Battery multi-state joint estimation method and device, electronic equipment and storage medium

By using the current charge and discharge cycle data and data-driven model, combined with convolutional neural network and other technologies, different time resolution features are extracted, the problem of the dependence of the multi-state joint estimation method on historical data is solved, and the accuracy and efficiency of the battery state is achieved. Joint estimates.

CN118964939APending Publication Date: 2024-11-15TSINGHUA UNIVERSITY
View PDF 0 Cites 7 Cited by

Patent Information

Application Number
CN202411024731.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing multi-state joint estimation method of battery is highly dependent on battery historical data, resulting in poor applicability in practical application scenarios and it is difficult to accurately evaluate the comprehensive state of the battery.

Method used

By using the current charge and discharge cycle data and the trained data-driven model, feature sequences with different time resolution characteristics are extracted, combined with convolutional neural network, multi-head self-attention mechanism and feature encoder, joint estimation of battery status indicators is performed. Reduce dependence on historical data.

Benefits of technology

This realizes the joint estimation of battery multi-state indicators accurately and efficiently without relying on battery historical data, which improves the applicability and accuracy of the battery multi-state joint estimation method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964939A_ABST
    Figure CN118964939A_ABST
Patent Text Reader

Abstract

The invention provides a battery multi-state joint estimation method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining an original feature sequence which is obtained according to the current charge-discharge cycle data of a battery; and inputting the original feature sequence into a trained data driving model to obtain target state indexes output by the trained data driving model, the target state indexes including at least two types of battery state indexes, and the at least two types of battery state indexes having different time resolution characteristics. According to the method, the target state index of the battery is obtained according to the current charge-discharge cycle data of the battery and the trained data driving model, the dependence on historical data of the battery is reduced, and the applicability of the battery multi-state joint estimation method is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the cross-technical field of machine learning and new energy batteries, and in particular to a battery multi-state joint estimation method, device, electronic device and storage medium. Background Art

[0002] In order to improve the use value of the battery throughout its life cycle, it is usually necessary to use the battery in a second-life cycle. That is, batteries that do not meet high-power requirements (such as electric vehicles) are used in low-power devices or energy system storage or conversion devices. In this case, accurately evaluating the comprehensive state of the battery is crucial to the safety and reliability of battery use and second-life utilization. At present, the battery multi-state joint estimation method based on physical models, parameter identification and filtering algorithms usually requires the acquisition of reliable battery historical data as a basis.

[0003] Since the existing battery multi-state joint estimation methods are too dependent on battery historical data, and battery historical data is difficult to obtain in most practical application scenarios, the current battery multi-state joint estimation methods have poor applicability. Summary of the invention

[0004] The present invention provides a battery multi-state joint estimation method, device, electronic device and storage medium, which are used to solve the defect of poor applicability of the battery multi-state joint estimation method in the prior art. The target state index of the battery is obtained according to the current charge and discharge cycle data of the battery and a trained data-driven model, which reduces the dependence on the battery historical data and improves the applicability of the battery multi-state joint estimation method.

[0005] The present invention provides a battery multi-state joint estimation method, comprising the following steps: Acquire an original feature sequence, where the original feature sequence is obtained based on current charge and discharge cycle data of the battery; The original feature sequence is input into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

[0006] According to a battery multi-state joint estimation method provided by the present invention, the trained data-driven model includes a feature extraction model, a cross-learning model, a feature encoder and a state estimator, and the original feature sequence is input into the trained data-driven model to obtain a target state indicator output by the trained data-driven model, including: Based on the feature extraction model, downsampling the input original feature sequence to obtain at least two initial feature sequences, wherein the feature extraction model includes at least two first convolutional neural networks, the at least two first convolutional neural networks include convolution kernels of different sizes, and the at least two initial feature sequences have different time resolution characteristics; Based on the cross-learning model, feature extraction and coupling relationship modeling are performed on the at least two initial feature sequences to obtain a target update sequence; Based on the feature encoder, projecting the target update sequence into multiple subspaces for encoding to obtain encoded features; State estimation is performed based on the state estimator and the encoded features to obtain the target state indicator.

[0007] According to a battery multi-state joint estimation method provided by the present invention, the at least two first convolutional neural networks include a first sampling convolutional neural network and a second sampling convolutional neural network, the convolution kernel size corresponding to the first sampling convolutional neural network is smaller than the convolution kernel size corresponding to the second sampling convolutional neural network, and the original feature sequence input is downsampled based on the feature extraction model to obtain at least two initial feature sequences, including: Downsampling the original feature sequence based on the first sampling convolutional neural network to obtain a first initial feature sequence, where the first initial feature sequence is used to characterize features with a smaller time resolution; The original feature sequence is downsampled based on the second sampling convolutional neural network to obtain a second initial feature sequence, where the second initial feature sequence is used to characterize features with a larger time resolution.

[0008] According to a battery multi-state joint estimation method provided by the present invention, the cross-learning model includes at least two multi-head self-attention mechanism units and at least two second convolutional neural networks. Based on the cross-learning model, feature extraction and coupling relationship modeling are performed on the at least two initial feature sequences to obtain a target update sequence, including: Mapping the at least two initial feature sequences through the at least two multi-head self-attention mechanism units to obtain at least two hidden sequences; Performing feature size transformation on the at least two hidden sequences by using the at least two one-dimensional second convolutional neural networks to obtain at least two transformed hidden sequences; Perform interactive modeling based on the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences; The at least two updated sequences are concatenated to obtain the target updated sequence.

[0009] According to a battery multi-state joint estimation method provided by the present invention, the at least two initial feature sequences include a first initial feature sequence and a second initial feature sequence, and the interactive modeling is performed according to the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences, including: Performing element-wise multiplication and addition and subtraction operations on the transformed hidden sequences corresponding to the first initial feature sequence and the second initial feature sequence to obtain a first updated sequence; Performing element-wise multiplication and addition and subtraction operations on the second initial feature sequence and the transformed hidden sequence corresponding to the first initial feature sequence to obtain a second updated sequence; The step of splicing the at least two updated sequences to obtain the target updated sequence includes: splicing the first updated sequence and the second updated sequence to obtain a spliced ​​updated sequence; Perform residual connection and batch normalization on the concatenated update sequence to obtain the target update sequence.

[0010] According to a battery multi-state joint estimation method provided by the present invention, the feature encoder includes a multi-head attention mechanism unit and a distillation operation unit. Based on the feature encoder, the target update sequence is projected into multiple subspaces for encoding to obtain the encoded features, including: The target update sequence is mapped to three tensors through multiple linear transformations by the multi-head attention mechanism unit to obtain multiple transformed vectors, each of the multiple transformed vectors includes a query vector, a key vector and a value vector, and the similarity between the query vector and the corresponding key vector is calculated for each transformed vector to obtain a weight coefficient of the value vector corresponding to each key vector, and after normalizing the weight coefficient of the value vector corresponding to each key vector, the weight coefficient and the corresponding value vector are weighted summed to obtain an attention value corresponding to each transformed vector, and the attention values ​​corresponding to each transformed vector are spliced ​​to obtain a spliced ​​feature; The distillation operation unit performs a distillation operation on the concatenated features to obtain the encoded features.

[0011] According to a battery multi-state joint estimation method provided by the present invention, the distillation operation unit includes a first one-dimensional convolutional neural network and a pooling layer, and the distillation operation unit performs a distillation operation on the spliced ​​features to obtain the encoded features, including: Down-sampling the concatenated features based on the first one-dimensional convolutional neural network and the distillation operation of the pooling layer to obtain sampled features; The encoded features are obtained according to the sampled features.

[0012] According to a battery multi-state joint estimation method provided by the present invention, the state estimator includes a deep neural network and a fully connected network, the deep neural network includes a two-layer network, and the state estimation is performed based on the state estimator and the encoded features to obtain the target state indicator, including: Extracting local features from the encoded features through a two-layer network of the deep neural network to obtain extracted features; The extracted features are passed through at least two fully connected layers of the fully connected network to obtain the target state indicator.

[0013] According to a battery multi-state joint estimation method provided by the present invention, the target state indicator includes a battery state of charge, the state estimator includes a gated recurrent unit, and the extracted features are respectively passed through at least two fully connected layers of the fully connected network to obtain the target state indicator, including: The extracted features are sequentially passed through the gated recurrent unit and the target fully connected layer to obtain the battery state of charge, wherein the target fully connected layer is a fully connected layer of the at least two fully connected layers, and the target fully connected layer corresponds to the battery state of charge.

[0014] According to a battery multi-state joint estimation method provided by the present invention, the obtaining of the original feature sequence includes: Performing standardization processing on the current charge-discharge cycle data to obtain standardized current charge-discharge cycle data; Position coding, global time coding and data coding are performed on the standardized current charge-discharge cycle data to obtain the original feature sequence, which includes a position matrix, a global time tag and a feature vector matrix.

[0015] According to a battery multi-state joint estimation method provided by the present invention, the position encoding, global time encoding and data encoding of the standardized current charge and discharge cycle data are performed to obtain the original feature sequence, including: Converting the standardized position information of the current charge-discharge cycle data by a fixed sine function or a cosine function to obtain the position matrix; Acquire the global time information of the standardized current charge-discharge cycle data through a learnable position embedding vector to obtain the global time tag; Converting the characteristic information of the standardized current charge and discharge cycle data through a second one-dimensional convolutional neural network to obtain the characteristic vector matrix; The position matrix, the global time stamp and the feature vector matrix are combined to obtain the original feature sequence.

[0016] The present invention also provides a battery multi-state joint estimation device, comprising the following modules: A feature acquisition module, used to acquire an original feature sequence, wherein the original feature sequence is obtained based on current charge and discharge cycle data of the battery; A feature processing module is used to input the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

[0017] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any one of the above-mentioned battery multi-state joint estimation methods is implemented.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the battery multi-state joint estimation method as described in any one of the above is implemented.

[0019] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the battery multi-state joint estimation method as described above is implemented.

[0020] The battery multi-state joint estimation method, device, electronic device and storage medium provided by the present invention obtain the target state index of the battery according to the current charge and discharge cycle data of the battery and the trained data-driven model, thereby reducing the dependence on the historical data of the battery and improving the applicability of the battery multi-state joint estimation method. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 It is a flow chart of the battery multi-state joint estimation method provided by the present invention.

[0023] Figure 2 It is a flow chart of the method for obtaining the original feature sequence provided by the present invention.

[0024] Figure 3 It is a flow chart of the method for obtaining target state indicators provided by the present invention.

[0025] Figure 4 It is a flow chart of the end-to-end battery multi-state joint estimation method provided by the present invention.

[0026] Figure 5 It is one of the comparison diagrams and error diagrams of the battery SOC estimation result and the true value using the technology of the present invention.

[0027] Figure 6 This is a comparison diagram of the battery SOH estimation result using the technology of the present invention and the true value, and the second error diagram.

[0028] Figure 7 It is a comparison diagram and an error diagram of the battery RUL estimation result using the technology of the present invention and the true value.

[0029] Figure 8 It is a structural schematic diagram of a battery multi-state joint estimation device provided by the present invention.

[0030] Fig. 9 It is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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.

[0032] As the national energy structure is transforming towards clean, low-carbon and electrified, more and more application scenarios require the participation of batteries. For example, with the increasing penetration rate of electric vehicles such as electric vehicles, the demand for power batteries has also increased significantly; due to intermittent problems such as wind power and photovoltaics, the demand for energy storage battery systems in renewable energy systems has also increased significantly. In order to improve the use value of batteries throughout their life cycle, it is usually necessary to use batteries in a cascade. Batteries that do not meet high-power requirements (such as electric vehicles) are used in low-power devices or energy system storage or conversion equipment. In this case, accurately evaluating the comprehensive status of the battery is crucial to the safety and reliability of battery use and cascade utilization.

[0033] The battery multi-state joint estimation method based on physical model, parameter identification and filtering algorithm usually requires reliable battery historical cycle data and accurate identification of intermediate parameters. Otherwise, the error accumulation will lead to large state estimation errors, which is not conducive to accurate online estimation of battery state. It can be seen that the existing battery multi-state joint estimation method is too dependent on battery historical data, and battery historical data is difficult to obtain in most practical application scenarios, resulting in poor applicability of the current battery multi-state joint estimation method.

[0034] In view of this, an embodiment of the present invention provides a battery multi-state joint estimation method, comprising obtaining an original feature sequence, wherein the original feature sequence is obtained based on the current charge and discharge cycle data of the battery; inputting the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis. The present invention simultaneously obtains multiple target state indicators of the battery based on the current charge and discharge cycle data of the battery and the trained data-driven model. This method can reduce the dependence on historical battery data and improve the applicability of the battery multi-state joint estimation method.

[0035] The technical solutions in the embodiments of the present invention will be described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] Figure 1 : is a flow chart of the battery multi-state joint estimation method provided by the present invention. The battery multi-state joint estimation method can be applied to electronic devices, which can be various types of devices with information processing capabilities during implementation. For example, the electronic device can include a personal computer, a laptop computer, a PDA or a server, etc.; the electronic device can also be a mobile terminal, for example, the mobile terminal can include a mobile phone, a car computer, a tablet computer or a projector, etc. Figure 1 As shown, the method may include the following steps 101 to 102: Step 101: Acquire an original feature sequence, where the original feature sequence is obtained based on current charge and discharge cycle data of the battery.

[0037] It should be noted that there are many ways to obtain the original feature sequence. For example, the feature extraction of the current charge and discharge cycle data of the battery can be performed through an algorithm, or the current charge and discharge cycle data of the battery can be standardized and then feature extracted through a model, etc. The present invention does not limit the method for obtaining the original feature sequence.

[0038] Step 102: Input the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

[0039] It should be noted that by inputting the original feature sequence into the trained data-driven model, at least two types of battery state indicators can be obtained, and the trained data-driven model is used to extract at least two initial feature sequences according to the original feature sequence, and then perform joint state estimation according to the coupling relationship between the at least two initial feature sequences to obtain the target state indicator. Wherein, at least two initial feature sequences have different time resolution characteristics.

[0040] Among them, there can be multiple model structures of the trained data-driven model, for example, it can include a convolutional neural network, etc. As long as it can extract at least two initial feature sequences based on the original feature sequence, and then perform joint state estimation based on the coupling relationship between the at least two initial feature sequences, the model structure of the target state indicator can be obtained.

[0041] In addition, at least two types of battery status indicators may include a first type of battery status indicator and a second type of battery status indicator, wherein the first type of battery status indicator may include a battery status indicator with a smaller time resolution, such as a battery state of charge (SOC), and the second type of battery status indicator may include a battery status indicator with a larger time resolution, such as a battery state of health (SOH), a battery remaining useful life (RUL), etc.

[0042] Exemplarily, in order to achieve the purpose of the invention, the present invention can use the current charge and discharge cycle data of the battery , which include the changes of time, voltage, current, temperature and power. A data-driven model is established and trained to obtain a trained data-driven model, which automatically extracts features of different time scales, models the coupling relationship between the features, jointly estimates the battery SOC and SOH, and performs battery RUL prediction.

[0043] Compared with the prior art, the battery multi-state joint estimation method provided by the present invention does not rely on historical battery data. It can accurately and efficiently realize the joint estimation of battery multi-state indicators using only current cycle battery data, thereby improving the applicability of the battery multi-state joint estimation method.

[0044] Figure 2 FIG. 1 is a flow chart of the method for obtaining the original feature sequence provided by the present invention. Figure 2 As shown, the step 101 of obtaining the original feature sequence may include: Step 201: performing standardization processing on the current charge-discharge cycle data to obtain standardized current charge-discharge cycle data.

[0045] It should be noted that there are many methods for standardizing the current charge and discharge cycle data, such as minimum-maximum standardization (Min-Max standardization), Z-score standardization, sum normalization, etc. The present invention standardizes the current charge and discharge cycle data to obtain the standardized current charge and discharge cycle data without limiting the method.

[0046] Step 202: Position encoding, global time encoding and data encoding are performed on the standardized current charge and discharge cycle data to obtain the original feature sequence, which includes a position matrix, a global time tag and a feature vector matrix.

[0047] It should be noted that position coding is used to add position information to each element in the input sequence to make up for the lack of sense of order that the model may lack when processing the sequence. Global time coding is used to uniformly encode or identify time on a global scale. Data coding refers to the conversion of one or more characters, data or information into other forms for use in transmission, storage or processing. The present invention can obtain the original feature sequence including the position matrix, the global time mark and the feature vector matrix by performing position coding, global time coding and data coding on the standardized current charge and discharge cycle data.

[0048] Furthermore, position encoding, global time encoding and data encoding are performed on the standardized current charge and discharge cycle data to obtain the original feature sequence, which may include: converting the position information of the standardized current charge and discharge cycle data through a fixed sine function or cosine function to obtain the position matrix; acquiring the global time information of the standardized current charge and discharge cycle data through a learnable position embedding vector to obtain the global time tag; converting the feature information of the standardized current charge and discharge cycle data through a second one-dimensional convolutional neural network to obtain the feature vector matrix; merging the position matrix, the global time tag and the feature vector matrix to obtain the original feature sequence.

[0049] Exemplarily, the standardized battery cycle data is feature transformed through position encoding of the embedding layer, global time encoding, and numerical encoding of one-dimensional convolution.

[0050] Among them, the position encoding converts the position information of the data into dimensional matrix, that is: , ,in, , and are the first and The global temporal encoding consists of a learnable position embedding vector , to obtain the global time tag. The numerical encoding can be transformed into a one-dimensional convolutional neural network with a convolution kernel size of 3. dimensional vector. Then the converted features are: Padding refers to the space between the border of an element and its content. This space can be controlled by the padding property of CSS, which is used to adjust the distance between the content of an element and its border, thus affecting the layout and appearance of the element.

[0051] It can be understood that the above method is used to extract features based on the current charge and discharge cycle data of the battery, so as to obtain an accurate original feature sequence, which lays a foundation for joint estimation of multiple states of the battery.

[0052] In some embodiments, the trained data-driven model may be composed of multiple models with different functions to obtain accurate target state indicators.

[0053] Figure 3 FIG. 1 is a flow chart of a method for obtaining a target state indicator provided by the present invention. Figure 3 As shown, the trained data-driven model may include a feature extraction model, a cross-learning model, a feature encoder and a state estimator. The step 102 inputs the original feature sequence into the trained data-driven model to obtain the target state indicator output by the trained data-driven model, which may include: Step 301: Based on the feature extraction model, downsample the input original feature sequence to obtain at least two initial feature sequences, wherein the feature extraction model includes at least two first convolutional neural networks, the at least two first convolutional neural networks include convolution kernels of different sizes, and the at least two initial feature sequences have different time resolution characteristics.

[0054] It should be noted that the trained data-driven model may include a feature extraction module, which is used to downsample the original feature sequence to obtain at least two initial feature sequences, and the feature extraction model may include at least two first convolutional neural networks, the at least two first convolutional neural networks include different convolution kernel sizes, and the at least two initial feature sequences have different time resolution characteristics. The feature extraction module can effectively capture features of different time scales, which is crucial to accurately jointly estimate the target state indicator of the battery.

[0055] Step 302: Based on the cross-learning model, feature extraction and coupling relationship modeling are performed on the at least two initial feature sequences to obtain a target update sequence.

[0056] It should be noted that the cross-learning model can perform deep feature extraction on at least two initial feature sequences, and model the coupling relationship to obtain a target update sequence, wherein the target update sequence includes features obtained by deep feature extraction and features generated by coupling between features. The cross-learning model can realize information exchange and coupling between features with different time resolutions.

[0057] Step 303: Based on the feature encoder, project the target update sequence onto multiple subspaces for encoding to obtain encoded features.

[0058] It should be noted that the feature encoder is used to perform deep feature encoding on the target update sequence obtained by the cross-learning model. By projecting the input sequence into multiple subspaces and calculating the corresponding weights to obtain the encoded features, the most informative features can be effectively extracted and combined to accurately estimate the battery state.

[0059] Among them, the feature encoder may include a head attention mechanism, etc., and the present invention does not limit the structure of the feature encoder.

[0060] Step 304: Perform state estimation based on the state estimator and the encoded features to obtain the target state indicator.

[0061] It should be noted that after obtaining the encoded features, the state estimator can perform multi-state estimation of the battery based on the encoded features, which can effectively extract and combine the most informative features, thereby accurately estimating the battery state. The state estimator may include a convolutional neural network, etc., and the present invention does not limit the type of the state estimator.

[0062] It can be understood that the trained data-driven module includes four models with different functions. The input original feature sequence is coordinated and processed by the four models with different functions to obtain the target state index, which can improve the efficiency and accuracy of the joint estimation of battery multi-states.

[0063] In some embodiments, the at least two first convolutional neural networks include a first sampling convolutional neural network and a second sampling convolutional neural network, the convolution kernel size corresponding to the first sampling convolutional neural network is smaller than the convolution kernel size corresponding to the second sampling convolutional neural network, and based on the feature extraction model, the input original feature sequence is downsampled to obtain at least two initial feature sequences, which may include: downsampling the original feature sequence based on the first sampling convolutional neural network to obtain a first initial feature sequence, and the first initial feature sequence is used to characterize features with smaller time resolution; downsampling the original feature sequence based on the second sampling convolutional neural network to obtain a second initial feature sequence, and the second initial feature sequence is used to characterize features with larger time resolution.

[0064] It should be noted that multi-time resolution downsampling can extract features of different time scales in the sequence. The original feature sequence can be downsampled by designing a convolutional neural network (CNN) with different convolution kernel sizes to obtain preliminary feature sequences at different time resolutions.

[0065] For example, in order to obtain features of different time resolutions in the original feature sequence, convolution kernel sequences of different sizes are designed. Downsampling is performed to obtain two feature sequences with different time resolutions. The small convolution kernel size is set to 5 and the step size is 1, which represents the features with smaller time resolution; the large convolution kernel size is set to 25 and the step size is 25, which represents the features with larger time resolution. After downsampling, the feature sequences with two different time resolutions can be expressed as: , Obviously, these two feature sequences and The lengths are different and contain features of different time resolutions. represents the intra-loop features at a smaller resolution, while Indicates the overall characteristics of the current cycle.

[0066] It can be understood that the above feature extraction model can effectively capture the features of different time scales, which is crucial for accurately jointly estimating battery SOC and SOH.

[0067] In some embodiments, the cross-learning model includes at least two multi-head self-attention mechanism units and at least two second convolutional neural networks. Based on the cross-learning model, the at least two initial feature sequences are subjected to feature extraction and coupling relationship modeling to obtain a target update sequence, which may include: mapping the at least two initial feature sequences through the at least two multi-head self-attention mechanism units to obtain at least two hidden sequences; performing feature size transformation on the at least two hidden sequences through the at least two second convolutional neural networks to obtain at least two transformed hidden sequences; performing interactive modeling based on the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences; and splicing the at least two updated sequences to obtain the target update sequence.

[0068] It should be noted that the cross-learning model performs further in-depth feature extraction on the two preliminary feature sequences obtained by downsampling based on the multi-head self-attention mechanism, and designs cross-learning to model the coupling between feature sequences of different time scales, and merges the two feature sequences to obtain an updated sequence including features of different time scales and coupling effects between features.

[0069] Furthermore, the at least two initial feature sequences include a first initial feature sequence and a second initial feature sequence, and interactive modeling is performed based on the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences, which may include: performing element-wise multiplication and addition and subtraction operations on the first initial feature sequence and the transformed hidden sequence corresponding to the second initial feature sequence to obtain a first updated sequence; performing element-wise multiplication and addition and subtraction operations on the second initial feature sequence and the transformed hidden sequence corresponding to the first initial feature sequence to obtain a second updated sequence.

[0070] Splicing the at least two updated sequences to obtain the target updated sequence may include: splicing the first updated sequence and the second updated sequence to obtain a spliced ​​updated sequence; and performing residual connection and batch normalization on the spliced ​​updated sequence to obtain the target updated sequence.

[0071] For example, interactive learning can be used to model the coupling between two feature sequences with different time resolutions and exchange information. The interactive learning operation can be as follows: two multi-head self-attention mechanism units with different parameters are used to model the coupling between two feature sequences with different time resolutions. and , two feature sequences with different time resolutions and is mapped to two hidden sequences. The parameters can be trained to be suitable for processing Characteristics, The parameters can be trained to be suitable for processing characteristics. and The most relevant features of each temporal resolution hidden sequence can be extracted. Then two second convolutional neural networks with different parameters are used and Transform the feature sizes of the two hidden sequences to facilitate information interaction and coupling modeling between the two hidden sequences. Used for The feature size of the hidden sequence is transformed to fit with To exchange information, Used for The feature size transformation of the hidden sequence is suitable for Finally, based on the element-level product and addition and subtraction operations and the interaction modeling with other hidden sequences, we can get two update sequences with different time resolutions as follows: , .in, Represents the product operation at the element level. The above interactive learning process makes the updated sequence contain both the original feature information of each time resolution sequence and the updated information from the features of another sequence.

[0072] Next, the two updated sequences of different time scales are concatenated, and through residual connection and batch normalization, the overall output of the cross-learning model is expressed as: .

[0073] It can be seen that the final updated feature sequence includes features of smaller time resolution and larger time resolution, as well as the interactive information between the two different time scale sequences. Residual connection and batch normalization can help the model training and learning process to be more stable and improve the performance of the overall model.

[0074] It can be understood that the cross-learning model is the core part of the model framework. It can realize the information exchange between features of different time resolutions. Since the battery state indicators are closely related and coupled to each other, for example, SOC and SOH are closely related and coupled to each other. At the same time, downsampling may also lead to information loss. The cross-learning model allows interaction and update between feature sequences with different time resolutions, and performs targeted updates of each time feature on the basis of preserving and utilizing feature information of different time scales, so as to obtain more accurate features.

[0075] Specifically, the multi-time resolution downsampling in the solution of the present invention can extract and process features at different time scales in battery cycle data. The large convolution kernel extracts the long time scale features corresponding to SOH and RUL, and the small convolution kernel extracts the short time scale characteristics corresponding to the SOC within the cycle. The interactive learning is used to model the interaction and coupling relationship between features of different time scales. In this way, the deficiencies of the prior art are solved, thereby improving the efficiency and accuracy of the joint estimation of multiple states of the battery.

[0076] In some embodiments, the feature encoder includes a multi-head attention mechanism unit and a distillation operation unit. Based on the feature encoder, the target update sequence is projected onto multiple subspaces for encoding to obtain encoded features, which may include: mapping the target update sequence into three tensors through multiple linear transformations through the multi-head attention mechanism unit to obtain multiple transformed vectors, each of the multiple transformed vectors includes a query vector, a key vector and a value vector, calculating the similarity between the query vector and the corresponding key vector for each transformed vector to obtain a weight coefficient of the value vector corresponding to each key vector, normalizing the weight coefficient of the value vector corresponding to each key vector, performing weighted summation on the weight coefficient and the corresponding value vector to obtain an attention value corresponding to each transformed vector, and splicing the attention values ​​corresponding to each transformed vector to obtain a spliced ​​feature; performing a distillation operation on the spliced ​​feature through the distillation operation unit to obtain the encoded feature.

[0077] It should be noted that through the feature encoder including the multi-head attention mechanism and the distillation operation, the update sequence obtained by the cross-learning model can be deeply feature encoded.

[0078] In some embodiments, the distillation operation unit includes a first one-dimensional convolutional neural network and a pooling layer. The distillation operation unit performs a distillation operation on the spliced ​​features to obtain the encoded features, which may include: down-sampling the spliced ​​features based on the distillation operation of the first one-dimensional convolutional neural network and the pooling layer to obtain sampled features; and obtaining the encoded features based on the sampled features.

[0079] For example, the multi-head self-attention mechanism can be used for feature extraction of the cross-learning model and feature encoder. Mapped into three tensors through linear transformation, namely: This transformation is done h times, which is called multi-head. The parameters of each linear transformation are , , They are different. The values ​​of specific parameters are learned during the model training process.

[0080] After the linear transformation, the query vector is calculated and key vector similarity or correlation, and obtain each key vector Corresponding value vector The weight coefficient is normalized by softmax, and the weight and the corresponding value vector Perform weighted summation to get the Attention value as follows: Among them, the factor It plays a regulatory role. After h calculations, the model is allowed to learn relevant information in different representation subspaces and concatenate them. Finally, the output of the attention mechanism model of the i-th layer is expressed as: The computational idea of ​​the multi-head attention mechanism is to effectively extract and combine the most informative features by projecting the input sequence into multiple subspaces and calculating the corresponding weights, thereby accurately estimating the battery state.

[0081] In the feature encoder, in order to improve computational efficiency and reduce memory consumption, a distillation operation based on the first one-dimensional convolutional neural network and the pooling layer is used to perform downsampling and reduce the output length, which is expressed as: After the distillation operation, the self-attention layer with half the input length is stacked to improve the robustness, wherein the distillation operation and the attention layer stacking can further enhance the model's ability to capture long-term dependencies and improve the robustness of the model.

[0082] In some embodiments, the state estimator includes a deep neural network and a fully connected network, the deep neural network includes a two-layer network, and the state estimation is performed based on the state estimator and the encoded features to obtain the target state indicator. Specifically: the local features in the encoded features can be extracted through the two-layer network of the deep neural network to obtain the extracted features; the extracted features are respectively passed through at least two fully connected layers of the fully connected network to obtain the target state indicator.

[0083] It should be noted that the joint estimation of multiple battery states such as SOC, SOH and RUL can be achieved through a state estimator based on a dilated convolutional neural network (DCNN).

[0084] Exemplarily, the state estimator can be composed of a two-layer deep convolutional neural network DCNN and a fully connected network FCN. There is a certain gap in the convolution kernel of DCNN, which can better capture the long-term relationship in time series data. A larger expansion d value can enable the expanded CNN to cover inputs longer than the kernel length. When d=1, DCNN is a standard convolution operation. Each DCNN layer includes a convolution layer and a pooling layer. After two layers of DCNN, the regularization technique dropout is used to feed the output to different FCN layers to avoid overfitting, so as to achieve the purpose of simultaneously estimating the battery SOH and RUL.

[0085] In some embodiments, the target state indicator includes a battery state of charge, the state estimator includes a gated recurrent unit, and the extracted features are passed through at least two fully connected layers of the fully connected network respectively to obtain the target state indicator, which may include: passing the extracted features through the gated recurrent unit and the target fully connected layer in sequence to obtain the battery state of charge, the target fully connected layer is a fully connected layer of the at least two fully connected layers, and the target fully connected layer corresponds to the battery state of charge.

[0086] It is understandable that the multi-state joint estimation of the battery based on the state estimator of the dilated convolutional neural network (DCNN) can improve the accuracy of the multi-state joint estimation. It should be noted that when performing SOC estimation within the cycle, in order to improve the accuracy, a gated recurrent unit (GRU) layer is added to improve the accuracy of SOC estimation.

[0087] The data-driven end-to-end algorithm provided by the present invention can avoid error accumulation to a certain extent and reduce the dependence on battery historical data, making it more suitable for online state estimation.

[0088] In some embodiments, the total loss function in the process of training the data-driven model can be defined as the weighted sum of the mean square error (MSE) between the SOC, SOH and RUL target values ​​and the model estimation results, that is: .

[0089] The end-to-end multi-time resolution battery multi-state joint estimation method proposed in the present invention combines multi-resolution downsampling, interactive learning, a feature encoder based on a multi-head self-attention mechanism, and a state estimator, which can effectively extract and utilize multi-time resolution features, and model and simulate the coupling relationship between features of different time scales, so as to achieve accurate joint estimation of at least two types of battery state indicators. At the same time, the present invention does not rely on historical battery data, but only based on partial data of the current charge and discharge cycle, it can achieve end-to-end joint estimation of at least two types of battery state indicators, and has good applicability for most practical application scenarios where historical battery data is difficult to obtain. Moreover, the method proposed in the present invention also has good robustness for batteries with different charge and discharge cycle depths.

[0090] It is understandable that the simultaneous estimation of the comprehensive state of the battery based on the data-driven algorithm needs to solve two major challenges: ① The comprehensive state evaluation of the battery, including SOC, SOH, etc., needs to consider the different time scale characteristics of multiple indicators. The battery SOC changes within each cycle, and the time resolution is small; while SOH and RUL gradually decay with the progress of the charge and discharge cycle, and the time resolution is large. ② There is coupling and mutual correlation between multiple indicators. The accurate estimation of the battery SOC requires the accurate estimation of the battery capacity, that is, SOH, and vice versa. Therefore, the present invention intends to provide an end-to-end data-driven method to achieve the joint estimation of multiple state indicators of the battery based only on the current cycle data by solving the above challenges.

[0091] Therefore, compared with the prior art, the present invention has a wider range of applications. On the one hand, the present invention can be more conveniently applied to the multi-state joint estimation of batteries without historical data. In this way, it is possible to perform comprehensive state estimation on batteries in the cascade utilization stage (such as energy storage batteries, etc.), provide a basis for the cascade utilization of batteries, improve the use value of the battery throughout its life cycle, and reduce the cost of battery use. On the other hand, the present invention can conveniently perform online estimation of the comprehensive state of the battery, perform multi-state monitoring of the battery in use, avoid battery use behaviors such as overcharging and over-discharging, and help to extend the battery service life in case of battery failure, thereby improving the safety and economy of battery use.

[0092] The following describes an exemplary application of an embodiment of the present invention in a practical application scenario.

[0093] Figure 4 FIG. 1 is a flow chart of the end-to-end battery multi-state joint estimation method provided by the present invention. Figure 4 As shown, the battery multi-state joint estimation method specifically includes the following four steps: Step 401: The original battery charge and discharge cycle data is preprocessed by standardization, and convolutional neural networks (CNNs) with different convolution kernel sizes are designed to downsample the original data to obtain preliminary feature sequences at different time resolutions.

[0094] Step 402: Further in-depth feature extraction is performed on the two preliminary feature sequences obtained by downsampling based on the multi-head self-attention mechanism, and cross-learning is designed to model the coupling effect between feature sequences of different time scales, and the two feature sequences are merged to obtain an updated sequence including features of different time scales and coupling effects between features.

[0095] Step 403: Deep feature encoding is performed on the update sequence obtained by the cross-learning module through a feature encoder including a multi-head attention mechanism and a distillation operation.

[0096] Step 404: A joint estimation of multiple battery states (SOC, SOH, and RUL) is achieved through a state estimator based on a dilated convolutional neural network (DCNN).

[0097] In the above embodiment, the core part is the multi-time resolution downsampling and cross-learning model in the method. Among them, the multi-time resolution downsampling downsamples the battery data sequence by using a convolutional neural network with convolution kernels of different lengths to obtain two preliminary feature sequences at different time resolutions. Interactive learning performs information interaction between two feature sequences with different time resolutions through feature extraction based on the self-attention mechanism and interactive modeling of different feature sequences.

[0098] The present invention achieves the purpose of the present invention through the above-mentioned scheme, that is, proposing an end-to-end multi-time scale battery multi-state joint estimation method based on attention mechanism and interactive learning, so as to solve the coupling problem between battery states of different time scales and the problem that battery historical data is difficult to obtain, and realizes the joint estimation of battery SOC, SOH and RUL based only on the current cycle data of the battery.

[0099] This embodiment uses the proposed end-to-end deep learning method to jointly estimate the battery SOC, SOH and RUL based on partial charge and discharge data in the MIT-Stanford public dataset.

[0100] Implementation data: This data is publicly available on the Experimental Data Platform (matr.io). The data includes charge and discharge cycle data of 124 lithium iron phosphate (LFP) 18650 batteries with a rated capacity of 1.1Ah. These batteries are fast charged at different rates in the 0%-80% SOC range. In the 80%-100% SOC range, the batteries are charged at the highest voltage constant voltage until the current drops below the charge cut-off current. During discharge, all batteries are discharged at a rate of 4C. The battery cycle depth is 0%-100%, and the cycle continues until the battery capacity decays to 80% of the rated capacity (that is, the usual definition of RUL).

[0101] The 124 batteries were divided into three parts: training set, validation set and test set. Some original charge and discharge data with a voltage range of [2.9, 3.6] were intercepted, including time t, current I, voltage V, temperature T, and capacity change as the input of the model, and the labeled state SOC, SOH and RUL values ​​of the training set batteries were used as output for model training. The trained model was used for multi-state joint estimation of the test set batteries, and the estimated results of battery SOC, SOH and RUL are shown below.

[0102] Figure 5 It is one of the comparison diagrams and error diagrams of the battery SOC estimation result and the true value using the technology of the present invention. Figure 5 (a) depicts the estimated SOC versus battery voltage for an example battery (#22) in the first and last cycles. The results show that despite the different SOH levels of the batteries (about 98% in the first cycle and about 80% in the last cycle), the estimated SOC can adapt to the change in SOH and remain consistent with the actual SOC value. Taking #22 and #101 batteries as examples, the mean absolute error (MAE) of the SOC within a specific cycle at the corresponding SOH level is shown in Figure 2. Figure 5 (b) As shown. It can be seen that the MAE values ​​of the SOC estimated by the model during a specific cycle are mostly below 0.5%. In addition, the SOC MAE value at a high SOH level is slightly smaller than the SOC MAE value at a low SOH level. Considering all batteries and all sampling points of each cycle, the SOC residuals estimated by the model are mostly concentrated in the error range of [-1%, 1%].

[0103] Figure 6 This is a comparison diagram of the battery SOH estimation result using the technology of the present invention and the true value, and the second error diagram. Figure 6(a) shows the estimated results of the SOH evolution curves of two example batteries (battery No. 22 and battery No. 101) as the number of cycles increases. It should be noted that the example battery selected to show the SOH results is the same as the example battery shown in the SOC results, in order to highlight the simultaneous estimation results of the battery SOC and SOH. Figure 6 The curve in (a) shows that as the number of battery cycles increases, the SOH decay curve is well consistent with the actual SOH decay curve. The inflection point where the battery SOH begins to degrade sharply can also be accurately estimated, with only a small difference between the estimated result and the actual test result. Figure 6 The SOH decay curves of the two batteries shown in (a) show that although the two batteries have the same chemistry, their decay curves are significantly different due to different cycling conditions. Nevertheless, the proposed model can effectively adapt to various charging strategies and battery decay curves, achieving accurate simultaneous SOC and SOH estimation using only raw cycling data. Figure 6 (b) shows the estimated SOH relative errors (RE) of two example batteries at specific SOHs. It can be seen that most of the estimated SOH REs are less than 1%. In general, the SOH estimation relative errors of the battery training set, validation set, and test set are all less than 0.5%.

[0104] Figure 7 The comparison chart and error chart of the battery RUL estimation result and the true value using the technology of the present invention are shown in Figure 2. The battery RUL-SOH curve results predicted by the method proposed by the present invention are shown in Figure 2. Figure 7 (a) is shown. Figure 7 For the two example batteries in (a) (battery No. 22 and battery No. 101), the estimated RUL-SOH curve is very close to the true curve. The predicted RUL value at each SOH state fluctuates around the actual value. Considering the RUL predicted values ​​of all test set batteries, Figure 7 (b) shows the absolute error of the RUL prediction value under different RUL ranges (0-500, 500-1000, 1000-1500, 1500-2000 and 2000-2500). The results show that the error of the RUL prediction value is relatively high in the larger RUL range. In the RUL range of [0,500], the average RUL absolute error is about 45 cycles, and in the range of [2000,2500] it is about 150 cycles.

[0105] Based on the foregoing embodiments, an embodiment of the present invention provides a battery multi-state joint estimation device, and each module included in the device and each unit included in each module can be implemented by a processor; of course, it can also be implemented by a specific logic circuit; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP) or a field programmable gate array (FPGA), etc.

[0106] The battery multi-state joint estimation device provided by the present invention is described below. The battery multi-state joint estimation device described below and the battery multi-state joint estimation method described above can be referenced to each other.

[0107] Figure 8 Schematic diagram of the structure of the battery multi-state joint estimation device provided by the present invention. Figure 8 As shown, the device 500 includes a feature acquisition module 501 and a feature processing module 502, wherein: A feature acquisition module 501 is used to acquire an original feature sequence, where the original feature sequence is obtained based on current charge and discharge cycle data of the battery; The feature processing module 502 is used to input the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

[0108] In some embodiments, the trained data-driven model includes a feature extraction model, a cross-learning model, a feature encoder, and a state estimator, and the feature processing module 502 includes a first processing unit, a second processing unit, a third processing unit, and a fourth processing unit, wherein:

[0109] The first processing unit is used to downsample the input original feature sequence based on the feature extraction model to obtain at least two initial feature sequences, wherein the feature extraction model includes at least two first convolutional neural networks, the at least two first convolutional neural networks include convolution kernels of different sizes, and the at least two initial feature sequences have different time resolution characteristics; The second processing unit is used to perform feature extraction and coupling relationship modeling on the at least two initial feature sequences based on the cross-learning model to obtain a target update sequence; The third processing unit is used to project the target update sequence into multiple subspaces for encoding to obtain encoded features; The fourth processing unit is used to perform state estimation based on the state estimator and the encoded features to obtain the target state indicator.

[0110] In some embodiments, the at least two first convolutional neural networks include a first sampling convolutional neural network and a second sampling convolutional neural network, the convolution kernel size corresponding to the first sampling convolutional neural network is smaller than the convolution kernel size corresponding to the second sampling convolutional neural network, and the first processing unit is specifically used to: downsample the original feature sequence based on the first sampling convolutional neural network to obtain a first initial feature sequence, and the first initial feature sequence is used to characterize features with smaller time resolution; downsample the original feature sequence based on the second sampling convolutional neural network to obtain a second initial feature sequence, and the second initial feature sequence is used to characterize features with larger time resolution.

[0111] In some embodiments, the cross-learning model includes at least two multi-head self-attention mechanism units and at least two second convolutional neural networks, and the second processing unit includes a feature mapping component, a feature transformation component, a feature interaction component, and a feature splicing component, wherein:

[0112] The feature mapping component is used to map the at least two initial feature sequences through the at least two multi-head self-attention mechanism units to obtain at least two hidden sequences; The feature transformation component is used to perform feature size transformation on the at least two hidden sequences through the at least two second convolutional neural networks to obtain at least two transformed hidden sequences; The feature interaction component is used to perform interactive modeling based on the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences; The feature splicing component is used to splice the at least two updated sequences to obtain the target updated sequence.

[0113] In some embodiments, the at least two initial feature sequences include a first initial feature sequence and a second initial feature sequence, and the feature interaction component is specifically used to: perform element-wise multiplication and addition and subtraction operations on the transformed hidden sequence corresponding to the first initial feature sequence and the second initial feature sequence to obtain a first updated sequence; perform element-wise multiplication and addition and subtraction operations on the second initial feature sequence and the transformed hidden sequence corresponding to the first initial feature sequence to obtain a second updated sequence; the feature splicing component is specifically used to: splice the first updated sequence and the second updated sequence to obtain a spliced ​​updated sequence; perform residual connection and batch normalization on the spliced ​​updated sequence to obtain the target updated sequence.

[0114] In some embodiments, the feature encoder includes a multi-head attention mechanism unit and a distillation operation unit, and the third processing unit includes an encoding component and a distillation component, wherein,

[0115] The encoding component is used to map the target update sequence into three tensors through multiple linear transformations through the multi-head attention mechanism unit to obtain multiple transformed vectors, each of the multiple transformed vectors includes a query vector, a key vector and a value vector, calculate the similarity between the query vector and the corresponding key vector for each transformed vector, obtain the weight coefficient of the value vector corresponding to each key vector, normalize the weight coefficient of the value vector corresponding to each key vector, perform weighted summation on the weight coefficient and the corresponding value vector, obtain the attention value corresponding to each transformed vector, and splice the attention value corresponding to each transformed vector to obtain the spliced ​​feature; The distillation component is used to perform a distillation operation on the spliced ​​features through the distillation operation unit to obtain the encoded features.

[0116] In some embodiments, the distillation operation unit includes a first one-dimensional convolutional neural network and a pooling layer, and the distillation component is specifically used to: downsample the spliced ​​features based on the distillation operation of the first one-dimensional convolutional neural network and the pooling layer to obtain sampled features; and obtain the encoded features based on the sampled features.

[0117] In some embodiments, the state estimator includes a deep neural network and a fully connected network, the deep neural network includes a two-layer network, and the fourth processing unit includes a feature extraction component and a data acquisition component, wherein,

[0118] The feature extraction component is used to extract local features from the encoded features through the two-layer network of the deep neural network to obtain extracted features; The data acquisition component is used to pass the extracted features through at least two fully connected layers of the fully connected network to obtain the target state indicator.

[0119] In some embodiments, the target state indicator includes the battery state of charge, the state estimator includes a gated loop unit, and the data acquisition component is specifically used to: pass the extracted features through the gated loop unit and the target fully connected layer in sequence to obtain the battery state of charge, the target fully connected layer is a fully connected layer of the at least two fully connected layers, and the target fully connected layer corresponds to the battery state of charge.

[0120] In some embodiments, the feature acquisition module 501 includes a normalization unit and an encoding unit, wherein:

[0121] The standardization unit is used to perform standardization processing on the current charge and discharge cycle data to obtain standardized current charge and discharge cycle data; The encoding unit is used to perform position encoding, global time encoding and data encoding on the standardized current charge and discharge cycle data to obtain the original feature sequence, which includes a position matrix, a global time mark and a feature vector matrix.

[0122] In some embodiments, the encoding unit is specifically used to: convert the position information of the standardized current charge and discharge cycle data through a fixed sine function or a cosine function to obtain the position matrix; The global time information of the standardized current charge and discharge cycle data is obtained through a learnable position embedding vector to obtain the global time tag; the feature information of the standardized current charge and discharge cycle data is converted through a second one-dimensional convolutional neural network to obtain the feature vector matrix; the position matrix, the global time tag and the feature vector matrix are merged to obtain the original feature sequence.

[0123] In an embodiment of the present invention, the target state index of the battery can be obtained according to the current charge and discharge cycle data of the battery and the trained data-driven model, which reduces the dependence on the historical data of the battery and improves the applicability of the battery multi-state joint estimation method.

[0124] Fig. 9 Schematic diagram of the physical structure of the electronic device provided by the present invention. Figure 8 As shown, the electronic device may include: a processor 601, a communication interface 602, a memory 603 and a communication bus 604, wherein the processor 601, the communication interface 602 and the memory 603 communicate with each other through the communication bus 604. The processor 601 may call the logic instructions in the memory 603 to execute a battery multi-state joint estimation method, the method comprising: obtaining an original feature sequence, the original feature sequence being obtained according to the current charge and discharge cycle data of the battery; inputting the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, the target state indicator comprising at least two types of battery state indicators, the at least two types of battery state indicators having different time resolution characteristics, wherein the trained data-driven model is obtained by extracting feature sequences having different time resolution characteristics from the original feature sequence for analysis.

[0125] In addition, the logic instructions in the above-mentioned memory 603 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0126] On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the battery multi-state joint estimation method provided by the above methods, and the method includes: obtaining an original feature sequence, and the original feature sequence is obtained based on the current charge and discharge cycle data of the battery; inputting the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, and the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

[0127] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0128] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the battery multi-state joint estimation method provided by the above methods, the method comprising: obtaining an original feature sequence, the original feature sequence being obtained according to current charge and discharge cycle data of the battery; The original feature sequence is input into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

[0129] The above-mentioned computer-readable storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, - but not limited to - an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM) or flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0130] Computer-readable signal media may include a data signal propagated in baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0131] The program code embodied on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wire, optical cable, radio frequency (RF), etc., or any suitable combination of the foregoing.

[0132] Computer program code for performing the operations of the present specification may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Python, Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0133] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0134] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery multi-state joint estimation method, characterized in that: include: Acquire an original feature sequence, where the original feature sequence is obtained based on current charge and discharge cycle data of the battery; The original feature sequence is input into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator based on the feature sequences with different time resolution characteristics by extracting feature sequences with different time resolution characteristics from the original feature sequence.

2. The battery multi-state joint estimation method according to claim 1, characterized in that: The trained data-driven model includes a feature extraction model, a cross-learning model, a feature encoder and a state estimator. The original feature sequence is input into the trained data-driven model to obtain the target state indicator output by the trained data-driven model, including: Based on the feature extraction model, downsampling the input original feature sequence to obtain at least two initial feature sequences, wherein the feature extraction model includes at least two first convolutional neural networks, the at least two first convolutional neural networks include convolution kernels of different sizes, and the at least two initial feature sequences have different time resolution characteristics; Based on the cross-learning model, feature extraction and coupling relationship modeling are performed on the at least two initial feature sequences to obtain a target update sequence; Based on the feature encoder, projecting the target update sequence into multiple subspaces for encoding to obtain encoded features; State estimation is performed based on the state estimator and the encoded features to obtain the target state indicator.

3. The battery multi-state joint estimation method according to claim 2, characterized in that: The at least two first convolutional neural networks include a first sampling convolutional neural network and a second sampling convolutional neural network, the convolution kernel size corresponding to the first sampling convolutional neural network is smaller than the convolution kernel size corresponding to the second sampling convolutional neural network, and the input original feature sequence is downsampled based on the feature extraction model to obtain at least two initial feature sequences, including: Downsampling the original feature sequence based on the first sampling convolutional neural network to obtain a first initial feature sequence, where the first initial feature sequence is used to characterize features with a smaller time resolution; The original feature sequence is downsampled based on the second sampling convolutional neural network to obtain a second initial feature sequence, where the second initial feature sequence is used to characterize features with a larger time resolution.

4. The battery multi-state joint estimation method according to claim 2, characterized in that: The cross-learning model includes at least two multi-head self-attention mechanism units and at least two second convolutional neural networks. Based on the cross-learning model, feature extraction and coupling relationship modeling are performed on the at least two initial feature sequences to obtain a target update sequence, including: Mapping the at least two initial feature sequences through the at least two multi-head self-attention mechanism units to obtain at least two hidden sequences; Performing feature size transformation on the at least two hidden sequences by using the at least two second convolutional neural networks to obtain at least two transformed hidden sequences; Perform interactive modeling based on the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences; The at least two updated sequences are concatenated to obtain the target updated sequence.

5. The battery multi-state joint estimation method according to claim 4, characterized in that: The at least two initial feature sequences include a first initial feature sequence and a second initial feature sequence, and the interactive modeling is performed according to the at least two initial feature sequences and the corresponding at least two transformed hidden sequences to obtain at least two updated sequences, including: Performing element-wise multiplication and addition and subtraction operations on the transformed hidden sequences corresponding to the first initial feature sequence and the second initial feature sequence to obtain a first updated sequence; Performing element-wise multiplication and addition and subtraction operations on the second initial feature sequence and the transformed hidden sequence corresponding to the first initial feature sequence to obtain a second updated sequence; The step of splicing the at least two updated sequences to obtain the target updated sequence includes: splicing the first updated sequence and the second updated sequence to obtain a spliced ​​updated sequence; Perform residual connection and batch normalization on the concatenated update sequence to obtain the target update sequence.

6. The battery multi-state joint estimation method according to claim 2, characterized in that: The feature encoder includes a multi-head attention mechanism unit and a distillation operation unit. Based on the feature encoder, the target update sequence is projected into multiple subspaces for encoding to obtain encoded features, including: The target update sequence is mapped to three tensors through multiple linear transformations by the multi-head attention mechanism unit to obtain multiple transformed vectors, each of the multiple transformed vectors includes a query vector, a key vector and a value vector, and the similarity between the query vector and the corresponding key vector is calculated for each transformed vector to obtain a weight coefficient of the value vector corresponding to each key vector, and after normalizing the weight coefficient of the value vector corresponding to each key vector, the weight coefficient and the corresponding value vector are weighted summed to obtain an attention value corresponding to each transformed vector, and the attention values ​​corresponding to each transformed vector are spliced ​​to obtain a spliced ​​feature; The distillation operation unit performs a distillation operation on the concatenated features to obtain the encoded features.

7. The battery multi-state joint estimation method according to claim 6, characterized in that: The distillation operation unit includes a first one-dimensional convolutional neural network and a pooling layer, and the distillation operation unit performs a distillation operation on the concatenated features to obtain the encoded features, including: Down-sampling the concatenated features based on the first one-dimensional convolutional neural network and the distillation operation of the pooling layer to obtain sampled features; The encoded features are obtained according to the sampled features.

8. The battery multi-state joint estimation method according to claim 2, characterized in that: The state estimator includes a deep neural network and a fully connected network, the deep neural network includes a two-layer network, and the state estimation is performed based on the state estimator and the encoded features to obtain the target state indicator, including: Extracting local features from the encoded features through a two-layer network of the deep neural network to obtain extracted features; The extracted features are passed through at least two fully connected layers of the fully connected network to obtain the target state indicator.

9. The battery multi-state joint estimation method according to claim 8, characterized in that: The target state indicator includes a battery state of charge, the state estimator includes a gated recurrent unit, and the extracted features are passed through at least two fully connected layers of the fully connected network to obtain the target state indicator, including: The extracted features are sequentially passed through the gated recurrent unit and the target fully connected layer to obtain the battery state of charge, wherein the target fully connected layer is a fully connected layer of the at least two fully connected layers, and the target fully connected layer corresponds to the battery state of charge.

10. The battery multi-state joint estimation method according to claim 1, characterized in that: The obtaining of the original feature sequence comprises: Performing standardization processing on the current charge-discharge cycle data to obtain standardized current charge-discharge cycle data; Position coding, global time coding and data coding are performed on the standardized current charge-discharge cycle data to obtain the original feature sequence, which includes a position matrix, a global time tag and a feature vector matrix.

11. The battery multi-state joint estimation method according to claim 10, characterized in that: The step of performing position coding, global time coding and data coding on the standardized current charge-discharge cycle data to obtain the original feature sequence includes: Converting the standardized position information of the current charge-discharge cycle data by a fixed sine function or a cosine function to obtain the position matrix; Acquire the global time information of the standardized current charge-discharge cycle data through a learnable position embedding vector to obtain the global time tag; Converting the characteristic information of the standardized current charge and discharge cycle data through a second one-dimensional convolutional neural network to obtain the characteristic vector matrix; The position matrix, the global time stamp and the feature vector matrix are combined to obtain the original feature sequence.

12. A battery multi-state joint estimation device, characterized in that: include: A feature acquisition module, used to acquire an original feature sequence, wherein the original feature sequence is obtained based on current charge and discharge cycle data of the battery; A feature processing module is used to input the original feature sequence into a trained data-driven model to obtain a target state indicator output by the trained data-driven model, wherein the target state indicator includes at least two types of battery state indicators, and the at least two types of battery state indicators have different time resolution characteristics, wherein the trained data-driven model obtains the target state indicator by extracting feature sequences with different time resolution characteristics from the original feature sequence for analysis.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the battery multi-state joint estimation method according to any one of claims 1 to 11 is implemented.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the battery multi-state joint estimation method according to any one of claims 1 to 11 is implemented.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the battery multi-state joint estimation method according to any one of claims 1 to 11 is implemented.

Citation Information

Cited By

  • System and method for rapidly estimating SOC (State of Charge) in combination with lightweight neural network

    CN119575199A

  • Power system disturbance lowest frequency prediction method based on physical information embedded layer

    CN119940107A

  • A power system disturbance lowest frequency prediction method based on a physical information embedded layer

    CN119940107B

  • Remaining service life prediction method and device and electronic equipment

    CN119959768A

  • Remaining service life prediction method, device and electronic equipment

    CN119959768B