Energy storage data processing method and system based on full-modal standardization and edge intelligence

Abstract

The embodiment of the invention provides an energy storage data processing method and system based on all-modal standardization and edge intelligence, and the method comprises the steps: collecting multi-source heterogeneous data of an energy storage system, carrying out the standardization processing, generating a unified data stream, and carrying out the encryption and anti-interference processing; carrying out decryption and security verification on the encrypted data; performing preprocessing and feature extraction on the verified data; performing time alignment processing on the preprocessed feature data, and performing hybrid fusion to obtain cross-modal fusion features; inputting the cross-modal fusion features into a lightweight artificial intelligence model for real-time reasoning, and outputting a fault risk probability; performing risk grading based on the fault risk probability, and generating a corresponding control instruction; and executing the control instruction, and performing feedback verification on an execution effect. According to the technical scheme provided by the invention, the availability, the safety and the intelligent level of the energy storage data can be remarkably improved, and a key technical support is provided for efficient operation of a novel power system.

Description

Technical Field

[0001] This invention belongs to the field of real-time energy storage data processing technology, and particularly relates to an energy storage data processing method and system based on full-modal standardization and edge intelligence. Background Technology

[0002] With the challenges of intermittency and volatility brought about by the large-scale grid connection of renewable energy (wind power and photovoltaics), data processing of energy storage systems (such as lithium-ion batteries, flow batteries, and compressed air energy storage) has become the core of system operation optimization. However, its multi-source heterogeneity, high-dimensional real-time requirements and strong security requirements mean that the existing "cloud-edge collaborative energy storage data standardization processing platform" still faces problems such as incomplete data standardization, contradiction between real-time performance and computing resources, security and privacy defects, and insufficient dynamic adaptability.

[0003] There are still four core problems in current energy storage data processing: First, existing technologies mainly focus on the standardized processing of structured electrochemical data such as voltage and current. However, for unstructured, multimodal data such as infrared thermal imaging, maintenance log text, and mechanical vibration signals, there is a lack of effective unified characterization and fusion methods. This makes it difficult to integrate data from different sources and formats, preventing true multimodal data collaborative analysis and limiting the accuracy of fault diagnosis and condition assessment. Second, deploying and running complex AI models (such as Transformer) on resource-constrained edge computing devices faces significant challenges. Computing bottlenecks lead to delays in analysis and decision-making in critical scenarios (such as frequency regulation and fault protection). Meeting real-time requirements (e.g., within 100 milliseconds) restricts the improvement of the rapid response and proactive safety control capabilities of energy storage systems. Third, existing systems often use plaintext transmission protocols (such as MQTT) and static access control mechanisms, which pose risks of data leakage and tampering. Specific security threats include falsification of battery state of charge values ​​and leakage of core process parameters (such as electrolyte formulations). There is a lack of a comprehensive and proactive security protection system covering identity authentication, transmission encryption, storage protection, and dynamic access control. Finally, standardized processing and analysis methods based on fixed rules are difficult to adapt to constantly changing external conditions and technological developments. The system cannot dynamically adjust its processing strategies, leading to performance degradation and a lack of long-term applicable flexibility and robustness. Summary of the Invention

[0004] The energy storage data processing method and system based on full-modal standardization and edge intelligence provided in this application can achieve unified data format, low-latency intelligent analysis, proactive defense security system and dynamic feature optimization, significantly improving the availability, security and intelligence level of energy storage data, and providing key technical support for the efficient operation of new power systems.

[0005] In a first aspect, embodiments of this application provide an energy storage data processing method based on full-modal standardization and edge intelligence, including:

[0006] Collect multi-source heterogeneous data from the energy storage system, standardize the multi-source heterogeneous data to generate a unified data stream, and then encrypt and anti-interference the unified data stream to generate encrypted data, which is then transmitted to the edge node.

[0007] Edge nodes decrypt the received encrypted data to obtain plaintext data, and perform security verification on the plaintext data; plaintext data that fails verification is filtered out, and plaintext data that passes verification is preprocessed and features are extracted to obtain preprocessed feature data;

[0008] The preprocessed feature data is time-aligned, and the time-aligned preprocessed feature data is then mixed and fused to obtain cross-modal fused features.

[0009] Cross-modal fusion features are input into a lightweight artificial intelligence model for real-time inference, and the failure risk probability is output. Risk is classified based on the failure risk probability, and corresponding control instructions are generated. The control instructions are executed, and the execution effect is verified by feedback.

[0010] In one alternative implementation, the multi-source heterogeneous data includes electrochemical parameters, thermal imaging data, vibration signals, and maintenance logs.

[0011] In one alternative implementation, the multi-source heterogeneous data is standardized to generate a unified data stream, including:

[0012] The protocol parsing function converts heterogeneous data from different industrial protocols into an intermediate representation.

[0013] The intermediate representation is transformed into a unified data stream through a standardized mapping function;

[0014] Configure a buffer queue to manage the instantaneous peak values ​​of a unified data stream.

[0015] In one alternative implementation, the unified data stream is encrypted and subjected to anti-interference processing, including:

[0016] Add device identifiers and timestamp metadata to the unified data stream and encode them in binary;

[0017] The encoded data is encrypted using the SM4 algorithm;

[0018] Differential transmission technology and cyclic redundancy check algorithm are introduced during data transmission.

[0019] In one optional implementation, the verified plaintext data undergoes preprocessing and feature extraction, including:

[0020] We use physical thresholding and statistical methods to filter out abnormal data and interpolate missing data to complete it.

[0021] For structured data, time-domain and frequency-domain features are extracted; for unstructured data, lightweight convolutional networks and natural language processing techniques are used to extract visual and semantic features.

[0022] In one optional implementation, time-domain and frequency-domain features are extracted for structured data, and visual and semantic features are extracted for unstructured data using lightweight convolutional networks and natural language processing techniques, including:

[0023] For electrochemical parameters and vibration signals, the mean and variance time-domain statistics are calculated using a sliding window, and the fundamental and harmonic frequency domain components are extracted using fast Fourier transform.

[0024] For thermal imaging data, a lightweight convolutional neural network is used to extract hotspot coordinates and temperature gradient visual features.

[0025] For operation and maintenance logs, key fault terms are extracted using natural language processing technology and transformed into structured text features.

[0026] In one optional implementation, time alignment processing is performed on the preprocessed feature data, including:

[0027] Using high-frequency sampled electrochemical parameters as a time reference, low-frequency sampled thermal imaging data is linearly interpolated and resampled.

[0028] By combining the timestamps of key events marked in the operation and maintenance logs, a spatiotemporal correlation matrix of abnormal events is established.

[0029] In one optional implementation, the time-aligned preprocessed feature data is mixed and fused, including:

[0030] Early feature-level fusion of electrochemical and thermal imaging features;

[0031] Vibration characteristics and operation and maintenance log characteristics are fused at the late decision level; the late decision level fusion adopts a dynamic weighting strategy.

[0032] In one alternative implementation, the lightweight AI model is a quantized and compressed TinyViT model, and the end-to-end latency of real-time inference is controlled within a preset threshold.

[0033] In one optional implementation, risk classification is performed based on the probability of failure risk, and corresponding control commands are generated, including:

[0034] Multiple risk probability thresholds are set to classify risks into different levels;

[0035] Different combinations of control commands and execution priorities are preset for different risk levels.

[0036] In one optional implementation, feedback verification of the execution effect includes:

[0037] Within the first preset time after the control command is issued, an inner loop verification is performed to check whether the execution result meets expectations.

[0038] The outer loop is periodically evaluated using the second preset time period to analyze the long-term effectiveness of the control strategy.

[0039] Based on the verification and evaluation results, the control parameters are automatically adjusted.

[0040] In one alternative implementation, it further includes:

[0041] Key data and events are uploaded to the cloud and securely stored using blockchain technology.

[0042] The lightweight artificial intelligence model is dynamically optimized based on cloud-based big data analysis results, and the optimized model is incrementally updated to the edge nodes.

[0043] In one alternative implementation, blockchain technology is used to securely store evidence of key data and events, including:

[0044] Use the SM3 hash algorithm to generate digital fingerprints of key data and events;

[0045] Digital fingerprints, timestamps, and device information are stored together in a blockchain block to form an immutable evidence storage chain.

[0046] Secondly, embodiments of this application provide an energy storage data processing system based on full-modal standardization and edge intelligence, including:

[0047] The data acquisition and encrypted transmission module is used to acquire multi-source heterogeneous data from the energy storage system, standardize the multi-source heterogeneous data, generate a unified data stream, encrypt and anti-interference the unified data stream to generate encrypted data, and then transmit it to the edge node.

[0048] The edge real-time preprocessing module is used to decrypt the received encrypted data to obtain plaintext data, perform security verification on the plaintext data, filter out plaintext data that fails verification, and perform preprocessing and feature extraction on the plaintext data that passes verification to obtain preprocessed feature data.

[0049] The multimodal data fusion module is used to perform time alignment processing on the preprocessed feature data, and then mix and fuse the time-aligned preprocessed feature data to obtain cross-modal fused features.

[0050] The real-time analysis and control module is used to input cross-modal fusion features into a lightweight artificial intelligence model for real-time inference and output the failure risk probability; perform risk classification based on the failure risk probability and generate corresponding control commands; execute the control commands and provide feedback verification on the execution effect.

[0051] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method provided in embodiments of this application.

[0052] Fourthly, embodiments of this application provide a computer-readable storage medium, characterized in that it stores a computer program thereon, which, when executed in a computer, causes the computer to perform the method provided in embodiments of this application.

[0053] The technical solution provided in this application has the following beneficial effects:

[0054] This invention proposes a real-time energy storage data processing method based on full-modal standardization and edge intelligence. By constructing a unified representation model for multi-source heterogeneous data, it achieves standardized processing of multi-modal data such as electrochemical parameters, thermal imaging, and vibration signals. A lightweight real-time analysis architecture on the edge side is designed, using optimized models such as TinyViT combined with 8-bit quantization and TensorRT acceleration technology to reduce inference latency to less than 8ms. Simultaneously, a dynamic weighted fusion algorithm is developed, which performs early feature fusion for strongly correlated data and late decision fusion for heterogeneous data, and introduces a risk-adaptive weighting mechanism. Data security is ensured through blockchain notarization and a four-fold encryption system (X.509 authentication + SM4 transmission encryption + AES-256 storage encryption + ABE access control). Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the energy storage data processing method based on full-modal standardization and edge intelligence provided in the embodiments of this application;

[0056] Figure 2 This is a schematic diagram of the edge real-time preprocessing process provided in the embodiments of this application;

[0057] Figure 3 This is a schematic diagram of the multimodal data fusion process provided in the embodiments of this application;

[0058] Figure 4 This is a flowchart illustrating the real-time analysis and control process provided in the embodiments of this application;

[0059] Figure 5 This is a flowchart illustrating the collaborative execution of control instructions by the actuators according to an embodiment of this application;

[0060] Figure 6 This is a schematic diagram of the cloud collaboration and evidence storage process provided in the embodiments of this application;

[0061] Figure 7 This is a schematic diagram of the architecture of the transmission protocol stack provided in an embodiment of this application;

[0062] Figure 8 This is a schematic diagram of the architecture of the analysis engine provided in the embodiments of this application;

[0063] Figure 9 This is a schematic diagram of the architecture of an energy storage data processing system based on full-modal standardization and edge intelligence provided in an embodiment of this application. Detailed Implementation

[0064] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] This application provides an energy storage data processing method based on full-modal standardization and edge intelligence. Figure 1 This is a flowchart illustrating the energy storage data processing method based on full-modal standardization and edge intelligence provided in this application embodiment. The method can be executed by an energy storage data processing system based on full-modal standardization and edge intelligence. The system can be implemented by software and / or hardware and can be configured in electronic devices such as computers.

[0066] like Figures 1-6 As shown, the technical solution provided in this application includes the following steps:

[0067] S110. Collect multi-source heterogeneous data from the energy storage system, standardize the multi-source heterogeneous data, generate a unified data stream, encrypt and anti-interference the unified data stream to generate encrypted data, and then transmit it to the edge node.

[0068] The classification of multi-source heterogeneous data from energy storage systems is shown in Table 1:

[0069] Table 1 Classification of Multi-Source Data

[0070]

[0071] In this embodiment, the multi-source heterogeneous data includes electrochemical parameters, thermal imaging data, vibration signals, and operation and maintenance logs.

[0072] In some embodiments, standardization processing is performed on multi-source heterogeneous data to generate a unified data stream, including:

[0073] The protocol parsing function converts heterogeneous data from different industrial protocols into an intermediate representation.

[0074] The intermediate representation is transformed into a unified data stream through a standardized mapping function;

[0075] Configure a buffer queue to manage the instantaneous peak values ​​of a unified data stream.

[0076] Specifically, the protocol parsing module first performs plug-and-play parsing of various industrial protocols such as CAN bus, Modbus, and MQTT, supporting flexible expansion with new protocols. Simultaneously, a data buffer queue is set up to specifically handle potential instantaneous peaks in high-frequency data (such as 1kHz sampled current data), ensuring the system can stably process data streams without information loss under high load. This includes:

[0077] Define the original data stream set Each of them It belongs to some kind of agreement ;d i Let P represent the i-th original data, n represent the total number of original data, and P represent the total number of original data. j Indicates the j-th communication protocol;

[0078] Through protocol parsing function The agreement Data d i Convert to intermediate representation ;

[0079] By standardizing the mapping function Generate a unified data stream ; where s i This represents the i-th data after processing.

[0080] The buffer queue model includes:

[0081] Set sampling frequency The amount of data arriving within the time window Δt is N= ;

[0082] Queue capacity ≥ ,satisfy ,in This indicates an indicator function, where t represents time, k represents the time index, and d represents the time index. k This represents the amount of data arriving at time k. This indicates the maximum expected peak value.

[0083] The complete conversion process is as follows:

[0084] st ;

[0085] in, This represents the compound operators for functions.

[0086] The constraints include:

[0087] Real-time requirements: ;

[0088] in, Indicates the parsing time. Indicates buffer time. Indicates processing time, i.e., the execution time of the function. The computation time for g;

[0089] Integrity constraint: ||S||≡||D||, meaning no data loss.

[0090] In some embodiments, the encryption scheme employs a multi-layered security mechanism to comprehensively ensure communication security. For data transmission encryption, the SM4 algorithm is selected, using 128-bit block cipher and CBC mode to ensure data confidentiality during transmission. For key management, quantum key distribution (QKD) technology is introduced to achieve highly secure key exchange, with session keys updated every 10 minutes to enhance resistance to cracking. For authentication, a two-way TLS 1.3 handshake protocol is used, and device certificates are based on the X.509 standard to achieve mutual authentication between communicating parties, preventing unauthorized access. For replay attack prevention, adding millisecond-level timestamps (Nonce values) to messages and combining them with random number signatures effectively prevents attackers from repeatedly sending old messages for malicious operations. The overall scheme integrates domestically developed cryptographic algorithms, quantum security technology, and modern network security protocols to construct an efficient, reliable encryption system that meets national information security requirements.

[0091] In some embodiments, the unified data stream is encrypted and subjected to anti-interference processing, including:

[0092] Add device identifiers and timestamp metadata to the unified data stream for subsequent tracking and management, and encode the data in binary (such as Apache Avro or Protocol Buffers) to reduce storage and transmission overhead;

[0093] To ensure data security, the SM4 algorithm can be used to encrypt the encoded data, and a secure channel can be established using the TLS protocol during transmission to prevent data leakage or tampering; specifically including:

[0094] Metadata injection:

[0095] ;

[0096] Where ⊕ represents data concatenation, || represents field concatenation, and D rawThis represents the raw data in the unified data stream, where ID represents the device identifier and timestamp represents the timestamp. This represents the data after metadata injection enhancement;

[0097] Binary encoding:

[0098] ;

[0099] in, This represents the encoded data. Represents binary encoding functions;

[0100] Encrypted transmission:

[0101] ;

[0102] Where K represents the encryption key, This indicates that the SM4 algorithm is used to encrypt the encoded data. S represents the encrypted data. cert It represents the certificate information of the two communicating parties, such as public keys and signatures. TLS represents the TLS protocol, and Channel represents the establishment of a secure transmission channel.

[0103] This ultimately forms a secure data processing chain:

[0104] ;

[0105] in, Represents function composition. Metadata refers to the data that is ultimately transmitted securely.

[0106] Data is transmitted to edge nodes via a 5G / fiber optic private network, ensuring low latency and high reliability. Edge nodes verify signatures before decryption, ensuring the credibility of data sources and the confidentiality of content, achieving end-to-end security.

[0107] To ensure reliable transmission of sensor signals in complex electromagnetic environments, the system employs differential transmission technology (such as RS485) to cancel common-mode interference through positive and negative dual-wire cancellation, enhancing anti-interference capabilities at the physical layer. Magnetic ring filters are installed on critical lines to effectively absorb high-frequency noise. During data transmission, the CRC-32 checksum algorithm is used to generate a unique checksum for each data packet. If the receiver fails to verify the data, a retransmission mechanism is automatically triggered, forming a closed loop of "transmission-verification-error correction," achieving a bit error rate of less than 10%. -9 Industrial-grade reliability; specifically including:

[0108] Differential transmission interference immunity:

[0109] Common-mode rejection ratio >60dB;

[0110] Among them, V 接收 V represents the effective signal voltage extracted by the receiving end. + V represents the positive phase differential signal voltage. - This represents the negative phase differential signal voltage;

[0111] Magnetic ring filtering effect:

[0112] ;

[0113] Where H(f) represents the frequency response function of the magnetic ring, f represents the input signal frequency, and f c This represents the cutoff frequency, and j represents the imaginary unit.

[0114] Data verification protocol:

[0115] ;

[0116] ;

[0117] Where CRC represents Cyclic Redundancy Check, and Data represents the data block to be checked. This indicates the generator polynomial, CRC_check indicates the check result calculated by the receiver, and retransmission indicates whether the retransmission mechanism is triggered.

[0118] Reliability metrics (bit error rate):

[0119] ;

[0120] in, This represents the probability of a bit error.

[0121] In some embodiments, the fault-tolerance strategy is designed to improve the stability and continuity of the system under abnormal conditions. Corresponding handling mechanisms are adopted for different fault types: when communication is interrupted, the system automatically activates the local caching function to temporarily store data in the local device, supporting data retention for up to 24 hours, and automatically uploads the data after communication is restored; if a sensor fails, data from a normally functioning sensor in an adjacent area is used as a substitute through an intelligent scheduling algorithm to ensure the continuity and integrity of monitoring data; in the event of a key update failure, the system automatically switches to a pre-set backup key to ensure uninterrupted encrypted communication. This fault-tolerance mechanism achieves multi-level redundancy protection from communication and sensing to security layers, effectively improving the robustness and availability of the system, and is suitable for application scenarios with high reliability requirements such as industrial IoT and smart cities.

[0122] Table 2 shows a comparison of the performance indicators of existing technologies and the present invention in terms of encrypted data transmission:

[0123] Table 2 Comparison of data encryption transmission performance indicators between existing technologies and the present invention.

[0124]

[0125] As shown in Table 2, this invention significantly outperforms existing technologies in terms of real-time data transmission, security, compatibility, and anti-interference capabilities, constructing a high-performance, highly secure, and highly reliable data communication system for industrial IoT scenarios. Specifically:

[0126] Extremely low transmission latency: The edge node receiving latency is less than 10ms, which is an order of magnitude higher than the existing technology (50~200ms), meeting the stringent real-time requirements of industrial control, intelligent sensing and other applications.

[0127] Higher encryption strength: It adopts the national cryptographic SM4 algorithm combined with quantum key distribution (QKD) to realize dynamic key updates, which has stronger anti-cracking ability and long-term security compared with AES-128 static keys, and meets the strategic requirements of national information security self-control.

[0128] Superior protocol compatibility: Supports more than 10 industrial communication protocols, breaking through the limitation of traditional systems that only support Modbus / MQTT, and realizing unified access and integrated management of multi-source heterogeneous devices;

[0129] Outstanding anti-interference capability: It has passed the GB / T17626EMC standard test, has a complete electromagnetic compatibility design, can operate stably in complex electromagnetic environments, and is suitable for strong interference industrial sites such as power, transportation, and manufacturing.

[0130] S120. The edge node decrypts the received encrypted data to obtain plaintext data, performs security verification on the plaintext data, filters out plaintext data that fails verification, and preprocesses and extracts features from the plaintext data that passes verification to obtain preprocessed feature data.

[0131] In some embodiments, the system employs a dual security mechanism of SM4 decryption and ECDSA signature verification: edge nodes first decrypt the data stream, then verify data integrity and source reliability through elliptic curve signature. Data that fails verification is automatically discarded, forming a "decryption-verification-filtering" security protection process. Specifically, this includes:

[0132] Decryption phase:

[0133] ;

[0134] Verification phase:

[0135] ;

[0136] Security Filtering:

[0137] ;

[0138] In the formula: This represents the encrypted data stream. This indicates the pre-stored session key. This represents the decrypted plaintext data. Indicates digital signature, Indicates the device's public key. This represents the verification function of the elliptic curve digital signature algorithm. This indicates the verification result flag, and Output indicates the final output data. This indicates that the data has been discarded.

[0139] In some embodiments, the verified plaintext data undergoes preprocessing and feature extraction, including:

[0140] Data cleaning: The system employs a dual data cleaning mechanism: 1) Physical thresholding to remove out-of-limit data (e.g., voltage > 4.2V); 2) The 3σ rule to filter abnormal fluctuations (e.g., temperature jumps of ±5℃). Missing data handling: Key parameters (voltage / current) are linearly interpolated within 10ms; non-key parameters (humidity) are filled with the moving average or left blank, forming a cleaning process of "threshold filtering → statistical denoising → intelligent completion". Specifically, this includes:

[0141] Dual filtering mechanism:

[0142] ;

[0143] Missing data completion:

[0144] ;

[0145] Flow function:

[0146] ;

[0147] In the formula: x represents the current sampling point, =2.8V indicates the lower limit of the hard threshold. =4.2V represents the hard threshold upper limit, μ represents the moving mean, σ represents the dynamic standard deviation, and μ and σ are dynamically updated. Indicates valid data output, MA n This represents an n-point moving average, where n represents the number of sampling points, for example, n=60. This represents the hard threshold filtering function. This represents the 3σ statistical filtering function. This represents a linear interpolation function (within 10ms). This represents the data after final cleaning. x represents the estimated value of missing data. t+1 x t-1 This indicates data from adjacent time points. Keep means to retain, and Discard means to remove.

[0148] For structured data, time-domain and frequency-domain features are extracted; for unstructured data, lightweight convolutional networks and natural language processing techniques are used to extract visual and semantic features.

[0149] For structured data (such as electrochemical parameters and mechanical vibrations), the system employs a multi-dimensional feature extraction strategy: it calculates time-domain statistics (mean and variance) through a sliding window (1s / 100ms), and combines this with FFT to extract the 50Hz fundamental frequency and the 3rd / 5th / 7th harmonics, constructing a "time-domain + frequency-domain" composite feature matrix to provide high-density feature input for equipment diagnostics. Specifically, this includes:

[0150] Time series characteristics (sliding window calculation):

[0151] ;

[0152] Where the window size N = 1 second of data points, such as 1000 points @ 1kHz sampling, the step size is 100ms, that is, the window slides 10% each time, x k Represents time-domain sampling points, σ represents the time-domain mean. 2 Represents the time-domain variance;

[0153] Frequency domain features (FFT extraction):

[0154] ;

[0155] Among them, A 50 A represents the amplitude of the 50Hz fundamental frequency. 150 A 250 A 350 These represent the amplitudes of the 3rd, 5th, and 7th harmonics, respectively, corresponding to amplitudes of 150, 250, and 350 Hz.

[0156] Fusion feature F:

[0157] ;

[0158] T represents the transpose operation, and the final output is a 6-dimensional feature vector.

[0159] For unstructured data, the system employs multimodal parsing technology: thermal imaging data is intelligently extracted using MobileNetV3 to obtain hotspot coordinates and temperature values, achieving image-to-spatial feature conversion; maintenance logs are processed using NLP to extract key fault terms (such as "overvoltage"), which are then converted into structured tags, constructing a text-to-semantic parsing channel, ultimately forming a quantifiable multimodal feature set. Specifically, this includes:

[0160] Thermal Imaging Analysis:

[0161] ;

[0162] Output: Hotspot coordinates + highest temperature;

[0163] Text parsing:

[0164] ;

[0165] Fusion feature F:

[0166] ;

[0167] In the formula: I represents the infrared image matrix; CNN represents a lightweight convolutional network; NLP represents a keyword extraction model; (x,y) represents the coordinates; T max The highest temperature value is indicated by the label, which represents a structured fault label, such as "overvoltage", "abnormal temperature", "communication interruption", etc.

[0168] Furthermore, the system employs spatiotemporal alignment technology to spatiotemporally correlate electrochemical voltage drops (>0.5V) with infrared hotspots, establishing an "electro-thermal" causal mapping. Through hybrid feature stitching, electrochemical parameters (voltage / internal resistance) and visual features (hotspot coordinates / temperature) are fused into a cross-modal matrix, which is then input into the lightweight TinyViT model, achieving an intelligent closed-loop collaborative diagnosis of "temporal-spatial-multi-source features." Specifically, this includes:

[0169] Spatiotemporal alignment model:

[0170] Define a collection of event timestamps:

[0171] ;

[0172] Alignment matching function:

[0173] ;

[0174] Feature extraction:

[0175] Electrochemical eigenvector f ec :

[0176] ;

[0177] Image feature vector f ir :

[0178] ;

[0179] Among them, t v t represents the moment of the voltage drop event. ir The value represents the moment when the infrared image was acquired, N represents the total number of voltage drop events, and M represents the total number of infrared image frames acquired. This represents the spatiotemporal alignment matching function. Indicates the maximum permissible time deviation. Represents the voltage surge, (x hot ,y hot) Indicates hotspot coordinates, This indicates the area of ​​the hot zone, that is, the extent to which the high-temperature region is covered in the image. , Represents the dimension of the real number space.

[0180] Lightweight AI model inference target

[0181] In some embodiments, the system employs the lightweight TinyViT-1.0 model (with only 1M parameters) for multimodal fault diagnosis. Its input consists of fused cross-modal features (dimension ≤ 100), and its output includes fault probability assessments (such as a continuous value of 0~1 for thermal runaway risk) and control decisions (such as a power reduction charging command). Through deep optimization with TensorRT, end-to-end inference with an ultra-low latency of 8ms is achieved on Jetson Nano edge devices, meeting real-time requirements. This architecture, while maintaining lightweight design, achieves a millisecond-level closed-loop response of "feature input → risk quantification → control output," with the single inference time consistently controlled within 5ms in typical scenarios. Specifically, it includes:

[0182] Model Inference:

[0183] ;

[0184] Where: the input is (Fusing features); the output is the risk probability p∈[0,1] and the control command vector u;

[0185] Real-time constraints:

[0186] ;

[0187] Lightweight metrics:

[0188] ;

[0189] Where σ represents the Sigmoid function, W1 and W2 represent the model weight matrices, and b1 and b2 represent the bias terms. This represents the end-to-end inference time, with typical latency being the average inference time.

[0190] Control command generation

[0191] When the system detects a thermal runaway risk value exceeding the safety threshold of 0.8, it will immediately activate a Level 3 emergency response mechanism: firstly, the charging power will be intelligently reduced to a safe level of 50%, and simultaneously, the maximum cooling airflow of the liquid cooling system will be automatically activated. The entire process, from risk detection to control execution, has a strictly controlled response time of less than 20 milliseconds. To ensure event traceability, the system will simultaneously encrypt and upload the warning information to the cloud management platform, and use blockchain technology for immutable evidence storage, forming a complete security closed loop of "real-time response - proactive protection - trusted evidence storage." Specifically, this includes:

[0192] Triggering conditions:

[0193] ;

[0194] In the formula, p∈[0,1] is the risk probability, representing the risk of thermal runaway;

[0195] Emergency control:

[0196] ;

[0197] Response speed: Detection → Execution time ≤ 20ms;

[0198] Secure Evidence Storage:

[0199] Cloud records = encrypted (alarm + timestamp), and the encryption uses the SM4 national cryptographic algorithm;

[0200] Blockchain evidence storage = hash (record), and the hash uses the SM3 national cryptographic algorithm.

[0201] Table 3 shows a comparison of the performance indicators of existing technologies and the present invention in data preprocessing:

[0202] Table 3 Comparison of data preprocessing performance indicators between existing technologies and the present invention.

[0203]

[0204] As shown in Table 3, this solution achieves a comprehensive breakthrough compared to existing technologies in terms of feature dimension, anomaly detection accuracy, and resource consumption. First, by fusing multimodal data (electrochemical + thermal imaging + text, etc.), a high-density feature vector of ≤100 dimensions is constructed, which is significantly superior to the low-dimensional representation of traditional single electrochemical modes, thus improving the system's ability to perceive complex faults. Second, the introduction of the lightweight TinyViT model for intelligent inference increases the anomaly detection F1 score from 0.85 in the traditional threshold rule to 0.95, greatly enhancing the recognition accuracy and robustness. Finally, the entire system is deployed on ARM architecture edge devices (such as Jetson Nano), without relying on x86 servers, achieving ultimate optimization of computing resources and decentralized deployment.

[0205] S130. Perform time alignment processing on the preprocessed feature data, and then mix and fuse the time-aligned preprocessed feature data to obtain cross-modal fusion features.

[0206] In some embodiments, time alignment processing is performed on the preprocessed feature data, including:

[0207] Using high-frequency sampled electrochemical parameters as a time reference, low-frequency sampled thermal imaging data is linearly interpolated and resampled.

[0208] By combining the timestamps of key events marked in the operation and maintenance logs, a spatiotemporal correlation matrix of abnormal events is established.

[0209] To address the cross-modal sampling rate mismatch between electrochemical data (1kHz) and infrared images (5Hz), the system employs a dual time alignment strategy: First, using the high-frequency electrochemical signal as a reference, low-frequency infrared data is resampled to 1kHz in real-time through linear interpolation to ensure strict time synchronization. Simultaneously, by combining key event markers in the operation and maintenance logs (such as manually recorded entries like "14:00:00 single-cell overvoltage"), a spatiotemporal correlation matrix of abnormal events is established based on precise timestamp matching, forming a double-insurance mechanism of "interpolation synchronization + log calibration." This solves the real-time alignment problem of continuous data streams while ensuring accurate correlation analysis of discrete, sudden events. Specifically, this includes:

[0210] Interpolation synchronization:

[0211] ;

[0212] In the formula: I(t) represents the raw infrared data (5Hz); This represents the time-aligned 1kHz data. t represents the current infrared frame acquisition time. next Indicates the time of the next infrared acquisition;

[0213] Event Relationship Matrix:

[0214] ;

[0215] In the formula: Δt represents the time tolerance, such as ±100ms. This represents the timestamp of the i-th electrochemical event. M represents the timestamp of the j-th log event. i,j Represents the elements of the event association matrix;

[0216] Double insurance conditions:

[0217] Effective association (Interpolation synchronization)∧( =1);

[0218] ∧ represents the logical AND relation;

[0219] The double-insurance condition means that an anomalous event is considered a "valid association" if and only if:

[0220] Infrared data has been synchronized and aligned with electrochemical signals via interpolation;

[0221] Furthermore, the event is recorded in the logs, and the time deviation is within the acceptable range;

[0222] The electrochemical data sampling frequency is 1 kHz; the infrared data sampling frequency is 5 Hz, and it is resampled to 1 kHz in real time through linear interpolation.

[0223] Furthermore, we first conceived of an idealized, complete model F that integrates all information. frused The system employs multimodal feature collaborative extraction: electrochemical data is processed using a sliding window to calculate time-domain features (mean, variance, slope) and FFT harmonic components (150Hz / 250Hz); thermal imaging data is processed using MobileNetV3 to extract hotspot coordinates and temperature gradients, combined with frame differencing to capture abrupt changes; vibration signals are processed using wavelet transform to extract the 50-100Hz fault frequency band; and maintenance logs are processed using NLP to parse keywords (such as "overvoltage") and quantify the urgency level. This ultimately forms a "time-frequency-thermal-semantic" fusion feature to support multimodal fault diagnosis. Specifically, this is represented as follows:

[0224] Feature extraction function:

[0225] ;

[0226] Feature fusion:

[0227] ;

[0228] Where: μ, These represent the mean and variance of the voltage calculated using the sliding window, respectively. ⊕ represents the FFT amplitude at frequency f; ⊕ represents the feature concatenation operation. Represents the feature weight matrix for each modality; OneHot represents the text keyword encoding; SentimentScore represents the sentiment analysis value (0-1). Represents the electrochemical eigenvectors. Represents the feature vector of thermal imaging. Represents the vibration eigenvector. Represents the text feature vector. Represents wavelet energy. This represents the fused feature vector.

[0229] However, this complete model is too complex and not conducive to real-time computation. Therefore, we designed an optimized, grouped processing hybrid fusion strategy, which greatly improves efficiency while maintaining accuracy. This hybrid fusion strategy can be viewed as the baseline model F frused This is an efficient and practical simplified implementation method.

[0230] In some embodiments, the time-aligned preprocessed feature data is mixed and fused, including:

[0231] Early feature-level fusion of electrochemical and thermal imaging features;

[0232] Vibration characteristics and operation and maintenance log characteristics are fused at the late decision level; the late decision level fusion adopts a dynamic weighting strategy.

[0233] This invention employs a hybrid multimodal fusion strategy, intelligently selecting the fusion method: Early feature fusion is used for strongly correlated electrochemical and thermal imaging data, concatenating electrical parameters and visual features before inputting them into the TinyViT model for joint analysis; late-decision fusion is used for highly heterogeneous vibration signals and maintenance logs, first calculating the vibration anomaly probability and text risk score separately, then dynamically weighting them to generate a comprehensive fault score. This hybrid architecture takes into account both cross-modal correlation and data type differences. Specifically:

[0234] Hybrid fusion strategy:

[0235] ;

[0236] Early feature fusion (electrochemical + thermal imaging):

[0237] ;

[0238] Late-stage decision fusion (vibration + log):

[0239] ;

[0240] In the formula: Concat represents feature concatenation while preserving the original associations; α and β represent dynamic weights, which can be set to 0.6 and 0.4 respectively, for example. , Let x1 and x2 represent the single-modal analysis function, and x1 and x2 represent the input feature vectors, respectively representing the two modal data to be fused. Represents the fusion function. Indicates early fusion output, Represents the electrochemical eigenvectors. This represents the thermal imaging feature vector, where d1 and d2 represent the feature dimensions. Indicates the probability of vibration anomalies. The text represents the risk score, and the score represents the overall failure score.

[0241] In some embodiments, the rules are generated based on rules engine control instructions. For example, the rules engine may include:

[0242] When the risk of thermal runaway is greater than 0.9 and the temperature is greater than 80°C, disconnect the charging circuit and activate the fire suppression system.

[0243] When an abnormality occurs and the text label "mechanical loosening" is displayed, an inspection work order is triggered.

[0244] In some embodiments, visual feedback can also be set, such as heat map overlay: marking high voltage fluctuation areas on infrared images; time axis analysis: synchronously displaying abnormal periods of voltage, temperature, and vibration signals, etc.

[0245] Table 4 shows a comparison of the performance indicators of existing technologies and the present invention in data multimodal fusion:

[0246] Table 4. Comparison of performance indicators of data multimodal fusion between existing technologies and the present invention.

[0247]

[0248] As shown in Table 4, the multimodal fusion fault diagnosis method proposed in this invention achieves significant improvements in several key indicators compared to traditional single-modal analysis. In terms of fault detection accuracy, the F1 score increases from 0.82 (based solely on voltage signals) to 0.95, indicating a stronger comprehensive discrimination capability. Regarding false alarm control, the false alarm rate decreases from 15% to 5%, significantly improving alarm reliability and user trust. In terms of real-time performance, the end-to-end latency is reduced from 50ms to 12ms (based on TinyViT model optimization), meeting the millisecond-level response requirements of industrial scenarios. Furthermore, when dealing with new energy storage devices, this solution supports automatic feature adaptation without retraining the model, whereas traditional methods require model reconstruction and training, resulting in poor adaptability and high costs.

[0249] S140. Input the cross-modal fusion features into the lightweight artificial intelligence model for real-time inference and output the failure risk probability; classify the risk based on the failure risk probability and generate corresponding control instructions; execute the control instructions and verify the execution effect.

[0250] In some embodiments, inputting cross-modal fusion features into a lightweight artificial intelligence model for real-time inference includes:

[0251] TinyViT is a lightweight multimodal fault diagnosis model that employs a dual-branch architecture to process image and time-series data separately: the image branch uses 16×16 convolutions for block embedding, while the time-series branch extracts signal features. The features from both modalities are then fused through a cross-modal attention mechanism, ultimately outputting a 5-class fault classification result. This model, optimized with 8-bit quantization compression (reduced from 3MB to 0.8MB) and TensorRT acceleration, achieves an inference latency of less than 8ms on Jetson Xavier edge devices, combining efficient computation with multimodal fusion capabilities, making it suitable for real-time fault detection in industrial scenarios. Specifically, it includes:

[0252] Model calculation process

[0253] Input: Image input and time series data ;

[0254] step:

[0255] Feature extraction: ;

[0256] Feature fusion: ;

[0257] Classified output: ;

[0258] Performance optimization:

[0259] Model compression: ;

[0260] Inference speed: Time 8ms (Jetson Xavier);

[0261] Overall representation:

[0262] ;

[0263] In the formula: ⊕ represents the feature fusion operation; LSTM represents image block convolution; Fimg represents the image feature vector, which is a 64-dimensional feature extracted by Conv; Fts represents the temporal feature vector, which is a 64-dimensional feature extracted by LSTM. Let W represent the fused feature vector, W represent the classification weight matrix, and p represent the classification output probability. This represents the overall model function, Size represents the model size, Time represents the inference delay, and P represents the risk probability.

[0264] In some embodiments, risk classification is performed based on the probability of failure risk, and corresponding control commands are generated, including:

[0265] Multiple risk probability thresholds are set to classify risks into different levels;

[0266] Different combinations of control commands and execution priorities are preset for different risk levels.

[0267] For example, this embodiment adopts a three-level risk classification and decision-making mechanism, which realizes intelligent and differentiated response based on risk probability P, to ensure that reasonable control strategies are adopted under different risk levels, taking into account both safety and operational efficiency:

[0268] High risk (P>0.9): Determined to be a serious fault that is highly likely to occur, immediately execute the "emergency shutdown" action, with a response time of less than 10ms, to prevent the accident from escalating;

[0269] Medium risk (0.6≤P≤0.9): There is a high probability of failure, triggering the "dilution operation" measure to reduce the equipment load to delay deterioration, with a response time of less than 50ms;

[0270] Low risk (P<0.6): The risk is low, no intervention is required, only "continuous monitoring" is carried out to accumulate data for trend analysis and preventive maintenance.

[0271] Dynamic weight adjustment: Based on the risk type (temperature or voltage), a preset combination of control weights is dynamically returned, with temperature risk receiving 70% primary weight and voltage risk receiving 50% secondary weight. Specifically:

[0272] Define the weight mapping function:

[0273] ;

[0274] In the formula: w( ) is the weight allocation function; the output is a triple (sovereign weight, secondary weight, auxiliary weight); for temperature-related risks (thermal), the sovereign weight = 0.7; for voltage-related risks (voltage), the secondary weight = 0.5.

[0275] Control command generation: The system adopts a three-level priority control strategy, namely safety commands (emergency shutdown, etc.), equipment protection commands (power derating, etc.), and economic optimization commands (frequency regulation response, etc.), to ensure that safe operation takes priority over equipment protection, and equipment protection takes priority over economic optimization.

[0276] Furthermore, feedback verification of the implementation results is conducted, including:

[0277] Within the first preset time after the control command is issued, an inner loop verification is performed to check whether the execution result meets expectations.

[0278] The outer loop is periodically evaluated using the second preset time period to analyze the long-term effectiveness of the control strategy.

[0279] Based on the verification and evaluation results, the control parameters are automatically adjusted.

[0280] Dual closed-loop verification: Inner loop: Verify the execution result within 100ms after the command is issued; Outer loop: Evaluate the effectiveness of the control strategy every 5 minutes. Specifically:

[0281] Inner ring (quick check):

[0282] ;

[0283] Outer ring (regular inspection):

[0284] ;

[0285] Automatic adjustment

[0286] ;

[0287] Exception handling:

[0288] Control verification function:

[0289] ;

[0290] In the formula: e represents the control effect evaluation value (0~1); threshold θ=0.8, used to determine whether abnormal handling is triggered; ∧ represents "and execute" two operations.

[0291] Table 5 shows a comparison of the performance indicators of existing technologies and the present invention in real-time analysis and control:

[0292] Table 5 Comparison of performance indicators of real-time analysis and control between existing technologies and the present invention.

[0293]

[0294] As shown in Table 5, this invention comprehensively surpasses traditional solutions in terms of performance indicators, achieving a comprehensive improvement in high precision, low latency, strong coordination, and rapid recovery. End-to-end latency is reduced from the traditional 200-500ms to 8-50ms, significantly enhancing the system's real-time response capability; control accuracy is improved from ±5% to ±1.2%, greatly improving control accuracy; multi-device coordination error is reduced from 300ms of asynchrony to less than 50ms, achieving high-precision timing alignment and linkage control; and anomaly recovery time is shortened from 2-5 seconds to only 0.5 seconds, possessing millisecond-level fault self-healing capability.

[0295] Furthermore, this solution embodiment also includes a security protection mechanism, specifically including a three-level emergency channel and a blockchain audit security protection mechanism, achieving end-to-end security protection from software to hardware. When software failure occurs, a level one response is triggered, automatically resetting via a hardware watchdog to ensure uninterrupted system operation; when communication is interrupted, a level two response is initiated, executing locally cached control strategies to ensure critical instructions can still be executed; in extreme cases such as multi-module failures, a level three response is triggered, with physical fuses forcibly cutting off power to prevent the accident from escalating. Simultaneously, all control instructions are stored on the blockchain, supporting full-process traceability and responsibility identification, achieving "intelligent decision-making traceability and safe actions verifiable," comprehensively improving the system's reliability, security, and compliance.

[0296] S150: Upload key data and events to the cloud and use blockchain technology to securely store key data and events; dynamically optimize the lightweight artificial intelligence model based on cloud big data analysis results, and incrementally update the optimized model to edge nodes.

[0297] Figure 7 This is a schematic diagram of the architecture of the transmission protocol stack provided in an embodiment of this application. Figure 8 This is a schematic diagram of the architecture of the analysis engine provided in the embodiments of this application. See also: Figure 7 and Figure 8In some embodiments, the core process includes: hierarchical data uploading. The hierarchical data uploading strategy is divided into three levels based on data importance: urgent data (such as fault alarms) is uploaded immediately and stored permanently to ensure real-time response and long-term traceability of critical events; important data (such as SOC changes) is uploaded every minute and retained for one year for operational status analysis and trend prediction; routine data (such as temperature monitoring) is uploaded every 15 minutes and retained for three months, balancing storage costs and monitoring needs. The transmission protocol stack adopts the MQTT+JSONSchema application layer protocol, supporting lightweight message publish-subscribe; the security layer uses SM4+QKD dynamic encryption to ensure data confidentiality and resistance to quantum attacks; the transport layer is based on TCP / UDP and 5G slicing technology to achieve highly reliable, low-latency network communication. This strategy realizes a closed loop of data management of "on-demand uploading, hierarchical control, and secure reliability," significantly improving the efficiency and security of industrial edge systems.

[0298] Blockchain-based evidence storage includes:

[0299] Block structure:

[0300] ;

[0301] Evidence Storage Chain:

[0302] ;

[0303] Safety features:

[0304] ;

[0305] Participant set P:

[0306] P = {grid nodes, regulatory agencies, ...};

[0307] In the formula: H represents the hash function; negl represents the ignoreable function; Indicates that there is a unique one; B i This represents the i-th block; i represents the block index; t i Represents a timestamp; h i This represents the hash value of the current block; This indicates the probability of tampering.

[0308] Cloud-based big data analytics includes typical analytical tasks such as cross-site battery health comparison, economic benefit evaluation of charging and discharging strategies, and optimization of energy storage scheduling based on meteorological data.

[0309] Model dynamic optimization

[0310] A global model is generated by cloud aggregation; incremental updates are then distributed to each node.

[0311] Model update verification function :

[0312] ;

[0313] In the formula: This indicates a new model that needs to be updated; This represents the model evaluation function (returning test accuracy); τ represents the accuracy threshold; Sign&Deploy represents edge deployment after multi-party signature, and Reject represents rejection of deployment.

[0314] Zero Trust Security System

[0315] This system constructs a four-layer security protection system: 1) X.509 two-way authentication ensures device identity trustworthiness; 2) SM4 encryption ensures transmission security; 3) AES-256 protects locally stored data; 4) ABE enables fine-grained access control. These four layers of protection cover the entire process of identity authentication, transmission encryption, data storage, and access control, forming end-to-end security protection. Specifically:

[0316] Security protection system: Each protective layer is defined as follows:

[0317] 1) Identity Authentication Layer (Auth):

[0318] Auth=X.509( , ) {0,1} (output 1 indicates successful two-way authentication);

[0319] 2) Trans encryption layer:

[0320] Trans = SM4(M) → Ct (plaintext message M → ciphertext Ct);

[0321] 3) Data storage layer: Store

[0322] Store = AES-256(D) → Cs (data D → ciphertext Cs);

[0323] 4) Access Control Layer:

[0324] (Access decision for user u to permission p);

[0325] Security attribute Secure:

[0326] Secure=Auth∧Trans∧Store∧Access;

[0327] In the formula: D represents data, M represents plaintext message, C represents ciphertext, u represents user, and p represents permission.

[0328] Table 6 shows a comparison of the performance indicators of existing technologies and the present invention in cloud collaboration and evidence preservation:

[0329] Table 6. Comparison of performance indicators of cloud collaboration and evidence storage between existing technologies and the present invention.

[0330]

[0331] As shown in Table 6, this invention significantly outperforms traditional solutions in cloud collaboration and evidence storage. Evidence storage latency is reduced from the traditional 2-5 seconds to less than 500 milliseconds, achieving near real-time data upload and greatly improving event response speed. Data analysis real-time performance is improved from hours to minutes, supporting faster situational awareness and decision-making. Model update efficiency is upgraded from once a week to on-demand triggering, enhancing the system's adaptability. Audit response speed is reduced from days to seconds, significantly improving the timeliness of security event tracking and tracing. Overall performance represents a leap from "periodic and lagging" to "real-time and dynamic," fully demonstrating the technological advancement and engineering practicality of this invention in intelligent operation and maintenance and trusted control.

[0332] In some embodiments, this embodiment also constructs a three-level fault tolerance and data repair mechanism to achieve efficient response to typical abnormal scenarios. For communication interruptions, an edge local caching and breakpoint resumption strategy is adopted to ensure no data loss and no service interruption. When data inconsistency occurs, the blockchain consensus verification mechanism automatically identifies and corrects errors to ensure data consistency. In the event of a cloud failure, the system can automatically switch to a backup regional center to achieve business continuity. Furthermore, the system has intelligent data repair capabilities: when missing evidence data is detected, a verification request is automatically initiated to neighboring nodes, and data recovery is completed based on the consensus voting mechanism of smart contracts, ensuring the integrity and reliability of on-chain information. This mechanism significantly improves the system's robustness, availability, and trustworthiness, making it suitable for industrial-grade application scenarios with high security and high reliability requirements.

[0333] This invention innovatively proposes a standardized and unified characterization method for multi-source heterogeneous data (including but not limited to electrochemical parameters, thermal imaging, mechanical vibration, and operation and maintenance logs) of energy storage systems. It achieves efficient fusion of multimodal data by constructing a general data model and effectively solves the problems of sampling frequency differences and timestamp asynchrony across sources using an adaptive time alignment algorithm, ensuring the temporal consistency of multi-dimensional data. Lightweight neural network models such as TinyViT are employed to achieve real-time feature extraction and online fault diagnosis of multimodal data including electrochemical, thermodynamic, and mechanical vibration data. Through 8-bit quantization compression and TensorRT acceleration optimization, millisecond-level (<8ms) inference response speeds are achieved on edge computing devices such as JetsonXavier. A hybrid multimodal fusion strategy is used to... Strongly correlated data (such as voltage-temperature) undergoes early feature-level fusion to preserve underlying correlations, while heterogeneous data (such as vibration-logs) employs late decision-level fusion to adapt to differences in data characteristics. A risk-sensitive adaptive weighting mechanism (e.g., temperature risk coefficient 0.7, voltage risk coefficient 0.5) dynamically optimizes the fusion effect. A blockchain-based secure evidence storage mechanism is designed to store critical operations such as SOC resets and control commands on the blockchain, employing the national cryptographic SM3 hash algorithm and X.509 digital certificate authentication technology to ensure the immutability of operation records. Simultaneously, a four-layer encryption protection system is constructed, including X.509-based device identity authentication, SM4 transmission channel encryption, AES-256 local data encryption, and attribute-based ABE access control, forming a comprehensive security protection architecture.

[0334] Figure 9 This is a schematic diagram of the architecture of an energy storage data processing system based on full-modal standardization and edge intelligence, as provided in an embodiment of this application. Figure 9 As shown, the system includes:

[0335] The data acquisition and encrypted transmission module is used to acquire multi-source heterogeneous data from the energy storage system, standardize the multi-source heterogeneous data, generate a unified data stream, encrypt and anti-interference the unified data stream to generate encrypted data, and then transmit it to the edge node.

[0336] The edge real-time preprocessing module is used to decrypt the received encrypted data to obtain plaintext data, perform security verification on the plaintext data, filter out plaintext data that fails verification, and perform preprocessing and feature extraction on the plaintext data that passes verification to obtain preprocessed feature data.

[0337] The multimodal data fusion module is used to perform time alignment processing on the preprocessed feature data, and then mix and fuse the time-aligned preprocessed feature data to obtain cross-modal fused features.

[0338] The real-time analysis and control module is used to input cross-modal fusion features into a lightweight artificial intelligence model for real-time inference and output the failure risk probability; perform risk classification based on the failure risk probability and generate corresponding control commands; execute the control commands and provide feedback verification on the execution effect.

[0339] Furthermore, it also includes cloud collaboration and evidence storage, which is used to upload key data and events to the cloud and use blockchain technology to securely store key data and events; dynamically optimize lightweight artificial intelligence models based on cloud big data analysis results, and incrementally update the optimized models to edge nodes.

[0340] The execution process of the system part of this application embodiment is the same as that of the method part of the embodiment described above, and will not be repeated here.

[0341] This application also provides an electronic device, including: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute any of the new energy-based power grid multi-resource coordinated control methods.

[0342] This application also proposes a computer storage medium storing a computer program, which, when executed by a processor, implements any one of the new energy-based power grid multi-resource coordinated control methods.

[0343] Computer storage media may be simply referred to as media. Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Dual Data SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Rambus Direct RAM (RDRAM), Direct Memory Bus Dynamic RAM (DRDRAM), and Memory Bus Dynamic RAM (RDRAM). The various embodiments described in this specification are presented in a progressive manner, and similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, for embodiments of apparatus, devices, and non-volatile computer storage media, since they are substantially similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to the description of the method embodiments.

[0344] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this application.

Claims

1. A data processing method for energy storage based on full-modal standardization and edge intelligence, characterized in that, include: Collect multi-source heterogeneous data from the energy storage system, standardize the multi-source heterogeneous data to generate a unified data stream, and encrypt and anti-interference the unified data stream to generate encrypted data, which is then transmitted to the edge node. The edge node decrypts the received encrypted data to obtain plaintext data, and performs security verification on the plaintext data; Plaintext data that fails verification is filtered out, and plaintext data that passes verification is preprocessed and features are extracted to obtain preprocessed feature data. The preprocessed feature data is time-aligned, and the time-aligned preprocessed feature data is then mixed and fused to obtain cross-modal fused features; The cross-modal fusion features are input into a lightweight artificial intelligence model for real-time inference, and the failure risk probability is output. Risk is classified based on the probability of the fault risk, and corresponding control instructions are generated; the control instructions are executed, and the execution effect is verified by feedback.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data includes electrochemical parameters, thermal imaging data, vibration signals, and operation and maintenance logs.

3. The method according to claim 2, characterized in that, The multi-source heterogeneous data is standardized to generate a unified data stream, including: The protocol parsing function converts heterogeneous data from different industrial protocols into an intermediate representation. The intermediate representation is converted into a unified data stream using a standardized mapping function; Configure a buffer queue to manage the instantaneous peak values ​​of the unified data stream.

4. The method according to claim 3, characterized in that, Encryption and anti-interference processing of the unified data stream includes: Add device identifiers and timestamp metadata to the unified data stream and encode them in binary; The encoded data is encrypted using the SM4 algorithm; Differential transmission technology and cyclic redundancy check algorithm are introduced during data transmission.

5. The method according to claim 2, characterized in that, The verified plaintext data undergoes preprocessing and feature extraction, including: We use physical thresholding and statistical methods to filter out abnormal data and interpolate missing data to complete it. For structured data, time-domain and frequency-domain features are extracted; for unstructured data, lightweight convolutional networks and natural language processing techniques are used to extract visual and semantic features.

6. The method according to claim 5, characterized in that, For structured data, time-domain and frequency-domain features are extracted. For unstructured data, lightweight convolutional networks and natural language processing techniques are used to extract visual and semantic features, including: For the electrochemical parameters and the vibration signal, the mean and variance time-domain statistics are calculated using a sliding window, and the fundamental and harmonic frequency domain components are extracted using a fast Fourier transform. The thermal imaging data is processed by a lightweight convolutional neural network to extract hotspot coordinates and temperature gradient visual features. The operation and maintenance logs are processed using natural language processing techniques to extract key fault terms and convert them into structured text features.

7. The method according to claim 2, characterized in that, The preprocessed feature data is time-aligned, including: Using the electrochemical parameters sampled at high frequency as a time reference, the thermal imaging data sampled at low frequency is linearly interpolated and resampled. By combining the key event timestamps marked in the operation and maintenance logs, a spatiotemporal correlation matrix of abnormal events is established.

8. The method according to claim 2, characterized in that, The time-aligned preprocessed feature data is then blended and fused, including: Early feature-level fusion of electrochemical and thermal imaging features; Vibration characteristics and operation and maintenance log characteristics are fused at the late decision level; wherein, the late decision level fusion adopts a dynamic weighting strategy.

9. The method according to claim 1, characterized in that, The lightweight AI model is a quantized and compressed TinyViT model, and the end-to-end latency of the real-time inference is controlled within a preset threshold.

10. The method according to claim 1, characterized in that, Risk classification is performed based on the aforementioned failure risk probability, and corresponding control commands are generated, including: Multiple risk probability thresholds are set to classify risks into different levels; Different combinations of control commands and execution priorities are preset for different risk levels.

11. The method according to claim 1, characterized in that, Feedback verification of the execution results includes: Within the first preset time after the control command is issued, an inner loop verification is performed to check whether the execution result meets expectations. The outer loop is periodically evaluated using the second preset time period to analyze the long-term effectiveness of the control strategy. Based on the verification and evaluation results, the control parameters are automatically adjusted.

12. The method according to claim 1, characterized in that, Also includes: Key data and events are uploaded to the cloud and securely stored using blockchain technology. The lightweight artificial intelligence model is dynamically optimized based on the results of big data analysis in the cloud, and the optimized model is incrementally updated to the edge node.

13. The method according to claim 12, characterized in that, Securely storing the aforementioned key data and events using blockchain technology includes: The SM3 hash algorithm was used to generate digital fingerprints of the key data and events. The digital fingerprint, along with the timestamp and device information, is stored in a blockchain block to form an immutable evidence storage chain.

14. An energy storage data processing system based on full-modal standardization and edge intelligence, characterized in that, include: The data acquisition and encrypted transmission module is used to acquire multi-source heterogeneous data from the energy storage system, standardize the multi-source heterogeneous data to generate a unified data stream, encrypt and anti-interference the unified data stream to generate encrypted data, and then transmit it to the edge node. An edge real-time preprocessing module is used to decrypt the received encrypted data to obtain plaintext data and to perform security verification on the plaintext data; Plaintext data that fails verification is filtered out, and plaintext data that passes verification is preprocessed and features are extracted to obtain preprocessed feature data. The multimodal data fusion module is used to perform time alignment processing on the preprocessed feature data, and to mix and fuse the time-aligned preprocessed feature data to obtain cross-modal fusion features; The real-time analysis and control module is used to input the cross-modal fusion features into a lightweight artificial intelligence model for real-time inference and output the probability of fault risk. Risk is classified based on the probability of the fault risk, and corresponding control instructions are generated; the control instructions are executed, and the execution effect is verified by feedback.

15. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, implements the method of any one of claims 1-13.

16. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to perform the method described in any one of claims 1-13.

Images