Pantograph wear prediction method and system oriented to wear time series and utilizing task model recommendation

Abstract

This invention discloses a method and system for predicting pantograph wear based on wear time series and utilizing task model recommendations. The method includes the following steps: acquiring complete pantograph wear curve data; constructing an offline task set and candidate model library; constructing a degradation state field; obtaining a state field sequence and inputting it into a dynamic task encoder to obtain a task representation; constructing a model latent capability representation; inputting the task representation and model latent capability representation into a relational latent variable encoder to construct task-model relational latent variables; calculating the score vector and ranking score of the candidate model for the current task; ranking the candidate models according to their ranking scores and selecting the model with the highest ranking score; and using the recommended model to predict pantograph wear for subsequent unknown segments. This technical solution utilizes degradation mechanism information from partially observed wear sequences to automatically recommend the prediction model best suited to the task, improving the accuracy and robustness of pantograph wear prediction.

Description

Technical Field

[0001] This invention belongs to the field of pantograph sliding plate wear prediction technology, and relates to a pantograph wear prediction method and system based on wear time series and recommended by task model. Background Technology

[0002] The pantograph contactor is a key component in the rail transit pantograph-catenary current collection system that directly contacts the contact wire. Its wear condition directly affects current collection stability, power supply safety, and maintenance cycle. In recent years, pantograph and contact wire condition monitoring technology has gradually shifted from manual inspection to online and intelligent methods. Correspondingly, the detection methods have also evolved from traditional measurement to machine vision, laser measurement, image segmentation, and multi-sensor fusion.

[0003] Regarding the wear of pantograph sliding plates, recent research has enabled the automatic extraction of remaining thickness, wear contours, and abnormal wear states using visual perception and intelligent recognition methods. This indicates that pantograph sliding plate wear has the technological foundation to move from "detectable" to "predictable".

[0004] In terms of wear prediction and life assessment, existing studies usually regard the thickness or wear of the pantograph plate as a degradation sequence that evolves over time, and use degradation process models, statistical learning methods or deep learning methods to predict its future evolution trend and remaining service life.

[0005] In the pantograph wear scenario, the construction of task representations based on some observed wear sequences, and the further realization of automatic matching and recommendation for candidate model libraries, already have a clear research foundation and engineering requirements.

[0006] Existing technologies can be mainly categorized into the following two types:

[0007] I. Prediction of pantograph sliding plate wear evolution and remaining life:

[0008] This type of technology typically treats the thickness or wear of the skateboard as a degradation sequence that evolves over time, and then uses statistical degradation models, stochastic process models, or neural network models for prediction.

[0009] II. Automatic model selection, meta-learning, and multi-objective model selection techniques for time series forecasting tasks:

[0010] This type of technology is not specifically designed for pantographs, but rather serves a more general time series prediction problem. For example, in the offline phase, a large amount of historical time series datasets are first collected, the meta-features of each dataset are calculated, and the performance of different prediction models on these datasets is statistically analyzed. Then, a meta-learner is trained, and in the online phase, after inputting the features of a new time series or a new time window, it quickly infers the most suitable prediction model without having to retrain and evaluate all candidate models.

[0011] While the aforementioned existing technologies have made progress in "lifetime prediction" and "automatic model selection," they still have the following shortcomings when addressing the specific problem of pantograph wear curve completion and model recommendation:

[0012] Existing skateboard wear / life prediction technologies typically rely on a fixed degradation model or fixed network structure, lacking the ability to adapt the model selection to the task.

[0013] The actual degradation process of the pantograph slide plate is affected by a variety of factors, such as the operating section, contact state, local anomalies, measurement noise, and current collection differences of slide plates at different locations. The statistical characteristics of different sequences are not consistent in the steady stage, the local mutation stage, and the accelerated wear stage. A single fixed model is often difficult to maintain stable accuracy in all tasks at the same time.

[0014] While existing general-purpose time series automatic model selection technologies have recognized that "different tasks are suitable for different models," there is still a significant gap between their task definitions and mechanical degradation scenarios. AutoForecast, AutoXPCR, and related dynamic model selection patents typically describe the task as a meta-feature description of "a dataset" or "a time window," and then predict the optimal model or optimal model set based on this. The features they use tend to be general statistics, spectral features, or resource complexity indicators, lacking explicit modeling of degradation mechanism information such as "slow trends, local fluctuations, abrupt disturbances, and changes during the observation phase" in the mechanical degradation process.

[0015] For typical degradation sequences like pantograph wear, it is difficult to fully characterize key issues such as "what degradation stage the current task is in, whether there are any abnormal fluctuations in the curve recently, and whether different prediction models are applicable to this stage" by using only general meta-features or general dataset attributes for model selection. Summary of the Invention

[0016] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a pantograph wear prediction method and system that is oriented towards wear time series and utilizes task model recommendations.

[0017] To achieve the above objectives, the basic solution of this invention is: a pantograph wear prediction method based on wear time series and recommended by a task model, comprising the following steps:

[0018] S1. Obtain complete pantograph wear curve data and construct an offline task set and candidate model library;

[0019] S2, construct the degradation state field for the observed wear sequence of each task to obtain the state field sequence;

[0020] S3, input the state field sequence into the dynamic task encoder to obtain the task representation;

[0021] S4. Based on the meta-information of candidate models and their applicability in historical stages, construct a representation of the potential capabilities of the models.

[0022] S5. For each candidate model, the task representation and the model latent capability representation are mapped to the relational interaction space respectively. The common activation terms and mismatch terms of the two are calculated. The common activation terms, mismatch terms, task representation and model latent capability representation are fused to obtain the task-model relational latent variables.

[0023] Input the task representation, the model's latent capability representation, and the latent variables of the task-model relationship into the matching scoring head, and output the candidate model's scoring vector and ranking score for the current task;

[0024] S6. Sort the candidate models according to their ranking scores and select the model with the highest ranking score.

[0025] S7, using the recommended model, the target pantograph is predicted to move at a certain step length. Predicting unknown wear sequences within the system; in the offline training task... This represents the length of the unknown future segment after the truncation point in the complete curve; in online applications, the actual data is sampled by vehicle number, day, kilometer, or test batch. This refers to the number of sampling points, the detection cycle, or the operating mileage interval. If the data sampling interval is... The corresponding actual prediction duration is .

[0026] The working principle and beneficial effects of this basic scheme are as follows: Under the condition that only the observed wear sequence of the target pantograph is obtained, this technical scheme automatically recommends a prediction model that is more suitable for the current task from the candidate prediction model library, providing a more suitable model selection basis for subsequent wear prediction of unknown sections.

[0027] To address the significant differences in model applicability across different task stages, fluctuation characteristics, and degradation modes in pantograph wear time series, this paper establishes a model recommendation method capable of extracting degradation state information from partial observed wear sequences, characterizing the task-model matching relationship, and outputting candidate model ranking results. This aims to improve the accuracy and stability of model selection in pantograph wear prediction scenarios. The method solves the problems of poor adaptability of fixed models and the lack of degradation mechanism modeling in general model selection methods in existing technologies, thereby improving the relevance and stability of model recommendation in pantograph wear scenarios.

[0028] Furthermore, the method for obtaining complete pantograph wear curve data in step S1 and constructing the offline task set and candidate model library is as follows:

[0029] Let the complete wear curve y of a certain pantograph be represented as:

[0030] ,

[0031] in, Indicates the first The wear observation value at a given moment, where T represents the total lifespan of the wear curve;

[0032] For any cutoff ratio The cutoff position is defined as:

[0033] ,

[0034] in, For a preset set of truncation ratios, Indicates the end position of the observable sequence in the current task; [] indicates rounding down.

[0035] Construct a task to be recommended:

[0036] ,

[0037] in, To represent a task, This indicates the curve number, vehicle number, or sequence identifier; the input for this task is the observed sequence. for:

[0038] ,

[0039] in, This indicates the wear observation value at the end position of the observable sequence in the current task;

[0040] The truth value of the corresponding unknown future segment for:

[0041] ,

[0042] Let the candidate model library be:

[0043] ,

[0044] in, Represents the set of candidate prediction models. Indicates the number of candidate models. Indicates the first There are 10 candidate prediction models.

[0045] By amplifying the training task through multi-proportion truncation, the diversity of training data was effectively increased under limited sample conditions, providing sufficient supervision signals for subsequent model recommendation learning.

[0046] Furthermore, step S2 involves constructing a degradation state field for each task's observed wear sequence to obtain the state field sequence.

[0047] For the observed wear sequence Perform trend, mutation, and noise decomposition:

[0048] ,

[0049] in, Indicates the first Wear observation values ​​at a given moment, Indicates the first Slowly degenerating skeleton at a given moment Indicates the first Sparse mutations at time 1; Indicates the first High-frequency noise term at each moment;

[0050] To solve and Establish optimization goals:

[0051] ,

[0052] in, Indicates the end position of the observable sequence in the current task. For robust loss function; This represents the second-order difference operator. , Used to encourage the sparse occurrence of mutations For trend smoothing weights, For mutation sparse weights;

[0053] In obtaining Then, in multi-scale assemblies The above constructs state field characteristics, for any scale Definition:

[0054] Multiscale slope :

[0055] ,

[0056] Indicates the first At any moment on the scale The rate of change of local trends on;

[0057] Multiscale curvature :

[0058] ,

[0059] Indicates the acceleration of a trend change;

[0060] Local fluctuation intensity :

[0061] ,

[0062] in, Representing scale The mean of the noise term within the lower window; This indicates the intensity of the fluctuation during that period; This represents the high-frequency noise term at time e;

[0063] Significance of mutations :

[0064] ,

[0065] in, To prevent extremely small positive numbers with a denominator of zero, The larger the value, the more likely a significant mutation is to occur at that moment;

[0066] Introducing relative time position :

[0067] ,

[0068] time The state field token is represented as :

[0069] ,

[0070] The entire observed sequence is thus represented as a state field sequence. :

[0071]

[0072] in, This represents the state field sequence indicating the termination position of the observable sequence in the current task.

[0073] By using trend-mutation-noise decomposition and multi-scale feature extraction, the original wear sequence is elevated into a state field representation with degradation mechanism implications, which can more effectively reflect key information such as slow degradation, local fluctuations and anomalous mutations.

[0074] Furthermore, in step S3, the state field sequence is input into the dynamic task encoder to obtain the task representation, as follows:

[0075] Let the dimension of the state field token be . To hide dimensions in subsequent coding networks Alignment, first perform linear mapping:

[0076] ,

[0077] in, For the input projection matrix, For bias vectors, For the first The initial embedding representation at each time step; For a moment State field token;

[0078] Let the set of multi-scale void ratios D be:

[0079] ,

[0080] For the first Individual residual blocks, scale Define the causal convolution branch output. for:

[0081] ,

[0082] in, Indicates from the first moment to the second moment. The hidden state of history at a certain moment Indicates the void ratio The branches consist of causal convolution, normalization, activation, and dropout;

[0083] Fusion is performed using scale attention weights, denoted as _i_. The weights for different scales at each time point are:

[0084] ,

[0085] in, The scale attention logits generated from the current hidden state, Indicates the first At any given moment, the scale The level of attention; for;

[0086] Fusion output for:

[0087] ,

[0088] After residual connection, we get:

[0089] ,

[0090] in, Representation layer normalization, This represents the residual projection matrix when the input and output dimensions are the same. Can take identity mapping; Indicates the first Task at a specific moment Hidden state after residual connection;

[0091] go through After several residual blocks, the hidden state sequence is obtained. :

[0092] ,

[0093] in, Indicates task The last moment of the observation sequence, Indicates task The hidden state at the last moment of the observation sequence;

[0094] Three aggregation methods are used:

[0095] Global aggregation:

[0096] ,

[0097] in, Indicates task The global aggregation representation; Indicates task The hidden state at time i;

[0098] Nearest-stage aggregation: Let the length of the nearest window be... ,but

[0099] ,

[0100] in, Indicates task The most recent stage aggregation representation;

[0101] Mutation-weighted aggregation: Let the significance of the mutation be . ,but

[0102] ,

[0103] ,

[0104] in, This is the amplification factor for the sudden change. Indicates the first The mutation weight at each time step; Indicates the significance of the mutation. Indicates task Mutation-weighted aggregation representation;

[0105] Construct task scalar feature vector Map it to a space of the same dimension as the sequence representation:

[0106] ,

[0107] in, This represents the scalar representation of the task. Let be the task scalar projection matrix. The task scalar bias vector;

[0108] Concatenate the three sequence representations:

[0109] ,

[0110] in, Indicates task Sequence degradation characterization;

[0111] Constructing gate vectors :

[0112] ,

[0113] in, This represents the Sigmoid function. The components are located at interval; This is the gated mapping matrix, also known as the gated weight matrix. This is the gate bias vector, used to adjust the reference offset of the gated output;

[0114] The final task is represented as:

[0115] ,

[0116] in, Represents element-wise product. Indicates task The final potential representation.

[0117] While satisfying the causal constraints of online scenarios, it can simultaneously capture long-term degradation trends, recent state changes, and abnormal shock information, and enhance the comprehensiveness and adaptability of task representation by fusing scalar features through a gating mechanism.

[0118] Furthermore, step S4, based on the meta-information of the candidate models and their applicability in historical stages, involves constructing a representation of the model's potential capabilities as follows:

[0119] For candidate models Construct its original model feature vector :

[0120] ,

[0121] in, Represents the static attribute features of the model. A signature indicating the model's historical applicability at different observation stages;

[0122] Suppose the source domain tasks are divided according to the observation ratio as follows: The first stage subset, the Model at each stage Statistical applicability of experience for:

[0123] ,

[0124] in, Indicates the first The set of tasks for each stage. This indicates the number of tasks in the set. Indicates task upper model Offline overall rating;

[0125] but:

[0126] ,

[0127] Mapping the original model features to the model capability space:

[0128] ,

[0129] in, Candidate models The model capability space, Here, represents the model capability projection matrix, and represents the learnable parameters. This represents both the model's capability bias and the learnable parameters. LN(⋅) is the mapping function, representing the layer normalization operation;

[0130] Introducing learnable residual vectors The final model is represented as:

[0131] ,

[0132] in, Representing candidate models The potential capabilities are represented.

[0133] By combining meta-information, empirical applicability signatures, and learnable residuals, a unified representation of the static attributes, historical performance, and individual differences of the model is achieved, thus solving the applicability representation problem when there are significant differences in model types.

[0134] Furthermore, for each candidate model, the task representation and the model's latent capability representation are mapped to the relational interaction space, respectively. The common activations and mismatches of the two are calculated, and the common activations, mismatches, task representation, and model latent capability representation are fused to obtain the task-model relational latent variables, as follows:

[0135] ,

[0136] in, This is the model mapping matrix, used to project the model representation onto a space that can interact multiplicatively with the task; This is the task mapping matrix, used to project the task representation onto a space that can be compared with the model; This represents the portion activated by both task requirements and model capabilities; This indicates the degree of mismatch between the task and the model; This indicates vector concatenation; It is a smooth non-negative activation function; Represents the latent variables of the task-model relationship.

[0137] The shared activation of the modeling task requirements and the model capabilities, and the degree of mismatch between the two, directly characterize the matching relationship, rather than simply splicing or comparing distances, thus improving the accuracy of the matching.

[0138] Furthermore, the task representation, the latent capability representation of the candidate model, and the latent variables of the task-model relationship are input into the matching scoring head to obtain the score vector and ranking score of the candidate model for the current task. The specific method is as follows:

[0139] Define the baseline for task difficulty:

[0140] ,

[0141] in, This is a task difficulty mapping matrix. For bias terms, This indicates the baseline difficulty of the current task across all scoring dimensions;

[0142] Define the overall bias of the model :

[0143] ,

[0144] in, For the model bias table, Representation Model The bias vector;

[0145] Relationship-driven scoring items Defined as:

[0146] ,

[0147] in, This is a relational rating mapping matrix;

[0148] Final score vector for:

[0149] ,

[0150] Weighted fusion is used:

[0151] ,

[0152] in, Indicates the sorted score; This is the weight vector;

[0153] according to Sort the data from largest to smallest to get the model's recommendation results.

[0154] The multi-objective score is generated by combining three parts: task difficulty baseline, overall model bias, and relationship-driven scoring items. It also takes into account the overall prediction error and shape preservation ability, so that the recommendation results are more in line with the engineering requirements for the rationality of wear curve shape.

[0155] Furthermore, it also includes the joint loss function L:

[0156] ,

[0157] in, Indicates the loss due to the return of supervision. Represents the sorting loss. Indicates the relationship between loss and comparison. These are the corresponding weighting coefficients;

[0158] Supervision regression loss:

[0159] ,

[0160] This represents the mean squared error between the predicted score vector and the offline supervision label; where For the rating vector, Let M represent a task, M be a candidate model library, and m be a candidate prediction model;

[0161] Ranking loss:

[0162] For the task The model with the higher overall score in the true mean is denoted as The lower model is denoted as Then the definition is:

[0163] ,

[0164] in, This indicates the model in the actual rating. Should be ranked Front; This represents the ranking score predicted by the model with the higher actual overall score in the recommendation network; This represents the ranking score predicted by the model with the lower actual overall score in the recommendation network; this loss is used to constrain the model with better actual performance to obtain a higher predicted ranking score.

[0165] Relational contrast loss:

[0166] set up Indicates task A collection of high-resolution models Let represent the set of low-scoring models, then we have:

[0167] ,

[0168] in, This is the direction vector of the relational quality. Temperature coefficient; Represents the latent variables of the task-model relationship.

[0169] By jointly optimizing supervised regression loss, ranking loss, and relational comparison loss, the system simultaneously ensures the accuracy of rating prediction, the consistency of model ranking, and the discriminativeness of relational latent variables, thereby enhancing the generalization ability of the recommendation system.

[0170] This invention also provides a pantograph wear prediction system based on the method described herein, oriented towards wear time series and utilizing a task model for recommendation, comprising:

[0171] The data acquisition module is used to acquire a portion of the observed wear sequence of the target pantograph;

[0172] The state field construction module is used to construct a degraded state field from the partially observed wear sequence to obtain a state field sequence.

[0173] The task encoding module includes a dynamic task encoder for encoding the state field sequence into a task representation;

[0174] The model representation module is used to store and provide a representation of the potential capabilities of each candidate model in the candidate model library;

[0175] The relation encoding module includes a relation latent variable encoder, used to construct task-model relation latent variables based on the task representation and the model potential capability representation of any candidate model;

[0176] The scoring and recommendation module includes a matching scoring header, which generates a scoring vector and ranking score for the candidate model based on the task representation, the model's potential ability representation, and the latent variables of the task-model relationship, and outputs a recommended model based on the ranking scores of each candidate model.

[0177] This system modularizes the methodology, achieving full-process automation from data acquisition, state field construction, task coding to model recommendation, and possesses excellent engineering deployability and scalability.

[0178] Furthermore, the dynamic task encoder includes an input projection layer, a multi-scale causal dilated convolution branch, a scale attention fusion layer, a residual normalization layer, a three-view aggregation layer, and a scalar gated fusion layer connected in sequence.

[0179] The input projection layer maps state field tokens to hidden dimensions; the multi-scale causal dilated convolution branch extracts degradation features at different time scales; the scale attention fusion layer weights and fuses the outputs of branches with different dilation rates; the three-view aggregation layer generates global aggregation representation, recent stage aggregation representation, and mutation-weighted aggregation representation, respectively; and the scalar gated fusion layer fuses sequence degradation representation and task scalar representation into a task representation. The relational latent variable encoder includes a task projection layer, a model projection layer, a co-activation computation unit, a mismatch computation unit, a splicing fusion unit, and a non-negative activation output unit.

[0180] The input of the task projection layer is connected to the task encoding module, the input of the model projection layer is connected to the model representation module, the common activation calculation unit and the mismatch calculation unit receive the task projection result and the model projection result respectively, the stitching and fusion unit receives the common activation term and the mismatch term, and the non-negative activation output unit outputs the latent variable of the task-model relationship.

[0181] The matching scoring head includes a task difficulty baseline branch, a model overall bias branch, and a relationship-driven scoring branch set in parallel. The outputs of the three branches are connected to a scoring vector fusion unit, which is connected to a ranking score calculation unit to generate candidate model ranking scores based on the scoring dimension weights.

[0182] The structure of each module is designed to facilitate use. Attached Figure Description

[0183] Figure 1 This is a flowchart illustrating the pantograph wear prediction method of the present invention, which is based on wear time series and utilizes a task model for recommendation.

[0184] Figure 2 This is a schematic diagram of the process for obtaining the state field sequence of the pantograph wear prediction method recommended by the task model based on wear time series in this invention.

[0185] Figure 3 This is a schematic diagram of the process for obtaining the task representation of the pantograph wear prediction method recommended by the task model based on wear time series in this invention;

[0186] Figure 4 This is a flowchart illustrating the process of obtaining the ranking scores of candidate models for the current task in the pantograph wear prediction method recommended by the present invention, which is oriented towards wear time series and utilizes task models. Detailed Implementation

[0187] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0188] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0189] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0190] To address the shortcomings of existing pantograph sliding plate wear evolution prediction technologies, which typically rely on a single fixed prediction model and struggle to adapt to different wear stages and sequence characteristics, as well as the lack of specific modeling for mechanical degradation mechanisms and pantograph wear scenarios in existing general-purpose time series automatic model selection technologies, this invention discloses a pantograph wear prediction method oriented towards wear time series and utilizing task model recommendations. This method automatically recommends prediction models that better match the current task from a candidate prediction model library, even when only the observed wear sequence of the target pantograph is available, thus providing a more suitable model selection basis for subsequent wear prediction of unknown segments.

[0191] like Figure 1 As shown, the pantograph wear prediction method based on wear time series and recommended by the task model includes the following steps:

[0192] S1. Obtain complete pantograph wear curve data and construct an offline task set and candidate model library;

[0193] S2, construct the degradation state field for the observed wear sequence of each task to obtain the state field sequence;

[0194] S3. Input the state field sequence into the dynamic task encoder to obtain the task representation. Considering the characteristics of slow degradation, local fluctuations, abrupt disturbances, and different degradation stages corresponding to different observation lengths that are common in pantograph wear sequences, construct a task representation that can reflect the true state of the current prediction task, rather than simply treating the observed sequence as a general static time window.

[0195] S4. Based on the meta-information of candidate models and their applicability in historical stages, construct a model potential capability representation. Under the condition that the number of candidate prediction models is limited but the model types are significantly different, establish a unified model representation that can characterize the applicability of the model, and further characterize the matching relationship between "current task requirements" and "model capability characteristics", rather than simply relying on fixed empirical rules or single statistical features for model selection.

[0196] S5. For each candidate model, the task representation and the model latent capability representation are mapped to the relational interaction space respectively. The common activation terms and mismatch terms of the two are calculated. The common activation terms, mismatch terms, task representation and model latent capability representation are fused to obtain the task-model relational latent variables.

[0197] Input the task representation, the model's latent capability representation, and the latent variables of the task-model relationship into the matching scoring head, and output the candidate model's scoring vector and ranking score for the current task;

[0198] S6. Sort the candidate models according to their ranking scores and select the model with the highest ranking score as the recommended model. Construct a scoring prediction mechanism suitable for pantograph wear prediction scenarios, so that it can comprehensively reflect multiple objective factors such as overall prediction error and shape retention capability, and perform more reasonable ranking and screening of candidate models, thereby outputting a reliable Top-K recommendation result.

[0199] S7, using the recommended model, the target pantograph is predicted to move at a certain step length. Predicting unknown wear sequences within the system; in the offline training task... This represents the length of the unknown future segment after the truncation point in the complete curve; in online applications, the actual data is sampled by vehicle number, day, kilometer, or test batch. For the number of sampling points, detection cycle, or running mileage interval, if the data sampling interval is... The corresponding actual prediction duration is In online scenarios where only preliminary observation data of the target pantograph is available and complete future true values ​​are lacking, model recommendation for new tasks can be implemented to improve the relevance, stability, and engineering usability of model selection, and reduce the decrease in prediction accuracy and maintenance decision bias caused by model misselection.

[0200] This invention balances overall prediction error with shape preservation capability, thereby improving the engineering effectiveness of the recommendation results. Instead of relying solely on a single error metric, this invention considers both overall error-related scores and shape preservation-related scores in both offline labeling and online scoring.

[0201] Since pantograph wear sequences require not only that future numerical values ​​closely approximate the actual trend, but also that the predicted curve conforms to the wear mechanism in terms of monotonicity, smoothness, local oscillations, and the reasonableness of abrupt changes, simply minimizing a single numerical error is insufficient to guarantee engineering usability. This invention creates a recommendation process driven by a multi-objective scoring vector and a weighted ranking score, ensuring that the recommended model not only performs better in terms of overall error but also better conforms to the physical laws of pantograph wear curves in terms of shape reasonableness. This reduces the probability of models with seemingly small numerical errors but significantly distorted curve shapes being incorrectly recommended, thus improving the engineering reliability of the recommendation results.

[0202] This invention enables cold-start model recommendation in online scenarios with only preliminary observation data, thereby enhancing the feasibility of practical deployment. In real-world maintenance scenarios, the target pantograph typically only has the observed wear sequence of the preliminary phase; the ground truth for future phases is unavailable, and it is not suitable to retrain all candidate models completely before comparison. This invention learns the task-model matching rules in the offline phase, and in the online phase, it only requires the input of the currently observed wear sequence of the target pantograph to generate a task representation, query the potential capabilities of candidate models, and output Top-K recommendation results. Since this method does not rely on the future ground truth of the target task and does not require costly online full comparison of all candidate models, it can significantly improve the efficiency of online model selection, meeting the real-time and deployability requirements of pantograph maintenance scenarios.

[0203] In a preferred embodiment of the present invention, the method for obtaining complete pantograph wear curve data in step S1 and constructing an offline task set and candidate model library is as follows:

[0204] Let the complete wear curve y of a certain pantograph be represented as:

[0205] ,

[0206] in, Indicates the first The wear observation value at a given moment, where T represents the total lifespan of the wear curve;

[0207] For any cutoff ratio The cutoff position is defined as:

[0208] ,

[0209] in, For a preset set of truncation ratios, Indicates the end position of the observable sequence in the current task; [] indicates rounding down.

[0210] Construct a task to be recommended:

[0211] ,

[0212] in, To represent a task, This indicates the curve number, vehicle number, or sequence identifier; the input for this task is the observed sequence. for:

[0213] ,

[0214] in, This indicates the wear observation value at the end position of the observable sequence in the current task;

[0215] The truth value of the corresponding unknown future segment for:

[0216] ,

[0217] Let the candidate model library be:

[0218] ,

[0219] in, Represents the set of candidate prediction models. Indicates the number of candidate models. Indicates the first There are 10 candidate prediction models.

[0220] For small sample sets, construct sufficient training tasks from a limited number of complete wear curves, and segment each complete curve according to multiple cutoff ratios. For example, you can set:

[0221] ,

[0222] Each complete curve can then generate multiple sub-tasks with different observation lengths. This can simulate the situation in actual engineering where "model selection is required only after observing the initial wear sequence".

[0223] The candidate model library can consist of various heterogeneous prediction models, such as statistical models, parameter degradation models, state-space models, probabilistic models, and deep learning models. To ensure that model-side information can be uniformly encoded, this invention creates a model meta-information vector for each candidate model, used to describe its modeling method, whether it has a monotonically degenerate prior, whether it supports online updates, training cost, inference cost, and whether the output contains uncertainty, among other attributes.

[0224] In a preferred embodiment of the present invention, such as Figure 2 As shown, step S2 involves constructing the degradation state field for each task's observed wear sequence to obtain the state field sequence.

[0225] The pantograph wear sequence is not a typical static time window, but a dynamic object with clear degradation mechanism characteristics. Its typical characteristics are: the main wear trend accumulates slowly over time; local fluctuations reflect changes in operating conditions and contact conditions; local anomalies or disturbances can lead to abrupt changes; and the same curve corresponds to different degradation stages at different observation lengths.

[0226] For the observed wear sequence Perform trend, mutation, and noise decomposition:

[0227] ,

[0228] in, Indicates the first Wear observation values ​​at a given moment, Indicates the first Slowly degenerating skeleton at a given moment Indicates the first Sparse mutations at time 1; Indicates the first High-frequency noise term at each moment;

[0229] To solve and Establish optimization goals:

[0230] ,

[0231] in, Indicates the end position of the observable sequence in the current task. For robust loss function; This represents the second-order difference operator. , Used to encourage the sparse occurrence of mutations For trend smoothing weights, For mutation sparse weights;

[0232] In obtaining Then, in multi-scale assemblies The above constructs state field characteristics, for any scale Definition:

[0233] Multiscale slope :

[0234] ,

[0235] Indicates the first At any moment on the scale The rate of change of local trends on;

[0236] Multiscale curvature :

[0237] ,

[0238] Indicates the acceleration of a trend change;

[0239] Local fluctuation intensity :

[0240] ,

[0241] in, Representing scale The mean of the noise term within the lower window; This indicates the intensity of the fluctuation during that period; This represents the high-frequency noise term at time e;

[0242] Significance of mutations :

[0243] ,

[0244] in, To prevent extremely small positive numbers with a denominator of zero, The larger the value, the more likely a significant mutation is to occur at that moment;

[0245] Introducing relative time position :

[0246] ,

[0247] time The state field token is represented as :

[0248] ,

[0249] The entire observed sequence is thus represented as a state field sequence. :

[0250]

[0251] in, This represents the state field sequence indicating the termination position of the observable sequence in the current task.

[0252] The original wear numerical sequence is upgraded into a state field representation with mechanistic interpretation, enabling the subsequent encoder to distinguish between "main trend changes", "local disturbances", "stage positions" and "abnormal mutations".

[0253] In a preferred embodiment of the present invention, such as Figure 3 As shown, step S3 inputs the state field sequence into the dynamic task encoder to obtain the task representation as follows:

[0254] Let the dimension of the state field token be . To hide dimensions in subsequent coding networks Alignment, first perform linear mapping:

[0255] ,

[0256] in, For the input projection matrix, For bias vectors, For the first The initial embedding representation at each time step; For a moment State field token;

[0257] Let the set of multi-scale void ratios D be:

[0258] ,

[0259] For the first Individual residual blocks, scale Define the causal convolution branch output. for:

[0260] ,

[0261] in, Indicates from the first moment to the second moment. The hidden state of history at a certain moment Indicates the void ratio The branches consist of causal convolution, normalization, activation, and dropout;

[0262] Fusion is performed using scale attention weights, denoted as _i_. The weights for different scales at each time point are:

[0263] ,

[0264] in, The scale attention logits generated from the current hidden state, Indicates the first At any given moment, the scale The level of attention; for;

[0265] Fusion output for:

[0266] ,

[0267] After residual connection, we get:

[0268] ,

[0269] in, Representation layer normalization, This represents the residual projection matrix when the input and output dimensions are the same. Can take identity mapping; Indicates the first Task at a specific moment Hidden state after residual connection;

[0270] go through After several residual blocks, the hidden state sequence is obtained. :

[0271] ,

[0272] in, Indicates task The last moment of the observation sequence, Representing the task The hidden state at the last moment of the observation sequence; task u is a time series that only observes a portion of the time points;

[0273] Three aggregation methods are used:

[0274] Global aggregation:

[0275] ,

[0276] in, Indicates task The global aggregation representation; Indicates task The hidden state at time i;

[0277] Nearest-stage aggregation: Let the length of the nearest window be... ,but

[0278] ,

[0279] in, Indicates task The most recent stage aggregation representation;

[0280] Mutation-weighted aggregation: Let the significance of the mutation be . ,but

[0281] ,

[0282] ,

[0283] in, This is the amplification factor for the sudden change. Indicates the first The mutation weight at each time step; Indicates the significance of the mutation. Indicates task Mutation-weighted aggregation representation;

[0284] Construct task scalar feature vector Examples include the proportion of observations, sequence length, location category, and local statistics. These are mapped to a space of the same dimension as the sequence representation.

[0285] ,

[0286] in, This represents the scalar representation of the task. Let be the task scalar projection matrix. The task scalar bias vector;

[0287] Concatenate the three sequence representations:

[0288] ,

[0289] in, Indicates task Sequence degradation characterization;

[0290] Constructing gate vectors :

[0291] ,

[0292] in, This represents the Sigmoid function. The components are located at interval; This is the gated mapping matrix, also known as the gated weight matrix. This is the gate bias vector, used to adjust the reference offset of the gated output;

[0293] The final task is represented as:

[0294] ,

[0295] in, Represents element-wise product. Indicates task The final potential representation.

[0296] In a preferred embodiment of the present invention, step S4, which involves constructing a representation of the model's potential capabilities based on the meta-information of the candidate model and its applicability across historical stages, is as follows:

[0297] For candidate models Construct its original model feature vector :

[0298] ,

[0299] in, This indicates the static attribute features of the model, such as model family, whether it is a probabilistic model, whether it has a monotonic prior, and whether it supports online updates. A signature indicating the model's historical applicability at different observation stages;

[0300] Phase applicability signatures can be obtained through offline statistics. Assume the source domain tasks are divided according to the observation ratio. The first stage subset, the Model at each stage Statistical applicability of experience for:

[0301] ,

[0302] in, Indicates the first The set of tasks for each stage. This indicates the number of tasks in the set. Indicates task upper model Offline overall rating;

[0303] but:

[0304] ,

[0305] Mapping the original model features to the model capability space:

[0306] ,

[0307] in, Candidate models The model capability space, Here, represents the model capability projection matrix, and represents the learnable parameters. This represents both the model's capability bias and the learnable parameters. LN(⋅) is the mapping function, representing the layer normalization operation;

[0308] To preserve individual model differences, a learnable residual vector is introduced. The final model is represented as:

[0309] ,

[0310] in, Representing candidate models The potential capabilities are represented.

[0311] This invention establishes a unified and physically meaningful model applicability characterization, thereby improving the model selection effect when the number of candidate models is limited but the types are very different.

[0312] This invention does not treat candidate prediction models as merely isolated numbers, but rather constructs a representation of the model's potential capabilities through model meta-information, historical stage applicability signatures, and learnable residuals. The model meta-information reflects the model's modeling paradigm, prior constraints, online update capability, and complexity characteristics; the stage applicability signature reflects the model's historical performance at different degradation stages; and the learnable residuals supplement individual model differences. Because this model representation method employs "explicit attribute priors + stage experience summaries + residual correction," it can more accurately express the applicability of each model even when the number of models is small and the differences in model types are significant, avoiding the coarse model selection problem caused by relying solely on fixed empirical rules or a single score mean.

[0313] In a preferred embodiment of the present invention, the method for recommending models for a specific task actually involves a matching relationship between "current task requirements" and "model capability features," rather than them being independent. For each candidate model, the task representation and the model's latent capability representation are mapped to a relational interaction space, respectively. The common activation terms and mismatch terms of the two are calculated, and the common activation terms, mismatch terms, task representation, and model latent capability representation are fused to obtain the task-model relational latent variables as follows:

[0314] ,

[0315] in, This is the model mapping matrix, used to project the model representation onto a space that can interact multiplicatively with the task; This is the task mapping matrix, used to project the task representation onto a space that can be compared with the model; This represents the portion activated by both task requirements and model capabilities; This indicates the degree of mismatch between the task and the model; This indicates vector concatenation; It is a smooth non-negative activation function; Represents the latent variables of the task-model relationship.

[0316] The latent variable of this relationship does not pre-specify the specific meaning of each dimension, but rather learns potential matching factors that "a certain type of task is more suitable for a certain type of model" automatically through training.

[0317] This invention improves the accuracy of candidate model ranking by directly learning matching patterns through latent variables of task-model relationships.

[0318] Existing general-purpose automatic model selection methods typically use general statistical meta-features to predict the optimal model, which struggles to characterize the more nuanced "demand-capability" correspondence between the mechanical degradation task and candidate models. This invention does not force the task representation and model representation to share a predefined set of semantic axes. Instead, it directly constructs task-model relational latent variables through a relational latent variable encoder and establishes a matching scoring head based on this. Since the relational latent variables simultaneously consider the common activation components of task requirements and model capabilities, as well as the degree of mismatch between them, the scoring process is no longer a simple addition or concatenation of task features and model features. Instead, it explicitly models the core question of "whether the current task is suitable for the current model." This more accurately reflects the suitability of different candidate models for the current pantograph wear task, improving the accuracy and reliability of the recommendation ranking.

[0319] In a preferred embodiment of the present invention, such as Figure 4 As shown, the task representation, the latent capability representation of the candidate model, and the latent variables of the task-model relationship are input into the matching scoring head to obtain the score vector and ranking score of the candidate model for the current task. The specific method is as follows:

[0320] Define the baseline for task difficulty:

[0321] ,

[0322] in, This is a task difficulty mapping matrix. For bias terms, This indicates the baseline difficulty of the current task across all scoring dimensions;

[0323] Define the overall bias of the model :

[0324] ,

[0325] in, For the model bias table, Representation Model The bias vector;

[0326] Relationship-driven scoring items Defined as:

[0327] ,

[0328] in, This is a relational rating mapping matrix;

[0329] Final score vector for:

[0330] ,

[0331] when hour:

[0332] To obtain the final ranking score, a weighted fusion method is used:

[0333] ,

[0334] in, Indicates the sorted score; This is the weight vector;

[0335] according to Sort the data from largest to smallest to get the model's recommendation results.

[0336] In a preferred embodiment of the present invention, in order to simultaneously ensure the accuracy of score prediction, the consistency of ranking, and the discriminative nature of relation representation, the pantograph wear prediction method further includes a joint loss function L:

[0337] ,

[0338] in, Indicates the loss due to the return of supervision. Represents the sorting loss. Indicates the relationship between loss and comparison. These are the corresponding weighting coefficients;

[0339] Supervision regression loss:

[0340] ,

[0341] This represents the mean squared error between the predicted score vector and the offline supervision label; where For the rating vector, Let M represent a task, M be a candidate model library, and m be a candidate prediction model;

[0342] Ranking loss:

[0343] For the task The model with the higher overall score in the true mean is denoted as The lower model is denoted as Then the definition is:

[0344] ,

[0345] in, This indicates the model in the actual rating. Should be ranked Front; This represents the ranking score predicted by the model with the higher actual overall score in the recommendation network; This represents the ranking score predicted by the model with the lower actual overall score in the recommendation network; this loss is used to constrain the model with better actual performance to obtain a higher predicted ranking score.

[0346] Relational contrast loss:

[0347] set up Indicates task A collection of high-resolution models Let represent the set of low-scoring models, then we have:

[0348] ,

[0349] in, This is the direction vector of the relational quality. Temperature coefficient; Represents the latent variables of the task-model relationship.

[0350] This loss causes the latent relation variables corresponding to the high-scoring model and the latent relation variables corresponding to the low-scoring model to become separable in the latent space.

[0351] For each task-model pair In the offline phase, a supervised scoring label vector can be obtained:

[0352] ,

[0353] In this embodiment, it is acceptable to... , respectively corresponding to: This indicates the overall prediction error related score; This represents a shape preservation-related score. The present invention learns a scoring function that, for any new task... It can output the predicted score vectors for each candidate model:

[0354] ,

[0355] And further, the ranking score is given:

[0356] ,

[0357] Based on this, a set of Top-K recommendation models is output.

[0358] The method of this invention is applicable to the following application scenarios:

[0359] For a given pantograph, given only the observed wear sequence of its initial segment, the goal is to automatically select a prediction model from a candidate model library that best matches the current task, and output the ranking and recommendation scores of the candidate models to provide a basis for model selection for subsequent wear prediction of unknown segments. The innovation of this method lies not in redesigning specific prediction models, but in constructing a task representation with degradation mechanism implications from a portion of the observed wear sequence, and further establishing a latent variable matching and scoring mechanism between the task and candidate models to achieve automatic model selection and Top-K recommendation for pantograph wear scenarios.

[0360] For example, suppose the length of a complete wear curve of a certain pantograph is... The observation ratio is taken ,but:

[0361]

[0362] Therefore, the length of the observed sequence for this task is 120, and the length of the unknown future segment is 80. Assume the candidate model library contains 10 models, denoted as:

[0363]

[0364] In the offline phase, for this task and similar historical tasks, the overall error and shape error of 10 models on the unknown segment are calculated to form a supervision scoring matrix. In the online phase, a degenerate state field is constructed for the first 120 observations of this target task. It is assumed that in the scale set... At each time step, four sets of state variables and one relative position variable are extracted, thus forming a fixed-dimensional state field token for each time step. The task representation is then obtained through a dynamic task encoder. Simultaneously read the model potential capability representations of 10 candidate models. Calculate the latent variables of the relationship one by one. Matching score .

[0365] If the final ranking scores of the 10 models are as follows:

[0366] ,

[0367] Sort in descending order, for example:

[0368] ,

[0369] When setting At that time, the Top-3 recommendation model output by the system is:

[0370] ,

[0371] This indicates that for the target task where only the first 60% of the wear curve is observed, the system believes... Ideally suited for subsequent wear prediction. and These are the second-best and third-best candidate models, respectively.

[0372] Compared with existing pantograph sliding plate wear prediction technologies that typically rely on a single fixed model and existing general time series automatic model selection technologies that lack explicit modeling of mechanical degradation mechanisms, this invention creates a technical route of "degradation state field task construction - dynamic task encoding - model potential capability representation - relational latent variable matching and scoring - Top-K recommendation", which can improve the ability to represent the true state of pantograph wear tasks, thereby enhancing the relevance of model recommendations.

[0373] Existing methods often treat partial observation sequences directly as general time windows or ordinary statistical samples, making it difficult to distinguish the main wear trend, local fluctuations, anomalous disturbances, and stage changes. This invention first performs trend-abrupt-noise decomposition on the observed wear sequences, and further constructs state field features such as multi-scale slope, curvature, local fluctuation intensity, anomalous significance, and relative time position. This technical solution elevates the original wear numerical sequence into a state field object with degradation mechanism implications, more fully reflecting the slow degradation, local anomalies, and stage transition characteristics in the pantograph wear process. This allows the subsequent recommendation process to no longer rely on coarse static window representations that only focus on appearances while ignoring the underlying mechanisms, thus establishing a task description that better conforms to the laws of mechanical degradation. This reduces model misselection problems caused by insufficient task representation and improves the consistency between model recommendations and the current wear state.

[0374] This invention also provides a pantograph wear prediction system based on the method described herein, oriented towards wear time series and utilizing a task model for recommendation, comprising:

[0375] The data acquisition module is used to acquire a portion of the observed wear sequence of the target pantograph;

[0376] The state field construction module is used to construct a degraded state field from the partially observed wear sequence to obtain a state field sequence.

[0377] The task encoding module includes a dynamic task encoder for encoding the state field sequence into a task representation;

[0378] The model representation module is used to store and provide a representation of the potential capabilities of each candidate model in the candidate model library;

[0379] The relation encoding module includes a relation latent variable encoder, used to construct task-model relation latent variables based on the task representation and the model potential capability representation of any candidate model;

[0380] The scoring and recommendation module includes a matching scoring header, which generates a scoring vector and ranking score for the candidate model based on the task representation, the model's potential ability representation, and the latent variables of the task-model relationship, and outputs a recommended model based on the ranking scores of each candidate model.

[0381] In a preferred embodiment of the present invention, the dynamic task encoder includes an input projection layer, a multi-scale causal dilated convolution branch, a scale attention fusion layer, a residual normalization layer, a three-view aggregation layer, and a scalar gated fusion layer connected in sequence.

[0382] The input projection layer is used to map the state field tokens to the hidden dimensions (i.e., a linear mapping). The multi-scale causal dilated convolutional branch is used to extract degradation features at different time scales; the scale attention fusion layer is used to weight and fuse the outputs of branches with different dilation rates; the three-view aggregation layer is used to generate global aggregation representation, recent stage aggregation representation, and mutation-weighted aggregation representation respectively; the scalar gated fusion layer is used to fuse the sequence degradation representation with the task scalar representation into a task representation; the relational latent variable encoder includes a task projection layer, a model projection layer, a common activation computation unit, a mismatch computation unit, a splicing fusion unit, and a non-negative activation output unit.

[0383] Task projection layer (i.e.: , The input of the task mapping matrix (used to project the task representation onto a space comparable to the model) is connected to the task encoding module, and the model projection layer (i.e.: , The model mapping matrix (used to project the model representation onto a space that can interact multiplicatively with the task) is connected to the input of the model representation module, jointly activating the computation unit ( (representing the part jointly activated by task requirements and model capabilities) and mismatched computational units ( (Indicating the degree of mismatch between the task and the model) receives the task projection results and the model projection results respectively, and then stitches and fuses them ( (Representation vector concatenation) receives common activation terms and mismatched terms, and the non-negative activation output unit outputs the latent variables of the task-model relationship;

[0384] The matching scoring head includes a task difficulty baseline branch, a model overall bias branch, and a relationship-driven scoring branch set in parallel. The outputs of these three branches are connected to a scoring vector fusion unit, which in turn is connected to a ranking score calculation unit (based on...). Sort from largest to smallest (to obtain the model recommendation result), and use it to generate candidate model ranking scores based on the weights of the scoring dimensions.

[0385] This invention can enhance the ability to identify task differences under different observation stages, different fluctuation modes, and different local perturbation conditions, thereby improving the stability of the recommendation results.

[0386] This invention employs a multi-scale causal attention TCN to encode the degenerate state field and constructs task representations from three perspectives: global aggregation, recent stage aggregation, and mutation-weighted aggregation. Since causal convolution only utilizes current position and historical information, it satisfies the constraint that future data cannot be used in online recommendation scenarios. Simultaneously, the multi-scale structure can capture both short-term local perturbations and long-term degradation trends, while mutation-weighted aggregation enhances sensitivity to outliers and stage mutations.

[0387] Therefore, compared with techniques that rely solely on a single time scale or a single statistical feature, this invention can more stably distinguish between different types of pantograph wear tasks, such as "early stable tasks," "mid-term fluctuating tasks," and "late-term accelerated degradation tasks," thus maintaining good recommendation robustness even when task status changes.

[0388] This invention addresses the significant differences in model applicability across different task stages, fluctuation characteristics, and degradation modes in pantograph wear time series. It establishes a model recommendation method capable of extracting degradation state information from partial observed wear sequences, characterizing task-model matching relationships, and outputting candidate model ranking results, thereby improving the accuracy and stability of model selection in pantograph wear prediction scenarios.

[0389] This invention does not apply the same prediction model to all tasks. Instead, it selects a more suitable model from a candidate model library based on the current degradation state, stage information, and local fluctuation characteristics of the target task. This reduces the amplification of prediction errors in future segments caused by "model-task mismatch." The model recommendation results are more consistent with the current task status, which helps to make subsequent wear trend judgments, lifespan estimations, and maintenance decisions more reasonable. Therefore, this invention can reduce prediction bias, maintenance timing bias, and the resulting risks of premature or delayed replacement caused by incorrect model selection in engineering.

[0390] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.

[0391] In the embodiments of this application, terms such as "fixed," "fixed connection," and "fixed connection" refer to common fixing methods in the prior art, such as welding, riveting, and screws. "Rotary connection" refers to common rotary connection methods in the prior art, such as hinges and bearing rotation. If electrical components are provided, the functions, control, and power supply methods of all electrical components are common technical means in the prior art. This application has not improved them and they are not within the protection scope of this application. Therefore, this application will not elaborate on them.

[0392] Furthermore, the selection of materials and strength limitations for all components in this application can be made and arranged by those skilled in the art based on the site environment and the requirements of relevant national or industry standards, and are not within the scope of protection of this application. Therefore, this application will not elaborate on these points.

Claims

1. A pantograph wear prediction method based on wear time series and utilizing a task model recommendation, characterized in that, Includes the following steps: S1. Obtain complete pantograph wear curve data and construct an offline task set and candidate model library; S2, construct the degradation state field for the observed wear sequence of each task to obtain the state field sequence; S3, Input the state field sequence into the dynamic task encoder to obtain the task representation; S4. Based on the meta-information of candidate models and their applicability in historical stages, construct a representation of the potential capabilities of the models. S5. For each candidate model, the task representation and the model latent capability representation are mapped to the relational interaction space respectively. The common activation terms and mismatch terms of the two are calculated. The common activation terms, mismatch terms, task representation and model latent capability representation are fused to obtain the task-model relational latent variables. Input the task representation, the model's latent capability representation, and the latent variables of the task-model relationship into the matching scoring head, and output the candidate model's scoring vector and ranking score for the current task; S6. Sort the candidate models according to their ranking scores and select the model with the highest ranking score. S7, using the recommended model, the target pantograph is predicted to move at a certain step length. Predicting unknown wear sequences within the range.

2. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 1, characterized in that, Step S1 involves obtaining complete pantograph wear curve data and constructing an offline task set and candidate model library as follows: Let the complete wear curve y of a certain pantograph be represented as: ， in, Indicates the first The wear observation value at a given moment, where T represents the total lifespan of the wear curve; For any cutoff ratio The cutoff position is defined as: ， in, For a preset set of truncation ratios, Indicates the end position of the observable sequence in the current task; [] indicates rounding down. Construct a task to be recommended: ， in, To represent a task, This indicates the curve number, vehicle number, or sequence identifier; the input for this task is the observed sequence. for: ， in, This indicates the wear observation value at the end position of the observable sequence in the current task; The corresponding truth value of the unknown future segment for: ， Let the candidate model library be: ， in, Represents the set of candidate prediction models. Indicates the number of candidate models. Indicates the first There are 10 candidate prediction models.

3. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 1, characterized in that, Step S2 involves constructing the degradation state field for each task's observed wear sequence to obtain the state field sequence. For the observed wear sequence Perform trend, mutation, and noise decomposition: ， in, Indicates the first Wear observation values ​​at a given moment, Indicates the first Slowly degenerating skeleton at a given moment Indicates the first Sparse mutations at time 1; Indicates the first High-frequency noise term at each moment; To solve and Establish optimization goals: ， in, Indicates the end position of the observable sequence in the current task. For robust loss function; This represents the second-order difference operator. , Used to encourage the sparse occurrence of mutations For trend smoothing weights, For mutation sparse weights; In obtaining Then, in multi-scale assemblies The above constructs state field characteristics, for any scale Definition: Multiscale slope : ， Indicates the first At any moment on the scale The rate of change of local trends on; Multiscale curvature : ， Indicates the acceleration of a trend change; Local fluctuation intensity : ， in, Representing scale The mean of the noise term within the lower window; This indicates the intensity of the fluctuation during that period; This represents the high-frequency noise term at time e; Significance of mutations : ， in, To prevent extremely small positive numbers with a denominator of zero, The larger the value, the more likely a significant mutation is to occur at that moment; Introducing relative time position : ， time The state field token is represented as : ， The entire observed sequence is thus represented as a state field sequence. : in, This represents the state field sequence indicating the termination position of the observable sequence in the current task.

4. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 1, characterized in that, Step S3 involves inputting the state field sequence into the dynamic task encoder to obtain the task representation, as follows: Let the dimension of the state field token be . To hide dimensions in subsequent coding networks Alignment, first perform linear mapping: ， in, For the input projection matrix, For bias vectors, For the first The initial embedding representation at each time step; For a moment State field token; Let the set of multi-scale void ratios D be: ， For the Individual residual blocks, scale Define the causal convolution branch output. for: ， in, Indicates from the first moment to the second moment. The hidden state of history at a certain moment The scale is represented as The branches consist of causal convolution, normalization, activation, and dropout; Fusion is performed using scale attention weights, denoted as _i_. The weights for different scales at each time point are: ， in, The scale attention logits generated from the current hidden state, Indicates the first At any given moment, the scale The level of attention; Fusion output for: ， After residual connection, we get: ， in, Representation layer normalization, This represents the residual projection matrix when the input and output dimensions are the same. Can take identity mapping; Indicates the first Task at a specific moment Hidden state after residual connection; go through After several residual blocks, the hidden state sequence is obtained. : ， in, Indicates task The last moment of the observation sequence, Indicates task The hidden state at the last moment of the observation sequence; Three aggregation methods are used: Global aggregation: ， in, Indicates task The global aggregation representation; Indicates task The hidden state at time i; Nearest-stage aggregation: Let the length of the nearest window be... ,but ， in, Indicates task The most recent stage aggregation representation; Mutation-weighted aggregation: Let the significance of the mutation be 1. ,but ， ， in, This is the amplification factor for the sudden change. Indicates the first The mutation weight at each time step; Indicates the significance of the mutation. Indicates task Mutation-weighted aggregation representation; Construct task scalar feature vector Map it to a space of the same dimension as the sequence representation: ， in, This represents the scalar representation of the task. Let be the task scalar projection matrix. The task scalar bias vector; Concatenate the three sequence representations: ， in, Indicates task Sequence degradation characterization; Constructing gate vectors : ， in, This represents the Sigmoid function. The components are located at interval; This is the gated mapping matrix, also known as the gated weight matrix. This is the gate bias vector, used to adjust the reference offset of the gated output; The final task is represented as: ， in, Represents element-wise product. Indicates task The final potential representation of .

5. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 1, characterized in that, Step S4 involves constructing a representation of the model's potential capabilities based on the meta-information of the candidate models and their applicability across historical stages. For candidate models Construct its original model feature vector : ， in, Represents the static attribute features of the model. A signature indicating the model's historical applicability at different observation stages; Suppose the source domain tasks are divided according to the observation ratio as follows: The first stage subset, the Model at each stage Statistical applicability of experience for: ， in, Indicates the first The set of tasks for each stage. This indicates the number of tasks in the set. Indicates task upper model Offline overall rating; but: ， Mapping the original model features to the model capability space: ， in, Candidate models The model capability space, Here, represents the model capability projection matrix, and represents the learnable parameters. This represents both the model's capability bias and the learnable parameters. LN(⋅) is the mapping function, representing the layer normalization operation; Introducing learnable residual vectors The final model is represented as: ， in, Representing candidate models The potential capabilities are represented.

6. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 1, characterized in that, For each candidate model, the task representation and the model's latent capability representation are mapped to the relational interaction space, respectively. The common activations and mismatches of the two are calculated, and the common activations, mismatches, task representation, and model latent capability representation are fused to obtain the task-model relational latent variables: ， in, This is the model mapping matrix, used to project the model representation onto a space that can interact multiplicatively with the task; This is the task mapping matrix, used to project the task representation onto a space that can be compared with the model; This represents the portion activated by both task requirements and model capabilities; This indicates the degree of mismatch between the task and the model; This indicates vector concatenation; It is a smooth non-negative activation function; Represents the latent variables of the task-model relationship.

7. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 6, characterized in that, The task representation, the candidate model's latent capability representation, and the task-model relationship latent variables are input into the matching scoring head to obtain the candidate model's score vector and ranking score for the current task. The specific method is as follows: Define the baseline for task difficulty: ， in, This is a task difficulty mapping matrix. For bias terms, This indicates the baseline difficulty of the current task across all scoring dimensions; Define the overall bias of the model : ， in, For the model bias table, Representation Model The bias vector; Relationship-driven scoring items Defined as: ， in, This is a relational rating mapping matrix; Final score vector for: ， Weighted fusion is used: ， in, Indicates the sorted score. This is the weight vector; according to Sort the data from largest to smallest to get the model's recommendation results.

8. The pantograph wear prediction method based on wear time series and recommended by a task model according to claim 1, characterized in that, It also includes the joint loss function L: ， in, Indicates the loss due to the return of supervision. Represents the sorting loss. Indicates the relationship between loss and comparison. These are the corresponding weighting coefficients; Supervision regression loss: ， This represents the mean squared error between the predicted score vector and the offline supervision label; where For the rating vector, Let M represent a task, M be a candidate model library, and m be a candidate prediction model; Ranking loss: For the task The model with the higher overall score in the true mean is denoted as The lower model is denoted as Then the definition is: ， in, This indicates the model in the actual rating. Should be ranked Front; This represents the ranking score predicted by the model with the higher actual overall score in the recommendation network; This represents the ranking score predicted by the model with the lower actual overall score in the recommendation network. Relational contrast loss: set up Indicates task A collection of high-resolution models Let represent the set of low-scoring models, then we have: ， in, This represents the direction vector of relational quality. Temperature coefficient; Represents the latent variables of the task-model relationship.

9. A pantograph wear prediction system based on the method of any one of claims 1-8, oriented towards wear time series and utilizing a task model for recommendation, characterized in that, include: The data acquisition module is used to acquire a portion of the observed wear sequence of the target pantograph; The state field construction module is used to construct a degraded state field from the partially observed wear sequence to obtain a state field sequence. The task encoding module includes a dynamic task encoder for encoding the state field sequence into a task representation; The model representation module is used to store and provide a representation of the potential capabilities of each candidate model in the candidate model library; The relation encoding module includes a relation latent variable encoder, used to construct task-model relation latent variables based on the task representation and the model potential capability representation of any candidate model; The scoring and recommendation module includes a matching scoring header, which generates a scoring vector and ranking score for the candidate model based on the task representation, the model's potential ability representation, and the latent variables of the task-model relationship, and outputs a recommended model based on the ranking scores of each candidate model.

10. The pantograph wear prediction system based on wear time series and recommended by a task model according to claim 9, characterized in that, The dynamic task encoder includes an input projection layer, a multi-scale causal dilated convolutional branch, a scale attention fusion layer, a residual normalization layer, a three-view aggregation layer, and a scalar gated fusion layer connected in sequence. The input projection layer maps state field tokens to hidden dimensions; the multi-scale causal dilated convolution branch extracts degradation features at different time scales; the scale attention fusion layer weights and fuses the outputs of branches with different dilation rates; the three-view aggregation layer generates global aggregation representation, recent stage aggregation representation, and mutation-weighted aggregation representation, respectively; and the scalar gated fusion layer fuses sequence degradation representation and task scalar representation into a task representation. The relational latent variable encoder includes a task projection layer, a model projection layer, a co-activation computation unit, a mismatch computation unit, a splicing fusion unit, and a non-negative activation output unit. The input of the task projection layer is connected to the task encoding module, the input of the model projection layer is connected to the model representation module, the common activation calculation unit and the mismatch calculation unit receive the task projection result and the model projection result respectively, the stitching and fusion unit receives the common activation term and the mismatch term, and the non-negative activation output unit outputs the latent variable of the task-model relationship. The matching scoring head includes a task difficulty baseline branch, a model overall bias branch, and a relationship-driven scoring branch set in parallel. The outputs of the three branches are connected to a scoring vector fusion unit, which is connected to a ranking score calculation unit to generate candidate model ranking scores based on the scoring dimension weights.

Images