Overhead line system facility maintenance scheduling and sequencing method

Abstract

The invention provides a catenary facility maintenance scheduling and sequencing method, and relates to the technical field of catenary facility safety. The method comprises the following steps: constructing a health degree ordered list by taking data pairs of contact network facilities as elements, and performing ascending sorting; a priority prediction model is constructed, and assignment is initialized randomly; inputting the feature vector of each contact network facility to be sorted into a priority prediction model to obtain a prediction value of a priority index; sequentially adding the data pair of each to-be-sorted contact network facility as a new element into a health degree ordered list; when a new element is added, all the data pairs are sorted in an ascending order again according to the priority indexes; wherein the priority index is the health degree of the contact network facility. According to the method, on the basis of avoiding existing data overfitting, effective generalization of new samples can be achieved, the accuracy, precision and efficiency of sorting of the health degrees of the contact network facilities are improved, and accurate overhauling of the contact network facilities can be achieved.

Description

Technical Field

[0001] This invention relates to the field of overhead contact line safety technology, and in particular to a method for scheduling and arranging the maintenance of overhead contact line facilities. Background Technology

[0002] Traditional overhead contact line maintenance follows a "periodic inspection" strategy, which involves sequentially checking and evaluating all facilities within a given timeframe. This approach suffers from problems such as long inspection cycles, low efficiency, and the risk of missed inspections, making it unsuitable for the significant application demands of the new era. To address these technical issues and achieve "condition-based precision maintenance," a sequencing method has been introduced by those skilled in the art into the overhead contact line maintenance process.

[0003] The ranking of overhead contact line facility maintenance is essentially a data-driven method for comparing and ranking the health status of facilities. In existing ranking methods, the latent variables involved in the prediction are unevenly distributed or even discontinuous in their spatial distribution, leading to prediction results occurring separately on various data silos. In this case, the model may overfit to existing data and suffer from technical problems in effectively generalizing to new samples. Furthermore, this results in limited accuracy, precision, and efficiency in ranking the health status of overhead contact line facilities, making it difficult to achieve precise maintenance of these facilities. Summary of the Invention

[0004] To address the aforementioned technical problems in the existing technology, this invention provides a method for scheduling and sequencing the maintenance of overhead contact line facilities.

[0005] Specifically, the technical solution provided by this sorting method includes the following steps:

[0006] Step S1: Construct an ordered list of health status for each data pair of overhead contact line facilities, i.e., the first data pair; if the ordered list of health status is not empty, sort all data pairs in ascending order according to priority indicators, i.e., sort all overhead contact line facilities; wherein, the first data pair includes the priority indicators and index of the overhead contact line facilities.

[0007] Construct a priority prediction model and perform random initialization assignment;

[0008] Step S2: For each overhead contact line facility to be sorted, input its feature vector into the priority prediction model to obtain the predicted value of the priority index;

[0009] Step S3: Sequentially add each data pair of the overhead contact line facility to be sorted, i.e., the second data pair, as a new element to the health-ordered list; wherein, the second data pair includes the predicted value of the priority indicator and the corresponding index;

[0010] Step S4: For each new element added, sort all data pairs in ascending order again according to the priority index;

[0011] Among them, the priority indicator is the health of the overhead contact system.

[0012] Preferably, an ordered list of health status is constructed for each element using data pairs of overhead contact line facilities, i.e., the first data pair, specifically including:

[0013] Data pairs based on overhead contact line facilities (s) i ,p i Construct an ordered list F of health status.

[0014] F = [(s1, p1), ..., (s i ,p i ...,(s n ,p n )];

[0015] In the formula, i is the serial number of the overhead contact line facility, i = 1, 2, ..., n, and n is the number of overhead contact line facilities. i For the health status of the i-th overhead contact line facility, p i The index for accessing the i-th overhead contact line facility.

[0016] Preferably, the quicksort algorithm is used to sort all data pairs in ascending order.

[0017] Preferred priority prediction models include:

[0018] The ISAB layer, PMA layer, and SAB layer are constructed based on the MAB module and connected sequentially.

[0019] The ISAB layer includes a first MAB module and a second MAB module; the first MAB module takes the feature vector of the overhead contact line facilities to be sorted as input; the second MAB module takes the feature vector of the overhead contact line facilities to be sorted and the output of the first MAB module as input.

[0020] The PMA layer includes an rFF module and a third MAB module; the rFF module takes the output of the second MAB module as its input; the third MAB module takes the output of the rFF module as its input.

[0021] The SAB layer includes a fourth MAB module; the fourth MAB module takes the output of the third MAB module as its input.

[0022] Preferably, the training method for the priority prediction model includes:

[0023] Step R1: Construct a sample data set X for overhead contact line facilities.

[0024]

[0025] In the formula, x j Let s be the feature vector of the j-th overhead contact line facility sample. j Let N be the priority index for the j-th overhead contact line facility sample, and N be the number of overhead contact line facility samples, where j = 1, 2, ..., N;

[0026] Step R2: Randomly sample m elements from the overhead contact system sample data set X to form a sample data set X'.

[0027] X′=(x j ,s j ):j∈B;

[0028] B∈N;

[0029] Step R3: Sort the elements in the sampled data set X' according to the priority index;

[0030] Step R4: Predict model f using priority settings θ For the feature vectors of m overhead contact line facility samples, priority indicators are predicted respectively.

[0031]

[0032] In the formula, μ j and Let represent the mean and variance of the feature vector of the j-th overhead contact line facility sample obtained by prediction, respectively;

[0033] Step R5: Calculate the prior regularization cost L for m predictions. prior ,

[0034] L prior =L mean +L var ;

[0035] In the formula, L mean For mean constraint, L var For variance constraints,

[0036]

[0037] In the formula, γj is the sorting parameter of the feature vector of the j-th overhead contact line facility sample, and μ γj Let be the mean of the feature vectors of the overhead contact system facility samples. Let ε be the variance of the feature vectors of the overhead contact system sample, and ε be the error.

[0038] Step R6: Calculate the importance weight w for each pair of overhead contact line facility sample data.

[0039] w = softmax(ηd);

[0040]

[0041] In the formula, η is the importance parameter. This represents the priority index of the γj-th overhead contact line facility sample. Priority indicators for samples of adjacent overhead contact line facilities The difference value;

[0042] Step R7: Calculate the priority index pairs of adjacent overhead contact line facility samples. Loss function L under variational learning likelihood ,

[0043]

[0044] In the formula, Here, θ represents the weighting coefficients, and θ represents the model parameters. Indicates according to the feature vector The probability of generating the priority index random variable s. for The probability of;

[0045] Step R8: Calculate the total loss function L.

[0046] L = L prior +L likelihood ;

[0047] Step R9: Set the learning rate and minimize the total loss function using the mini-batch stochastic gradient descent algorithm to obtain the optimized parameters θ of the priority prediction model. (t+1) ,

[0048]

[0049] In the formula, θ (t) These are the parameters for the priority prediction model;

[0050] Step R10: Repeat steps R3 to S9 until the error of the priority prediction model before and after parameter optimization is no greater than the maximum permissible error, and output parameter θ. (t+1) Save the application and complete the training of the priority prediction model.

[0051] Preferably, before step R1, the feature vectors of the overhead contact line facility samples are preprocessed, specifically as follows:

[0052] Remove outliers, missing items, and duplicates;

[0053] The feature vectors of all overhead contact line facility samples are processed using a noise reduction algorithm;

[0054] Calculate the standard deviation σ of the eigenvectors of all overhead contact line facility samples and the mean μ of each channel;

[0055] For each overhead contact line facility sample, the feature vector x j Perform range conversion.

[0056]

[0057] In the formula, Let be the feature vector of the j-th overhead contact line facility sample after range transformation.

[0058] Preferably, the feature vector includes the maintenance cycle, maintenance history, and evaluation indicators of the overhead contact system facilities, as well as grayscale statistics, Gabor features, wavelet features, and ridge wave features of the overhead contact system facility images.

[0059] The technical solution provided by this invention can clarify the probability distribution of the priority of each ranking index through a priority prediction model. It can effectively generalize to new samples while avoiding overfitting of existing data, thereby improving the accuracy, precision and efficiency of ranking the health of overhead contact line facilities and helping to achieve precise maintenance of overhead contact line facilities. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the structure of the Multihead Attention Block (MAB) module in the priority prediction model of this invention.

[0061] Figure 2 This is a schematic diagram of the priority prediction model in this invention. Detailed Implementation

[0062] The technical solution provided by the present invention will be further described in detail below with reference to the accompanying drawings.

[0063] Electrified high-speed railways use an overhead contact system to provide traction power to trains. Because the overhead contact system is the weakest link in the entire rail transit system, it is susceptible to environmental factors and lacks a backup system. A failure in this system can cause significant economic losses and safety risks to the railway transportation system. Therefore, maintaining the condition of the overhead contact system is a key focus for railway safety. With the continuous expansion of high-speed rail lines, the number of overhead contact system facilities has increased dramatically. Traditional overhead contact system maintenance follows a "periodic inspection" strategy, which involves sequentially inspecting and evaluating all facilities within a certain time period. This method suffers from problems such as long inspection cycles, low efficiency, and the risk of missed inspections, making it difficult to meet the demands of new applications. For critical facilities, a maintenance method based on facility service life is also used, where the order of maintenance is determined by the degree of aging caused by the facility's service life. This maintenance scheduling method is predetermined and lacks consideration for different occasions, time periods, and environmental factors. In recent years, deep learning-based video recognition technology has been introduced into overhead contact line inspection, significantly improving the frequency and accuracy of identifying key facilities. However, these identification results generate massive amounts of information to be processed, still requiring guidance from priority ranking methods for maintenance scheduling. Since the window of opportunity for high-speed railway maintenance is limited to the early morning hours when no trains are running, time is extremely limited. Against the backdrop of continuously increasing operational intensity and frequency, there is an urgent need to develop a service life ranking technology based on facility condition characteristics, thereby forming an intelligent maintenance solution for overhead contact line facilities.

[0064] The sorting of overhead contact line facilities for maintenance is essentially a data-driven method for comparing and ranking the health status of facilities. Sorting is the process of rearranging one or more sets of data according to the magnitude of one or more keys. A sorting algorithm is an algorithm that rearranges this data according to specific requirements. After sorting, the originally disordered data will be reorganized into an ascending or descending list following certain rules. The regularity of the sorting results helps in data filtering and calculation, and can greatly improve computational efficiency in subsequent processing. Sorting algorithms are widely valued in many fields, especially in the field of computer data processing. Generally speaking, traditional computer data sorting functions rearrange an arbitrary sequence of data records into a sequence with ordered keys. It can be said that an excellent sorting algorithm can save a lot of computing storage and computing resources, making it possible for computer programs to undertake large-scale data service tasks. Traditionally, sorting algorithms are divided into internal sorting and external sorting. Internal sorting sorts data records in memory, while external sorting is used when the data to be sorted is too large to fit in memory at once, requiring access to external memory during the sorting process. Common internal sorting algorithms include: insertion sort, shell sort, selection sort, bubble sort, merge sort, quick sort, heap sort, and radix sort. Analyzing the performance of these sorting algorithms involves several evaluation metrics, including time and space complexity, stability, and applicability, covering multiple aspects. The time complexity of a sorting algorithm refers to the total time spent from the initial state of the sequence to the final result after transformations and shifts. The space complexity of a sorting algorithm is the space overhead incurred from the initial state to the final state after sorting and shifting. The stability of a sorting algorithm means that when two elements with the same value appear in the sequence, their relative positions should not change before and after sorting. In short, a stable sorting algorithm does not change the relative order of elements, while an unstable sorting algorithm often changes this order, leading to unnecessary swaps and thus increasing the algorithm's time complexity. Furthermore, the applicable scenarios for sorting algorithms are also a factor to consider. Because there are many sorting algorithms, and different types of sorting algorithms are suitable for different types of scenarios, sometimes it may be necessary to save space and not have much time requirement, while other times more time is considered and space requirement is lower. In short, the choice is usually based on one aspect.

[0065] In general, the overhead contact line facility maintenance sorting algorithm should not only meet the comprehensive requirements of low time and space complexity, good stability and wide applicability, but also adapt to the implicit expression characteristics of the condition characteristics to be inspected in the historical data of the overhead contact line facility condition inspection.

[0066] In the scheduling of overhead contact line facility maintenance, high-risk facilities with limited remaining service life need to be prioritized. Specifically, based on historical inspection data, a comparison of the health status of components such as the cantilever arms, catenary wires, contact wires, and droppers on each support of the contact line can be established. This creates a ranking list from most urgent to least urgent maintenance, forming a scheduling sequence for contact line maintenance. However, the service life of an overhead contact line facility is a comprehensive concept and cannot be directly obtained from physical measurements. Therefore, there are no readily available explicit ranking indicators for the facilities to be maintained. Traditionally, the ranking task is completed manually by experts, for example, ranking several images of a facility according to their severity. This transforms the ranking task into establishing a mapping from sample features to a set of sequence numbers. After collecting a large number of labeled ranking samples, considering various data limitations and specifications, algorithms can be used to learn the priority prediction of the samples. However, for a set of n elements, there are n! possible rankings of all elements. As n increases, the computational burden becomes unbearable. The mainstream solution is to predict a priority index for each sample. This index is a numerical value, thus transforming the original problem into ranking the predicted priority indices. This can be achieved using existing sorting algorithms, such as quicksort, mergesort, or insertionsort. Furthermore, due to the large variety and number of facilities on the overhead contact line, for ease of description, this paper will refer to them uniformly as "facilities." Therefore, the core of data-driven sample ranking is fitting a mapping function from samples to priorities, i.e., selecting an appropriate dimensionality reduction mathematical model and obtaining optimized model parameters.

[0067] While establishing a data-driven sample-priority mapping function is not difficult, optimizing its parameters presents a technical challenge, primarily due to the lack of a directly applicable optimization objective. Therefore, current research commonly employs surrogate loss functions, where an additional loss function is designed to optimize model parameters and improve ranking accuracy. These surrogate loss functions fall into three categories.

[0068] The first method is based on a single sample point (point-wise). It compares the predicted priority of each data point with the true priority value given in the dataset using Euclidean distance, and minimizes this distance in a statistically significant way. Therefore, the first method is essentially a regression model.

[0069] The second method is based on the relative differences between sample pairs. It borrows the idea from RankNet's ranking approach, predicting the priority of each sample, inputting the priority difference into a logistic function, and finally using the function's output as the probability that the former sample is prioritized over the latter. Since the two samples belong to the same dataset and their priority order is known, negative log likelihood (NLL) can be used for optimization. Method two is essentially a relative classification problem of sample pairs. Because its training only requires knowing the order of the ground truth priority values, without needing to know the specific values ​​of the ground truth priority values, method two is easier to implement and train than method one.

[0070] Method three builds upon method two by expanding the batch size of data during model training, thus transforming it into a list-wise approach. Because method three utilizes more data in a single training iteration, it can capture a more balanced and effective model optimization trend, improving data utilization efficiency. Another challenge in ranking and prediction algorithms is that the latent variables involved in the prediction may be unevenly distributed or even discontinuous in their space, leading to separate relationships in different data silos. This can cause the model to overfit on existing data and result in poor generalization performance on new samples. This invention aims to address this difficulty by proposing the following variational learning-based ranking method for overhead contact line facility maintenance.

[0071] The overall concept of this invention is:

[0072] First, the processing object of the ranking algorithm's proxy objective function is expanded from single data points or data pairs to a list of data points, and importance sampling point weights are added during this process to strengthen the processing results of key parts. Second, when estimating the priority of samples, the model predicts the probability distribution function parameters of the priority, rather than predicting the expected priority. Finally, in terms of data processing, the prediction model should be able to handle set-type data, supporting data of arbitrary size and dimensions, while maintaining the input order of the prediction. Guided by the above ideas, this invention proposes a maintenance ranking algorithm based on historical inspection feature data for the task of ranking the service health status of numerous facilities in the overhead contact system. Specifically, the topology of the model construction is discussed first, then the training process of combining the neural network with the loss function is discussed, and finally the method of releasing the model after training is explained.

[0073] Model building is a prerequisite for model parameter training and model deployment, and therefore will be stated first. The model built in this invention is a neural network model based on Set-Transformer. This model is chosen because its task is to process set-type data. Therefore, the model must meet two key requirements: First, it should be permutation invariant, meaning that the model's output should not change under any permutation of the elements in the input set. Second, such a model should be able to handle input sets of various sizes. These requirements stem from the characteristic that sorting methods deal with set-type data. Clearly, this specific requirement is incompatible with the construction of most traditional neural networks (such as MLPs, Multi-Layer Perceptrons). The Set-Transformer, which has emerged in recent years, provides a solution to this problem. Therefore, the encoder architecture is based on Set-Transformer to handle set-type inputs. It should be noted that if the sorting task is viewed as a sequence-to-sequence (seq-to-seq) learning problem, it can be easily handled using a recurrent neural network (RNN). However, in reality, the input to the sorting task is not a sequence, but a set. Therefore, Set-Transformer is more suitable for sorting problems than recurrent processing networks. The model used in this invention includes one ISAB module, one PMA module, and one SAB module in a series configuration. These network modules are all built upon the MAB module, and their topology is as follows: Figure 1 As shown. Here, rFF is the feedforward neural network module used in the MAB and PMA blocks, Q, K, and V are the three input ports of the multi-head attention module, and H(·) represents the intermediate computational cost. The above four modules—MAB, SAB, ISAB, and PMA—were proposed by the Set Transformer framework and all follow the usage specifications in the encoder network.

[0074] like Figure 2 As shown, the details of the above four modules are as follows:

[0075] MAB (Multihead Attention Block) is a multihead attention module. It takes two vectors X and Y as input and outputs a single vector. It's important to note that the input and output of MAB are of the same size. The MAB module forms the basis for other modules.

[0076] ISAB (Induced Self-Attention Block) is an induced self-attention module. Its processing satisfies ISAB(X) = MAB(X, MAB(I, X)). The induced self-attention module introduces an implicit function I to reduce computational complexity, thus avoiding the quadratic time complexity O(N) of the SAB module. 2 (N>>1). In short, the induced self-attention module performs better when the scale of the ensemble data is large. The ISAB module receives features of the overhead contact system as system input and then transmits the processed data to the PMA module.

[0077] PMA (Pooling of Multihead Attention) is a multihead attention pooling module. Its purpose is to aggregate a set of learnable seed vectors S and a set of latent vectors X of different sizes. The operation of PMA is as follows: PMA(X) = MAB(S, rFF(X)). With the multihead attention pooling module, the data stream during processing can appropriately change its shape and size. The PMA module receives input from the ISAB module, processes it, and then transmits it to the SAB module.

[0078] SAB (Self-Attention Block) is a self-attention processing module. This module reads a set and performs self-attention among the elements in the set, thereby producing an output of the same size as the set, i.e., SAB(X) = MAB(X,X). SAB receives input from PMA and outputs the mean and variance of the overhead contact line facility health index after processing.

[0079] Here, feature represents the input features, and priority represents the optimal output result.

[0080] The model employed in this invention uses ISAB block stacking to compress sample feature information into a low-dimensional latent space, and then transforms it through PMA and SAB blocks to produce a final prediction with the desired shape. The entire process utilizes the attention of deformable modules. Although the data size and output size are fixed in the description of this invention, the priority index prediction model of this invention is actually universal in terms of both output and input, and therefore can be applied to input feature arrays of any dimension.

[0081] The algorithm of this invention is implemented through a model, the parameters of which need to be trained with data. In railway inspection, a railway section is taken as the macro-unit, containing numerous overhead contact line supports, contact wires, catenary cables, droppers, cantilever arms, and other facilities. Summarizing these facilities and assigning them numbers constitutes the source of the overhead contact line data for a specific section. The model parameter training process is as follows:

[0082] 1. Prepare the dataset Where x i It is the feature vector of the i-th overhead contact line inspection historical feature sample, whose content includes, but is not limited to, grayscale statistics of facility images, Gabor features, wavelet features, ridge wave features, facility maintenance cycle, facility maintenance history, and evaluation indicators, etc., and other diverse and heterogeneous information. i The value of s has been converted to the range of 0 to 1 through proportional changes. i ∈R is the priority index for the i-th sample, which is manually specified by railway experts and is only needed for training data. Here, N is the number of samples, i.e., the total number of overhead contact line facilities to be monitored in this section of the railway.

[0083] 2. Parameter θ of the service health priority prediction model for overhead contact line facilities (t) Perform random initialization assignment θ (t) =θ (0) .

[0084] 3. Randomly sample n samples X = {(x i ,s i ):i∈B}, where This represents the ordered set of selected sample numbers.

[0085] 4. Utilize the priority parameters from these n historical feature samples of overhead contact line inspections. Sort these n samples using the quicksort algorithm γ = argsort(s), where γ represents the sample number after sorting by health level.

[0086] 5. Construct a prediction model f using the Set-Transformer neural network described earlier. θ That is, the priority prediction model, which uses the feature parameters x of these n samples. i Predicting priority indicators of historical characteristic samples from overhead contact line inspections.

[0087]

[0088] Where, μ i , Let represent the predicted mean and variance of each feature of the i-th overhead contact line facility sample, respectively.

[0089] 6. Calculate the prior regularization cost L for these n prediction results. prior =L mean +L var , where L mean L var The meanings are mean constraint and variance constraint, respectively, and the expressions are as follows:

[0090]

[0091] In the formula, γi is the sorting parameter of the feature vector of the i-th overhead contact line facility sample, and μ γi Let be the mean of the feature vectors of the overhead contact system facility samples. Let ε be the variance of the feature vectors of the overhead contact line facility samples, and ε be the error.

[0092] 7. Calculate the importance weight w of the contact network facility health status sequence sample pairs according to the following formula, i.e., w = softmax(ηd), where d = [d γ1 ,…,d γn ] is the importance vector, while d γi =s γi+1 -s γi This represents the priority difference value of the priority ranking samples; in addition, η is an importance parameter. The larger the value, the more it emphasizes the importance of samples with relatively large differences in the ranking. If it is zero, all samples are treated equally.

[0093] 8. Calculate adjacent sample pairs according to the following formula. Loss function L under variational learning likelihood ,

[0094]

[0095] in, The following formula is satisfied.

[0096]

[0097] in, Here, θ represents the weighting coefficients, and θ represents the parameters of the priority prediction model. Indicates according to the feature vector Generate priority index random variables The probability, for The probability of.

[0098] 9. Combining likelihood loss L likelihood and prior loss L prior Calculate the overall loss function value L.

[0099] L = L prior +L likelihood ;

[0100] 10. Let r∈[0,1] be the learning rate, and minimize the loss function L using the mini-batch stochastic gradient descent algorithm.

[0101]

[0102] 11. If the parameters of the priority prediction model are updated, θ(t+1) Compared to the previous θ (t) If the difference is still greater than the set maximum allowable error ∈, then return to step 3.

[0103] 12. Calculate the training parameters θ (t+1) The output is recorded to external storage for long-term preservation.

[0104] 13. End.

[0105] Once the model is trained, the trained model parameters can be used to process any new data, enabling the prediction of priority metrics for the new data. Before executing the prediction process, the model structure needs to be built in memory, and appropriate weights need to be assigned to the model parameters. This preparation process can be obtained from the end state of model training, or by starting a new model and reading the model parameters from the hard drive.

[0106] In summary, after establishing an appropriate model topology and obtaining the model parameters, model prediction can be performed. The processing steps are as follows:

[0107] 1. Create an ordered list F = [(s1, p1), ..., (s n ,p n ], where each element s of the list i p represents the health status of the i-th overhead contact line facility. i This represents the index for accessing the i-th overhead contact line facility.

[0108] 2. If the sorted list is not empty, then use the quicksort algorithm to sort each element according to its priority index s. i Sort the results to obtain an ascending order; the quicksort algorithm can be replaced by any algorithm that can achieve the above sorting.

[0109] 3. Read a new test data x that has been converted to the specified value range. new .

[0110] 4. Read the optimized model parameters from external memory to obtain the trained model f. θ (·), then the data x new Input the data into the model and read its output. And As this data x new The predicted value of the priority indicator.

[0111] 5. Predict the value of this priority indicator. Add to the list of predicted results.

[0112]

[0113] 6. If there is still new data that has not been processed, return to step 2.

[0114] 7. The order of the contents of the ordered list F is the sequence number of the corresponding test data in this sorting operation.

[0115] index(x new ;F)←argmin i∈{1,…,|F|} (F i -f θ (x new )) 2 ,

[0116] Among them, F i This represents the i-th element of the ordered list F, which corresponds to the i-th facility of the overhead contact line for a certain section.

[0117] 8. End.

[0118] This invention can be applied to the data processing center for monitoring the overhead contact system of electrified railways. Its implementation process is divided into four stages: data preprocessing, model topology construction, model parameter training, and model deployment. These four stages are described in detail below:

[0119] (1) During the catenary monitoring data preprocessing stage, the following preprocessing is required for all data in the dataset, including:

[0120] Read the entire overhead contact system data set traversal information of the railway section from external storage, and read samples based on the traversal information;

[0121] First, remove samples containing outliers, missing items, and duplicates;

[0122] Then, a noise reduction algorithm is used to process the samples to reduce the impact of noise.

[0123] Calculate the mean μ of each channel of the sample data, and then calculate the standard deviation σ of the sample.

[0124] Transform the range of each sample x as follows:

[0125]

[0126] The range-converted samples are saved to external storage.

[0127] During the model topology construction phase, data engineers design the model's topology, including the number of layers in the neural network, the number of neurons in each layer, and the parameters computed by each neuron. After completing the model topology construction, the model topology is recorded in the model configuration file as descriptive text (such as JSON or XML).

[0128] During the model training phase, the transformed data is used as the training dataset, and the model parameters are optimized using the method described above; the training result is the optimized model parameters θ. * To satisfy

[0129]

[0130] Once the model has been trained, it can be deployed to predict priority metrics for any new sample. Therefore, the model deployment phase includes:

[0131] The model structure is read from the model configuration file, and the corresponding model algorithm module is built in memory.

[0132] Read model parameters from the hard drive and assign the read values ​​to the corresponding parameters of the model.

[0133] First, the mean μ and variance σ obtained in the data preprocessing stage are used to process the new data x. new Perform a range conversion.

[0134]

[0135] Using the method described above, predictions are made on the new data.

[0136]

[0137] Using binary search, the predicted value Add it to the existing priority-sorted queue.

[0138]

[0139] The ranking result of a new sample can be determined by its position in the priority ranking queue.

[0140] The first, second, and third stages only need to be completed once at the beginning of the algorithm call, while the fourth stage can be executed repeatedly, thus continuously processing new data.

[0141] In summary, the technical solution provided by this invention can clarify the probability distribution of the priority of each ranking index through a priority prediction model. It can effectively generalize to new samples while avoiding overfitting of existing data, thereby improving the accuracy, precision and efficiency of ranking the health of overhead contact line facilities and helping to achieve precise maintenance of overhead contact line facilities.

[0142] Furthermore, in addition to the aforementioned beneficial effects, the preferred solutions also achieve the following beneficial effects: the quicksort algorithm can improve sorting speed; training the priority prediction model can improve the accuracy and efficiency of sorting; multi-dimensional feature vector preprocessing helps to improve the sorting accuracy of the priority prediction model; all of the above methods help to improve the maintenance efficiency of overhead contact line facilities.

Claims

1. A method for scheduling and sequencing maintenance of a catenary installation, characterized in that, The method comprises the following steps: Step S1: constructing a health degree ordered list with data pairs of the overhead contact line facilities as elements, i.e. a first data pair, and if the health degree ordered list is not empty, sorting all data pairs in ascending order according to the priority index, i.e. sorting all overhead contact line facilities; wherein the first data pair comprises the priority index and an index of the overhead contact line facility; constructing a priority prediction model and performing random initialization assignment; Step S2: inputting the feature vector of each overhead contact line facility to be sorted into the priority prediction model to obtain a predicted value of the priority index; Step S3: sequentially adding the data pair of each overhead contact line facility to be sorted, i.e. a second data pair, as a new element into the health degree ordered list; wherein the second data pair comprises the predicted value of the priority index and a corresponding index; Step S4: after adding a new element, sorting all data pairs in ascending order according to the priority index again; The priority index is the health degree of the overhead contact line facility.

2. The method of claim 1, wherein, The health degree ordered list with data pairs of the overhead contact line facilities as elements, i.e. a first data pair, is constructed in detail as follows: constructing an ordered list F of healthiness based on data of the catenary installation (s i ,p i ) F = [(s1, p1),..., (s i ,p i )...,(s n ,p n )] ; where i is the serial number of the overhead contact line facility, i = 1, 2,..., n, n is the number of overhead contact line facilities, s i is the health degree of the i-th overhead contact line facility, p i is the index for accessing the i-th overhead contact line facility.

3. The method of claim 1, wherein, The sorting is performed in ascending order using a quicksort algorithm.

4. The method of claim 1, wherein, The priority prediction model comprises: an ISAB layer, a PMA layer and an SAB layer connected in sequence based on MAB modules; The ISAB layer comprises a first MAB module and a second MAB module; the first MAB module takes the feature vector of the overhead contact line facility to be sorted as input; the second MAB module takes the feature vector of the overhead contact line facility to be sorted and the output of the first MAB module as input; The PMA layer comprises an rFF module and a third MAB module; the rFF module takes the output of the second MAB module as input; the third MAB module takes the output of the rFF module as input; The SAB layer comprises a fourth MAB module; the fourth MAB module takes the output of the third MAB module as input.

5. The method of claim 1, wherein, The training method of the priority prediction model comprises: Step R1: constructing a set X of overhead contact line facility sample data pairs, In the formula, x j is a feature vector of the jth overhead contact line facility sample, s j is a priority index of the jth overhead contact line facility sample, N is the number of overhead contact line facility samples, and j=1, 2, …, N. Step R2: randomly sampling m elements from the set X of overhead contact line facility sample data pairs to form a set X' of sampled data pairs, X' = (xj,s j ): j e B; B∈N; Step R3: sorting the elements in the set X' of sampled data pairs according to the priority index; Step R4: predicting the priority index for the feature vectors of the m overhead contact line facility samples respectively through the priority prediction model fθ, where μ j and respectively represent the mean and variance of the feature vector of the jth catenary facility sample predicted. Step R5: Compute the prior regularized cost L of m predictions prior , L prior = L mean + L var ; wherein L mean is the mean constraint, L var is the variance constraint, wherein γj is the ranking parameter of the feature vector of the jth catenary facility sample, μ γj is the mean of the feature vector of the catenary facility sample, is the variance of the feature vector of the catenary facility sample, and ε is the error. Step R6: calculating the importance weight w of the overhead contact line facility sample data pairs, w=softmax(ηd); wherein η is an importance parameter, denotes the priority indicator of the jth catenary installation sample the difference value to the priority indicator of the adjacent catenary installation sample denotes the difference value to the priority indicator of the adjacent catenary installation sample Step R7: Calculate priority indicator pairs for adjacent overhead line system samples Loss function L under variational learning likelihood , wherein are weight coefficients, and θ is a model parameter, denotes the probability of a feature vector generating a priority indicator random variable s, is the probability of is the probability of Step R8: calculating the total loss function L, L = L prior + L likelihood ; Step R9: Set the learning rate, and minimize the total loss function according to the mini-batch stochastic gradient descent algorithm to obtain the optimized parameters θ of the priority prediction model (t+1) , In the formula, θ (t) is a parameter of the priority prediction model; Step R10: Steps R3 to S9 are repeatedly executed until the error before and after the parameter optimization of the priority prediction model is not greater than the maximum allowed error, and the parameter θ is output (t+1) and save the application, completing the training of the priority prediction model.

6. The method of claim 5, wherein the method further comprises: Before Step R1, the feature vectors of the overhead contact line facility samples are preprocessed in detail as follows: removing outliers, missing items and duplicate items; processing the feature vectors of all overhead contact line facility samples through a denoising algorithm; calculating the standard deviation σ and the mean value μ of each channel of the feature vectors of all overhead contact line facility samples; a feature vector x for each catenary installation sample j performing a range conversion, In the formula, is the range-transformed feature vector of the jth catenary installation sample.

7. The method of claim 1 to 6, characterized in that, The feature vector comprises the maintenance period, maintenance history and evaluation index of the overhead contact line facility, and the gray scale statistics, Gabor features, wavelet features and ridge wave features of the overhead contact line facility image.

Images