM is the training parameter of the LSTM, h t is the time series hidden feature at time t, h t-1 is the time series hidden feature at time t - 1, and f is the recursive function of the LSTM;

[0107] S33. Input H LSTM into the Transformer module, and calculate the global feature through the multi-head self-attention mechanism:

[0108] H Trans = Concat(head 1 , head 2 ,..., head n ) · W O ;

[0109] where:

[0110]

[0111] where H Trans is the global feature, Concat is to splice the output features of n attention heads, W O is the weight matrix of the output layer, and head i is the output feature of the i-th attention head, headn is the output feature of the n-th attention head, W Q,i , W K,i , W V,i is the weight matrix of the i-th attention head, is the feature dimension normalization factor, and Softmax is the normalization function;

[0112] S34. Map the global feature H Trans output by the Transformer to the device health index HI:

[0113] HI = σ(W 1 ·H Trans + b 1 );

[0114] where σ is the activation function, W 1 and b 1 are the linear mapping parameters respectively;

[0115] S35. Combine the exponential smoothing and weighted sliding window algorithms to smooth the device health index HI and enhance the trend:

[0116]

[0117] where HI smooth is the processed device health index, HI t is the health index at the current time t, HI t-i is the health index at the i-th past time, α is the exponential smoothing factor, w i is the weighted coefficient of historical data, and N is the sliding window size;

[0118] S36. Output the processed device health index HI smooth and combine the historical trend of device operation to predict the potential failure risk and time window.

[0119] In this embodiment, the specific steps of S4 include:

[0120] S41. Input the multi-dimensional device operation data X = {x 1 , x 2 ,..., x t}, the device health index HI smooth , and the historical failure mode data H = {H 1 , H 2 ,..., H N} accumulated by the LMD system through long-term monitoring into the rule optimization module as the basic input for monitoring rule optimization;

[0121] S42. Use the Bayesian optimization algorithm to extract key fault modes from historical data H and calculate the posterior distribution of the monitoring rule parameters θ:

[0122]

[0123] where P(θ∣H,HI smooth ) is the joint probability of historical fault mode data and health index, θ is the set of monitoring rule parameters, P(θ) is the prior distribution of the rule parameters, and Θ is the search space of the rule parameters;

[0124] S43. Optimize the rule parameters by the maximum a posteriori probability method:

[0125] θ opt = argmax θ P(H,HI smooth ∣θ)·P(θ);

[0126] where θ opt is the optimized monitoring rule parameter, and argmax θ is the parameter value corresponding to the maximum a posteriori probability of θ;

[0127] S44. Combine the equipment health index HI smooth and the optimized monitoring rule parameter θ opt , and dynamically adjust the rule parameter through the exponential weight mechanism and linear correction to generate the real-time optimized rule parameter θ real-time :

[0128]

[0129] where μ is the exponential adjustment factor, ∈ is the linear adjustment factor, e is the base of the natural logarithm, and HI t is the real-time health index of the equipment at the current moment;

[0130] S45. According to the real-time optimized rule parameter θ real-time , optimize the alarm logic through non-linear combination

[0131]

[0132] where is the current alarm logic, is the updated alarm logic, θ prev is the rule parameter of the previous cycle, δ is the smoothing parameter to avoid the denominator being zero, ρ is the exponential amplification factor, τ is the alarm logic update factor, and ln is the natural logarithm;

[0133] S46. The real-time optimized rule parameter θ real-time and the alarm logic Update to the rule library to generate personalized monitoring rules.

[0134] In this embodiment, S5 specifically includes:

[0135] S51. Input the rule parameters θ real-time optimized in real time and the alarm logic

[0136] into the status classification module as the basic basis for abnormal status recognition; smooth S52. Combine the device health index HI 1 and the multi-dimensional device operation data X = {x 2 , x t ,..., x

[0137] } to construct a device operation feature graph G = (V, E); real-time S53. Input the device operation feature graph G and the rule parameters θ 1 optimized in real time 2 into the graph neural network, extract the deep relationship between features through the message passing mechanism, and adjust the feature embedding weights in combination with the rule parameters to generate the device operation status feature embedding H = {h m}:

[0138]

[0139] where is the feature embedding of vertex v i at the k-th layer, is the feature embedding of vertex v i at the (k + 1)-th layer, N(i) is the neighbor set of vertex v i , deg(i) is the degree of vertex v i , deg(j) is the degree of vertex v j , W (k) is the weight matrix of the k-th layer of the graph neural network, θ is the activation function, j is the neighbor node of vertex v i , and w ij is the dynamic weight of the relationship between vertex v i and v j ;

[0140] S54. Use the device operation status feature embedding H, combine the device health index HI smooth and the rule parameters θ real-time optimized in real time

[0141] s i = sigmoid(W s ·ω(∑j∈N(i) α ij ·(h j +θ real-time ·h i )) + HI smooth ·δ + b s );

[0142] Among them, s i is the classification result of node i, sigmoid is to map the classification result to a probability value, W s is the weight matrix of the classification layer, ω is the non-linear activation function, α ij is the attention weight of neighbor node j to central node i, h i and h j are the embedding features of node i and neighbor node j, δ is the influence factor of the device health index, b s is the bias term of the classification layer;

[0143] S55. According to the status classification result S, combined with the device health index HI smooth , the rule parameter θ real-time optimized in real time, and the device operating status feature embedding H, calculate the alarm level A of the abnormal event:

[0144]

[0145] Among them, λ 1 is the health index weight factor, λ 2 is the weighted factor of the status classification result, λ 3 is the rule parameter weight factor, K is the total number of detected abnormal states, s k is the k-th classification result, max(HI smooth , H is to take the larger value of the device health index HI smooth and the device operating status feature embedding H;

[0146] S56. Label the alarm level A of the abnormal event, divide it into multiple alarm categories according to the preset classification rules, and trigger the corresponding alarm information.

[0147] In this embodiment, the S6 specifically includes:

[0148] S61. Combine the alarm level A of the abnormal event and the real-time geographical location L i of the device to construct the objective function of the dynamic programming model:

[0149]

[0150] Among them, C is the total cost of scheduling, N is the total number of alarm devices to be scheduled, D i is the geographical distance from the resource center to alarm device i, Ti The time required for the response device i, γ 1 , γ 2 , γ 3 is the weight factor, and Minimize means taking the minimum value of the objective function C;

[0151] S62. Through the real-time position (x i , y i , z i ) of the alarm device i and the position of the resource center (x r , y r , z r ), combined with the path complexity weight w d and the vertical distance weight ω p , calculate the geographical distance D i :

[0152]

[0153] S63. Dynamically estimate the response time of the alarm device, combining geographical distance, road congestion, resource movement speed, and historical response data:

[0154]

[0155] where v r is the average speed of resource scheduling, C i is the real-time road congestion factor, ω a is the influence weight of the historical correction term, mean(T history ) is the mean of the historical response time, std(T history ) is the standard deviation of the historical response time, is a small positive number to prevent the denominator from being zero, and T i is the response time of the alarm device i;

[0156] S64. Dynamically calculate the priority weight by combining the alarm level, geographical location, and historical event frequency of the abnormal event:

[0157]

[0158] where P i is the priority weight of device i, exp is the exponential function, α 1 is the attenuation factor of the geographical distance, f i is the historical event frequency of device i, f j is the historical event frequency of device j, N is the total number of all alarm devices, A i is the alarm level of device i, A j is the alarm level of device j, D iis the geographical distance from device i to the resource center, D j is the geographical distance from device j to the resource center;

[0159] S65. According to the priority weight P i , the response time T of the alarm device i and the real-time geographical location of the device, generate a resource scheduling plan using the dynamic programming model:

[0160]

[0161] where C total is the total cost of the scheduling plan, min is the optimization objective, γ 1 is the weight of the distance factor, γ 2 is the weight of the response time factor, γ 3 is the weight of the priority factor;

[0162] S66. Output the resource scheduling plan, including task priority, resource allocation, path and completion time.

[0163] Reference Figure 2 , a system for monitoring the quality of electric train equipment based on the LMD system, including the following modules:

[0164] Data acquisition and transmission module, used to collect multi-dimensional data from device sensors, and combine the vehicle-mounted GPS and Beidou systems to obtain the real-time geographical location information of the device;

[0165] Data processing and analysis module, used to clean, fill in missing values, align time, normalize and extract features from the collected data;

[0166] Rule generation and optimization module, used to dynamically adjust monitoring rules and alarm logic through the Bayesian optimization method;

[0167] Status monitoring and alarm module, used to monitor the running status in real time through a classification model, identify anomalies and generate alarm information;

[0168] Resource scheduling and path optimization module, used to comprehensively consider the device alarm level, real-time geographical location and response time, and generate a resource scheduling plan using the dynamic programming model;

[0169] GIS visualization and task execution module, used to display the location of alarm devices, scheduling paths and task status in real time through the GIS module, and dispatch tasks to maintenance personnel;

[0170] Historical data management and model optimization module, used to store the historical operation data, alarm records and maintenance feedback of the device, and continuously optimize the performance of the system model through offline training and deployment.

[0171] Example 1:

[0172] To verify the feasibility of the present invention in implementation, the present invention is applied to a certain railway transportation. During the railway transportation process, the abnormal operation of the on-board electro-mechanical equipment frequently occurs on multiple lines in a certain area due to the increased operation density and equipment aging, seriously affecting the punctuality rate and operation safety of the train. Due to the lag in data collection, single alarm rules, and slow dispatching response in the prior art, the equipment maintenance is delayed, and the operation risk cannot be effectively controlled. To solve the above problems, the present invention constructs a monitoring and dispatching optimization method for on-board electro-mechanical equipment based on the LMD system to comprehensively monitor and manage the operation status of the on-board electro-mechanical equipment in this area.

[0173] First, in the equipment operation monitoring stage, the LMD system is installed on the locomotive, and real-time operation status data of the equipment is collected through a variety of sensors, including physical indicators such as voltage, current, temperature, vibration, and the real-time geographical location provided by the on-board GPS and Beidou systems. These data are transmitted to the ground system through the railway intranet MTUP platform. During the one-month monitoring, an average of about 5GB of data is generated by a single device per day, and the proportion of abnormal data is 4.2%. After data cleaning, missing value filling, and normalization processing, standardized time series data is formed, laying a foundation for further analysis.

[0174] In the data analysis stage, the present invention combines the LSTM-Transformer hybrid model to calculate the equipment health index and predict the occurrence time of potential faults. Taking equipment number A003 as an example, the analysis shows that its health index drops from 0.85 to 0.65. At the same time, combining historical data and real-time monitoring, the model predicts that the probability of an electrical system failure of this equipment within the next 48 hours is 85%. According to the results generated by the model, the system issues an early warning in advance and marks the high-priority alarm level.

[0175] Through the personalized rule optimization module of the present invention, the alarm rule parameters are dynamically adjusted, improving the accuracy of the alarm. Taking current fluctuation as an example, the false alarm rate of a certain equipment was as high as 12% due to the fixed threshold rule before optimization. After optimization, the false alarm rate dropped to 2%, and the missed alarm rate dropped to 0.5%. This process significantly improves the reliability and pertinence of the system alarm and reduces unnecessary maintenance intervention.

[0176] In the resource scheduling stage, the system combines the alarm level, equipment health index, and geographical location to generate a scheduling plan through the dynamic programming algorithm. In a certain scheduling, 5 devices and 3 groups of maintenance personnel are involved. Before the scheduling optimization, the total time is 125 minutes, and the total path coverage distance is 73 kilometers; after optimization, the total time is reduced to 82 minutes, and the total path coverage distance is reduced to 45 kilometers. At the same time, through the GIS visualization module, the scheduling path and the equipment alarm location are displayed on the map in real time, and the maintenance personnel can quickly understand the task distribution.

[0177] After the task is completed, the maintenance personnel upload the maintenance feedback through the LMD system, including the cause of the failure, the maintenance process, and the completion time. Combining the feedback information, the system further optimizes the health index and alarm rules. The statistical data for a quarter shows that after the optimization, the overall failure rate of the equipment in the area has dropped from 7.5% to 3.2%, the alarm accuracy has increased by 25%, and the dispatching efficiency has increased by 30%.

[0178] Table 1 Equipment Health Index and Failure Prediction Results

[0179]

[0180]

[0181] Table 2 Comparison of Dispatching Optimization Effects

[0182]

[0183] Table 1 shows the changes in the health index of different equipment and the failure prediction results. From the data, it can be seen that the downward trend of the equipment health index is closely related to the possibility of failure. For example, the health index of equipment number A003 has dropped from the initial 0.85 to 0.65, and at the same time, the probability of it failing within the next 48 hours is as high as 85%. Similarly, the health index of A005 has dropped from 0.86 to 0.60, and the failure probability is even higher, reaching 90%. These data indicate that through the dynamic monitoring of the health index and the prediction model of the present invention, high-risk equipment can be effectively identified, providing a basis for early warning and rapid response. At the same time, the equipment with a relatively high health index has a relatively low corresponding failure probability, further proving the accuracy of the model prediction. This precise prediction ability significantly improves the scientific nature and reliability of equipment management.

[0184] Table 2 reflects the optimization effects of resource dispatching in different scenarios. From the data, it can be seen that after the introduction of the optimization model, both the total dispatching time and the path coverage distance have decreased significantly. For example, in Scenario 1, the total dispatching time before optimization was 125 minutes and the total distance was 73 kilometers; after optimization, the time dropped to 82 minutes and the total distance decreased to 45 kilometers, with the efficiency increasing by 34.4%. Similarly, the dispatching times in Scenario 2 and Scenario 3 decreased by 44 minutes and 40 minutes respectively, and the path coverage distances decreased by 23 kilometers and 31 kilometers respectively.

[0185] As can be seen from the above table, the present invention has significant advantages in equipment operation status monitoring, fault prediction, and resource scheduling optimization. The dynamic monitoring of the health index enables the accurate identification of high-risk equipment, and the prediction results provide support for taking preventive maintenance measures in advance. At the same time, the optimized scheduling significantly improves the maintenance efficiency and reduces the waste of time and paths. These data fully illustrate the practical value and innovative significance of the present invention in enhancing the intelligent management level of railway signal equipment.

[0186] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A method for monitoring the quality of electric vehicle-mounted equipment based on an LMD system, characterized in that: The steps include: S1. Collect equipment data through the LMD system on the locomotive, combine with GPS and Beidou system, transmit the data to the ground system through the railway intranet MTUP platform, and generate multi-dimensional equipment operation data; S2, preprocessing the device data transmitted via the MTUP platform to generate multi-dimensional time series data; S3, combining multi-dimensional equipment operation data and multi-dimensional time series data, using LSTM-Transformer hybrid model to analyze equipment operation status and generate equipment health index; S4. Combine multi-dimensional equipment operation data, multi-dimensional time series data and equipment health index, use Bayesian optimization algorithm to dynamically adjust monitoring rule parameters and alarm logic, and generate personalized monitoring rules; S5. Use personalized monitoring rules to identify abnormal equipment status through graph neural networks, classify and mark abnormal equipment status according to warning levels, and trigger alarm information; S6. Combine equipment data and alarm information, analyze task priority, resource distribution and equipment location based on dynamic programming model, and generate resource scheduling plan; S7. Use the LMD system GIS module to display resource scheduling plans and alarm information in real time, assign tasks to maintenance personnel, and upload feedback after maintenance is completed; S8. Compare the maintenance feedback information with the equipment health index and alarm information, optimize the LSTM-Transformer hybrid model parameters, and update the equipment status history records and monitoring rules.

2. The method for monitoring the quality of electric vehicle-mounted equipment based on the LMD system according to claim 1 is characterized in that: The S3 specifically includes: S31, multi-dimensional equipment operation data X={x1,x2,...,x t } and multidimensional time series data Z = {z1,z2,...,z t } are merged into input data I, where I={I1,I2,...,I t }, which represents comprehensive data combining equipment physical parameters and time series characteristics; S32, input data I into the LSTM module for time-dependent modeling, the hidden state is H LSTM ={h1,h2,...,h t }, specifically: h t =f(h t-1 ,I t ;θ LSTM ); Among them, θ LSTM is the training parameter of LSTM, h t is the implicit feature of the time series at time t, h t-1 is the implicit feature of the time series at time t-1, and f is the recursive function of LSTM; S33, H LSTM Input to the Transformer module, and calculate the global features through the multi-head self-attention mechanism: H Trans =Concat(head1,head2,...,head n )·W O ; in: Among them, H Trans is the global feature, Concat is the concatenation of the output features of n attention heads, and W O is the weight matrix of the output layer, head i is the output feature of the i-th attention head, head n is the output feature of the nth attention head, W Q,i ,W K,i ,W V,i is the weight matrix of the i-th attention head, is the feature dimension normalization factor, and Softmax is the normalization function; S34, the global feature H output by Transformer Trans Mapped to the device health index HI: HI=σ(W1·H Trans +b1); Among them, σ is the activation function, W1 and b1 are linear mapping parameters respectively; S35. Combine exponential smoothing and weighted sliding window algorithms to smooth and enhance the trend of the equipment health index HI: Among them, HI smooth is the equipment health index after processing, HI t is the health index at the current time t, HI t-i is the health index at the past i-th moment, α is the exponential smoothing factor, and w i is the weighting coefficient of historical data, and N is the sliding window size; S36: The processed equipment health index HI smooth Output, combined with the historical trend of equipment operation, predicts potential failure risks and time windows.

3. The method for monitoring the quality of electric vehicle-mounted equipment based on the LMD system according to claim 1 is characterized in that: The S4 specifically includes: S41, multi-dimensional equipment operation data X={x1,x2,...,x t } and device health index HI smooth , and the historical failure mode data H={H1,H2,...,H N } Input to the rule optimization module as the basic input for monitoring rule optimization; S42. Using the Bayesian optimization algorithm, extract the key failure mode through the historical data H, and calculate the posterior distribution of the monitoring rule parameter θ: Among them, P(θ|H,HI smooth ) is the joint probability of historical failure mode data and health index, θ is the parameter set of monitoring rules, P(θ) is the prior distribution of rule parameters, and Θ is the search space of rule parameters; S43. Optimize rule parameters by maximum a posteriori probability method: θ opt 1argmax θ P(H,HI smooth ∣θ)·P(θ) Among them, θ opt is the optimized monitoring rule parameter, argmax θ To find the parameter value corresponding to the maximum posterior probability of θ; S44, combined with equipment health index HI smooth and the optimized monitoring rule parameter θ opt , dynamically adjust the rule parameters through the exponential weight mechanism and linear correction to generate real-time optimized rule parameters θ real-time : Where μ is the exponential adjustment factor, ∈ is the linear adjustment factor, e is the base of the natural logarithm, HI t The real-time health index of the device at the current moment; S45, according to the real-time optimized rule parameter θ real-time , optimize alarm logic through nonlinear combination in, is the current alarm logic, is the updated alarm logic, θ prev is the rule parameter of the previous cycle, δ is the smoothing parameter to avoid the denominator being zero, ρ is the exponential amplification factor, τ is the alarm logic update factor, and ln is the natural logarithm; S46, the real-time optimized rule parameter θ real-time and alarm logic Update the rule base to generate personalized monitoring rules.

4. The method for monitoring the quality of electric vehicle-mounted equipment based on the LMD system according to claim 1 is characterized in that: The S5 specifically includes: S51, optimize the rule parameter θ in real time real-time and alarm logic Input state classification module as the basis for abnormal state identification; S52, combined with equipment health index HI smooth and multi-dimensional equipment operation data X={x1,x2,...,x t }, construct the equipment operation characteristic graph G = (V, E); S53, the equipment operation characteristic graph G and the real-time optimization rule parameter θ real-time The graph neural network is input to extract the deep relationship between features through the message passing mechanism, and the feature embedding weight is adjusted in combination with the rule parameters to generate the device operation status feature embedding H = {h1,h2,...,h m }: in, For vertex v i At the k-th layer of feature embedding, For vertex v i In the feature embedding of the k+1th layer, N(i) is the vertex v i The neighbor set of vertex v, deg(i) is i The degree of vertex v j Degree, W (k) is the weight matrix of the kth layer of the graph neural network, is the activation function, j is the vertex v i Neighbor nodes, w ij For vertex v i and v j The dynamic weight of the relationship between them; S54, using the equipment operation status feature embedding H, combined with the equipment health index HI smooth and the rule parameter θ for real-time optimization real-time , the state classification result S is generated through a nonlinear classification model with attention weights: s i =sigmoid(W s ·ω(∑ j∈N(i) a ij ·(h j +θ real-time ·h i ))+HI smooth ·d+b s ); Among them, s i is the classification result of node i, sigmoid is to map the classification result to the probability value, W s is the classification layer weight matrix, ω is the nonlinear activation function, α ij is the attention weight of neighbor node j to center node i, h i and h j is the embedded feature of node i and its neighbor node j, δ is the influencing factor of the device health index, b s is the classification layer bias term; S55. Based on the status classification result S, combined with the equipment health index HI smooth , real-time optimization rule parameter θ real-time The equipment operation status characteristics are embedded in H, and the alarm level A of abnormal events is calculated: Among them, λ1 is the health index weight factor, λ2 is the weight factor of the state classification result, λ3 is the rule parameter weight factor, K is the total number of abnormal states detected, and s k is the k-th classification result, max(HI smooth ,H) is the equipment health index HI smooth and the larger value of the equipment operation status feature embedding H; S56. Mark the alarm level A of the abnormal event, classify it into multiple alarm categories according to preset classification rules, and trigger corresponding alarm information.

5. The method for monitoring the quality of electric vehicle-mounted equipment based on the LMD system according to claim 1 is characterized in that: The S6 specifically includes: S61. Combine the alarm level A of the abnormal event and the real-time geographic location L of the device i , construct the objective function of the dynamic programming model: Among them, C is the total cost of dispatch, N is the total number of alarm devices to be dispatched, and D i is the geographical distance from the resource center to the alarm device i, T i is the time required to respond to device i, γ1, γ2, γ3 are weight factors, and Minimize is to minimize the objective function C; S62, through the real-time location of the alarm device i (x i ,y i ,z i ) and resource center location (x r ,y r ,z r ), combined with the path complexity weight w d and the vertical distance weight ω p , calculate the geographical distance D i : S63. Dynamically estimate the response time of the alarm device, taking into account geographic distance, road congestion, resource movement speed and historical response data: Among them, v r is the average speed of resource scheduling, C i is the real-time road congestion factor, ω a is the impact weight of the historical correction term, mean(T history ) is the mean of historical response time, std(T history ) is the standard deviation of historical response time, To prevent small positive numbers with zero denominators, T i is the response time of alarm device i; S64. Dynamically calculate the priority weight based on the alarm level, geographic location, and historical event frequency of the abnormal event: Among them, P i is the priority weight of device i, exp is the exponential function, α1 is the attenuation factor of the geographical distance, and f i is the historical event frequency of device i, f j is the historical event frequency of device j, N is the total number of all alarm devices, A i is the alarm level of device i, A j is the alarm level of device j, D i is the geographical distance from device i to the resource center, D j is the geographical distance from device j to the resource center; S65, according to the priority weight P i , Response time of alarm equipment T i And the real-time geographic location of the device, using the dynamic programming model to generate a resource scheduling plan: Among them, C total is the total cost of the scheduling plan, min is the optimization target, γ1 is the weight of the distance factor, γ2 is the weight of the response time factor, and γ3 is the weight of the priority factor; S66. Output resource scheduling plan, including task priority, resource allocation, path and completion time.

6. A system for monitoring the quality of electric vehicle-mounted equipment based on an LMD system, a method for monitoring the quality of electric vehicle-mounted equipment based on an LMD system according to any one of claims 1 to 5, characterized in that: Includes the following modules: Data collection and transmission module, used to collect multi-dimensional data from equipment sensors and obtain real-time geographic location information of the equipment in combination with vehicle-mounted GPS and Beidou system; Data processing and analysis module, used for cleaning, missing value filling, time alignment, normalization and feature extraction of collected data; The rule generation and optimization module is used to dynamically adjust the monitoring rules and alarm logic through the Bayesian optimization method; The status monitoring and alarm module is used to monitor the operating status in real time through the classification model, identify abnormalities and generate alarm information; Resource scheduling and path optimization module, which integrates equipment alarm levels, real-time geographic locations, and response times, and uses dynamic planning to generate resource scheduling plans; GIS visualization and task execution module, which is used to display the location of alarm equipment, dispatch path and task status in real time through the GIS module, and dispatch tasks to maintenance personnel; The historical data management and model optimization module is used to store historical equipment operation data, alarm records and maintenance feedback, and continuously optimize system model performance through offline training and deployment.

Images