A multi-modal time sequence big data fusion storage and health monitoring method, system, computer and storage medium for an electrochemical energy storage system

Abstract

This invention relates to the field of energy storage technology and network database technology, and particularly to a method, system, computer, and storage medium for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems. The method includes: edge data acquisition and preprocessing, lightweight edge anomaly data detection, cloud-edge data transmission, cloud data processing, data indexing and fusion, establishing a dual-dimensional hash index based on device identifiers and timestamps, and generating a multimodal dataset. This invention reduces storage and bandwidth costs and improves data processing and query efficiency through edge differential encoding and cloud dual compression. It meets sampling frequency requirements by preprocessing delay edge cleaning and differential compression. The dual-dimensional index significantly improves query efficiency across three layers compared to traditional multi-database queries, balancing real-time anomaly detection with accuracy. The pruning of isolated forest algorithm reduces inference latency, and the cloud-based daily model updates significantly improve performance compared to traditional fixed models.

Description

Technical Field

[0001] This invention relates to the field of energy storage technology and network database technology, and in particular to a method, system, computer and storage medium for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems. Background Technology

[0002] Electrochemical energy storage systems are rapidly gaining penetration in areas such as grid peak shaving and renewable energy consumption. A single energy storage power station typically contains 100-1000 battery clusters, each equipped with 200-1000 sensors, generating time-series data at a frequency of 100ms-1s, with an average daily data volume of 5-50GB. This data is accompanied by multimodal data such as equipment files and maintenance records. These data exhibit significant differences in dimensionality, sampling frequency, time synchronization, and physical meaning, forming a typical multimodal time-series big data scenario.

[0003] Existing electrochemical energy storage systems have the following problems:

[0004] (1) Low efficiency of multi-source data fusion: Static, time-series and semi-structured data are scattered in three types of databases without a unified associated index. Querying time-series data and maintenance records in a certain battery cluster requires three cross-database queries. Data preparation takes more than 10 seconds during fault diagnosis, while battery anomalies require millisecond-level response, which cannot support real-time decision-making. Traditional database architecture is difficult to balance the integrity of static data and the efficiency of time-series data.

[0005] (2) High cost and bandwidth waste of time-series storage: The existing TSDB only relies on cloud-based compression and does not perform preprocessing at the edge. Redundant data accounts for more than 60%, and the annual storage cost of a single power station is high, resulting in high cost and bandwidth waste of time-series storage. At the same time, the full upload of raw data occupies an average of 5GB of bandwidth per day, putting great pressure on the network of energy storage power stations in remote areas.

[0006] (3) Poor real-time performance of anomaly detection and model fixation: The hardware computing power of energy storage power stations is low, such as a single microcontroller with 64MB memory. Existing anomaly detection mostly uses complex algorithms such as cloud LSTM and Kalman filtering, with inference latency exceeding 500ms, which cannot meet the "millisecond-level early warning" of energy storage systems. After the model is trained, it remains unchanged for a long time. As the battery ages, the accuracy of anomaly detection is seriously reduced.

[0007] (4) Control linkage disconnection: Existing electrochemical energy storage systems only output early warning information, requiring manual operation of BMS / EMS to reduce power / shut down. From the discovery of the anomaly to the execution of the control, it takes more than 5 minutes, which may cause the fault handling window to be missed, resulting in control linkage disconnection.

[0008] Therefore, it is necessary to design a multimodal time-series big data fusion storage and health monitoring method, system, computer and storage medium for electrochemical energy storage systems that features low latency in anomaly early warning, high data correlation efficiency, low time-series storage cost and strong real-time anomaly detection. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method, system, computer, and storage medium for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems includes:

[0012] S1: Data Acquisition and Database Deployment. Collect time-series data from sensors related to electrochemical energy storage devices. Deploy the non-time-series static database, the semi-structured database, and the time-series database at the edge and cloud respectively. The time-series data includes at least a timestamp (Time), device identifier (Device_ID), sensor identifier (Sensor_ID), measured value (Raw_Value), and anomaly score (Quality_Flag). The non-time-series static data includes at least a sensor identifier (Sensor_ID) and a device identifier (Device_ID) field. The semi-structured data includes at least a timestamp (Time) and a device identifier (Device_ID) field. Establish a two-dimensional hash index of Device_ID and Timestamp (Time) among the three.

[0013] S2: Edge data preprocessing, cleaning outliers in the time series data at the edge, performing differential encoding on the cleaned time series data, and storing the compressed time series data into the edge time series database;

[0014] S3: Lightweight anomaly detection at the edge: A pruned isolated forest algorithm model is loaded at the edge to calculate anomaly scores on the time-series data and the non-time-series static data and semi-structured data stored at the edge.

[0015] S4: Cloud-edge data transmission, the edge terminal periodically calculates the time-series data to generate an aggregation result, and transmits the abnormal time-series data that generates the abnormal score and the aggregation result data together to the cloud;

[0016] S5: Cloud data processing: The uploaded abnormal time series data and the aggregation results are compressed in the cloud and stored in the cloud time series database;

[0017] S6: Data fusion, generating multimodal datasets in the cloud.

[0018] According to the above technical solution, preferably, the non-time-series static data includes a battery cluster file table and a sensor parameter table, wherein the battery cluster file table includes at least a Device_ID field and the sensor parameter table includes at least a Sensor_ID field; the semi-structured data includes maintenance logs and equalization correction data.

[0019] According to the above technical solution, preferably, in step S3, the pruned isolated forest algorithm model selects a maximum of 3 types of core features, and performs lightweight processing on the data corresponding to the core features.

[0020] According to the above technical solution, preferably, the lightweight processing includes normalizing the measurement value Raw_Value to reduce the numerical scale; converting the data type of the device identifier Device_ID or sensor identifier Sensor_ID, converting the string ID to an integer; and extracting the periodic features of the timestamp Time.

[0021] According to the above technical solution, preferably, in step S3, the pruned isolated forest algorithm model uses a two-dimensional array to store tree nodes; it shares and stores duplicate data; it prunes useless nodes and only retains key nodes that distinguish between normal and abnormal nodes.

[0022] According to the above technical solution, preferably, in step S3, the differential encoding is calculated as the difference between the current period's time series data and the previous period's data sequence. If the absolute value of the current period's difference is |X... i If |>N*X, then the previous difference is used to replace the current difference, where X represents the mean of the differences in the previous M periods, and N and M are both constants.

[0023] According to the above technical solution, preferably, in step S5, Zstandard compression is used, and level 5-8 is selected.

[0024] According to the above technical solution, preferably, in step S4, the aggregation result includes short-term mean, maximum value, and minimum value.

[0025] According to the above technical solution, preferably, step S5 further includes model training and updating, whereby model training is performed on the cloud based on the uploaded time-series data, the non-time-series static data and the semi-structured data stored in the cloud, anomaly scores are calculated, and the trained model parameters are sent to the edge device.

[0026] According to the above technical solution, preferably, in step S5, the model training data is divided into a training set and a validation set according to a certain ratio. When the anomaly detection accuracy of the validation set meets the requirements, the model parameters are sent to the edge.

[0027] According to the above technical solution, preferably, in step S5, the model training uses full data or incremental data.

[0028] According to the above technical solution, preferably, it also includes S7, early warning and control linkage, when the abnormal score is between 0.6 and 0.9, an early warning command is triggered, and the early warning command sends a control command to the control layer through the data protocol.

[0029] According to the above technical solution, preferably, in step S7, the abnormal score comes from the edge or the cloud.

[0030] According to the above technical solution, preferably, in step S7, the aggregation result corresponding to the abnormal score is verified a second time to trigger an early warning command.

[0031] According to the above technical solution, preferably, S1 also includes associated scheduling, caching recently queried hash indexes at the edge, and caching recently active hash indexes in the cloud.

[0032] A multimodal time-series big data fusion storage and health monitoring system for electrochemical energy storage systems, comprising the following steps using the method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems:

[0033] The data acquisition and database deployment module collects time-series data from sensors related to electrochemical energy storage devices. It then analyzes the non-time-series static database, the semi-structured database, and the time-series database deployed at the edge and in the cloud. The time-series data includes at least a timestamp (Time), device ID (Device_ID), sensor ID (Sensor_ID), measured value (Raw_Value), and anomaly score (Quality_Flag). The non-time-series static data includes at least a sensor ID (Sensor_ID) and device ID (Device_ID) field. The semi-structured data includes at least a timestamp (Time) and device ID (Device_ID) field. A two-dimensional hash index based on the device ID (Device_ID) and timestamp (Time) is established among the three databases.

[0034] The edge data preprocessing module cleans outliers in the time series data at the edge, performs differential encoding on the cleaned time series data, and stores the compressed time series data into the edge time series database.

[0035] The lightweight anomaly detection module at the edge loads a pruned isolated forest algorithm model at the edge to calculate anomaly scores on the time-series data and the non-time-series static data and semi-structured data stored at the edge.

[0036] The cloud-edge data transmission module periodically calculates the time-series data at the edge to generate an aggregation result, and transmits the abnormal time-series data that generates the abnormal score along with the aggregation result data to the cloud.

[0037] The cloud data processing module compresses the uploaded abnormal time-series data and the aggregation results in the cloud and stores them in the cloud time-series database.

[0038] The data fusion module generates multimodal datasets in the cloud.

[0039] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems.

[0040] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems.

[0041] The beneficial effects of this invention are as follows:

[0042] (1) Significantly reduced storage and bandwidth costs: edge differential coding and cloud dual compression achieve a compression ratio of 15:1, which significantly reduces the annual storage cost of a single energy storage power station; at the same time, only abnormal data and aggregation results are uploaded, greatly reducing the average daily upload volume from 5GB to 0.8G, which can adapt to the weak network environment in remote areas.

[0043] (2) Significantly improved data processing and query efficiency: The total latency of edge data cleaning and differential compression preprocessing can meet the sampling frequency requirement of 100ms; the dual-dimensional index makes the latency of cross-three-layer data query ≤100ms, which is significantly more efficient than traditional multi-database query (>800ms); it supports writing massive data bars per second at the edge and cloud, adapts to large-scale energy storage power stations with thousands of battery clusters, and greatly improves data throughput;

[0044] (3) Balance between real-time performance and accuracy in anomaly detection: The edge-pruned isolated forest algorithm has an inference latency of ≤80ms, which is 84% ​​higher than the cloud-based LSTM algorithm (>500ms), meeting the requirement of "early warning within 1 second" for battery thermal runaway; the cloud-based model is updated daily, and the anomaly detection accuracy is stable at ≥95%, with a false alarm rate of ≤3%, which is significantly improved compared to the traditional fixed model;

[0045] (4) Control linkage closed loop improves system security: The total response delay from anomaly detection to BMS execution command is reduced to less than 500ms, which is nearly twice as efficient as manual operation and effectively avoids thermal runaway; it supports the operation of multiple industrial protocols and is compatible with mainstream BMS / EMS protocols without the need to modify existing equipment; the edge can be deployed on domestic hardware, reducing the cost of running memory and hardware, and meeting the requirements of independent control. Attached Figure Description

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] Figure 1 This is a schematic diagram of the overall structure of a multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems according to the present invention.

[0048] Figure 2 This is a flowchart of the process of loading isolated pruned forest in a multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems according to the present invention.

[0049] Figure 3 This is a schematic diagram of a multimodal data fusion model for a multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems according to the present invention.

[0050] Figure 4 This invention relates to a timing diagram of anomaly detection and control linkage in a multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. It should be noted that, unless otherwise expressly specified and limited, the terms "generate," "process," and "detect" should be interpreted broadly. For those skilled in the art, the specific meaning of the above terms in this patent can be understood according to the specific circumstances.

[0052] Example 1

[0053] Please see Figures 1-4 The following table shows a structural diagram of a multimodal time-series big data fusion storage and health monitoring method, system, computer, and storage medium for electrochemical energy storage systems, as provided in an embodiment of the present invention. The hardware architecture for implementing the multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems includes:

[0054]

[0055] The storage architecture adopts a three-tier linkage and edge-cloud collaborative design:

[0056] 1. Non-time-series static database: Stores non-time-series static data such as battery cluster files and sensor parameters. The non-time-series static data includes a battery cluster file table and a sensor parameter table. The battery cluster file table includes at least a Device_ID field to ensure data integrity. The battery cluster file table also includes battery model, rated capacity, installation time, and cabinet number. The sensor parameter table includes at least a Sensor_ID field. The sensor parameter table also includes measurement range, calibration cycle, and a foreign key associated with Device_ID.

[0057] 2. Time Series Database: 7-day hot data is stored at the edge, and monthly cold data is stored in the cloud; the table structure includes fields such as "Time (millisecond-level timestamp), Device_ID, Sensor_ID, Raw_Value (measurement value), and Quality_Flag (0=normal / 1=abnormal / 2=missing)"; the monthly time partitioning strategy of the time series database is as follows: partitions are named by natural month (e.g., "202509" corresponds to September 2025), and data within a partition is sorted in three levels: "Device_ID→Sensor_ID→Time". Multi-sensor data under the same Device_ID are physically stored continuously, improving the efficiency of batch queries.

[0058] 3. Semi-structured Database: Stores semi-structured data, including maintenance logs and load balancing data. Maintenance logs include fields such as Time, Device_ID, maintenance content, maintenance personnel, and fault type. Load balancing data includes fields such as Time, Device_ID, single-module calibration coefficient, load balancing duration, and calibration error. A joint index of "Time + Device_ID" is used to achieve cross-layer association. The JSON document of the semi-structured data layer supports dynamically expanded fields: maintenance logs can add fields such as "Spare Part Model" and "Maintenance Time," and load balancing data can add a "Ambient Temperature" field. Expanding fields does not require modifying the table structure, adapting to the data storage needs of various scenarios in energy storage systems.

[0059] 4. Association Scheduling: Deployed at the edge and in the cloud, with a built-in "Device_ID - Time" two-dimensional hash index, supporting joint queries across three layers of data, with an index update cycle of ≤1s and a query latency of ≤100ms.

[0060] A method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems includes:

[0061] S1: The edge device acquires data via the Modbus-RTU protocol, collecting time-series data of relevant sensors in the electrochemical energy storage device. Preferably, the relevant sensors include voltage sensors, current sensors, and temperature sensors, generating one data entry in the format "Device_ID+Sensor_ID+Time+Raw_Value" every 100ms.

[0062] S2: Edge data preprocessing:

[0063] S21: Data Cleaning

[0064] Outliers in the time series data are cleaned at the edge. The data cleaning is based on the 3σ criterion to remove outliers. The mean μ and standard deviation σ of the sensor data for the past hour are calculated respectively. Data exceeding [μ-3σ, μ+3σ] are removed. For data with ≤3 consecutive missing points, linear interpolation is used to complete the data. For example, if t1=3.21V and t3=3.23V, t2=3.22V is completed.

[0065] S22: Data Compression

[0066] The cleaned time-series data is differentially encoded, and the compressed time-series data is stored in the edge-end time-series database. The differential encoding generates one formatted data entry every 100ms. The difference between the currently received formatted data sequence and the formatted data sequence received in the previous 100ms is calculated in real time and stored. If the absolute value of the difference is greater than the average of the differences calculated over the previous X 100ms intervals × 2, it is determined to be an abnormal jump, and the difference calculated in the current interval is replaced by the difference calculated in the previous 100ms interval. Here, the value of X is at least 10, such as in voltage sequences [3.2]. [1, 3.22, 3.23] → [3.21, 0.01, 0.01], the differential compression method only calculates "current value - previous value", and the processing time for a single data item is ≤1ms, which is far lower than the 100ms sampling period and is completely acceptable; since more than 70% of the energy storage data has small fluctuations, at least 60% of the redundant data is filtered out by differential compression. After differential compression, only "initial value + difference" needs to be stored, which not only greatly simplifies the differential compression logic, but also reduces the pressure on the subsequent uploading and compression process, avoids data distortion after compression, and has great benefits from redundancy filtering.

[0067] S3: Lightweight Anomaly Detection at the Edge:

[0068] A pruned isolated forest algorithm model is loaded at the edge to calculate anomaly scores on the time-series data and the non-time-series static and semi-structured data stored at the edge. The specific process of processing and calculating in the pruned isolated forest algorithm model includes:

[0069] Tree structure collection:

[0070] A forest is constructed consisting of fifty independent pruned isolated trees. Each pruned isolated tree has node information and a pruning marker. The node information is stored in a two-dimensional array, representing the leaf nodes and non-leaf nodes of the pruned isolated trees respectively. The leaf nodes store the number of samples with corresponding feature dimensions for the time-series data, non-time-series static data, and semi-structured data. The non-leaf nodes store a split feature index and a split threshold. The split feature index is used to obtain the feature dimension, such as the measurement value corresponding to the sensor ID. The split threshold is used to mark and locate the sample positions of the time-series data, non-time-series static data, and semi-structured data with corresponding feature dimensions. Each node contains a pruning flag to indicate whether the node has been pruned. Pruned nodes are directly treated as leaf nodes and do not split further.

[0071] The Fifty Trees model achieves an anomaly detection accuracy of 95.3%, reduces inference latency by 60%, adapts to low computing power at the edge, and solves the problems of conventional models being unable to run at the edge and having poor real-time performance.

[0072] Model metadata:

[0073] This includes the mean normal sample path length, feature dimension information, and pruning parameter records from the training phase. The mean normal sample path length is used for anomaly score standardization; the feature dimensions match the format data of the sensor, such as the device type mapped by Device_ID and the physical quantity dimension corresponding to Sensor_ID; the pruning parameter records include the threshold for the path length corresponding to the leaf node and the threshold for the number of samples in the leaf node, used for model interpretation.

[0074] Model loading steps:

[0075] File reading involves retrieving the model binary file from local storage and loading it directly into memory via DMA (Direct Memory Access) to avoid CPU usage.

[0076] Structural analysis involves sequentially reading the node data of each pruned isolated tree, storing splitting features and thresholds in the non-leaf nodes, and directly marking pruned nodes as leaf nodes.

[0077] Memory optimization employs arrays instead of linked lists to store tree nodes. For example, a two-dimensional array `nodes` is used for fast access via tree index and node index, eliminating the need for pointer positioning and saving 30% of memory. Duplicate thresholds or split feature indexes are shared and stored, with other trees only referencing this value, further saving 20% ​​of memory. Unnecessary nodes are pruned, retaining only key nodes that distinguish between normal and abnormal nodes, reducing the number of tree nodes by 50%. This memory optimization approach reduces the overall model's memory usage from 8MB to ≤2MB, adapting the complex anomaly detection model to edge microcontrollers with limited memory.

[0078] Initialize the verification process to check the compatibility between the model metadata and the sensor's format data, such as whether the feature dimensions match. If they do not match, trigger the model redeployment process.

[0079] Inference calculation: For each pruned isolated tree, traverse from the root node and calculate the path length of the sample. If the current node is a pruned node or a leaf node, return the current depth, i.e., the path length. If it is a non-leaf node, move to the left subtree (less than the threshold) or the right subtree (greater than or equal to the threshold) based on the comparison between the sample's splitting feature value and the node threshold, incrementing the depth by 1, repeating until a leaf node or pruned node is reached. Forest anomaly score aggregation: Calculate the average path length avg_length of 50 trees and substitute it into the anomaly score formula.

[0080] score = 2^(-avg_length / c(n));

[0081] Where c(n) is the average path length of normal samples when the training sample size is n;

[0082] Abnormal scores are standardized so that they are between 0 and 1. If the score is greater than 0.8, Quality_Flag=1 is set to trigger an abnormal alarm at the edge, ensuring that the response is more than 500ms in advance and that the false alarm rate is kept below 3%.

[0083] S4: Cloud-edge data transmission:

[0084] Every 5 minutes, the edge device calculates the "1-minute average, maximum, and minimum values" of the time-series data to generate an aggregated result. The abnormal time-series data that generated the abnormal score and the aggregated result data are transmitted to the cloud together via the MQTT protocol.

[0085] Uploading only abnormal time-series data and aggregation results reduces the amount of uploaded data by 99.6%, bandwidth usage by 84%, and data volume from 5GB / day to 0.8GB / day. Stable transmission is possible even in the weak network environment of remote energy storage power stations, avoiding the situation where "data cannot be transmitted and the cloud cannot issue timely warnings." The cloud only analyzes aggregation results and abnormal data, reducing the "data preparation time" for anomaly detection from 10s (traditional multi-database query) to ≤100ms, ensuring that control commands (such as load reduction) can be executed within 500ms.

[0086] S5: Cloud Data Processing

[0087] S51: Data Compression

[0088] To address the application scenarios of low computing power, weak network transmission capabilities, and high data accuracy at the energy storage edge, the storage unit processes the received abnormal data and aggregated result data using the Zstandard algorithm and writes them to the corresponding monthly partition of the time-series database. A 32KB sliding window is used for the formatted data uploaded from the edge. A minimum repeating fragment length threshold of 3 bytes is set, and repeating fragments exceeding this threshold are compressed. The 3-byte threshold captures valid repeating fragments (such as "BAT" and "202") while filtering out meaningless short repeats (such as single digits "3" and "5"), avoiding invalid calculations. Compared to the 2-byte threshold, it reduces short matching calculations by approximately 30%, while capturing approximately 20% more valid repeating fragments than the 4-byte threshold, balancing speed and compression ratio.

[0089] Huffman coding is performed on the position of repeating sequences, the length of repeating sequences, and non-repeating data. Short codes represent high-frequency data, and long codes represent low-frequency data. For common characters in sensor data, such as numbers 0-9, letters A and Z, and separators "+" and ":", a basic Huffman code table is pre-generated to avoid rebuilding the code table for each compression, saving about 40% of the encoding time. For low-frequency special characters, such as "Q" and "F" in the outlier marker "Quality_Flag=1", the code table is dynamically fine-tuned to ensure that the compression ratio is not lost, thereby further improving the compression ratio.

[0090] The "difference granularity" (e.g., 0.01V) of the edge differential processing matches the "32KB sliding window" of the cloud-based Zstandard. A 32KB window can cover nearly 1000 differential data entries, precisely capturing segments with "Device_ID repetition" and "difference repetition" (e.g., 100 consecutive + 0.01V), resulting in a 30% improvement in compression ratio compared to a 16KB window. If the cloud-based Zstandard window size (e.g., 8KB) is too small, it cannot cover the repeated differences after differential processing, causing a sharp drop in compression ratio; if the window is too large (256KB), processing time increases, making it unsuitable for the edge's 5-minute upload frequency.

[0091] Employing Zstandard 6-level compression with a compression ratio of 15:1, for example, the original 3.4MB becomes 0.23MB after Zstandard 6-level compression. Single data processing time is ≤2ms. A 3-node cloud cluster can handle 100,000 writes per second, while the amount of data uploaded to the edge decreases from 5GB / day to 0.8GB / day, bandwidth usage decreases by 84%, and storage costs decrease by 58%. Differential compression with anomaly jump protection is used at the edge, and cloud-edge compression parameters work together to achieve a perfect balance between storage, bandwidth, and latency.

[0092] S52: Model Training and Update

[0093] Deploy the TensorFlow framework to automatically train the model daily. On the cloud, based on the uploaded time-series data, the cloud-stored non-time-series static data, and the semi-structured data, train an isolated forest model using the aggregated results of the past 30 days and the full set of outlier data (training time ≤ 30 minutes). The model training employs cross-validation, dividing the 30 days of data into training and validation sets in a 7:3 ratio. When the cross-validation accuracy is detected to be ≥ 95%, approximately 5MB of model decision parameters, such as tree splitting threshold and feature weights, are distributed to all edge devices.

[0094] The model was trained using aggregated data from the past 30 days and the full set of outlier data. The raw data contained minute sensor fluctuations (such as voltage jumps of 0.01V), which, if used directly for training, would cause the model to "learn incorrectly" (mistaking normal fluctuations for anomalies). The aggregated data filtered out noise and preserved the battery's true operating trend (e.g., as capacity decays, the average voltage slowly decreases from 3.2V to 3.0V), resulting in more accurate model training. The aggregated data also captured the gradual changes in battery aging (e.g., after 1000 cycles, the maximum current decreases from 200A to 180A), whereas conventional models trained with "raw data over a fixed time period" cannot adapt to aging, and their accuracy drops from 90% to below 70%.

[0095] In this embodiment, after the model is updated daily, the anomaly detection accuracy remains stable at over 95%. Even after 1000 battery cycles, the accuracy still reaches 94.8%, completely solving the problem of model solidification.

[0096] S6: Data Indexing and Fusion

[0097] A two-dimensional hash index of Device ID (Device_ID) and Time is established between the non-time-series static database, the semi-structured database, and the time-series database at both the edge and the cloud. The two-dimensional hash index is called in the cloud to associate "time-series data (InfluxDB) + battery cluster archive (non-time-series static data MySQL) + maintenance log (semi-structured data MongoDB)" to generate a multimodal fusion dataset, such as "BAT-001 battery 2025-09-15 10:00 voltage 3.21V + 2025-09-14 maintenance log".

[0098] S7: Early Warning and Control Linkage:

[0099] The edge device performs initial anomaly detection based on real-time collected single data points using a pruned isolated forest model. The aggregated results are then used to verify the anomaly score detected initially. If the edge device detects a single voltage jump with an anomaly score > 0.8, but the aggregated voltage value remains within the normal range, it is considered temporary interference, and no warning is triggered. If the anomaly score > 0.8 and the aggregated result exceeds the range, an anomaly is confirmed, triggering a warning command. An alarm message, including Device_ID, anomaly type, and time, is sent to maintenance personnel via SMS gateway. The warning command sends a "reduce power to 50%" or "shut down" command to the control layer BMS / EMS via the OPC UA protocol. The BMS / EMS executes the command and returns a confirmation signal. This mechanism, which combines single-detection with aggregated verification, completes anomaly detection and verification in milliseconds, reducing the false alarm rate from the usual 5% to 0.2%. At the same time, it does not require uploading to the cloud and directly triggers local alerts. The entire process from data acquisition to control command execution has a latency of ≤500ms, balancing real-time performance and reliability, and meeting the real-time requirements of core high-priority anomalies such as battery overcharging and thermal runaway precursors.

[0100] This embodiment achieves a total latency of ≤500ms from anomaly detection to BMS instruction execution, a 98% improvement over manual operation, effectively preventing battery overcharging and thermal runaway; it supports industrial protocols such as OPC UA / Modbus, is compatible with mainstream BMS / EMS, and requires no modification to existing equipment; the edge can be deployed on domestically produced hardware such as RK3568, with ≤256MB of RAM, reducing hardware costs by 40% and meeting the requirements for independent control.

[0101] Example 2

[0102] Based on Example 1, an improved method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems is proposed.

[0103] In step S3, the pruned isolated forest algorithm model selects a maximum of three core features, choosing different core features based on the anomaly detection scenario. For example, in energy storage scenarios, voltage, current, and charge / discharge cycle features directly correspond to the core characteristics of battery faults. For instance, a sudden voltage drop corresponds to a short circuit, current fluctuations correspond to battery aging, and abnormal cycle features correspond to abnormal charge / discharge. These three types of features have an AUC (Area Under the Curve) ≥ 0.92 for anomaly detection, covering over 95% of battery fault types. If a slight anomaly is observed in the core features, secondary verification is performed using the aggregated results to avoid missed detections due to a limited number of core features.

[0104] The reasoning speed of 3 types of core features is about 2.5 times faster than that of 8 types of core features. At the same time, it avoids interference from irrelevant features and controls the false alarm rate of anomaly detection to 0.2%, solving the problems of slow calculation and high false alarm rate caused by conventional full features.

[0105] Lightweight preprocessing of the data corresponding to the core features includes:

[0106] Convert the input "Device_ID+Sensor_ID+Time+Raw_Value" data into the feature vectors required by the model:

[0107] Device type encoding: Convert the string ID into an integer from 0 to N, such as temperature and humidity sensor = 0, voltage sensor = 1, avoiding the time-consuming string comparison, and making it 2 times faster;

[0108] Time features: Extract periodic features such as hours and minutes from the timestamp, such as the voltage fluctuation pattern of charging during the day and discharging at night. For example, if the time is 10 o'clock, use 10%24=10 to represent the 10th hour of the day. This helps the model quickly distinguish between "normal periodic fluctuations" and "abnormal jumps" without recalculating, reducing invalid calculations and false alarms.

[0109] Raw value normalization: The mean μ and standard deviation σ calculated by the cleaning unit are normalized to Raw_Value ((Raw_Value - μ) / σ) to keep it consistent with the feature distribution during model training, avoiding calculation deviations caused by large differences in the range of current and voltage values, while reducing the numerical scale and making the calculation speed 1.5 times faster.

[0110] Example 3

[0111] Based on Example 1, an improved method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems is proposed.

[0112] In S4, the ZStandrad compression level selection adapts to the "storage-bandwidth-latency" triangle requirements of energy storage systems. If the edge computing power is weaker, such as 32MB of memory, ZStandrad level 5 is selected, with a compression ratio of 13:1 and a single-transaction processing latency of ≤1ms, suitable for extremely low computing power scenarios. If the edge computing power is slightly higher, such as 128MB of memory, ZStandrad levels 7-8 are selected, with a compression ratio of 17:1 and a single-transaction processing latency of 3-4ms, suitable for scenarios that are more sensitive to storage costs and have a slightly higher tolerance for latency.

[0113] Example 4

[0114] Based on Example 1, an improved method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems is proposed.

[0115] The model is trained only on incremental data. For example, only some decision trees are retrained on newly added data from the last 24 hours, reducing the training time from 30 minutes to 5 minutes and reducing cloud resource consumption.

[0116] Example 5

[0117] Based on Example 1, an improved method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems is proposed.

[0118] In S1, the edge caches the Device_ID index that has been queried in the last 24 hours, and the cloud caches the Device_ID index that has been active in the last 3 months. The cache hit rate is ≥90%. When querying, the data location is directly located, avoiding cross-database scanning and further reducing query latency.

[0119] Example 6

[0120] Based on Example 1, an improved method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems is proposed.

[0121] Set up a cloud-based visualization platform to display the integrated data in real time.

[0122] The cloud analyzes the fused data through multiple devices to identify global anomalies that cannot be judged by the edge alone, such as aging correlations across multiple clusters of devices, and then sends warning instructions to the edge / control layer to achieve more comprehensive risk management.

[0123] Example 7

[0124] Based on Example 1, an improved method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems is proposed.

[0125] For scenarios where the fluctuations in detection data are relatively weak, such as slight battery aging and a slow voltage drop, the anomaly score threshold is lowered to 0.6-0.7 to avoid missed detections. For scenarios with extremely high detection risks, such as extreme faults like short circuits, the anomaly score threshold is raised to 0.9 to reduce false alarms and ensure accurate early warnings.

[0126] Example 8

[0127] Based on the same concept as the above embodiments, embodiments of the present invention also provide a multimodal time-series big data fusion storage and health monitoring system for electrochemical energy storage systems. The steps of the multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems include:

[0128] The data acquisition and database deployment module collects time-series data from sensors related to electrochemical energy storage devices. The edge data acquisition module also supports fault self-recovery: if the acquisition is interrupted for ≥5 seconds, the Modbus communication port is automatically restarted; if the restart fails, a hardware reset is triggered, with a recovery success rate of ≥98%.

[0129] The non-time-series static database, the semi-structured database, and the time-series database are deployed at the edge and in the cloud, respectively. The time-series data includes at least the timestamp (Time), device ID (Device_ID), sensor ID (Sensor_ID), measurement value (Raw_Value), and anomaly score (Quality_Flag) fields. The non-time-series static data includes at least the sensor ID (Sensor_ID) and device ID (Device_ID) fields. The semi-structured data includes at least the timestamp (Time) and device ID (Device_ID) fields. A two-dimensional hash index of device ID (Device_ID) and timestamp (Time) is established among the three databases.

[0130] The edge data preprocessing module cleans outliers in the time series data at the edge, performs differential encoding on the cleaned time series data, and stores the compressed time series data into the edge time series database.

[0131] The lightweight anomaly detection module at the edge loads a pruned isolated forest algorithm model at the edge to calculate anomaly scores on the time-series data and the non-time-series static data and semi-structured data stored at the edge.

[0132] The cloud-edge data transmission module periodically calculates the time-series data at the edge to generate an aggregation result, and transmits the abnormal time-series data that generates the abnormal score along with the aggregation result data to the cloud.

[0133] The cloud data processing module compresses the uploaded abnormal time-series data and the aggregation results in the cloud and stores them in the cloud time-series database.

[0134] The data fusion module generates multimodal datasets in the cloud.

[0135] Example 9

[0136] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the steps of the method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems.

[0137] Example 10

[0138] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the multimodal time-series big data fusion storage and health monitoring method for electrochemical energy storage systems.

[0139] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems, characterized in that, include: S1: Data acquisition and database deployment, collecting time-series data from sensors related to electrochemical energy storage devices; The non-time-series static database, the semi-structured database, and the time-series database are deployed at the edge and in the cloud, respectively. The time-series data includes at least the timestamp (Time), device ID (Device_ID), sensor ID (Sensor_ID), measurement value (Raw_Value), and anomaly score (Quality_Flag) fields. The non-time-series static data includes at least the sensor ID (Sensor_ID) and device ID (Device_ID) fields. The semi-structured data includes at least the timestamp (Time) and device ID (Device_ID) fields. A two-dimensional hash index of device ID (Device_ID) and timestamp (Time) is established among the three databases. S2: Edge data preprocessing, cleaning outliers in the time series data at the edge, performing differential encoding on the cleaned time series data, and storing the compressed time series data into the edge time series database; S3: Lightweight anomaly detection at the edge: A pruned isolated forest algorithm model is loaded at the edge to calculate anomaly scores on the time-series data and the non-time-series static data and semi-structured data stored at the edge. S4: Cloud-edge data transmission, the edge terminal periodically calculates the time-series data to generate an aggregation result, and transmits the abnormal time-series data that generates the abnormal score and the aggregation result data together to the cloud; S5: Cloud data processing: The uploaded abnormal time series data and the aggregation results are compressed in the cloud and stored in the cloud time series database; S6: Data fusion, generating multimodal datasets in the cloud.

2. The method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems according to claim 1, characterized in that: The non-time-series static data includes a battery cluster file table and a sensor parameter table. The battery cluster file table includes at least a Device_ID field, and the sensor parameter table includes at least a Sensor_ID field. The semi-structured data includes maintenance logs and equalization calibration data.

3. The method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems according to claim 1, characterized in that, In S3, the pruned isolated forest algorithm model selects a maximum of 3 core features and performs lightweight processing on the data corresponding to the core features.

4. The method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems according to claim 3, characterized in that, The lightweighting process includes normalizing the measured value (Raw_Value) to reduce its numerical scale; converting the data type of the device identifier (Device_ID) or sensor identifier (Sensor_ID) to an integer; and extracting the periodic features of the timestamp (Time).

5. The method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems according to claim 1, characterized in that, In S3, the pruned isolated forest algorithm model uses a two-dimensional array to store tree nodes; it shares the storage of duplicate data; it prunes useless nodes and retains only the key nodes that distinguish between normal and abnormal nodes.

6. The method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 1, characterized in that, In step S2, the differential encoding is calculated as the difference between the current period's time series data and the previous period's data sequence. If the absolute value of the current period's difference is |X... i If |>N*X, then the previous difference is used to replace the current difference, where X represents the mean of the differences in the previous M periods, and N and M are both constants.

7. The method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 1, characterized in that, In S5, Zstandard compression is used, with levels 5-8 selected.

8. The method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 1, characterized in that, In S4, the aggregation results include short-term mean, maximum value, and minimum value.

9. The method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 1, characterized in that, S5 also includes model training and updating, whereby the model is trained in the cloud based on the uploaded time-series data, the non-time-series static data and the semi-structured data stored in the cloud, anomaly scores are calculated, and the trained model parameters are sent to the edge device.

10. A method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 9, characterized in that, In S5, the data used for model training is divided into a training set and a validation set according to a certain ratio. When the anomaly detection accuracy of the validation set meets the requirements, the model parameters are sent to the edge.

11. A method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 9, characterized in that, In step S5, the model training uses either full data or incremental data.

12. The method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems according to claim 1, characterized in that, It also includes S7, which is a warning and control linkage mechanism. When the abnormal score is between 0.6 and 0.9, a warning command is triggered, and the warning command is sent to the control layer through the data protocol.

13. A method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 12, characterized in that, In S7, the anomaly score comes from the edge or the cloud.

14. The method for multimodal time-series big data fusion storage and health monitoring of electrochemical energy storage systems according to claim 12, characterized in that, In step S7, the aggregated result corresponding to the abnormal score is verified a second time, triggering an early warning command.

15. A method for multimodal time-series big data fusion storage and health monitoring for electrochemical energy storage systems according to claim 1, characterized in that, S1 also includes associated scheduling, which caches recently queried hash indexes at the edge and recently active hash indexes in the cloud.

16. A multimodal time-series big data fusion storage and health monitoring system for electrochemical energy storage systems, characterized in that, The steps of employing the method according to any one of claims 1-15 include: The data acquisition and database deployment module collects time-series data from sensors related to electrochemical energy storage devices. It deploys the non-time-series static database, the semi-structured database, and the time-series database at the edge and cloud respectively. The time-series data includes at least a timestamp (Time), device identifier (Device_ID), sensor identifier (Sensor_ID), measured value (Raw_Value), and anomaly score (Quality_Flag). The non-time-series static data includes at least a sensor identifier (Sensor_ID) and a device identifier (Device_ID) field. The semi-structured data includes at least a timestamp (Time) and a device identifier (Device_ID) field. A two-dimensional hash index based on the device identifier (Device_ID) and timestamp (Time) is established among the three databases. The edge data preprocessing module cleans outliers in the time series data at the edge, performs differential encoding on the cleaned time series data, and stores the compressed time series data into the edge time series database. The lightweight anomaly detection module at the edge loads a pruned isolated forest algorithm model at the edge to calculate anomaly scores on the time-series data and the non-time-series static data and semi-structured data stored at the edge. The cloud-edge data transmission module periodically calculates the time-series data at the edge to generate an aggregation result, and transmits the abnormal time-series data that generates the abnormal score along with the aggregation result data to the cloud. The cloud data processing module compresses the uploaded abnormal time-series data and the aggregation results in the cloud and stores them in the cloud time-series database. The data fusion module generates multimodal datasets in the cloud.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-15.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-15.

Images