Battery health assessment method and system based on cloud-edge collaboration

Abstract

The application discloses a kind of based on cloud edge cooperation's battery health degree evaluation method and system, the method includes, edge device detects short-time discharge event in series battery pack operation, collects the voltage time series of each single battery, discharge current and battery temperature, calculates voltage first derivative and second derivative after filtering, extracts absolute derivative characteristic value, and according to the deviation of each single battery relative to group reference value, difference derivative characteristic value is calculated.According to discharge current and battery temperature, the above-mentioned characteristic value is normalized and temperature corrected, and the corresponding state of charge interval is marked.The cloud end writes the characteristic value of multiple short-time discharge events according to state of charge interval in the cumulative characteristic table of each single battery, inputs the cumulative characteristic table into regression model, and obtains the residual capacity estimation value of each single battery.The application can continuously evaluate battery health status using short-time discharge data generated during operation, reduce the dependence on nuclear capacity test.

Description

Technical Field

[0001] This application relates to the field of battery status monitoring and health assessment technology, specifically to a battery health assessment method and system based on cloud-edge collaboration. Background Technology

[0002] Battery banks are an important component of substation DC power systems, typically consisting of multiple individual batteries connected in series. They are used to continuously supply power to protection, control, and communication equipment when AC power is interrupted or station loads are abnormal. Batteries are in a float-charge state for extended periods, and with increasing usage time, problems such as active material degradation, increased internal resistance, and decreased actual capacity may occur. Furthermore, the degree of degradation may vary significantly between individual cells within the same bank; therefore, it is essential to monitor the health status of each individual battery cell in a timely manner.

[0003] Current battery health assessments primarily employ capacity testing, internal resistance testing, or condition assessment based on float charge voltage. Capacity testing requires prolonged deep discharge, has a long implementation cycle, and places high demands on on-site operation and personnel, making it difficult to conduct frequently. Internal resistance and float charge voltage are easily affected by temperature, state of charge, connection status, and measurement errors, and do not consistently correspond to the actual remaining capacity. Some assessment methods based on operational data require relatively complete charge-discharge curves or fixed test conditions. However, substation batteries typically only experience short-duration, shallow-depth discharge events during normal operation, and the capacity information contained in a single event is limited, making it difficult to directly form reliable assessment results.

[0004] Furthermore, the discharge current, ambient temperature, and state of charge vary depending on the event, and the relationship between voltage and state of charge changes with aging during long-term operation, making it difficult to uniformly compare and effectively utilize data collected at different times. Therefore, how to utilize the short-term discharge data naturally generated during the daily operation of batteries to achieve continuous assessment of the remaining capacity of individual cells while reducing the burden of on-site testing remains a problem that existing technologies need to solve. Summary of the Invention

[0005] This application provides a method and system for assessing battery health based on cloud-edge collaboration, in order to at least solve some of the technical problems existing in the related technologies described above.

[0006] According to a first aspect of the embodiments of this application, a method for assessing the health of a battery based on cloud-edge collaboration is provided, comprising: an edge device detecting a short-term discharge event occurring during the operation of a series-connected battery pack; extracting the voltage time series of each individual battery cell during the short-term discharge event period, and recording the discharge current and battery temperature; filtering the voltage time series and calculating the first and second derivatives of the voltage of each individual battery cell, and extracting the absolute derivative feature value of each individual battery cell from it; calculating the absolute derivative feature value of each individual battery cell relative to a reference value within the same short-term discharge event. The differential derivative characteristic value of each individual cell is obtained; the absolute derivative characteristic value and the differential derivative characteristic value are normalized and temperature-corrected according to the discharge current and the battery temperature, and the state of charge interval corresponding to the short-term discharge event is marked; the cloud fills the normalized and temperature-corrected absolute derivative characteristic value and the differential derivative characteristic value of multiple short-term discharge events into the cumulative characteristic table of each individual cell according to the state of charge interval; the cloud inputs the absolute derivative characteristic value and the differential derivative characteristic value in the cumulative characteristic table into the regression model to obtain the remaining capacity estimate of each individual cell.

[0007] As an optional approach, the absolute derivative characteristic values ​​include the mean of the first derivative within the initial discharge time period, the mean of the second derivative within the entire discharge period, and the ratio of the mean of the first derivative within the final discharge time period to the mean of the first derivative within the initial discharge time period; the differential derivative characteristic values ​​include the deviation of the mean of the first derivative of each individual cell relative to the reference value within the group and the deviation of the mean of the second derivative of each individual cell relative to the reference value of the second derivative within the group.

[0008] As an optional approach, when filling the cumulative feature table with the absolute derivative characteristic values ​​and differential derivative characteristic values ​​of multiple short-term discharge events, if different short-term discharge events overlap in the state of charge interval, the cloud calculates the systematic deviation between the normalized temperature-corrected absolute derivative characteristic values ​​of the two events within the overlapping portion. Based on the ratio of the systematic deviation to the rate of change of the absolute derivative characteristic values ​​along the state of charge direction within the overlapping portion, the correction amount for the state of charge label of the later event is determined. The state of charge label of the later event is then shifted along the state of charge direction by the correction amount before being filled into the cumulative feature table.

[0009] As an optional solution, the cloud uses historical data of batteries of the same model to calculate the normal range of systematic deviations within the overlapping area; when the systematic deviations exceed the normal range, the cloud clears the earlier event data in the cumulative feature table of the individual battery and retains only the more recent event data.

[0010] As an optional approach, the normalization specifically involves calculating the absolute derivative eigenvalues ​​and the differential derivative eigenvalues ​​based on the discharge current; the temperature correction involves calculating the normalized absolute derivative eigenvalues ​​and the differential derivative eigenvalues ​​based on the temperature correction coefficient of the Arrhenius relation, and uniformly correcting them to the reference temperature; the activation energy parameter in the temperature correction coefficient is determined by the cloud through regression fitting based on the operating data of the same type of battery under multiple temperature conditions.

[0011] As an optional approach, when multiple event observations exist within the same state of charge interval in the cumulative feature table, the multiple observations are weighted and averaged according to the time decay weight, with events closer to the current time being assigned a higher weight.

[0012] As an optional approach, the regression model is a gradient boosting regression tree model, which is composed of multiple classification regression trees stacked sequentially. The training labels of the gradient boosting regression tree model are derived from the actual capacity percentage of each individual battery cell measured in the nuclear capacity test, and the training inputs are the absolute derivative feature value and the differential derivative feature value of the corresponding individual battery cell in the cumulative feature table at the time of the nuclear capacity test.

[0013] As an optional approach, the cloud calculates a confidence index based on the proportion of the number of state-of-charge intervals covered by existing observation data in the cumulative feature table to the total number of state-of-charge intervals; when the confidence index is lower than a set threshold, the interval estimate of the remaining capacity estimate is output; when the confidence index reaches the set threshold, the point estimate and error range of the remaining capacity estimate are output.

[0014] As an optional approach, after the site performs the capacity test, the measured capacity value is transmitted back to the cloud. The cloud compares the measured capacity value with the corresponding estimated remaining capacity value, and uses the measured capacity value and the corresponding cumulative feature table status as new training samples to update the regression model. The cloud verifies the consistency between the differential derivative feature value and the capacity ranking based on the measured capacity ranking of each individual cell in the same group. For inconsistent individual cells, the contribution of their differential derivative feature values ​​is reduced in subsequent training, and the updated model parameters are sent to the edge device.

[0015] According to a second aspect of the embodiments of this application, a cloud-edge collaborative battery health assessment system is also provided, including an edge device and a cloud communicationally connected to the edge device. The edge device includes: an event detection and feature extraction unit configured to detect short-term discharge events occurring during the operation of a series-connected battery pack, extract the voltage time series of each individual battery cell during the short-term discharge event period, and record the discharge current and battery temperature; calculate the first and second derivatives of the voltage of each individual battery cell after filtering the voltage time series, and extract the absolute derivative feature values ​​of each individual battery cell from them; and a differential feature calculation unit configured to calculate the absolute derivative feature values ​​of each individual battery cell in the same short-term discharge event relative to a reference value within the group. The system includes: a deviation calculation unit for the differential derivative characteristic value of each individual cell; a feature correction and interval labeling unit configured to normalize and temperature-correct the absolute derivative characteristic value and the differential derivative characteristic value based on the discharge current and the battery temperature, and label the state of charge interval corresponding to the short-term discharge event; and a cloud-based system including: a feature accumulation unit configured to fill the normalized and temperature-corrected absolute derivative characteristic value and the differential derivative characteristic value of multiple short-term discharge events into the cumulative characteristic table of each individual cell according to the state of charge interval; and a capacity estimation unit configured to input the absolute derivative characteristic value and the differential derivative characteristic value from the cumulative characteristic table into a regression model to obtain the remaining capacity estimate of each individual cell.

[0016] This application utilizes naturally occurring short-term discharge events during battery pack operation to conduct health assessments, eliminating the need for frequent deep discharge capacity testing and reducing the impact of on-site testing on the normal operation of the DC power supply system. By extracting the first and second derivatives of the voltage of each individual battery cell and their variation relationships, and constructing differential derivative features relative to the intra-group benchmark value using the shared discharge current of each cell in the series-connected battery pack, the interference of discharge condition differences on individual cell state comparisons can be reduced, improving the identification capability of degraded cells. After current normalization and temperature correction of the derivative features, data from multiple short-term discharge events are accumulated according to the state of charge interval, and the state of charge labeling deviation is corrected based on feature changes within overlapping intervals, enabling local data obtained under different periods and operating conditions to form continuous capacity features. The cloud utilizes multi-site data for model training and updates, while edge devices handle event detection and feature calculation. Under controlled communication and computational load, it can continuously output individual battery remaining capacity estimates and provide corresponding confidence information based on data coverage, thus providing a more reliable basis for battery maintenance and replacement.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Furthermore, no embodiment in this disclosure is required to achieve all the effects described above. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0019] Figure 1 A flowchart illustrating a cloud-edge collaborative battery health assessment method provided in this embodiment of the disclosure; Figure 2 A flowchart illustrating the extraction of absolute derivative feature values ​​for each individual cell provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating the process of labeling state of charge intervals as provided in this embodiment of the disclosure; Figure 4 A schematic block diagram of a cloud-edge collaborative battery health assessment system provided in this embodiment of the disclosure; Figure 5 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0021] This embodiment provides a cloud-edge collaborative battery health assessment method applicable to battery banks composed of multiple valve-regulated lead-acid batteries connected in series in a substation DC power supply system. Online monitoring equipment is deployed at the substation side to continuously collect the terminal voltage of each individual battery, the total discharge current of the battery bank, and the battery surface temperature. The sampling interval can be configured from 1 to 5 seconds. The online monitoring equipment also acts as an edge computing node, possessing data acquisition, local storage, and conventional numerical calculation capabilities. The cloud platform connects multiple substation edge nodes via a communication network, aggregating battery operation data from each site and completing cross-site model training and data analysis. Substation batteries are in a float-charge state for extended periods during normal operation. Under conditions such as AC power interruption, station load fluctuations, or protection device activation, they experience various short-term discharge events. These events are short in duration and shallow in depth, and traditionally, they are only recorded as operational events and not included in health assessments. This method extracts voltage derivative features from these naturally occurring short-term discharge events, calculates differential derivative features using the condition that the same group of cells in the series topology share the discharge current, and accumulates and splices the features of multiple events in the cloud according to the state of charge interval, and outputs the estimated remaining capacity of each cell through a regression model.

[0022] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0023] Please see Figure 1 , Figure 1 This is a flowchart of a battery health assessment method based on cloud-edge collaboration according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps: S1, the edge device detects short-term discharge events that occur during the operation of the series-connected battery pack, extracts the voltage time series of each individual battery during the short-term discharge event period, and records the discharge current and battery temperature; after filtering the voltage time series, it calculates the first and second derivatives of the voltage of each individual battery, and extracts the absolute derivative characteristic value of each individual battery from it. S2, calculate the differential derivative characteristic value of each individual cell based on the deviation of the absolute derivative characteristic value of each individual cell in the same short-term discharge event from the reference value within the group; S3, normalize and temperature-correct the absolute derivative characteristic value and the differential derivative characteristic value based on the discharge current and the battery temperature, and mark the state of charge interval corresponding to the short-term discharge event; S4, the cloud fills the normalized and temperature-corrected absolute derivative characteristic values ​​and differential derivative characteristic values ​​of multiple short-time discharge events into the cumulative characteristic table of each individual cell according to the state of charge interval; S5, the cloud inputs the absolute derivative feature value and the differential derivative feature value from the cumulative feature table into the regression model to obtain the estimated remaining capacity of each individual battery cell.

[0024] In some embodiments, for step S1, please refer to Figure 2 , Figure 2 A flowchart illustrating the extraction of absolute derivative eigenvalues ​​of individual cells provided in an embodiment of this disclosure is shown, as follows: Figure 2 As shown in box 201, the edge device detects short-term discharge events that occur during the operation of the series-connected battery pack and extracts the voltage time sequence of each individual battery cell during the short-term discharge event period.

[0025] The edge device monitors the total discharge current of the battery pack in real time. When the current direction changes from charging to discharging and the discharge duration exceeds the minimum duration threshold, a short-term discharge event is determined to have occurred.

[0026] Among them, the minimum duration threshold By sampling interval And the ratio of the change of the derivative The minimum number of data points required is determined jointly. Let the number of effective data points required for calculating the first derivative using the central difference method be... (Right now , , The smooth length of the sliding window is (Taking odd numbers), to reliably calculate the average derivative values ​​of the initial and final stages of discharge, each stage requires at least... There are 10 effective derivative values. Therefore, the total number of effective data points is at least 100. The corresponding discharge duration is at least To leave room for error, set seconds, of which, This is the length of the sliding window; the default value is... .

[0027] For example, when the sampling interval is 2 seconds, at least 15 consecutive sampling points are required to complete the differential operation and feature extraction, and the corresponding minimum duration threshold is set to 30 seconds. The edge device automatically captures the voltage time series of all individual cells during the period from the start of discharge to the end of discharge, and simultaneously records the total discharge current series and the surface temperature series of each individual cell. For cells with lower capacity, the ion concentration gradient at the electrodes is established more quickly in the early stage of discharge, and the nonlinear characteristics of the voltage drop rate appear earlier. Although the discharge depth of a single short-term discharge event is shallow and the amount of information contained is limited, after multiple events accumulate in different state of charge intervals, a complete feature profile related to the capacity state can be gradually pieced together.

[0028] The edge devices employ an event-triggered data processing model. During normal periods without discharge events, only the voltage and current data from the most recent few sampling points are cached in a circular buffer, without performing derivative calculations. Once a discharge event is detected, the complete discharge period data is extracted from the buffer for subsequent feature extraction operations. After the calculation is complete, the memory space occupied by the original voltage sequence is immediately released to control the memory overhead of the edge devices.

[0029] In box 202, the first and second derivatives of the voltage of each individual cell are calculated after filtering the voltage time series. For the voltage time series extracted from each individual cell, a moving average filter is first applied to suppress sampling noise. The filter window length can be configured to 5 to 11 sampling points. Let the first sampling point be... Only single cell battery in the first The filtered voltage at each sampling time is The sampling interval is (Unit: s), calculate the first and second derivatives using the central difference method: in This represents the first derivative of voltage, with dimensions in V / s; The second derivative of voltage is represented by V / s². The derivative sequence is smoothed by applying a sliding mean of the same window length to reduce the high-frequency noise amplified by the difference operation. The central difference method is a standard discrete approximation of continuous derivatives. The rate of change is obtained by dividing the difference between two adjacent sampling points by the time interval.

[0030] In box 203, the absolute derivative feature values ​​of each individual cell are extracted from the first and second derivatives of the voltage. Three absolute derivative feature values ​​are extracted from the smoothed derivative sequence, the first of which is the mean of the first derivative during the initial discharge phase. Collect all data within the first 30 seconds after the discharge begins. The arithmetic mean of the terms reflects the rate of voltage drop of the cell in the initial stage of discharge; the second term is the mean of the second derivative. Take all during the entire discharge period The arithmetic mean reflects whether the rate of voltage change itself is accelerating or slowing down during the voltage drop process; the third term is the derivative change ratio. Defined as the mean of the first derivative within the last 30 seconds of the discharge phase. The average first derivative within the first 30 seconds of the discharge The ratio: , Since it is a dimensionless quantity, this feature is only calculated when the duration of the discharge event exceeds 60 seconds; for events lasting less than 60 seconds, only the first two features are extracted, and the average discharge current of this event is recorded. (Unit: A) Average battery surface temperature (Unit: K), discharge duration and float charge steady-state voltage of each cell before the event. All of the above calculations are standard numerical calculations and can be performed on edge devices.

[0031] In some embodiments, for step S2, the individual cells of the battery pack are connected in series, and the current flowing through all cells is exactly the same during the same short-term discharge event. The temperature difference between the cells in the battery compartment is typically no more than 2 to 3 degrees Celsius. Under these conditions, the differences in the voltage derivative responses of the individual cells during the same event almost entirely stem from the differences in the electrochemical states of each cell, and the effects of current and temperature on different cells are naturally canceled out when comparing them.

[0032] After extracting the absolute derivative feature values, the edge device calculates the deviation of each individual cell in the same event relative to the group's benchmark value, obtaining the differential derivative feature value. The group's benchmark value is the median of the absolute derivative feature values ​​of all individual cells in the same short-term discharge event. Specifically, taking the mean of the first derivative as an example, the mean of the first derivative of all cells in the initial stage of discharge in that event is taken. The median is the benchmark value. The eigenvalues ​​of the difference first derivative of each monomer are: Similarly, take the average of the second derivatives of all individual variables in this event. the median of Using the baseline value, the eigenvalues ​​of the differential second derivative of each monomer are: The median, rather than the arithmetic mean, is chosen as the benchmark because typically only a few cells in a battery pack exhibit significant degradation. The median is not sensitive to these outliers and better represents the normal level of the pack. (Difference first derivative eigenvalue) The dimension is V / s, and the eigenvalue of the second derivative of the difference is... Its dimension is V / s².

[0033] In some embodiments, for step S3, please refer to Figure 3 , Figure 3A flowchart illustrating the state-of-charge (SOC) interval labeling process provided in an embodiment of this disclosure is shown, as follows: Figure 3 As shown in box 301, normalization is performed based on the discharge current. Different short-term discharge events have different discharge currents; for batteries of the same capacity, the voltage drop rate will inevitably be faster under a larger discharge current. To eliminate the influence of current differences on the eigenvalues, the obtained absolute derivative eigenvalues ​​and the obtained differential derivative eigenvalues ​​are divided by the average discharge current of the event, respectively. Taking the mean of the first derivative during the initial stage of discharge as an example, after normalization, it is: Derivative change ratio Since it is a dimensionless ratio, the discharge current has already canceled out in the numerator and denominator, so no normalization is required.

[0034] In box 302, temperature correction is performed based on the battery temperature. As temperature increases, ion migration in the electrolyte accelerates, internal resistance decreases, and the voltage drop rate slows down under the same capacity and current conditions. This is to unify the characteristic values ​​under different temperature conditions to the same reference temperature. (For example, take 298.15K, i.e. 25°C), and multiply the normalized absolute derivative eigenvalues ​​and difference derivative eigenvalues ​​by a temperature correction factor based on the Arrhenius relation. : in The activation energy of the reaction (unit: J / mol). The universal gas constant is 8.314 J / (mol·K). The average battery surface temperature (in K) for this event, where the exponential term contains The dimensions of each item are J / mol and J / (mol·K)·K=J / mol, respectively. The activation energy is a dimensionless coefficient. Based on operational data of the same battery model under known health conditions at multiple temperature conditions, the cloud performs analysis... and The slope of the linear regression fit is determined, and this slope is multiplied by... Get The typical value range is 15 to 25 kJ / mol. For initial deployment, the value can be calculated using the temperature coefficient provided by the battery manufacturer, which will not be elaborated here.

[0035] The corrected normalized first derivative eigenvalues ​​are denoted as The dimension remains V / (A·s); the temperature correction method for the other derivative characteristics is the same.

[0036] In box 303, the state of charge interval corresponding to the short-term discharge event is marked. The float charge steady-state voltage of each cell before the event occurs is also noted. Approximating the open circuit voltage (OCV), and based on the correspondence between the OCV and the state of charge (SOC) of valve-regulated lead-acid batteries, the state of charge at the start of discharge is estimated by referring to a table. The OCV-SOC correspondence represents the known characteristic data of this battery model, typically provided by the manufacturer at the time of shipment or obtained through calibration tests, and stored in the edge device in the form of a lookup table. The state of charge at the moment of discharge termination... Estimated using the ampere-hour integral method: in Discharge duration (in hours). Given the battery's rated capacity (in Ah), the observed SOC interval for this event is determined as follows: .

[0037] In actual implementation, the scope of the SOC of interest is divided into equal parts. The event is divided into eight sub-intervals, for example, from 60% to 100%, with each sub-interval spanning 5 percentage points. The normalized temperature-corrected absolute derivative eigenvalues ​​and differential derivative eigenvalues ​​of each event are assigned to the corresponding sub-intervals covered by its observed SOC interval. After the above processing, each short-term discharge event generates a set of characteristic records for each individual, including: the SOC sub-interval number covered by the event, and three normalized temperature-corrected absolute derivative eigenvalues ​​(…). , , ) and two normalized temperature-corrected differential derivative eigenvalues ​​( , The edge device records and uploads this feature to the cloud.

[0038] In some embodiments, for step S4, the cloud maintains a cumulative feature table for each individual battery cell, i.e., a two-dimensional data table organized according to SOC intervals. Each SOC sub-interval is indexed by its row number and by its column index by its five derivative feature types. The system stores the normalized temperature-corrected feature values ​​of each event uploaded within each SOC sub-interval. Each time the cloud receives a feature record of an event, it fills the corresponding SOC sub-interval row with its five feature values.

[0039] When a SOC subinterval already contains records of a previous event, multiple observations within the same subinterval are weighted and averaged using time decay weights. The weighting method is as follows: Let the number of days since a given event be . The time decay weight of this event is ,in The decay time constant (unit: days); The value is determined based on the average frequency of short-term discharge events at the site. Specifically, this is done by calculating the average interval in days between short-term discharge events at the site over the past year, and then multiplying this interval by 3 to 5 times. Initial value; sites with low event frequency Larger values ​​are chosen to retain more historical information, while smaller values ​​are chosen when the event frequency is high to make recent observations more prominent. The typical range is 30 to 90 days.

[0040] As an example, suppose that in the cumulative feature table of a certain single battery cell, SOC sub-interval number 4 (corresponding to SOC 80% to 85%) received observations of 3 events in the past 60 days, respectively 12 days, 35 days, and 58 days from the current date. If the time decay period is set to 60 days, then the time decay weights for the three events are respectively... , , The eigenvalues ​​of the first derivative after normalized temperature correction for these three events are weighted and averaged according to the above weights, and then filled into the corresponding positions in the sub-interval.

[0041] As time goes on, short-term discharge events covering different SOC sub-intervals gradually accumulate, and the rows of the cumulative feature table are gradually filled.

[0042] In one embodiment, during the accumulation of multiple events, the observed SOC intervals of different events may sometimes partially overlap. For example, if event A covers SOC sub-interval numbers 3 to 5 (corresponding to SOC 75% to 90%), and event B covers sub-interval numbers 4 to 6 (corresponding to SOC 80% to 95%), then sub-intervals 4 and 5 are the overlapping intervals of the two.

[0043] In the aforementioned steps, the SOC labeling relies on the float charge voltage to estimate the SOC. However, the correspondence between the float charge voltage and SOC will shift as the battery ages, causing inconsistent errors in the SOC labeling for different events. Therefore, this embodiment can correct this deviation by utilizing the consistency of derivative characteristics in the overlapping interval. Within the overlapping interval, the absolute derivative characteristic values ​​of the two events after normalized temperature correction should be close. If there is a systematic deviation, it indicates that the SOC labeling is misaligned.

[0044] Specifically, within the overlapping SOC sub-intervals, the mean of the difference between the eigenvalues ​​of the first derivative after normalized temperature correction for two events is calculated: in The number of overlapping subintervals. and Events A and B are respectively in the th... Normalized temperature-corrected first-order derivative eigenvalues ​​of overlapping subintervals The dimension is V / (A·s). At the endpoints of the overlapping interval, the rate of change of the normalized first derivative eigenvalues ​​along the SOC direction is estimated by dividing the difference in eigenvalues ​​of event A by the corresponding SOC span. The dimensionless value is V / (A·s) per percentage point of SOC. and The ratio is used as the SOC annotation translation correction for event B. : , Representing the SOC percentage points, all SOC labels in event B are shifted along the SOC axis. Then, the cumulative feature table is filled in. This correction uses the continuity and monotonicity of the derivative feature along the SOC axis as an alignment basis, and does not completely depend on the SOC estimation accuracy of the float charge voltage; for example, when for V / (A·s) for V / (A·s) per percentage point of SOC The correction is 2 percentage points, meaning that the SOC label of event B shifted by 2 percentage points towards the high SOC direction. After correction, it was pulled back by 2 percentage points.

[0045] In some embodiments, the cloud platform statistically analyzes the normal range of systematic deviations within overlapping intervals based on historical data of batteries of the same model. Specifically, it collects the deviation values ​​of overlapping intervals for batteries of the same model at multiple sites, calculates their mean and standard deviation, and uses the mean plus or minus twice the standard deviation as the boundary of the normal range. When the value exceeds this normal range, it indicates that the capacity state of the individual battery may have changed rapidly between the two events, rather than due to SOC labeling error. In this case, the cloud will clear the earlier event data from the cumulative feature table of the individual battery and retain only the more recent event data to start accumulating again, thus avoiding mixing feature values ​​from different degradation stages.

[0046] In some embodiments, for step S5, the cloud extracts the model input from the accumulated feature table of each individual battery cell, wherein the input is a feature vector, derived from... It is composed of features from SOC sub-intervals, each sub-interval containing 5 feature values: the mean of the first derivative after normalized temperature correction. Mean of second derivative The ratio of the change of the derivative , first derivative deviation , difference second derivative deviation .when When the value is 8, the total dimension of the feature vector is 40. If some SOC sub-intervals have no observed data, the corresponding positions are marked as missing values, and the model output is a scalar. This indicates the percentage of the remaining capacity of the single battery cell relative to its rated capacity (unit: %).

[0047] As battery capacity decreases, the amount of active material on the electrode decreases, the ion concentration gradient on the electrode surface is established more quickly during discharge, the voltage drop rate increases, and the nonlinear characteristics become more pronounced. The magnitude of the aforementioned derivative characteristic values ​​and the shape of their variation along the SOC axis directly reflect the differences in these electrochemical processes. Therefore, there is a correlation between the input characteristics and the remaining output capacity.

[0048] In one embodiment, the regression model employs Gradient Boosting Regression Tree (GBRT), an ensemble learning method consisting of multiple Classification and Regression Trees (CART) stacked sequentially. The model's predictions are a weighted sum of the outputs of each tree. in The input feature vector (dimension) ), For the first Output of a CART regression tree (in %) This is the learning rate (also known as the shrinkage factor), with a value ranging from 0.01 to 0.1. The total number of trees; the basic principle of gradient boosting is that a new tree is added in each iteration, and the target tree fits is the previous one. The negative gradient direction of the residual between the predicted value and the true label is accumulated across multiple trees, thereby gradually reducing the overall prediction error. CART regression trees are tree structures that fit target values ​​by recursively bisecting the input feature space. Each internal node selects a feature and a splitting threshold to divide the samples into two groups, and the leaf nodes output the mean of the samples falling into that region as the predicted value.

[0049] In one embodiment, the parameters of each CART regression tree are configured such that the maximum depth can be configured to 4 to 8 layers, the minimum number of leaf nodes can be configured to 5 to 20 samples, the splitting criterion adopts Friedman mean squared error, and each iteration randomly selects a subset from the training samples to construct the current tree, with the sampling ratio configured to 0.7 to 0.9 to reduce the risk of overfitting.

[0050] When CART regression trees split nodes, missing value samples can be automatically handled through a substitution splitting mechanism. That is, when a certain feature is missing, other features that are most consistent with the splitting result of that feature are used as substitutes for splitting. This feature allows GBRT to directly accept inputs with missing feature values ​​in some SOC sub-intervals without additional missing value imputation steps, which is compatible with the actual data situation of gradually filling the cumulative feature table.

[0051] The training process uses the Huber loss function: This represents the actual remaining capacity percentage (in %). The model predicted value (in %). Transition threshold (in %). Compared to mean squared error, Huber loss affects the threshold for bias exceeding... Applying a linear rather than quadratic penalty to the samples reduces the impact of a few noisy training samples on the model parameters. The method for determining the value is as follows: first, train a preliminary GBRT model using mean squared error loss, calculate the absolute value of the prediction residuals for each sample on the validation set, and take the 80th percentile as the value. The typical value range is 3% to 8%.

[0052] In one embodiment, the training data originates from historical data from multiple stations aggregated in the cloud. For batteries with both short-term discharge event records and capacity test results, an input feature vector is constructed using the absolute derivative and differential derivative eigenvalues ​​from the cumulative feature table of the battery at the time of the capacity test, and a training sample is formed using the actual capacity percentage measured in the capacity test as the label. Specifically, after a station completes a capacity test, the time of the capacity test is used as the base time. The cloud retrieves the most recently updated cumulative feature table of the battery up to that point in time. This feature table is a single-dimensional feature table. The two-dimensional feature matrix is ​​flattened into a matrix of length . One-dimensional feature vector This serves as the input for the training samples. Simultaneously, the actual remaining capacity percentage of the battery, measured in this capacity test, is used. (Unit: %) serves as the label for this sample. In this way, training samples are constructed one by one from all historical data aggregated in the cloud, forming the original training set.

[0053] During training It can be configured to 0.05. The number of trees can be configured from 200 to 500. An early stopping strategy is adopted, that is, 20% of the training data is retained as a validation set. After each iteration, the Huber loss on the validation set is calculated. If the validation set loss does not decrease for 20 consecutive iterations, training stops, and the model with the last decrease in loss is taken as the final model. Optionally, during training, some feature values ​​of SOC sub-intervals in the input feature vector can be randomly set to missing values ​​to enhance the model's adaptability to incomplete cumulative feature tables.

[0054] Understandably, the training and updating process of the GBRT model can be completed entirely in the cloud. Only a lightweight model inference engine is deployed on the edge device to construct feature vectors and perform forward computation based on the model parameters (i.e., the structure of each CART tree and the weights of the leaf nodes) sent from the cloud to output the remaining capacity estimate. This design can ensure that the computing power consumption and storage overhead of the edge device are at an extremely low level, which is compatible with the hardware conditions of embedded online monitoring equipment in substations.

[0055] In some embodiments, the cumulative characteristic table of each individual cell has different fill levels at different times, and the cloud calculates a confidence index while outputting the remaining capacity estimate. : in The total number of SOC sub-intervals, For the first The number of existing observed events in each sub-interval. For indicator functions, when If the value is 1, then the value is 0; otherwise, the value is 0. For the first The quality weights of each sub-interval. If the observation data of this sub-interval has undergone the overlapping interval cross-validation in the aforementioned steps, then... Take 1.0; unverified. Take 0.7, The value is determined based on the fact that validated SOC annotations have smaller errors and more reliable feature values, thus contributing more to the confidence calculation. This value can be further adjusted based on the ratio of the model estimation error under validated conditions to that under unvalidated conditions. It is a dimensionless quantity, and its value ranges from 0 to 1.

[0056] when When the capacity is below a set threshold, the cloud only outputs an interval estimate of the remaining capacity; when... When the set threshold is reached, output the specific point estimate. And the corresponding error range; this threshold is determined by the cloud based on historical verification data, i.e., statistically different The mean of the model estimation error is selected based on the value, choosing the minimum value that makes the mean estimation error lower than an acceptable level (e.g., ±5 percentage points). The value is used as a threshold, for example, 0.5.

[0057] In one embodiment, the model update process after the capacity test includes: after a capacity test is performed at a site, the measured capacity values ​​of each individual cell are transmitted back to the cloud. The cloud compares the measured capacity values ​​with the estimated remaining capacity of the corresponding cell, and incorporates the measured capacity values ​​along with the cumulative feature table state of the cell at the time of the capacity test into the training set as new training samples to retrain the GBRT model. The capacity test also provides the actual capacity ranking of each individual cell in the same group. The cloud uses this to verify the consistency between the differential derivative feature values ​​and the capacity ranking, i.e., checking whether cells with large differential derivative deviations in the cumulative feature table correspond to cells with low capacity in the capacity test. If an inconsistency occurs (e.g., a cell has normal capacity but abnormal voltage derivative due to loose terminal connection), the cloud marks the cell and reduces its contribution to the differential derivative feature values ​​in subsequent training. The updated model parameters and activation energy parameters in temperature correction are then updated accordingly. The OCV-SOC correspondence table is distributed from the cloud to each edge device.

[0058] This method starts with the short-term discharge events that naturally occur in substation batteries during operation. Utilizing the sensitivity of voltage derivatives to capacity status, it accumulates and stitches together the limited information of shallow discharge characteristics from a single event across SOC intervals, gradually constructing the capacity-related characteristic profiles of each individual battery in the cloud. Differential derivative features under series topology conditions further eliminate the common-mode effects of current and temperature, improving the discriminative power of capacity differences between individual cells. Cross-validation of overlapping intervals uses the continuity of derivative features to correct SOC labeling errors caused by aging-related shifts in float charge voltage, enabling reliable stitching of event data from different periods. The regression model based on these accumulated features can continuously output the remaining capacity estimate of each individual battery without performing capacity testing, providing a reference for the health management of substation battery banks.

[0059] Please see Figure 4 , Figure 4 This is a structural block diagram of a cloud-edge collaborative battery health assessment system provided in an embodiment of this application. Figure 4As shown, the system includes an edge device 401 and a cloud 402 communicatively connected to the edge device. The edge device 401 includes: an event detection and feature extraction unit 4010, configured to detect short-term discharge events occurring during the operation of the series-connected battery pack, extract the voltage time series of each individual battery cell during the short-term discharge event period, and record the discharge current and battery temperature; calculate the first and second derivatives of the voltage of each individual battery cell after filtering the voltage time series, and extract the absolute derivative feature value of each individual battery cell from it; and a differential feature calculation unit 4011, configured to calculate the differential derivative of each individual battery cell based on the deviation of the absolute derivative feature value of each individual battery cell in the same short-term discharge event from the reference value within the group. The feature value; feature correction and interval labeling unit 4012 is configured to normalize and temperature-correct the absolute derivative feature value and the differential derivative feature value based on the discharge current and the battery temperature, and label the state of charge interval corresponding to the short-term discharge event; the cloud 402 includes: a feature accumulation unit 4021, configured to fill the normalized and temperature-corrected absolute derivative feature value and differential derivative feature value of multiple short-term discharge events into the cumulative feature table of each individual battery according to the state of charge interval; a capacity estimation unit 4022, configured to input the absolute derivative feature value and differential derivative feature value in the cumulative feature table into the regression model to obtain the remaining capacity estimate of each individual battery.

[0060] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0061] Based on the same inventive concept, this application also provides an electronic device, the method corresponding to which can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. For example... Figure 5 As shown, Figure 5 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods and / or technical solutions of the foregoing embodiments of the present application.

[0062] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processor, it performs the functions defined in the methods of this application.

[0063] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application.

[0064] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A method for assessing battery health based on cloud-edge collaboration, characterized in that, include: The edge device detects short-term discharge events that occur during the operation of the series-connected battery pack, extracts the voltage time sequence of each individual battery cell during the short-term discharge event period, and records the discharge current and battery temperature. After filtering the voltage time series, the first and second derivatives of the voltage of each individual cell are calculated, and the absolute derivative characteristic value of each individual cell is extracted from them. Based on the deviation of the absolute derivative characteristic value of each individual cell from the group's reference value in the same short-term discharge event, the differential derivative characteristic value of each individual cell is calculated. The absolute derivative characteristic value and the differential derivative characteristic value are normalized and temperature-corrected according to the discharge current and the battery temperature, and the state of charge interval corresponding to the short-term discharge event is marked. The cloud-based system fills the normalized and temperature-corrected absolute derivative characteristic values ​​and differential derivative characteristic values ​​of multiple short-term discharge events into the cumulative characteristic table of each individual cell according to the stated state of charge interval. The cloud inputs the absolute derivative feature value and the differential derivative feature value from the cumulative feature table into the regression model to obtain the estimated remaining capacity of each individual battery cell.

2. The method according to claim 1, characterized in that, The absolute derivative characteristic values ​​include the mean of the first derivative within the initial discharge time period, the mean of the second derivative within the entire discharge period, and the ratio of the mean of the first derivative within the final discharge time period to the mean of the first derivative within the initial discharge time period. The differential derivative characteristic values ​​include the deviation of the mean first derivative of each individual cell relative to the benchmark value within the group, and the deviation of the mean second derivative of each individual cell relative to the benchmark value of the second derivative within the group.

3. The method according to claim 1, characterized in that, When filling the cumulative feature table with the absolute derivative characteristic values ​​and differential derivative characteristic values ​​of multiple short-term discharge events, if different short-term discharge events overlap in the state of charge interval, the cloud calculates the systematic deviation between the normalized temperature-corrected absolute derivative characteristic values ​​of the two events within the overlapping part. Based on the ratio of the systematic deviation to the rate of change of the absolute derivative characteristic values ​​along the state of charge direction within the overlapping part, the correction amount for the state of charge label of the later event is determined. The state of charge label of the later event is then shifted along the state of charge direction by the correction amount before being filled into the cumulative feature table.

4. The method according to claim 3, characterized in that, The cloud platform calculates the normal range of systematic deviations within the overlapping portion based on historical data of batteries of the same model. When the systematic deviations exceed the normal range, the cloud platform clears the earlier event data in the cumulative feature table of the individual battery and retains only the more recent event data.

5. The method according to claim 1, characterized in that, The normalization specifically involves calculating the absolute derivative eigenvalues ​​and the differential derivative eigenvalues ​​based on the discharge current, respectively; the temperature correction involves calculating the normalized absolute derivative eigenvalues ​​and differential derivative eigenvalues ​​based on the temperature correction coefficient of the Arrhenius relation, and uniformly correcting them to the reference temperature; the activation energy parameter in the temperature correction coefficient is determined by the cloud through regression fitting based on the operating data of the same type of battery under multiple temperature conditions.

6. The method according to claim 1, characterized in that, When multiple event observations exist within the same state of charge interval in the cumulative feature table, the multiple observations are weighted and averaged according to the time decay weight, with events closer to the current time being assigned a higher weight.

7. The method according to claim 1, characterized in that, The regression model is a gradient boosting regression tree model, which is composed of multiple classification regression trees stacked in sequence. The training labels of the gradient boosting regression tree model come from the actual capacity percentage of each individual battery cell measured in the nuclear capacity test. The training input is the absolute derivative feature value and the differential derivative feature value of the corresponding individual battery cell in the cumulative feature table at the time of the nuclear capacity test.

8. The method according to claim 1, characterized in that, The cloud calculates a confidence index based on the proportion of the number of state-of-charge intervals covered by existing observation data in the cumulative feature table to the total number of state-of-charge intervals; when the confidence index is lower than a set threshold, it outputs an interval estimate of the remaining capacity estimate; when the confidence index reaches the set threshold, it outputs a point estimate and error range of the remaining capacity estimate.

9. The method according to claim 1, characterized in that, After the site performs the capacity test, the measured capacity value is transmitted back to the cloud. The cloud compares the measured capacity value with the corresponding estimated remaining capacity value, and uses the measured capacity value and the corresponding cumulative feature table status as new training samples to update the regression model. The cloud verifies the consistency between the differential derivative feature value and the capacity ranking based on the measured capacity ranking of each cell in the same group. For inconsistent cells, the contribution of their differential derivative feature values ​​is reduced in subsequent training, and the updated model parameters are sent to the edge device.

10. A battery health assessment system based on cloud-edge collaboration, characterized in that, The system includes an edge device and a cloud communicationally connected to the edge device. The edge device includes: an event detection and feature extraction unit configured to detect short-term discharge events occurring during the operation of the series-connected battery pack, extract the voltage time series of each individual battery cell during the short-term discharge event period, and record the discharge current and battery temperature; calculate the first and second derivatives of the voltage of each individual battery cell after filtering the voltage time series, and extract the absolute derivative feature value of each individual battery cell from it; and a differential feature calculation unit configured to calculate the differential derivative feature value of each individual battery cell based on the deviation of the absolute derivative feature value of each individual battery cell in the same short-term discharge event from the reference value within the group. The feature correction and interval labeling unit is configured to normalize and temperature-correct the absolute derivative feature value and the differential derivative feature value based on the discharge current and the battery temperature, and label the state of charge interval corresponding to the short-term discharge event; the cloud includes: a feature accumulation unit, configured to fill the normalized and temperature-corrected absolute derivative feature value and differential derivative feature value of multiple short-term discharge events into the cumulative feature table of each individual battery according to the state of charge interval; and a capacity estimation unit, configured to input the absolute derivative feature value and differential derivative feature value in the cumulative feature table into a regression model to obtain the remaining capacity estimate of each individual battery.

Images