Battery data analysis method and platform based on cloud edge collaboration

Abstract

The invention relates to the technical field of data processing, and provides a battery data analysis method and platform based on cloud edge collaboration, and the method comprises the steps: obtaining battery operation data collected at an edge side, carrying out the preprocessing, obtaining an edge side processing result, carrying out the feature reconstruction, obtaining a multi-dimensional trend feature and a cross validation index, and carrying out the noise reduction and pairing. After the feedback identifier is obtained, performing increment adjustment with the edge side processing result to obtain an adjusted edge side processing result; and analyzing and verifying the feature analysis result and the adjusted edge side processing result to obtain a comprehensive analysis result. Through fusion analysis and verification of a feature analysis result and an adjusted edge side processing result, the problems of low accuracy and poor reliability of an analysis result caused by the fact that an existing scheme is difficult to capture an edge side short-period dynamic state in time when a battery operation state is frequently changed or an edge side environment is greatly fluctuated are solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data processing, in particular to a battery data analysis method and platform based on cloud-edge collaboration. BACKGROUND

[0002] With the wide popularity of electric vehicles, renewable energy storage systems and mobile devices, the importance of performance state monitoring and operation management of lithium batteries as a key energy carrier is increasingly prominent. Since the battery operation data usually has high-frequency collection, high-dimensional characteristics and strong time sequence correlation, it is difficult to meet the accuracy and response requirements of intelligent analysis by relying only on local devices for processing, therefore, the battery data analysis method based on cloud-edge collaboration has gradually become a research hotspot and industrial application direction.

[0003] In the related technical means, the cloud-edge collaboration framework processes battery data mainly by collecting basic operation data such as voltage, current and temperature of the battery, and performing data cleaning, noise reduction and compression processing to reduce the upload burden. Subsequently, these data are transmitted to the cloud, and the cloud side relies on its powerful computing resources and data integration capabilities to centrally process and uniformly model the data of different battery devices, realizing the collaborative cooperation of edge computing and cloud intelligence, and improving the analysis efficiency and the coverage ability of the global perspective.

[0004] For the above technical solution, although the collaborative structure of edge-side collection and cloud-side centralized analysis can realize cross-device data fusion and intelligent analysis modeling, when the battery operation state changes frequently or the edge-side environment fluctuates greatly, the existing scheme is difficult to capture the short-period dynamics of the edge side in time, and the unified analysis model generated by the cloud is difficult to accurately respond to the specific battery or local state, thereby affecting the accuracy and reliability of the analysis results. SUMMARY

[0005] In order to improve the problem that when the battery operation state changes frequently or the edge-side environment fluctuates greatly, the existing scheme is difficult to capture the short-period dynamics of the edge side in time, resulting in low accuracy and poor reliability of the analysis results, the application provides a battery data analysis method and platform based on cloud-edge collaboration.

[0006] The application provides a battery data analysis method based on cloud edge cooperation, comprising: obtaining battery operation data collected at an edge side, preprocessing the battery operation data to obtain an edge side processing result; uploading the edge side processing result to a cloud end for feature reconstruction to obtain a feature analysis result and a reliability evaluation value; structurally deconstructing the feature analysis result to obtain prediction bias data, using the prediction bias data to correct errors of the reliability evaluation value to generate a judgment factor, performing edge side rapid return and local calibration on the judgment factor to obtain a feedback identifier; based on the feedback identifier and the edge side processing result, performing incremental adjustment to obtain an adjusted edge side processing result; analyzing and verifying the feature analysis result and the adjusted edge side processing result to obtain a comprehensive analysis result.

[0007] As a preferred scheme, the step of obtaining battery operation data collected at an edge side and preprocessing the battery operation data to obtain an edge side processing result comprises: collecting battery operation data in real time through an edge device, time aligning and redundantly compressing the battery operation data to obtain a synchronous data sequence and a compression mapping relationship, performing fluctuation detection and extreme value extraction on the synchronous data sequence to obtain an abnormal point index; performing edge cleaning on the compression mapping relationship and the abnormal point index to obtain cleaned data, and performing short-period change analysis and correlation window construction on the cleaned data to obtain an edge side processing result.

[0008] As a preferred scheme, the step of uploading the edge side processing result to a cloud end for feature reconstruction to obtain a feature analysis result and a reliability evaluation value comprises: uploading the edge side processing result to the cloud end for feature reconstruction to generate a time sequence matrix and a space coupling matrix; extracting a short-term trend, a medium-term trend and a long-term mode based on the time sequence matrix, and constructing a multi-dimensional trend feature according to the short-term trend, the medium-term trend and the long-term mode; performing consistency testing and trend coordination scoring on the space coupling matrix to obtain a cross-validation index, and jointly denoising and dimension pairing the multi-dimensional trend feature and the cross-validation index to generate a feature analysis result and a reliability evaluation value.

[0009] As a preferred scheme, the step of jointly denoising and dimension pairing the multi-dimensional trend feature and the cross-validation index to generate a feature analysis result and a reliability evaluation value comprises: jointly denoising and dimension pairing the multi-dimensional trend feature and the cross-validation index to generate a trend relationship graph and a confidence weight table; constructing a multi-source fusion path based on the trend relationship graph to obtain a trend fusion feature, using the confidence weight table to evaluate the reliability and sensitivity of the trend fusion feature to output a reliability initial value; comparing and analyzing the trend fusion feature and the reliability initial value to generate a feature analysis result and a reliability evaluation value.

[0010] As a preferred solution, the step of structurally deconstructing the feature analysis result to obtain prediction deviation data, using the prediction deviation data to correct errors of the credibility evaluation value to generate a decision factor, and performing edge side fast backhaul and local calibration on the decision factor to obtain a feedback identifier, comprises: performing dimension splitting and time sequence segmentation processing on the feature analysis result to obtain feature contribution degree and trend abnormality segment, using the trend abnormality segment to perform interval mapping on the feature contribution degree to obtain prediction deviation data; correcting high-sensitive features in the credibility evaluation value by using the prediction deviation data to obtain a corrected credibility value and an adjustment factor, performing weight mapping on the corrected credibility value and the adjustment factor to generate a decision factor; performing edge backhaul on the decision factor, and based on historical feedback data, performing fast local calibration on the backhauled decision factor to obtain a response delay parameter and a consistency threshold; constructing a calibration rule table based on the response delay parameter and the consistency threshold, using the calibration rule table to perform secondary analysis on the decision factor to generate a feedback identifier.

[0011] As a preferred solution, the step of performing incremental adjustment based on the feedback identifier and the edge side processing result to obtain an adjusted edge side processing result comprises: analyzing response delay, data drift and trust interval change based on the feedback identifier, generating a drift factor group and a compensation coefficient based on the response delay, the data drift and the trust interval change; using the drift factor group and the compensation coefficient to perform residual fitting on the edge side processing result to obtain a compensation residual sequence and a correction frequency diagram, cross-comparing the compensation residual sequence and edge side historical state features to identify high fluctuation dimensions and high consistency dimensions; constructing a residual mapping curve based on the high fluctuation dimensions, extracting residual prediction weights based on the high consistency dimensions, using the residual mapping curve and the residual prediction weights to construct an enhanced feature set, performing quantization encoding and weight reallocation on the enhanced feature set to generate an enhanced encoding stream; fusing the enhanced encoding stream and the original edge side processing result to obtain a multi-dimensional fusion vector, performing incremental adjustment on the multi-dimensional fusion vector to obtain an adjusted edge side processing result.

[0012] As a preferred solution, the step of analyzing and verifying the feature analysis result and the adjusted edge side processing result to obtain a comprehensive analysis result comprises: extracting a trend fusion feature in the feature analysis result, extracting an enhanced encoding stream in the adjusted edge side processing result, performing dimension correspondence and semantic mapping on the trend fusion feature and the enhanced encoding stream to obtain a difference parameter; performing time section division and fluctuation reorganization on the difference parameter to generate a dynamic consistency profile, performing adaptive segmentation analysis on the dynamic consistency profile to extract a key consistent area and a conflict segment, marking and tracking the key consistent area to obtain a consistency marking value, and performing entropy value calculation and rollback inspection on the conflict segment to obtain a calibration threshold; and performing bidirectional harmonization on the trend fusion feature and the enhanced encoding stream using the consistency marking value and the calibration threshold to output a comprehensive analysis result.

[0013] The application also provides a battery data analysis platform based on cloud edge collaboration, comprising: an acquisition module configured to acquire battery operation data collected at an edge side, and to preprocess the battery operation data to obtain an edge side processing result; an uploading module configured to upload the edge side processing result to a cloud end for feature reconstruction to obtain a feature analysis result and a credibility evaluation value; a correction module configured to structurally deconstruct the feature analysis result to obtain prediction bias data, to use the prediction bias data to correct errors of the credibility evaluation value to generate a decision factor, to perform edge side rapid feedback and local calibration on the decision factor to obtain a feedback identifier; a fitting module configured to perform incremental adjustment based on the feedback identifier and the edge side processing result to obtain an adjusted edge side processing result; and a verification module configured to analyze and verify the feature analysis result and the adjusted edge side processing result to obtain a comprehensive analysis result.

[0014] Compared with the prior art, the application has the following beneficial effects: high accuracy and strong reliability. After the battery operation data collected is preprocessed to obtain the edge side processing result, the structural deconstruction and error correction mechanism of the feature analysis result are used to make the generated decision factor have stronger adaptability and interpretability, and feedback is realized through rapid feedback and local calibration. In addition, the drift factor group and the compensation coefficient generated based on the feedback identifier can dynamically adjust the edge side processing result, improve the accuracy and consistency of local analysis, and improve the problem that the existing scheme is difficult to capture the edge side short period dynamics in a timely manner when the battery operation state frequently changes or the edge side environment fluctuates greatly, resulting in low accuracy and poor reliability of the analysis result. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0016] The structures, proportions, sizes, etc. shown in the drawings of the present specification are only used to cooperate with the content disclosed in the present specification, to be understood and read by those skilled in the art, and are not used to limit the conditions that the present application can be implemented, so they do not have technical significance. Any modification of structure, change of proportion relationship or adjustment of size, without affecting the effects that the present application can produce and the purposes that the present application can achieve, should still fall within the scope of the technical content disclosed by the present application.

[0017] Figure 1 is a flowchart of a battery data analysis method based on cloud edge collaboration provided by an embodiment of the present application; Figure 2 is a structural schematic block diagram of a battery data analysis platform based on cloud edge collaboration provided by an embodiment of the present application.

[0018] Legend of reference signs: 10, battery data analysis platform based on cloud edge collaboration; 11, acquisition module; 12, uploading module; 13, correction module; 14, fitting module; 15, verification module. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. 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.

[0020] The flowchart shown in the drawings is only an example description, not necessarily including all contents and operations / steps, and not necessarily executed in the order described. For example, some operations / steps can also be decomposed, combined or partially merged, so the actual execution order may be changed according to the actual situation.

[0021] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, unless otherwise clear from the context, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0022] It should be further understood that the term "and / or" used in the description and claims of the application means one or more of the associated listed items as well as all possible combinations of the items and includes the combinations.

[0023] The technical solutions of the present application will be further illustrated below in conjunction with the drawings and through specific embodiments.

[0024] Example 1: As Figure 1 The present application provides a battery data analysis method based on cloud edge collaboration, including steps S100 to S500.

[0025] Step S100, acquiring battery operation data collected at the edge side, preprocessing the battery operation data to obtain an edge side processing result.

[0026] In this step, the battery operation data from the battery management system is collected in real time by the edge device, including voltage, current, temperature, charging and discharging state, and residual capacity, etc. original data; Specifically, first, the battery operation data is time stamped and compressed for multi-source redundancy to eliminate timing misalignment and data redundancy problems, to obtain a synchronization data sequence and a compression mapping relationship; Then, based on the synchronization data sequence, a sliding window fluctuation detection is performed, a mutation point is extracted and an extreme value vector is constructed, forming an abnormal point index; The compression mapping relationship and the abnormal point index are used as edge cleaning input, and a rule filtering and median smoothing operation is performed to obtain cleaned data; Next, a periodic trend analysis algorithm (such as autocorrelation coefficient analysis) is used to detect short period changes in the cleaned data, and a sliding correlation window is constructed based on the analysis result, thereby generating an edge side processing result and a preliminary correlation graph; Finally, the stable factor group (such as temperature stability index, internal resistance change range) and the change frequency index (such as voltage fluctuation frequency) in the edge side processing result are extracted, and the two are aggregated for features, to construct a local feature abstract, and the preliminary correlation graph is fused to generate context correlation information with correlation structure.

[0027] For example, in 24 hours of continuous sampling of battery data, the time series is aligned using 5 minutes as a unit, the charging fluctuation point is captured by the extreme value detection algorithm with a sliding window size of 12, and the compressed data is generated based on the moving average method, completing the edge preprocessing and context modeling.

[0028] Step S200, uploading the edge side processing result to the cloud for feature reconstruction, to obtain a feature analysis result and a credibility evaluation value.

[0029] In this step, firstly, the local feature summary uploaded to the cloud and the context association information are uniformly formatted, and the time series matrix and the spatial coupling matrix are generated through feature reconstruction. Specifically, by using dimension compression methods such as principal component analysis (PCA), the multi-dimensional original index is mapped to the main trend axis, the short-term trend (such as voltage jitter), the medium-term trend (such as temperature interval change), and the long-term mode (such as capacity degradation path) are extracted, and the multi-dimensional trend feature is formed. Then, the spatial coupling matrix is subjected to collaborative testing and correlation scoring, and the cross-validation index is generated by combining the state coupling degree between devices and the historical behavior consistency. On this basis, the multi-dimensional trend feature and the cross-validation index are jointly denoised by using empirical mode decomposition (EMD) and wavelet denoising method, and the dimension pairing is performed by using the trend correlation degree and the index complementarity, to obtain the feature analysis result and the reliability evaluation value.

[0030] For example, after 10-dimensional feature compression, it is found that the temperature and voltage variation of multiple batteries have a coupling relationship in the charging state, and by setting the similarity threshold to 0.85 for cross-scoring, a device feature distribution graph with a reliability evaluation value higher than 0.9 is obtained.

[0031] Step S300, the structural deconstruction of the feature analysis result is performed to obtain the prediction bias data, the error of the reliability evaluation value is corrected by using the prediction bias data, the decision factor is generated, the decision factor is quickly returned to the edge side and locally calibrated, and the feedback identifier is obtained.

[0032] In this step, the feature analysis result is subjected to dimension splitting processing, which is segmented according to time segments and feature channels, and the area with a large change rate is identified as a trend abnormality segment. Specifically, the high sensitivity segment is labeled by using the change rate threshold screening method, and interval mapping is performed with the contribution degree of each dimension feature, so as to obtain the prediction bias data. The feature item with serious correlation weight offset in the reliability evaluation value is corrected by using the prediction bias data, to generate the corrected reliability value and the adjustment factor. The above two values are weighted and fused to map the decision factor. The decision factor is quickly returned to the edge device, combined with the historical feedback data on the edge side, and subjected to local rapid calibration to output the response delay parameter and the consistency threshold. On this basis, the calibration rule table is constructed, secondary analysis is performed, and the feedback identifier is generated.

[0033] For example, it is found that the temperature-internal resistance feature pair has a significant impact on the prediction bias, and by high-frequency abnormality detection, it is found that the misjudgment frequency of a certain type of device is higher than 10%. The confidence interval is remapped to improve the decision weight.

[0034] Step S400, based on the feedback identifier and the edge side processing result, the incremental adjustment is performed to obtain the adjusted edge side processing result.

[0035] In this step, the three problems of response delay, data drift and confidence interval change existing in the current edge side device are analyzed by feedback identification; specifically, the delay index curve is extracted according to the historical response time sequence, the drift factor group is generated by combining the offset trend extraction algorithm, and the compensation coefficient is fitted and generated by using the confidence boundary sliding window method; the drift factor group and the compensation coefficient are applied to the edge side processing result, residual fitting operation is performed, and the compensation residual sequence and the correction frequency diagram are generated; the compensation residual sequence and the edge side historical state characteristics are cross compared to identify high fluctuation dimensions and high consistency dimensions, and the residual mapping curve and the residual prediction weight are constructed; the enhanced feature set is formed by feature fusion, and quantization coding and weight reallocation are performed to generate the enhanced coding stream; finally, the enhanced coding stream and the original edge side processing result are fused and incremental adjustment is performed to output the adjusted edge side processing result.

[0036] For example, for the feedback delay of a certain battery pack in a high temperature environment, the voltage change prediction accuracy is improved by fitting the residual trend to realize dynamic adaptive compensation.

[0037] Step S500, analyze and verify the feature analysis result and the adjusted edge side processing result to obtain a comprehensive analysis result.

[0038] In this step, the trend fusion feature is extracted from the feature analysis result, the enhanced coding stream is extracted from the adjusted edge side processing result, the dimension correspondence and semantic mapping are performed to form a difference parameter; specifically, the difference parameter is divided into several time segments by inter-frame consistency verification, fluctuation reorganization and entropy analysis are performed on each segment, and a dynamic consistency profile is constructed; based on the profile, adaptive segmentation analysis is performed to extract key consistent regions and conflict segments, and respectively perform marker tracking and rollback inspection; finally, the consistency marker value and the calibration threshold are applied to the fusion result to output the comprehensive analysis result.

[0039] For example, it is detected that the edge side frequency encoding and the cloud trend fitting have more than 93% consistent segments in the initial stage of capacity attenuation, thereby improving the accuracy of overall life cycle determination.

[0040] In this embodiment, by acquiring the battery operation data collected by the edge side, the battery operation data is preprocessed to obtain the edge side processing result, and the local feature summary and context association information are generated based on the edge side processing result; then the local feature summary and context association information are uploaded to the cloud, the cloud extracts multi-dimensional trend features and cross-validation indicators, and the multi-dimensional trend features and cross-validation indicators are denoised and paired to obtain feature analysis results and credibility evaluation values; then the feature analysis results are structurally deconstructed to obtain prediction bias data, and the credibility evaluation values are error corrected based on the prediction bias data to generate a decision factor, and the decision factor is quickly returned to the edge side and calibrated locally to generate a feedback identifier; further, based on the feedback identifier and the edge side processing result, the adjusted edge side processing result is obtained; finally, the feature analysis results and the adjusted edge side processing result are analyzed and verified, and the comprehensive analysis result is output.

[0041] Through the battery data analysis method based on cloud edge collaboration of the application, the multi-dimensional trend extraction and verification can be realized by the cloud computing capability while retaining the real-time advantage of edge side processing, forming a high-credibility analysis result; at the same time, through the structural deconstruction and error correction mechanism of the feature analysis result, the generated decision factor has stronger adaptability and interpretability, and then feedback is realized through fast return and local calibration; in addition, through the drift factor group and compensation coefficient generated based on the feedback identifier, the edge side processing result can be dynamically adjusted to improve the accuracy and consistency of local analysis; through the fusion analysis and verification of the feature analysis result and the adjusted edge side processing result, not only the robustness and dynamic consistency of the result are enhanced, but also the accuracy, reliability and intelligent level of the entire battery operation state analysis system are significantly improved. When the battery operation state frequently changes or the edge side environment fluctuates greatly, the existing scheme is difficult to capture the short-period dynamics of the edge side in time, resulting in low accuracy and poor reliability of the analysis result.

[0042] Embodiment 2: In step S100, the battery operation data collected by the edge side is acquired, and the battery operation data is preprocessed to obtain the edge side processing result, which specifically includes: The battery operation data is collected by the edge device in real time, the battery operation data is time-aligned and redundantly compressed to obtain a synchronous data sequence and a compression mapping relationship, and the synchronous data sequence is subjected to fluctuation detection and extreme value extraction to obtain an abnormal point index.

[0043] The edge device deployed near the battery management system (BMS) is configured to collect battery operation data in real time; specifically, the collected data includes time series parameters such as voltage, current, temperature, charge and discharge state, internal resistance, etc. of the battery, and the sampling frequency is set to 1Hz to 10Hz according to the application scenario. The edge device has a time synchronization module embedded, which aligns the full amount of timestamps of the collected data and constructs a unified synchronous data sequence. In order to reduce bandwidth and computing resource overhead, the synchronous data sequence is further compressed using a sliding window compression algorithm, and the mapping relationship between the original and compressed data is recorded to form a compression mapping relationship. Subsequently, a fluctuation detection algorithm based on weighted moving average (WMA) is used to calculate the fluctuation amplitude of each window, and combined with the local extreme value detection method to extract mutation points or abnormal peak points, to obtain an abnormal point index list containing index positions and corresponding abnormal amplitudes.

[0044] For example, after collecting the current data of a certain battery pack running for 30 minutes, first align the timestamps of the current sequence with 1 second as the unit; then use a sliding window of every 10 seconds as a group for redundancy compression, and perform weighted moving average operation on the current difference value in each group of windows. When the fluctuation amplitude of a group exceeds the set threshold value 5A or the current extreme value exceeds the rated value, it is recorded as an abnormal point and stored in the abnormal point index list.

[0045] The compressed mapping relationship and abnormal point index are edge cleaned to obtain cleaned data, and the cleaned data is subjected to short period change analysis and associated window construction to obtain edge side processing results and associated graph.

[0046] By constructing an edge cleaning mechanism based on a rule base and threshold strategy, data drift, repeated records or missing items in the compressed mapping relationship and abnormal point index are repaired and removed; specifically, if there are data points with the same value or significantly deviating from the historical mean in consecutive multiple windows, the system will mark them as redundant or error data and delete or repair them to obtain the cleaned data sequence. Subsequently, short period change analysis is performed on the cleaned data, i.e. in a sliding interval of 10 seconds to 1 minute, the variation degree and relative change rate of voltage, current and temperature characteristics are calculated, and the time period sensitive to change rate is extracted. Based on the above change characteristics, an associated window containing time windows and coupling relationships between variables is further constructed, and a covariant network between features is established to form an associated graph with a topological structure as the basis for subsequent context information generation.

[0047] For example, in a certain operation data sequence, it is found that the voltage signal is maintained at 3.6V for 10 seconds continuously, but the current suddenly jumps from 2A to 6A. The system first cleans up the abnormal unchanged segment, and then constructs an analysis window within 30 seconds before and after the current jump point, compares the voltage and temperature changes, and draws the correlation map therebetween to generate the edge side processing result containing the co-variation relationship among the current-voltage-temperature and the corresponding correlation graph.

[0048] Based on the edge side processing result, a stable factor group and a change frequency index are extracted, and a local feature summary is constructed based on the stable factor group and the change frequency index. The local feature summary and the correlation graph are fused to generate context-related information with a correlation structure.

[0049] By evaluating the fluctuation degree of each type of battery feature in the edge side processing result within a set time window, the feature items that show low volatility and stable trend in multiple windows are extracted to form a stable factor group. Specifically, a double judgment mechanism based on standard deviation and range is used to calculate the variation index of temperature, voltage, internal resistance and other features within each 10-minute sliding window, and a threshold is set to select features that meet the stability requirements. In parallel, the number of times each type of feature appears in a unit of time is counted to construct a change frequency index set as an important dimension for subsequent trend analysis. Then, the stable factor group and the change frequency index are used to construct a local feature summary with time labels and feature labels. The feature summary is fused with the established correlation graph to generate context-related information with a correlation structure and context-dependent relationship, which is used for feature reconstruction and trend analysis in the cloud.

[0050] For example, after continuously monitoring the data of a battery module for 12 hours, it is found that the temperature variation range is always maintained at ±1.2°C and the voltage variation is not more than ±0.03V, so it can be determined that temperature and voltage are stable factors. At the same time, the current appears more than 15 times per hour exceeding the rated value, which is recorded as a high-frequency change index. The system labels and encodes the above features as a local feature summary, and combines the co-occurrence topology of "voltage-current-temperature" to form complete context-related information for cloud modeling.

[0051] In step S200, the edge side processing result is uploaded to the cloud for feature reconstruction to obtain the feature analysis result and the credibility evaluation value. The step specifically includes: After uploading the edge side processing result to the cloud, feature reconstruction is performed to generate a time series matrix and a space coupling matrix.

[0052] The local feature summary generated on the edge side is uploaded to the cloud feature modeling module in association with the context-related information. Specifically, the cloud first receives the local feature summary marked with a time label and a feature label, and performs field analysis and standardization processing on the stable factors and change frequency indicators contained therein. Then, the stable factors are reordered according to the time dimension to construct a time series matrix. At the same time, in combination with the association graph structure in the context-related information, the coupling degree and influence path between voltage, current, temperature, internal resistance and other features are identified, and a spatial coupling matrix is constructed based on the co-occurrence frequency to describe the spatial coordination relationship between different features.

[0053] For example, the uploaded local feature summary records the temperature change frequency and voltage stability factor of a certain battery pack in different time periods. After arranging them in chronological order, the cloud system constructs a two-dimensional time series matrix of temperature-voltage. According to the topology structure in the context-related information, the high-frequency co-occurrence relationship of "voltage-temperature-internal resistance" is extracted, and the corresponding spatial coupling matrix is constructed to represent the spatial coordination characteristics of the structure.

[0054] Based on the time series matrix, short-term trends, medium-term trends and long-term patterns are extracted, and multi-dimensional trend features are constructed according to the short-term trends, medium-term trends and long-term patterns.

[0055] Through multi-scale sliding window extraction operation on the time series matrix, short-term (such as 1 hour), medium-term (such as 6 hours) and long-term (such as 24 hours) analysis windows are generated. Specifically, within each time scale, the system uses weighted moving average (WMA) and difference fluctuation analysis method to extract the change trend, slope change and periodic pattern of the corresponding scale. On this basis, the trend features under different time scales are spliced and coded to construct a multi-dimensional trend feature set with hierarchical structure.

[0056] For example, when analyzing the voltage data of a certain battery pack in the past 24 hours, the system extracts the short-term fluctuation trend in the 1-hour sliding window, detects the stable downward trend in the 6-hour sliding window, and identifies the diurnal periodic change pattern in the 24-hour window. These information are integrated into the multi-dimensional trend feature of "voltage short-term fluctuation strong, medium-term downward, long-term periodic fluctuation", which is used for subsequent cross-validation and credibility evaluation.

[0057] The spatial coupling matrix is subjected to consistency test and trend coordination scoring to obtain cross-validation indicators. The multi-dimensional trend features and cross-validation indicators are jointly denoised and dimensionally paired to generate feature analysis results and credibility evaluation values.

[0058] The consistency degree is judged by using co-integration analysis and mutual information evaluation by testing the synergistic variation relationship between different characteristic pairs in the spatial coupling matrix. Specifically, the system calculates the correlation coefficient and coupling delay of each pair of characteristics, selects high consistency pairs, and assigns trend synergy scores. These scores are used as cross-validation indicators to screen and weight multi-dimensional trend characteristics. Then, principal component analysis (PCA) and multidimensional scaling analysis (MDS) algorithms are used to reduce noise and dimension pairing of the joint feature set, remove redundant information, and enhance the representation ability of key trend features, and finally generate feature analysis results and reliability evaluation values.

[0059] For example, when analyzing the synergistic relationship of current-temperature-internal resistance of a certain battery, it is found that the current and temperature are strongly positively correlated when the load is increased, but the internal resistance responds laggingly. The system identifies that "current-temperature" is a high-consistency coupling feature group and assigns a high synergy score. Combined with the abnormal fluctuation value of the temperature trend feature, the system obtains a reliability evaluation value of 0.38 in a low confidence state after noise reduction compression, and forms the final feature analysis result.

[0060] The step of jointly reducing noise and dimension pairing of the multi-dimensional trend features and the cross-validation indicators to generate the feature analysis result and the reliability evaluation value includes: Jointly reducing noise and dimension pairing of the multi-dimensional trend features and the cross-validation indicators to generate a trend relationship graph and a confidence weight table.

[0061] The multi-dimensional trend features and the cross-validation indicators are fused, and a joint score matrix is established based on the feature synergy score. Specifically, the local weighted regression (LOWESS) method is used to fit and reduce noise of the joint features, eliminating high-frequency disturbance terms. Then, the remaining high-confidence dimensions and coupling-degree-high feature dimensions are paired to form a trend association group. Based on the above processing results, a trend relationship graph is constructed to visualize the trend path relationship between features, and the stability and change sensitivity of each feature in trend transmission are extracted to generate a confidence weight table.

[0062] For example, in a certain battery data set, after noise reduction processing, it is found that temperature, internal resistance and current form a continuous trend path, and temperature change has a significant impact on internal resistance prediction. The system draws a trend relationship graph of "temperature-internal resistance-current", and assigns a confidence weight of 0.82 to temperature, indicating that it has high prediction value in this path.

[0063] Based on the trend relationship graph, a multi-source fusion path is constructed to obtain trend fusion features, and the reliability and sensitivity of the trend fusion features are evaluated using the confidence weight table to output an initial value of reliability.

[0064] By analyzing the path connection structure in the trend relationship diagram, a fusion path network is constructed. The system identifies several fusion paths in the diagram according to the transmission directionality and coupling strength of the features. Specifically, the features on each path are superimposed and modeled to form a trend fusion feature set. Then, the fusion features are weighted and scored one by one through a confidence weight table to extract reliability indicators and volatility response features, which are used to determine the confidence interval range of the trend change, and then output the initial confidence value as a further correction reference.

[0065] For example, in the temperature-resistance-current path, the system generates a fusion feature "temperature-resistance-current dynamic response group" based on the confidence weight of each node, and finds that its historical fluctuation range is ±2% without obvious deviation, determining that the initial confidence value is 0.91, representing that the prediction under this path has high stability.

[0066] The trend fusion features are compared and analyzed with the initial confidence value to generate feature analysis results and confidence evaluation values.

[0067] By comparing the current state of the trend fusion features with the historical evolution law, and the consistency analysis of the corresponding initial confidence value and confidence weight; Specifically, the system introduces a sliding evaluation window to calculate the trend change rate and historical deviation degree, and combines the current initial confidence value to correct the error, finally generates the corrected feature analysis results, and synchronously outputs the final confidence evaluation value, for subsequent error compensation and feedback mechanism calling.

[0068] For example, in the scenario of continuously monitoring the abnormal temperature rise of a certain battery pack while the internal resistance is stable, the system finds that the current trend fusion feature deviates from the historical template by 0.25, with an initial confidence value of 0.91, which is reduced to 0.68 after comparison. The final feature analysis result indicates that the temperature trend is no longer the dominant path prediction, and an additional drift factor should be introduced for adjustment.

[0069] In step S300, the feature analysis result is structurally deconstructed to obtain prediction bias data, which is used to correct the error of the confidence evaluation value to generate a decision factor. The decision factor is quickly returned on the edge side and calibrated locally to obtain a feedback identifier. The steps include: The feature analysis result is dimensionally split and time-series segmented, and the analysis area is divided according to the feature change rate and time window boundary to obtain the feature contribution degree and trend abnormality segment. The trend abnormality segment is used to map the feature contribution degree to obtain the prediction bias data.

[0070] By splitting the feature analysis results by feature dimension, an independent time series data stream is constructed for each dimension, and combined with the trend change rate and the edge side set sliding time window for segmentation processing; Specifically, the fluctuation gradient of each feature sequence is calculated per unit time according to its change rate, and if the fluctuation gradient exceeds a certain threshold (such as the current change rate exceeding 1A / s), it is set as a trend abnormal point; The system divides multiple analysis regions by taking these points as boundaries. Subsequently, by integrating and counting the slope, fluctuation intensity and other parameters of each feature before and after the change in each analysis region, the degree of influence of the region on the overall trend is evaluated, and the feature contribution degree is generated; According to the position and frequency of the trend abnormal segment, the feature contribution degree is interval mapped to evaluate the influence of the feature on the overall trend prediction deviation in different time periods, thereby forming the prediction deviation data.

[0071] For example, when analyzing the voltage-internal resistance-temperature three-dimensional feature, the system segments the temperature sequence with a 30-minute window, and after finding that the temperature slope rises from 0.2 to 0.7 and then falls sharply between the 3rd and 4th windows, it determines that it is a trend abnormal segment; Combined with the stable performance of voltage and internal resistance before and after the change, the high feature contribution degree of temperature in this segment is evaluated, and it is mapped as prediction deviation data for the next step of credibility correction.

[0072] By correcting the high-sensitive features in the prediction deviation data, the corrected credibility value and the adjustment factor are obtained, and the corrected credibility value and the adjustment factor are weight mapped to generate the decision factor.

[0073] By cross-comparing the prediction deviation data with the original credibility evaluation value, the high-sensitive features (i.e. features with high contribution degree and frequent trend abnormalities) are error-corrected; Specifically, the system uses a feature-by-feature correction mechanism to compare the consistency between the trend direction reflected by the deviation data and the prediction direction in the original credibility value. If the consistency is insufficient (for example, the trend is judged to be rising but the credibility value shows a decrease), the correction is adjusted downward according to the deviation degree. At the same time, the system calculates the correction amplitude of each item and generates the corresponding adjustment factor in combination with the upper and lower boundaries of the original credibility. The corrected credibility value and the adjustment factor are normalized mapped according to the historical weight of each feature dimension to generate a decision factor with context consistency, trend sensitivity and decision accuracy.

[0074] For example, if the credibility evaluation value of the temperature feature in the previous stage is 0.76, and the prediction deviation shows that it has continuous deviation in multiple windows and the trend is wrong, the system corrects the credibility to 0.61 and records the correction amplitude as -0.15, and the corresponding adjustment factor generated is "trend reversal + high fluctuation", which is combined and mapped with the weight factor 0.8 of the historical temperature feature to finally form a decision factor with "temperature-adjustment weight-deviation label" as the core field.

[0075] The decision factor is returned to the edge, and the returned decision factor is quickly and locally calibrated based on historical feedback data to obtain a response delay parameter and a consistency threshold; wherein the historical feedback data refers to a time series set of calibration log records, decision response results and parameter mismatch conditions formed in the same type of equipment or the same battery life cycle on the edge side, used to represent the deviation trend between the historical decision behavior and the true state.

[0076] By quickly returning the decision factor from the cloud to the corresponding edge device node after compression encoding, the system retrieves historical feedback data on the edge side for matching comparison and deviation correction; specifically, the system aligns the returned decision factor in time with the stored decision response records and device calibration logs on the edge side, and analyzes the semantics to calculate the deviation trend, response time lag and decision reliability index. On this basis, the response delay (unit: s) and the judgment consistency threshold (confidence difference) of the current feature and the historical feature in the decision process are extracted to guide the next calibration strategy.

[0077] For example, if there is a 3-second response delay of a certain battery in the edge-side historical record under the condition of a sudden temperature rise, and the corresponding confidence and actual trend consistency difference is 0.18, the current system will set the temperature-related decision factor delay response time to 3 seconds according to the historical calibration template, and apply a consistency threshold of 0.18 for the next rule construction.

[0078] Based on the response delay parameter and the consistency threshold, a calibration rule table is constructed, and the decision factor is analyzed again using the calibration rule table to generate a feedback identifier.

[0079] By integrating the response delay parameter and the consistency threshold, a dynamic calibration rule table is constructed, which records the decision tolerance, response time lag correction value and correction strategy of different features under different states; specifically, the system constructs a mapping relationship according to the classification label, deviation mode and threshold range of each feature, and sets the correction condition and reconciliation method; then, the returned decision factor is analyzed again according to the rule table, to evaluate whether its current state meets the calibration trigger condition, and to select the calibration weight, feedback label type and feature category for encoding according to the rule, and finally generate a feedback identifier for subsequent residual compensation and model correction.

[0080] For example, the calibration rule table of the temperature feature is set as follows: if the response delay is greater than 2 seconds and the consistency threshold exceeds 0.15, the feedback identifier type is "high delay - medium mismatch", and an adjustment label with priority 2 is assigned. After the system analyzes the temperature decision factor that meets the condition for the second time, it generates a feedback identifier with the code "Temp_HD_Mismatch_P2", indicating that this feature needs to be prioritized for drift fitting and residual compensation in subsequent processing.

[0081] In step S400, based on the feedback identifier, the incremental adjustment is performed on the edge side processing result to obtain an adjusted edge side processing result, including: Based on the feedback identifier, the response delay, data drift and confidence interval change are analyzed, and the drift factor group and the compensation coefficient are generated based on the response delay, data drift and confidence interval change.

[0082] By analyzing and classifying the feedback identifier, the response delay label, data drift direction and amplitude feature of each feature dimension, and the upper and lower limit change information of the confidence interval are extracted; specifically, the system locates back the historical time series according to the feature label in the feedback identifier, synchronously compares the features marked as "high delay", "low confidence", "trend reversal" and the like in the feedback with the original edge side processing result, calculates the actual response delay value (such as second level unit), drift slope (such as value change rate per unit time) and confidence upper and lower limit adjustment amplitude in a specific time period, respectively as response delay factor, trend offset factor and confidence interval correction factor, and combines to form a drift factor group. Based on the weight of each factor and the correlation of feature dimensions, the compensation coefficient corresponding to the multi-dimensional feature is generated through weighted average and normalization for subsequent residual fitting and feature correction.

[0083] For example, if the feedback identifier of the voltage dimension contains "response delay 3 seconds", "trend reversal upward" and "confidence lower limit below threshold" labels, the system will extract the average response delay (3 seconds), drift trend slope (such as +0.15 V / min) and confidence lower limit (such as -0.08) between the original sequence and the feedback time point, respectively corresponding to the response delay factor, the positive drift factor and the confidence correction factor, and combined with the historical weight (for example, the weight ratio is 3:2:1) for weighting, finally forming the compensation coefficient of the voltage dimension 0.126, which is one item in the drift factor group.

[0084] The drift factor group and the compensation coefficient are used to fit the residual of the edge side processing result to obtain a compensated residual sequence and a modified frequency diagram. The compensated residual sequence and the edge side historical state feature are cross-compared to identify high volatility dimensions and high consistency dimensions; wherein the edge side historical state feature refers to the feature change trajectory and stability label data recorded by the edge device within a set time period, which is related to the battery operating state, including but not limited to temperature, voltage change rate, internal resistance fluctuation characteristics and current response delay sequence.

[0085] By applying residual correction based on the drift factor group and the compensation coefficient to the original edge side processing result, the fitting residual of each feature dimension is calculated and its frequency fluctuation in unit time is recorded; Specifically, the system is based on the feature time series, and the compensation coefficient is applied to the original feature value in the time sliding window to obtain the residual between the predicted correction value and the original value, and generate a residual sequence. Then the average fluctuation frequency and change amplitude of the sequence in the sliding window are counted to form a correction frequency diagram. The system cross-comparisons the residual sequence with the historical state features (such as temperature, internal resistance, voltage change rate, current delay, etc.), calculates the correlation coefficient and cooperative fluctuation index of each feature dimension, and if the fluctuation frequency exceeds the set threshold and is strongly correlated with the historical state change, it is marked as a high fluctuation dimension; If the residual sequence is stable in different feedback periods and the change trend is highly consistent with the historical state, it is classified as a high consistency dimension.

[0086] For example, when processing the temperature feature, the residual sequence obtained after applying the compensation coefficient fluctuates by ±0.8°C in every 15-minute window, with a frequency of 4 times / hour, and has a high positive correlation (r=0.91) with the historical high-temperature internal resistance change rate. The system marks the temperature dimension as "high fluctuation" accordingly; While the current response delay dimension has minimal residual fluctuation (<0.02A) and is highly consistent with the trend in past feedback periods, it is marked as a "high consistency" dimension for the next step of enhanced modeling.

[0087] Based on the high fluctuation dimension, a residual mapping curve is constructed, and based on the high consistency dimension, a residual prediction weight is extracted. The enhanced feature set is constructed using the residual mapping curve and the residual prediction weight, and the enhanced feature set is quantized and weighted. The enhanced coding stream is generated.

[0088] By performing trend line fitting on the residual sequence of the high fluctuation dimension, a residual mapping curve reflecting its change path is constructed, and the residual prediction weight is extracted based on the stability of the residual mean and deviation of the high consistency dimension; Specifically, the system uses a local weighted regression (LOWESS) algorithm to smooth fit the high fluctuation residual, extracts its inflection point position, change rate and abnormal window; For the high consistency dimension, the residual mean (such as average residual = 0.02), standard deviation and confidence interval are counted to represent its prediction stability. Then, the two types of information are spliced to generate an enhanced feature set, which is then quantized by an encoding algorithm (such as dictionary mapping-based hash encoding) and the weight is redistributed according to the relative contribution of each type of feature in the error composition. Finally, a structured enhanced coding stream is generated that can be used for model optimization and anomaly identification.

[0089] For example, the residual mapping curve generated by the LOWESS algorithm for the two high fluctuation dimensions of voltage and temperature shows that there are two abnormal inflection points in 2 hours, respectively, and the system records the change rate (such as 0.12 V / h, 0.35°C / h) into the enhanced features; while the current response delay is always less than 0.01 in the residual mean of six cycles, the confidence level is above 95%, and the prediction weight is 0.92. After splicing and encoding, the enhanced encoding stream field is generated: “V:0.12,W1;T:0.35,W2;I:0.01,W3”, which is used for subsequent fusion adjustment.

[0090] The multi-dimensional fusion vector is incrementally adjusted to obtain an adjusted edge side processing result; wherein the original edge side processing result refers to a preliminary processing result set formed by directly cleaning, analyzing and feature extracting based on battery operation data before feedback adjustment, including basic time series features, window statistical indicators and low-dimensional feature abstracts, etc.

[0091] By splicing the enhanced encoding stream and the original edge side processing result, the incremental adjustment model is used to reconstruct and fit the regression of the fusion vector, and finally the adjusted edge side processing result is output; specifically, the system uses an incremental linear regression (Incremental Linear Regression) model to merge the original feature vector and the field after weight redistribution in the enhanced encoding stream into a new set of input features, real-time updates the feature coefficients, and re-estimates the prediction output according to the incremental learning result. This operation can effectively make up for the missing fluctuation trend or delayed response in the original processing result, and dynamically adapt to new feedback information on the premise that the model parameters remain convergent, thereby improving the credibility and stability of the edge side result.

[0092] For example, the voltage prediction value in the original edge side processing result is 3.75V, and the temperature trend change rate is 0.1°C / h. After adjustment by the enhanced encoding stream, the system incrementally regresses to calculate a new voltage prediction value of 3.82V, and the temperature change rate is increased to 0.14°C / h. The final new edge processing result vector is [voltage=3.82V, temperature rate=0.14°C / h, internal resistance=56mΩ, delay=2.1s], which is used for subsequent analysis and verification.

[0093] In step S500, the feature analysis result and the adjusted edge side processing result are analyzed and verified to obtain a comprehensive analysis result, specifically including: The trend fusion features in the feature analysis result are extracted, the enhanced encoding stream in the adjusted edge side processing result is extracted, the trend fusion features and the enhanced encoding stream are dimensionally corresponding and semantically mapped, and the difference parameters are obtained.

[0094] By extracting key dimension information from trend fusion features and enhanced encoding stream respectively, and one-to-one corresponding them by feature category, timestamp and attribute label, a multi-dimensional semantic mapping relationship is constructed. Specifically, the system first identifies the time series indicators (such as voltage trend, temperature trend, internal resistance trend, etc.) in the trend fusion features and their statistical characteristics (such as mean, variance, frequency component), and analyzes the semantic meaning of the encoding field in the enhanced encoding stream (such as residual change rate, drift factor weight, calibration label, etc.), and through feature dimension name, time alignment and attribute similarity matching, the corresponding features of the two data sources are paired one by one. Based on this mapping relationship, the deviation of the corresponding features in value and semantics is calculated, and a difference parameter containing the difference measurement of each feature is generated, and the matrix elements reflect the relative difference and change amplitude of the corresponding features in the current time window.

[0095] For example, the voltage trend fusion feature contains a mean change of 0.12V in the past 1 hour, and the corresponding encoding field in the enhanced encoding stream reflects a voltage residual change rate of 0.10V. After matching by time window and feature name, the difference value 0.02V is calculated and filled in the corresponding position of the difference parameter. Similarly, the difference between the temperature trend and the residual weight field is 0.03°C, and the difference between the internal resistance is 0.005Ω, forming a complete multi-dimensional difference parameter.

[0096] The difference parameter is divided into time sections and fluctuation is reorganized to generate a dynamic consistency profile. The dynamic consistency profile is adaptively segmented and analyzed to extract key consistent regions and conflict segments. The key consistent regions are labeled and tracked to obtain a consistency label value. The conflict segments are subjected to entropy value calculation and rollback test to obtain a calibration threshold.

[0097] By segmenting the difference parameter by time axis, the numerical fluctuation is aggregated and reorganized to construct a dynamic consistency profile reflecting the change trend of multi-dimensional features. Specifically, the system performs weighted averaging and trend superposition on each element in the difference parameter with a set time window (such as 10 minutes) as the unit, generating a profile graph representing the overall difference fluctuation in each time interval. Then, the profile graph data is adaptively segmented using a clustering algorithm (such as DBSCAN density clustering) to divide the key consistent regions with stable changes and the conflict segments with significant differences. For the key consistent regions, the system labels and continuously tracks their time evolution, calculates the similarity and stability of multi-dimensional features in the region, and generates a consistency label value. For the conflict segments, the information complexity and uncertainty are evaluated using entropy value calculation method, and rollback test is performed combined with historical calibration data to determine whether it exceeds the preset tolerance threshold, and finally the threshold parameter for dynamic calibration is determined.

[0098] For example, in the difference parameters of the past 2 hours, the system divides the first hour into a consistent area, marked as "Area A", where the differences in voltage, temperature, internal resistance, etc. are stable and have low fluctuations, with a marked value of 0.85 (close to 1 indicating high consistency); while the second hour is determined to be a conflict segment "Segment B", with an entropy value of 0.65, exceeding the preset threshold of 0.6, triggering a rollback verification, confirming that there are abnormal fluctuations in some characteristics in this time period, and subsequent calibration needs to be strengthened.

[0099] The consistency mark value and the calibration threshold are used to bidirectionally harmonize the trend fusion features and the enhanced encoding stream, and output the comprehensive analysis results.

[0100] By taking the consistency mark value and the calibration threshold as weight parameters, the corresponding feature values in the trend fusion features and the enhanced encoding stream are dynamically adjusted, realizing bidirectional data harmonization; specifically, for areas with high consistency, the system increases the weight of trend fusion features and reduces the influence of abnormal residuals in the encoding stream; for conflict segments, it increases the role of calibration compensation factors in the encoding stream to correct the deviation of fusion features. This process is completed through a weighted average algorithm, ensuring that the final output comprehensive analysis results reflect both edge side adjustment information and the overall perspective of cloud trend features, achieving consistency, accuracy and stability of data.

[0101] For example, when the consistency of "Area A" is high, the system sets the voltage and temperature weights in the trend fusion features to 0.7 and the enhanced encoding stream weights to 0.3, making the analysis results stable and reflecting the true trend; when the conflict in "Segment B" is large, the drift compensation weight in the encoding stream is increased to 0.6 and the trend fusion weight is reduced to 0.4, correcting the deviation and outputting the adjusted voltage and temperature values, finally generating a comprehensive analysis report.

[0102] In this embodiment, the battery operation data is collected in real time by the edge device, and combined with time alignment and redundancy compression technology, high-quality synchronous data sequence and compression mapping relationship are obtained, further fluctuation detection and extreme value extraction are carried out to identify abnormal points, combined with edge cleaning and short period change analysis, accurate processing and correlation window construction of edge side data are realized, local feature abstract with stable factor group and change frequency index is formed, and it is fused with correlation graph to generate context correlation information. In the cloud, the uploaded local feature abstract and context correlation information are reconstructed by features to form time series matrix and space coupling matrix, through multi-level trend extraction and consistency test, joint denoising and dimension pairing, accurate feature analysis result and credibility evaluation value are generated. For the feature analysis result, the structural deconstruction technology is used to segment the dimension and time series, combined with the trend abnormal segment and feature contribution degree mapping, the prediction bias data is obtained, and based on this, the credibility evaluation value is corrected to generate the judgment factor, and the historical feedback data is used to realize the fast edge backhaul and local calibration, obtain the response delay parameter and consistency threshold, construct the calibration rule table, and output the feedback identifier.

[0103] Based on the feedback identifier, the response delay, data drift and confidence interval change are extracted, the drift factor group and compensation coefficient are formed, the residual fitting is performed on the edge side processing result, the residual mapping curve and residual prediction weight are constructed by identifying the high fluctuation and high consistency dimension, the enhanced encoding stream is generated, and the incremental adjustment model optimization fusion vector is output. The adjusted edge side processing result is output. Finally, the feature analysis result and the adjusted edge side processing result are dimensionally corresponding and semantically mapped to construct the difference parameter and perform dynamic consistency profile analysis, realize the bidirectional harmonization of trend fusion features and enhanced encoding stream, and output the comprehensive analysis result. The overall scheme realizes the collaborative optimization of edge and cloud, effectively improves the accuracy, timeliness and robustness of battery data analysis, and significantly enhances the ability of the system to detect abnormalities, drift compensation and result calibration.

[0104] Embodiment 3: As shown in Figure 2 The application also provides a battery data analysis platform 10 based on cloud edge collaboration, including an acquisition module 11, an upload module 12, a correction module 13, a fitting module 14 and a verification module 15.

[0105] The acquisition module 11 is mainly used for acquiring battery operation data collected by the edge side, and pre-processing the battery operation data to obtain an edge side processing result.

[0106] The acquisition module 11 uses the edge device to collect battery operation data in real time, and pre-processes the original data by combining time alignment, anomaly detection and edge cleaning technology to generate high-quality edge side processing results.

[0107] The uploading module 12 is mainly used for uploading the edge side processing result to the cloud for feature reconstruction to obtain feature analysis result and credibility evaluation value.

[0108] The uploading module 12 is responsible for efficiently transmitting the local feature digest and context association information to the cloud, constructing a time series matrix and a spatial coupling matrix, using a multi-level trend extraction algorithm and a consistency verification mechanism to complete the generation of multi-dimensional trend features and cross-validation indicators, and obtaining accurate feature analysis results and credibility evaluation values through joint noise reduction and dimension pairing technology.

[0109] The correction module 13 is mainly used for structural deconstruction of the feature analysis result to obtain prediction bias data, error correction of the credibility evaluation value using the prediction bias data, generation of a decision factor, edge side fast feedback and local calibration of the decision factor, and obtaining a feedback identifier.

[0110] The correction module 13 performs structural deconstruction and time series segmentation processing on the feature analysis result, corrects the credibility evaluation value using the prediction bias data, generates a decision factor, performs fast edge feedback and local calibration using historical feedback data, forms a response delay parameter and a consistency threshold, and then outputs a feedback identifier, thereby improving the real-time and accuracy of the decision result.

[0111] The fitting module 14 is mainly used for incremental adjustment based on the feedback identifier and the edge side processing result to obtain an adjusted edge side processing result.

[0112] The fitting module 14 extracts a drift factor group and a compensation coefficient based on the feedback identifier, dynamically adjusts the edge side processing result using a residual fitting method, constructs a residual mapping curve and a prediction weight by identifying high volatility and high consistency dimensions, generates an enhanced encoding stream, and optimizes the accuracy and robustness of the edge side processing result.

[0113] The verification module 15 is mainly used for analyzing and verifying the feature analysis result and the adjusted edge side processing result to obtain a comprehensive analysis result.

[0114] The verification module 15 performs dimension correspondence and semantic mapping on the feature analysis result and the adjusted edge side processing result, constructs a difference parameter and conducts dynamic consistency profile analysis, realizes bidirectional harmonization of trend fusion features and enhanced encoding stream, and finally outputs a comprehensive analysis result.

[0115] In this embodiment, the platform fully utilizes the powerful computing capability of the cloud and the real-time data processing advantage of the edge side, effectively improves the accuracy, response speed and overall intelligence level of the battery state monitoring, and significantly enhances the safety and reliability of the battery operation.

[0116] It should be noted that, for the convenience and brevity of description, the specific working processes of the platform and each module described above can be referred to the corresponding processes in the foregoing embodiment 1, and will not be described here.

[0117] The structures, proportions, sizes, etc. shown in the drawings of the present specification are merely used to cooperate with the disclosed content, to be understood and read by those skilled in the art, and do not define the limiting conditions for the implementation of the present application, and therefore do not have substantial technical significance. Any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects and purposes that can be achieved by the present application, should still fall within the scope of the disclosed technical content.

[0118] The above-described embodiments are merely used to illustrate the technical solutions of the present application, rather than limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A battery data analysis method based on cloud-edge collaboration, characterized in that: include: Acquire battery operation data collected on the edge side, preprocess the battery operation data, and obtain edge side processing results; Uploading the edge-side processing results to the cloud for feature reconstruction to obtain feature analysis results and credibility evaluation values; Structurally deconstructing the feature analysis results to obtain prediction deviation data, using the prediction deviation data to perform error correction on the credibility assessment value to generate a determination factor, and performing edge-side rapid feedback and local calibration on the determination factor to obtain a feedback identifier; Performing incremental adjustment based on the feedback identifier and the edge-side processing result to obtain an adjusted edge-side processing result; The feature analysis result and the adjusted edge-side processing result are analyzed and verified to obtain a comprehensive analysis result.

2. The battery data analysis method based on cloud-edge collaboration according to claim 1 is characterized in that: The step of obtaining the battery operation data collected on the edge side, preprocessing the battery operation data, and obtaining the edge side processing result includes: Collect battery operation data in real time through edge devices, perform time alignment and redundancy compression on the battery operation data to obtain a mapping relationship between a synchronized data sequence and compression, perform fluctuation detection and extreme value extraction on the synchronized data sequence to obtain an abnormal point index; Edge cleaning is performed on the compression mapping relationship and the abnormal point index to obtain cleaned data, and short-term change analysis and associated window construction are performed on the cleaned data to obtain edge-side processing results.

3. The battery data analysis method based on cloud-edge collaboration according to claim 1 is characterized in that: The step of uploading the edge-side processing results to the cloud for feature reconstruction to obtain feature analysis results and credibility evaluation values ​​includes: The edge-side processing results are uploaded to the cloud and feature reconstruction is performed to generate a time series matrix and a spatial coupling matrix; Extracting short-term trends, medium-term trends, and long-term patterns based on the time series matrix, and constructing multidimensional trend features based on the short-term trends, the medium-term trends, and the long-term patterns; The spatial coupling matrix is ​​subjected to consistency check and trend collaborative scoring to obtain a cross-validation index, and the multidimensional trend feature and the cross-validation index are subjected to joint noise reduction and dimension pairing to generate feature analysis results and credibility evaluation values.

4. The battery data analysis method based on cloud-edge collaboration according to claim 3 is characterized in that: The step of performing joint noise reduction and dimension pairing on the multidimensional trend feature and the cross-validation index to generate feature analysis results and credibility evaluation values ​​includes: Performing joint noise reduction and dimension pairing on the multidimensional trend features and the cross-validation indicators to generate a trend relationship graph and a confidence weight table; Constructing a multi-source fusion path based on the trend relationship graph to obtain a trend fusion feature, evaluating the reliability and sensitivity of the trend fusion feature using the confidence weight table, and outputting an initial credibility value; The trend fusion feature is compared and analyzed with the initial credibility value to generate a feature analysis result and a credibility evaluation value.

5. The battery data analysis method based on cloud-edge collaboration according to claim 1 is characterized in that: The steps of structurally deconstructing the feature analysis results to obtain prediction deviation data, using the prediction deviation data to perform error correction on the credibility evaluation value to generate a determination factor, and performing edge-side rapid feedback and local calibration on the determination factor to obtain a feedback identifier include: Performing dimension splitting and time series segmentation processing on the feature analysis results to obtain feature contribution and trend change segments, and performing interval mapping on the feature contribution using the trend change segments to obtain prediction deviation data; Correcting the highly sensitive features in the credibility evaluation value using the prediction deviation data to obtain a corrected credibility value and an adjustment factor, and performing weight mapping on the corrected credibility value and the adjustment factor to generate a determination factor; The decision factor is edge-transmitted, and the returned decision factor is quickly and locally calibrated based on historical feedback data to obtain a response delay parameter and a consistency threshold; A calibration rule table is constructed based on the response delay parameter and the consistency threshold, and the determination factor is subjected to a secondary analysis using the calibration rule table to generate a feedback identifier.

6. The battery data analysis method based on cloud-edge collaboration according to claim 3 is characterized in that: The step of performing incremental adjustment based on the feedback identifier and the edge-side processing result to obtain an adjusted edge-side processing result includes: Analyzing response delay, data drift, and trustworthy interval change based on the feedback identifier, and generating a drift factor group and a compensation coefficient based on the response delay, the data drift, and the trustworthy interval change; Performing residual fitting on the edge-side processing results using the drift factor group and the compensation coefficient to obtain a compensated residual sequence and a corrected frequency graph, cross-comparing the compensated residual sequence with edge-side historical state characteristics to identify high-fluctuation dimensions and high-consistency dimensions; Constructing a residual mapping curve based on the high-fluctuation dimension, extracting residual prediction weights based on the high-consistency dimension, constructing an enhanced feature set using the residual mapping curve and the residual prediction weights, performing quantization encoding and weight redistribution on the enhanced feature set, and generating an enhanced coding stream; The enhanced coding stream is fused with the original edge side processing result to obtain a multi-dimensional fusion vector, and the multi-dimensional fusion vector is incrementally adjusted to obtain an adjusted edge side processing result.

7. The battery data analysis method based on cloud-edge collaboration according to claim 6 is characterized in that: The step of analyzing and verifying the feature analysis result and the adjusted edge-side processing result to obtain a comprehensive analysis result includes: Extracting trend fusion features from the feature analysis results, extracting enhanced coding streams from the adjusted edge-side processing results, performing dimensional correspondence and semantic mapping on the trend fusion features and the enhanced coding streams to obtain difference parameters; Divide the difference parameters into time segments and reorganize the fluctuations to generate a dynamic consistency profile, perform adaptive segmentation analysis on the dynamic consistency profile, extract key consistent regions and conflicting segments, mark and track the key consistent regions to obtain a consistency marking value, calculate the entropy value and perform a backoff test on the conflicting segments to obtain a calibration threshold; The trend fusion feature and the enhanced coding stream are bidirectionally reconciled using the consistency mark value and the calibration threshold, and a comprehensive analysis result is output.

8. A battery data analysis platform based on cloud-edge collaboration, characterized in that: include: An acquisition module is used to acquire battery operation data collected by the edge side, pre-process the battery operation data, and obtain edge side processing results; An uploading module is used to upload the edge-side processing results to the cloud for feature reconstruction to obtain feature analysis results and credibility evaluation values; a correction module, configured to structurally deconstruct the feature analysis results to obtain prediction deviation data, use the prediction deviation data to perform error correction on the credibility assessment value to generate a determination factor, and perform edge-side rapid feedback and local calibration on the determination factor to obtain a feedback identifier; a fitting module, configured to perform incremental adjustment based on the feedback identifier and the edge-side processing result to obtain an adjusted edge-side processing result; The verification module is used to analyze and verify the feature analysis result and the adjusted edge-side processing result to obtain a comprehensive analysis result.

Images