Energy storage battery inconsistency diagnosis method based on comparative learning and multi-modal weighting

Through comparative learning and multimodal weighting methods, multi-view features are constructed and multi-head self-attention network is used, which solves the problem of insufficient accuracy of battery cell incompatibility detection in energy storage battery packs and achieves efficient incompatibility incompatibility in the battery pack Diagnosis and management.

CN118914860APending Publication Date: 2024-11-08JIANGSU NENGTAN SMART TECHNOLOGY CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411024238.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The inconsistency between battery cells in energy storage battery packs is difficult to be effectively discovered and solved by the existing technology, resulting in insufficient operation and maintenance management complexity and diagnostic accuracy.

Method used

Using a method based on contrast learning and multimodal weighting, multi-view features are constructed by collecting voltage, current and temperature data, and feature representation is performed using a multi-head self-attention network. Combining genetic algorithm and KL divergence calculation, unsupervised battery singles are realized. Detection of body inconsistency.

Benefits of technology

It improves the accuracy and robustness of battery cell inconsistency diagnosis, can detect potential inconsistency problems without large amounts of labeled data, and improves the reliability of the diagnostic system through multimodal weighting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118914860A_ABST
    Figure CN118914860A_ABST
Patent Text Reader

Abstract

The invention discloses an energy storage battery inconsistency diagnosis method based on comparative learning and multi-modal weighting, and the method comprises the steps: 1, collecting voltage, current and temperature data from a battery module, and carrying out the dimension alignment processing; step 2, constructing a maximum target function of difference between single batteries in the module to obtain multi-view features of the battery module; 3, learning feature representations in different views by using a multi-head self-attention network; 4, defining a loss function of model training; and 5, detecting the inconsistency in the battery module. According to the method, the advanced thought of comparative learning and the data processing method of multi-modal weighting are ingeniously combined, so that efficient and accurate detection of the inconsistency of the single batteries is realized in a complex and changeable actual environment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the application field of deep learning in energy storage systems, and in particular to an energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting. Background Art

[0002] At present, the system form of the power system has changed from the three elements of "source, grid, load" to the four elements of "source, grid, load, storage". Based on the demand for energy storage elements in the new power system, it has begun to play an increasingly important role.

[0003] Energy storage batteries store electricity during low electricity consumption periods and release it during peak electricity consumption periods, thereby helping to balance fluctuations between energy supply and demand, improving the stability and reliability of the power system, and increasing energy utilization. In addition, energy storage batteries can also be used for backup power, emergency power support, and smooth output of renewable energy, promoting efficient energy utilization and energy conservation and emission reduction. In the context of energy transformation and sustainable development, the importance of energy storage batteries is becoming increasingly prominent.

[0004] However, there are many energy storage devices and cells, and the operation and maintenance management is complex. The inconsistency problem caused by the synchronous control of thousands of cells is prominent. Most methods now use a cell balancing management system. It dynamically balances the cells in the battery pack according to the charge and discharge status and performance characteristics of each cell to ensure balanced energy distribution between the cells, but this method cannot detect cells with inconsistent status.

[0005] The energy storage battery monomer inconsistency diagnosis system based on contrastive learning and multimodal weighting is an innovative technology that aims to solve the inconsistency problem caused by the different characteristics and states of different battery monomers in the battery pack of the energy storage system. It uses the idea of ​​contrastive learning to learn features by comparing the similarities and differences between different samples, so that it can effectively discover the inconsistency in the battery pack under unsupervised conditions. Multimodal weighting combines information from multiple data sources (such as voltage, current, temperature, etc.), comprehensively considers the importance of different data through weight distribution, and improves the accuracy and robustness of the detection system. Summary of the invention

[0006] In order to overcome the deficiencies in the prior art, the present invention provides an energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting, which is used to detect the inconsistency of battery cells in a battery pack.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted to solve the technical problem is as follows:

[0008] The energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting includes the following steps:

[0009] Step 1: Collect voltage, current and temperature data from the battery module and perform dimension alignment;

[0010] Step 2: Construct the maximum objective function of the battery cell difference within the module to obtain the multi-view features of the battery module;

[0011] Step 3: Use a multi-head self-attention network to learn feature representations in different views;

[0012] Step 4: Define the loss function for model training;

[0013] Step 5: Perform inconsistency detection within the battery module.

[0014] Furthermore, step 1 includes the following:

[0015] For each battery cell, a voltage sensor is used to collect voltage data V every 1 second, a current sensor is used to collect current data I, and a temperature sensor is used to collect temperature data T every 5 seconds;

[0016] The collected temperature data is processed by linear interpolation. Linear interpolation is based on the linear relationship between adjacent data points. The temperature value at any time point is estimated by calculating the time ratio. The formula is as follows:

[0017]

[0018] Where t1 and t2 are the times of two known temperature data points, and T1 and T2 are the corresponding temperatures.

[0019] Furthermore, step 2 includes the following:

[0020] First, construct the maximum objective function D of the battery cell difference within the module max , the formula is as follows:

[0021]

[0022] Wherein, N is the number of battery cells in the same module, i is the battery cell number, T is the battery cell temperature data, U is the battery cell voltage data, and I is the battery cell current data;

[0023] Select a genetic algorithm to solve the objective function and find the weight combination α1, β1, γ1 that maximizes the difference between battery cells in the module. Here, α1, β1, γ1 are local optimal solutions.

[0024] Similarly, construct the minimum objective function D of the battery cell difference within the module min , the formula is as follows:

[0025]

[0026] Wherein, N is the number of battery cells in the same module, i is the battery cell number, T is the battery cell temperature data, U is the battery cell voltage data, and I is the battery cell current data;

[0027] Similarly, the genetic algorithm is selected to find the weight combination α2, β2, and γ2 that minimizes the difference between battery cells in the module. Here, α2, β2, and γ2 are local optimal solutions.

[0028] Then, the temperature, voltage, and current characteristics of the battery cells are spliced ​​according to the weight ratio, and the formula is as follows:

[0029] P i =Concat[α1T i β1U i ;γ1I i ]

[0030] N i =Concat[α1T i β1U i ;γ1I i ]

[0031] Among them, Concat[A; B] represents matrix concatenation, and this process can obtain two mixed views of the battery cell, namely, mixed view P and mixed view N.

[0032] Furthermore, step 3 includes the following:

[0033] For each battery cell, a multi-head self-attention network is used to learn representations in different views, where the number of encoder layers is M and the hidden state dimension is d model , the number of attention heads is H, and the feature extraction process is as follows:

[0034] For each view, multi-head attention first encodes it, and the encoded view dimension is P∈ and Then use multi-head attention weights to compute view representation;

[0035] First, initialize the Query and Key of the P view:

[0036]

[0037] in, Indicates Query, represents Key, H is the number of heads of multi-head attention, is the learnable parameter matrix;

[0038] Then, calculate the attention weight score:

[0039]

[0040] Among them, Softmax() is a normalization function, Concat[] represents matrix concatenation, is the learnable parameter matrix;

[0041] Finally, the attention mechanism score is applied to the original feature map to obtain different view representations of the same battery cell:

[0042]

[0043] Among them, ⊙ represents element-by-element multiplication;

[0044] Similarly, perform the same operation on view N:

[0045] Initialize Query and Key:

[0046]

[0047] in, Indicates Query, represents Key, H is the number of heads of multi-head attention, is the learnable parameter matrix;

[0048] Then, calculate the attention weight score:

[0049]

[0050] Among them, Softmax() is a normalization function, Concat[] represents matrix concatenation, is the learnable parameter matrix;

[0051] Finally, the attention mechanism score is applied to the original feature map to obtain different view representations of the same battery cell:

[0052]

[0053] Where ⊙ represents element-by-element multiplication.

[0054] Furthermore, step 4 includes the following:

[0055] The loss function formula is defined as follows:

[0056]

[0057] Among them, X represents the input time series, Stopgrad() stops gradient calculation, P and N are two different view representations of the same object, and KL() is the KL divergence, which is used to measure the difference between two feature maps. The formula is as follows:

[0058]

[0059] Where i represents the battery cell number, P(i) and N(i) represent the probabilities of the features of battery i in views P and N, respectively. The larger the divergence value, the greater the difference between the two probability distributions. The divergence value is 0 if and only if P and N are completely consistent.

[0060] Furthermore, step 5 includes the following:

[0061] By calculating the KL divergence of the feature graphs of the same battery cell from two different perspectives, normal batteries and inconsistent batteries are distinguished. First, the inconsistency score of the battery is calculated. The formula is as follows:

[0062]

[0063] Where X represents the input time series, KL() is the KL divergence, which is used to measure the difference between two feature maps;

[0064] Finally, the working condition of the battery is determined, and the inconsistency standard is set by the prior threshold δ, and the formula is as follows:

[0065]

[0066] Among them, δ[0,1] is the prior threshold of the inconsistency standard;

[0067] If the inconsistent score of the battery cell does not exceed the threshold, the battery cell is considered normal, otherwise, the battery is determined to be inconsistent.

[0068] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0069] The present invention uses the idea of ​​contrastive learning to compare the similarities and differences between different samples, so that the system can learn the characteristic performance of battery cells under various operating conditions. This unsupervised learning method allows the system to effectively discover potential inconsistency problems in the battery pack without relying on a large amount of labeled data. At the same time, in order to further improve the accuracy and robustness of the diagnostic system, the present invention also introduces a multimodal weighted data processing method. In practical applications, the state of the battery is often affected by many factors, such as voltage, current, temperature, etc. Therefore, the system comprehensively considers these multimodal data and assigns different weights to each type of data to reflect their importance in the diagnosis of battery cell inconsistency. This weight distribution method enables the system to more accurately capture changes in battery cell performance, thereby improving the accuracy and reliability of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0071] Figure 1 It is a system structure diagram corresponding to the energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting of the present invention;

[0072] Figure 2 It is a diagram of the energy storage battery management system in the present invention;

[0073] Figure 3 It is a flow chart of constructing multi-view features by genetic algorithm in the present invention;

[0074] Figure 4 This is the structure diagram of the multi-head attention mechanism in the present invention. DETAILED DESCRIPTION

[0075] The technical solution of the present invention will be described clearly and completely below 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 creative work are within the scope of protection of the present invention.

[0076] like Figure 1 As shown, this embodiment discloses a method for diagnosing inconsistency of energy storage batteries based on contrastive learning and multi-modal weighting, comprising the following steps:

[0077] Step 1: Collect voltage, current and temperature data from the battery module and perform dimension alignment;

[0078] Furthermore, step 1 includes the following:

[0079] like Figure 2 As shown, taking a certain energy storage power station as an example, the energy storage power station contains 9 clusters, each cluster contains 9 battery modules, and each battery module has 45 battery cells densely arranged inside. For each battery cell, a voltage sensor is used to collect its voltage data V every 1 second, and a current sensor is used to collect its current data I. At the same time, a temperature sensor is used to collect its temperature data T every 5 seconds;

[0080] Due to technical cost and other limitations, the acquisition frequency of temperature sensors is relatively low. In order to make up for this deficiency, the collected temperature data is processed by linear interpolation. Linear interpolation is based on the linear relationship between adjacent data points and estimates the temperature value at any time point by calculating the time ratio. The formula is as follows:

[0081]

[0082] Where t1 and t2 are the times of two known temperature data points, and T1 and T2 are the corresponding temperatures.

[0083] Step 2: Construct the maximum objective function of the battery cell difference within the module to obtain the multi-view features of the battery module;

[0084] Furthermore, step 2 includes the following:

[0085] First, construct the maximum objective function D of the battery cell difference within the module max , the formula is as follows:

[0086]

[0087] Wherein, N is the number of battery cells in the same module, i is the battery cell number, T is the battery cell temperature data, U is the battery cell voltage data, and I is the battery cell current data;

[0088] Select the genetic algorithm to solve the objective function. The genetic algorithm process is as follows: Figure 3 As shown, find the weight combination α1, β1, γ1 that maximizes the difference between battery cells in the module. Here, α1, β1, γ1 are local optimal solutions;

[0089] Similarly, construct the minimum objective function D of the battery cell difference within the module min , the formula is as follows:

[0090]

[0091] Wherein, N is the number of battery cells in the same module, i is the battery cell number, T is the battery cell temperature data, U is the battery cell voltage data, and I is the battery cell current data;

[0092] Similarly, the genetic algorithm is selected to find the weight combination α2, β2, and γ2 that minimizes the difference between battery cells in the module. Here, α2, β2, and γ2 are local optimal solutions.

[0093] Then, the temperature, voltage, and current characteristics of the battery cells are spliced ​​according to the weight ratio, and the formula is as follows:

[0094] P i =Concat[α1T i β1U i ;γ1I i ]

[0095] N i =Concat[α1T i β1Ui ;γ1I i ]

[0096] Among them, Concat[A;B] represents matrix concatenation, and this process can obtain two mixed views of the battery cell, namely mixed view P and mixed view N. This representation enlarges the difference between normal points and inconsistent points, which is conducive to the subsequent use of attention mechanism to extract features and the use of contrastive learning ideas to detect battery cell inconsistency.

[0097] Step 3: Use a multi-head self-attention network to learn feature representations in different views;

[0098] Furthermore, step 3 includes the following:

[0099] For each battery cell, a multi-head self-attention network is used to learn representations in different views, where the number of encoder layers is M and the hidden state dimension is d model , the number of attention heads is H, and the feature extraction process is as follows:

[0100] like Figure 4 As shown, for each view, multi-head attention is first encoded, and the encoded view dimension is and Then use multi-head attention weights to compute view representation;

[0101] First, initialize the Query and Key of the P view:

[0102]

[0103] in, Indicates Query, represents Key, H is the number of heads of multi-head attention, is the learnable parameter matrix;

[0104] Then, calculate the attention weight score:

[0105]

[0106] Among them, Softmax() is a normalization function, Concat[] represents matrix concatenation, is the learnable parameter matrix;

[0107] Finally, the attention mechanism score is applied to the original feature map to obtain different view representations of the same battery cell:

[0108]

[0109] Among them, ⊙ represents element-by-element multiplication;

[0110] Similarly, perform the same operation on view N:

[0111] Initialize Query and Key:

[0112]

[0113] in, Indicates Query, represents Key, H is the number of heads of multi-head attention, is the learnable parameter matrix;

[0114] Then, calculate the attention weight score:

[0115]

[0116] Among them, Softmax() is a normalization function, Concat[] represents matrix concatenation, is the learnable parameter matrix;

[0117] Finally, the attention mechanism score is applied to the original feature map to obtain different view representations of the same battery cell:

[0118]

[0119] Where ⊙ represents element-by-element multiplication.

[0120] It is worth noting that it can be shared among multiple attention heads.

[0121] Step 4: Define the loss function for model training;

[0122] Furthermore, step 4 includes the following:

[0123] The loss function formula is defined as follows:

[0124]

[0125] Among them, X represents the input time series, Stopgrad() stops gradient calculation, P and N are two different view representations of the same object, and KL() is the KL divergence, which is used to measure the difference between two feature maps. The formula is as follows:

[0126]

[0127] Where i represents the battery cell number, P(i) and N(i) represent the probabilities of the features of battery i in views P and N, respectively. The larger the divergence value, the greater the difference between the two probability distributions. The divergence value is 0 if and only if P and N are completely consistent.

[0128] Step 5: Perform inconsistency detection within the battery module.

[0129] Furthermore, step 5 includes the following:

[0130] For normal points, due to their strong correlation, they will share the same potential pattern even at different viewing angles in most cases. However, inconsistent points are rare and have no clear pattern, so the correlation between inconsistent points and other points is weak, and it is difficult for them to share potential patterns with normal points or with each other. Therefore, by calculating the KL divergence of the feature maps of the same battery cell at two different viewing angles, normal batteries and inconsistent batteries are distinguished. First, the inconsistency score of the battery is calculated, and the formula is as follows:

[0131]

[0132] Where X represents the input time series, KL() is the KL divergence, which is used to measure the difference between two feature maps;

[0133] Finally, the working condition of the battery is determined, and the inconsistency standard is set by the prior threshold δ, and the formula is as follows:

[0134]

[0135] Among them, δ[0,1] is the prior threshold of the inconsistency standard;

[0136] If the inconsistent score of the battery cell does not exceed the threshold, the battery cell is considered normal, otherwise, the battery is judged to be inconsistent. In actual situations, δ can be set to 0.95.

[0137] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for diagnosing inconsistency of energy storage batteries based on contrastive learning and multimodal weighting, characterized in that: The following steps are involved: Step 1: Collect voltage, current and temperature data from the battery module and perform dimension alignment; Step 2: Construct the maximum objective function of the battery cell difference within the module to obtain the multi-view features of the battery module; Step 3: Use a multi-head self-attention network to learn feature representations in different views; Step 4: Define the loss function for model training; Step 5: Perform inconsistency detection within the battery module.

2. The energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting according to claim 1 is characterized in that: Step 1 includes the following: For each battery cell, a voltage sensor is used to collect voltage data V every 1 second, a current sensor is used to collect current data I, and a temperature sensor is used to collect temperature data T every 5 seconds; The collected temperature data is processed by linear interpolation. Linear interpolation is based on the linear relationship between adjacent data points. The temperature value at any time point is estimated by calculating the time ratio. The formula is as follows: Where t1 and t2 are the times of two known temperature data points, and T1 and T2 are the corresponding temperatures.

3. The energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting according to claim 1 is characterized in that: Step 2 includes the following: First, construct the maximum objective function D of the battery cell difference within the module max , the formula is as follows: Wherein, N is the number of battery cells in the same module, i is the battery cell number, T is the battery cell temperature data, U is the battery cell voltage data, and I is the battery cell current data; Select a genetic algorithm to solve the objective function and find the weight combination α1, β1, γ1 that maximizes the difference between battery cells in the module. Here, α1, β1, γ1 are local optimal solutions. Similarly, construct the minimum objective function D of the battery cell difference within the module min , the formula is as follows: Wherein, N is the number of battery cells in the same module, i is the battery cell number, T is the battery cell temperature data, U is the battery cell voltage data, and I is the battery cell current data; Similarly, the genetic algorithm is selected to find the weight combination α2, β2, and γ2 that minimizes the difference between battery cells in the module. Here, α2, β2, and γ2 are local optimal solutions. Then, the temperature, voltage, and current characteristics of the battery cells are spliced ​​according to the weight ratio, and the formula is as follows: P i =Concat[α1T i ;β1U i ;γ1I i ] N i =Concat[α1T i ;β1U i ;γ1I i ] Among them, Concat[A; B] represents matrix concatenation, and this process can obtain two mixed views of the battery cell, namely, mixed view P and mixed view N.

4. The energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting according to claim 1 is characterized in that: Step 3 includes the following: For each battery cell, a multi-head self-attention network is used to learn representations in different views, where the number of encoder layers is M and the hidden state dimension is d model , the number of attention heads is H, and the feature extraction process is as follows: For each view, multi-head attention first encodes it, and the encoded view dimension is and Then use multi-head attention weights to compute view representation; First, initialize the Query and Key of the P view: in, Indicates Query, represents Key, H is the number of heads of multi-head attention, is the learnable parameter matrix; Then, calculate the attention weight score: Among them, Softmax() is a normalization function, Concat[] represents matrix concatenation, is the learnable parameter matrix; Finally, the attention mechanism score is applied to the original feature map to obtain different view representations of the same battery cell: Among them, ⊙ represents element-by-element multiplication; Similarly, perform the same operation on view N: Initialize Query and Key: in, Indicates Query, represents Key, H is the number of heads of multi-head attention, is the learnable parameter matrix; Then, calculate the attention weight score: Among them, Softmax() is a normalization function, Concat[] represents matrix concatenation, is the learnable parameter matrix; Finally, the attention mechanism score is applied to the original feature map to obtain different view representations of the same battery cell: Where ⊙ represents element-by-element multiplication.

5. The energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting according to claim 1 is characterized in that: Step 4 includes the following: The loss function formula is defined as follows: Among them, X represents the input time series, Stopgrad() stops gradient calculation, P and N are two different view representations of the same object, and KL() is the KL divergence, which is used to measure the difference between two feature maps. The formula is as follows: Where i represents the battery cell number, P(i) and N(i) represent the probabilities of the features of battery i in views P and N, respectively. The larger the divergence value, the greater the difference between the two probability distributions. The divergence value is 0 if and only if P and N are completely consistent.

6. The energy storage battery inconsistency diagnosis method based on contrastive learning and multi-modal weighting according to claim 1 is characterized in that: Step 5 includes the following: By calculating the KL divergence of the feature graphs of the same battery cell from two different perspectives, normal batteries and inconsistent batteries are distinguished. First, the inconsistency score of the battery is calculated. The formula is as follows: Where X represents the input time series, KL() is the KL divergence, which is used to measure the difference between two feature maps; Finally, the working condition of the battery is determined, and the inconsistency standard is set by the prior threshold δ, and the formula is as follows: Among them, δ[0,1] is the prior threshold of the inconsistency standard; If the inconsistent score of the battery cell does not exceed the threshold, the battery cell is considered normal, otherwise, the battery is determined to be inconsistent.

Citation Information

Cited By

  • Method for estimating residual capacity of lithium battery of energy storage system based on new energy electric field

    CN120779237A

  • Battery cell inconsistency detection method and system based on multi-mode time sequence analysis

    CN121385649A