An unsupervised fault diagnosis method for energy storage batteries based on similarity measurement

Through an unsupervised fault diagnosis method based on similarity measurement, the maximum value point, trend term characteristics and outlier measurement of energy storage battery cells are extracted, and combined with the D-S evidence theory, the difficulty in detecting battery cell failures in the prior art is solved. The problem is that high efficiency and high precision fault detection and positioning are achieved, and are suitable for large-scale energy storage battery systems.

CN114355206BActive Publication Date: 2024-11-01ZHEJIANG ZERO CARBON DIGITAL ENERGY TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210009321.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-05
Publication Date
2024-11-01
Estimated Expiration
2042-01-05

AI Technical Summary

Technical Problem

The existing energy storage battery fault diagnosis methods are mainly based on the operation of the battery cluster or the overall operation, and it is difficult to effectively detect and locate battery cell failures. Especially in a system composed of thousands of battery cells, the faulty battery cell affects the system. Safe and stable, and the charging and discharging characteristics and environmental factors of lithium batteries increase the difficulty of diagnosis.

Method used

Using an unsupervised fault diagnosis method based on similarity measurement, an unsupervised fault diagnosis data set is constructed by monitoring and collecting battery signals, using a sliding average filter to remove noise, extract the maximum value points, inflection points and trend terms characteristics, and calculate the single The outlier measurement of bulk batteries is combined with D-S evidence theory for information fusion to achieve high efficiency and high accuracy fault detection and positioning.

Benefits of technology

It realizes rapid detection of faulty batteries under unmanned monitoring conditions and provides maintenance suggestions. It has low calculation overhead and good robustness. It is suitable for different battery objects and charging and discharging scenarios. It can identify abnormal batteries online without manual calibration.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114355206B_ABST
    Figure CN114355206B_ABST
Patent Text Reader

Abstract

The present invention provides an unsupervised fault diagnosis method for energy storage batteries based on similarity measurement, which belongs to the field of energy storage battery fault diagnosis. It solves the problem that it is impossible to detect and locate the fault of a battery pack in series with high efficiency. The present invention monitors, collects, and summarizes the signals of the energy storage battery charging and discharging process to construct an unsupervised fault diagnosis data set; cleans the collected signals, removes abnormal points, and constructs a filter to suppress the noise mixed in the sampling; constructs a key point sequence for each single cell according to the signal characteristics of the charging and discharging process; extracts the segmented trend item features of the monitoring signal, divides the complete monitoring signal into time segments, and performs segmented linearization; constructs a single cell outlier value calculation module; realizes the information fusion of the three standards of Euclidean distance, key point distribution, and trend item difference, obtains the degree of outlier as an important measurement indicator of the occurrence of the fault, and then identifies the faulty battery. It can quickly troubleshoot faulty batteries and provide maintenance suggestions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of energy storage battery fault diagnosis, and in particular to an energy storage battery unsupervised fault diagnosis method based on similarity measurement. Background Art

[0002] Promoting the rapid development of new energy storage is an important technology and basic equipment to support the new power system.

[0003] On the power supply side, energy storage can ensure the efficient absorption and utilization of new energy, provide capacity support and a certain peak-shaving capability for the power system; on the grid side, it can improve the system's flexible adjustment capabilities and safety and stability levels after large-scale, high-proportion new energy and large-capacity DC access; at the end of the grid and in remote areas, energy storage or wind and solar power stations are built to improve the power supply capacity of the grid; mobile or fixed energy storage can improve emergency power supply capabilities or delay the need for transmission and transformation upgrades; on the user side of distributed new energy, microgrids, big data centers, 5G base stations, charging facilities, industrial parks, etc., energy storage is integrated to form new application scenarios.

[0004] Most new energy storage systems use lithium battery cells as basic units. The energy storage batteries are grouped in such a way that multiple battery cells are connected in series and parallel to form a module, multiple modules are connected in series and parallel to form a battery cluster, and multiple battery clusters are connected in parallel to form a battery system unit. This battery system unit contains thousands of battery cells. Large energy storage power stations on the power supply side, the grid side, and some user sides will be composed of multiple of the above-mentioned battery system units. In such a system composed of tens of thousands of battery cells, if one cell fails, it will have a certain impact on the safe, stable, and efficient operation of the entire system. Therefore, the diagnosis of the health status of battery cells is particularly important.

[0005] At present, most methods for energy storage battery fault diagnosis are based on the operation of the battery as a whole or a cluster of batteries, and rarely based on the analysis of battery cells. A large number of battery cells, the charging and discharging characteristics of lithium batteries, and environmental factors at the re-entry point all make battery fault diagnosis difficult. This patent provides an unsupervised fault diagnosis method for energy storage batteries based on similarity measurement, which is used to autonomously realize high-efficiency and high-precision fault detection and positioning of series battery packs, quickly troubleshoot faulty batteries, and provide maintenance suggestions. Summary of the invention

[0006] The present invention provides a method, device and system for identifying and diagnosing energy storage battery cell faults. The main goal is to autonomously achieve high-efficiency and high-precision energy storage battery cell fault detection and positioning under unmanned monitoring conditions, quickly troubleshoot faulty batteries, and provide maintenance suggestions. To achieve the above goals, the present invention provides an unsupervised fault diagnosis method for energy storage battery cells. The method can be applied to distributed data acquisition and analysis terminals. The technical solution includes:

[0007] S1. Signal monitoring, collection and aggregation during the charging and discharging process of energy storage batteries to build an unsupervised fault diagnosis data set;

[0008] S2. Clean the collected signals to remove abnormal points caused by sensor fluctuations, sudden changes in the operating environment, etc., and use the sliding average method to build a filter to suppress the noise mixed in the sampling. The sliding time window slides forward once every sampling time. The length of the time window is L, the number of samples is n, and therefore it slides nL times in total.

[0009] S3. According to the signal characteristics of the charging and discharging process, and the timing characteristics such as the maximum (minimum) value point and the inflection point, a key point sequence of each single battery is constructed;

[0010] Since the monitoring signal of the series battery pack is a discrete signal, its first-order difference and second-order difference are calculated, and the moment when the first-order difference and the second-order difference are zero is taken as the key point, so as to construct a key point sequence consisting of maximum (minimum) values ​​and inflection points.

[0011] S4. Extract the segmented trend item features of the monitoring signal. Based on the key point sequence of each single cell, the complete monitoring signal is divided into several time segments. Perform segmented linearization, replace the original nonlinear signal with a linear function in each time segment, take the slope in each time segment as the trend item, and finally construct the trend item sequence of each single cell.

[0012] S5. Based on the similarity analysis principle, a single cell outlier value calculation module is constructed. This module can measure the degree of outlier of each single cell compared with other cells during the charging and discharging process. For this purpose, the following three outlier measurement submodules are constructed;

[0013] S501, constructing a Euclidean distance submodule. For a single battery, calculate the bi-norm of its monitoring signal and all other batteries one by one, and take the average value as the outlier value of the single battery under the Euclidean distance standard.

[0014] S502, constructing a key point distribution submodule. For a single battery, using the dynamic time warping method, calculate the distance between its key point sequence and all other batteries one by one, and take the average value as the outlier value of the single battery under the key point distribution standard.

[0015] S503, constructing a trend item difference submodule. For a single battery, using the dynamic time warping method, calculate the distance between its trend feature sequence and all other batteries one by one, and take the average value as the outlier value of the single battery under the trend item difference standard.

[0016] S6. The three outlier measurement submodules are regarded as three different basic probability distribution functions, and the existence of a battery fault is regarded as the basic hypothesis space. Then, the DS evidence theory is used to realize the information fusion of the three outlier measurement submodules, and the outlier degree of each single cell compared with other single cells is obtained as an important measurement indicator of the occurrence of failure; if the outlier value exceeds a certain set threshold, the battery is determined to be faulty.

[0017] After adopting the above technical scheme, the unsupervised fault diagnosis method of energy storage battery based on similarity measurement of the present invention has the following advantages compared with the prior art:

[0018] This method measures the similarity between two single cells based on distance, with low computational overhead, good robustness and easy engineering implementation. It constructs multiple outlier metrics to capture battery characteristics from multiple angles such as the battery's charging and discharging stages and real-time power, effectively improving the ability to extract timing information. Dynamic time warping is introduced to address the characteristic differences between different single cells, and this method has strong adaptability to different battery objects and charging and discharging scenarios. This method belongs to unsupervised learning, which can not only capture factors such as the charging and discharging stages and the real-time status of the battery without manual calibration, but also identify abnormal batteries when fault samples are unknown, thereby achieving online diagnosis.

[0019] Other features and advantages of the present invention will be set forth in the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a general flow chart of an embodiment of the present invention based on similarity measurement.

[0021] Figure 2 is a diagram of the original monitoring signal of the embodiment.

[0022] Figure 3 Schematic diagram of an outlier curve of the embodiment based on the Euclidean distance standard.

[0023] Figure 4 Schematic diagram of a curve of the original monitoring signal of the embodiment after the first-order derivative.

[0024] Figure 5 Schematic diagram of the curve of the original monitoring signal after the second-order derivative of the embodiment.

[0025] Figure 6 It is the final result obtained after information fusion is performed by DS evidence theory in the embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0027] S1. Collect the voltage signals of the battery pack during several charge and discharge cycles, including the data of normal batteries and faulty batteries. Assuming that the detected battery pack contains m single cells, there are n sampling samples in several charge and discharge cycles, and build a monitoring matrix for the battery pack. ,in .

[0028] S2. Connect the data matrix to the outlier processing module, which can filter and smooth the monitoring signal. For the monitoring signal at time t, the outlier processing module constructs a time window , the time window length is L, so the window contains data For each single cell monitoring signal, due to noise, sensor mutation and other reasons, the signal needs to be smoothed to remove abnormal points. At the same time, the time-varying factors caused by the state change of the battery during the charging and discharging stage need to be retained as much as possible. First, calculate the mean value of the signal in the time window

[0029]

[0030] and standard deviation

[0031]

[0032] in According to the 68–95–99.7 rule, the probability of falling within 3 standard deviations of the mean of the signal in the window is 99.73%, so the upper and lower 3 standard deviations of the mean constitute the normal range of the signal at the current moment.

[0033] For time t, the normal range at that time is The signal value outside the interval will be regarded as an abnormal point. If the monitoring signal at time t If it is greater than the upper bound, it is replaced by the upper bound. If it is less than the lower bound, it is replaced by the lower bound. If the monitoring signal If the value is within the normal range, the sliding average value is used instead, that is,

[0034]

[0035] The time window slides on all the data for nL times to obtain a smoothed signal. All subsequent signals are smoothed signals.

[0036] S3. In order to extract the timing regularity of the signal, a sequence of specific time points is constructed, which can reflect whether the actual battery object is in a charging state or a discharging state during the charging and discharging process, as well as the change from a fast stage to a slow stage during the charging and discharging process.

[0037] For each single cell j and its monitoring signal , calculate the first-order difference at each moment and second-order differences , For the obtained first-order difference matrix F and second-order difference matrix G, according to the matrix F and matrix G The corresponding moments with zero values ​​in the column constitute the A sequence of specific time points of monitoring signals ,in For the The number of specific time points of a monitoring signal.

[0038] Define the key point as an isolated point in the sequence of specific time points, that is, a specific time point is a key point if and only if . like , then discard the specific time point. , can get the first The key point sequence of the monitoring signal ,in For the Similarly, the key point sequences of the remaining monitoring signals can be obtained according to the above steps.

[0039] S4. Extract the segmented trend item features of the monitoring signal. According to the key point sequence of each single battery, the complete monitoring signal is divided into For convenience, the time segment Monitoring signal of each single cell for , its key point sequence

[0040]

[0041] In the In each time segment, a linear function is used to replace the original nonlinear signal to achieve piecewise linearization. The expression of the linear function is The slope If the time segment involves two endpoints, the endpoint values ​​are used instead. For the first time segment, , for the time segment, , and finally constructed the The trend item sequence of a single cell is Similarly, a sequence of trend terms for the remaining cells in the battery pack can be constructed.

[0042] S5. Construct a calculation module for the degree of outlier of a single cell, which is divided into three submodules as follows.

[0043] S501, outlier calculation module 1: Calculate the outlier degree of a single cell based on Euclidean distance. For a single cell j and its monitoring signal , and the degree of outlier from another single cell k is

[0044]

[0045] Under the Euclidean distance standard, the outlier degree of single cell j is defined as the average of the outlier degrees of all other single cells, that is,

[0046]

[0047] in is the normalized exponential function, .

[0048] S502, outlier calculation module 2: Calculate the outlier degree of a single cell based on the key point distribution. For a single cell j and its key point sequence , and the degree of outlier from another single cell k is

[0049] in is the distance obtained using the dynamic time warping method. Different from the Euclidean distance, the dynamic time warping method can measure the distance between two time series with inconsistent lengths. Under the key point distribution standard, the outlier degree of single cell j is defined as the average of the outlier degrees of all other single cells, that is,

[0050]

[0051] S503, outlier calculation module 3: Calculate the outlier degree of the single cell based on the trend item distribution. Similar to the dynamic time warping standard, for a single cell j and its trend item sequence , and the degree of outlier from another single cell k is ,in is the distance obtained using the dynamic time warping method. Under the trend term distribution standard, the outlier degree of single cell j is defined as the average of the outlier degrees of all other single cells, that is,

[0052] .

[0053] The outputs of the three outlier calculation modules are three matrices: , the matrix dimension is related to the characteristics of the actual signal.

[0054] S6. According to the DS evidence theory, an outlier fusion module is constructed. Its function is to fuse the outliers under three different standards and effectively overcome the decision-making difficulties caused by the discrepancies between the three standards. The three standards are m1, m2 and m3, and the corresponding outliers are , the batteries in the battery pack are B1, B2, …, B N , which constitute the recognition framework , where N is the number of single cells, m i (B j ) is the single cell B under the i-th standard j The degree of outlier, due to the output characteristics of the softmax function, satisfies Obviously

[0055] The outlier fusion module first calculates the normalization constant for , the normalization constant calculation formula is

[0056]

[0057] For a single cell B i The outlier fusion module calculates the fusion results of outliers under the three standards according to the following formula:

[0058]

[0059] The output of the outlier fusion module is the single cell B i The failure probability of

[0060]

[0061] If the failure probability of a single battery exceeds the set threshold of 50%, it is judged to be faulty and the recommended information is transmitted to the human-computer interaction system.

[0062] This example uses monitoring data from a BMS system in Hefei to verify the unsupervised fault diagnosis method for energy storage batteries based on similarity measurement. The specific process is as follows:

[0063] (1) The BMS collects voltage signals of the battery pack during several charge and discharge processes, with a total of 216 single cells and 1,439 samples;

[0064] (2) During the charge and discharge process, the voltage signal of the faulty battery has characteristics such as abnormal increase and high level retention, which is different from that of a normal battery. Figure 2 Therefore, the battery status can be mathematically described by extracting key points, trend items and other features, as shown in the attached Figure 3-5 shown.

[0065] (3) Attachment Figure 6 The result of fault diagnosis. The failure probability of single cell BMS2N4U007 in the battery pack is significantly higher than that of other batteries, exceeding the threshold of 50%, so it is determined to be faulty. Under this technical solution, the failure probability of the faulty battery BMS2N4U007 is much higher than that of other normal batteries, and the diagnosis result has a high confidence level.

Claims

1. An unsupervised fault diagnosis method for energy storage batteries based on similarity measurement, characterized in that: The steps include: S1. Signal monitoring, collection and aggregation during the charging and discharging process of energy storage batteries to build an unsupervised fault diagnosis data set; S2. Clean the collected signals to remove abnormal points caused by sensor fluctuations and sudden changes in the operating environment. At the same time, use the sliding average method to build a filter to suppress the noise mixed in the sampling; S3. According to the signal characteristics of the charging and discharging process, and the timing characteristics of the maximum point, minimum point, and inflection point, a key point sequence for each single cell is constructed; S4, extracting the segmented trend item features of the monitoring signal, dividing the complete monitoring signal into several time segments based on the key point sequence of each single battery, and performing segmented linearization; S5. Based on the principle of similarity analysis, a single cell outlier value calculation module is constructed, which includes three outlier measurement submodules: Euclidean distance, key point distribution, and trend item difference; S6. Utilize DS evidence theory to realize information fusion of three outlier measurement submodules: Euclidean distance, key point distribution, and trend item difference. Obtain the degree of outlier of each single cell compared with other single cells as an important measurement indicator of failure, and then identify the faulty battery.

2. According to claim 1, the unsupervised fault diagnosis method for energy storage batteries based on similarity measurement is characterized in that: In step S1, the following steps are specifically included: S101, using a voltage sensor to collect voltage signals of a battery pack during several charge and discharge cycles; S102, the detected battery pack includes m single cells, and there are n sampling samples in a number of charge and discharge cycles, and a monitoring matrix of the battery pack is constructed.

3. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 1 is characterized in that: In step S2, the following steps are specifically included: S201, calculating the mean and standard deviation of the local signal, and constructing a normal range of the monitoring signal at the current moment; S202: Smoothing the monitoring signals that are outside the normal range.

4. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 3 is characterized in that: In step S201, a time series segment is constructed using a time window to reflect local features and remove outliers. The time window W t Contains data [X t-L ,X t-L+2 ,...,X t-1 ] (1) Where L is the length of the time window, and the time window moves backward one moment at a time.

5. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 1 is characterized in that: In step S3, the following steps are specifically included: S301, constructing matrices F and G respectively according to the extreme points and inflection points of the single cell monitoring signal, thereby obtaining a specific time point sequence; S302: Constructing a key point sequence of the kth single cell according to the specific time point sequence Among them, M k is the number of specific time points of the kth monitoring signal.

6. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 5 is characterized in that: In step S301, for discrete signals, the first-order difference and second-order differences To obtain extreme points and inflection points, where i = 2, 3, ..., n-1, j = 1, 2, ..., m, and then to obtain specific time points based on the extreme points and inflection points being zero.

7. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 5 is characterized in that: In step S302, the sequence of specific time points is traversed to search for isolated points in the time dimension in the sequence to obtain the key point sequence of the kth monitoring signal.

8. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 1 is characterized in that: In step S4, the following steps are specifically included: S401, according to the key point sequence obtained in step S3, the monitoring signal is divided into several segments, and a linear function is used to replace the original nonlinear signal in each segment to achieve piecewise linearization; S402: extracting the slope of each linear function to reflect the trend of each time series segment and forming a segmented trend item feature sequence.

9. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 1 is characterized in that: In step S5, multiple outlier standards are used to construct calculation modules for the outlier degree of each single cell, and the outlier value of each single cell is calculated, which specifically includes the following steps: S501, calculating the outlier value of each single battery one by one according to the Euclidean distance standard; S502, calculating the outlier value of each single cell one by one according to the key point distribution standard; S503 , calculating the outlier value of each single cell one by one according to the trend item distribution standard.

10. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 9, characterized in that: In step S501, the degree of outlier between the monitoring signal of a single cell and another single cell k is measured by the Euclidean distance, and finally mapped to an outlier value through a normalization function, that is, Where softmax(·) is a normalized exponential function, and the Euclidean distance is calculated as 11. The unsupervised fault diagnosis method for energy storage batteries based on similarity measurement according to claim 9 is characterized in that: In step S502, the degree of outlier between the monitoring signal of a single cell and another single cell k is measured by the key point distribution, and finally mapped to an outlier value through a normalization function, that is, Where D jk =Distance(T j ,T k ), Distance(·,·) is the distance obtained by the dynamic time warping method; in step S503, the degree of outlier between the monitoring signal of a single cell and another single cell k is measured by the trend term distribution, and finally mapped to the outlier value through the normalization function, that is, Trend jk =Distance(T j , T k ); In step S6, the DS evidence theory is introduced to integrate multiple outlier criteria and comprehensively calculate the outlier value of each single cell, which specifically includes the following steps: S601, constructing an identification framework for battery pack fault diagnosis according to the definitions of three standards and their outliers; S602, constructing an outlier fusion module to fuse the outliers of each single battery under the three standards and output the failure probability; S603: Determine whether the battery fails based on the failure probability and the set threshold.