Carbon brush failure prediction method and system

Abstract

The application discloses a kind of carbon brush fault prediction method and system, method includes collecting carbon brush relevant data, and the data collected is preprocessed;The data after pre-processing is input into the optimized prediction model, and three kinds of prediction results such as carbon brush temperature anomaly occurrence probability, carbon brush remaining wear life, bolt loosening risk level are output simultaneously.By collecting carbon brush multi-source heterogeneous data and microsecond time synchronization and classification preprocessing, build exclusive feature subset according to fault type;For temperature anomaly, wear overrun, bolt loosening three kinds of faults, respectively design working condition adaptive fusion, prior constraint regression, frequency band sensitive classification customized prediction submodel, realize multi-fault parallel accurate prediction and graded early warning, support equipment intelligent state maintenance.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring and intelligent operation and maintenance technology for rotating electric machines, specifically to a method and system for predicting carbon brush failures. Background Technology

[0002] Carbon brushes are core, easily damaged components in power transmission equipment such as rotating electric machines and excitation generator sets. They bear the critical functions of current conduction and commutation, and their operating status directly determines the reliability and power supply stability of the entire equipment. During long-term operation, carbon brushes are prone to typical faults such as abnormal temperature burn-out, excessive wear, and loose brush holder bolts. If these faults are not detected and addressed in a timely manner, they can easily lead to equipment downtime, arcing, or even safety accidents. As industrial equipment operation and maintenance shifts towards intelligent and condition-based maintenance, data-driven carbon brush fault prediction technology has become a core technical means to improve operation and maintenance efficiency and reduce the risk of unplanned downtime.

[0003] Currently, carbon brush failure prediction technology has made some progress, but there are still many shortcomings in practical industrial applications: First, the accuracy of multi-source heterogeneous data fusion is insufficient. Existing solutions mostly directly collect multimodal data such as temperature, vibration, and images for prediction. However, the sampling rates of different types of sensors vary significantly, and there is a lack of a high-precision unified time synchronization mechanism, which easily leads to temporal misalignment of multimodal data and distortion of feature correlation. At the same time, data preprocessing mostly adopts a generalized and unified process, without designing differentiated processing schemes for the characteristics of numerical time series, surface images, and vibration signals. As a result, data noise and acquisition bias cannot be filtered out in a targeted manner, and the data quality is difficult to support high-precision prediction.

[0004] Secondly, the feature set and prediction model have a low matching degree. Existing technologies often extract all features and directly input them into the prediction model without classifying and aggregating features according to fault type. Features of irrelevant faults are prone to noise interference, which increases the computational complexity of the model and weakens the specificity of single-type fault prediction. At the same time, most solutions use a single general model to adapt to all fault types, which cannot match the completely different occurrence mechanisms and data characteristics of three types of faults: abnormal temperature, wear and deterioration, and loose bolts. It is difficult to ensure the prediction accuracy of multiple types of faults at the same time.

[0005] Third, single-type fault prediction algorithms have significant technical shortcomings. In temperature anomaly prediction, many algorithms employ single regression algorithms or fixed-weight fusion algorithms, failing to fully consider the significant impact of changes in motor load, speed, and other operating conditions on carbon brush temperature. This leads to a substantial increase in prediction errors under varying operating conditions, resulting in high false alarm and false negative rates for temperature anomaly warnings. In wear life prediction, many algorithms rely solely on image or time-series features for single-dimensional prediction, ignoring the inherent gradual physical decay of carbon brush wear. This easily leads to abrupt changes in predicted life values. Furthermore, the loss function does not account for the tolerance of noise from on-site measurements, making the model prone to overfitting to minor acquisition fluctuations, resulting in insufficient stability and rationality in long-term life prediction. In bolt loosening identification, many algorithms directly classify and train the entire frequency band vibration spectrum. A large amount of irrelevant environmental noise in these bands can drown out the weak characteristic signals of bolt loosening, resulting in a low signal-to-noise ratio for fault features. Simultaneously, treating all fault samples equally without distinguishing the severity levels of different degrees of loosening leads to a high false negative rate for high-risk, severe loosening faults, making it difficult to meet on-site safety production requirements.

[0006] Fourth, the models have poor long-term operational adaptability. Most existing prediction models adopt an offline training and fixed deployment mode. Although some solutions have incremental updates, they mostly use full retraining at fixed intervals, which cannot adapt to changes in data distribution caused by aging field equipment and shifting operating conditions. Moreover, incremental updates do not have reasonable triggering mechanisms and parameter protection mechanisms. Frequent updates will increase the computational burden, and untimely updates will lead to continuous degradation of model performance, resulting in high maintenance costs for long-term field operation of equipment.

[0007] In summary, existing carbon brush fault prediction technologies have significant shortcomings in areas such as accurate synchronous fusion of multimodal data, fault-oriented feature and model adaptation, scenario-specific prediction algorithm optimization, and adaptive long-term iteration, making it difficult to meet the application requirements of high precision, high reliability, and long cycle for carbon brush condition monitoring in industrial settings. Summary of the Invention

[0008] The main objective of this invention is to provide a carbon brush failure prediction method and system to address the shortcomings of related technologies.

[0009] To achieve the above objectives, according to a first aspect of the present invention, a carbon brush failure prediction method is provided, comprising collecting carbon brush-related data and preprocessing the collected data; inputting the preprocessed data into an optimized prediction model, and simultaneously outputting three types of prediction results: the probability of carbon brush temperature anomaly, the remaining wear life of the carbon brush, and the risk level of bolt loosening. Before inputting the preprocessed data into the optimized prediction model, features corresponding to different failure types are extracted from the preprocessed data, and the extracted features are classified and aggregated according to the failure type to form subsets of temperature anomaly features, wear anomaly features, and bolt loosening features. Specifically, the carbon brush temperature anomaly prediction uses a weighted fusion model of multiple linear regression and support vector machine regression (SVR); the carbon brush wear amount and remaining service life prediction uses a cascaded fusion model of convolutional neural network (CNN) and SVR; and the carbon brush holder bolt loosening prediction uses a CNN classification model based on vibration spectrum.

[0010] According to a second aspect of the present invention, a carbon brush failure prediction system is provided, comprising a preprocessing unit for collecting carbon brush-related data and preprocessing the collected data; and a prediction unit for inputting the preprocessed data into an optimized prediction model and simultaneously outputting three prediction results: the probability of carbon brush temperature anomaly, the remaining wear life of the carbon brush, and the risk level of bolt loosening. Before inputting the preprocessed data into the optimized prediction model, features corresponding to different failure types are extracted from the preprocessed data, and the extracted features are classified and aggregated according to the failure type to form subsets of temperature anomaly features, wear anomaly features, and bolt loosening features. Specifically, the carbon brush temperature anomaly prediction uses a weighted fusion model of multiple linear regression and support vector machine regression (SVR); the carbon brush wear amount and remaining service life prediction uses a cascaded fusion model of convolutional neural network (CNN) and SVR; and the carbon brush holder bolt loosening prediction uses a CNN classification model based on vibration spectrum.

[0011] According to a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing the computer to perform the method described in any one of the first aspects.

[0012] According to a fourth aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method described in any implementation of the first aspect.

[0013] This embodiment presents a carbon brush fault prediction method and system. The method includes collecting carbon brush-related data and preprocessing the collected data. The preprocessed data is then input into an optimized prediction model, simultaneously outputting three prediction results: the probability of abnormal carbon brush temperature, the remaining wear life of the carbon brush, and the risk level of bolt loosening. Before inputting the preprocessed data into the optimized prediction model, features corresponding to different fault types are extracted from the preprocessed data. These extracted features are then classified and aggregated according to the fault type, forming subsets for abnormal temperature features, abnormal wear features, and bolt loosening features. By collecting multi-source heterogeneous carbon brush data and performing microsecond-level time synchronization and classification preprocessing, specific feature subsets are constructed according to fault type. For the three types of faults—abnormal temperature, excessive wear, and bolt loosening—customized prediction sub-models are designed using adaptive fusion of operating conditions, prior constraint regression, and frequency band sensitive classification, respectively. This enables parallel and accurate prediction and hierarchical early warning of multiple faults, supporting intelligent condition-based maintenance of equipment. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the carbon brush failure prediction method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] According to embodiments of the present invention, a method is provided, such as Figure 1 As shown, steps 101 to 102 are included below: Step 101: Collect relevant data on the carbon brushes and preprocess the collected data; In this step, carbon brush body operation data, unit operating data, and environmental data are collected respectively. The carbon brush body operation data includes carbon brush surface temperature, contact current, surface wear image, and brush holder vibration signal. The unit operating data includes motor load percentage, speed, stator voltage, and stator current. The environmental data includes ambient temperature, relative humidity, and dust concentration. All data acquisition devices are connected to a unified time synchronization system, and hardware-level clock synchronization is achieved through the Beidou pulse-per-second (PPS) signal. A unified timestamp with an accuracy of no less than 1 microsecond is added to each collected data sample. To address the sampling rate differences of different data types, a time alignment interpolation algorithm is used to unify all data to the same time reference. For the synchronized numerical time-series data (including carbon brush surface temperature, contact current, unit operating data, and environmental data), outliers are first removed using the 3σ principle combined with the sliding window midpoint method. Then, linear interpolation is used to fill in short-term missing values, and K-nearest neighbor interpolation is used to fill in long-term missing values. Finally, the min-max normalization method is used to map all numerical data to the [0,1] interval.

[0020] For the synchronized image data (carbon brush surface wear images), a threshold-based segmentation method is used to extract the carbon brush contact area, Gaussian filtering is used to remove image noise, and adaptive histogram equalization is used to enhance wear edge features. For the synchronized vibration signal (brush grip vibration signal), a short-time Fourier transform is used to convert the time-domain vibration signal into a two-dimensional time-frequency spectrum, and the time resolution of the spectrum is consistent with the sampling interval of the numerical time-series data.

[0021] By achieving microsecond-level timestamp synchronization through a unified time synchronization system, coupled with a differentiated classification preprocessing mechanism, the common industry problems of large differences in multimodal data sampling rates, inconsistent time bases, and heterogeneous modal features are solved. This ensures that numerical, image, and vibration data are strictly aligned in the time dimension, eliminates prediction errors caused by time series deviations, and provides a high-quality data foundation for subsequent multimodal feature extraction and model training.

[0022] Step 102: Input the preprocessed data into the optimized prediction model and simultaneously output three prediction results: the probability of carbon brush temperature anomaly, the remaining wear life of the carbon brush, and the risk level of bolt loosening. Before inputting the preprocessed data into the optimized prediction model, features corresponding to different fault types are extracted from the preprocessed data. The extracted features are classified and aggregated according to the fault type to form temperature anomaly feature subset, wear anomaly feature subset, and bolt loosening feature subset, respectively.

[0023] In this embodiment, a dedicated feature subset is constructed according to the fault type to form a clearly categorized multi-dimensional feature library, avoiding feature redundancy and interference caused by the mixing of multiple fault features; each prediction sub-model only calls the dedicated feature subset corresponding to the fault, which greatly reduces the model input dimension and computational load, while reducing the interference of irrelevant features on the model decision, and improving the pertinence and computational efficiency of single-type fault prediction.

[0024] Multiple models are used to adapt to different fault mechanisms, avoiding the shortcomings of a single algorithm that cannot take into account multi-dimensional faults. Each model extracts thermal, image, and vibration features in a targeted manner, resulting in higher prediction accuracy. Parallel computing enables the simultaneous output of the three types of indicators. Based on the cross-validation of the prediction results by fault coupling relationship, false alarms and false negatives are greatly reduced. Multi-dimensional equipment health assessments are generated simultaneously, providing a complete quantitative basis for graded early warning and condition-based maintenance.

[0025] As an optional implementation of this embodiment, the temperature anomaly training feature subset is divided into a first training set and a first validation set. The first training set is used to train the multiple linear regression submodule and the SVR submodule respectively. During the training process, the working condition correlation weight factor is introduced to adjust the contribution of the two submodules to the temperature prediction result in order to obtain the temperature prediction value. The first validation set is used to calculate the prediction error of the two submodules under different working conditions. The fusion coefficient matrix corresponding to each working condition interval is determined based on the error reciprocal weighting method to obtain the trained carbon brush temperature anomaly prediction submodel.

[0026] As an optional implementation in this embodiment, the temperature anomaly training feature subset includes carbon brush temperature features and corresponding motor load and speed operating condition features. During the training process, an operating condition correlation weight factor is introduced, and the contribution of the two sub-modules to the temperature prediction result is dynamically adjusted according to the real-time operating condition features. This includes: identifying the preset operating condition interval to which the current operating condition belongs based on the real-time operating condition features, and then obtaining the operating condition correlation weight factor corresponding to the operating condition interval; dynamically weighting and fusing the intermediate prediction results of the two sub-modules according to the operating condition correlation weight factor to obtain the final temperature prediction value.

[0027] During model training, temperature anomaly training feature subsets, wear anomaly training feature subsets, and bolt loosening training feature subsets corresponding one-to-one with each fault type are extracted from the labeled historical fault dataset. To address the imbalance of carbon brush fault samples, feature-level data augmentation based on SMOTE is performed on the fault samples in each training feature subset to generate virtual fault feature samples and balance the positive and negative sample ratios of each training feature subset.

[0028] When determining the operating condition correlation weight factor, the operating conditions of the motor are pre-divided into three intervals: low load, medium load, and high load. For each operating condition interval, the average prediction error of the multiple linear regression submodule and the SVR submodule on historical data is calculated. The reciprocal of the average prediction error of the two submodules is used as the corresponding operating condition correlation weight factor for that operating condition interval.

[0029] The subset of temperature anomaly characteristics includes time-domain statistical features of temperature, such as the mean, peak, valley, and standard deviation of carbon brush surface temperature, the rate of temperature rise per unit time, the temperature gradient, and the temperature fluctuation coefficient; electrical correlation features, such as the mean carbon brush contact current, current fluctuation rate, current imbalance, and contact voltage drop; operating condition matching features, such as motor load percentage, real-time speed, stator current, and stator voltage; and environmental influence features, such as ambient temperature and relative humidity.

[0030] For example, a subset of temperature anomaly features and samples of corresponding temperature labels are used to form a subset of temperature anomaly training features. The subset of temperature anomaly training features is randomly divided into a first training set and a first validation set according to a preset ratio (usually 8:2). The division process uses a stratified sampling method to ensure that the sample distribution of each operating condition interval in the training set and the validation set is consistent. All features in the first training set and the first validation set are standardized, and the mean and standard deviation of each feature in the training set are calculated. All features are converted into a standard normal distribution with a mean of 0 and a variance of 1.

[0031] Furthermore, a predefined rule for dividing motor operating condition ranges is established, creating a mapping relationship between load percentage and operating condition ranges: load < 30% is the low load range, 30% ≤ load < 70% is the medium load range, and load ≥ 70% is the high load range. A unique range identifier (e.g., 0, 1, 2) is assigned to each operating condition range. All samples in the first training set and the first validation set are traversed, and a corresponding operating condition range label is added to each sample based on its load feature value.

[0032] Further configure basic training hyperparameters such as batch size, number of training epochs, and learning rate. Define that after every N training epochs or after processing M batches of samples, update the fusion coefficients for each operating condition interval (the fusion coefficient is the decision weight ratio of the two prediction submodules (multivariate linear regression and SVR) in the final temperature prediction result under a single operating condition, including two values: w_LR (weight of linear regression submodule) and w_SVR (weight of SVR submodule), and satisfying w_LR + w_SVR = 1). Set initial fusion coefficients for all operating condition intervals, typically setting the initial weights of both the multiple linear regression submodule and the SVR submodule to 0.5. Randomly select a batch of samples from the first training set according to the batch size; extract the feature vector, true temperature label, and operating condition interval label of the batch of samples; input the batch feature vector into both the multiple linear regression submodule and the SVR submodule simultaneously. The multiple linear regression submodule calculates the intermediate predicted temperature value T_LR for each sample through linear transformation, while the SVR submodule calculates the intermediate predicted temperature value T_SVR for each sample through kernel function mapping and regression. Iterate through each sample in the batch, and based on its operating condition interval label, obtain the corresponding current fusion coefficients w_LR and w_SVR from the fusion coefficient matrix (the fusion coefficient matrix is ​​a structured set of fusion coefficients corresponding to all operating condition intervals, divided into low, medium, and high load operating condition intervals, so the fusion coefficient matrix has a fixed structure of 3 rows and 2 columns: each row corresponds to an operating condition interval (row 1 low load, row 2 medium load, row 3 high load); each column corresponds to the weight of a sub-module (column 1 w_LR, column 2 w_SVR). (In each training batch, based on the sample's operating condition interval label, extract the corresponding row's fusion coefficient from the matrix, and complete the weighted fusion calculation for each sample). Calculate the final predicted temperature value of each sample according to the formula T_pred = w_LR×T_LR + w_SVR ×T_SVR, and generate the final predicted temperature vector for all samples in the batch.

[0033] Based on the final predicted temperature vector and the true temperature label vector, the mean squared error (MSE) loss function value is calculated. Gradient descent is then used for backpropagation to update the weight parameters of the multiple linear regression submodule and the support vectors and kernel function parameters of the SVR submodule. During backpropagation, the fusion coefficients are treated as constants in the calculation and are not updated using gradients.

[0034] When the preset weight update cycle is reached, the two sub-modules in the current training state are evaluated using the first validation set. The first validation set is divided into three subsets according to the working condition interval label. The mean absolute error (MAE) of the multiple linear regression sub-module and the SVR sub-module on each working condition subset is calculated. The fusion coefficients w_LR = (1 / MAE_LR) / (1 / MAE_LR + 1 / MAE_SVR) and w_SVR = (1 / MAE_SVR) / (1 / MAE_LR + 1 / MAE_SVR) for each working condition interval are recalculated using the error inverse weighting method. The fusion coefficient matrix is ​​updated with the newly calculated fusion coefficients, and the next training iteration begins.

[0035] Training is terminated early when the mean absolute error on the validation set decreases by less than a preset threshold over K consecutive weight update cycles to prevent overfitting. After training is terminated, the two trained sub-modules are evaluated using the complete first validation set; the final fusion coefficients for each operating condition interval are recalculated to generate the final operating condition fusion coefficient matrix.

[0036] This implementation addresses the issues of carbon brush temperature being greatly affected by load and rotation speed, and the poor adaptability of a single algorithm. It adopts a dual-algorithm weighted fusion architecture, combining operating condition interval division with a dynamic fusion coefficient matrix of error reciprocal weighting, enabling the model to automatically adjust algorithm weights according to real-time operating conditions. Compared with a fixed-weight fusion scheme, it significantly reduces temperature prediction errors under varying operating conditions, especially under complex operating conditions such as high load, significantly improving prediction accuracy and ensuring the reliability of temperature anomaly warnings.

[0037] As an optional implementation of this embodiment, when training the convolutional neural network (CNN) and SVR cascade fusion model, the method includes: dividing the wear anomaly training feature subset into a second training set and a second validation set; using the second training set to train the CNN submodule to extract deep semantic features from the wear image; concatenating the deep semantic features with the temporal wear features in the wear anomaly training feature subset and inputting the result into the SVR submodule for regression training; wherein, during the training process, a priori curve of carbon brush wear life decay is introduced as a constraint condition to penalize prediction results that deviate from the normal wear pattern; and using the second validation set to adjust the kernel function parameters and penalty coefficient of the SVR submodule to obtain the trained carbon brush wear amount and remaining service life prediction submodel.

[0038] The subset of wear anomaly features integrates two dimensions: visual morphology and temporal evolution. It corresponds to the cross-modal input of the wear life prediction sub-model and specifically includes three categories: image dimension features such as carbon brush contact area ratio, wear edge roughness, surface texture uniformity, wear depth equivalent value, and contact area eccentricity; and temporal evolution features such as cumulative runtime, historical average wear rate, recent periodic wear increment, and cumulative wear amount.

[0039] As an optional implementation of this embodiment, the CNN submodule is trained using a second training set to extract deep semantic features from the wear image. The deep semantic features are then concatenated with the temporal wear features in the wear anomaly training feature subset and input into the SVR submodule for regression training. This includes: extracting carbon brush surface image features from the second training set and inputting them into the CNN submodule. The CNN submodule employs a three-branch parallel convolutional structure, using three different sized convolutional kernels to extract wear texture features at different scales. After concatenating the feature maps output from the three branches, a unified-dimensional deep semantic feature is generated through a global average pooling layer. Temporal wear features are extracted from the second training set, and the temporal wear features are concatenated with the deep semantic feature vector in the feature dimension to obtain a joint feature vector that integrates image semantics and temporal change information. The joint feature vector is then input into the SVR submodule for regression training, using the actual wear amount of the carbon brush as the label and an ε-insensitive loss function as the base loss.

[0040] In the above-mentioned optional implementation, all carbon brush wear image samples in the second training set are read, and after uniform size and grayscale preprocessing, they are fed into a three-branch parallel CNN structure. The three independent convolutional paths are configured with 3×3, 5×5, and 7×7 convolutional kernels, respectively, to simultaneously extract image dimensional features. Each branch completes convolution and pooling operations to output channel feature maps. Global and local wear information is spliced ​​and fused along the channel dimension, and then compressed by global average pooling to generate a fixed-dimensional deep image semantic vector without spatial redundancy.

[0041] Furthermore, temporal evolution features, such as cumulative carbon brush runtime and historical wear rate, are extracted synchronously from the same training samples. These one-dimensional temporal features are then concatenated with the output image semantic vector along the feature dimension to form a joint feature vector that integrates visual wear morphology and time decay patterns. This unifies the input dimension to adapt to the subsequent regression network. Training is conducted using the joint feature vector as input and the measured carbon brush wear thickness as the label. The basic loss employs an ε-insensitive loss, calculating the loss only for predicted values ​​exceeding the allowable error. Simultaneously, the wear decay prior curve for the corresponding carbon brush model is retrieved, and the standard wear amount for the same runtime is read. The deviation between the predicted value and the standard value is calculated. If the deviation exceeds the limit, a secondary penalty term is added. The basic loss and prior penalty are integrated to obtain the total loss function, constraining the model to conform to the physical wear patterns.

[0042] The ε-insensitive loss is the SVR loss function, which sets an error tolerance threshold ε: when the absolute value of the difference between the predicted value and the true label is less than or equal to ε, no loss is calculated; loss is only calculated when the deviation exceeds ε. Its core purpose is to allow the model to have small, normal measurement errors, without forcing a fit to every minute fluctuation. Carbon brush images and current acquisition contain normal measurement noise at the micrometer level; ε corresponds to the allowable measurement error in the field. Minor wear fluctuations are normal acquisition deviations; forcing a fit to all of them would cause the model to overfit. This loss ignores small noise within ε, only penalizing truly significant wear deviations, allowing the model to focus on the actual wear trend of the carbon brush and weakening the interference of acquisition disturbances.

[0043] Furthermore, a priori curve of carbon brush wear life decay is introduced as a constraint during training to penalize prediction results that deviate from the normal wear pattern. The kernel function parameters and penalty coefficient of the SVR submodule are adjusted using a second validation set to obtain the trained carbon brush wear amount and remaining service life prediction sub-model. Based on the measured wear data of different models of carbon brushes throughout their entire life cycle, a standardized wear life prior curve is generated by fitting an exponential decay function to establish a one-to-one mapping relationship between runtime and standard wear amount. Before training, the curve numerical model and matching rules for each model are stored in the model reading interface. During training, the theoretical standard wear amount at that moment is retrieved in real time according to the carbon brush model and cumulative runtime labeled in the sample.

[0044] Furthermore, the SVR submodule outputs the predicted wear amount for a single sample, and retrieves the corresponding standard wear amount by combining the device runtime t carried by the sample. The absolute deviation between the predicted value and the theoretical value is calculated, and this deviation characterizes the degree to which the predicted result deviates from the inherent wear and decay law of the carbon brush. A preset allowable wear deviation threshold is set; if the deviation is less than or equal to the threshold, it means the prediction conforms to the normal wear law of the carbon brush, no constraint needs to be applied, and the prior loss is 0; if the deviation is greater than the threshold, it indicates that the prediction violates the physical decay characteristics, and a secondary penalty formula is introduced.

[0045] Furthermore, using the SVR's native ε-insensitive base loss as the underlying fitting loss, and superimposing the aforementioned prior regularization loss term, a total loss is synthesized. During the training backpropagation phase, the total loss is used as the optimization objective. When updating model parameters, both the fitting accuracy of the measured samples and the objective physical laws of carbon brush wear are taken into account, suppressing abnormal wear prediction outputs that deviate from reality.

[0046] The aforementioned basic loss is responsible for fitting the measured wear data of a single sample. The prior regularization term serves as a global physical constraint, limiting the model from learning prediction results that violate the life decay trend due to local noise and abnormal samples, avoiding overfitting, and improving the stability and rationality of the long-term prediction of the remaining life of the carbon brush.

[0047] Furthermore, a second validation set is used for model verification. All preset parameter combinations of the radial basis kernel γ and penalty coefficient C are traversed by grid search. After training each set of parameters, the average absolute error of wear prediction on the validation set is statistically analyzed. The parameter combination corresponding to the minimum error is selected and solidified to complete the training of the integrated prediction sub-model for carbon brush wear and remaining life.

[0048] A multi-scale CNN and SVR cascade architecture is adopted, which integrates image wear morphology and temporal wear patterns. ε-insensitive loss is introduced to filter measurement noise and avoid model overfitting to minor acquisition fluctuations. The wear life decay prior curve is superimposed as a physical regularization constraint to penalize prediction results that deviate from normal wear patterns, effectively suppressing jumps and abnormal outputs in life prediction, and significantly improving the stability, rationality, and long-term prediction accuracy of carbon brush remaining life prediction.

[0049] As an optional implementation in this embodiment, the bolt loosening training feature subset is divided into a third training set and a third validation set. The bolt loosening training feature subset consists of different frequency band spectrogram features obtained by converting vibration signals. Before training, the spectrogram features are ranked by frequency band importance, and the k feature frequency bands with the highest correlation to bolt loosening faults are selected. The CNN classifier is trained using the third training set. During training, higher weights are assigned to the selected high-correlation frequency band features, and differentiated loss function weights are set for fault samples with different degrees of loosening. The convolution kernel size and pooling layer parameters of the CNN are adjusted using the third validation set to obtain the trained carbon brush grip bolt loosening prediction sub-model.

[0050] As an optional implementation in this embodiment, the frequency band importance of the spectrogram features is ranked before training, and the k feature frequency bands with the highest correlation to bolt loosening faults are selected. A third training set is used to train the CNN classifier. During training, higher weights are assigned to the selected high-correlation frequency band features, and differentiated loss function weights are set for fault samples with different degrees of loosening. The third validation set is used to adjust the convolution kernel size and pooling layer parameters of the CNN to obtain the trained carbon brush holder bolt loosening prediction sub-model. Based on the absolute value of the Pearson correlation coefficient of the k high-correlation feature frequency bands, the feature weight coefficient of each frequency band is calculated. The calculated frequency band weight coefficients are multiplied by the spectrogram features of the corresponding frequency bands to obtain the weighted spectrogram features. The weighted spectrogram features are then input into the CNN. The classifier uses a predefined weighted cross-entropy loss function as the optimization objective. The backpropagation algorithm is used to iteratively train the CNN classifier and update the weight parameters of each layer of the network. The trained CNN classifier is validated using a third validation set. The convolution kernel size and pooling layer parameters of the CNN are adjusted by grid search. The parameter combination that gives the highest F1 score on the validation set is selected to obtain the trained carbon brush holder bolt loosening prediction sub-model.

[0051] In the above-mentioned optional implementation methods, the bolt loosening feature subset extracts all vibration time-frequency spectrum features, performs non-repeating random partitioning according to a fixed layer ratio, and generates an independent third training set and a third validation set. The spectral feature data distribution of the two datasets is kept consistent to prevent data distribution shift from interfering with the model training effect. All spectral features are generated from the original vibration signal through short-time Fourier transform, and are uniformly divided into multiple independent frequency band features according to frequency intervals, forming the complete set of original input features for the model. The importance ranking of frequency bands is completed based on the Pearson correlation coefficient: the correlation coefficient between the spectral amplitude sequence of a single frequency band and the bolt loosening degree label is calculated one by one, and the absolute value of the coefficient is used to represent the fault correlation strength of the frequency band. All frequency bands are sorted from high to low values, and the top three groups of frequency bands are selected as high-correlation core frequency bands, while the remaining frequency bands are classified as ordinary low-correlation frequency bands.

[0052] The absolute values ​​of the Pearson coefficients corresponding to the three highly correlated frequency bands are extracted, and a weight conversion rule is established: the basic weight of the low-correlation frequency band is fixed at 1, and the weight of the high-correlation frequency band is linearly mapped to the range of 2-5 times according to the corresponding correlation coefficient. The higher the correlation coefficient, the larger the weight value is assigned. The spectral feature tensor of all frequency bands is weighted channel by channel. The original feature of a single frequency band is multiplied element by element with its own calculated weight to complete the adaptive scaling of the feature amplitude. This amplifies the expression of fault-sensitive frequency band data and suppresses irrelevant frequency band noise information. All weighted frequency band tensors are merged to form a complete weighted spectral feature matrix, which is used as the unified input material for CNN.

[0053] Based on bolt preload loss standards, four categories are defined: normal, slightly loose, moderately loose, and severely loose, sorted from low to high severity. Independent loss weights are assigned to each category, with higher severity samples receiving larger weights, achieving differentiated control of the loss contribution of fault samples. Using standard cross-entropy as the basic framework, corresponding fault level weights are bound to each sample's loss calculation term, forming a weighted cross-entropy loss function. During model training, severely loose samples will generate larger loss values ​​due to classification errors, forcing the network to prioritize learning high-risk fault features and reducing the probability of missing major loosening faults.

[0054] The weighted spectral features output are batch-input into the CNN network. The network performs convolutional feature extraction, pooling dimensionality reduction, and fully connected classification layer by layer to obtain the predicted probability distribution of each category. The predicted probability is compared with the true fault label of the sample, and the total batch loss is calculated by substituting it into the weighted cross-entropy formula. The backpropagation algorithm is used to calculate the derivative along the network chain, and the gradient of the parameters of each convolutional layer and fully connected layer is calculated layer by layer. The weight parameters of all networks are updated in conjunction with the preset learning rate. The batch forward inference and backward gradient update are performed iteratively to complete multiple rounds of training.

[0055] A grid search global optimization mechanism is adopted to traverse all parameter combinations of preset convolution kernel size and pooling window. After completing the complete training process under each parameter combination, the F1 score of the model is calculated by inputting the third validation set. The F1 score is used as the sole evaluation index of the model's comprehensive recognition ability. After traversing all parameter combinations, the convolution and pooling parameter configurations corresponding to the maximum F1 score of the validation set are selected. The network structure and all training weight parameters are fixed, and the final bolt looseness classification prediction sub-model is output.

[0056] This implementation addresses the issues of high noise levels in vibration signals and the tendency for bolt loosening features to be overlooked. It uses Pearson correlation coefficient to filter high-correlation frequency bands and weights them to enhance the signal-to-noise ratio of fault features. Combined with a graded differentiated loss function, loss weights are allocated according to the severity of the fault, guiding the model to prioritize learning the features of severe loosening faults. This significantly reduces the missed detection rate of high-risk bolt loosening faults and improves the overall accuracy and comprehensive performance of classification and recognition.

[0057] Parallel reasoning of three types of fault sub-models synchronously outputs three types of results: abnormal temperature probability, remaining wear life, and bolt loosening risk, which can comprehensively cover the core fault modes of carbon brushes. Combined with the cross-validation mechanism of fault physical correlation, the probability of false alarms and missed alarms is further reduced. The multi-dimensional quantitative results can provide accurate data support for equipment graded early warning and planned maintenance, thereby improving the intelligence level of carbon brush operation and maintenance and the safety of equipment operation.

[0058] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0059] According to an embodiment of the present invention, a method is also provided.

[0060] According to embodiments of the present invention, the present invention also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the methods described in any of the above embodiments.

[0061] According to embodiments of the present invention, the present invention also provides a readable storage medium storing computer instructions that enable a computer to perform the methods described in any of the above embodiments when executed.

[0062] According to embodiments of the present invention, the present invention also provides a computer program product that, when executed by a processor, can implement the methods described in any of the above embodiments.

[0063] Figure 2 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0064] like Figure 2 As shown, the electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0065] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0066] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as the object matching method. For example, in some embodiments, the object matching method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of the methods described above may be performed.

[0067] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0068] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0069] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

Claims

1. A method for predicting carbon brush failure, characterized in that, include: Collect relevant data on carbon brushes and preprocess the collected data; The preprocessed data is input into the optimized prediction model, and three prediction results are output simultaneously: the probability of carbon brush temperature anomaly, the remaining wear life of carbon brush, and the risk level of bolt loosening. Before inputting the preprocessed data into the optimized prediction model, features corresponding to different fault types are extracted from the preprocessed data. The extracted features are classified and aggregated according to the fault type to form temperature anomaly feature subset, wear anomaly feature subset, and bolt loosening feature subset, respectively. Among them, the carbon brush temperature anomaly prediction adopts a weighted fusion model of multiple linear regression and support vector machine regression (SVR); the carbon brush wear and remaining service life prediction adopts a cascaded fusion model of convolutional neural network (CNN) and SVR; and the carbon brush holder bolt loosening prediction adopts a CNN classification model based on vibration spectrum.

2. The carbon brush failure prediction method according to claim 1, characterized in that, The method for training the multiple linear regression submodule and the SVR submodule includes: The temperature anomaly training feature subset is divided into a first training set and a first validation set. The first training set is used to train the multiple linear regression submodule and the SVR submodule respectively. During the training process, the working condition correlation weight factor is introduced to adjust the contribution of the two submodules to the temperature prediction results in order to obtain the temperature prediction value. The first validation set is used to calculate the prediction error of the two submodules under different working conditions. The fusion coefficient matrix corresponding to each working condition interval is determined based on the error reciprocal weighting method, so as to obtain the trained carbon brush temperature anomaly prediction submodel.

3. The carbon brush failure prediction method according to claim 1, characterized in that, The methods for training a cascaded fusion model of a convolutional neural network (CNN) and an SVR include: The abnormal wear training feature subset is divided into a second training set and a second validation set. The second training set is used to train the CNN submodule to extract deep semantic features from the wear images. The deep semantic features are then concatenated with the temporal wear features in the abnormal wear training feature subset and input into the SVR submodule for regression training. During the training process, a priori curve of carbon brush wear life decay is introduced as a constraint to penalize prediction results that deviate from the normal wear pattern. The second validation set is used to adjust the kernel function parameters and penalty coefficient of the SVR submodule to obtain the trained carbon brush wear amount and remaining service life prediction submodel.

4. The carbon brush failure prediction method according to claim 1, characterized in that, The methods used in training CNN classification models include: The bolt loosening training feature subset was divided into a third training set and a third validation set. The bolt loosening training feature subset consisted of spectrogram features of different frequency bands obtained from vibration signal conversion. Before training, the spectrogram features were ranked by frequency band importance, and the k feature frequency bands with the highest correlation to bolt loosening faults were selected. The CNN classifier was trained using the third training set. During training, higher weights were assigned to the selected high-correlation frequency band features, and differentiated loss function weights were set for fault samples with different degrees of loosening. The convolution kernel size and pooling layer parameters of the CNN were adjusted using the third validation set to obtain the trained carbon brush grip bolt loosening prediction sub-model.

5. The carbon brush failure prediction method according to claim 2, characterized in that, The temperature anomaly training feature subset includes carbon brush temperature features and corresponding operating condition features; During training, a working condition-related weighting factor is introduced to dynamically adjust the contribution of the two sub-modules to the temperature prediction results based on real-time working condition characteristics. Identify the preset operating condition interval to which the current operating condition belongs based on real-time operating condition characteristics, and obtain the operating condition association weight factor corresponding to the operating condition interval; The final temperature prediction value is obtained by dynamically weighting and fusing the intermediate prediction results of the two sub-modules based on the working condition correlation weight factor.

6. The carbon brush failure prediction method according to claim 3, characterized in that, The second training set was used to train the CNN submodule to extract deep semantic features from the wear image; The deep semantic features are concatenated with the temporal wear features in the wear anomaly training feature subset and then input into the SVR submodule for regression training, including: The carbon brush surface image features are extracted from the second training set and input into the CNN submodule. The CNN submodule adopts a three-branch parallel convolution structure, and uses three different sizes of convolution kernels to extract wear texture features at different scales. After the feature maps output by the three branches are concatenated by channels, a deep semantic feature with uniform dimension is generated through a global average pooling layer. The wear amount temporal features are extracted from the second training set. The wear amount temporal features are then concatenated with the deep semantic feature vector in the feature dimension to obtain a joint feature vector that integrates image semantics and temporal change information. The joint feature vector is input into the SVR submodule for regression training, with the actual wear of the carbon brush as the label and the ε-insensitive loss function as the base loss.

7. The carbon brush failure prediction method according to claim 4, characterized in that, Before training, the frequency band importance of the spectrogram features was ranked, and the k feature frequency bands with the highest correlation to bolt loosening faults were selected. The CNN classifier was trained using the third training set. During the training process, higher weights were assigned to the selected high-relevance frequency band features, and different loss function weights were set for fault samples with different degrees of loosening. The kernel size and pooling layer parameters of the CNN were adjusted using the third validation set to obtain the trained carbon brush holder bolt loosening prediction sub-model. Based on the absolute values ​​of the Pearson correlation coefficients of k highly correlated feature frequency bands, the feature weight coefficients of each frequency band are calculated; the calculated frequency band weight coefficients are multiplied by the spectrum features of the corresponding frequency bands to obtain the weighted spectrum features. The weighted spectrogram features are input into the CNN classifier. The predefined weighted cross-entropy loss function is used as the optimization objective. The backpropagation algorithm is used to iteratively train the CNN classifier and update the weight parameters of each layer of the network. The CNN classifier in training was validated using a third validation set. The kernel size and pooling layer parameters of the CNN were adjusted by grid search. The parameter combination that gives the highest F1 score on the validation set was selected to obtain the trained carbon brush holder bolt loosening prediction sub-model.

8. A carbon brush failure prediction system, characterized in that, include: The preprocessing unit is used to collect carbon brush-related data and preprocess the collected data. The prediction unit is used to input the preprocessed data into the optimized prediction model and simultaneously output three prediction results: the probability of carbon brush temperature anomaly, the remaining wear life of the carbon brush, and the risk level of bolt loosening. Before inputting the preprocessed data into the optimized prediction model, features corresponding to different fault types are extracted from the preprocessed data. The extracted features are classified and aggregated according to the fault type to form temperature anomaly feature subsets, wear anomaly feature subsets, and bolt loosening feature subsets, respectively. Among them, the carbon brush temperature anomaly prediction adopts a weighted fusion model of multiple linear regression and support vector machine regression (SVR); the carbon brush wear and remaining service life prediction adopts a cascaded fusion model of convolutional neural network (CNN) and SVR; and the carbon brush holder bolt loosening prediction adopts a CNN classification model based on vibration spectrum.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1-7.

10. An electronic device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method according to any one of claims 1-7.

Images