A pantograph carbon strip abrasion prediction method, system and device

Abstract

This invention discloses a method, system, and device for predicting the wear of a pantograph carbon sliding plate, relating to the fields of fault diagnosis and intelligent operation and maintenance technology. The method includes: constructing a training dataset based on the mileage and wear data of the pantograph carbon sliding plate; constructing a meta-learning model using a model-independent meta-learning architecture; optimizing model parameters through inner and outer loop training methods to enable it to quickly adapt to new tasks under limited sample conditions; fine-tuning the model using prediction data; and outputting the predicted wear value of the carbon sliding plate. This method, by combining data-driven and meta-learning techniques, significantly improves the accuracy and generalization ability of carbon sliding plate wear prediction, effectively solving the problems of low prediction accuracy and poor adaptability under limited sample conditions, and improving prediction efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and intelligent operation and maintenance technology, specifically to a method, system and device for predicting the wear of a pantograph carbon sliding plate. Background Technology

[0002] In modern railway transportation systems, the health of the pantograph's carbon contactor is crucial for ensuring safe train operation. As a key component of the pantograph, the carbon contactor directly contacts the overhead contact line and is responsible for transmitting electrical energy. However, due to the nonlinear characteristics of carbon contactor wear and its susceptibility to various external factors (such as train speed, temperature, humidity, current, and voltage), traditional carbon contactor wear monitoring methods often struggle to accurately predict the wear status of the contactor. This results in delays and inaccurate fault warnings and wear analysis during equipment maintenance, thereby increasing maintenance costs and potential safety risks for the railway system.

[0003] In recent years, data-driven and meta-learning-based carbon skateboard wear prediction methods have received widespread attention. These methods can automatically extract features from a large amount of historical data and make accurate wear predictions. Compared with traditional methods, they have higher accuracy and stronger adaptability. In particular, in the context of big data, deep learning methods can identify key factors affecting carbon skateboard wear by learning from a large amount of data and provide accurate prediction results.

[0004] However, most current methods focus on using deep learning to process large-scale datasets. In practical applications, due to problems such as incomplete data collection and scarce labeled samples, wear prediction models often have difficulty being effectively trained under conditions with few samples, resulting in poor generalization ability of wear prediction models and thus low accuracy of wear prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies in predicting carbon pantograph wear based on limited sample data, such as low accuracy and poor model adaptability, this invention proposes a method, system, and device for predicting pantograph carbon pantograph wear, thereby solving the problems existing in the prior art.

[0006] A method for predicting the wear of a pantograph carbon sliding plate includes the following steps: Mileage data and corresponding wear label data of multiple pantograph carbon skateboard samples were collected to construct a training dataset. The training dataset was divided into multiple task data subsets according to the carbon skateboard wear prediction task, and each task data subset was divided into a non-overlapping support set and query set. Each task data subset corresponds to a set of pantograph carbon skateboard samples and their corresponding wear label data. The training dataset is used to train a meta-learning prediction model based on a model-independent meta-learning architecture. The specific steps include: In the inner loop training process, mileage data from the support sets of each task are input into a base prediction model with shared initialization parameters to obtain wear prediction values; based on the loss between the wear prediction values ​​and the wear label data in the support sets, the shared initialization parameters are updated using gradients to obtain the model parameters for each task; In the outer loop training process, mileage data from the query sets of each task are input into the model trained in the inner loop to obtain query set prediction values; The meta-loss between the predicted query values ​​and wear labels for each task is calculated, and based on the meta-loss, gradient descent is used to optimize the shared initialization parameters of the base prediction model to complete the training of the meta-learning model. The mileage data of the carbon sliding plate of the pantograph to be tested is input into the trained meta-learning prediction model to output the corresponding wear prediction value of the carbon sliding plate.

[0007] Furthermore, after collecting mileage data and corresponding wear tag data from multiple sets of pantograph carbon slider samples, the mileage data and wear tag data are preprocessed. The preprocessing process includes the following steps: Linear interpolation was used to fill in missing values ​​in mileage and wear label data; Outlier removal was performed on the mileage and wear label data after missing value imputation using the standard score method. The mileage data and wear label data after outlier removal are normalized, and a training dataset is constructed based on the processed mileage data and wear label data.

[0008] Furthermore, the gradient update formula for the inner loop is expressed as: ; in, These are the initial model parameters. The learning rate for the inner loop. It is the loss function of the task. These are the updated parameters of the model corresponding to the task. Represents the loss function pair The gradient.

[0009] Furthermore, the gradient update formula for the outer loop is expressed as: ; in, It is the learning rate of the outer loop. It is the loss function of the task. These are the updated initial model parameters. This represents the initial model parameters of the outer loop. This represents the model parameters after the outer loop update.

[0010] Furthermore, the step of inputting the mileage data of the carbon sliding plate of the pantograph under test into the trained meta-learning prediction model to output the corresponding wear prediction value of the carbon sliding plate specifically includes the following steps: The historical mileage data and corresponding wear data of the target pantograph carbon slide plate are obtained to form the target support set; The target support set is input into the trained meta-learning prediction model, and the shared initialization parameters are quickly updated to obtain a task adaptation model that adapts to the target pantograph carbon skateboard. The mileage data of the target pantograph carbon slider is input into the task adaptation model, and the corresponding wear prediction value is output.

[0011] Furthermore, the basic prediction model in the meta-learning prediction model is a backpropagation neural network; the backpropagation neural network includes: The input layer is used to receive mileage data from the training dataset; One or more hidden layers are used to transform and extract features from the input mileage data through a non-linear activation function to generate intermediate feature representations; each hidden layer consists of fully connected neurons. The output layer is used to output continuous carbon slide plate wear predictions based on the intermediate feature representations.

[0012] Furthermore, it also includes using four metrics—accuracy, precision, recall, and F1 score—to evaluate the performance of the trained meta-learning prediction model.

[0013] The present invention also includes a pantograph carbon slider wear prediction system, comprising: The data acquisition module is used to collect mileage data and corresponding wear label data of multiple pantograph carbon skateboard samples to construct a training dataset. The training dataset is divided into multiple task data subsets according to the carbon skateboard wear prediction task, and each task data subset is divided into a non-overlapping support set and query set. Each task data subset corresponds to a set of pantograph carbon skateboard samples and their corresponding wear label data. The model training module is used to train a meta-learning prediction model based on a model-independent meta-learning architecture using a training dataset. Specifically, it includes the following steps: During the inner loop training, mileage data from the support sets of each task are input into a base prediction model with shared initialization parameters to obtain wear prediction values; based on the loss between the wear prediction values ​​and the wear label data in the support sets, the shared initialization parameters are updated using gradients to obtain the model parameters for each task; During the outer loop training, mileage data from the query sets of each task are input into the model trained in the inner loop to obtain query set prediction values; the meta-loss between the predicted query values ​​and wear labels for each task is calculated, and based on the meta-loss, gradient descent is used to optimize the shared initialization parameters of the base prediction model to complete the training of the meta-learning model. The prediction module is used to input the mileage data of the carbon sliding plate of the pantograph under test into the trained meta-learning prediction model, so as to output the wear prediction value of the corresponding carbon sliding plate.

[0014] The present invention also includes a computer device for predicting the wear of a pantograph carbon sliding plate, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the pantograph carbon sliding plate wear prediction method.

[0015] The present invention also includes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the pantograph carbon slide plate wear prediction method.

[0016] This invention provides a method for predicting the wear of a pantograph carbon sliding plate, which has the following beneficial effects: This invention employs a model-independent meta-learning framework. Its training objective is not to directly fit a single dataset, but rather to perform meta-training on a large number of heterogeneous tasks (corresponding to different carbon skateboards and different operating conditions). An inner loop independently updates gradients on each task, minimizing the loss function to obtain task-specific model parameters. Simultaneously, an outer loop optimizes the initial parameters of the meta-learning model based on the loss values ​​of all tasks, enabling the model to quickly adapt to new tasks and improving its generalization ability and adaptability. This method effectively solves the problems of low prediction accuracy and poor adaptability under limited sample conditions. Through the meta-learning architecture, the model can learn common knowledge from multiple related tasks, thereby quickly achieving high performance on new tasks. This significantly improves prediction efficiency and accuracy, providing reliable assurance for the safe operation of railway systems, reducing maintenance costs, and demonstrating broad application prospects and practical value. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating the carbon skateboard wear prediction method based on data-driven and meta-learning in an embodiment of the present invention. Figure 2 This is a regression curve of the meta-learning model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the pantograph carbon slide plate wear prediction method in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] This invention proposes a method for predicting the wear of carbon sliding plates on pantographs. The method constructs a training dataset based on the mileage and wear data of the carbon sliding plates and preprocesses the data. A model-independent meta-learning (MAML) architecture is used to build a meta-learning model, and the model parameters are optimized through inner and outer loop training to enable it to quickly adapt to new tasks under limited sample conditions. The model is then fine-tuned using the predicted data to output the predicted wear value of the carbon sliding plate. By combining data-driven and meta-learning techniques, this invention significantly improves the accuracy and generalization ability of carbon sliding plate wear prediction, especially demonstrating superior adaptability under data-scarce conditions. This system can be widely applied in the railway transportation field, providing accurate predictive support for the maintenance of carbon sliding plates, reducing maintenance costs and improving safety.

[0020] like Figure 3 As shown, the method specifically includes the following steps: S1. Construct a training dataset based on the mileage data (input matrix) and wear data (target vector) of the pantograph carbon skateboard.

[0021] We collected and organized mileage and wear data of the pantograph carbon skateboard during actual operation to construct a comprehensive training dataset (support set) and test dataset (query set). Mileage data reflects the frequency and intensity of use of the carbon skateboard, while wear data directly reflects the degree of wear. By combining these two types of data, we can capture the intrinsic relationship between carbon skateboard wear and use, providing a rich information foundation for subsequent model training.

[0022] Building training and testing datasets is fundamental to the entire prediction process, and their quality and diversity directly impact the model's prediction accuracy and generalization ability. A comprehensive and diverse dataset helps the model learn multiple patterns and influencing factors of carbon skateboard wear, thereby improving the model's accuracy and reliability in practical applications.

[0023] The mileage and wear data were preprocessed, including missing value imputation, outlier removal, and data normalization. A training dataset was constructed based on the preprocessed mileage and wear data of the pantograph carbon slide plate. Missing values ​​were imputed using linear interpolation, and outliers were removed using the standard score method.

[0024] S2. Based on multiple carbon skateboard wear prediction tasks and combined with an irrelevant meta-learning architecture, a meta-learning model is constructed. The meta-learning model is trained using a training dataset. The training method is as follows: First, an inner loop is used to train the meta-learning prediction model. Based on the shared initial model parameters (referring to the same learnable initialization vector, which serves as the starting point of the inner loop for all tasks), the gradient is updated for each carbon skateboard wear prediction task using the training dataset. The loss function of the carbon skateboard wear prediction task is minimized to obtain the model parameters for each carbon skateboard wear prediction task.

[0025] The gradient update formula for the inner loop is as follows: ; in, These are the initial model parameters. The learning rate for the inner loop. It is the loss function of the task. These are the updated parameters of the model corresponding to the task. It is a loss function pair The gradient.

[0026] Then, the gradient of the initial model parameters of all carbon skateboard wear prediction tasks and the meta-learning prediction model is determined by the outer loop. The initial model parameters of the meta-learning prediction model are optimized by the gradient descent method to obtain the optimized initial model parameters of the meta-learning prediction model.

[0027] In-task losses (support set): ; Inner loop update: ; Unexpected losses (query set): ; Summary of all mission losses: ; in, i Task Index i =1,…, =1,…,N; N: Total number of tasks (e.g., multiple wear prediction tasks formed by different working conditions / lines / humidity batches, etc.); :Task i The support set (the set of samples used for inner loop adaptation / update); :Task iThe query set (a sample set used for outer loop optimization / evaluation); : by parameters θ The predictive model outputs the input; θ The initial model parameters for meta-learning (a starting point shared by all tasks; the object to be optimized in the outer loop); Task adaptive parameters, determined by the inner loop within the task. i Support set from θ The update is obtained (for evaluation on the query set); α Inner loop learning rate (step size); :Task i Loss on the support set; :Task i Loss on the query set; Meta-goal / Total loss (average loss across all task query sets).

[0028] Based on the loss values ​​of all carbon skateboard wear prediction tasks, the gradient of the loss with respect to the initial model parameters of the meta-learning prediction model is calculated. The initial model parameters of the meta-learning prediction model are then updated and optimized using a multi-round gradient descent method to obtain the optimized initial model parameters of the meta-learning prediction model.

[0029] The gradient update formula for the outer loop is as follows: ; in, It is the learning rate of the outer loop. It is the loss function of the task. These are the updated initial model parameters. This represents the initial model parameters of the outer loop. This represents the model parameters after the outer loop update.

[0030] This invention employs a Model-Agnostic Meta-Learning (MAML) architecture, training the meta-learning model by constructing multiple carbon skateboard wear prediction tasks. The training process consists of two parts: an inner loop and an outer loop. The inner loop performs gradient updates for each specific task, minimizing the loss function for that task to obtain the model parameters for each task. The outer loop optimizes the initial parameters of the meta-learning model based on the loss values ​​of all tasks, enabling it to quickly adapt to new tasks. This training method allows the model to quickly learn and adapt to new wear prediction tasks under limited sample conditions.

[0031] The basic prediction model in the meta-learning prediction model is a backpropagation neural network. The backpropagation neural network optimizes the initial parameters through inner and outer loops on multiple tasks. The backpropagation neural network includes: an input layer that receives training data and inputs it into the first hidden layer. The first hidden layer performs a linear transformation on the input data through a weight matrix and a bias term; the hidden layer consists of multiple fully connected neurons, and the output of each layer undergoes a non-linear transformation through an activation function to achieve feature extraction and mapping, generating an intermediate feature representation; based on the feature representation extracted by the hidden layer, the output layer completes the regression prediction of the carbon skateboard wear value.

[0032] The training method for backpropagation neural networks is as follows: the backpropagation algorithm is used to calculate the loss function, and the weights and biases of the backpropagation neural network are updated by gradient descent, so that the prediction results of the backpropagation neural network get closer and closer to the true value, thus completing the training of the backpropagation neural network.

[0033] By training with a meta-learning model, the model can learn the commonalities and differences between different carbon skateboard wear prediction tasks, thereby extracting generalized knowledge. This generalized knowledge enables the model to quickly adjust and achieve high prediction performance when faced with new and unseen wear prediction tasks, greatly improving the model's adaptability and generalization ability.

[0034] S3. Optimize the initial model parameters of the meta-learning prediction model based on the prediction data. The optimized meta-learning prediction model outputs the wear prediction value of the carbon skateboard.

[0035] After obtaining the initial model parameters optimized by the meta-learning prediction model, these parameters are further optimized using the prediction data (i.e., the carbon skateboard mileage data to be predicted). Through fine-tuning, the model uses a small learning rate to make small adjustments to the initial parameters to optimize the model's performance on the specific dataset. The fine-tuned model can better adapt to the characteristics of the prediction data, thus outputting more accurate carbon skateboard wear predictions.

[0036] The optimization method is as follows: ; in, These are the optimized model parameters. For the loss function in the optimization phase, To optimize the learning rate.

[0037] Further improve the prediction accuracy of the meta-learning prediction model on specific datasets. Through fine-tuning, the model can fully utilize the feature information of the prediction data and finely adjust the initial parameters, thereby providing more accurate and reliable prediction results when facing actual wear prediction tasks. This is of great significance for ensuring the safe operation of trains and reducing maintenance costs.

[0038] In some embodiments, a training dataset is constructed based on the mileage and wear data of the pantograph carbon sliding plate. This includes preprocessing the collected mileage and wear data, which includes missing value imputation, outlier removal, and data normalization. The training dataset is then constructed based on the preprocessed mileage and wear data. Missing values ​​are imputed using linear interpolation, and outliers are removed using a standard score method to ensure data integrity and consistency. To avoid scale differences between different features, all numerical features are normalized, uniformly adjusting the data range to [0,1]. After preprocessing, the training dataset is constructed.

[0039] In some embodiments, the carbon skateboard wear prediction task is performed using a backpropagation neural network (BP), which is employed to execute different carbon skateboard wear prediction tasks. The BP neural network serves as the foundation of the meta-learning model, employing multiple layers of neurons for feature learning and prediction. In this model, the BP neural network is responsible for feature extraction and initial prediction. To improve the model's adaptability, especially under conditions of limited sample data, a Model-Independent Meta-Learning (MAML) architecture is introduced. A single BP neural network is used to optimize initial parameters across multiple tasks, forming BP neural networks tailored to different tasks. Optimization is achieved through internal and external loop training processes, enabling the model to better adapt to different tasks and improve its prediction ability under limited sample conditions. Based on the model training results, this invention further optimizes the model through fine-tuning. During fine-tuning, a smaller learning rate is used to locally adjust the model parameters, improving the model's prediction accuracy for new data. Through fine-tuning, the model can quickly adapt to specific tasks and datasets, thereby improving the accuracy and real-time performance of carbon skateboard wear prediction.

[0040] This invention also employs a model-independent meta-learning architecture and fine-tuning method, which effectively overcomes the problems of data scarcity and insufficient samples, optimizes training performance under conditions of few samples, and enhances the model's adaptability. This technological innovation enables the model to provide accurate wear prediction results even when faced with limited data, facilitating efficient wear monitoring in complex and incomplete data environments.

[0041] Compared with traditional detection methods, this invention can not only accurately predict the wear condition of carbon sliding plates on individual trains, but also has strong generalization ability, enabling its application to wear condition analysis and maintenance prediction for different trains, thus possessing broad application prospects. By deeply analyzing the key factors affecting carbon sliding plate wear, it can provide data support for subsequent maintenance decisions, reducing the maintenance cost of railway systems while improving the accuracy and timeliness of maintenance work.

[0042] This invention also employs a model-independent meta-learning architecture and fine-tuning method, which effectively overcomes the problems of data scarcity and insufficient samples, optimizes training performance under conditions of few samples, and enhances the model's adaptability. This technological innovation enables the model to provide accurate wear prediction results even when faced with limited data, facilitating efficient wear monitoring in complex and incomplete data environments.

[0043] Based on the above inventive concept, see [reference] Figure 1 As shown, this invention proposes a carbon skateboard wear prediction method based on data-driven and model-independent meta-learning, using carbon skateboard wear data collected from a subway station as an example. The method includes the following steps: S1. Construct a training dataset based on subway mileage and wear data, and preprocess the training dataset to obtain the preprocessed training dataset.

[0044] S1.1 Collect mileage and wear data of a subway line to construct a training dataset.

[0045] The training dataset mainly contains mileage data and corresponding wear data of carbon skateboards. The collected data covers different time periods as much as possible to ensure the diversity and comprehensiveness of the data, providing a foundation for subsequent model training.

[0046] S1.2, Fill in the missing values ​​in the training dataset.

[0047] For the collected mileage data, considering that the data may be missing, linear interpolation is used to fill in the missing values ​​that may exist during the data collection process.

[0048] For each feature, if missing values ​​exist, linear interpolation is used to estimate the missing values ​​based on the two nearest non-missing values. The basic formula for linear interpolation is: ; in, These are estimates of the missing values. It indicates the location of the missing value. and It represents the positions of two known data points before and after the missing value. and These are the known values ​​for these two positions.

[0049] S1.3 Remove outliers from the training data.

[0050] Outliers in the data are detected and removed using the standard score method. Outliers are values ​​that deviate significantly from other data points in the dataset, usually due to faults, measurement errors, or special circumstances.

[0051] In this embodiment, outliers are defined as data points whose wear values ​​differ significantly from other carbon skid plates within the same mileage range. For example, if multiple trains have similar mileage, but one train has a significantly higher carbon skid plate wear value than the others, then that wear value may be considered an outlier.

[0052] In this embodiment, the standard score method is used to identify such outliers. The z-score formula is:

[0053] ; in, It is a data point in the dataset (i.e., the wear value of the carbon skateboard of a train). It is the mean of all data for this feature (i.e., the wear value of the carbon skateboard). This is the standard deviation of the feature. According to the formula, the z-score describes the degree to which a data point deviates from the mean, expressed in standard deviation. When the absolute value of the standard score is... If the value exceeds the set threshold, the data point is considered an outlier and must be removed.

[0054] S1.4 Normalize the training data obtained in S1.3.

[0055] To ensure the effectiveness and accuracy of model training, all samples in the training dataset are normalized to a range between [0,1]. The normalization formula is:

[0056] ; in, It is the normalized value. It is the original value. and These are the minimum and maximum values ​​of all data for this feature, respectively.

[0057] S2. Based on multiple carbon skateboard wear prediction tasks and combined with a model-independent meta-learning architecture, a meta-learning model is constructed. The meta-learning model is trained using a training dataset. The training method is as follows: An inner loop is used to update the gradient of each carbon skateboard wear prediction task based on the shared initial model parameters and the support set of each task, so as to minimize the loss function of the carbon skateboard wear prediction task and obtain the model parameters of each carbon skateboard wear prediction task. An outer loop is used to calculate the query set loss on the query set of each task and to obtain the meta-gradient of the loss relative to the shared initial model parameters. Based on the gradient, the initial model parameters of the meta-learning prediction model are optimized by the gradient descent method to obtain the optimized initial model parameters of the meta-learning prediction model.

[0058] This invention introduces a model-independent meta-learning (MAML) architecture. MAML is a gradient-update-based meta-learning method that aims to enable the model to quickly adapt to new tasks and achieve high performance by training on multiple tasks.

[0059] S2.1. A backpropagation neural network is used to perform different carbon skateboard wear prediction tasks. The backpropagation neural network learns and predicts features through neurons at multiple levels. The structure of the backpropagation neural network includes an input layer, a hidden layer, and an output layer.

[0060] Input Layer: The input layer receives preprocessed and normalized mileage data and feeds it into the first hidden layer. The first hidden layer performs a linear transformation on the input data using a weight matrix and a bias term. Normalization is used to ensure that all input features are within the same scale range. Typically, the minimum value of the mileage data is mapped to 0, and the maximum value is mapped to 1 to eliminate dimensional differences between different features.

[0061] Hidden layers: Hidden layers consist of multiple fully connected neurons. Each neuron in a layer performs non-linear mapping and feature learning through an activation function (such as the ReLU function). The ReLU (Rectified Linear Unit) activation function is defined as follows:

[0062] ; Output Layer: The output layer generates the prediction result, namely the predicted wear value of the carbon skateboard. Since the wear value of the carbon skateboard is a continuous value, the output layer uses a linear activation function, the formula of which is:

[0063] ; in As weight, As input features, This is the bias term. Linear activation functions can generate continuous predictions and are suitable for regression tasks.

[0064] During the training of a backpropagation neural network, the backpropagation algorithm is used to calculate the loss function, and the model's weights and biases are updated using gradient descent. The loss function uses the mean squared error, and its formula is as follows: ; in, Let be the predicted value for the i-th sample. This is the actual value. Let be the number of samples. By minimizing the loss function, gradient descent gradually updates the weights and biases of the neural network, making the model's predictions increasingly closer to the true values.

[0065] S2.2 Divide the training dataset into multiple sub-training sets, and use inner and outer loops to train the meta-learning prediction model through multiple sub-training sets; 1) Internal circulation training The inner loop training can also be understood as feature task training. During the inner loop training process, the corresponding carbon skateboard wear prediction task is trained according to the support set. The support set performs gradient updates on the carbon skateboard wear prediction task, minimizes the loss function of the sub-carbon skateboard wear prediction task, and obtains the model parameters of each carbon skateboard wear prediction task.

[0066] Through its inner loop, the meta-learning prediction model can quickly adapt to specific tasks, adjusting its parameters to optimize task-specific performance. This process enables the model to learn and predict more efficiently when faced with new tasks.

[0067] The gradient update formula for the inner loop is as follows: ; in, These are the initial model parameters. The learning rate for the inner loop. It is a task-specific loss function. These are the parameters updated after the task-specific model.

[0068] 2) External circulation The outer training, also known as meta-training, aims to optimize the initial model parameters of the meta-learning prediction model trained in the inner loop. Based on the loss values ​​of all carbon skateboard wear prediction tasks, it calculates the gradient of the loss relative to the initial model parameters on the query set. Multiple rounds of gradient descent are used to update and optimize the initial model parameters, resulting in optimized initial model parameters that enable the meta-learning prediction model to learn generalized knowledge from multiple tasks. In the outer loop, the meta-learning prediction model learns how to adapt and converge more quickly in new tasks by optimizing the task-specific parameters obtained after the inner loop training. The gradient update formula for the outer loop is as follows:

[0069] ; in, It is the learning rate of the outer loop. It is a task-specific loss function. These are the parameters updated in the inner loop. Finally, through multiple rounds of gradient updates in the outer loop, an optimized initial model parameter is obtained, enabling the model to quickly adapt to different tasks and exhibit high performance.

[0070] S3. Input the data to be predicted into the trained meta-learning prediction model, optimize the initial model parameters of the meta-learning prediction model based on the data to be predicted, and output the wear prediction value of the carbon skateboard after optimization.

[0071] Fine-tuning further enhances the predictive ability of the meta-learning prediction model on specific datasets. During the fine-tuning phase, the meta-learning prediction model uses a small learning rate to make minor adjustments to the initial model parameters, optimizing the model's performance on the specific dataset. The goal of fine-tuning is to achieve higher prediction accuracy on new driving mileage data, especially under conditions of few samples, allowing the model to quickly adapt to new data. The fine-tuning formula is as follows:

[0072] ; in, These are the model parameters after fine-tuning. The loss function during the fine-tuning phase. This is the fine-tuned learning rate. Through this process, the model can better adapt to specific tasks, thereby improving the accuracy of carbon skateboard wear prediction.

[0073] K-fold cross-validation was used to evaluate the performance of the meta-learning prediction model in carbon skateboard wear prediction. Labeled data was divided into K mutually exclusive subsets according to the task. One subset was used as the validation set each time, and the remaining K-1 subsets were used for training and meta-optimization; this process was repeated K times to obtain K sets of validation results. For each subset, accuracy, precision, recall, and F1 score were calculated, and the mean ± standard deviation of each metric was reported as the overall performance. The trained meta-learning prediction model was evaluated using four metrics: accuracy, precision, recall, and F1 score. The difference between the predicted result and the actual label for each sample was calculated to comprehensively evaluate the model's performance in carbon skateboard wear prediction.

[0074] Accuracy: Accuracy refers to the proportion of samples correctly predicted by the model out of the total number of samples. The formula is: ;

[0075] in, TP These are true positives. TN These are true negatives. FP These are false positives. FN These are false negatives.

[0076] Precision: Precision refers to the proportion of samples that are predicted as positive by the model but are actually positive.

[0077] The accuracy formula is: .

[0078] Recall: Recall refers to the proportion of samples that are actually positive that can be correctly predicted as positive by the model.

[0079] The recall formula is: .

[0080] F1 Fraction( F1Score ): F1 The score is the harmonic mean of precision and recall, used to comprehensively measure the model's accuracy and recall capabilities.

[0081] F1 The formula for a fraction is: .

[0082] The trained model is evaluated using metrics such as accuracy and recall. F1 The accuracy and effectiveness of the model in predicting the wear condition of carbon skateboards are evaluated using indicators such as scores. Through comprehensive analysis of these indicators, the predictive performance of the model in practical applications is confirmed, ensuring the model's reliability and practicality; for example... Figure 2 The curve shown is the model regression curve, and the curve color represents the result after fine-tuning with different timings.

[0083] After training and evaluation, the trained deep learning model is applied to real-world scenarios to predict the wear status of carbon skateboards in real time. The wear prediction results are then combined with the model's output to analyze the wear trend of the carbon skateboards and provide maintenance suggestions. By calculating the contribution of each feature to the model's prediction accuracy, a basis is provided for further optimizing the use and maintenance of carbon skateboards.

[0084] This embodiment presents a data-driven and model-independent meta-learning-based method for predicting carbon skateboard wear. It utilizes mileage data as input features, combines a backpropagation neural network for feature extraction and prediction, and then optimizes the model's performance under limited sample conditions using a model-independent meta-learning architecture and fine-tuning methods. This method effectively overcomes the challenges posed by data scarcity and provides reliable support for the safe operation of trains.

[0085] This invention ensures the comprehensiveness and accuracy of training data by collecting and preprocessing mileage and wear data of carbon skateboards. It utilizes a backpropagation neural network to perform different wear prediction tasks, independently updating gradients on each task through an inner loop to minimize the loss function and obtain task-specific model parameters. Then, an outer loop optimizes the initial parameters of the meta-learning model based on the loss values ​​of all tasks, enabling the model to quickly adapt to new tasks and improving its generalization and adaptability. This method effectively solves the problems of low prediction accuracy and poor adaptability under limited sample conditions. The meta-learning architecture allows the model to learn common knowledge from multiple related tasks, thus quickly achieving high performance on new tasks and significantly improving prediction efficiency and accuracy. Simultaneously, fine-tuning methods are used to finely adjust model parameters, further improving the model's prediction accuracy on specific datasets. This provides a reliable guarantee for the safe operation of railway systems, reduces maintenance costs, and demonstrates broad application prospects and practical value.

[0086] Based on the same inventive concept, this invention also proposes a pantograph carbon sliding plate wear prediction system, comprising: The data acquisition module is used to collect mileage data and corresponding wear label data of multiple pantograph carbon skateboard samples to construct a training dataset. The training dataset is divided into multiple task data subsets according to the carbon skateboard wear prediction task. Each task data subset is further divided into a non-overlapping support set and a query set. Each task data subset corresponds to a set of pantograph carbon skateboard samples and their corresponding wear label data.

[0087] The model training module is used to train a meta-learning prediction model based on a model-independent meta-learning architecture using a training dataset. Specifically, it includes the following steps: During the inner loop training, mileage data from the support sets of each task are input into a base prediction model with shared initialization parameters to obtain wear prediction values; based on the loss between the wear prediction values ​​and the wear label data in the support sets, the shared initialization parameters are updated using gradients to obtain the model parameters for each task; During the outer loop training, mileage data from the query sets of each task are input into the model trained in the inner loop to obtain query set prediction values; the meta-loss between the query value predictions and wear labels for each task is calculated, and based on the meta-loss, gradient descent is used to optimize the shared initialization parameters of the base prediction model to complete the training of the meta-learning model.

[0088] The prediction module is used to input the mileage data of the carbon sliding plate of the pantograph under test into the trained meta-learning prediction model, so as to output the wear prediction value of the corresponding carbon sliding plate.

[0089] The present invention also proposes a computer device for predicting the wear of a pantograph carbon sliding plate, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the pantograph carbon sliding plate wear prediction method.

[0090] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the pantograph carbon slide plate wear prediction method.

[0091] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the wear of a pantograph carbon sliding plate, characterized in that, Includes the following steps: Mileage data and corresponding wear label data of multiple pantograph carbon skateboard samples were collected to construct a training dataset; The training dataset is divided into multiple task data subsets according to the carbon skateboard wear prediction task, and each task data subset is divided into a non-overlapping support set and query set; wherein, each task data subset corresponds to a set of pantograph carbon skateboard samples' mileage data and their corresponding wear label data. The training dataset is used to train a meta-learning prediction model based on a model-independent meta-learning architecture. The specific steps include: In the inner loop training process, mileage data from the support sets of each task are input into a base prediction model with shared initialization parameters to obtain wear prediction values; based on the loss between the wear prediction values ​​and the wear label data in the support sets, the shared initialization parameters are updated using gradients to obtain the model parameters for each task; In the outer loop training process, mileage data from the query sets of each task are input into the model trained in the inner loop to obtain query set prediction values; The meta-loss between the predicted query values ​​and wear labels for each task is calculated, and based on the meta-loss, gradient descent is used to optimize the shared initialization parameters of the base prediction model to complete the training of the meta-learning model. The mileage data of the carbon sliding plate of the pantograph to be tested is input into the trained meta-learning prediction model to output the corresponding wear prediction value of the carbon sliding plate.

2. The method for predicting the wear of a pantograph carbon sliding plate according to claim 1, characterized in that, This also includes preprocessing the mileage data and corresponding wear label data of multiple sets of pantograph carbon slider samples. The preprocessing process includes the following steps: Linear interpolation was used to fill in missing values ​​in mileage and wear label data; Outlier removal was performed on the mileage and wear label data after missing value imputation using the standard score method. The mileage data and wear label data after outlier removal are normalized, and a training dataset is constructed based on the processed mileage data and wear label data.

3. The method for predicting the wear of a pantograph carbon sliding plate according to claim 1, characterized in that, The gradient update formula for the inner loop is expressed as: ； in, These are the initial model parameters. The learning rate for the inner loop. It is the loss function of the task. These are the updated parameters of the model corresponding to the task. Represents the loss function pair The gradient.

4. The method for predicting the wear of a pantograph carbon sliding plate according to claim 3, characterized in that, The gradient update formula for the outer loop is expressed as: ； in, It is the learning rate of the outer loop. It is the loss function of the task. These are the updated initial model parameters. This represents the initial model parameters of the outer loop. This represents the model parameters after being updated via the outer loop.

5. The method for predicting the wear of a pantograph carbon sliding plate according to claim 1, characterized in that, The process of inputting the mileage data of the carbon sliding plate of the pantograph under test into the trained meta-learning prediction model to output the corresponding wear prediction value of the carbon sliding plate specifically includes the following steps: The historical mileage data and corresponding wear data of the target pantograph carbon slide plate are obtained to form the target support set; The target support set is input into the trained meta-learning prediction model, and the shared initialization parameters are quickly updated to obtain a task adaptation model that adapts to the target pantograph carbon skateboard. The mileage data of the target pantograph carbon slider is input into the task adaptation model, and the corresponding wear prediction value is output.

6. The method for predicting the wear of a pantograph carbon sliding plate according to claim 1, characterized in that, The basic prediction model in the meta-learning prediction model is a backpropagation neural network; the backpropagation neural network includes: The input layer is used to receive mileage data from the training dataset; One or more hidden layers are used to transform and extract features from the input mileage data through a non-linear activation function to generate intermediate feature representations; each hidden layer consists of fully connected neurons. The output layer is used to output continuous carbon slide plate wear predictions based on the intermediate feature representations.

7. The method for predicting the wear of a pantograph carbon sliding plate according to claim 1, characterized in that, It also includes using four metrics—accuracy, precision, recall, and F1 score—to evaluate the performance of the trained meta-learning prediction model.

8. A pantograph carbon sliding plate wear prediction system, characterized in that, include: The acquisition module is used to collect mileage data and corresponding wear label data of multiple pantograph carbon skateboard samples to construct a training dataset. The training dataset is divided into multiple task data subsets according to the carbon skateboard wear prediction task, and each task data subset is divided into a non-overlapping support set and query set; wherein, each task data subset corresponds to a set of pantograph carbon skateboard samples' mileage data and their corresponding wear label data. The model training module is used to train a meta-learning prediction model based on a model-independent meta-learning architecture using a training dataset. Specifically, it includes the following steps: During the inner loop training, mileage data from the support sets of each task are input into a base prediction model with shared initialization parameters to obtain wear prediction values; based on the loss between the wear prediction values ​​and the wear label data in the support sets, the shared initialization parameters are updated using gradients to obtain the model parameters for each task; During the outer loop training, mileage data from the query sets of each task are input into the model trained in the inner loop to obtain query set prediction values; the meta-loss between the predicted query values ​​and wear labels for each task is calculated, and based on the meta-loss, gradient descent is used to optimize the shared initialization parameters of the base prediction model to complete the training of the meta-learning model. The prediction module is used to input the mileage data of the carbon sliding plate of the pantograph under test into the trained meta-learning prediction model, so as to output the corresponding wear prediction value of the carbon sliding plate.

9. A computer device for predicting the wear of a pantograph carbon sliding plate, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the pantograph carbon slide plate wear prediction method according to any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which includes program instructions that, when executed by a processor, are used to perform the steps of the pantograph carbon slide plate wear prediction method according to any one of claims 1-7.

Images