A monomer battery consistency detection algorithm based on a vehicle networking big data platform

Abstract

The application discloses a kind of monomer battery consistency detection algorithm based on Internet of Vehicles big data platform, including the following steps: S1, receive data and pre-process;S2, utilize trained hidden Markov model to judge the voltage interval section to be detected;S3, extract the subsequence near each monomer voltage sequence feature point, calculate the Mahalanobis distance between corresponding subsequence;S4, do anomaly detection to the feature matrix formed by characteristic subsequence with isolated forest algorithm;S5, respectively to the highest monomer voltage sequence and the lowest monomer voltage sequence and average voltage sequence do KS test;S6, calculate the Wasserstein distance between abnormal monomer voltage sequence and average monomer voltage sequence;Avoid relying on the error, noise and time delay of the voltage measurement value of some single point moment, can eliminate the noise and error of single point moment, avoid the interference of a large number of same subsequence in full time sequence distance calculation, greatly reduce the calculation amount compared with full time sequence distance calculation and basic shapelet search algorithm based on brute force search.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery detection, and particularly relates to a single battery consistency detection algorithm based on a vehicle networking big data platform. BACKGROUND

[0002] Lithium ion batteries are widely used in new energy vehicles and large-scale energy storage fields. During the charging and discharging cycle process, different use environments, use time and use conditions will significantly affect the battery performance, the inconsistency between single batteries will increase, and the performance of the battery pack will decline, thereby reducing the overall service life. Therefore, a reliable consistency detection technology is the key to ensuring good consistency during the use of the battery pack, and plays an important role in improving the operation performance and safety of the battery pack and prolonging the service life.

[0003] In the prior art, the single battery consistency detection based on the BMS is generally based on the statistical indicators such as the threshold values of the mean value, variance, slope and difference of the highest single voltage and the lowest single voltage at a certain moment to determine whether the single voltage difference is too large, or uses a single simple clustering algorithm to determine the abnormality. Generally, only the voltage difference between each single battery at each discrete moment is simply evaluated, which can represent the consistency between each single battery to a certain extent, but is easily affected by the measurement time delay, noise, environmental temperature and other interference factors. Moreover, all the battery data are not collected and uploaded at the same moment, and the data collected by the last battery in the same frame of uploaded data is about 100 ms later than the data collected by the first battery. The data sequence time axis obtained is often asynchronous, and it is difficult to directly compare, which leads to that the intrinsic characteristics of the single batteries in the battery pack cannot be truly described, and the error is large. When the existing data is used for abnormal data identification through fault clustering, the distance method is often used to directly determine the fault data. However, the clustering distance threshold of the fault point is generally given by experience estimation, and there is a great contingency. Once the clustering point set is too large or too small, it is easy to cause false alarm of fault warning. There is a lack of comprehensive evaluation index for the single battery voltage difference in a period of time, and the consistency evaluation index obtained is not comprehensive. Generally, the whole time subsequence in a period of time is detected, and all possible time subsequences are searched to distinguish the abnormal battery voltage, which has a large amount of calculation and is not easy to find the single battery difference of a few key points.

[0004] For example, a kind of battery level group consistency monitoring method is disclosed in Chinese patent literature, the publication number of which is CN112068007A, which discloses that the consistency is screened by detecting the voltage drop of the battery level group, and has the above-mentioned defects. SUMMARY

[0005] To this end, the application provides a single battery consistency detection algorithm based on a vehicle networking big data platform, which avoids errors, noises and time delays by detecting single battery voltage discharge sequences in a time segment during vehicle discharge, eliminates noises and errors at a single point in time, focuses on abnormal subsequence intervals to avoid interference of a large number of same subsequences in full time sequence distance calculation, and greatly reduces calculation amount.

[0006] In order to achieve the above object, the application provides the following technical scheme.

[0007] A single battery consistency detection algorithm based on a vehicle networking big data platform comprises the following steps.

[0008] S1, receiving data and preprocessing;

[0009] S2, judging a voltage interval segment to be detected by using a trained hidden Markov model

[0010] S3, extracting a subsequence near a feature point of each single battery voltage sequence, and calculating Mahalanobis distances between corresponding subsequences;

[0011] S4, performing abnormality detection on a feature matrix composed of feature subsequences by using an isolated forest algorithm;

[0012] S5, performing KS test on the highest single battery voltage sequence and the lowest single battery voltage sequence and the average voltage sequence respectively;

[0013] S6, calculating a Wasserstein distance between an abnormal single battery voltage sequence and an average single battery voltage sequence. By constructing an unsupervised shapelet Mahalanobis distance feature matrix, using an isolated forest algorithm to detect voltage change dynamic trend, and using KS test and wasserstein distance to provide a comprehensive evaluation index for cell voltage in a detection time window.

[0014] Preferably, S1 comprises accepting data streams from a message queue and obtaining single battery voltage data of each vehicle in a certain time window. Data can be quickly and batch obtained, and detection efficiency is improved.

[0015] Preferably, S1 further comprises detecting single cell voltage signals of all online vehicles and obtaining a parameter matrix, and forming a sliding observation window matrix for each frame of data Ui of single battery voltage of a vehicle to be concerned during driving. By using a hidden Markov model, it can be avoided that a fixed interval specified by a person cannot accurately cover the to-be-detected interval of each single battery of each vehicle.

[0016] As preferred, the voltage and SOC of several cycle samples in S2 are included as input variables to train a supervised hidden Markov model, and the output label is the corresponding interval segment, denoted as s1, s2, s3; the trained model can automatically determine the battery discharge interval segment by inputting the total battery voltage and SOC, and the detection algorithm is started in the middle segment s2. The maximum likelihood estimation algorithm is used for training.

[0017] As preferred, S3 includes extracting the average single cell voltage sequence Smean and the interval of about 1 minute before and after the feature point of each single cell voltage time sequence Si ∈ [S1, S2, …, Sn] as a shapelet candidate set. Through unsupervised shapelet extraction, asynchronous signals can be aligned. The average single cell voltage Vmean_i, at each time in the time window is calculated to form the average single cell voltage sequence Smean, i ∈ [1, t]. The feature points refer to the maximum value point Vmax, the minimum value point Vmin, the mean value point (or the point closest to the mean) Vmin, the time series first derivative (or the closest) zero point dV0, the time series first derivative maximum value point dVmax, the time series first derivative minimum value point dVmin, the time series second derivative (or the closest) zero point ddV0, the time series second derivative maximum value point ddVmax, and the time series second derivative minimum value point ddVmin.

[0018] As preferred, S3 includes continuing data processing on the shapelet candidate set and calculating the Mahalanobis distance MD of each single cell feature subsequence set and the same type of subsequence in the average voltage feature subsequence set. Only the subsequence near the key point of the single cell voltage time sequence is extracted to calculate the distance, reducing the calculation amount. The continuing data processing includes that if the number of a certain type of feature point k >= 2 in the sequence, the features at the first and last time points of this type of feature point are taken; if there is no certain type of feature point in the sequence, the substitution method of this type of feature point t_sub in the sequence is that the average value tf_mean of the time points [tf_k, tf_l, …, tf_n] of this type of feature point of other sequences is taken. If all sequences have no certain type of feature point, this type of feature point is removed. That is, for each single cell voltage sequence, the value of each type of feature point is at least null and at most 2, and a subsequence segment of a certain length is extracted near the feature point. For the average voltage sequence Smean and each single cell voltage sequence Si, the feature subsequence set SSi = {ss1, ss2, …, ssp} is extracted, and p is the maximum value of the number of extracted feature points in all single cell voltage sequences. Focusing on the subsequence interval prone to abnormality avoids the interference of a large number of same subsequences in the full time sequence distance calculation. The shapelet extraction and abnormality detection algorithm based on Mahalanobis distance used can eliminate the noise and errors of single time points.

[0019] As preferred, S4 includes parameters in the adjustment algorithm, and the adjustment parameters include the maximum height of the tree and the number of trees. The Isolation Forest algorithm based on Mahalanobis distance can accurately detect abnormal battery cells.

[0020] As preferred, S5 includes verifying the highest single battery voltage sequence and the average voltage sequence in S3 and S4, and verifying the lowest single battery voltage sequence and the average voltage sequence. The KS two-sample test gives a probability estimate for the difference between the highest single battery voltage sequence and the lowest single battery voltage sequence. The KS test can quickly calculate the wassastein distance used for the voltage distribution offset of the battery cell, providing a cumulative index of the voltage difference of the battery cell for a period of time, and providing a reliable feedback quantity for subsequent battery cell equalization control.

[0021] As preferred, S6 includes calculating the wasserstein distance between the abnormal single battery voltage sequence and the average voltage sequence for the abnormal detection in S3, and calculating the wasserstein distance between the highest or lowest single battery voltage sequence and the average single battery voltage sequence for the abnormal detection in S4. By calculating the Mahalanobis distance between the shapelets, abnormal shapelets can be more significantly distinguished, the detection accuracy can be improved, and the false detection rate can be reduced. The wasserstein distance can be used to measure the voltage difference of the single battery cell for a period of time.

[0022] The present application has the following advantages:

[0023] (1) By detecting the single battery voltage discharge sequence for a period of time during the discharge process of the vehicle, the error, noise and time delay of relying on detecting the voltage measurement value at a single point in time are avoided. The feature point shapelet extraction and the abnormal detection algorithm based on Mahalanobis distance can eliminate the noise and error at a single point in time, focus on the subsequence interval prone to abnormality to avoid the interference of a large number of same subsequences in the full time sequence distance calculation, and the Isolation Forest algorithm based on Mahalanobis distance can accurately detect abnormal battery cells; (2) The calculation amount is greatly reduced compared with the full time sequence distance calculation and the basic shapelet search algorithm based on brute force search; (3) The KS test can quickly calculate the wassastein distance used for the voltage distribution offset of the battery cell, providing a cumulative index of the voltage difference of the battery cell for a period of time, and providing a reliable feedback quantity for subsequent battery cell equalization control. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0025] Figure 1 is a method step diagram of the present application. DETAILED DESCRIPTION

[0026] The specific embodiments of the present application are described below by way of specific examples, and other advantages and effects of the present application can be easily understood by those skilled in the art from the contents disclosed in the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] As shown in FIG. 1, in a preferred embodiment, the present application discloses a single battery consistency detection algorithm based on a vehicle networking big data platform, comprising the following steps:

[0028] 1. Using the streaming batch processing function of the big data processing framework spark, the spark streaming accepts data stream from the message queue kafka, and obtains the single battery voltage data of each vehicle within a certain time window, which can be divided into an investigation mode and a tracking mode. Investigation mode: In the window time before the current time, the single battery voltage signal of all online vehicles is detected. The single battery voltage of each vehicle in the time window forms an n*t matrix [U1, U2, …, Ut], where Ui=[u1, u2, …, un], i∈[1,t], n is the number of single batteries in the battery pack, and t is the length of the time window. After transposition, it becomes a t*n matrix [S1, S2, …, Sn], and each column Si=[v1, v2, …, vt] corresponds to the voltage value of each time point of an electric core. Tracking mode: A fixed maximum length matrix [U1, U2, …, Ut] is created, where t is the length of the time window, and Ui=[u1, u2, …, un]. The sliding observation window is formed by the single voltage of each frame data Ui of the vehicle to be concerned in driving.

[0029] 2. The high, medium and low interval sections of the voltage during the battery discharge process are marked by artificial marking. The voltage and SOC of several cycle samples are used as input variables to train a supervised hidden Markov model. The output label is the corresponding interval section, denoted as s1, s2, s3. The label of the interval section corresponds to the hidden state of the hidden Markov model. The trained model can automatically judge the battery discharge interval section by inputting the total voltage and SOC of the battery. The detection algorithm is started in the middle section s2.

[0030] 3. The average voltage Vmean_i of each single battery in the time window is calculated to form an average single battery voltage sequence Smean, i∈[1,t].

[0031] 4. Extract the average monomer voltage sequence Smean and the monomer cell voltage time sequence Si ∈ [S1, S2……Sn] feature points before and after about 1 minute interval to extract subsequence as shapelet candidate set, the feature points are the maximum value point Vmax, the minimum value point Vmin, the mean value point (or the closest point) Vmin, the time series first derivative (or the closest) zero point dV0, the time series first derivative maximum value point dVmax, the time series first derivative minimum value point dVmin, the time series second derivative (or the closest) zero point ddV0, the time series second derivative maximum value point ddVmax, the time series second derivative minimum value point ddVmin, etc.

[0032] 5. If the number of a certain feature point in the sequence k >= 2, take the feature of the first and last time point of this feature point; if there is no certain feature point in the sequence, the alternative method of this kind of feature point t_sub in the sequence is: the average value tf_mean of the time of this feature of other sequences with this feature is taken in the sequence. If all sequences have no certain feature point, remove this kind of feature point. That is, for each monomer cell voltage sequence, the value of each kind of feature point is at least null and at most 2.

[0033] 6. After the above data processing, extract several time length subsequence fragments near the feature points, for the average voltage sequence Smean and each monomer cell voltage sequence Si, extract the feature subsequence set SSi={ss1, ss2,……ssp}, p is the maximum value of the number of extracted feature points in all monomer cell voltage sequences.

[0034] 7. Calculate the Mahalanobis distance MD (SSi, SSmean) of each monomer cell feature subsequence set and the same kind of feature subsequence in the average voltage feature subsequence set, form the monomer voltage feature distance matrix M, that is [MD1, MD2, …MDn]T, MDi = [D1, D2…Dp], n is the number of monomer cells, and p is the number of extracted features.

[0035] 8. Use the Isolation Forest algorithm to detect anomalies in the feature distance matrix, adjust the parameters in the algorithm: the maximum height and the number of trees to appropriate values, and give the results of anomaly detection with appropriate anomaly score as the standard.

[0036] 9. Use Kolmogorov-Smironv two-sample test to test the highest monomer voltage sequence and the average voltage sequence, which can be used simultaneously with steps 7-8, and test the lowest monomer cell voltage sequence and the average voltage sequence, and use appropriate confidence threshold to detect the cumulative experience function KS statistic as a judgment of whether the data distribution of the highest or lowest monomer voltage and the average monomer voltage is the same or different.

[0037] 10. In step 8, if the abnormal result is detected, calculate the wasserstein distance between the abnormal result and the average voltage sequence; in step 9, if the abnormal result is detected, calculate the wasserstein distance between the highest or lowest voltage sequence and the average voltage sequence.

[0038] In another embodiment, the present application can also use Euclidean distance or dynamic time warping (DTW) distance as the similarity calculation metric when calculating the u-shapelets similarity. When detecting the distance feature matrix of shaplets, a support vector machine, a local outlier factor (LOF), or an elliptical model fitting can also be used. When calculating the difference of the single battery voltage sequence distribution, JS divergence or KL divergence can also be used.

[0039] Although the present application has been described in detail with general description and specific embodiments above, some modifications or improvements can be made on the basis of the present application, which is obvious to those skilled in the art. Therefore, these modifications or improvements made on the basis of not deviating from the spirit of the present application, all belong to the scope of the present application.

Claims

1. A monomer battery consistency detection algorithm based on a vehicle networking big data platform, characterized by comprising the following steps: S1, receiving data and preprocessing; S2, using a trained hidden Markov model to determine the voltage interval segment to be detected; By manually labeling the high, medium and low interval segments of the voltage during the battery discharge process, the voltage and SOC of several cycle samples are used as input variables to train a supervised hidden Markov model, and the output label is the corresponding interval segment, denoted as s1, s2, s3; By using the maximum likelihood estimation algorithm, the trained model can automatically determine the battery discharge interval segment by inputting the total voltage and SOC of the battery, and the detection algorithm is started in the middle segment s2; S3, calculating the Mahalanobis distance MD of the same feature subsequence in each monomer battery feature subsequence set and the average voltage feature subsequence set to form a monomer voltage feature distance matrix M; S4, using the isolated forest algorithm to detect the feature distance matrix, and adjusting the parameters in the algorithm; S5, using the Kolmogorov-Smirnov two-sample test to test the highest monomer voltage sequence and the average voltage sequence, and testing the lowest monomer battery voltage sequence and the average voltage sequence, and using a suitable confidence threshold to detect the cumulative empirical function KS statistic as a judgment of whether the data distribution of the highest or lowest monomer voltage and the average monomer voltage is the same or different; S6, if S4 detects an anomaly, calculate the wasserstein distance between the abnormal monomer voltage sequence and the average voltage sequence; if S5 detects an anomaly, calculate the wasserstein distance between the highest or lowest monomer voltage sequence and the average monomer voltage sequence. 2.The single battery consistency detection algorithm based on the Internet of Vehicles big data platform of claim 1, wherein, The S1 includes accepting data stream from message queue, and obtaining monomer battery voltage data of each vehicle within a certain time window. 3.The single battery consistency detection algorithm based on the Internet of Vehicles big data platform of claim 2, characterized in that, S1 also includes detecting the monomer cell voltage signal of all online vehicles and obtaining the parameter matrix, and forming a sliding observation window matrix for each frame of data Ui of the monomer voltage of the vehicle to be concerned during driving. 4.The single battery consistency detection algorithm based on the Internet of Vehicles big data platform of claim 1, wherein, The S3 includes extracting the average monomer voltage sequence Smean and the 1-minute interval before and after the feature point of each monomer battery voltage time sequence Si∈[S1, S2……Sn] as a shapelet candidate set. 5.The monomer battery consistency detection algorithm based on the Internet of Vehicles big data platform of claim 4, characterized in that, S3 includes further data processing of the shapelet candidate set and calculating the Mahalanobis distance MD of the same feature subsequence in each monomer battery feature subsequence set and the average voltage feature subsequence set. 6.The monomer battery consistency detection algorithm based on the Internet of Vehicles big data platform of claim 1 or 5, characterized in that, The S4 includes adjusting the parameters in the algorithm, and the adjustment parameters include the maximum height of the tree and the number of trees.

7. The monomer battery consistency detection algorithm based on the Internet of Vehicles big data platform according to claim 6, characterized in that, The S5 includes testing the highest monomer voltage sequence and the average voltage sequence in S3 and S4, and testing the lowest monomer battery voltage sequence and the average voltage sequence. 8.The single battery consistency detection algorithm based on the Internet of Vehicles big data platform of claim 6, wherein, The S6 includes calculating the wasserstein distance between the abnormal monomer voltage sequence and the average voltage sequence for the abnormal test in S3, and calculating the wasserstein distance between the highest or lowest monomer voltage sequence and the average monomer voltage sequence for the abnormal test in S4.

Images