A multi-time sequence parallel InceptionNeXt-GGhostNet super capacitor residual service life classification prediction method

Abstract

The application provides a multi-time sequence parallel InceptionNeXt-GGhostNet super capacitor residual service life classification prediction method, and the steps of the method are as follows: firstly, the charging and discharging capacity data of the super capacitor are respectively subjected to exponential moving average, multi-curve atlas and multi-time sequence integral processing; then the data subjected to the exponential moving average processing is input into an open-end convolutional network for feature extraction, and the data subjected to the multi-atlas curve processing is input into a graphic processing ghost network for feature extraction; finally, the features of the open-end convolutional network and the graphic processing ghost network and the data subjected to the multi-time sequence integral processing are input into a deep full connection layer network for optimized classification prediction results. The method solves the error caused by the super capacitor capacity data fluctuation to prediction, enhances the feature extraction capability and the time sequence dependence relationship in the captured data, and improves the super capacitor residual service prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of supercapacitors and relates to a classification and prediction method for the remaining service life of multi-time-parallel InceptionNeXt-GGhostNet supercapacitors, which is applicable to the prediction of the remaining service life of supercapacitors. Background Technology

[0002] Supercapacitors are widely used energy storage components, characterized by high power density, fast response speed, and long cycle life, and are extensively used in electric vehicles and energy storage systems. The purpose of predicting the remaining lifespan of supercapacitors is to replace them before they reach their failure threshold. Furthermore, when a supercapacitor is nearing failure, accurate prediction of its remaining lifespan is crucial to minimize the impact on the power system caused by untimely replacement of aging supercapacitors due to prediction errors.

[0003] Furthermore, many parameters characterize supercapacitor aging when predicting the remaining lifespan of supercapacitors, such as peak power, peak current, and charge / discharge capacity. Charge / discharge capacity data is widely used for predicting supercapacitor lifespan due to its advantage of continuous degradation. However, due to the complex internal structure of supercapacitors and data sampling errors, the charge / discharge capacity data fluctuates. This fluctuation affects the accuracy of the predicted remaining lifespan. In addition, in the early stages of supercapacitor degradation, the data sampled in adjacent charge / discharge cycles shows little change, resulting in insignificant sample differences and impacting the accuracy of the predicted remaining lifespan.

[0004] Therefore, in order to improve the accuracy of prediction and replace the supercapacitor in time before it fails before its remaining service life, a multi-time-parallel InceptionNeXt-GGhostNet supercapacitor remaining service life classification prediction method is proposed. Summary of the Invention

[0005] This invention proposes a multi-temporal parallel InceptionNeXt-GGhostNet supercapacitor remaining lifespan classification and prediction method. It divides the remaining lifespan of the supercapacitor into 50 lifespan intervals for classification and prediction, reducing errors caused by data fluctuations and losses due to prediction errors. It uses the exponential moving average method to address the problem of supercapacitor capacity data fluctuations, reducing errors caused by these fluctuations. It employs a multi-curve graph method to address the issue of indistinct differences in input sample features, improving classification accuracy. It uses a multi-temporal integral processing method to capture the long-term dependencies of temporal features in the supercapacitor capacity data. It uses open convolutional networks and graph processing phantom networks to extract features, improving feature extraction efficiency. Finally, it uses a deep fully connected layer network to integrate open convolutional networks, graph processing phantom networks, and temporal features, optimizing the classification and prediction results and improving prediction accuracy. The steps in the process are as follows:

[0006] Step (1): Obtain the supercapacitor charge and discharge capacity data. The method is as follows: First, record the charging capacity or discharging capacity once per second in a complete charge and discharge cycle to obtain the charge and discharge capacity data of one cycle; then, sample the charge and discharge capacity data of one cycle equally 60 times to obtain the charge and discharge capacity sequence of one cycle; then, generate a sample from the charge and discharge capacity sequence of every 10 consecutive cycles until the remaining service life of the supercapacitor first drops to the failure threshold; therefore, the size of a single sample is 10×60.

[0007] Step (2): Process the collected samples using the exponential moving average method, the multi-curve graph method, and the multi-time series integral processing method respectively;

[0008] The smoothed value at time t in the exponential moving average method is:

[0009] p t =ω×x t +(1-ω)×p t-1 (1)

[0010] Where t is the index number at time p t p is the smoothed value at time t. t-1 Let ω be the smoothed value at time t-1, and let x be the value of x. t The decay weight, (1-ω) is p t-1 The decay weight, x t The true value at time t;

[0011] The specific content of the multi-curve graph method is as follows: 60 sampled data points from each complete charge-discharge cycle are plotted into curves, with each curve containing 60 data points; one sample contains 10 curves.

[0012] The specific content of the multi-time series integration processing method is as follows: calculate the maximum value, minimum value, average value, variance and average difference of 60 sampled data in each complete charge-discharge cycle data to obtain a multi-time series integration sequence; generate a multi-time series integration sample every 10 consecutive cycles of the multi-time series integration sequence, and each multi-time series integration sample has a size of 10×5;

[0013] The variance is calculated as follows:

[0014]

[0015] Where θ is the variance, m is the total number of variables, and m-1 is the number of variables excluding y. k The total number of variables, k is the index number of the data, y k As variables, The average value of the variable. For variable y k Compared with the average The difference, For variable y k Compared with the average The square of the difference, where ∑ is the summation symbol;

[0016] The method for calculating the average difference is as follows:

[0017]

[0018] Where ψ is the mean difference, m is the total number of variables, k is the index number of the data, and y k As variables, The average value of the variable. For variable y k Compared with the average The difference, For variable y k Compared with the average The absolute value of the difference, where ∑ is the summation symbol;

[0019] Step (3): Input the samples processed using the exponential moving average method into the open convolutional network for feature extraction; input the samples processed using the multi-curve graph method into the graph processing phantom network for feature extraction;

[0020] The structure of the open-ended convolutional network model is as follows:

[0021] enter;

[0022] The number of access channels is 96, and the stride is 4×4 convolution.

[0023] Then connect 3 more initial convolutional modules;

[0024] The number of access channels is 192, and the stride is 2×2 convolution.

[0025] Then connect 3 more initial convolutional modules;

[0026] The number of access channels is 384, and the stride is 2×2 convolution.

[0027] Then connect 9 more initial convolutional modules;

[0028] The number of access channels is 768, and the stride is 2×2 convolution.

[0029] Then connect 3 more initial convolutional modules;

[0030] Then add another 1×1 global average pooling;

[0031] Then connect two more fully connected layers;

[0032] Add another probability distribution layer;

[0033] Output;

[0034] The structure of the opening convolutional module is as follows:

[0035] enter;

[0036] Divided into 2 branches;

[0037] Branch 1 is input that has not undergone any processing;

[0038] Branch 2 connects to a separate layer;

[0039] It is further divided into 4 branches;

[0040] Branch 1 is performed without any processing;

[0041] Branch 2 connects to an 11×1 depthwise separable convolution;

[0042] Branch 3 connects to a 1×11 depthwise separable convolution;

[0043] Branch 4 connects to a 3×3 depthwise separable convolution;

[0044] Then connect branch 1, branch 2, branch 3 and branch 4 to a feature fusion layer;

[0045] Then connect one more normalization layer;

[0046] Connect another multilayer sensor;

[0047] Then add the results from branch 1 and the multilayer perceptron output together;

[0048] Output;

[0049] The structure of the graphics processing phantom network is as follows:

[0050] enter;

[0051] A 3×3 convolution with 16 access channels;

[0052] Then connect a graphics processing phantom residual module with 24 channels;

[0053] Divided into 2 branches;

[0054] Branch 1 is a graphics processing phantom residual module with 24 access channels;

[0055] Branch 2 is a 1×1 convolution with 24 access channels;

[0056] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0057] Then connect a graphics processing phantom residual module with 40 channels;

[0058] Divided into 2 branches;

[0059] Branch 1 is a graphics processing phantom residual module with 40 access channels;

[0060] Branch 2 is a 1×1 convolution with 40 access channels;

[0061] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0062] Then connect to the graphics processing phantom residual module with 80 channels;

[0063] Divided into 2 branches;

[0064] Branch 1 connects to 5 cascaded graphics processing phantom residual modules with a total of 80 channels;

[0065] Branch 2 is a 1×1 convolution with 80 access channels;

[0066] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0067] Divided into 2 branches;

[0068] Branch 1 connects to 5 cascaded graphics processing phantom residual modules with a total of 160 channels;

[0069] Branch 2 is a 1×1 convolution with 160 access channels;

[0070] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0071] Reintegrate a 1×1 convolution with 960 channels;

[0072] Then add another 1×1 global average pooling;

[0073] Then connect two more fully connected layers;

[0074] Add another probability distribution layer;

[0075] Output;

[0076] The structure of the graphics processing phantom residual module is as follows:

[0077] enter;

[0078] Divided into 2 branches;

[0079] Branch 1 is input that has not undergone any processing;

[0080] Branch 2 involves connecting a 1×1 convolution;

[0081] Then add another 3×3 depthwise separable convolution;

[0082] Then add another 1×1 convolution;

[0083] Then add branch 1 and branch 2 together;

[0084] Output;

[0085] The classification result in the probability distribution layer is z. i The probability is:

[0086]

[0087] Where i is the index number of the classification category, z i For the classification result, S(z) i The classification result is z. i The probability, For z i The exponent is base e, n is the total number of categories, and ∑ is the summation symbol;

[0088] Step (4): Input the features extracted by the initial convolutional network, the features extracted by the image processing phantom network, and the multi-temporal integral samples after multi-temporal integral processing into the deep fully connected layer network to optimize the classification prediction results and obtain the classification prediction results of the remaining service life of the supercapacitor.

[0089] The specific structure of the deep fully connected layer network is as follows:

[0090] enter;

[0091] Connect to a fully connected layer with 600 channels;

[0092] Then connect another fully connected layer with 300 channels;

[0093] Then connect another fully connected layer with 150 channels;

[0094] Then connect another fully connected layer with 100 channels;

[0095] Then connect another fully connected layer with 50 channels;

[0096] Add another probability distribution layer;

[0097] Add one more category layer;

[0098] Output.

[0099] The present invention has the following advantages and effects compared with the prior art:

[0100] (1) This invention divides the remaining service life of supercapacitors into 50 service life intervals for classification and prediction, thereby reducing the impact of data fluctuations on the prediction of the remaining service life of supercapacitors. At the same time, the service life intervals can reduce the large-scale losses to the power system caused by prediction errors when predicting the remaining service life of supercapacitors that are about to fail.

[0101] (2) This invention uses the exponential moving average method to smooth the fluctuating supercapacitor capacity data, reducing the error caused by the fluctuation of capacity data in the prediction of remaining service life; it uses the multi-curve graph method to solve the problem of indistinct differences in input sample features and improve the accuracy of classification; it uses the multi-time series integral processing method to capture the long-term dependence of time features in supercapacitor capacity data and improve the accuracy of prediction.

[0102] (3) This invention uses a front-end convolutional network to extract samples processed by the exponential moving average method and a graph processing phantom network to extract samples processed by the multi-curve graph method. The two networks can mine deep features in the data and improve the accuracy of classification. At the same time, a deep fully connected layer network is used to integrate the front-end convolutional network, the graph processing phantom network and the temporal features to optimize the classification prediction results and improve the accuracy of classification prediction. Attached Figure Description

[0103] Figure 1 This is a flowchart of a multi-time-parallel InceptionNeXt-GGhostNet supercapacitor classification and prediction method in this invention.

[0104] Figure 2 This is a structural diagram of the initial convolutional network model in this invention.

[0105] Figure 3 This is a structural diagram of the starting convolution module in this invention.

[0106] Figure 4This is a structural diagram of the phantom network model for graphics processing in this invention.

[0107] Figure 5 This is a structural diagram of the graphics processing phantom residual module in this invention.

[0108] Figure 6 This is a structural diagram of the deep fully connected layer network in this invention. Detailed Implementation

[0109] This invention proposes a classification and prediction method for the remaining lifespan of multi-time-parallel InceptionNeXt-GGhostNet supercapacitors, which is described in detail below with reference to the accompanying drawings:

[0110] Figure 1 This is a flowchart of a multi-temporal parallel InceptionNeXt-GGhostNet supercapacitor remaining lifespan classification and prediction method according to the present invention. First, the supercapacitor's charge and discharge capacity data are processed using exponential moving average, multi-curve graphs, and multi-temporal integration, respectively. Then, the data processed by exponential moving average is input into an open convolutional network for feature extraction, and the data processed by multi-curve graphs is input into a graph processing phantom network for feature extraction. Finally, the features from the open convolutional network and the graph processing phantom network, along with the data processed by multi-temporal integration, are input into a deep fully connected layer network to optimize the classification and prediction results.

[0111] Figure 2 This is a structural diagram of the open-ended convolutional network model in this invention. The structure of the open-ended convolutional network model is as follows:

[0112] enter;

[0113] The number of access channels is 96, and the stride is 4×4 convolution.

[0114] Then connect 3 more initial convolutional modules;

[0115] The number of access channels is 192, and the stride is 2×2 convolution.

[0116] Then connect 3 more initial convolutional modules;

[0117] The number of access channels is 384, and the stride is 2×2 convolution.

[0118] Then connect 9 more initial convolutional modules;

[0119] The number of access channels is 768, and the stride is 2×2 convolution.

[0120] Then connect 3 more initial convolutional modules;

[0121] Then add another 1×1 global average pooling;

[0122] Then connect two more fully connected layers;

[0123] Add another probability distribution layer;

[0124] Output;

[0125] Figure 3 This is a structural diagram of the starting convolution module in this invention. The structure of the starting convolution module is as follows:

[0126] enter;

[0127] Divided into 2 branches;

[0128] Branch 1 is input that has not undergone any processing;

[0129] Branch 2 connects to a separate layer;

[0130] It is further divided into 4 branches;

[0131] Branch 1 is performed without any processing;

[0132] Branch 2 connects to an 11×1 depthwise separable convolution;

[0133] Branch 3 connects to a 1×11 depthwise separable convolution;

[0134] Branch 4 connects to a 3×3 depthwise separable convolution;

[0135] Then connect branch 1, branch 2, branch 3 and branch 4 to a feature fusion layer;

[0136] Then connect one more normalization layer;

[0137] Connect another multilayer sensor;

[0138] Then add the results from branch 1 and the multilayer perceptron output together;

[0139] Output.

[0140] Figure 4 This is a structural diagram of the graphics processing phantom network model in this invention. The structure of the graphics processing phantom network model is as follows:

[0141] enter;

[0142] A 3×3 convolution with 16 access channels;

[0143] Then connect a graphics processing phantom residual module with 24 channels;

[0144] Divided into 2 branches;

[0145] Branch 1 is a graphics processing phantom residual module with 24 access channels;

[0146] Branch 2 is a 1×1 convolution with 24 access channels;

[0147] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0148] Then connect a graphics processing phantom residual module with 40 channels;

[0149] Divided into 2 branches;

[0150] Branch 1 is a graphics processing phantom residual module with 40 access channels;

[0151] Branch 2 is a 1×1 convolution with 40 access channels;

[0152] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0153] Then connect to the graphics processing phantom residual module with 80 channels;

[0154] Divided into 2 branches;

[0155] Branch 1 connects to 5 cascaded graphics processing phantom residual modules with a total of 80 channels;

[0156] Branch 2 is a 1×1 convolution with 80 access channels;

[0157] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0158] Divided into 2 branches;

[0159] Branch 1 connects to 5 cascaded graphics processing phantom residual modules with a total of 160 channels;

[0160] Branch 2 is a 1×1 convolution with 160 access channels;

[0161] Then perform feature fusion processing on branch 1 and branch 2 once more;

[0162] Reintegrate a 1×1 convolution with 960 channels;

[0163] Then add another 1×1 global average pooling;

[0164] Then connect two more fully connected layers;

[0165] Add another probability distribution layer;

[0166] Output.

[0167] Figure 5 This is a structural diagram of the graphics processing phantom residual module in this invention. The structure of the graphics processing phantom residual module is as follows:

[0168] enter;

[0169] Divided into 2 branches;

[0170] Branch 1 is input that has not undergone any processing;

[0171] Branch 2 involves connecting a 1×1 convolution;

[0172] Then add another 3×3 depthwise separable convolution;

[0173] Then add another 1×1 convolution;

[0174] Then add branch 1 and branch 2 together;

[0175] Output.

[0176] Figure 6 This is a structural diagram of the deep fully connected layer network in this invention. The structure of the deep fully connected layer network is as follows:

[0177] enter;

[0178] Connect to a fully connected layer with 600 channels;

[0179] Then connect another fully connected layer with 300 channels;

[0180] Then connect another fully connected layer with 150 channels;

[0181] Then connect another fully connected layer with 100 channels;

[0182] Then connect another fully connected layer with 50 channels;

[0183] Add another probability distribution layer;

[0184] Add one more category layer;

[0185] Output.

[0186] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for classifying and predicting the remaining lifespan of multi-time-parallel InceptionNeXt-GGhostNet supercapacitors, characterized in that, The remaining lifespan of supercapacitors is divided into 50 categories for classification and prediction to reduce losses caused by prediction errors. The exponential moving average method is used to address the fluctuation problem of supercapacitor capacity data, reducing errors caused by capacity data fluctuations; a multi-curve graph method is used to address the problem of indistinct differences in input sample features, improving classification accuracy; a multi-temporal integral processing method is used to capture the long-term dependencies of temporal features in supercapacitor capacity data; open convolutional networks and graph processing phantom networks are used to extract features, improving feature extraction efficiency; a deep fully connected layer network is used to integrate open convolutional networks, graph processing phantom networks, and temporal features to optimize classification prediction results and improve prediction accuracy; the steps in the process are as follows: Step (1): Obtain the supercapacitor charge and discharge capacity data. The method is as follows: First, record the charging capacity or discharging capacity once per second in a complete charge and discharge cycle to obtain the charge and discharge capacity data of one cycle; then, sample the charge and discharge capacity data of one cycle equally 60 times to obtain the charge and discharge capacity sequence of one cycle; then, generate a sample from the charge and discharge capacity sequence of every 10 consecutive cycles until the remaining service life of the supercapacitor first drops to the failure threshold; therefore, the size of a single sample is 10×60. Step (2): Process the collected samples using the exponential moving average method, the multi-curve graph method, and the multi-time series integral processing method respectively; The smoothed value at time t in the exponential moving average method is: p t =ω×x t +(1-ω)×p t-1 (1) Where t is the index number at time p t p is the smoothed value at time t. t-1 Let ω be the smoothed value at time t-1, and let x be the value of x. t The decay weight, (1-ω) is p t-1 The decay weight, x t The true value at time t; The specific content of the multi-curve graph method is as follows: 60 sampled data points from each complete charge-discharge cycle are plotted into curves, with each curve containing 60 data points; one sample contains 10 curves. The specific content of the multi-time series integration processing method is as follows: calculate the maximum value, minimum value, average value, variance and average difference of 60 sampled data in each complete charge-discharge cycle data to obtain a multi-time series integration sequence; generate a multi-time series integration sample every 10 consecutive cycles of the multi-time series integration sequence, and each multi-time series integration sample has a size of 10×5; The variance is calculated as follows: Where θ is the variance, m is the total number of variables, and m-1 is the number of variables excluding y. k The total number of variables, k is the index number of the data, y k As variables, The average value of the variable. For variable y k Compared with the average The difference, For variable y k Compared with the average The square of the difference, where ∑ is the summation symbol; The method for calculating the average difference is as follows: Where ψ is the average difference. For variable y k Compared with the average The absolute value of the difference, where ∑ is the summation symbol; Step (3): Input the samples processed using the exponential moving average method into the open convolutional network for feature extraction; input the samples processed using the multi-curve graph method into the graph processing phantom network for feature extraction; The structure of the open-ended convolutional network model is as follows: enter; The number of access channels is 96, and the stride is 4×4 convolution. Then connect 3 more initial convolutional modules; The number of access channels is 192, and the stride is 2×2 convolution. Then connect 3 more initial convolutional modules; The number of access channels is 384, and the stride is 2×2 convolution. Then connect 9 more initial convolutional modules; The number of access channels is 768, and the stride is 2×2 convolution. Then connect 3 more initial convolutional modules; Then add another 1×1 global average pooling; Then connect two more fully connected layers; Add another probability distribution layer; Output; The structure of the opening convolutional module is as follows: enter; Divided into 2 branches; Branch 1 is input that has not undergone any processing; Branch 2 connects to a separate layer; It is further divided into 4 branches; Branch 1 is processed without any intervention; Branch 2 connects to an 11×1 depthwise separable convolution; Branch 3 connects to a 1×11 depthwise separable convolution; Branch 4 connects to a 3×3 depthwise separable convolution; Then connect branch 1, branch 2, branch 3 and branch 4 to a feature fusion layer; then connect to a normalization layer; Connect another multilayer sensor; Then add the results from branch 1 and the multilayer perceptron output together; Output; The structure of the graphics processing phantom network is as follows: enter; A 3×3 convolution with 16 access channels; Then connect a graphics processing phantom residual module with 24 channels; Divided into 2 branches; Branch 1 is a graphics processing phantom residual module with 24 access channels; Branch 2 is a 1×1 convolution with 24 access channels; Then perform feature fusion processing on branch 1 and branch 2 once more; Then connect a graphics processing phantom residual module with 40 channels; Divided into 2 branches; Branch 1 is a graphics processing phantom residual module with 40 access channels; Branch 2 is a 1×1 convolution with 40 access channels; Then perform feature fusion processing on branch 1 and branch 2 once more; Then connect to the graphics processing phantom residual module with 80 channels; Divided into 2 branches; Branch 1 connects to 5 cascaded graphics processing phantom residual modules with 80 channels; Branch 2 connects to a 1×1 convolution with 80 channels. Then perform feature fusion processing on branch 1 and branch 2 once more; Divided into 2 branches; Branch 1 connects to 5 cascaded graphics processing phantom residual modules with a total of 160 channels; Branch 2 is a 1×1 convolution with 160 access channels; Then perform feature fusion processing on branch 1 and branch 2 once more; Reintegrate a 1×1 convolution with 960 channels; Then add another 1×1 global average pooling; Then connect two more fully connected layers; Add another probability distribution layer; Output; The structure of the graphics processing phantom residual module is as follows: enter; Divided into 2 branches; Branch 1 is input that has not undergone any processing; Branch 2 involves connecting a 1×1 convolution; Then add another 3×3 depthwise separable convolution; Then add another 1×1 convolution; Then add branch 1 and branch 2 together; Output; The classification result in the probability distribution layer is z. i The probability is: Where i is the index number of the classification category, z i For the classification result, S(z) i The classification result is z. i The probability, For z i An exponent with base e, where n is the total number of categories; Step (4): Input the features extracted by the initial convolutional network, the features extracted by the image processing phantom network, and the multi-temporal integral samples after multi-temporal integral processing into the deep fully connected layer network to optimize the classification prediction results and obtain the classification prediction results of the remaining service life of the supercapacitor. The specific structure of the deep fully connected layer network is as follows: enter; Connect to a fully connected layer with 600 channels; Then connect another fully connected layer with 300 channels; Then connect another fully connected layer with 150 channels; Then connect another fully connected layer with 100 channels; Then connect another fully connected layer with 50 channels; Add another probability distribution layer; Add one more category layer; Output.

Images