Battery health state real-time evaluation method based on data flow clustering

Abstract

The invention discloses a battery health state real-time evaluation method based on data flow clustering, which comprises the following steps: collecting operation record data characteristics of a battery, grouping, dynamically segmenting according to a working voltage fluctuation range of the battery, and constructing a data set representing an initialized health state of the battery. The real-time operation record data of the battery is compared with the record of the existing health state data set, and the health state is quantified in combination with the pseudo tag, so that the evaluation result of the health state of the battery is comprehensively obtained, and the evaluation accuracy is improved. Meanwhile, in consideration of different operation states of the battery in different environments, a dynamically updated data set maintenance method and a health state quantification method based on a neural network are adopted to evaluate and quantify the operation record data of the battery in real time, whether the operation record data is similar to an existing health state data set or not is observed, and the operation record data can be added; if the health state changes, a blank health state data set is newly established, and the characteristics of the health state are initialized; and meanwhile, the quantitative model is trained by utilizing historical data in real time, so that the detection accuracy is further improved. According to the method, a large amount of data is summarized by adopting dynamic segmentation and the health state, and the calculated amount of the evaluation process is reduced, so that the real-time evaluation of the health state of the battery is realized.

Description

Technical Field

[0001] The present invention belongs to the field of energy storage battery health management, and more specifically, relates to a real-time battery health status assessment method based on data stream clustering. Background Art

[0002] How to accurately assess the health status of batteries is of great significance to the normal operation of batteries, early warning of faults and prevention of safety accidents.

[0003] Traditional battery health assessment methods, such as parameter identification methods or threshold-based assessment techniques, typically estimate and predict the properties of a battery or battery pack, capturing and calculating the battery's operating characteristics and health level. These methods often rely on expert knowledge to analyze battery data and identify specific characteristics for fault detection. This heavy reliance on expertise and data quality limits their applicability in real-world scenarios. Real-time recording of battery operating data and rapid, accurate analysis and assessment of health status within limited time and memory are core challenges facing the current field of battery health management. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a real-time battery health status assessment method based on data stream clustering to achieve rapid assessment of the health status of energy storage batteries.

[0005] To achieve the above-mentioned object, the present invention provides a real-time battery health status assessment method based on data stream clustering, which includes the following steps:

[0006] (1) Collect the characteristics of the energy storage battery or battery pack;

[0007] (2) Simply group each battery according to its health type in battery characteristics;

[0008] (3) Dynamically segment the overall data according to the fluctuation range of the battery voltage characteristics;

[0009] (4) For each independent battery, an initial state data set consisting of a certain number of records representing the initial normal health state of the battery is initialized;

[0010] (5) For the current operation record data of the battery obtained, find a health status data set that is closest to the current battery operation status among the many health status data sets of the battery, and judge whether the health status of the battery has changed based on the difference between the record and the current operation status record; if it has not changed, insert the current operation status as a record into the current status data set of the battery; if the health status has changed, otherwise, create a new health status data set and save the current operation status as the latest health status data set;

[0011] (6) Dynamic dataset maintenance: A set of features is maintained for each health status dataset: the mean and radius of all data records in the dataset (the data record with the farthest distance from the mean is acceptable). Each normal operating status record is regarded as a point in the feature vector space, and the dataset is regarded as a micro-cluster. When a new battery data operation record is added to the dataset, the overall mean and radius of the dataset are updated;

[0012] At the same time, a health status pseudo-label fitting method based on a neural network is used to assign a pseudo-label to each operating data state. Specifically, for the operating record data of each battery, a large amount of historical data is used to train the neural network for pseudo-label fitting. The pseudo-label takes into account the battery's life span, health status, and fault conditions during operation. In this way, a health status assessment can be performed on a certain health status data set to quantify a health status score. Based on the change in the health status score, the average health status level of the entire data set can be obtained. At the same time, by comparing different health status data sets, the change in the health status level of the battery during operation can also be obtained. After each battery data set is initialized according to step (4), the overall data set is evaluated and updated in real time according to step (5).

[0013] The object of the present invention is achieved in this way.

[0014] This method, based on data stream clustering, provides a real-time battery health assessment method. By collecting the characteristics of each battery, dynamically segmenting it, and constructing an initial health dataset representing the battery's normal state, the method uses dynamic dataset maintenance and comparisons of new operating records with the health dataset's characteristics to comprehensively determine the battery's current health status, improving detection accuracy. Furthermore, considering that the battery's health status varies over time and requires quantification to compare changes in health levels, the method utilizes a neural network-based health status pseudo-label fitting method to assess and quantify the health dataset and examine its evolutionary trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of a specific implementation of the method for real-time battery health status assessment based on data stream clustering of the present invention;

[0016] Figure 2 Schematic diagram of dynamic data segmentation in the present invention, wherein each black dot represents corresponding battery data, and the data points framed by red dotted lines constitute a segment;

[0017] Figure 3 This is a schematic diagram of segmented representation data obtained by segmenting the original data in the present invention. The small black dots on the left represent the multi-dimensional operating data of the battery, and the larger black dots on the right represent the segmented representation data, which represents the average health level of the operating records within the segment.

[0018] Figure 4 Schematic diagram of the health status pseudo-label fitting based on the neural network in the present invention, wherein X and S represent the battery operation record data and the score quantified for the health status, respectively, and Model is the model that maps the health status to the quantified score;

[0019] Figure 5 This is a system framework diagram for evaluating battery health status according to a specific embodiment of the method for real-time evaluation of battery health status based on data stream clustering of the present invention. DETAILED DESCRIPTION

[0020] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0021] Figure 1 This is a flowchart of a specific implementation of the method for real-time battery health status assessment based on data stream clustering of the present invention.

[0022] In this embodiment, if Figure 1 As shown, the real-time battery health status assessment method based on data stream clustering of the present invention includes a step:

[0023] S1: Collecting battery operating characteristics

[0024] Battery characteristics include basic information during operation, such as battery voltage, current, temperature, working status, and performance indicators such as battery capacity and charge and discharge rate.

[0025] Taking the lithium battery of a new energy vehicle as an example, the battery characteristics include the battery's output power, charging power, operating voltage, current, battery pack temperature, ambient temperature, and vehicle operating status such as acceleration and braking. i represents multiple characteristics of the battery at time i, Represents the jth feature of the battery at the i-th moment. i=[48,0,360], which means that the output power of the battery at that time was 48 kVA (the jth=1 feature), the charging power was 0 (the jth=2 feature), and the operating voltage was 360 V (the jth=3 feature).

[0026] S2: Battery Group

[0027] In the specific implementation process, the battery grouping can be simply grouped based on whether a fault occurs, or clustering methods such as K-means, spectral clustering, DBSCAN, etc. can be used to group the batteries according to multiple characteristics of the battery.

[0028] S3: Dynamically segment the overall data according to voltage

[0029] For each battery's operating record data, the data is segmented based on the battery's operating voltage fluctuation range. Specifically, the voltage fluctuation range is divided into N regions. Continuous data within the same data range is treated as a segment, and each data segment is processed independently. The data segments are then aggregated based on the battery's charge and discharge behavior, enabling analysis of state evolution from a multi-granular perspective.

[0030] S4: Initialize the normal operation status data set

[0031] For each battery, a running record dataset consisting of a certain number of records representing the battery's normal health status is initialized. The initial state datasets obtained from all initialized batteries together constitute the overall health status initial dataset, and this dataset is dynamically updated during subsequent battery health assessments.

[0032] S5: Real-time evaluation of battery health status

[0033] Once a battery's current operating record data is obtained, the health state closest to the operating record data can be found based on the corresponding health state dataset. The difference between the operating record and the health state record can then be used to determine whether the battery's health state has changed. In this embodiment, the data stream clustering method uses the characteristics of data points and microclusters to perform a distance-based health state assessment (the difference is represented by distance) based on their center points and radius. If no state change has occurred, the current operating state is inserted as a record into the latest health state dataset for the battery. If a health state shift has occurred, a new health state dataset is created and the current operating record data is inserted into it.

[0034] In this embodiment, the specific method for real-time evaluation of battery health status is as follows:

[0035] 5.1) Find the current running status x respectively k The data point closest to the health status data set corresponding to the battery is the health status record xh , calculate the difference d according to the following formula:

[0036] d=dist(x k ,c h )-r h

[0037] Wherein, dist is the distance function. In this embodiment, Euclidean distance is used, c h is the distance x k The center point of the health status data set where the most recent running status record is located, that is, the mean of the health status feature vector, r h is x h The radius of the health status dataset. The health status dataset containing N running records is c h and r h The calculation method is:

[0038]

[0039] Where LS is the linear sum of all running records in the health status data set, and SS is the square sum of all running records in the health status data set.

[0040] 5.2) Health status change judgment

[0041] After obtaining the distance d between the current running state and the nearest health state data set, the health state is determined using the following rules:

[0042] First, if d is less than 0, the battery's health status has not changed; insert the running status record into the corresponding health status data set, and update the center point c of the health status data set at the same time. h and radius r h ;

[0043] Second, if d is greater than 0, it indicates that the battery's health status has changed. It is necessary to create a new empty health status data set, initialize the center point and radius of the health status data set, and insert the current operating record data into the health status data set.

[0044] S6: Dynamic Dataset Maintenance

[0045] 6.1) Real-time update of battery health status records

[0046] Because the battery health status data set contains a large amount of operating record data, and most of the operating record data in the same health state are relatively similar, the two characteristics of the health status data set can be used to summarize and describe it: the center point c h and radius rh , where the center point can be considered as a unified health state feature vector, and the radius is the operating record data area covered by the health state. Whenever new operating record data is inserted into the battery health state set, the center point and radius of the health state need to be dynamically updated.

[0047] In this embodiment, the method for real-time updating of battery health status records is as follows:

[0048] 6.1.1) Each running status record data in the health status set (expressed by feature vector x i ) is considered as a point (an object) in the eigenvector space;

[0049] 6.1.2) There is a new object x in the health status set containing N objects new When adding, the feature vector represented by the object is used to first update the center point of the health state. The update formula is:

[0050]

[0051] where c old represents the old center point of the health state set, c new Indicates the updated health status set center point; use c new Update the radius of the health state set. The update formula is:

[0052]

[0053] 6.1.3) After multiple interactions, similar operating status records will be clustered together and have the same health status. At the same time, the health status from different batteries may be relatively similar. Considering the dynamic nature of data, computing efficiency and resource limitations, you can choose to merge similar health status. For similar health states a and b, first calculate the Euclidean distance between the center points of the health states. If it is less than the radius of a certain health state, it may be merged. You can compare the variance, distribution shape, etc. to ensure that the structure is consistent after the merger. When merging, you need to calculate the characteristics of the merged health state. First, you need to perform a weighted average of the center points:

[0054]

[0055] Among them, c merge is the center point of the merged health status, c a and c b are the center points of two similar health states, N a and N bare the number of battery operation record data for two batteries with similar health states; for the radius, it can be approximated as the maximum distance of the merged data distribution, or a conservative estimate based on the maximum distance:

[0056] r merge =max(r a ,r b )+||c a -c b ||

[0057] Among them, r a and r b The new radius covers the maximum radius of the original micro-cluster and adds the distance between the two centroids to ensure that all points are included. When the number of micro-clusters in memory exceeds the upper limit, you can choose to force merge the most similar micro-clusters.

[0058] 6.1.4) Finally, for all the health states obtained from the historical operation record data, the data of the battery health state evolution can be compared based on the center point of the health state data set. Specifically, assuming that the initial health state s0 of the battery can be represented by its center point c0, the health state set generated based on the real-time battery operation record data can be expressed as:

[0059] S={s0,s1,s2,…}

[0060] where s i The health state of the battery i can be expressed as its center point c i The entire battery health status set can be represented as the battery health evolution data.

[0061] Given that this data stream-based battery health state extraction method can re-record and summarize new battery health records by adjusting the radius of the health state data set, it is theoretically possible to perform unlimited real-time health state assessments. Therefore, by dynamically maintaining the data set, it is possible to manage potentially unlimited and real-time battery health record data and accurately extract the battery safety status.

[0062] 6.2) Quantification of battery health status

[0063] To intuitively assess the battery's state of health, a Health Risk Score (HRS) was proposed based on the evolution of health status data. This scoring method aims to quantify the battery's health status through deep feature extraction of battery operating data. Similar to the State of Health (SOH) metric, the HRS score ranges from 0 to 1, but unlike the SOH, higher HRS values ​​indicate greater safety risks.

[0064] In this invention, a neural network-based health status assessment and quantification method is used throughout the dynamic data maintenance process to quantitatively assess the battery's real-time operating record data and the health status represented by the data set. Specifically, throughout the process, a pseudo-label is generated for the historical data based on the actual health status. The operating record data and the pseudo-label are then used to train the neural network. The pseudo-label generation function is:

[0065]

[0066] Where t represents the time scale of the data point in the sequence, and f represents the fault condition of the data point. f = 0 indicates no fault, and f is assigned a specific value when there is a fault. T is a hyperparameter that controls the time scale, and α and β are hyperparameters that control the time scale of the healthy state and the fault weight, respectively.

[0067] In terms of fitting models, you can choose classic regression models such as Linear Regression or Neural Network. Here, the classic neural network Multilayer Perceptron (MLP) is used to perform regression fitting on pseudo labels. Multilayer Perceptron is a feedforward artificial neural network that consists of an input layer, at least one hidden layer, and an output layer. It uses nonlinear activation functions to approximate complex functions and is widely used in classification and regression tasks. Its basic model form is as follows: Figure 3 As shown in the Model.

[0068] In this embodiment, the real-time health status quantitative assessment method is as follows:

[0069] 6.2.1) Record the battery's operating status and calculate a pseudo-label based on its actual health status;

[0070] 6.2.2) Use historical data and pseudo labels to perform batch gradient backpropagation training on the neural network. The batch size of the training data depends on the specific application (for example, a batch contains 1000 records). The training loss function uses the mean square error:

[0071]

[0072] Where n is the amount of training data, Quantify the health status of the model;

[0073] 6.2.3) For the real-time operation record data of the battery and the generated health status, the model is used to perform a quantitative assessment of the health status. In this embodiment, Figure 3 As shown, the center point X of the battery real-time operation record data or health status data set is input into the model Model to obtain the estimated result. Considered as a quantitative score of the battery's health status.

[0074] Figure 4 This is a system framework diagram for quantitative evaluation of battery health status according to a specific embodiment of the method for real-time evaluation of battery health status based on data stream clustering of the present invention.

[0075] In this embodiment, if Figure 4 As shown in FIG, when the health status evaluation system is running, the following steps are included: (1) for each battery operation record data, the data is segmented according to the battery operating voltage fluctuation range; (2) the latest health status data set result is obtained by comparing the real-time operation record data with the existing health status data set; (3) the center point of the battery real-time operation record data and the health status data set is quantified to evaluate the overall and latest battery health status; (4) the battery health status is analyzed by combining the historical health status and current health status evaluation results.

[0076] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A real-time battery health status assessment method based on data stream clustering, characterized in that: The following steps are involved: (1) Collect the characteristics of the energy storage battery or battery pack; (2) Simply group each battery according to its health type in battery characteristics; (3) Dynamically segment the overall data according to the fluctuation range of the battery voltage characteristics; (4) For each independent battery, an initial state data set consisting of a certain number of records representing the initial normal health state of the battery is initialized; (5) For the current operating record data of the battery obtained, find a health status data set that is closest to the current battery operating status among the many health status data sets of the battery, and determine whether the health status of the battery has changed based on the degree of difference between the record and the current operating status record; If no change occurs, the current operating status is inserted as a record into the current status data set of the battery. If the health status changes, a new health status data set is created and the current operating status is saved as the latest health status data set. (6) Dynamic dataset maintenance: A set of features is maintained for each health status dataset: the mean and radius of all data records in the dataset (the data record with the farthest distance from the mean is acceptable). Each normal operating status record is regarded as a point in the feature vector space, and the dataset is regarded as a micro-cluster. When a new battery data operation record is added to the dataset, the overall mean and radius of the dataset are updated; At the same time, a health status pseudo-label fitting method based on a neural network is used to assign a pseudo-label to each operating data state. Specifically, for the operating record data of each battery, a large amount of historical data is used to train the neural network for pseudo-label fitting. The pseudo-label takes into account the battery's life span, health status, and fault conditions during operation. In this way, a health status assessment can be performed on a certain health status data set to quantify a health status score. Based on the change in the health status score, the average health status level of the entire data set can be obtained. At the same time, by comparing different health status data sets, the change in the health status level of the battery during operation can also be obtained. After each battery data set is initialized according to step (4), the overall data set is evaluated and updated in real time according to step (5).

2. The battery health status assessment method according to claim 1, characterized in that: In step (5), the health status data set corresponding to the battery is used to find the health status closest to the operation record data, and the difference between the operation record and the health status record is used to determine whether the health status of the battery has changed: 5.1) Find the current running status x respectively k The data point closest to the health status data set corresponding to the battery is the health status record x h , calculate the difference d according to the following formula: d=dist(x k ,c h )-r h Wherein, dist is the distance function. In this embodiment, Euclidean distance is used, c h is the distance x k The center point of the health status data set where the most recent running status record is located, that is, the mean of the health status feature vector, r h is x h The radius of the health status dataset. The health status dataset containing N running records is c h and r h The calculation method is: Where LS is the linear sum of all running records in the health status data set, and SS is the square sum of all running records in the health status data set. 5.2) Health status change judgment After obtaining the distance d between the current running state and the nearest health state data set, the health state is determined using the following rules: First, if d is less than 0, the battery's health status has not changed; insert the running status record into the corresponding health status data set, and update the center point c of the health status data set at the same time. h and radius r h ; Second, if d is greater than 0, it indicates that the battery's health status has changed. It is necessary to create a new empty health status data set, initialize the center point and radius of the health status data set, and insert the current operating record data into the health status data set.

Images