A super capacitor residual life classification method with multi-fractional integral attention convolution

Abstract

The application provides a super capacitor residual life classification method based on multi-fractional integral attention convolution, which combines a multi-fractional integral method with a transform convolution network and is used for super capacitor residual service life prediction. Firstly, the multi-fractional integral method in the method is used for processing super capacitor capacity data, capturing time sequence dependence in the super capacitor capacity data, and inputting the data processed by the multi-fractional integral into an attention mechanism, thereby enhancing the long-distance important feature extraction capability of the transform convolution network. Secondly, the transform convolution network in the method is used for mining deep features in the super capacitor capacity data. The super capacitor residual life classification method based on multi-fractional integral attention convolution can capture time sequence dependence in super capacitor charging and discharging capacity data, enhance feature extraction capability, and improve super capacitor residual service life prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power systems, electric vehicles, new energy, big data, artificial intelligence and super capacitors, and relates to a super capacitor residual life classification method combining artificial intelligence in the traditional control field, which is suitable for predicting the residual service life of super capacitors in power grids or electric vehicles. BACKGROUND

[0002] When predicting the residual service life of super capacitors, the charge-discharge capacity data of super capacitors is widely used as an aging parameter for predicting the residual service life of super capacitors due to its continuous degradation advantage. In the early stage of degradation of super capacitors, the charge-discharge capacity data sampled in the charge-discharge cycle experiment does not change significantly or even does not change at all in adjacent cycles, resulting in prediction errors. At the same time, when the super capacitor is about to fail, accurate prediction is needed, otherwise the prediction error will cause the failed super capacitor to not be replaced in time, resulting in losses to the system.

[0003] In addition, the charge-discharge capacity data sampled in the super capacitor charge-discharge cycle experiment contains a lot of time sequence information, and capturing the time sequence features in the data can improve the prediction accuracy. The existing network for processing time series, long short-term memory network, is widely used. Although the long short-term memory network solves the problems of gradient disappearance and gradient explosion, it is still very difficult to deal with long sequences. The long short-term memory network consumes a lot of computing resources and takes a long time to calculate as the network depth increases. At the same time, the long short-term memory network is not as good as the convolutional neural network in extracting deep features. Although the existing method uses a combination of convolutional neural networks and long short-term memory networks, the method of combining the two networks increases the complexity of the network model and occupies a large amount of memory resources. In addition, the method of combining the two networks requires a large number of experiments to adjust the parameters of the network to obtain the expected effect.

[0004] Therefore, a super capacitor residual life classification method with multi-fractional integral attention convolution is proposed to solve the problem of prediction error caused by insignificant changes in capacity data. A simple mathematical method is used instead of a long short-term memory network to capture the time sequence information in the data. At the same time, the multi-fractional discrete integral method described in reference [1] Wu N. Finite difference approximation of time fractional Fokker-Planck equation [D]. Lanzhou University, 2021. DOI: 10.27204 / d.cnki.glzhu.2021.001510 is used to solve the problem that continuous multi-fractional integral is not suitable for sampled super capacitor capacity data, improving the prediction speed and accuracy. SUMMARY

[0005] The application provides a super capacitor residual life classification method based on multi-fractional integral attention convolution, which divides the residual service life of the super capacitor into life intervals for classification and prediction, reduces the prediction error caused by the insignificant change of the capacity data in the early degradation of the super capacitor, and reduces the large-scale loss of the power system caused by the prediction error when predicting the residual service life of the super capacitor about to fail in the late degradation of the super capacitor; the multi-fractional integral method is used to process the super capacitor charge and discharge capacity data sampled in the experiment, to capture the time sequence dependence in the super capacitor capacity data, and the data processed by the multi-fractional integral method is used as the input of the attention mechanism to enhance the long-distance important feature extraction capability of the change convolution network; the change convolution network is used to mine the deep features in the data and enhance the classification capability.

[0006] The steps in use are:

[0007] Step (1): Obtain the super capacitor charge and discharge capacity data, the obtaining method is: first, record the charge capacity or discharge capacity every second in a complete charge and discharge cycle to obtain a cycle charge and discharge capacity data; then, equivalently sample n times in a cycle charge and discharge capacity data to obtain a cycle charge and discharge capacity sequence, and the single cycle charge and discharge capacity sequence is described as:

[0008] A k =[a 1,k a 2,k ...a n,k ] (1)

[0009] Wherein, A k is the sampled charge and discharge capacity sequence in the kth cycle, k is the index sequence number of the cycle number, a 1,k , a 2,k and a n,k are the 1st sampling value, the 2nd sampling value and the n th sampling value in the kth cycle, 1, 2 and n are the sampling times; then, a sampling sample is generated every continuous m cycles of charge and discharge capacity sequence until the residual service life of the super capacitor first decreases to the failure threshold; a single sampling sample is described as:

[0010]

[0011] Wherein, ω i is the i th sampling sample, i is the index sequence number of the sample number, a 1,k+1 , a 2,k+1 and a n,k+1 are the 1st sampling value, the 2nd sampling value and the n th sampling value in the k+1th cycle, a 1,k+m-1 , a 2,k+m-1 and a n,k+m-1These are the first, second, and nth sampled values ​​in the (k+m-1)th iteration, respectively, where m is the index number of the iteration.

[0012] Step (2): Process the supercapacitor charge / discharge capacity data samples sampled in step (1) using a multi-fractional integration method. The multi-fractional integration method is as follows: For a single sample ω i Each charge / discharge capacity sequence in the sequence is λ1, λ2, ... λ h Processing the integrals of order λ, where λ1, λ2 and λ h All are fractional orders, and h is the total number of fractional-order calculations performed, resulting in multi-fractional-order integral feature samples θ after multi-fractional-order integral processing. i :

[0013]

[0014] Where ψ is the fractional discrete integral operation function, ψ(A k ,λ1)ψ(A k ,λ2) and ψ(A k ,λ h ) represent λ1, λ2, and λ3, respectively, the charge / discharge capacity sequence in the k-th cycle. h Order integral, ψ(A) k+1 ,λ1)ψ(A k+1 ,λ2) and ψ(A k+1 ,λ h ) represent λ1, λ2, and λ3, respectively, the charge / discharge capacity sequence in the (k+1)th cycle. h Order integral, ψ(A) k+m-1 ,λ1)ψ(A k+m-1 ,λ2) and ψ(A k+m-1 ,λ h ) represent λ1, λ2, and λ3, respectively, the charge / discharge capacity sequence in the (k+m-1)th cycle. h Fractional integrals; in multi-fractional integral feature samples, each element is calculated according to the following fractional discrete integral operation function:

[0015]

[0016] Where τ is the sampling interval, Γ() is the gamma function, and Γ(2-λ) h ) indicates that the independent variable is 2-λ h The gamma function value, and All are λ corresponding to the sampling points h Order coefficient, a n,k a z,k and a 1,kare sampling points at the kth cycle, z is an index number of sampling times, and ∑ is a summation symbol; the calculation method of the gamma function is:

[0017]

[0018] where e is a base number of a natural logarithm function, and t is an independent variable; the λ of the sampling point is h The calculation formula of the order coefficient is:

[0019]

[0020] where l is an index number of the sampling point;

[0021] Step (3): converting the sampling data sample ω i into a sampling picture sample, the size of a single sampling picture sample is m x n x 3, then the size of the sampling picture sample is adjusted to 224 x 224 x 3 to match the input size of the network model, then the sampling picture sample after the size adjustment is input into the multi-fractional integral attention convolution network for classification prediction; meanwhile, the multi-fractional integral feature data sample θ i is converted into a picture sample, the size of a single multi-fractional integral picture sample is h x n x 3, then the multi-fractional integral processing picture sample is input into the attention module to enhance the important feature extraction ability of the network in a long distance; and an output prediction result is output.

[0022] The structure of the multi-fractional integral attention convolution network model is:

[0023] The input sampling sample;

[0024] 1 A 2x2 convolution with a step of 2 and a channel number of 72 is accessed;

[0025] Then, 3 transform convolution modules with a channel number of 72 are accessed;

[0026] Then, 1 A 2x2 convolution with a step of 2 and a channel number of 144 is accessed;

[0027] Then, the multi-fractional integral feature sample with a size of 28x28x144 after the size adjustment is added;

[0028] Then, 3 transform convolution modules with a channel number of 144 are accessed;

[0029] Then, 1 A 2x2 convolution with a step of 2 and a channel number of 288 is accessed;

[0030] Then, the multi-fractional integral feature sample with a size of 14x14x288 after the size adjustment is added;

[0031] Then, 12 transform convolution modules with a channel number of 144 are accessed;

[0032] Again access 1 step 2 channel number 576 2x2 convolution;

[0033] Again add the multi-fractional integral feature sample resized to 7x7x576;

[0034] Again access 3 transform convolution modules with 576 channels;

[0035] Again access 1 global average pooling;

[0036] Again access 2 fully connected layers;

[0037] Again access 1 classification layer;

[0038] Output;

[0039] The structure of the transform convolution module is:

[0040] Input;

[0041] Divided into 3 branches;

[0042] Branch 1 is input without any processing;

[0043] Branch 2 is access 1x1 convolution;

[0044] Branch 3 is access 1x1 convolution;

[0045] Again access 1 3x3 depth separable convolution;

[0046] Again multiply branch 2 and branch 3;

[0047] Again access 1x1 convolution;

[0048] Again add branch 1;

[0049] Output.

[0050] The present application has the following advantages and effects relative to the prior art:

[0051] (1) The present application divides the remaining service life of the super capacitor into life intervals for classification prediction. In the early stage of degradation of the super capacitor, the life interval can reduce the prediction error caused by the insignificant change of the capacity data. In the late stage of degradation of the super capacitor, the life interval contains multiple life values, which can reduce the large-scale loss of the power system caused by prediction error when predicting the remaining service life of the super capacitor about to fail.

[0052] (2) The application uses multi-fractional integral processing to process the sampled charge and discharge capacity data, the multi-fractional integral processing method can capture the time sequence dependence in the sampled capacity data, and uses a cheap mathematical operation method to replace the long short-term memory network to capture the time sequence dependence in the sampled capacity data, thereby reducing the operation time and improving the prediction speed; the transformed convolution network model is used to mine deep features in the data, the transformed convolution network model is a kind of convolution neural network model, has strong feature extraction capability, good classification performance, and high prediction accuracy.

[0053] (3) The application uses the data processed by the multi-fractional integral as the input of the attention mechanism, introduces the attention mechanism into the transformed convolution network model, enhances the important feature extraction capability of the network under long distance, and improves the prediction accuracy.

[0054] (4) The application first combines the idea of proportional integral derivative controller in the control field with the idea of fractional order controller and applies it to the prediction and classification field. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is the structure diagram of the multi-fractional integral attention convolution network model in the application.

[0056] Figure 2 is the structure diagram of the transformed convolution module in the application. DETAILED DESCRIPTION

[0057] The application provides a supercapacitor residual life classification method based on multi-fractional integral attention convolution, and the application is described in detail in combination with the drawings as follows:

[0058] Figure 1 is the structure diagram of the multi-fractional integral attention convolution network model in the application. The structure of the multi-fractional integral attention convolution network model is as follows:

[0059] input sampling samples;

[0060] access 1 2x2 convolution with a step length of 2 and a channel number of 72;

[0061] access 3 transformed convolution modules with a channel number of 72;

[0062] access 1 2x2 convolution with a step length of 2 and a channel number of 144;

[0063] add the multi-fractional integral feature samples with a size of 28x28x144 after adjustment;

[0064] access 3 transformed convolution modules with a channel number of 144;

[0065] access 1 2x2 convolution with a step length of 2 and a channel number of 288;

[0066] Again add the multi-fractional integral feature samples resized to 14x14x288;

[0067] Again access 12 transform convolution modules with 144 channels;

[0068] Again access 1 2x2 convolution with 2 channels and 2 steps;

[0069] Again add the multi-fractional integral feature samples resized to 7x7x576;

[0070] Again access 3 transform convolution modules with 576 channels;

[0071] Again access 1 global average pooling;

[0072] Again access 2 fully connected layers;

[0073] Again access 1 classification layer;

[0074] Output;

[0075] Figure 2 is the structure diagram of the transform convolution module in the method of the application. The structure of the transform convolution module is:

[0076] Input;

[0077] Divided into 3 branches;

[0078] Branch 1 is input without any processing;

[0079] Branch 2 is access to 1x1 convolution;

[0080] Branch 3 is access to 1x1 convolution;

[0081] Again access 1 3x3 depth separable convolution;

[0082] Again multiply branch 2 and branch 3;

[0083] Again access 1x1 convolution;

[0084] Again add branch 1;

[0085] Output.

[0086] The above only describes the preferred embodiments of the application, and does not limit the patent scope of the application, and any equivalent structure or equivalent flow transformation using the content of the application specification and drawings, or directly or indirectly applied to other related technical fields, are also included in the patent protection scope of the application.

Claims

1. A method for supercapacitor remaining life classification with multi-fractional integral attention convolution, characterized in that, The remaining service life of the super capacitor is divided into life intervals for classified prediction, so as to reduce the prediction error caused by the insignificant change of the capacity data in the early degradation of the super capacitor, and reduce the large-scale loss of the power system caused by the prediction error when predicting the remaining service life of the super capacitor about to fail in the late degradation of the super capacitor; the multi-fractional integral method is used to process the super capacitor charging and discharging capacity data sampled in the experiment, so as to capture the time sequence dependence in the super capacitor capacity data, and the data processed by the multi-fractional integral is used as the input of the attention mechanism, so as to enhance the long-distance important feature extraction capability of the change convolution network; the change convolution network is used to mine the deep features in the data, so as to enhance the classification capability; The steps in use are as follows: Step (1): Obtain the super capacitor charging and discharging capacity data, and the obtaining method is as follows: first, record the charging capacity or discharging capacity every second in a complete charging and discharging cycle to obtain a cycle charging and discharging capacity data; then, equivalently sample n times in a cycle charging and discharging capacity data to obtain a cycle charging and discharging capacity sequence, and a single cycle charging and discharging capacity sequence is described as: A k = [a 1,k a 2,k ... a n,k ] (1) wherein A k is the sampled charge-discharge capacity sequence at the kth cycle, k is the index number of cycle number, a 1,k , a 2,k and a n,k are the 1st, 2nd and nth sampling values at the kth cycle, 1, 2 and n are all the sampling number; then a sampling sample is generated for every continuous m cycles of charge-discharge capacity sequence until the remaining service life of the super capacitor first drops to the failure threshold; a single sampling sample is described as: wherein ω i is the i-th sampling sample, i is the sample number index, a 1,k+1 , a 2,k+1 and a n,k+1 are the 1st, 2nd and nth sampling values at the k+1th cycle, a 1,k+m-1 , a 2,k+m-1 and a n,k+m-1 are the 1st, 2nd and nth sampling values at the k+m-1th cycle, and m is the index of the cycle number. Step (2): processing the super capacitor charge-discharge capacity data samples sampled in step (1) using a multi-fractional integral method, the content of the multi-fractional integral processing method being: performing λ1, λ2, … λ i order integral processing on each of the single sampling samples ω h , wherein λ1, λ2, and λ h are all fractional orders, h is the total number of fractional order calculations, and a multi-fractional integral feature sample θ i after multi-fractional integral processing is obtained: wherein ψ is a fractional order discrete integral operation function, ψ(A k , λ1), ψ(A k , λ2) and ψ(A k , λ h ) are λ1, λ2 and λ h order integrals of the charge and discharge capacity sequence in the kth cycle respectively, ψ(A k+1 , λ1), ψ(A k+1 , λ2) and ψ(A k+1 , λ h ) are λ1, λ2 and λ h order integrals of the charge and discharge capacity sequence in the k+1th cycle respectively, ψ(A k+m-1 , λ1), ψ(A k+m-1 , λ2) and ψ(A k+m-1 , λ h ) are λ1, λ2 and λ h order integrals of the charge and discharge capacity sequence in the k+m-1th cycle respectively; in the multi-fractional integral characteristic sample, each element is calculated according to the following fractional order discrete integral operation function: where τ is the sampling interval, Γ() is the gamma function, Γ(2-λ h ) represents the gamma function value of the independent variable 2-λ h and are the λ h order coefficients corresponding to the sampling points, a n,k , a z,k , and a 1,k are the sampling points at the kth cycle, z is the index number of the sampling times, and ∑ is the summation symbol; the calculation method of the gamma function is:​ where e is the base of the natural logarithm function, t is the independent variable; and λ h order coefficients The calculation formula is: Wherein, l is the sampling point index number; Step (3): converting the sampling data sample ω i into a sampling picture sample, the size of a single sampling picture sample being m×n×3, then adjusting the size of the sampling picture sample to 224×224×3 to match the input size of the network model, and then inputting the resized sampling picture sample into the multi-fractional integral attention convolution network for classification prediction; meanwhile, converting the multi-fractional integral feature data sample θ i into a picture sample, the size of a single multi-fractional integral picture sample being h×n×3, then inputting the multi-fractional integral processing picture sample as an input of an attention module to enhance the important feature extraction capability of the network in a long distance; and outputting a prediction result; The structure of the multi-fractional integral attention convolution network model is as follows: Input the sampling sample; Access 1 2×2 convolution with a step of 2 and a channel number of 72; Then access 3 transform convolution modules with a channel number of 72; Then access 1 2×2 convolution with a step of 2 and a channel number of 144; Then add the multi-fractional integral feature sample with the size adjusted to 28×28×144; Then access 3 transform convolution modules with a channel number of 144; Then access 1 2×2 convolution with a step of 2 and a channel number of 288; Then add the multi-fractional integral feature sample with the size adjusted to 14×14×288; Then access 12 transform convolution modules with a channel number of 144; Then access 1 2×2 convolution with a step of 2 and a channel number of 576; Then add the multi-fractional integral feature sample with the size adjusted to 7×7×576; Then access 3 transform convolution modules with a channel number of 576; Then access 1 global average pooling; Then access 2 fully connected layers; Then access 1 classification layer; Output; The structure of the transform convolution module is as follows: Input; Divided into 3 branches; Branch 1 is input without any processing; Branch 2 is accessed 1 1×1 convolution; Branch 3 is accessed 1 1×1 convolution; Then access 1 3×3 depth separable convolution; then multiply branch 2 and branch 3; Then access 1 1×1 convolution; Then add branch 1; Output.

Images