A large parameter model adaptive updating method for power equipment state prediction

Abstract

The present invention discloses a large parameter model adaptive updating method for power equipment state prediction, which relates to the field of model training technology. The present invention uses machine training to obtain a state prediction model of power equipment; calculates the feature novelty and condition coverage of new data; formulates a new data screening mechanism to judge and screen the new data value evaluation indicators during the monitoring process; calculates the task relevance and contribution attenuation factor of historical data; formulates a historical data screening mechanism, and uses the historical data screening mechanism to screen the historical data during the monitoring process; uses the initial learning rate to formulate a dynamic learning rate update strategy, and dynamically adjusts the learning rate of the state prediction model; constructs a hybrid training strategy, and uses the hybrid training strategy to update the state prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of model training, and in particular to a large parameter model adaptive updating method for power equipment state prediction. Background Art

[0002] During operation, the state of power equipment is affected by numerous factors, including high-voltage electric fields, heat, mechanical forces, operating conditions, and the meteorological environment. Equipment state changes and fault evolution patterns are embodied in a wealth of state information, including live testing, online monitoring, and patrol tests, as well as operating conditions, environmental climate, and grid operations. Accurately understanding the state of power equipment and promptly detecting potential faults are crucial for ensuring the safe and stable operation of power systems. With the development of smart grids, equipment testing methods are constantly expanding, and the amount of data generated by grid operation and equipment testing is growing exponentially. At the same time, the presence of a large amount of anomalies in the collected data complicates power status prediction and places higher demands on data processing and analysis methods. However, the continuous advancement of artificial intelligence and machine learning technologies has provided new approaches and methods for power equipment status prediction. Algorithms such as deep learning can automatically extract features from large amounts of data, build complex nonlinear models, and accurately predict power equipment status.

[0003] In order to increase the breadth and accuracy of models today, large-parameter models based on artificial intelligence need to be updated in real time. However, there are deficiencies in the updating process. On the one hand, the focus is on training new data, which may cause the model to forget historical data; on the other hand, the focus is on historical data, and the training of new data is not thorough, which makes the model unable to make accurate predictions for new features. Therefore, it is crucial to balance the new data and historical data when updating the model. Summary of the Invention

[0004] The purpose of the present invention is to provide a large parameter model adaptive updating method for power equipment state prediction to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A large parameter model adaptive updating method for power equipment state prediction, the method comprising the following steps:

[0007] S100, collecting equipment data from historical operations of the power equipment, analyzing all equipment data when the power equipment is in normal and abnormal states, extracting feature data reflecting the state of the power equipment, and using the feature data for machine training to obtain a state prediction model for the power equipment;

[0008] Furthermore, the specific steps for using feature data for machine training to obtain a state prediction model for power equipment are as follows:

[0009] S101. Collect equipment data from the historical operation of power equipment. The equipment data includes sensor data, equipment logs, and equipment environment data. According to the equipment logs, the collected equipment data is divided into two states: normal state and abnormal state, and labels are generated. Specifically, the normal state Y=0, the abnormal state Y=1; for each continuous time series equipment data, the sample generated with the window length L and step size is: [X t （i） ,Y （i） ∈(0,1)],X t （i） represents the i-th type of continuous device data based on time series, Y （i） Indicates the two device status labels for the i-th device data and standardizes all sample data;

[0010] S102. Extract device features from all device data using a neural network encoder, calculate statistical features of all collected device features, including mean, variance, peak, and crest factor; extract frequency domain features of the device features through Fourier transform; calculate the covariance matrix of the device features, and extract the upper triangular elements in the covariance matrix as correlation features; combine the statistical features, frequency domain features, and correlation features to obtain state features of the power equipment; calculate the mutual information between each state feature and label y, sort the mutual information of all state features and label y from largest to smallest, and select the top M state features in the sort as training features, where M represents a screening threshold, and M is set based on model training experience;

[0011] S103. Use the training features of historical device data to construct a training set and a validation set, use the temporal convolutional network and the attention mechanism to build a state prediction model, set the loss function, use the training set to train the state prediction model, and use the validation set to verify the trained state prediction model.

[0012] S200, using a state prediction model to monitor the state of the power equipment in real time while the power equipment is operating, collecting real-time data of the power equipment operation as new data during the monitoring process, and calculating feature novelty and condition coverage of the new data;

[0013] Furthermore, the specific steps for combining the two data to obtain the value assessment index of the new data are as follows:

[0014] S201, using the state prediction model to monitor the state of the power equipment in real time when the power equipment is working, and collecting the real-time equipment data of the power equipment as new data x during the monitoring process. new, the neural network encoder is used to extract the device features in the new data, and the device features of the new data and the device features of the historical device data are used to calculate the feature novelty of the new data. The formula is:

[0015] ;

[0016] In the formula, N novelty represents the novelty of the features of the new data, m represents the number of samples of the new data, T(x u new ) represents the device characteristics of the u-th new data sample, p hist Represents the historical device feature mean, u ranges from 1 to m, The square of the Euclidean distance between the new data device feature and the mean of the historical device feature;

[0017] S202: Extract device features of different dimensions from the device features of historical device data, extract the maximum and minimum values of the device features in each dimension, use the maximum and minimum values to construct the historical interval B of the device features in each dimension, and extract the device feature data values v of different dimensions in the new data. new , using the historical interval to judge the mechanical energy of the equipment characteristic data value of the new data, when v new ∉B, define the output result J=1, when v new ∈B, define the output result J = 0; use the output result J to calculate the conditional coverage of the new data, the formula is:

[0018] ;

[0019] In the formula, C cover represents the conditional coverage of the new data, D represents the total dimension of the device features, and J d Represents the output result of the device feature of the d-th dimension, where d is from 1 to D.

[0020] By calculating the feature novelty and condition coverage of new data, we ensure that the abnormal features hidden in the new data (such as early insulation degradation and partial discharge) are quickly identified, shortening the model update cycle.

[0021] S300, combining the feature novelty and condition coverage of the new data to obtain a new data value assessment index, and formulating a new data screening mechanism to judge and screen the new data value assessment index during the monitoring process;

[0022] Furthermore, the specific steps for formulating a new data screening mechanism to judge and screen the new data value assessment indicators during the monitoring process are as follows:

[0023] S301. Combining the feature novelty and condition coverage of the new data to obtain the new data value evaluation index, the formula is:

[0024] ;

[0025] In the formula, V new Represents the value assessment index of new data, w1 represents the weight of the novelty of the new data features, and w2 represents the weight of the coverage of the new data conditions; w1 and w2 are set manually;

[0026] Set the value evaluation index threshold τ, and use the value evaluation index threshold τ to build a new data screening mechanism: V new >τ; When the new data screening mechanism is met, the model update is started and the new data is retained;

[0027] S302. Collect the accuracy change of the state prediction model after real-time update. Accuracy is the difference between the predicted value of the state prediction model and the actual value. Let the accuracy change before and after the real-time update of the state prediction model be △Z. When the accuracy change is positive, the output result is J'=-1. When the accuracy change is negative, the output result is J'=1. When the accuracy change is 0, the output result is J'=0. Use the accuracy change to formulate a value assessment indicator threshold update strategy, specifically:

[0028] ;

[0029] In the formula, τ' represents the updated value assessment indicator threshold.

[0030] Automatically tighten or relax data access conditions based on changes in validation set accuracy to avoid model degradation caused by fluctuations in data quality.

[0031] S400, analyzing historical data for training the power equipment state prediction model during the monitoring process, and calculating the task relevance and contribution attenuation factor of the historical data;

[0032] Furthermore, the specific steps for calculating the task relevance and contribution attenuation factor of historical data are as follows:

[0033] S401. Collect the model parameters θ in the state prediction model of the power equipment during the training process. For each historical data sample and new data sample, calculate the partial derivative of the loss function based on the model parameters, use the partial derivative as the loss gradient, and then convert the loss gradient into a one-dimensional vector. Let the loss gradient of the historical data sample converted into a one-dimensional vector be g hist , the loss gradient of the new data sample is converted into a one-dimensional vector g new ;

[0034] S402. Calculate the task relevance of historical data using the one-dimensional vector of the loss gradient of each historical data sample and the new data sample. The formula is:

[0035] ;

[0036] In the formula, R rel represents the task relevance of historical data, n represents the total number of historical data samples, g j hist Represents the one-dimensional vector of the loss gradient of the j-th historical data sample, cos represents the cosine similarity, and j ranges from 1 to n;

[0037] S403. Extract the time from the acquisition of historical data to the real-time monitoring of power equipment as the data age of the historical data. Collect the gradient L of the validation set with respect to the model parameters during state prediction model training. Calculate the contribution attenuation factor of the historical data using the data age of the historical data and the validation set gradient. The formula is:

[0038] ;

[0039] In the formula, A decay It represents the contribution attenuation factor of historical data, t represents the data age of historical data, β represents the attenuation coefficient, and α represents the attenuation rate. Both the attenuation coefficient and the attenuation rate are manually set based on the model training experience.

[0040] By calculating the task relevance of historical data, we eliminate interference from irrelevant historical data and avoid false triggering caused by outdated data. We also dynamically eliminate low-value historical data through the contribution attenuation factor, significantly reducing storage costs.

[0041] S500, combining the task relevance and contribution attenuation factor of the historical data to obtain a value assessment index for the historical data, formulating a historical data screening mechanism, and utilizing the historical data screening mechanism to screen the historical data during the monitoring process;

[0042] Furthermore, the specific steps for using the historical data screening mechanism to screen historical data during the monitoring process are as follows:

[0043] S501. Combining the task relevance and contribution attenuation factor of historical data, the value evaluation index of historical data is obtained. The formula is:

[0044] ;

[0045] In the formula, V hist Represents the value evaluation index of historical data; the historical data screening mechanism is constructed as follows:

[0046] Manually set the screening threshold E, set the value assessment indicators of all historical data to be sorted from large to small, and select the first E historical data to be retained.

[0047] S600: For the state prediction model of the power equipment, collect the initial learning rate of the model, formulate a dynamic learning rate update strategy using the initial learning rate, and dynamically adjust the learning rate of the state prediction model;

[0048] Furthermore, the specific steps for dynamically adjusting the learning rate of the state prediction model are as follows:

[0049] S601. For the state prediction model of the power equipment, collect the initial learning rate of the model and use the initial learning rate to formulate a dynamic learning rate update strategy. The formula is:

[0050] ;

[0051] In the formula, Lr s represents the updated learning rate, Lt0 represents the initial learning rate, r represents the number of training times, e -0.1r It indicates exponential decay during training r times.

[0052] S700: Based on the adjusted learning rate, the filtered new data and historical data are combined to construct a hybrid training strategy, and the state prediction model is updated using the hybrid training strategy.

[0053] Furthermore, the specific steps for updating the state prediction model using the hybrid training strategy are as follows:

[0054] S701, using the new data screening mechanism and the historical data screening mechanism to obtain the new data set after screening is F new , the filtered historical data set is F hist After judging the new data and starting the model update program, the dynamic learning rate update strategy is used to update the model's learning rate in real time. The state prediction model mixes the screened new data set and the historical data set according to the updated learning rate to formulate a mixed training strategy. The formula is:

[0055] ;

[0056] In the formula, LF z represents the total loss function, LF(f) represents the loss function of a single new data sample, LF(h) represents the loss function of a single historical data sample, and b represents the dynamic weight coefficient;

[0057] Calculating the average loss for a new dataset directly reflects the model's ability to adapt to the new data. A hybrid training strategy forces the model to prioritize learning device state changes reflected in the new data. If the new data contains signs of a fault, high loss terms drive the model to quickly adjust parameters to capture abnormal patterns.

[0058] S702. Use the value evaluation index of the new data to adjust the dynamic weight coefficient in real time. The formula is:

[0059] ;

[0060] In the formula, b represents the dynamic weight coefficient; the state prediction model is updated in real time using a hybrid training strategy. The fixed term 11 ensures that historical data is always included in the training to prevent complete forgetting. log(1+V new ) with V new Growth is increasing, but the growth rate is slowing down;

[0061] In a hybrid training strategy, when new data contains significant new features, increasing b forces the model to adapt to the new features while also strengthening its learning of relevant historical patterns. Regularization using historical data suppresses the model's oversensitivity to new data. If the new data is correlated with some historical data patterns, hybrid training enhances the model's ability to model continuous state changes.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. The present invention directly quantifies the contribution direction of historical data to the current parameter optimization by calculating the gradient cosine similarity, which improves the accuracy of correlation judgment in complex working conditions.

[0064] 2. The present invention combines the physical aging law of equipment with data-driven value by calculating the contribution attenuation factor, thereby solving the defect of insufficient modeling of equipment life cycle by pure data-driven methods.

[0065] 3. The present invention solves the industry pain points of low data utilization, severe model forgetting and high update costs in power equipment status prediction, and achieves the coordinated optimization of prediction accuracy, real-time performance and long-term stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of the steps of a large parameter model adaptive updating method for power equipment state prediction according to the present invention;

[0067] Figure 2 The present invention is a flow chart of a large parameter model adaptive updating method for power equipment state prediction. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0069] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0070] A large parameter model adaptive updating method for power equipment state prediction, the method comprising the following steps:

[0071] S100, collecting equipment data from historical operations of the power equipment, analyzing all equipment data when the power equipment is in normal and abnormal states, extracting feature data reflecting the state of the power equipment, and using the feature data for machine training to obtain a state prediction model for the power equipment;

[0072] The specific steps for using feature data for machine training to obtain a state prediction model for power equipment are as follows:

[0073] S101. Collect equipment data from the historical operation of power equipment. The equipment data includes sensor data, equipment logs, and equipment environment data. According to the equipment logs, the collected equipment data is divided into two states: normal state and abnormal state, and labels are generated. Specifically, the normal state Y=0, the abnormal state Y=1; for each continuous time series equipment data, the sample generated with the window length L and step size is: [X t （i） ,Y （i） ∈(0,1)],X t （i） represents the i-th type of continuous device data based on time series, Y （i） Indicates the two device status labels for the i-th device data and standardizes all sample data;

[0074] S102. Extract device features from all device data using a neural network encoder, calculate statistical features of all collected device features, including mean, variance, peak, and crest factor; extract frequency domain features of the device features through Fourier transform; calculate the covariance matrix of the device features, and extract the upper triangular elements in the covariance matrix as correlation features; combine the statistical features, frequency domain features, and correlation features to obtain state features of the power equipment; calculate the mutual information between each state feature and label y, sort the mutual information of all state features and label y from largest to smallest, and select the top M state features in the sort as training features, where M represents a screening threshold, and M is set based on model training experience;

[0075] S103. Use the training features of historical device data to construct a training set and a validation set, use the temporal convolutional network and the attention mechanism to build a state prediction model, set the loss function, use the training set to train the state prediction model, and use the validation set to verify the trained state prediction model.

[0076] S200, using a state prediction model to monitor the state of the power equipment in real time while the power equipment is operating, collecting real-time data of the power equipment operation as new data during the monitoring process, and calculating feature novelty and condition coverage of the new data;

[0077] The specific steps for combining the two types of data to obtain the value assessment index of the new data are:

[0078] S201, using the state prediction model to monitor the state of the power equipment in real time when the power equipment is working, and collecting the real-time equipment data of the power equipment as new data x during the monitoring process. new , the neural network encoder is used to extract the device features in the new data, and the device features of the new data and the device features of the historical device data are used to calculate the feature novelty of the new data. The formula is:

[0079] ;

[0080] In the formula, N novelty represents the novelty of the features of the new data, m represents the number of samples of the new data, T(x u new ) represents the device characteristics of the u-th new data sample, p hist Represents the historical device feature mean, u ranges from 1 to m, The square of the Euclidean distance between the new data device feature and the mean of the historical device feature;

[0081] S202: Extract device features of different dimensions from the device features of historical device data, extract the maximum and minimum values of the device features in each dimension, use the maximum and minimum values to construct the historical interval B of the device features in each dimension, and extract the device feature data values v of different dimensions in the new data. new , using the historical interval to judge the mechanical energy of the equipment characteristic data value of the new data, when v new ∉B, define the output result J=1, when v new ∈B, define the output result J = 0; use the output result J to calculate the conditional coverage of the new data, the formula is:

[0082] ;

[0083] In the formula, C cover represents the conditional coverage of the new data, D represents the total dimension of the device features, and J d Represents the output result of the device feature of the d-th dimension, where d is from 1 to D.

[0084] By calculating the feature novelty and condition coverage of new data, we ensure that the abnormal features hidden in the new data (such as early insulation degradation and partial discharge) are quickly identified, shortening the model update cycle.

[0085] S300, combining the feature novelty and condition coverage of the new data to obtain a new data value assessment index, and formulating a new data screening mechanism to judge and screen the new data value assessment index during the monitoring process;

[0086] The specific steps for formulating a new data screening mechanism to judge and screen new data value assessment indicators during the monitoring process are as follows:

[0087] S301. Combining the feature novelty and condition coverage of the new data to obtain the new data value evaluation index, the formula is:

[0088] ;

[0089] In the formula, V new Represents the value assessment index of new data, w1 represents the weight of the novelty of the new data features, and w2 represents the weight of the coverage of the new data conditions; w1 and w2 are set manually;

[0090] Set the value evaluation index threshold τ, and use the value evaluation index threshold τ to build a new data screening mechanism: V new >τ; When the new data screening mechanism is met, the model update is started and the new data is retained;

[0091] S302. Collect the accuracy change of the state prediction model after real-time update. Accuracy is the difference between the predicted value of the state prediction model and the actual value. Let the accuracy change before and after the real-time update of the state prediction model be △Z. When the accuracy change is positive, the output result is J'=-1. When the accuracy change is negative, the output result is J'=1. When the accuracy change is 0, the output result is J'=0. Use the accuracy change to formulate a value assessment indicator threshold update strategy, specifically:

[0092] ;

[0093] In the formula, τ' represents the updated value assessment indicator threshold.

[0094] Automatically tighten or relax data access conditions based on changes in validation set accuracy to avoid model degradation caused by fluctuations in data quality.

[0095] S400, analyzing historical data for training the power equipment state prediction model during the monitoring process, and calculating the task relevance and contribution attenuation factor of the historical data;

[0096] The specific steps for calculating the task relevance and contribution attenuation factor of historical data are as follows:

[0097] S401. Collect the model parameters θ in the state prediction model of the power equipment during the training process. For each historical data sample and new data sample, calculate the partial derivative of the loss function based on the model parameters, use the partial derivative as the loss gradient, and then convert the loss gradient into a one-dimensional vector. Let the loss gradient of the historical data sample converted into a one-dimensional vector be g hist , the loss gradient of the new data sample is converted into a one-dimensional vector g new ;

[0098] S402. Calculate the task relevance of historical data using the one-dimensional vector of the loss gradient of each historical data sample and the new data sample. The formula is:

[0099] ;

[0100] In the formula, R rel represents the task relevance of historical data, n represents the total number of historical data samples, g j hist Represents the one-dimensional vector of the loss gradient of the j-th historical data sample, cos represents the cosine similarity, and j ranges from 1 to n;

[0101] S403. Extract the time from the acquisition of historical data to the real-time monitoring of power equipment as the data age of the historical data. Collect the gradient L of the validation set with respect to the model parameters during state prediction model training. Calculate the contribution attenuation factor of the historical data using the data age of the historical data and the validation set gradient. The formula is:

[0102] ;

[0103] In the formula, A decay It represents the contribution attenuation factor of historical data, t represents the data age of historical data, β represents the attenuation coefficient, and α represents the attenuation rate. Both the attenuation coefficient and the attenuation rate are manually set based on the model training experience.

[0104] By calculating the task relevance of historical data, we eliminate interference from irrelevant historical data and avoid false triggering caused by outdated data. We also dynamically eliminate low-value historical data through the contribution attenuation factor, significantly reducing storage costs.

[0105] S500, combining the task relevance and contribution attenuation factor of the historical data to obtain a value assessment index for the historical data, formulating a historical data screening mechanism, and utilizing the historical data screening mechanism to screen the historical data during the monitoring process;

[0106] The specific steps for filtering historical data during monitoring using the historical data filtering mechanism are as follows:

[0107] S501. Combining the task relevance and contribution attenuation factor of historical data, the value evaluation index of historical data is obtained. The formula is:

[0108] ;

[0109] In the formula, V hist Represents the value evaluation index of historical data; the historical data screening mechanism is constructed as follows:

[0110] Manually set the screening threshold E, set the value assessment indicators of all historical data to be sorted from large to small, and select the first E historical data to be retained.

[0111] S600: For the state prediction model of the power equipment, collect the initial learning rate of the model, formulate a dynamic learning rate update strategy using the initial learning rate, and dynamically adjust the learning rate of the state prediction model;

[0112] The specific steps for dynamically adjusting the learning rate of the state prediction model are:

[0113] S601. For the state prediction model of the power equipment, collect the initial learning rate of the model and use the initial learning rate to formulate a dynamic learning rate update strategy. The formula is:

[0114] ;

[0115] In the formula, Lr s represents the updated learning rate, Lt0 represents the initial learning rate, r represents the number of training times, e -0.1r It indicates exponential decay during training r times.

[0116] S700: Based on the adjusted learning rate, the filtered new data and historical data are combined to construct a hybrid training strategy, and the state prediction model is updated using the hybrid training strategy.

[0117] The specific steps for updating the state prediction model using the hybrid training strategy are:

[0118] S701, using the new data screening mechanism and the historical data screening mechanism to obtain the new data set after screening is F new , the filtered historical data set is F hist After judging the new data and starting the model update program, the dynamic learning rate update strategy is used to update the model's learning rate in real time. The state prediction model mixes the screened new data set and the historical data set according to the updated learning rate to formulate a mixed training strategy. The formula is:

[0119] ;

[0120] In the formula, LF z represents the total loss function, LF(f) represents the loss function of a single new data sample, LF(h) represents the loss function of a single historical data sample, and b represents the dynamic weight coefficient;

[0121] Calculating the average loss for a new dataset directly reflects the model's ability to adapt to the new data. A hybrid training strategy forces the model to prioritize learning device state changes reflected in the new data. If the new data contains signs of a fault, high loss terms drive the model to quickly adjust parameters to capture abnormal patterns.

[0122] S702. Use the value evaluation index of the new data to adjust the dynamic weight coefficient in real time. The formula is:

[0123] ;

[0124] In the formula, b represents the dynamic weight coefficient; the state prediction model is updated in real time using a hybrid training strategy. The fixed term 11 ensures that historical data is always included in the training to prevent complete forgetting. log(1+V new ) with V new Growth is increasing, but the growth rate is slowing down;

[0125] In a hybrid training strategy, when new data contains significant new features, increasing b forces the model to adapt to the new features while also strengthening its learning of relevant historical patterns. Regularization using historical data suppresses the model's oversensitivity to new data. If the new data is correlated with some historical data patterns, hybrid training enhances the model's ability to model continuous state changes.

[0126] Example 1: Assume that a transformer monitoring system in a converter station needs to evaluate whether three newly collected samples trigger a model update. The feature dimensions of the new data are feature 1 and feature 2, and the historical intervals of the two feature dimensions are [0, 1.2] and [0.5, 2.0], respectively. The device feature data values in the two feature dimensions of sample 1 are 0.6 and 1.3, the device feature data values in the two feature dimensions of sample 2 are 0.3 and 0.8, respectively, and the device feature data values in the two feature dimensions of sample 3 are 1.2 and 2.1, respectively. Assume that the historical device feature data means are 0.5 and 1.2.

[0127] The feature novelty of the new data is calculated to be (0.02+0.2+1.3) / 3=0.5;

[0128] The output of the three samples of the new data based on the feature 1 dimension is 0, and the output based on the feature 2 dimension is 1; the conditional coverage of the new data is calculated to be 2;

[0129] Assume w1 and w2 are 0.7 and 0.3 respectively; the value evaluation index of the new data is calculated to be 0.95.

[0130] Example 2: Assume that in a transformer monitoring system, the new data is the detection of an abnormal increase in winding temperature, and the historical data is the precursor data of three similar temperature rise events in the past. The dynamic weight coefficient is calculated as follows: b = 1 + log (1 + 2.5) = 1 + 1.216 = 2.216. In the total loss LF, the weight of the historical data loss item is 2.216 times higher than the weight of the new data item. Through the hybrid training strategy, while fitting the new temperature rise pattern, the learning of similar historical events is strengthened to identify common features.

[0131] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A large parameter model adaptive updating method for power equipment state prediction, characterized by: The method comprises the following steps: S100, collecting equipment data from historical operations of the power equipment, analyzing all equipment data when the power equipment is in normal and abnormal states, extracting feature data reflecting the state of the power equipment, and using the feature data for machine training to obtain a state prediction model for the power equipment; S200, using a state prediction model to monitor the state of the power equipment in real time while the power equipment is operating, collecting real-time data of the power equipment operation as new data during the monitoring process, and calculating feature novelty and condition coverage of the new data; S300, combining the feature novelty and condition coverage of the new data to obtain a new data value assessment index, and formulating a new data screening mechanism to judge and screen the new data value assessment index during the monitoring process; The specific steps for formulating a new data screening mechanism to judge and screen new data value assessment indicators during the monitoring process are as follows: S301. Combining the feature novelty and condition coverage of the new data to obtain the new data value evaluation index, the formula is: In new =w1×N nowelty +w2×C cover ; In the formula, V new Represents the value assessment index of new data, w1 represents the weight of the novelty of the new data features, and w2 represents the weight of the coverage of the new data conditions; w1 and w2 are set manually; Set the value evaluation index threshold τ, and use the value evaluation index threshold τ to build a new data screening mechanism: V new >τ; When the new data screening mechanism is met, the model update is started and the new data is retained; S302. Collect the accuracy change of the state prediction model after real-time update. Accuracy is the difference between the predicted value of the state prediction model and the actual value. Let the accuracy change before and after the real-time update of the state prediction model be ΔZ. When the accuracy change is positive, the output result is J'=-1. When the accuracy change is negative, the output result is J'=1. When the accuracy change is 0, the output result is J'=0. Use the accuracy change to formulate a value assessment indicator threshold update strategy, specifically: τ′=τ+J′×0.1×τ; In the formula, τ' represents the updated value assessment index threshold; S400, analyzing historical data for training the power equipment state prediction model during the monitoring process, and calculating the task relevance and contribution attenuation factor of the historical data; S500, combining the task relevance and contribution attenuation factor of the historical data to obtain a value assessment index for the historical data, formulating a historical data screening mechanism, and utilizing the historical data screening mechanism to screen the historical data during the monitoring process; S600: For the state prediction model of the power equipment, collect the initial learning rate of the model, formulate a dynamic learning rate update strategy using the initial learning rate, and dynamically adjust the learning rate of the state prediction model; S700: Based on the adjusted learning rate, the filtered new data and historical data are combined to construct a hybrid training strategy, and the state prediction model is updated using the hybrid training strategy.

2. A large parameter model adaptive updating method for power equipment state prediction according to claim 1, characterized in that: The specific steps of using the feature data to perform machine training to obtain the state prediction model of the power equipment in S100 are: S101. Collect equipment data from the historical operation of power equipment. The equipment data includes sensor data, equipment logs, and equipment environment data. According to the equipment logs, the collected equipment data is divided into two states: normal state and abnormal state, and a label is generated. Specifically, the normal state Y=0, the abnormal state Y=1; for each continuous time series equipment data, the sample generated with the window length L and step size is: [X t (i) ,Y (i) ∈(0,1)],X t (i) represents the i-th type of continuous device data based on time series, Y (i) Indicates the two device status labels for the i-th device data and standardizes all sample data; S102. Extract device features of all device data using a neural network encoder, calculate statistical features of all collected device features, including mean, variance, peak, and crest factor; and extract frequency domain features of the device features using Fourier transform. Calculate the covariance matrix of the equipment characteristics and extract the upper triangular elements in the covariance matrix as correlation features; combine the statistical features, frequency domain features and correlation features to obtain the state characteristics of the power equipment; Calculate the mutual information between each state feature and label y, sort the mutual information between all state features and label y from large to small, and select the top M state features in the sorting as training features. M represents the screening threshold, and M is set according to model training experience. S103. Use the training features of historical device data to construct a training set and a validation set, use the temporal convolutional network and the attention mechanism to build a state prediction model, set the loss function, use the training set to train the state prediction model, and use the validation set to verify the trained state prediction model.

3. The method for adaptively updating a large parameter model for power equipment state prediction according to claim 2, characterized in that: The specific steps of combining the two data to obtain the value assessment index of the new data in S200 are: S201, using the state prediction model to monitor the state of the power equipment in real time when the power equipment is working, and collecting the real-time equipment data of the power equipment as new data x during the monitoring process. new , the neural network encoder is used to extract the device features in the new data, and the device features of the new data and the device features of the historical device data are used to calculate the feature novelty of the new data. The formula is: In the formula, N novelty represents the novelty of the features of the new data, m represents the number of samples of the new data, T(x u new ) represents the device characteristics of the u-th new data sample, p hist Represents the historical device feature mean, u ranges from 1 to m, The square of the Euclidean distance between the new data device feature and the mean of the historical device feature; S202: Extract device features of different dimensions from the device features of historical device data, extract the maximum and minimum values of the device features in each dimension, use the maximum and minimum values to construct the historical interval B of the device features in each dimension, and extract the device feature data values v of different dimensions in the new data. new , using the historical interval to judge the mechanical energy of the equipment characteristic data value of the new data, when When the output result J=1, when v new ∈B, define the output result J = 0; use the output result J to calculate the conditional coverage of the new data, the formula is: In the formula, C cover represents the conditional coverage of the new data, D represents the total dimension of the device features, and J d Represents the output result of the device feature of the d-th dimension, where d is from 1 to D.

4. The method for adaptively updating a large parameter model for power equipment state prediction according to claim 2, characterized in that: The specific steps of calculating the task relevance and contribution attenuation factor of the historical data in S400 are: S401. Collect the model parameters θ in the state prediction model of the power equipment during the training process. For each historical data sample and new data sample, calculate the partial derivative of the loss function based on the model parameters, use the partial derivative as the loss gradient, and then convert the loss gradient into a one-dimensional vector. Let the loss gradient of the historical data sample converted into a one-dimensional vector be g hist , the loss gradient of the new data sample is converted into a one-dimensional vector g new ; S402. Calculate the task relevance of historical data using the one-dimensional vector of the loss gradient of each historical data sample and the new data sample. The formula is: In the formula, R rel represents the task relevance of historical data, n represents the total number of historical data samples, g j hist Represents the one-dimensional vector of the loss gradient of the j-th historical data sample, cos represents the cosine similarity, and j ranges from 1 to n; S403. Extract the time from the acquisition of historical data to the real-time monitoring of power equipment as the data age of the historical data. Collect the gradient L of the validation set with respect to the model parameters during state prediction model training. Calculate the contribution attenuation factor of the historical data using the data age of the historical data and the validation set gradient. The formula is: A decay =α×e -βt +(1-a)×L; In the formula, A decay It represents the contribution attenuation factor of historical data, t represents the data age of historical data, β represents the attenuation coefficient, and α represents the attenuation rate. Both the attenuation coefficient and the attenuation rate are manually set based on the model training experience.

5. The method for adaptively updating a large parameter model for power equipment state prediction according to claim 4, characterized in that: The specific steps of using the historical data screening mechanism to screen historical data during the monitoring process in S500 are: S501. Combining the task relevance and contribution attenuation factor of historical data, the value evaluation index of historical data is obtained. The formula is: V hist =R rel ×A decay ; In the formula, V hist Represents the value evaluation index of historical data; the historical data screening mechanism is constructed as follows: Manually set the screening threshold E, set the value assessment indicators of all historical data to be sorted from large to small, and select the first E historical data to be retained.

6. The method for adaptively updating a large parameter model for power equipment state prediction according to claim 2, characterized in that: The specific steps of dynamically adjusting the learning rate of the state prediction model in S600 are: S601. For the state prediction model of the power equipment, collect the initial learning rate of the model and use the initial learning rate to formulate a dynamic learning rate update strategy. The formula is: In the formula, Lr s represents the updated learning rate, Lt0 represents the initial learning rate, r represents the number of training times, e -0.1r It indicates exponential decay during training r times.

7. A large parameter model adaptive updating method for power equipment state prediction according to claim 6, characterized in that: The specific steps of updating the state prediction model using the hybrid training strategy in S700 are: S701, using the new data screening mechanism and the historical data screening mechanism to obtain the new data set after screening is F new , the filtered historical data set is F hist After judging the new data and starting the model update program, the dynamic learning rate update strategy is used to update the model's learning rate in real time. The state prediction model mixes the screened new data set and the historical data set according to the updated learning rate to formulate a mixed training strategy. The formula is: In the formula, LF z represents the total loss function, LF(f) represents the loss function of a single new data sample, LF(h) represents the loss function of a single historical data sample, and b represents the dynamic weight coefficient; S702. Use the value evaluation index of the new data to adjust the dynamic weight coefficient in real time. The formula is: b=1+log(1+V new ); In the formula, b represents the dynamic weight coefficient; the state prediction model is updated in real time using a hybrid training strategy.

Images