Operation and maintenance data statistical analysis and prediction method based on large model

Abstract

The invention relates to an operation and maintenance data statistical analysis and prediction method based on a large model, and relates to the technical field of rail transit operation and maintenance data management. Comprising the steps of constructing a rail transit operation and maintenance database, performing layered arrangement on operation and maintenance data, constructing a tree-shaped mapping network, calculating influence weight factors among points, constructing component state feature vectors and operation and maintenance means to train an operation and maintenance analysis and prediction model, predicting the operation and maintenance means by using the model, and evaluating the health degree of each level. According to the method, the block chain is constructed to summarize the operation and maintenance data of a plurality of operators, and the size of an abnormal data sample is expanded; the objective weight and the subjective weight are combined to obtain an influence weight factor of each element between adjacent hierarchies, and high coupling between the elements of each hierarchies can be avoided when health degrees of different hierarchies are subsequently carried out; and meanwhile, the prediction frequency of the operation and maintenance analysis and prediction model is controlled, so that the problems of calculation power loss caused by frequent prediction implementation and weak risk management and control strength caused by regular prediction are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of rail transit operation and maintenance data management, and specifically provides a method for statistical analysis and prediction of operation and maintenance data based on a large model. Background Art

[0002] As an important means of public transportation, rail transit provides great convenience for people's daily travel. As a huge operating system, the number of hardware devices and software logics involved in rail transit is extremely large. If only emergency operation and maintenance are carried out according to the immediate alarm method during management and control, it will lead to a long response time, greatly affecting the smoothness of the system operation, and further affecting people's travel experience.

[0003] There are obvious deficiencies in the accumulation and discovery of abnormal samples in the existing rail transit operation and maintenance management technology. As a result, there is a lack of data support for abnormal prediction, leading to low warning accuracy. In addition, due to the lack of reasonable classification and sorting of equipment and the weight allocation of each element in the same classification level in the existing operation and maintenance management, it is difficult to comprehensively evaluate the health of each element due to its high coupling. Therefore, we provide a method for statistical analysis and prediction of operation and maintenance data based on a large model. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for statistical analysis and prediction of operation and maintenance data based on a large model.

[0005] The technical problems solved by the present invention are as follows:

[0006] (1) How to summarize the operation and maintenance data of multiple operators by constructing a blockchain to expand the size of abnormal data samples and solve the problem of insufficient abnormal samples in the existing technology;

[0007] (2) How to classify and layer the hardware and software in the entire rail transit system, and then construct a tree-shaped mapping network and the influence weight factor between adjacent levels to solve the problems of high coupling between each element node and difficult health evaluation in the existing technology;

[0008] (3) How to construct a component status feature vector through personalized tags and common tags, and predict and match the corresponding operation and maintenance means through the constructed operation and maintenance analysis and prediction model, and control the prediction frequency, so as to solve the problems of computing power loss caused by prediction implementation and weak risk control intensity caused by regular prediction.

[0009] The present invention can be realized through the following technical solutions: A method for statistical analysis and prediction of operation and maintenance data based on a large model, including the following steps:

[0010] Step 1: Build a rail transit operation and maintenance database based on blockchain, aggregate the operation and maintenance data of each operator, and obtain a large number of abnormal samples;

[0011] Step 2: Stratify and organize the operation and maintenance data in the rail transit operation and maintenance database according to the system layer, module layer, unit layer, and component layer;

[0012] Step 3: Use a dot-line graph to build a tree-like mapping network between levels. In the tree-like mapping network, construct the influence weight factor of each node in the lower level relative to the upper level;

[0013] Step 4: Define the tracking feature labels of components and construct the component status feature vector. The tracking feature labels include common labels and individual labels;

[0014] Step 5: Use the component status feature vector and the corresponding operation and maintenance means as the input and output respectively to construct an operation and maintenance analysis and prediction model based on the BP neural network;

[0015] Step 6: Use the operation and maintenance analysis and prediction model to predict the operation and maintenance means required for the corresponding components;

[0016] Step 7: Update the health of each level based on the working status log of the components and the predicted operation and maintenance means.

[0017] A further technical improvement of the present invention is that in Step 1, the rail transit operation and maintenance database is a distributed ledger copy set at each node. Each operator node regularly uploads the operation and maintenance data to the blockchain, and the operation and maintenance data is synchronously updated to the rail transit operation and maintenance database.

[0018] A further technical improvement of the present invention is that the influence weight factor = w1 * subjective weight factor + w2 * objective weight factor, where w1 and w2 are the distribution ratios of the subjective weight factor and the objective weight factor respectively, w1 + w2 = 1 and w1 < w2;

[0019] The acquisition method of the subjective weight factor is as follows: Based on expert experience, use the 1-9 scale method to mark the relative importance degree between elements in the same level, organize the marking results into a judgment matrix, and calculate the eigenvector of the judgment matrix. After normalization, obtain the subjective weight factor of each node in the lower level relative to a certain node in the upper level;

[0020] The acquisition method of the objective weight factor is as follows: Conduct index scoring on each element in the same level, and perform a ratio operation on the index score of each element and the total index score of all elements in this level.

[0021] A further technical improvement of the present invention is that the calculation method of the element index score is as follows:

[0022]

[0023] Among them, V represents the intrinsic value of the device or apparatus corresponding to a certain node, L d represents the power or load-bearing capacity, E represents the average operation and maintenance frequency, W represents the failure influence range, that is, the number of affected nodes in the tree-shaped mapping network, L f represents the service life of the corresponding device or apparatus, ρ represents the setting density of the same device or apparatus at this level; A, B, C, D, α, β are respectively the compensation ratio coefficients of the above corresponding influence parameters, e is the natural constant in mathematics, and the above influence parameters have all been dimensionless processed; i represents the element number in the corresponding level.

[0024] A further technical improvement of the present invention lies in that: the component status feature vector in step four is constructed by the top N1 personality tags with the highest correlation degree in the personality tags and all N2 common tags to obtain a component status feature vector of (N1 + N2) dimensions Among them, j represents the component number, and the correlation degree between the personality tag and the component is set through expert experience.

[0025] A further technical improvement of the present invention lies in that: in step six, the operation and maintenance analysis and prediction model is used to predict the current operation and maintenance status of the component and match the current operation and maintenance means, and the prediction frequency is controllable. The calculation method of the frequency is as follows: Among them, V j represents the proportion of the value of the corresponding component j in the value of all components at the corresponding level, and has been normalized; T 有效 represents the total duration of the corresponding component in the working state, T ～ represents the service life of the corresponding component.

[0026] A further technical improvement of the present invention lies in that: when there is data update in each prediction, the health degree of each level is synchronously updated. The evaluation process of the health degree includes:

[0027] Assign corresponding influence degrees to different operation and maintenance means;

[0028] Convert the operation and maintenance means of the corresponding component in the component layer into an influence degree assignment, and multiply the influence degree assignment by the influence weight factor of this level and sum to obtain the health degree of the upper level;

[0029] Calculate the health degree of each module layer according to the health degree of a certain level according to the influence weight factor of this level relative to the upper level, and so on.

[0030] A further technical improvement of the present invention lies in: in each layer, a monitoring threshold for health is set. When the health of the corresponding layer exceeds the monitoring threshold, a warning signal is generated and sent to the operation and maintenance management department. The operation and maintenance management department responds according to the warning signal, marks the influence degree assignment of the component layer in the tree-like mapping network and marks the health in the nodes of other layers. The operation and maintenance personnel perform priority operation and maintenance processing according to the occurrence position of the warning signal and the position with the largest value in its next layer. After the processing is completed, the health is updated. When the health still exceeds the monitoring threshold, the above operation and maintenance selection is repeated until the health returns to the monitoring threshold range.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The present invention aggregates the operation and maintenance data of multiple operators by constructing a blockchain, expands the size of abnormal data samples, so that sufficient data support can be provided for subsequent prediction under sufficient abnormal samples, thereby improving the accuracy of prediction.

[0033] 2. The present invention combines objective weight and subjective weight to obtain the influence weight factor of each element between adjacent layers. When the health of different layers is carried out subsequently, the high coupling between the elements of each layer can be avoided, so that the monitoring of health has reference value and provides a trigger condition for the operation and maintenance of the whole system.

[0034] 3. The present invention constructs a component state feature vector through personalized labels and common labels, aggregates data on the state characteristics of components from two aspects and multiple parameters to obtain a vector, so that this state feature has high representativeness. Thus, a prediction model is trained with the support of a large amount of operation and maintenance data, and the required operation and maintenance means are matched to provide data guidance for subsequent health assessment and operation and maintenance process; in addition, the frequency of prediction is controlled to avoid the problem of computing power loss caused by frequent prediction and the weak risk control intensity caused by regular prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] For the convenience of those skilled in the art to understand, the present invention will be further described below with reference to the accompanying drawings.

[0036] Figure 1 It is a schematic diagram of the execution process of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0037] In order to further elaborate the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following describes in detail the specific implementation manner, structure, features and effects of the present invention with reference to the accompanying drawings and preferred embodiments.

[0038] Please refer to Figure 1As shown in the figure, a method for statistical analysis and prediction of operation and maintenance data based on a large model includes the following steps:

[0039] Step 1: Construct an operation and maintenance database for rail transit based on blockchain

[0040] Since the operation and maintenance data of each rail transit operator are stored in its own database, but due to the extremely small proportion of abnormal samples in the massive operation data, there will also be some valuable abnormal samples scattered among different operation entities, resulting in the inability to obtain a sufficient number and as many different types of abnormal samples as possible, thus making it impossible to analyze and externalize the value of abnormal samples;

[0041] Considering the privacy and high data throughput characteristics of rail transit operation and maintenance data, a consortium chain is selected as the underlying framework of the corresponding blockchain, nodes are built and node permissions are set, smart contracts and encryption algorithms are designed for decentralized automatic execution control, and the structure of the rail transit operation and maintenance database and access control policies are designed. Among them, the structure of the rail transit operation and maintenance database includes business name rules, parameter types, dimensions, etc.;

[0042] Each operator node regularly uploads operation and maintenance data to the blockchain, and the operation and maintenance data are synchronously updated to the rail transit operation and maintenance database. The rail transit operation and maintenance database is specifically a distributed ledger copy of each node, with interaction records of all data in the blockchain, so that a large amount of abnormal sample data can be collected for analysis;

[0043] Step 2: Classify and organize the rail transit operation and maintenance database

[0044] The operation and maintenance data are hierarchically organized according to the system layer, module layer, unit layer, and component layer. The system layer includes the vehicle system, track system, and in-station system; each system layer includes a mechanical module, an electrical module, and a communication module; the mechanical module includes a power output unit, a functional unit, and an intermediate connection unit, the electrical module includes an electrical output unit, an electrical transmission unit, and a functional unit, and the communication module includes a signal sending unit and a signal receiving unit; the component layer is the specific component in the corresponding unit;

[0045] Step 3: Construct the influence weight factor between every two adjacent levels

[0046] According to the classification and organization results in Step 2 above, a tree-like mapping network between levels is built using a dot-line graph. The nodes correspond to the corresponding levels in the steps. For example, there are three system layer nodes and nine module layer nodes in the network, etc.; the length of the node connection line between adjacent levels represents the influence weight factor of the lower level relative to the upper level between the corresponding two levels;

[0047] The influence weight factors include subjective weight factors and objective weight factors. Among them, the subjective weight factors are based on expert experience and relevant specifications, and the analytic hierarchy process is used to obtain the weights of sub-elements in each level relative to the upper level:

[0048] Compare each pair of sub-elements in the same level, use the 1-9 scale method to make judgments and mark the results. Finally, organize the marked results into a judgment matrix. The elements in the matrix represent the relative importance degrees between sub-elements in each level. Calculate the eigenvector of this judgment matrix, and this eigenvector is the weight vector of each sub-element. Normalize the eigenvector to obtain the subjective weight factor of each node in the lower level relative to a certain node in the upper level;

[0049] Calculate the objective weight factors of each sub-element in the same level: By scoring the indicators of each sub-element, and performing a ratio operation on the indicator scores of each sub-element relative to a certain common node in the upper level and the total indicator scores of all sub-elements in this level. Specifically:

[0050] The calculation formula for the indicator score of a sub-element is:

[0051] Among them, V represents the intrinsic value of a device or apparatus corresponding to a certain node, L d represents the power or load it bears, E represents the average maintenance frequency, W represents the fault influence range, and the influence range is represented by the number of affected nodes in the tree-shaped mapping network. L f represents the service life of the corresponding device or apparatus, ρ represents the setting density of the same device or apparatus at this level, that is, the proportion of the number of the same device or apparatus corresponding to this node in all devices or apparatuses;

[0052] A, B, C, D, α, β are respectively the compensation ratio coefficients of the above corresponding influence parameters, e is the natural constant in mathematics, and the above influence parameters have all been dimensionless processed;

[0053]

[0054] Influence weight factor = w1 * subjective weight factor + w2 * objective weight factor;

[0055] Among them, w1 and w2 are respectively the distribution ratio of the subjective weight factor and the distribution ratio of the objective weight factor, w1 + w2 = 1 and w1 < w2.

[0056] Step Four: Define the tracking feature labels of components and construct the component state feature vectors

[0057] Since the operation and maintenance work will ultimately be implemented on specific components, tracking feature tags are set for the operation and maintenance data generated by the components in the component layer. Also, due to the differences in the functions, types, and usage scenarios of the components, the tracking feature tags are divided into common tags and individual tags;

[0058] The common tags include the running duration, average load, peak load, overload duration, and energy loss efficiency between two adjacent operation and maintenance events. Among them, the energy loss efficiency represents the ratio of the total energy consumption (such as electric energy) within the running duration to the running duration;

[0059] The individual tags are feature tags highly relevant to the functions when the corresponding components are performing functions, such as the wear amount, vibration frequency, and amplitude of transmission components; the heat generation efficiency, execution response duration, and electrical circuit stability of electrical components; the packet loss rate, traffic transmission size, and network response duration of communication components.

[0060] Select the top N1 individual tags with the highest correlation from the individual tags, and together with all N2 common tags, construct a component state feature vector of (N1 + N2) dimensions Among them, j represents the component number, and the correlation between the individual tags and the components is set through expert experience.

[0061] Step Five: Build an operation and maintenance analysis and prediction model

[0062] Obtain the operation logs of the corresponding components within a certain time length before and after each operation and maintenance from the rail transit operation and maintenance database, calculate the mean value of the component state feature data of the corresponding components within a certain time length before and after each operation and maintenance, filter out the singular point data during the mean value calculation to improve the data credibility, and use the obtained mean value to fill the component state feature vector. Among them, the data spanning the duration of two adjacent operation and maintenance cycles in the common tags is filled after being obtained according to the corresponding mean value calculation method;

[0063] Associate each constructed component state feature vector with the operation and maintenance means adopted in the current operation and maintenance. The operation and maintenance means include replacement, repair, debugging, maintaining the current, etc.; Subsequently, based on the BP neural network, using the component state feature vector and the operation and maintenance means as the input and output respectively, construct a training set and a test set to obtain an operation and maintenance analysis and prediction model;

[0064] Step Six: Use the operation and maintenance analysis and prediction model to match the operation and maintenance means required for the corresponding components

[0065] Based on the component value, service life, and the total duration of the corresponding component in the working state, control the frequency of constructing the component state feature vector and applying the operation and maintenance analysis and prediction model for the operation logs of the corresponding components within each fixed period (one day, one week, or one month). The calculation method of this frequency is: Where Vj It represents the proportion of the value of the corresponding component j in the value of all components at the corresponding level, and is normalized; T 有效 It represents the total duration of the corresponding component in the working state, T ～ It represents the service life of the corresponding component;

[0066] After determining the frequency of using the model to predict components, the number of prediction times obtained by the components is evenly distributed in a fixed period, and the component state feature vectors are automatically constructed at the evenly distributed time points and the model is used for prediction to match the current operation and maintenance means.

[0067] Step Seven: Update the health of each level based on the working state of the components and the predicted operation and maintenance means

[0068] When there is data update in each prediction in Step Six, the health of each level is synchronously updated; the health of each level is comprehensively evaluated according to the prediction results of each element in the next level according to their respective influence weight factors. The specific process is as follows:

[0069] Set the corresponding influence degree assignment according to the operation and maintenance means: {maintain current, debug, repair, replace} ~ {0, 1, 2, 3};

[0070] Taking the component layer as an example, when obtaining the current operation and maintenance means of the corresponding component, it is converted into the corresponding influence degree assignment, and then the product of each influence degree assignment and the influence weight factor is calculated and summed to obtain the health of each unit layer; and so on, the health of each module layer can be calculated according to the health of each unit layer according to the influence weight factor of this level relative to the upper level, and similarly the health of the system layer can be calculated;

[0071] Set monitoring thresholds for the health of each level. When the health of the corresponding level exceeds the monitoring threshold, an early warning signal is generated and sent to the operation and maintenance management department. The operation and maintenance management department responds according to the early warning signal, marks the influence degree assignment of the component layer in the tree-shaped mapping network and marks the health in the nodes of other levels. The operation and maintenance personnel perform priority operation and maintenance processing according to the occurrence position of the early warning signal and the position with the largest value in its next level. After the processing is completed, the health is updated. When the health still exceeds the monitoring threshold, repeat the above operation and maintenance selection until the health returns to the monitoring threshold range.

[0072] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the above-disclosed technical content within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for statistical analysis and prediction of operation and maintenance data based on a large model, characterized in that: The steps include: Step 1: Build a rail transit operation and maintenance database based on blockchain, aggregate the operation and maintenance data of each operator, and obtain a large number of abnormal samples; Step 2: organize the operation and maintenance data in the rail transit operation and maintenance database into layers according to the system layer, module layer, unit layer and component layer; Step 3: Use a point-line graph to build a tree-like mapping network between levels, and in the tree-like mapping network, construct an influence weight factor of each node in the next level relative to the previous level; Step 4: define the tracking feature label of the component and construct the component state feature vector. The tracking feature label includes a common label and a personalized label. Step 5: Using the component status feature vector and the corresponding operation and maintenance means as input and output respectively, construct an operation and maintenance analysis and prediction model based on BP neural network; Step 6: Use the operation and maintenance analysis prediction model to predict the operation and maintenance means required for the corresponding components; Step 7: Update the health of each layer based on the component's working status log and predicted operation and maintenance measures.

2. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 1, characterized in that: In step one, the rail transit operation and maintenance database is a distributed ledger copy set up at each node. Each operator node regularly uploads the operation and maintenance data to the blockchain, and the operation and maintenance data is synchronously updated to the rail transit operation and maintenance database.

3. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 1, characterized in that: The impact weight factor = w1*subjective weight factor + w2*objective weight factor, wherein w1 and w2 are the subjective weight factor allocation ratio and the objective weight factor allocation ratio respectively, w1+w2=1 and w1<w2; The subjective weight factor is obtained by: based on expert experience, using the 1-9 scale method to mark the relative importance of each element in the same level, organizing the marking results into a judgment matrix, and calculating the eigenvector of the judgment matrix. After normalization, the subjective weight factor of each node in the next level relative to a node in the previous level is obtained; The objective weight factor is obtained by scoring each element in the same level and performing a ratio operation on the score of each element and the total score of all elements in the level.

4. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 3 is characterized in that: The element indicator score is calculated as follows: Among them, V represents the intrinsic value of the device or device corresponding to a node, L d represents the power or load size, E represents the average operation and maintenance frequency, W represents the scope of fault impact, that is, the number of affected nodes in the tree mapping network, L f represents the service life of the corresponding equipment or device, ρ represents the setting density of the same equipment or device at this level; A, B, C, D, α, β are the compensation ratio coefficients of the above corresponding influencing parameters respectively, e is a natural constant in mathematics, and the above influencing parameters are all dimensionless; i represents the element number in the corresponding level.

5. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 1, characterized in that: The component state feature vector in step 4 is constructed by combining the top N1 most correlated individual tags with all N2 common tags to obtain a component state feature vector of (N1+N2) dimension. Among them, j represents the part number, and the correlation between the personalized label and the part is set based on expert experience.

6. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 1, characterized in that: In step 6, the operation and maintenance analysis prediction model is used to predict the current operation and maintenance status of the component and match the current operation and maintenance means. The frequency of prediction is controllable, and the frequency is calculated as follows: Among them, V j T represents the value of the corresponding component j as a percentage of the value of all components in the corresponding level, and is normalized; 有效 Indicates the total time that the corresponding component is in working condition, T ～ Indicates the service life of the corresponding component.

7. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 6 is characterized in that: When there is data update in each prediction, the health of each level is updated synchronously. The health evaluation process includes: Assign corresponding impact values ​​to different operation and maintenance methods; Convert the operation and maintenance means of the corresponding components in the component layer into impact values, and multiply and sum the impact values ​​with the impact weight factors of the layer to obtain the health of the previous layer; The health of a certain layer is calculated according to the influence weight factor of the layer relative to the previous layer to calculate the health of each module layer, and so on.

8. The method for statistical analysis and prediction of operation and maintenance data based on a large model according to claim 7 is characterized in that: At each level, a monitoring threshold for health is set. When the health of the corresponding level exceeds the monitoring threshold, an early warning signal is generated and sent to the operation and maintenance management department. The operation and maintenance management department responds according to the early warning signal, marks the impact value of the component layer in the tree mapping network, and marks the health of the nodes in other levels. The operation and maintenance personnel perform priority operation and maintenance processing based on the location of the early warning signal and the location with the largest value in the next level. After the processing is completed, the health is updated. When the health still exceeds the monitoring threshold, the above operation and maintenance selection is repeated until the health returns to the monitoring threshold range.

Images