A data preprocessing method based on edge computing for a lithium battery energy storage station

Abstract

The application discloses a kind of lithium battery energy storage station based on edge computing data preprocessing method.It includes the following steps: obtaining wavelet base function is discretized, and discrete wavelet function is obtained;Discrete wavelet transform is carried out to arbitrary input acoustic signal, and approximate value and noise value are decomposed to discrete wavelet;Threshold size is determined according to sample estimation;Wavelet coefficient is reorganized according to threshold, and the inverse transform of wavelet energy spectrum after wavelet coefficient processing is carried out, reconstructs time signal;Denoising evaluation index function is established, for the one that fails to meet, reconstructed signal is returned again decomposition, and second threshold determination is carried out again, and reconstruction is carried out again, if the signal after twice reconstruction still cannot meet, it will be discarded;The reconstructed signal after wavelet denoising and meeting the requirement signal-to-noise ratio evaluation index is extracted by principal component analysis PCA eigenvalue.The beneficial effects of the application are: it can realize denoising dimension reduction processing to obtain eigenvalue.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery energy storage stations, and in particular to a data preprocessing method for lithium battery energy storage stations based on edge computing. BACKGROUND

[0002] With the proposal of the double carbon target and the development of digital traction in the construction of new power systems, electrochemical energy storage stations play a very important role in green energy conversion due to their relatively mature technology and high cost performance. However, as the installed capacity of energy storage stations continues to increase, safety hazards of lithium battery energy storage stations begin to emerge. Therefore, safety monitoring of lithium battery energy storage stations is urgently needed.

[0003] In order to fully perceive the operating state of the lithium battery energy storage station, all operating parameters related to the energy storage station are integrated, including BMS, battery metering device, temperature sensor, acoustic sensor, thermal sensor, gas sensor, infrared detector, smoke detector, video sensor, and ion release sensor. If these sensor data are not processed and directly reported to the intelligent fire control cloud platform, the intelligent fire control cloud platform will be under great load pressure in the case of a large number of energy storage stations in multiple regions. Due to the limitations of bandwidth and computing load, the feedback of the fire control command will be delayed, which is unacceptable for time-sensitive fire control systems.

[0004] The collected multiple sensor signals involve multiple sources and are complex and variable in the time domain, with strong randomness and contingency. Among the characterization parameters related to early warning of battery failure, acoustic, thermal, temperature, and gas are the four most important reference parameters. According to the time dimension, from the occurrence of battery failure thermal runaway to thermal abuse, the four parameters are acoustic -> gas -> temperature -> thermal. Therefore, real-time monitoring and identification of these four parameters directly affect the early warning of battery failure. The present method focuses on processing the collected parameter data of the operating state of the energy storage station on these four signals. Although the four battery failure characterization parameters belong to low-frequency signals in the frequency space distribution, they are strongly disturbed by background interference during the operation of the energy storage station, such as safety valve acoustic, which is easily affected by current noise and human movement noise in the prefabricated cabin, and is also disturbed by Gaussian white noise. SUMMARY

[0005] The present application is to overcome the above-mentioned deficiencies in the prior art and provides a lithium battery energy storage station data preprocessing method based on edge computing that can realize denoising and dimensionality reduction processing to obtain characteristic values.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A data preprocessing method for lithium battery energy storage stations based on edge computing specifically includes the following steps:

[0008] (1) Obtain the wavelet basis functions and discretize them to obtain discrete wavelet functions;

[0009] (2) For any input acoustic signal f(t)∈L 2 (R) performs discrete wavelet transform and decomposes the discrete wavelet into four levels based on the approximation value and the noise value;

[0010] (3) Determine the threshold size based on the sample estimation. The selection and setting of the threshold is the key in the wavelet denoising process. By removing wavelet coefficients that do not meet the set threshold requirements and retaining the remaining coefficients, a good denoising effect can be achieved.

[0011] (4) Based on the threshold, the wavelet coefficients are reconstructed, and the wavelet energy spectrum after wavelet coefficient processing is inversely transformed to reconstruct the time signal;

[0012] (5) Establish a denoising evaluation index function, introduce the signal-to-noise ratio to evaluate the signal denoising performance evaluation index. For denoising effects that fail to meet the requirements, return the reconstructed signal to step (2) for decomposition again, and perform a second threshold judgment before reconstructing. If the reconstructed signal still fails to meet the evaluation index requirements, discard the signal.

[0013] (6) The reconstructed signal after wavelet denoising and meeting the required signal-to-noise ratio evaluation index needs to be extracted by principal component analysis (PCA).

[0014] This invention proposes an improved data preprocessing method suitable for edge computing under resource-constrained conditions. This method uses wavelet denoising and PCA principal component analysis to denoise and reduce the dimensionality of raw data collected by sensors, transforming a dataset with multiple variables into one with fewer new variables, thereby enabling the denoising and dimensionality reduction process to obtain feature values.

[0015] Preferably, in step (1), the specific steps are as follows:

[0016] Setting wavelet transform L 2 (R) represents a signal with finite energy, whose Fourier transform is Ψ(w); when Conditions met:

[0017]

[0018] say The mother wavelet is defined as Ψ(ω), which is the mother wavelet's spectral function, where w represents frequency and t represents time. A wavelet sequence can be obtained through appropriate linear transformation:

[0019]

[0020] In the above formula, a and b are the scaling and translation factors of the mother wavelet respectively; the discrete wavelet function is obtained by discretizing the above formula as follows:

[0021]

[0022] where j and k are positive integers.

[0023] As preferred, in step (2), the discrete wavelet transform is as follows:

[0024]

[0025] The discrete wavelet is decomposed into four layers according to the approximation value and the noise value, as shown in the following formula:

[0026] w f (j, k) = C A4 + C D4 + C D3 + C D2 + C D1

[0027] In the formula, C Ai is the approximation part of the wavelet decomposition, C Di represents the part where the noise is located, and the wavelet function layering refers to decomposing the part where the noise is located into four specific parts, i.e., the low-frequency part and the high-frequency parts in three directions (horizontal, vertical, and diagonal).

[0028] As preferred, in step (3), according to the background noise characteristics of the battery prefabricated cabin and considering the computing capacity of the edge side gateway, a fixed maximum minimum criterion method is used to determine the threshold value, so as to reduce the computing overhead and improve the computing efficiency; the maximum minimum criterion threshold method is shown in the following formula:

[0029]

[0030] In the above formula, θ is the standard deviation of the noise, and n is the number of wavelet coefficients at each scale.

[0031] As preferred, in step (4), the wavelet transform of a given signal is to expand the signal according to a certain wavelet function cluster, i.e., to express the signal as a linear combination of a series of wavelet functions with different scales and different shifts, where the coefficient of each term is called the wavelet coefficient, and specifically, the wavelet coefficients greater than the set threshold value are retained, and the remaining ones are set to zero, as shown in the following formula:

[0032]

[0033] wherein x is the wavelet coefficient, and T is the threshold value;

[0034] The inverse transform of the wavelet energy spectrum after processing the wavelet coefficient is performed to reconstruct the time signal:

[0035]

[0036] wherein a = 2 j , b = 2 j k, j, and k are positive integers.

[0037] As preferred, in step (5), specifically, the relationship between the reconstructed time signal in step (4) and the evaluation index function is shown in the following formula, that is, the square difference is summed and averaged, and then logarithmic operation is performed, and the evaluation index function is as follows:

[0038]

[0039] wherein n is the number of wavelet coefficients at each scale, and f(n) refers to the reconstructed time signal with n as the variable.

[0040] As preferred, in step (6), the specific steps of the principal component analysis method are as follows:

[0041] (61) Assuming that the input data source has n samples and p indexes, an n x p sample matrix x can be constructed:

[0042]

[0043] (62) The average value of the characteristic vector is calculated:

[0044]

[0045] (63) The covariance matrix S of the characteristic vector is calculated, that is:

[0046]

[0047] (64) The eigenvalue and eigenvector of the normalized sample covariance are calculated:

[0048] Eigenvalue: λ1, λ2,..., λ n , eigenvector: v1, v2,..., v n

[0049] (65) The contribution rate η i and the cumulative contribution rate β k of each component are calculated:

[0050]

[0051]

[0052] (66) Construct the feature vector mapping matrix.

[0053] Preferably, in step (66), the specific steps are as follows:

[0054] Sort the principal components by their contribution rates, select the p largest principal components that meet the requirements, and then use their corresponding p eigenvectors to form an eigenvector matrix V as a mapping matrix. Transform the denoised and reconstructed signal into the feature space, i.e.:

[0055] Y = V T X

[0056] Where Y is the projected data matrix, i.e., the principal component matrix of the original signal after PCA processing; V is the eigenvector matrix, V = [v1, v2, ..., v p ], V T Let X be the transpose of the matrix, and X be the average of the eigenvectors.

[0057] The beneficial effects of this invention are: under limited resource conditions, the original data collected by the sensor is denoised and dimensionality reduced based on wavelet denoising and PCA principal component analysis, transforming a dataset with multiple variables into a dataset with fewer new variables, thereby enabling the denoising and dimensionality reduction processing to obtain feature values. Attached Figure Description

[0058] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0059] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.

[0060] like Figure 1 In the embodiments described above, a data preprocessing method for a lithium battery energy storage station based on edge computing specifically includes the following steps:

[0061] (1) Obtain the wavelet basis functions and discretize them to obtain discrete wavelet functions; specifically:

[0062] Setting wavelet transform L 2 (R) represents a signal with finite energy, whose Fourier transform is Ψ(w); when Conditions met:

[0063]

[0064] say The mother wavelet is defined as Ψ(ω), which is the mother wavelet's spectral function, where w represents frequency and t represents time. A wavelet sequence can be obtained by proper linear transformation:

[0065]

[0066] In the above formula, a and b are the scaling and translation factors of the mother wavelet respectively; the discrete wavelet function is obtained by discretizing the above formula as follows:

[0067]

[0068] where j and k are positive integers.

[0069] (2) For any input acoustic signal f(t) ∈ L 2 (R), a discrete wavelet transform is performed, and the discrete wavelets are decomposed into four layers according to the approximation value and the noise value; the discrete wavelet transform is as follows:

[0070]

[0071] The discrete wavelets are decomposed into four layers according to the approximation value and the noise value, as shown in the following formula:

[0072] w f (j, k) = C A4 + C D4 + C D3 + C D2 + C D1

[0073] where C Ai is the approximation part of the wavelet decomposition, and C Di represents the part where the noise is located; the wavelet function layering refers to decomposing the part where the noise is located into four specific parts, i.e., a low-frequency part and three high-frequency parts in the horizontal, vertical, and diagonal directions, which is one of the steps of the standard wavelet transform.

[0074] (3) The threshold size is determined according to the sample estimation; the selection and setting of the threshold are the key in the wavelet denoising process; by removing the wavelet coefficients that do not meet the set threshold requirements and retaining the remaining coefficients, a good denoising effect is achieved; according to the background noise characteristics of the battery prefabricated cabin and considering the computing capacity of the edge gateway, a fixed maximum-minimum criterion method is used to determine the threshold to reduce the computing overhead and improve the computing efficiency; the maximum-minimum criterion threshold method is as follows:

[0075]

[0076] In the above formula, θ is the standard deviation of the noise, and n is the number of wavelet coefficients at each scale.

[0077] (4) According to the threshold recombination wavelet coefficient, the wavelet energy spectrum after processing the wavelet coefficient is inversely transformed, and the time signal is reconstructed; the wavelet transform of a given signal is to expand the signal according to a certain wavelet function cluster, that is, to express the signal as a linear combination of a series of wavelet functions with different scales and different time shifts, wherein the coefficient of each term is called a wavelet coefficient, and the specific is: the wavelet coefficient greater than the set threshold is left, and the rest is zero processing, as shown in the following formula:

[0078]

[0079] Wherein, is the wavelet coefficient set after determination, x is the wavelet coefficient, and T is the threshold;

[0080] The wavelet energy spectrum after processing the wavelet coefficient is inversely transformed, and the time signal is reconstructed:

[0081]

[0082] Wherein a = 2 j , b = 2 j k, j, k are positive integers. Note: j, k in steps (1) to (4) above have no actual meaning, which represents the serial number parameter.

[0083] (5) Establish a denoising evaluation index function, introduce the signal-to-noise ratio to evaluate the signal denoising performance evaluation index, for the denoising effect that fails to meet the requirements, the reconstructed signal is returned to step (2) again for decomposition, and the second threshold determination is carried out before reconstruction, and the signal after the second reconstruction cannot meet the evaluation index requirements, then the signal is discarded; the relationship between the reconstructed time signal in step (4) and the evaluation index function is shown in the following formula, that is: the average number of square difference summation is logarithmically operated, and the evaluation index function is as follows:

[0084]

[0085] Wherein: n is the number of wavelet coefficients at each scale, and f(n) refers to the reconstructed time signal with n as the variable.

[0086] (6) The reconstructed signal after wavelet denoising and meeting the signal-to-noise ratio evaluation index needs to be subjected to principal component analysis PCA for eigenvalue extraction; the specific steps of principal component analysis are:

[0087] (61) Assuming that the input data source has n samples and p indexes (because wavelet denoising and PCA eigenvalue extraction are two relatively independent parts, here n and p are only assumed parameters, not corresponding to the previous steps), an n×p sample matrix x can be constructed:

[0088]

[0089] (62) Calculate the mean of the eigenvectors:

[0090]

[0091] (63) Obtain the covariance matrix S of the eigenvectors, that is:

[0092]

[0093] (64) Calculate the eigenvalues and eigenvectors of the normalized sample covariance:

[0094] Eigenvalues: λ1, λ2,..., λ n , Eigenvectors: v1, v2,..., v n

[0095] (65) Calculate the contribution rate η i and cumulative contribution rate β k of each component:

[0096]

[0097]

[0098] (66) Construct the eigenvector mapping matrix, specifically:

[0099] Sort the principal components by contribution rate, select the largest p principal components that meet the requirements (cumulative contribution rate higher than 85%), then form the eigenvector matrix V corresponding to the p principal components as the mapping matrix, and transform the denoised reconstructed signal to the feature space, that is:

[0100] Y = V T X

[0101] Where Y is the projected data matrix, that is, the principal component matrix of the original signal after PCA processing; V is the eigenvector matrix, V = [v1, v2,..., v p ], V T is the transpose matrix, and X is the mean of the eigenvectors.

[0102] In the current early warning research of lithium batteries, some domestic research literature methods are based on the temperature of energy storage batteries for early warning, the internal temperature of the battery is analyzed according to the battery thermal model and electrochemical impedance spectrum, and a corresponding model is established to predict the overall temperature distribution of the battery. The time difference of the temperature of the electric field from the inside to the surface is several seconds, and the calculation process of the thermal model is relatively complex. There are also methods of detecting the content of specific gas in the battery cabin to distinguish the thermal runaway of the battery, sampling and detecting the gas after thermal decomposition, and establishing a multi-level early warning system for energy storage power stations. Due to the slow diffusion of gas caused by the blocking of the battery rack, the early warning sensitivity is poor, and the location information of the faulty battery cannot be directly obtained. In the system of detecting and analyzing acoustic signals, an acoustic emission measurement system is applied, and the correlation between the battery fault signal and the acoustic emission spectrum is statistically analyzed, the acoustic signal after denoising is extracted, and the recognition classifier is constructed according to the extracted characteristic value. However, in the process of extracting the lithium battery fault signal, the internal information of the signal cannot be obtained, and it is also easy to be disturbed by noise.

[0103] Although the four battery fault characteristic parameters belong to low-frequency signals in the frequency space distribution, they have strong background interference during the operation of the energy storage station, such as safety valve acoustic waves, which are easily affected by the current noise and human walking noise in the prefabricated cabin, and are also disturbed by Gaussian white noise. Therefore, before entering the characteristic value extraction, a certain denoising treatment is also needed. The present application proposes an improved data preprocessing method suitable for edge computing under limited resource conditions. This method is based on wavelet denoising and PCA principal component analysis to denoise and reduce the dimension of the original data collected by the sensor, and to convert the data set of multiple variables into a data set of fewer new variables, so as to realize denoising and dimensionality reduction to obtain characteristic values, and at the same time, the principal components are normalized and reported to the intelligent fire control cloud platform for training and dynamic recognition and diagnosis.

Claims

1. A lithium battery energy storage station edge computing-based data preprocessing method, characterized in that, Specifically comprising the following steps: (1) obtaining a wavelet base function, discretizing the same to obtain a discrete wavelet function; (2) for any input acoustic signal f(t) e L 2 (R) to perform a discrete wavelet transform and to perform a four-level decomposition of the discrete wavelet according to approximation values and noise values; (3) determining a threshold size according to sample estimation, the selection and setting of the threshold are key in the wavelet denoising process, by removing wavelet coefficients that do not meet the set threshold requirement and retaining the remaining coefficients, a good denoising effect is achieved; (4) recombining wavelet coefficients according to the threshold, inversely transforming the wavelet energy spectrum after wavelet coefficient processing, and reconstructing a time signal; (5) establishing a denoising evaluation index function, introducing a signal-to-noise ratio to evaluate the signal denoising performance evaluation index, for the denoising effect that fails to meet the requirement, the reconstructed signal is returned to step (2) again for decomposition, second threshold determination and reconstruction, if the twice reconstructed signal still fails to meet the evaluation index requirement, the signal is discarded; (6) the reconstructed signal after wavelet denoising and meeting the required signal-to-noise ratio evaluation index needs to be subjected to principal component analysis PCA for eigenvalue extraction.

2. The data preprocessing method based on edge computing for a lithium battery energy storage station according to claim 1, characterized in that, In step (1), specifically: Setting up a wavelet transform L 2 (R) denotes a signal with finite energy whose Fourier transform is Ψ(w); when the condition is met: say The mother wavelet is defined as Ψ(ω), which is the mother wavelet's spectral function, where w represents frequency and t represents time. A wavelet sequence can be obtained through appropriate linear transformation: In the formula, a and b are the scaling and translation factors of the mother wavelet; the discrete wavelet function obtained by discretizing the formula is: Wherein, j, k are positive integers.

3. The data preprocessing method for a lithium battery energy storage station based on edge computing according to claim 2, characterized in that, In step (2), the discrete wavelet transform is: The discrete wavelet is decomposed into four layers according to the approximate value and the noise value, as shown in the following formula: w f (j, k) = C A4 +C D4 +C D3 +C D2 +C D1 where C Ai is the approximation part of the wavelet decomposition, C Di represents the part where the noise is located, and the wavelet function decomposition levels refer to the decomposition of the part where the noise is located into four specific parts, namely the low frequency part and the high frequency parts in three directions (horizontal, vertical, diagonal).

4. The data preprocessing method for a lithium battery energy storage station based on edge computing according to claim 3, characterized in that, In step (3), according to the background noise characteristics of the battery prefabricated cabin and considering the computing capacity of the edge gateway, a fixed maximum minimum criterion method is used to determine the threshold, so as to reduce the computing overhead and improve the computing efficiency; the maximum minimum criterion threshold method is as follows In the formula, θ is the standard deviation of the noise, and n is the number of wavelet coefficients at each scale.

5. The data preprocessing method for a lithium battery energy storage station based on edge computing according to claim 4, characterized in that, In step (4), wavelet transform of a given signal is to expand the signal according to a certain wavelet function cluster, that is, to express the signal as a linear combination of wavelet functions of different scales and different time shifts, wherein each coefficient is called a wavelet coefficient, and the specific process is as follows: the wavelet coefficients greater than the set threshold are retained, and the remaining coefficients are set to zero, as shown in the following formula: wherein is the set of wavelet coefficients after thresholding, x is a wavelet coefficient, and T is a threshold value. Inverse transform the wavelet energy spectrum after wavelet coefficient processing to reconstruct the time signal: where a = 2 j b = 2 j k, j, k are positive integers.

6. The data preprocessing method for a lithium battery energy storage station based on edge computing according to claim 5, characterized in that, In step (5), the relationship between the reconstructed time signal in step (4) and the evaluation index function is shown in the following formula, that is: after square difference summation and averaging, logarithmic operation is performed, and the evaluation index function is as follows: Wherein: n is the number of wavelet coefficients at each scale, and f(n) refers to the reconstructed time signal with n as the variable.

7. The data preprocessing method for a lithium battery energy storage station based on edge computing according to claim 1 or 2 or 3 or 4 or 5 or 6, characterized in that, In step (6), the specific steps of principal component analysis are as follows: (61) assuming that the input data source has n samples and p indexes, a sample matrix x of size n x p can be constructed: (62) calculating the average value of the characteristic vector: (63) calculating the covariance matrix S of the characteristic vector, that is: (64) calculating the eigenvalue and eigenvector of the normalized sample covariance: Eigenvalues: λ1, λ2,..., λ n , eigenvectors: v1, v2,..., v n (65) calculating the contribution rate η of each component i and the cumulative contribution rate β k : (66) constructing a characteristic vector mapping matrix.

8. The data preprocessing method for a lithium battery energy storage station based on edge computing according to claim 7, characterized in that, In step (66), specifically: The principal components are sorted according to the contribution rate, the largest p principal components meeting the requirements are selected, then the corresponding p eigenvectors are combined to form an eigenvector matrix V as a mapping matrix, and the denoised reconstructed signal is transformed to the feature space, i.e. Y = V T X Wherein, Y is the data matrix after projection, that is, the principal component matrix of the original signal after PCA processing; V is the eigenvector matrix, V = [v1, v2,..., vn], V p ] is the transpose matrix, and X is the average value of the eigenvector. T ​

Images