Self-adaptive health sensing method for degradation full life of energy storage cell stack

Abstract

The invention discloses an energy storage cell stack degradation full-life self-adaptive health sensing method, which comprises the following steps of: firstly, constructing a circulating window for standardized analysis through feature fusion and circulating window construction, combining a dual-scale non-stationary degradation analysis architecture, and then carrying out decoupling characterization through a designed degradation composition decoupling algorithm, so as to realize self-adaptive health sensing of the degradation full-life of an energy storage cell stack. An input signal is decomposed into inherent characteristics which are not changed in the degradation process and drifting characteristics which are obviously changed in the degradation process, and real abnormal and normal degradation phenomena are distinguished through extracted degradation decoupling representation information, so that a self-adaptive monitoring scheme with degradation sensing capability for the full life cycle of the lithium battery stack is realized. According to the method, a self-adaptive health sensing scheme under the degeneration decoupling characterization and full life cycle is established for the first time, and degeneration and real abnormity are decoupled in detail by identifying aging behaviors, so that the fault detection rate is effectively improved, the false alarm rate is reduced, and practical support is provided for fine health management and safe operation of the energy storage cell stack.

Description

Technical Field

[0001] The present invention relates to the technical field of health management of energy storage power stations, and in particular to the technical field of continuous degradation of lithium-ion battery stacks in energy storage power stations during operation, and more particularly to a full-life adaptive health perception method for degradation of energy storage battery stacks. Background Art

[0002] Benefiting from higher energy density and less pollution, lithium-ion batteries have become the most widely used choice in energy storage systems. However, safety issues of lithium-ion batteries have also continued to emerge, and their safety issues have also attracted much attention. In recent decades, the rapid development of the Internet of Things and artificial intelligence technologies has spawned many data-based lithium-ion battery abnormality monitoring methods. In these mainstream methods, a normal range is determined through historical data. Once the operating data deviates from the preset range, the abnormal behavior of the lithium battery can be identified in time.

[0003] However, lithium battery modules inevitably experience performance degradation during long-term operation, such as a decrease in capacity and power, which can lead to deviations in their operating characteristics. As a result, the operating data will experience slow time-varying problems, and the monitoring models developed for new batteries may gradually become ineffective, causing a large number of false alarms. Although adaptive monitoring methods have been significantly developed in other industrial fields, these methods face challenges in terms of "whether to update". On the one hand, these methods have difficulty in distinguishing whether the degraded online samples represent normal data drift or anomalies. In normal data drift, updates are required, and in abnormal cases, updates should be terminated and alarms should be triggered. On the other hand, these methods rely on surface observations of online data to judge "whether to update", which leads to ambiguity in phenomenon identification and unreliability in judgment. In addition, these methods are not suitable for modeling lithium battery modules. Lithium battery modules usually undergo frequent charge and discharge cycles during operation, showing significant non-stationary characteristics. Existing methods have shown efficiency in separating non-stationary components from steady states, but they fail to further identify non-stationary trends driven by different factors, such as degradation or charge and discharge behavior. Therefore, how to correctly characterize the degradation process of lithium batteries throughout their life cycle and establish a reliable adaptive monitoring model remains an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a full-life adaptive health perception method for degradation of energy storage battery stacks in view of the shortcomings of the adaptive anomaly monitoring methods in the prior art in capturing non-stationary data trends caused by degradation and distinguishing anomalies.

[0005] The objective of the present invention is achieved through the following technical solution: a method for adaptive health perception of energy storage battery stack degradation over its entire life cycle, comprising the following steps: (1) Obtain the original offline data of the entire life cycle of the historical operation of the energy storage battery stack, and perform cluster feature fusion and cyclic window construction processing on the original offline data in the normal state to obtain fused feature sample data of multiple cyclic windows; (2) Using the fused feature sample data of the cycle window, the dual-scale non-stationary degradation analysis framework is used for analysis. First, at the inner cycle scale, the charging behavior or discharging behavior in each charge and discharge cycle is identified as an inner cycle non-stationary trend through the Gaussian mixture model. Then, at the cross-cycle scale, the degradation component decoupling algorithm is used to decouple the fused feature sample data to obtain the intrinsic projection matrix and the drift projection matrix. (3) Obtain the historical service data of the energy storage battery stack to be monitored, and perform cluster feature fusion processing on it to obtain the corresponding fused feature data; use the inherent projection matrix and the drift projection matrix to obtain the inherent features and drift features corresponding to the fused feature data, and calculate the monitoring statistics of the inherent features and drift features respectively through the Gaussian mixture model, and use kernel density estimation to obtain the control limits of the monitoring statistics corresponding to the inherent features and drift features; (4) When performing online adaptive abnormal monitoring on the energy storage battery stack to be monitored, the service data collected in real time is analyzed with the battery module as the basic unit, and the status of the energy storage battery stack to be monitored is judged according to the control limits of the monitoring statistics corresponding to the inherent characteristics and drift characteristics; and an adaptive model update strategy is designed to determine whether an update is needed based on the aging degree of the online service data.

[0006] Furthermore, the acquisition of original offline data of the entire life cycle of the historical operation of the energy storage battery stack specifically includes: The energy storage battery stack includes a plurality of battery modules, with the battery module being the smallest analysis unit. Each battery module is composed of a plurality of battery cells connected in series and parallel. Each battery cell is a lithium-ion battery, which is the smallest unit of energy storage. The service data of all battery modules during the entire life cycle of historical operation are collected. The service data of the battery module includes the service status variable data of each battery cell in the battery module. The service status variables include charge state indicators and measurement point variables. The measurement point variables include temperature, voltage, and current.

[0007] Furthermore, the cluster feature fusion and cyclic window construction processing are performed on the original offline data in the normal state to obtain fused feature sample data of multiple cyclic windows, specifically including: First, the original offline data in the normal state is subjected to cluster feature fusion processing, specifically: the service state variable of the battery module at any sampling time is recorded as ,in Indicates the total number of battery cells in the battery module. Represents the total number of service status variables of the battery cell; the service status variables of the battery cell are sequentially spliced ​​to obtain the individual characteristics of the battery cell, which is expressed as ,in represents the individual feature vector of the i-th battery cell in the battery module, Represents the jth service state variable of the ith battery cell in the battery module; the individual characteristics of all battery cells in the battery module are sequentially spliced ​​to obtain the individual characteristics of the battery module ; Calculate statistical indicators based on the individual characteristics of the battery module, and extract the cluster characteristics of the battery module based on the statistical indicators ,in Indicates A column vector of statistical indicators, Indicates the total number of statistical indicator types; individual characteristics of battery modules and its corresponding cluster characteristics Splice to form fusion features, expressed as ,in The total number of features representing the fused features; Then, the obtained fusion features are used to construct a cycle window to create a cycle window containing multiple charge and discharge cycles. All fused feature samples collected in the charge-discharge cycle are included in one cycle window, and fused feature sample data of multiple cycle windows are obtained, where The fusion feature sample data of the cycle window is recorded as , Indicates The total number of fused feature samples in the cycle window, Indicates the index of the first charge-discharge cycle in the cycle window, Indicates The number of fused feature samples in a charge-discharge cycle.

[0008] Furthermore, in the inner cycle scale, the charging behavior or discharging behavior in each charge and discharge cycle is identified as an inner cycle non-stationary trend through a Gaussian mixture model, specifically including: The charging or discharging behavior in each charge-discharge cycle is identified as an internal cycle non-stationary trend, and the Gaussian mixture model is used to describe the data distribution of each cycle window. Fusion feature sample data of cyclic windows , use the Gaussian mixture model to model it, and give the probability density function of the modeled Gaussian mixture model, which is expressed as:

[0009] in, represents the probability density function of the Gaussian mixture model, Indicates A single fusion feature sample data in a cycle window, , and Respectively represent The mixture weight, mean vector and covariance matrix of the mth Gaussian cost in the charge-discharge cycle, M represents the total number of Gaussian components, represents the Gaussian probability density function; where the mixing weight Satisfy constraints .

[0010] Furthermore, the degenerate component decoupling algorithm is used to decouple the fused feature sample data under the cross-cycle scale to obtain the intrinsic projection matrix and the drift projection matrix, which specifically includes:

[0011] First, a cross-loop intrinsic Gaussian mixture model is constructed based on the probability density function of the Gaussian mixture model of each loop window under the inner loop scale, so that the intrinsic features obtained by the original fusion feature sample data Similar data distribution is presented in different cycle windows; the probability density function of the cross-cycle intrinsic Gaussian mixture model is:

[0012] in, represents the probability density function of the cross-cyclic intrinsic Gaussian mixture model, , and denote the mixture weight, mean vector and covariance matrix of the mth Gaussian cost of the cross-cyclic intrinsic Gaussian mixture model respectively; Then the original model parameters of the cross-cyclic intrinsic Gaussian mixture model for each cyclic window are given by ,in Indicates The original model parameters of the cross-cyclic intrinsic Gaussian mixture model for the cyclic window, , and Respectively represent The original mixing weights, original mean vectors and original covariance matrices of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cycle window are obtained; the model parameters of the cross-cyclic intrinsic Gaussian mixture model of each cycle window are estimated by the expectation maximization algorithm. After Q iterative estimates, the final model parameters of the cross-cyclic intrinsic Gaussian mixture model of each cycle window are obtained as follows: , which is simplified to ,in , and Respectively represent The final mixing weights, final mean vector and final covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cyclic window; Secondly, the linear projection coefficient matrix is ​​solved according to the final model parameters of the cross-cyclic intrinsic Gaussian mixture model for all cycle windows , its optimization problem and constraints are equivalent to:

[0013] in, represents the trace of the matrix, represents the identity matrix, W is the total number of loop windows, and denote the mean vector average and covariance matrix average of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model, respectively. They are all intermediate parameters with no practical significance; Finally, the singular value decomposition method is used to solve the optimization problem under the constraints of the linear projection coefficient matrix, and the closed-form solution of the optimization problem in the above formula is obtained. The closed-form solution is feature vectors, in Select the eigenvectors corresponding to the smallest R eigenvalues ​​from the eigenvectors and use them as the intrinsic projection matrix, expressed as ;Will The remaining eigenvectors in the eigenvectors form the drift projection matrix, which is expressed as ;in express The rth coefficient vector in , R is the dimension of the intrinsic feature, is the dimension of the drift feature.

[0014] Furthermore, the model parameters of the cross-cycle intrinsic Gaussian mixture model of each cycle window are estimated by an expectation maximization algorithm, specifically including: First, set the expected step, specifically: In the cycle window, calculate the In the training iteration Fusion feature samples The posterior probability from the mth Gaussian component , and its calculation formula is:

[0015] in, , and Respectively represent The estimated value of the first The mixing weights, mean vector and covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cyclic window; Then set the maximum step, specifically: according to the In iteration The value of The Gaussian mixture model parameters for the iteration , and its calculation formula is:

[0016] in, , and Respectively represent The estimated value of the first The mixing weights, mean vector, and covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model for a cyclic window.

[0017] Furthermore, the step (3) specifically includes the following sub-steps: (3.1) Obtaining historical service data of the energy storage battery stack to be monitored; wherein the energy storage battery stack includes a plurality of battery modules, with the battery module being the smallest analysis unit, and the service data of the battery module includes service status variable data of each single battery in the battery module, and the service status variable includes a state of charge index and a measurement point variable; (3.2) Perform cluster feature fusion processing on the historical service data of each battery module to be monitored to obtain the corresponding fusion feature data ,in Indicates the total number of fused feature data of the battery module to be monitored; (3.3) Using the intrinsic projection matrix and the drift projection matrix Obtain the inherent characteristics and drift characteristics of the battery module to be monitored, and the calculation formulas are:

[0018]

[0019] in, and They respectively represent the inherent characteristics and drift characteristics of the battery module to be monitored; (3.4) Calculate the intrinsic features separately through the Gaussian mixture model and drift characteristics The monitoring statistic is the Bayesian inference distance, and its calculation formula is:

[0020]

[0021] in, is the Bayesian inference distance indicator, representing the inherent characteristics or drift characteristics Monitoring statistics of represents the Gaussian mixture model parameters of the mth Gaussian component, represents the mixing weight in the Gaussian mixture model parameters of the mth Gaussian component; When it indicates an inherent characteristic, When indicates drift characteristics; express To The Mahalanobis distance of the Gaussian components, Indicates the posterior probability corresponding to the inherent feature or drift feature; (3.5) The monitoring statistics of the inherent characteristics and drift characteristics calculated in step (3.4) are statistically analyzed using the kernel density estimation method to obtain the control limits of the monitoring statistics corresponding to the inherent characteristics and drift characteristics.

[0022] Furthermore, the step (4) specifically includes: When performing online adaptive abnormal monitoring on the energy storage battery stack to be monitored, the service data collected in real time is analyzed with the battery module as the basic unit. The service data of the battery module includes the service status variable data of each battery cell. The service status variables include the state of charge index and the measurement point variable. When a new service data sample arrives, the cluster feature fusion processing is first performed to obtain a new fusion feature vector. , and using the intrinsic projection matrix and the drift projection matrix Calculate its inherent characteristics and drift characteristics, and the calculation formulas are:

[0023]

[0024] in, and Represent the new fusion feature vectors The inherent characteristics and drift characteristics of Recalculate the inherent characteristics and drift characteristics The monitoring statistic is calculated as follows:

[0025] in, The monitoring statistic that represents the inherent characteristics of the new service data sample, A monitoring statistic representing the drift characteristics of a new sample of service data; represents the total number of Gaussian components in the intrinsic Gaussian mixture model, Represents the total number of Gaussian components in the drift Gaussian mixture model; represents the model parameters of the mth Gaussian component in the intrinsic Gaussian mixture model, Represents the model parameters of the mth Gaussian component in the drift Gaussian mixture model; express The Mahalanobis distance to the mth Gaussian component in the intrinsic Gaussian mixture model, express The Mahalanobis distance to the mth Gaussian component in the drift Gaussian mixture model; represents the posterior probability of the intrinsic Gaussian mixture model corresponding to the intrinsic feature distribution, Represents the posterior probability of the drift Gaussian mixture model corresponding to the drift feature; Then determine the monitoring statistic of the inherent characteristics of the calculated new service data sample Is it greater than the control limit of the monitoring statistic corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristics of the new service data sample? Is it greater than the monitoring statistic control limit corresponding to the drift characteristic? If is less than or equal to the control limit of the monitoring statistic corresponding to the inherent characteristic, and If the value is less than or equal to the control limit of the monitoring statistic corresponding to the drift characteristic, the state of the energy storage battery stack to be monitored is considered normal; otherwise, the state of the energy storage battery stack to be monitored is considered abnormal.

[0026] Furthermore, the design of the adaptive model update strategy determines whether an update is required based on the aging degree of the online service data, specifically including: For the intrinsic features, a Gaussian mixture model for monitoring is designed, which is called the intrinsic Gaussian mixture model; for the drift features, a recursive Gaussian mixture model is established, which is called the drift Gaussian mixture model; Design an adaptive model update strategy for the intrinsic Gaussian mixture model and the drifting Gaussian mixture model, specifically by analyzing the degradation components to determine whether the intrinsic Gaussian mixture model or the drifting Gaussian mixture model needs to be updated and alarmed; Among them, the adaptive model update strategy specifically includes: Case 1: When the monitoring statistic of the inherent characteristics of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristic is less than or equal to the monitoring statistic control limit corresponding to the drift characteristic, the state of the energy storage battery stack to be monitored at this time is regarded as a normal state, and new service data samples are collected to update the model parameters of the drift Gaussian mixture model. When the size of the collected data sample block reaches the preset value When , the model parameters of the drift Gaussian mixture model and the monitoring statistic control limits corresponding to the drift characteristics are updated recursively; Case 2: When the calculated monitoring statistic of the inherent characteristics of the service data sample is greater than the monitoring statistic control limit corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristics is less than or equal to the monitoring statistic control limit corresponding to the drift characteristics, the state of the energy storage battery stack to be monitored at this time is regarded as an abnormal state, the update of the drift Gaussian mixture model is stopped and an alarm is triggered; Case 3: When the monitoring statistic of the inherent characteristic of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the inherent characteristic, and the monitoring statistic of the drift characteristic is greater than the monitoring statistic control limit corresponding to the drift characteristic, the collected service data samples are used for a recursive strategy, and then the monitoring statistic of the drift characteristic is recalculated and the monitoring statistic control limit of the drift characteristic is determined again; thereafter, if the monitoring statistic of the drift characteristic of the calculated service data sample is still greater than the monitoring statistic control limit corresponding to the drift characteristic, an alarm is triggered; if the monitoring statistic of the drift characteristic of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the drift characteristic, it is considered that the energy storage battery stack to be monitored is in a normal state; Case 4: When the calculated monitoring statistic of the inherent characteristic of the service data sample is greater than the monitoring statistic control limit corresponding to the inherent characteristic, and the monitoring statistic of the drift characteristic is greater than the monitoring statistic control limit corresponding to the drift characteristic, the update of the Gaussian mixture model is stopped and an alarm is triggered.

[0027] The beneficial effects of the present invention are as follows: the present invention establishes for the first time a degradation decoupling characterization and an adaptive health perception scheme under the whole life cycle, and by identifying aging behavior, performs a detailed decoupling of degradation and real abnormalities, provides a basis for model updating, and thereby effectively improves the fault detection rate and reduces the false alarm rate, and provides practical support for the fine health management and safe operation of energy storage battery stacks; the present invention can achieve clear observation of the degradation process while jointly identifying and processing the non-stationary trends caused by charging behavior or discharging behavior and aging; on this basis, the present invention establishes an adaptive monitoring scheme with reliable distinction between decay and abnormalities, achieves a clear representation of the degradation process of lithium battery cells, and provides a comprehensive update strategy for the monitoring model based on aging perception, which maintains a high sensitivity recognition capability for real abnormalities after the lithium battery ages. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the adaptive health perception method for energy storage battery stack degradation throughout its life cycle of the present invention; Figure 2 The architecture flow chart of the adaptive model update strategy of the present invention; Figure 3This is an example diagram of the monitoring results of the present invention in Example 1; wherein, Figure 3 (a) is the monitoring result diagram of the inherent characteristics. Figure 3 (b) is the monitoring result diagram of drift characteristics; Figure 4 This is an example diagram of the monitoring results of the present invention on Example 2; wherein, Figure 4 (a) is the monitoring result diagram of the inherent characteristics. Figure 4 (b) is the monitoring result diagram of drift characteristics; Figure 5 This is a monitoring effect diagram of the Gaussian mixture model method of the present invention on Example 1; Figure 6 This is a monitoring effect diagram of the slow feature analysis method of the present invention in Example 1; Figure 7 This is a monitoring effect diagram of the Gaussian mixture model method of the present invention on Example 2; Figure 8 This is a diagram showing the monitoring effect of the slow feature analysis method of the present invention on Example 2. DETAILED DESCRIPTION

[0029] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0030] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0031] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0032] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0033] See also Figure 1 The energy storage battery stack degradation full life self-adaptive health perception method of the present invention specifically comprises the following steps: (1) Obtain the original offline data of the entire life cycle of the historical operation of the energy storage battery stack, and perform cluster feature fusion and cyclic window construction processing on the original offline data in the normal state to obtain fused feature sample data of multiple cyclic windows for subsequent analysis.

[0034] It should be understood that when performing cluster feature fusion and cyclic window construction processing on the original offline data, the above processing is only performed on the original offline data in the normal state, and the discrete data of abnormal / fault conditions are not included. This is because the normal state data is used for offline modeling, and the control limits are defined according to the normal state data, which is convenient for the subsequent online application stage. By judging whether the sample exceeds the limit, the health of the battery can be judged online.

[0035] Furthermore, the original offline data of the entire life cycle of the historical operation of the energy storage battery stack is obtained, specifically including: the energy storage battery stack includes multiple battery modules, with the battery module as the smallest analysis unit, each battery module is composed of multiple battery cells connected in series and parallel, each battery cell is a lithium-ion battery, and is the smallest unit of energy storage; the service data of all battery modules during the entire life cycle of historical operation are collected, the service data of the battery module includes the service status variable data of each battery cell in the battery module, the service status variable includes the charge state index and the measurement point variable, and the measurement point variable includes temperature, voltage, current, etc.

[0036] Exemplarily, in this embodiment, offline modeling samples are obtained: 1.5 years of original offline data are collected from a battery module as offline training data, and the battery module is composed of 16 battery cells connected in series. Each battery cell is a lithium-ion battery, which is the smallest unit of energy storage; the sampling interval is set to 1 minute, and the service data of the battery module is collected at regular intervals. The service data of the battery module includes service status variable data of the 16 battery cells, and the service status variables include charge state indicators and measurement point variables (such as temperature, voltage, current, etc.).

[0037] Furthermore, cluster feature fusion and cyclic window construction processing are performed on the original offline data in the normal state to obtain fused feature sample data of multiple cyclic windows, specifically including: firstly, cluster feature fusion processing is performed on the original offline data in the normal state, specifically: the service state variable of the battery module at any sampling time is recorded as ,in Indicates the total number of battery cells in the battery module. Represents the total number of service status variables of the battery cell; the service status variables of the battery cell are sequentially spliced ​​to obtain the individual characteristics of the battery cell, which is expressed as ,in represents the individual feature vector of the i-th battery cell in the battery module, Represents the jth service state variable of the ith battery cell in the battery module; the individual characteristics of all battery cells in the battery module are sequentially spliced ​​to obtain the individual characteristics of the battery module In addition, in order to describe the overall state of the entire battery module, statistical indicators such as mean and variance are calculated based on the individual characteristics of the battery module, and the cluster characteristics of the battery module are extracted based on the statistical indicators. ,in Indicates A column vector of statistical indicators, Indicates the total number of statistical indicator types; individual characteristics of battery modules and its corresponding cluster characteristics Splice to form fusion features, expressed as ,in Indicates the total number of features of the fused features. In this embodiment, is 16; is 3; four statistical indicators, namely mean, variance, maximum value and minimum value, are used to extract the cluster characteristics of the battery module, namely ; is 60. Then the cycle window is constructed for the obtained fusion features to create a cycle window containing multiple charge and discharge cycles as the basic unit for subsequent analysis. All fused feature samples collected in the charge-discharge cycle are included in one cycle window, and fused feature sample data of multiple cycle windows are obtained, where The fusion feature sample data of the cycle window is recorded as , Indicates The total number of fused feature samples in the cycle window, Indicates the index of the first charge-discharge cycle in the cycle window, Indicates In this embodiment, 4 cycle windows are evenly sampled from the offline data of the battery module, and each window contains 12 charge and discharge cycles, that is, is 12.

[0038] (2) The fused feature sample data of the cycle window is analyzed through a dual-scale non-stationary degradation analysis framework. First, at the inner cycle scale, the charging behavior or discharging behavior in each charge and discharge cycle is identified as an inner cycle non-stationary trend through a Gaussian mixture model (GMM). Then, at the cross-cycle scale, the fused feature sample data is decoupled using a degradation component decoupling algorithm to obtain an intrinsic projection matrix and a drift projection matrix. The dual-scale non-stationary degradation analysis framework includes an inner cycle scale and a cross-cycle scale.

[0039] Furthermore, at the inner cycle scale, the charging behavior or discharging behavior in each charge-discharge cycle is identified as an inner cycle non-stationary trend through the Gaussian mixture model, specifically including: identifying the charging behavior or discharging behavior in each charge-discharge cycle as an inner cycle non-stationary trend, using the Gaussian mixture model to describe the data distribution of each cycle window, and for the first Fusion feature sample data of cyclic windows , use the Gaussian mixture model to model it, and give the probability density function of the modeled Gaussian mixture model, which is expressed as:

[0040] in, represents the probability density function of the Gaussian mixture model, Indicates A single fusion feature sample data in a cycle window, , and Respectively represent The mixture weight, mean vector and covariance matrix of the mth Gaussian cost in the charge-discharge cycle, M represents the total number of Gaussian components, represents the Gaussian probability density function; where the mixing weight Satisfy constraints .

[0041] Furthermore, at the cross-cycle scale, the degenerate component decoupling algorithm is used to decouple the fused feature sample data to obtain the intrinsic projection matrix and the drift projection matrix, which specifically includes: firstly, a set of original feature fusion sample data is found, and a cross-cycle intrinsic Gaussian mixture model is constructed based on the probability density function of the Gaussian mixture model of each cycle window at the corresponding inner cycle scale, so that the intrinsic features obtained from the original fused feature sample data Similar data distribution is presented in different cycle windows; the cross-cycle inherent Gaussian mixture model can reveal the inherent characteristics of the fused feature sample data in different cycle windows, and the probability density function of the cross-cycle inherent Gaussian mixture model is:

[0042] in, represents the probability density function of the cross-cyclic intrinsic Gaussian mixture model, , and denote the mixture weight, mean vector and covariance matrix of the mth Gaussian cost of the cross-cycle intrinsic Gaussian mixture model respectively. Then the original model parameters of the cross-cycle intrinsic Gaussian mixture model for each cycle window are given as ,in Indicates The original model parameters of the cross-cyclic intrinsic Gaussian mixture model for the cyclic window, , and Respectively represent The original mixing weights, original mean vectors and original covariance matrices of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cycle window are obtained; the model parameters of the cross-cyclic intrinsic Gaussian mixture model of each cycle window are estimated by the expectation maximization algorithm. After Q iterative estimates, the final model parameters of the cross-cyclic intrinsic Gaussian mixture model of each cycle window are obtained as follows: , which is simplified to ,in , and Respectively represent The final mixing weights, final mean vectors, and final covariance matrices of the mth Gaussian components of the cross-cycle intrinsic Gaussian mixture model of the cycle window. After Q parameter estimations, the original data distribution can be revealed by the cross-cycle intrinsic Gaussian mixture model. Secondly, assume that the linear projection coefficient matrix from the original fusion feature sample data to the intrinsic features and drift features is , use the corresponding linear projection coefficient matrix to obtain the corresponding intrinsic features and drift features, that is, solve the linear projection coefficient matrix according to the final model parameters of the cross-cycle intrinsic Gaussian mixture model of all cycle windows , its optimization problem and constraints are equivalent to:

[0043] in, represents the trace of the matrix, represents the identity matrix, W is the total number of loop windows, and denote the mean vector average and covariance matrix average of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model, respectively. are all intermediate parameters with no practical significance. Finally, the singular value decomposition method is used to solve the optimization problem under the constraints of the linear projection coefficient matrix, and the closed-form solution of the optimization problem in the above formula is obtained. The closed-form solution is feature vectors, in Select the eigenvectors corresponding to the smallest R eigenvalues ​​from the eigenvectors and use them as the intrinsic projection matrix, expressed as ;Will The remaining eigenvectors in the eigenvectors form the drift projection matrix, which is expressed as ;in express The rth coefficient vector in , R is the dimension of the intrinsic feature, is the dimension of the drift feature.

[0044] It should be noted that the intrinsic projection matrix is ​​obtained and the drift projection matrix Then, use the intrinsic projection matrix Get fusion feature sample data Inherent characteristics , using the drift projection matrix Get fusion feature sample data Drift characteristics Among them, the intrinsic characteristics and drift characteristics obtained show significant deviations in the degradation process of lithium-ion batteries; the intrinsic characteristics are used to characterize the long-term constant characteristics of lithium-ion batteries, and the drift characteristics are used to characterize the degradation behavior of lithium-ion batteries. In addition, through the intrinsic projection matrix and the drift projection matrix The inherent characteristics and drift characteristics of the fused feature sample data can be obtained, and its data distribution can also be described by a Gaussian mixture model, where the model parameters of the cross-cycle inherent Gaussian mixture model are: , , ; The model parameters of the cross-cycle drift Gaussian mixture model are: , , .

[0045] Furthermore, the model parameters of the cross-cycle intrinsic Gaussian mixture model of each cycle window are estimated by the expectation maximization algorithm, which specifically includes: first setting the expected step, specifically: In the cycle window, calculate the In the training iteration Fusion feature samples The posterior probability from the mth Gaussian component , and its calculation formula is:

[0046] in, , and Respectively represent The estimated value of the first The mixing weights, mean vectors and covariance matrices of the mth Gaussian component of the cross-cycle intrinsic Gaussian mixture model of the cycle window. Then set the maximum step, specifically: according to the In iteration The value of The Gaussian mixture model parameters for the iteration , and its calculation formula is:

[0047] in, , and Respectively represent The estimated value of the first The mixing weights, mean vector, and covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model for a cyclic window.

[0048] (3) Obtain the historical service data of the energy storage battery stack to be monitored, and perform cluster feature fusion processing on it to obtain the corresponding fused feature data; use the inherent projection matrix and the drift projection matrix to obtain the inherent features and drift features corresponding to the fused feature data, and calculate the monitoring statistics of the inherent features and drift features respectively through the Gaussian mixture model, and use kernel density estimation to obtain the control limits of the monitoring statistics corresponding to the inherent features and drift features.

[0049] (3.1) Obtain the historical service data of the energy storage battery stack to be monitored. Similarly, the energy storage battery stack includes multiple battery modules, with the battery module as the smallest analysis unit. The service data of the battery module includes the service status variable data of each single battery in the battery module. The service status variable includes the charge state index and the measurement point variable, which is the same as the offline data.

[0050] (3.2) Perform cluster feature fusion processing on the historical service data of each battery module to be monitored to obtain the corresponding fusion feature data ,in Indicates the total number of fused feature data of the battery module to be monitored.

[0051] (3.3) Using the intrinsic projection matrix and the drift projection matrix Obtain the inherent characteristics and drift characteristics of the battery module to be monitored, and the calculation formulas are:

[0052]

[0053] in, and They respectively represent the inherent characteristics and drift characteristics of the battery module to be monitored.

[0054] (3.4) Calculate the intrinsic features separately through the Gaussian mixture model and drift characteristics The monitoring statistic is the Bayesian inference distance, and its calculation formula is:

[0055]

[0056] in, is the Bayesian inference distance indicator, representing the inherent characteristics or drift characteristics Monitoring statistics of represents the Gaussian mixture model parameters of the mth Gaussian component, represents the mixing weight in the Gaussian mixture model parameters of the mth Gaussian component; When it indicates an inherent characteristic, When indicates drift characteristics; express To The Mahalanobis distance of the Gaussian components, Represents the posterior probability corresponding to the inherent feature or drift feature.

[0057] (3.5) The monitoring statistics of the inherent characteristics and drift characteristics calculated in step (3.4) are statistically analyzed using a kernel density estimation method to obtain the monitoring statistic control limits corresponding to the inherent characteristics and drift characteristics. In this embodiment, the confidence levels of the monitoring statistic control limits corresponding to the inherent characteristics and drift characteristics are both 99.9%.

[0058] (4) When performing online adaptive abnormal monitoring on the energy storage battery stack to be monitored, the service data collected in real time is analyzed by taking the battery module as the basic analysis unit, and judging the status of the energy storage battery stack to be monitored according to the control limits of the monitoring statistics corresponding to the inherent characteristics and drift characteristics; and designing an adaptive model update strategy to judge whether an update is needed based on the aging degree of the online service data, such as Figure 1 shown.

[0059] Specifically, when performing online adaptive abnormal monitoring on the monitored energy storage battery stack, the service data collected in real time is analyzed with the battery module as the basic unit. The service data of the battery module includes the service status variable data of each battery cell. The service status variables include the state of charge index and the measurement point variable, which are consistent with the offline data. When a new service data sample arrives, the cluster feature fusion processing is first performed to obtain a new fusion feature vector , and using the intrinsic projection matrix and the drift projection matrix Calculate its inherent characteristics and drift characteristics, and the calculation formulas are:

[0060]

[0061] in, and Represent the new fusion feature vectors The inherent characteristics and drift characteristics of and drift characteristics The monitoring statistic is calculated as follows:

[0062] in, The monitoring statistic that represents the inherent characteristics of the new service data sample, A monitoring statistic representing the drift characteristics of a new sample of service data; represents the total number of Gaussian components in the intrinsic Gaussian mixture model, Represents the total number of Gaussian components in the drift Gaussian mixture model; represents the model parameters of the mth Gaussian component in the intrinsic Gaussian mixture model, Represents the model parameters of the mth Gaussian component in the drift Gaussian mixture model; express The Mahalanobis distance to the mth Gaussian component in the intrinsic Gaussian mixture model, express The Mahalanobis distance to the mth Gaussian component in the drift Gaussian mixture model; represents the posterior probability of the intrinsic Gaussian mixture model corresponding to the intrinsic feature distribution, The posterior probability of the drift Gaussian mixture model corresponding to the drift feature is represented; the inherent Gaussian mixture model is a Gaussian mixture model designed for monitoring the inherent features, and the drift Gaussian mixture model is a recursive Gaussian mixture model established for the drift features. In this embodiment, the size of m is determined by the Bayesian information criterion. Then, the monitoring statistic of the inherent features of the calculated new service data sample is determined. Is it greater than the control limit of the monitoring statistic corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristics of the new service data sample? Is it greater than the monitoring statistic control limit corresponding to the drift characteristic? If is less than or equal to the control limit of the monitoring statistic corresponding to the inherent characteristic, and If the value is less than or equal to the control limit of the monitoring statistic corresponding to the drift characteristic, the state of the energy storage battery stack to be monitored is considered normal; otherwise, in any other case, the state of the energy storage battery stack to be monitored is considered abnormal.

[0063] It should be understood that the modeling of the inherent Gaussian mixture model and the drift Gaussian mixture model is similar to that of the cross-cycle inherent Gaussian mixture model and the cross-cycle drift Gaussian mixture model, and the model parameters included are also similar, both of which include mixing weights, mean vectors and covariance matrices, but the values ​​of their respective model parameters are different.

[0064] Furthermore, an adaptive model update strategy is designed to determine whether an update is needed based on the aging degree of the online service data, such as Figure 2 As shown, specifically, for the inherent features, a Gaussian mixture model for monitoring is designed, which is called the inherent Gaussian mixture model; for the drift features, a recursive Gaussian mixture model is established, which is called the drift Gaussian mixture model. An adaptive model update strategy is designed for the inherent Gaussian mixture model and the drift Gaussian mixture model, specifically by analyzing the degradation components to determine whether the inherent Gaussian mixture model or the drift Gaussian mixture model needs to be updated and alarmed. Among them, the adaptive model update strategy specifically includes the following four main situations and their corresponding model update strategies:

[0065] Case 1: When the monitoring statistic of the inherent characteristics of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristic is less than or equal to the monitoring statistic control limit corresponding to the drift characteristic, the state of the energy storage battery stack to be monitored at this time is regarded as a normal state, and new service data samples are collected to update the model parameters of the drift Gaussian mixture model. When the size of the collected data sample block reaches the preset value When , the model parameters of the drift Gaussian mixture model and the monitoring statistic control limits corresponding to the drift characteristics are updated recursively.

[0066] Furthermore, the model parameters of the drift Gaussian mixture model and the monitoring statistic control limits corresponding to the drift characteristics are updated recursively. Specifically, in the In the iterations, the mixing weights will be based on each drift feature in the data block refresh:

[0067] in, is the forgetting ratio, represents the parameters in the bth iteration. The mixing weights are normalized to satisfy the constraint The mean and covariance of the mth Gaussian component are updated as:

[0068] in, represents the mean of the mth Gaussian component in the bth iteration, represents the covariance of the m-th Gaussian component in the b-th iteration. In this embodiment, is 200, is 0.01.

[0069] Case 2: When the calculated monitoring statistic of the inherent characteristics of the service data sample is greater than the monitoring statistic control limit corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristic is less than or equal to the monitoring statistic control limit corresponding to the drift characteristic, this indicates that an abnormal situation that breaks the long-term stable operation law may have occurred. The state of the energy storage battery stack to be monitored at this time is regarded as an abnormal state, the update of the drift Gaussian mixture model is stopped and an alarm is triggered.

[0070] Case 3: When the monitoring statistic of the inherent characteristic of the calculated service data sample is less than or equal to the control limit of the monitoring statistic corresponding to the inherent characteristic, and the monitoring statistic of the drift characteristic is greater than the control limit of the monitoring statistic corresponding to the drift characteristic, it indicates that the long-term stable operation law remains normal, but the drift part has a significant deviation. In this case, there are two possibilities: one is that the operation deviation caused by the decay obviously exceeds the control limit, and the other is that an abnormality occurs, resulting in an abnormal data feature of the drift characteristic. Therefore, it is necessary to further identify the operating state. Specifically, a recursive strategy is performed using the collected service data samples, and then the monitoring statistic of the drift characteristic is recalculated and the control limit of the monitoring statistic of the drift characteristic is determined again; thereafter, if the monitoring statistic of the drift characteristic of the calculated service data sample is still greater than the control limit of the monitoring statistic corresponding to the drift characteristic, an alarm is triggered; if the monitoring statistic of the drift characteristic of the calculated service data sample is less than or equal to the control limit of the monitoring statistic corresponding to the drift characteristic, the energy storage battery stack to be monitored is considered to be in a normal state.

[0071] Case 4: When the calculated monitoring statistic of the inherent characteristic of the service data sample is greater than the monitoring statistic control limit corresponding to the inherent characteristic, and the monitoring statistic of the drift characteristic is greater than the monitoring statistic control limit corresponding to the drift characteristic, it indicates that abnormal operation may have occurred in the inherent part and drift part of the energy storage battery stack to be monitored, and the update of the Gaussian mixture model is stopped and an alarm is triggered.

[0072] Exemplarily, the data of two different battery modules (ie, battery module 2 and battery module 3) are used to perform abnormal monitoring and analysis to illustrate the effect of the method of the present invention.

[0073] A total of 8331 non-aged normal operation samples were collected from battery module 2 and battery module 3 for initialization of the monitoring model.

[0074] In terms of preventing false alarms for normal aging samples, Example 1 contains 1124 samples from battery module 2, collected after one year of operation, to test the adaptive monitoring scheme. Figure 3 (a) and Figure 3 As shown in (b), although there is no adaptive scheme for the statistics of the inherent features, relatively few false alarms are observed, which proves the reliable invariance of the inherent features during the aging process of the battery module. At the same time, the drift features also adapt to the aging situation by continuously updating the model, thus avoiding false alarms caused by aging. In terms of abnormality monitoring, Example 2 contains 1443 samples from battery module 3. The initial stage of these samples contains the gradual abnormality of the voltage of a specific cell starting from the 812th sample. Figure 4 (a) and Figure 4 As shown in (b), the vertical dotted line indicates the time when the anomaly starts. The method of the present invention can timely distinguish the abnormal situation of abnormal increase of voltage caused by different causes. Therefore, the statistics based on the inherent characteristics timely identify the anomaly. In addition, the anomaly monitoring based on the inherent characteristics avoids the improper adaptation of the changing part to the abnormal sample.

[0075] In order to more clearly demonstrate the superiority of the method described in the present invention in the abnormal monitoring task, the Gaussian mixture model (GMM) method and the slow feature analysis (SFA) method are selected here for comparison with the method described in the present invention. The monitoring statistic BID of the GMM method on two instances is used to verify the fault monitoring effect of the GMM method. The results are as follows: Figure 5 and Figure 7 As shown; the statistical SFA method is used to calculate the significant difference between the two examples. To verify the fault monitoring effect of the SFA method, the results are as follows Figure 6 and Figure 8 As shown in the figure, it can be found that in Example 1, the GMM method will have some false alarms from about the 700th to the 800th samples due to the data offset caused by decay; in Example 2, both traditional methods failed to detect the fault and alarm in time.

[0076] For a more intuitive comparison, Table 1 lists the false alarm rate and monitoring accuracy of different methods on instance 1, and Table 2 lists the false alarm rate and monitoring accuracy of different methods on instance 2.

[0077] Table 1: False alarm rate and monitoring accuracy of different methods on Example 1

[0078] Table 2: False negative rate and monitoring accuracy of different methods on Example 2

[0079] As shown in Table 1, in Example 1, the method of the present invention can effectively reduce the abnormal false alarms caused by aging compared with other methods. As shown in Table 2, in Example 2, the method of the present invention can more accurately detect abnormalities compared with other methods, and the number of missed samples is significantly reduced. These two examples prove the effectiveness of the method of the present invention.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. 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 full-life adaptive health perception method for energy storage battery stack degradation, characterized in that: The following steps are involved: (1) Obtain the original offline data of the entire life cycle of the historical operation of the energy storage battery stack, and perform cluster feature fusion and cyclic window construction processing on the original offline data in the normal state to obtain fused feature sample data of multiple cyclic windows; (2) Using the fused feature sample data of the cycle window, the dual-scale non-stationary degradation analysis framework is used for analysis. First, at the inner cycle scale, the charging behavior or discharging behavior in each charge and discharge cycle is identified as an inner cycle non-stationary trend through the Gaussian mixture model. Then, at the cross-cycle scale, the degradation component decoupling algorithm is used to decouple the fused feature sample data to obtain the intrinsic projection matrix and the drift projection matrix. (3) Obtain the historical service data of the energy storage battery stack to be monitored, and perform cluster feature fusion processing on it to obtain the corresponding fusion feature data; The inherent features and drift features corresponding to the fused feature data are obtained using the inherent projection matrix and the drift projection matrix, and the monitoring statistics of the inherent features and drift features are calculated respectively through the Gaussian mixture model, and the control limits of the monitoring statistics corresponding to the inherent features and drift features are obtained using kernel density estimation; (4) When performing online adaptive abnormal monitoring on the energy storage battery stack to be monitored, the service data collected in real time is analyzed with the battery module as the basic unit, and the status of the energy storage battery stack to be monitored is judged according to the control limits of the monitoring statistics corresponding to the inherent characteristics and drift characteristics; and an adaptive model update strategy is designed to determine whether an update is needed based on the aging degree of the online service data.

2. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 1 is characterized in that: The acquisition of the original offline data of the entire life cycle of the historical operation of the energy storage battery stack specifically includes: The energy storage battery stack includes a plurality of battery modules, with the battery module being the smallest analysis unit. Each battery module is composed of a plurality of battery cells connected in series and parallel. Each battery cell is a lithium-ion battery, which is the smallest unit of energy storage. The service data of all battery modules during the entire life cycle of historical operation are collected. The service data of the battery module includes the service status variable data of each battery cell in the battery module. The service status variables include charge state indicators and measurement point variables. The measurement point variables include temperature, voltage, and current.

3. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 1 is characterized in that: The cluster feature fusion and cyclic window construction processing are performed on the original offline data in the normal state to obtain fused feature sample data of multiple cyclic windows, specifically including: First, the original offline data in the normal state is subjected to cluster feature fusion processing, specifically: the service state variable of the battery module at any sampling time is recorded as ,in Indicates the total number of battery cells in the battery module. Represents the total number of service status variables of the battery cell; the service status variables of the battery cell are sequentially spliced ​​to obtain the individual characteristics of the battery cell, which is expressed as ,in represents the individual feature vector of the i-th battery cell in the battery module, Represents the jth service state variable of the ith battery cell in the battery module; the individual characteristics of all battery cells in the battery module are sequentially spliced ​​to obtain the individual characteristics of the battery module ; Calculate statistical indicators based on the individual characteristics of the battery module, and extract the cluster characteristics of the battery module based on the statistical indicators ,in Indicates A column vector of statistical indicators, Indicates the total number of statistical indicator types; individual characteristics of battery modules and its corresponding cluster characteristics Splice to form fusion features, expressed as ,in The total number of features representing the fused features; Then, the obtained fusion features are used to construct a cycle window to create a cycle window containing multiple charge and discharge cycles. All fusion feature samples collected in the charge-discharge cycle are included in one cycle window, and fusion feature sample data of multiple cycle windows are obtained, where the first The fusion feature sample data of the cycle window is recorded as , Indicates The total number of fused feature samples in the cycle window, Indicates the index of the first charge-discharge cycle in the cycle window, Indicates The number of fused feature samples in a charge-discharge cycle.

4. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 1 is characterized in that: In the inner cycle scale, the charging behavior or discharging behavior in each charge and discharge cycle is identified as an inner cycle non-stationary trend through a Gaussian mixture model, specifically including: The charging or discharging behavior in each charge-discharge cycle is identified as an internal cycle non-stationary trend, and the Gaussian mixture model is used to describe the data distribution of each cycle window. Fusion feature sample data of cyclic windows , use the Gaussian mixture model to model it, and give the probability density function of the modeled Gaussian mixture model, which is expressed as: ； in, represents the probability density function of the Gaussian mixture model, Indicates A single fusion feature sample data in a cycle window, , and Respectively represent The mixture weight, mean vector and covariance matrix of the mth Gaussian cost in the charge-discharge cycle, M represents the total number of Gaussian components, represents the Gaussian probability density function; where the mixing weight Satisfy constraints .

5. The method for adaptive health perception of energy storage battery stack degradation over its entire life cycle according to claim 1 is characterized in that: The method uses a degenerate component decoupling algorithm to decouple the fused feature sample data under the cross-cycle scale to obtain an intrinsic projection matrix and a drift projection matrix, specifically including: First, a cross-loop intrinsic Gaussian mixture model is constructed based on the probability density function of the Gaussian mixture model of each loop window under the inner loop scale, so that the intrinsic features obtained by the original fusion feature sample data Similar data distribution is presented in different cycle windows; the probability density function of the cross-cycle intrinsic Gaussian mixture model is: ； in, represents the probability density function of the cross-cyclic intrinsic Gaussian mixture model, , and denote the mixture weight, mean vector and covariance matrix of the mth Gaussian cost of the cross-cyclic intrinsic Gaussian mixture model respectively; Then the original model parameters of the cross-cyclic intrinsic Gaussian mixture model for each cyclic window are given by ,in Indicates The original model parameters of the cross-cyclic intrinsic Gaussian mixture model for the cyclic window, , and Respectively represent The original mixing weights, original mean vectors and original covariance matrices of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cycle window are obtained; the model parameters of the cross-cyclic intrinsic Gaussian mixture model of each cycle window are estimated by the expectation maximization algorithm. After Q iterative estimates, the final model parameters of the cross-cyclic intrinsic Gaussian mixture model of each cycle window are obtained as follows: , which is simplified to ,in , and Respectively represent The final mixing weights, final mean vector and final covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cyclic window; Secondly, the linear projection coefficient matrix is ​​solved according to the final model parameters of the cross-cyclic intrinsic Gaussian mixture model for all cycle windows , its optimization problem and constraints are equivalent to: ； in, represents the trace of the matrix, represents the identity matrix, W is the total number of loop windows, and denote the mean vector average and covariance matrix average of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model, respectively. They are all intermediate parameters with no practical significance; Finally, the singular value decomposition method is used to solve the optimization problem under the constraints of the linear projection coefficient matrix, and the closed-form solution of the optimization problem in the above formula is obtained. The closed-form solution is feature vectors, in Select the eigenvectors corresponding to the smallest R eigenvalues ​​from the eigenvectors and use them as the intrinsic projection matrix, expressed as ;Will The remaining eigenvectors in the eigenvectors form the drift projection matrix, which is expressed as ;in express The rth coefficient vector in , R is the dimension of the intrinsic feature, is the dimension of the drift feature.

6. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 5, characterized in that: The model parameters of the cross-cycle intrinsic Gaussian mixture model of each cycle window are estimated by an expectation maximization algorithm, specifically including: First, set the expected step, specifically: In the cycle window, calculate the In the training iteration Fusion feature samples The posterior probability from the mth Gaussian component , and its calculation formula is: ； in, , and Respectively represent The estimated value of the first The mixing weights, mean vector and covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model of the cyclic window; Then set the maximum step, specifically: according to the In iteration The value of The Gaussian mixture model parameters for the iteration , and its calculation formula is: ； in, , and Respectively represent The estimated value of the first The mixing weights, mean vector, and covariance matrix of the mth Gaussian component of the cross-cyclic intrinsic Gaussian mixture model for a cyclic window.

7. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 1, characterized in that: The step (3) specifically includes the following sub-steps: (3.1) Obtaining historical service data of the energy storage battery stack to be monitored; wherein the energy storage battery stack includes a plurality of battery modules, with the battery module being the smallest analysis unit, and the service data of the battery module includes service status variable data of each single battery in the battery module, and the service status variable includes a state of charge index and a measurement point variable; (3.2) Perform cluster feature fusion processing on the historical service data of each battery module to be monitored to obtain the corresponding fusion feature data ,in Indicates the total number of fused feature data of the battery module to be monitored; (3.3) Using the intrinsic projection matrix and the drift projection matrix Obtain the inherent characteristics and drift characteristics of the battery module to be monitored, and the calculation formulas are: ； ； in, and They respectively represent the inherent characteristics and drift characteristics of the battery module to be monitored; (3.4) Calculate the intrinsic features separately through the Gaussian mixture model and drift characteristics The monitoring statistic is the Bayesian inference distance, and its calculation formula is: ； ； in, is the Bayesian inference distance indicator, representing the inherent characteristics or drift characteristics Monitoring statistics of represents the Gaussian mixture model parameters of the mth Gaussian component, represents the mixing weight in the Gaussian mixture model parameters of the mth Gaussian component; When it indicates an inherent characteristic, When indicates drift characteristics; express To The Mahalanobis distance of the Gaussian components, Indicates the posterior probability corresponding to the inherent feature or drift feature; (3.5) The monitoring statistics of the inherent characteristics and drift characteristics calculated in step (3.4) are statistically analyzed using the kernel density estimation method to obtain the control limits of the monitoring statistics corresponding to the inherent characteristics and drift characteristics.

8. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 1, characterized in that: The step (4) specifically includes: When performing online adaptive abnormal monitoring on the energy storage battery stack to be monitored, the service data collected in real time is analyzed with the battery module as the basic unit. The service data of the battery module includes the service status variable data of each battery cell. The service status variables include the state of charge index and the measurement point variable. When a new service data sample arrives, the cluster feature fusion processing is first performed to obtain a new fusion feature vector. , and using the intrinsic projection matrix and the drift projection matrix Calculate its inherent characteristics and drift characteristics, and the calculation formulas are: ； ； in, and Represent the new fusion feature vectors The inherent characteristics and drift characteristics of Recalculate the inherent characteristics and drift characteristics The monitoring statistic is calculated as follows: ； in, The monitoring statistic that represents the inherent characteristics of the new service data sample, A monitoring statistic representing the drift characteristics of a new sample of service data; represents the total number of Gaussian components in the intrinsic Gaussian mixture model, Represents the total number of Gaussian components in the drift Gaussian mixture model; represents the model parameters of the mth Gaussian component in the intrinsic Gaussian mixture model, Represents the model parameters of the mth Gaussian component in the drift Gaussian mixture model; express The Mahalanobis distance to the mth Gaussian component in the intrinsic Gaussian mixture model, express The Mahalanobis distance to the mth Gaussian component in the drift Gaussian mixture model; represents the posterior probability of the intrinsic Gaussian mixture model corresponding to the intrinsic feature distribution, Represents the posterior probability of the drift Gaussian mixture model corresponding to the drift feature; Then determine the monitoring statistic of the inherent characteristics of the calculated new service data sample Is it greater than the control limit of the monitoring statistic corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristics of the new service data sample? Is it greater than the monitoring statistic control limit corresponding to the drift characteristic? If is less than or equal to the control limit of the monitoring statistic corresponding to the inherent characteristic, and If the value is less than or equal to the control limit of the monitoring statistic corresponding to the drift characteristic, the state of the energy storage battery stack to be monitored is considered normal; otherwise, the state of the energy storage battery stack to be monitored is considered abnormal.

9. The method for adaptively sensing the degradation of energy storage battery stack over its entire life cycle according to claim 8, characterized in that: The design of the adaptive model update strategy determines whether an update is required based on the aging degree of the online service data, and specifically includes: For the intrinsic features, a Gaussian mixture model for monitoring is designed, which is called the intrinsic Gaussian mixture model; for the drift features, a recursive Gaussian mixture model is established, which is called the drift Gaussian mixture model; Design an adaptive model update strategy for the intrinsic Gaussian mixture model and the drifting Gaussian mixture model, specifically by analyzing the degradation components to determine whether the intrinsic Gaussian mixture model or the drifting Gaussian mixture model needs to be updated and alarmed; Among them, the adaptive model update strategy specifically includes: Case 1: When the monitoring statistic of the inherent characteristics of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristic is less than or equal to the monitoring statistic control limit corresponding to the drift characteristic, the state of the energy storage battery stack to be monitored at this time is regarded as a normal state, and new service data samples are collected to update the model parameters of the drift Gaussian mixture model. When the size of the collected data sample block reaches the preset value When , the model parameters of the drift Gaussian mixture model and the monitoring statistic control limits corresponding to the drift characteristics are updated recursively; Case 2: When the calculated monitoring statistic of the inherent characteristics of the service data sample is greater than the monitoring statistic control limit corresponding to the inherent characteristics, and the monitoring statistic of the drift characteristics is less than or equal to the monitoring statistic control limit corresponding to the drift characteristics, the state of the energy storage battery stack to be monitored at this time is regarded as an abnormal state, the update of the drift Gaussian mixture model is stopped and an alarm is triggered; Case 3: When the monitoring statistic of the inherent characteristic of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the inherent characteristic, and the monitoring statistic of the drift characteristic is greater than the monitoring statistic control limit corresponding to the drift characteristic, the collected service data samples are used for a recursive strategy, and then the monitoring statistic of the drift characteristic is recalculated and the monitoring statistic control limit of the drift characteristic is determined again; thereafter, if the monitoring statistic of the drift characteristic of the calculated service data sample is still greater than the monitoring statistic control limit corresponding to the drift characteristic, an alarm is triggered; if the monitoring statistic of the drift characteristic of the calculated service data sample is less than or equal to the monitoring statistic control limit corresponding to the drift characteristic, it is considered that the energy storage battery stack to be monitored is in a normal state; Case 4: When the calculated monitoring statistic of the inherent characteristic of the service data sample is greater than the monitoring statistic control limit corresponding to the inherent characteristic, and the monitoring statistic of the drift characteristic is greater than the monitoring statistic control limit corresponding to the drift characteristic, the update of the Gaussian mixture model is stopped and an alarm is triggered.

Images