A battery life prediction and health management system based on big data analysis

Abstract

The application relates to the technical field of battery health management, and specifically discloses a battery life prediction and health management system based on big data analysis, which comprises a hardware adaptation layer, an edge-cloud collaborative scheduling module and a core algorithm layer; the hardware adaptation layer is used for collecting multi-source data of a battery through a standardized interface, and performing protocol conversion on the multi-source data to generate a standardized data stream; the edge-cloud collaborative scheduling module is used for monitoring the resource availability state of an edge in real time, and dynamically adjusting a model updating strategy based on the resource availability state. Through the cooperation of the hardware adaptation layer and the core algorithm layer, the application realizes the standardization and deep fusion of multi-source data; the weight of a data credibility score dynamic adjustment mechanism model and a data-driven model is adjusted, the physical constraint ability of the mechanism model under low credibility is retained, the nonlinear fitting advantage of the data-driven model under high credibility is exerted, and the prediction robustness under all working conditions is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, specifically to a battery life prediction and health management system based on big data analysis. Background Technology

[0002] With the large-scale application of new energy vehicles and energy storage power stations, battery management systems (BMS) are placing higher demands on the prediction accuracy of battery state of health (SOH) and remaining lifetime (RUL). Traditional BMS mainly rely on fixed equivalent circuit models or electrochemical mechanism models. These model parameters are usually calibrated in laboratory environments and are difficult to adapt to the complex nonlinear aging characteristics throughout the battery's entire life cycle. While emerging pure data-driven methods have strong predictive capabilities, they consume huge amounts of computing and storage resources and heavily rely on massive amounts of high-quality tag data, making them difficult to deploy in real time on edge devices with limited computing power. Furthermore, their model generalization ability is weak, making it difficult to cope with changing real-world operating conditions.

[0003] More importantly, existing technologies face serious confidence conflicts and resource-security paradoxes in actual operation:

[0004] On the one hand, when a sensor experiences non-physical drift, such as zero-point shift, the collected data may still statistically exhibit a high signal-to-noise ratio and high integrity, causing the pure data-driven model to be misled by anomalous data with high statistical confidence. The system lacks a circuit breaker mechanism based on physical mechanism constraints to correct this statistical deception.

[0005] On the other hand, in order to alleviate the transmission and computing pressure at the edge, existing systems usually adopt a static feature filtering strategy. However, this is inherently contradictory to the weakness of early battery fault signals. That is, early critical anomalies are often hidden in data fluctuations that are judged to be of low importance. Static filtering will cause the system to blind faults in order to pursue efficiency, thus failing to capture dynamically evolving abnormal signals without increasing the computing power burden significantly. Summary of the Invention

[0006] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a battery life prediction and health management system based on big data analysis, to achieve accurate perception and closed-loop management of the battery's safety status throughout its entire lifecycle.

[0007] To achieve the above objectives, the first aspect of the present invention proposes a battery life prediction and health management system based on big data analysis, including a hardware adaptation layer, an edge-cloud collaborative scheduling module, and a core algorithm layer.

[0008] The hardware adaptation layer is used to collect multi-source data from the battery through a standardized interface, and to perform protocol conversion on the multi-source data to generate a standardized data stream.

[0009] The edge-cloud collaborative scheduling module is used to monitor the resource availability status of the edge in real time, dynamically adjust the model update strategy based on the resource availability status, and trigger the core algorithm layer to execute the prediction task.

[0010] The core algorithm layer is used to calculate battery health status indicators based on the standardized data stream through a dynamic collaborative prediction unit.

[0011] The operation of the dynamic collaborative prediction unit includes: calculating the data credibility score of the standardized data stream in real time; dynamically adjusting the weight allocation of the mechanism model layer and the data-driven model layer according to the data credibility score, and synthesizing the final battery life prediction result; and automatically triggering the model fine-tuning process based on transfer learning in response to the data credibility score being lower than a preset fine-tuning trigger threshold.

[0012] To achieve the above objectives, a second aspect of the present invention proposes a method for battery life prediction and health management based on big data analysis, comprising the following steps:

[0013] Multi-source data from the battery is collected through a standardized interface, and the multi-source data is converted according to a protocol to generate a standardized data stream;

[0014] Real-time monitoring of resource availability status at the edge, dynamic adjustment of model update strategy based on resource availability status, and triggering of battery life prediction task;

[0015] Based on the standardized data stream, the battery life prediction task is executed to calculate battery health status indicators.

[0016] The steps of performing the battery life prediction task to calculate battery health status indicators specifically include: calculating the data credibility score of the standardized data stream in real time; dynamically adjusting the weight allocation of the mechanism model layer and the data-driven model layer according to the data credibility score to synthesize the final battery life prediction result; and automatically triggering the model fine-tuning process based on transfer learning in response to the data credibility score being lower than a preset fine-tuning trigger threshold.

[0017] To achieve the above objectives, a third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for battery life prediction and health management based on big data analysis.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] The battery life prediction and health management system based on big data analysis in this invention achieves standardization and deep integration of multi-source data through the collaboration of the hardware adaptation layer and the core algorithm layer; it dynamically adjusts the weights of the mechanism model and the data-driven model by using data credibility scores, which not only retains the physical constraint capability of the mechanism model under low credibility, but also gives full play to the nonlinear fitting advantage of the data-driven model under high credibility, thus significantly improving the prediction robustness under all operating conditions.

[0020] In particular, this invention introduces a secondary monitoring and physical consistency verification mechanism, which effectively resolves the contradiction between static feature reduction and dynamic abnormal evolution. Without increasing the computing power burden at the edge, it captures early faults that are being masked through a bypass wake-up mechanism. At the same time, it establishes a veto power of physical laws over statistical scoring, preventing model misjudgments caused by sensor drift. Thus, while ensuring computational efficiency, it achieves accurate perception and closed-loop management of the safety status of the battery throughout its entire life cycle. Attached Figure Description

[0021] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0022] Figure 1 This is a schematic diagram illustrating the implementation of the battery life prediction and health management system based on big data analysis provided by the present invention.

[0023] Figure 2 This is a comparison curve of temperature sensor data before and after two-dimensional deviation correction in the battery life prediction and health management system based on big data analysis provided by this invention;

[0024] Figure 3 This is a comparison chart of the prediction accuracy of the global teacher model and the lightweight student model after model distillation in the battery life prediction and health management system based on big data analysis provided by this invention.

[0025] Figure 4 This is a schematic diagram illustrating the dynamic evolution of data credibility score and mechanism / data-driven model weights in the battery life prediction and health management system based on big data analysis provided by this invention.

[0026] Figure 5 This is a comparison chart of the SOH index calculated by the system in the battery life prediction and health management system based on big data analysis provided by this invention and the prediction results of the traditional voltage and current method;

[0027] Figure 6This is a simulation diagram of the secondary monitoring subunit capturing early minor fault fluctuations and priority transitions in the battery life prediction and health management system based on big data analysis provided by this invention.

[0028] Figure 7 This is a flowchart illustrating the battery life prediction and health management method based on big data analysis provided by the present invention.

[0029] Figure 8 This is a schematic diagram of the electronic device provided by the present invention. Detailed Implementation

[0030] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0031] The following description, with reference to the accompanying drawings, describes a battery life prediction and health management system, method, and electronic device based on big data analysis, according to embodiments of the present invention.

[0032] Example 1:

[0033] This embodiment discloses a battery life prediction and health management system based on big data analysis. The system achieves closed-loop management across the entire chain, from underlying physical sensing to upper-level intelligent decision-making, by constructing a layered architecture comprising a hardware adaptation layer, an edge-cloud collaborative scheduling module, and a core algorithm layer.

[0034] like Figure 1 As shown, the system in this embodiment includes the following:

[0035] Specifically, the system described in this embodiment mainly consists of three core parts in its logical architecture: a hardware adaptation layer, an edge-cloud collaborative scheduling module, and a core algorithm layer. These three parts do not exist in isolation, but are tightly coupled through standardized data and control flows. Among them, the hardware adaptation layer, as a bridge connecting the physical world and the digital world, undertakes extremely critical foundational work.

[0036] For example, the hardware adaptation layer's function is not limited to simple data transmission; it is used to collect multi-source data from the battery through standardized interfaces and perform protocol conversion on this multi-source data to generate a standardized data stream. In practical applications, battery systems are often assembled from cells, sensors, and battery management systems (BMS) from different manufacturers. These hardware devices typically use different communication protocols, such as CAN, Modbus, and MQTT, and different data formats. To address this heterogeneity issue, the hardware adaptation layer incorporates a protocol parsing engine and a format conversion unit. When different types of hardware devices are connected, the hardware adaptation layer can automatically identify their communication protocols and convert them into a standardized format common to the system. This standardized interface design greatly improves the system's compatibility and scalability, allowing subsequent core algorithm layers to focus solely on data analysis without needing to consider the differences in the underlying hardware.

[0037] Specifically, the multi-source data collected by the hardware adaptation layer is not a single-dimensional voltage or current data, but rather covers three dimensions of data that can comprehensively reflect the battery state, including internal chemical state data collected by micro-sensors, external environmental data, and traditional operating data.

[0038] It is also worth noting that, in order to delve into the microscopic mechanisms of battery aging, this embodiment incorporates microsensors embedded within the battery. These microsensors can penetrate deep into the electrolyte or electrode surface to directly collect key indicators reflecting the rate of chemical reactions and the state of mass transport within the battery—the aforementioned internal chemical state data. Simultaneously, considering the significant impact of environmental factors on battery performance, the hardware adaptation layer also collects external environmental data, such as environmental vibration frequency, humidity, and pollutant concentrations, through environmental sensors deployed around the battery pack. Furthermore, traditional operating data such as voltage, current, and temperature collected by the battery management system (BMS) are also an important component of the multi-source data.

[0039] Specifically, after collecting the above three types of data, the hardware adaptation layer performs a data fusion process. Since different sensors often have different sampling frequencies—for example, voltage data might be sampled at the millisecond level, while ambient humidity data might be sampled at the second level—direct data fusion would lead to severe timing misalignment. This embodiment employs timestamp alignment technology to achieve time-dimension consistency fusion of the internal chemical state data, the external environmental data, and the traditional operational data. This process includes establishing a unified, high-precision system clock and assigning a precise timestamp to each collected data packet. For data with low sampling frequencies, the system uses linear interpolation or spline interpolation algorithms to fill in missing values ​​at high-frequency time points; for data with excessively high sampling frequencies, the system uses moving average or downsampling processing to ultimately ensure that all dimensions of data are strictly aligned on the time axis, forming a multi-dimensional time series matrix.

[0040] Optionally, to further improve data accuracy and eliminate interference from sensor errors, the hardware adaptation layer also constructs a two-dimensional deviation correction model. This model fully considers the coupled influence of battery state and environmental factors on sensor readings. Specifically, the model uses state-of-charge parameters, ambient temperature parameters, and electrolyte concentration parameters to calculate the temperature sensor deviation value, and uses the temperature sensor deviation value to perform online calibration of the original measurement data.

[0041] Specifically, the mathematical expression logic of the two-dimensional deviation correction model is as follows: The system defines the temperature sensor deviation value as... This deviation value is not a fixed constant, but rather a function that dynamically changes with the battery's operating conditions. In this embodiment, The calculation relies on a linear or nonlinear combination of three key variables: first, the state of charge parameter, usually denoted as SOC, which reflects the current remaining capacity of the battery; second, the ambient temperature parameter, denoted as... It reflects the baseline state of the external thermal field; finally, there is the electrolyte concentration parameter, denoted as... This reflects the internal chemical environment of the battery. Through extensive experimental calibration and online identification, the system determined the mapping relationship between these three parameters and temperature deviation.

[0042] For example, in a simplified linear model, the temperature sensor deviation value It can be represented as:

[0043] ;

[0044] Where 'a' represents the correction coefficient for the state of charge (SOC) on temperature measurement, 'b' represents the influence coefficient of ambient temperature on temperature measurement, and 'c' represents the coupling coefficient of electrolyte concentration on temperature measurement. During real-time operation, the hardware adaptation layer will integrate the real-time collected SOC data... and Substituting into the above model, calculate the current... The deviation value is then subtracted from or added to the original acquired temperature value to obtain the calibrated true temperature data. This online calibration mechanism significantly reduces measurement errors caused by sensor drift or environmental interference, laying a solid data foundation for subsequent lifetime prediction.

[0045] like Figure 2 This demonstrates the significant advantages of the two-dimensional bias correction model in handling sensor measurement errors. In the figure, the horizontal axis represents the battery's continuous operating time in seconds, and the vertical axis represents the battery's real-time temperature in degrees Celsius.

[0046] Figure 2 The original measured temperature curve, presented as a red dashed line, exhibits a significant numerical rise and nonlinear drift during operation. This fully verifies that the sensor is subject to measurement deviations due to coupling interference from factors such as state of charge, ambient temperature, and electrolyte concentration under actual operating conditions, resulting in measured values ​​that are significantly higher than the actual physical temperature of the battery.

[0047] Figure 2 The actual reference temperature curve, represented by black dotted lines, serves as a benchmark, representing the expected temperature state of the battery under its current state of charge and heat generation rate. Figure 2 The corrected temperature curve, presented as a solid blue line, uses a two-dimensional deviation correction model built into the hardware adaptation layer to perform real-time online calibration of the original data. Its trajectory achieves a high degree of fit with the real reference temperature curve, successfully filtering out spurious temperature fluctuations introduced by sensor hardware limitations or changes in the external electrochemical environment.

[0048] This comparison demonstrates that by constructing a correction function containing multiple key operating parameters, the present invention can effectively eliminate sensor zero-point offset and physical drift, ensuring the physical authenticity of the generated standardized data stream, thereby laying a solid and reliable data foundation for the core algorithm layer to calculate high-precision battery health status indicators.

[0049] After high-quality data collection and preprocessing are completed, the data stream will enter the processing and analysis stage. Considering that the computing resources (CPU, memory, bandwidth) of vehicle-mounted edge devices or energy storage station edge devices are extremely limited and cannot run large-scale deep learning models for a long time, this embodiment introduces an edge-cloud collaborative scheduling module.

[0050] Specifically, the edge-cloud collaborative scheduling module is used to monitor the resource availability status of the edge devices in real time, dynamically adjust the model update strategy based on the resource availability status, and trigger the core algorithm layer to execute prediction tasks. This module integrates a resource monitor that can poll the CPU utilization, remaining memory space, and current network uplink / downlink bandwidth of edge devices at a frequency of seconds.

[0051] For example, resource availability can be assessed using a comprehensive scoring model that weights and combines CPU idle rate, memory idle rate, and network quality into a score between 0 and 1. When resource availability is high, such as when the vehicle is stationary charging and connected to a Wi-Fi network, the scheduling module will adopt a more aggressive model update strategy, allowing large-scale model parameter downloads or complex gradient uploads. When resource availability is low, such as when the vehicle is traveling at high speed, processing power is strained, and the network is unstable, the scheduling module will switch to a conservative strategy, allowing only a very small number of key features to be transmitted or pausing model updates, prioritizing real-time prediction tasks for the core algorithm layer.

[0052] Specifically, the model update strategy employs advanced federated learning and model distillation mechanisms, aiming to achieve continuous evolution of model capabilities while protecting user privacy data by keeping it locally. The edge-cloud collaborative scheduling module is configured to perform the following operations:

[0053] First, the control edge trains a lightweight local model based on local data and generates knowledge representations of intermediate layer feature maps. Here, local data refers to historical operational data and real-time collected data stored on the edge device. Because this data contains sensitive privacy information such as user driving habits and location data, it is not directly uploaded to the cloud. The edge device utilizes local computing power to train or fine-tune a lightweight neural network, such as a low-parameter LSTM or CNN, for several rounds. After training, the system does not directly upload the original model weights but extracts the output of the model's intermediate layers (Feature Maps) as knowledge representations. These knowledge representations contain the model's high-dimensional understanding of data features but cannot directly deduce the original data.

[0054] Secondly, to further enhance security, the knowledge representation is uploaded to the cloud after undergoing homomorphic encryption. The application of homomorphic encryption allows the cloud server to directly perform aggregation operations, such as addition or multiplication, on data uploaded from different edge devices in encrypted form, without decryption. This means that even if the cloud server is compromised, attackers cannot obtain any valuable plaintext information.

[0055] Finally, the cloud aggregates all knowledge representations uploaded from edge devices to train a global teacher model, and generates a lightweight student model through model distillation, which is then distributed to the edge devices. In the cloud, the system gathers tens of thousands of encrypted knowledge representations uploaded from edge devices, leveraging this rich collective intelligence to train a massive, highly accurate global teacher model. However, this teacher model is too large to be directly deployed on the edge. Therefore, the cloud employs knowledge distillation technology, allowing the massive teacher model to teach a streamlined student model, enabling the student model to mimic the teacher model's output behavior as closely as possible with minimal parameters. Ultimately, this lightweight student model generated through knowledge distillation is distributed to various edge devices via over-the-air (OTA) download technology, replacing or updating the original local models, thus achieving a co-evolutionary model between the cloud and the edge.

[0056] like Figure 3 The results demonstrate the consistency in prediction accuracy between the global teacher model and the lightweight student model after knowledge distillation. Figure 3 The horizontal axis represents the number of battery cycles, showing the long-term process of the battery from its initial state to a thousand cycles, while the vertical axis represents the percentage of battery health.

[0057] Figure 3 The solid black line in the graph clearly shows the non-linear degradation trend of the battery over time, with the value gradually decreasing from 100% to approximately 80%. The dashed red line in the graph represents the global teacher model prediction curve, a highly complex model trained by aggregating knowledge representations from various edge devices on a cloud server. Its prediction trajectory almost perfectly matches the actual health state curve, demonstrating the teacher model's superior fitting ability when processing massive amounts of multi-source data.

[0058] Figure 3 The lightweight student model prediction curve, represented by the blue dotted line, is an optimized model generated in the cloud and distributed to the edge using model distillation technology. It aims to mimic the output behavior of the teacher model with a very small parameter scale.

[0059] Observing the waveform transformation in the figure reveals that although the lightweight student model has undergone structural simplification to adapt to the limited computing power at the edge, its predicted trajectory still accurately follows the fluctuation characteristics of the teacher model. The prediction deviation between the two is strictly controlled within a very small range, and there is no significant collapse in accuracy due to model compression. This result proves that the model distillation mechanism described in this invention can map the generalization features of the global model in the cloud to the edge node, achieving accurate perception and closed-loop management of the battery's health status throughout its entire life cycle without increasing the computing power burden at the edge.

[0060] When the standardized data stream is transmitted to the core algorithm layer, it first needs to be processed by the feature engineering unit to extract the most valuable features for battery life prediction. However, facing massive amounts of high-dimensional data, if the features are not filtered and compressed, it will impose a huge computational burden on the subsequent prediction model. To address this, this embodiment designs a feature processing mechanism with tiered computing power.

[0061] Specifically, the core algorithm layer also includes a feature engineering unit, which performs feature processing according to computing power levels. This process is not a one-size-fits-all approach, but rather tailored to the actual capabilities of the hardware environment.

[0062] First, this unit calculates the feature importance score for each feature dimension in the standardized data stream using a self-supervised contrastive learning module. Self-supervised contrastive learning is an advanced deep learning method that does not require manual labeling. The system trains a feature extraction network by constructing positive sample pairs (such as different augmented views of data from the same time point) and negative sample pairs (such as data from different time points), enabling it to identify the features that are most sensitive and discriminative as the battery's aging state changes. After training, this module outputs a value between 0 and 1 for each input feature, such as voltage variance, temperature gradient, and ion concentration change rate, i.e., the feature importance score. The higher the score, the greater the contribution of that feature to the prediction of SOH or RUL.

[0063] It is also important to note that the feature engineering unit obtains the current computing power level of the edge controller in real time and selects the corresponding feature compression scheme based on the current computing power level. The system presets three levels: low computing power, medium computing power, and high computing power.

[0064] Specifically, if the feature importance score is greater than or equal to a first preset threshold, such as 0.6, it indicates that the feature is a core, key feature, and its most complete information must be retained. In this case, the system uses a high-fidelity lossless compression scheme to process the feature. Although this scheme has a relatively low compression rate, it ensures that the decompressed data is substantially consistent with the original data, without losing any minute details, thus guaranteeing the usability of high-value features.

[0065] Optionally, if the feature importance score is between the first preset threshold and the second preset threshold, such as between 0.3 and 0.6, it indicates that the feature has certain reference value, but a small loss of accuracy is allowed in exchange for a higher compression ratio. In this case, the system uses an autoencoder to process the feature using a standard compression scheme. An autoencoder is a neural network that attempts to compress the input into a low-dimensional latent vector (Encoder process) and then reconstruct the input from that vector (Decoder process). In this mode, feature data is mapped to a low-dimensional space for transmission and storage. Although there will be small errors during reconstruction, they are acceptable in scenarios of moderate importance.

[0066] Specifically, if the feature importance score is less than the second preset threshold, such as 0.3, it indicates that the feature is mainly redundant information or noise, and its contribution to the model prediction is negligible. To save resources to the greatest extent, the system marks this feature as a feature to be filtered and performs a filtering removal operation. This means that these data will not enter the subsequent prediction model and will be discarded directly in the feature engineering stage, thereby greatly reducing data throughput.

[0067] The carefully selected feature data, after feature engineering, is finally entered into the dynamic collaborative prediction unit.

[0068] Specifically, the core algorithm layer is used to calculate battery health status indicators based on the standardized data stream through a dynamic collaborative prediction unit. The operation of the dynamic collaborative prediction unit includes the following rigorous logical steps:

[0069] The first step is to calculate the data credibility score of the standardized data stream in real time. In the field of data analysis, the quality of data determines the upper limit of model prediction. To quantify the quality of the current input data, this system introduces multi-dimensional evaluation metrics. Specifically, the process of generating the data credibility score includes: calculating the data integrity index, signal-to-noise ratio index, and inter-domain KL divergence index of the standardized data stream, respectively.

[0070] Among these metrics, data integrity reflects whether there is packet loss or interruption. If the proportion of consecutive missing data points exceeds a certain limit within a certain period, this metric will significantly decrease. Signal-to-noise ratio (SNR) reflects the ratio of effective signal to background noise. High-frequency electromagnetic interference can reduce the SNR, thus affecting data purity. The Kullback-Leibler-Divergence metric measures the difference between the distribution of the currently collected data and the distribution of the source domain data used during model training. If the battery operates under extreme conditions, such as extremely cold weather or ultra-fast charging and discharging, its data distribution may differ significantly from the training data, i.e., domain drift occurs. In this case, the KL divergence increases, and the corresponding reliability metric should decrease.

[0071] After obtaining these three sub-indicators, the system performs a weighted summation of the data integrity index, the signal-to-noise ratio index, and the inter-domain KL divergence index to obtain the final data credibility score. This score is a scalar between 0 and 1, which comprehensively reflects whether the current data is trustworthy.

[0072] The second step involves dynamically adjusting the weight distribution of the mechanistic model layer and the data-driven model layer based on the data credibility score, and synthesizing the final battery life prediction result. The prediction model in this embodiment is a two-layer hybrid architecture: one layer is a mechanistic model layer based on electrochemical equations or equivalent circuits, which follows strict physical and chemical laws. Although its accuracy may be limited by parameter identification, it effectively reduces the probability of outputting prediction results that violate physical laws, such as predicting a sudden increase in battery capacity; the other layer is a data-driven model layer based on deep neural networks, which makes predictions by learning nonlinear mapping relationships in massive amounts of data. It has high accuracy but is easily misled by abnormal data.

[0073] Specifically, the dynamic adjustment of the weight allocation between the mechanistic model layer and the data-driven model layer includes the following: When the data confidence score is greater than or equal to a preset high confidence threshold, such as 0.8, it means that the current data quality is extremely high and the operating conditions are consistent with the model's understanding. In this case, the system tends to trust the more accurate data-driven model, thus reducing the weight of the mechanistic model layer and increasing the weight of the data-driven model layer. For example, the weight of the data-driven model can be adjusted to a relatively high level, such as 0.7, while the weight of the mechanistic model can be adjusted to a relatively low level, such as 0.3. Conversely, when the data confidence score is low, it indicates poor data quality or abnormal operating conditions. In this case, the data-driven model may give an outrageous prediction, and the system will automatically increase the weight of the mechanistic model layer, using the constraints of physical laws to provide a safety net and ensure the security of the prediction results.

[0074] like Figure 4It demonstrates how the dynamic collaborative prediction unit intervenes in the model decision-making process in real time based on data quality, fundamentally solving the conflict between statistical high confidence and physical inconsistency.

[0075] Figure 4 The horizontal axis represents the operating time cycle of the battery system, and the vertical axis represents the index score and weight coefficient.

[0076] The data credibility score curve depicted by the solid green line in the figure reflects the real-time trustworthiness of the standardized data stream generated by the hardware adaptation layer, which comprehensively evaluates indicators such as data integrity, signal-to-noise ratio, and inter-domain divergence.

[0077] In the initial stage of operation, when the green curve is in the high confidence range above 0.8, the weight of the data-driven model (represented by the blue dashed line) is configured to a high proportion of 0.7, while the weight of the mechanistic model layer (represented by the red dotted line) remains at an auxiliary level of 0.3. At this time, the system mainly uses the deep learning model for nonlinear fitting. As the operating environment changes or interference increases, the waveform in the figure shows a significant decrease in the confidence score, entering the medium range of 0.3 to 0.8. At this point, the system automatically completes the weight switching, increasing the weight of the mechanistic model layer to 0.6 to introduce physical law constraints.

[0078] In the later stages of operation, an extreme scenario was simulated where the data reliability score dropped below 0.3. At this point, the system identified a physical violation and triggered a forced overwrite operation. The red dotted line rapidly climbed and locked at an absolute dominant score of 0.9, while the weight of the data-driven model was suppressed to an extremely low level of 0.1. This safety margin, based on the physical laws of the mechanistic model, prevented model misjudgments caused by sensor drift. This dynamic evolution of the weights clearly demonstrates the veto power of the physical mechanism established in this invention over the statistical score, ensuring the robustness of lifespan prediction throughout the battery's entire lifespan.

[0079] Third, in response to the data confidence score falling below a preset fine-tuning trigger threshold, such as 0.3, it indicates that the current model can no longer adapt to the new data distribution or operating conditions, and simply adjusting the weights is insufficient to solve the problem. At this point, the system automatically triggers a model fine-tuning process based on transfer learning. This process uses a small amount of newly collected data, i.e., target domain data, to quickly update the parameters of the data-driven model, enabling it to adapt to the new data distribution and thereby improve confidence in subsequent predictions.

[0080] In the aforementioned dynamic prediction architecture, a key output is the State of Health (SOH) index. This embodiment presents an innovative calculation method based on microscopic sensor data.

[0081] Specifically, the logical definition of the core algorithm layer for calculating the battery health status index is as follows: First, obtain the data of each of the microsensors in... The difference between the real-time electrolyte concentration collected at each step and the initial electrolyte concentration. We define the first... A miniature sensor in The real-time electrolyte concentration collected at various times is The initial electrolyte concentration of the battery at the time of manufacture is The difference between the two It directly reflects the degree of electrolyte consumption or deterioration and is direct chemical evidence of battery aging.

[0082] Then, multiply the difference by a preset electrolyte concentration weighting coefficient to obtain the first product, whereby the weighting coefficient is defined as... Because the ability of microsensors at different locations to characterize the overall SOH varies, The value of was determined through fitting a large amount of experimental data. Therefore, the first product can be expressed as Next, the ion migration velocity collected by each of the microsensors is acquired, and this ion migration velocity is multiplied by a preset ion migration velocity weighting coefficient to obtain a second product. Define the first... The ion migration rate of each sensor is , For the preset first The weighting coefficients for the ion migration rates of each microsensor characterize the impact of ion migration characteristics at that location on the overall health of the battery. Ion migration rate reflects the battery's internal kinetics; slower rates result in higher internal resistance and lower state of equilibrium (SOH). Therefore, the second product is expressed as... .

[0083] Finally, the first product and the second product corresponding to all the microsensors are summed to obtain the microscopic electrochemical health factor. Assume the total number of microsensors is... The complete formula for calculating microscopic electrochemical health factors is:

[0084] ;

[0085] Furthermore, the core algorithm layer is also configured to acquire the external operating characteristics of the battery. The external operating characteristics specifically cover a multi-dimensional feature set reflecting the macroscopic degradation performance of the battery, including but not limited to: the ampere-hour throughput and equivalent full cycle count calculated based on historical operating data, the characteristic peak area and position offset extracted by incremental capacity analysis (ICA) or differential voltage analysis (DVA) using the current charge-discharge curve, and the macroscopic internal resistance growth rate identified online by Ohm's law.

[0086] The dynamic collaborative prediction unit combines the microscopic electrochemical health factors and the external operating characteristics to comprehensively evaluate the final battery state of health (SOH) index. Specifically, since battery health is a comprehensive evaluation system encompassing capacity decay, power reduction, and safety degradation, simple microscopic mechanism deduction or macroscopic data statistics have limitations. This system uses the microscopic electrochemical health factors obtained by summing the above-mentioned factors as a physical prior to characterize internal degradation, and deeply cascades and integrates them with high-dimensional external operating characteristics. This process ensures that the final output battery health index not only anchors to the fundamental physicochemical causes of battery aging, such as electrolyte deterioration and ion transport obstruction, but also fully accommodates the macroscopic evolution laws of capacity and internal resistance exhibited externally. This fundamentally overcomes the one-sidedness of using a single microscopic parameter to characterize the overall SOH, significantly improving the universality and accuracy of health status assessment.

[0087] Through the above formula, this system quantifies the invisible and intangible microscopic chemical changes into specific numerical indicators. Compared with the traditional method of estimating SOH based solely on voltage and current, its accuracy and physical interpretability have achieved a qualitative leap.

[0088] like Figure 5 The performance comparison between the prediction algorithm based on microscopic sensor data used in this system and traditional estimation methods is shown. Figure 5 The horizontal axis represents the number of battery cycles, showing the aging process of the battery from its initial state to one thousand cycles, while the vertical axis represents the percentage of the battery's health status.

[0089] Figure 5 The true reference value, represented by the solid black line, represents the actual degradation trajectory of the battery, which exhibits a non-linear characteristic of accelerating decline with increasing cycle count. The traditional voltage-current method prediction results, represented by the dashed red line, show severe waveform oscillations and obvious systematic biases. After 500 cycles, its prediction error gradually amplifies, verifying the limitation of traditional methods that rely solely on external macroscopic electrical parameters and are unable to capture complex aging characteristics.

[0090] In comparison, Figure 5 The predicted values ​​of this system, represented by the solid blue line, closely match the actual reference values. The curve is smooth and accurately pinpoints the evolution of the actual health status. This significant improvement in accuracy is attributed to the core algorithm layer of this invention, which directly acquires key internal chemical state data such as changes in electrolyte concentration and ion migration rate through micro-sensors. By multiplying the internal chemical state data by weighting coefficients and summing them, the microscopic physicochemical changes are quantified into specific health status indicators.

[0091] By observing the specific values ​​in the graph, it can be seen that at the critical point of approximately 800 cycles, the predicted value of the traditional method drops below 80%, while the predicted value of this system remains near the true level of 85%. This comparison graph demonstrates the accuracy of this invention in perceiving battery health status throughout its entire life cycle, overcoming the shortcomings of traditional methods such as high computational consumption and weak generalization ability, and achieving high-precision prediction and reliable management of battery life.

[0092] After obtaining accurate SOH prediction results, this system further intervenes and optimizes the operation of the battery through model predictive control (MPC) and graded anomaly response mechanism.

[0093] Specifically, the core algorithm layer also integrates a model predictive control multi-objective optimization module, used to construct a cost function that includes a temperature tracking error term, an actuator energy consumption term, a health state decay rate term, and a control quantity change term. The core idea of ​​MPC is to solve for the optimal control sequence in a finite time domain within each control cycle, thereby minimizing the predefined cost function.

[0094] The cost function defined in this embodiment The optimization objectives cover four dimensions:

[0095] The first term is the temperature tracking error term, denoted as... It measures the deviation between the battery's actual temperature and its optimal operating temperature range. The smaller the deviation, the better the battery performance.

[0096] The second item is the actuator energy consumption item, denoted as... It measures the energy consumption of cooling systems such as fans, water pumps, or heaters. Energy consumption should be reduced as much as possible while meeting temperature control requirements.

[0097] The third term is the rate of decline in health status, denoted as... It predicts the amount of SOH decrease over a future period using a model. Control strategies should aim to minimize this decrease, i.e., extend battery life.

[0098] The fourth item is the change in control quantity, denoted as It limits the intensity of control actions, such as the rate of change of current, to protect the actuator.

[0099] In summary, the total cost value This can be expressed as a weighted sum of the individual components:

[0100] ;

[0101] in, These are the weighting coefficients for each item, which are adjusted according to the current operating strategy, such as long lifespan mode or high performance mode.

[0102] It should also be noted that the model predictive control multi-objective optimization module, based on the cost function, solves for the optimal control strategy to minimize the total cost value of the cost function, provided that the battery temperature state constraints and charge / discharge input constraints are satisfied. The state constraints ensure that the battery temperature is always within a safe range, such as 0°C to 55°C; the input constraints ensure that the charging current and depth of discharge do not exceed physical limits, such as the maximum charging rate not exceeding 1C.

[0103] In addition to conventional optimization and control, the system must also be able to handle unexpected situations. Specifically, the core algorithm layer also includes a multi-scale anomaly response unit for executing a graded response process. This unit monitors key indicators such as the rate of change of battery health status, the amount of change in electrolyte concentration, and the rate of increase in internal resistance in real time. Once an anomaly is detected, graded handling is immediately initiated.

[0104] For example, if an anomaly is identified as Level 1, such as a slight non-linear decline in State of Harm (SOH), it indicates a potential risk that has not yet jeopardized safety. In this case, the system control edge adjusts the charging rate: for example, reducing it to 0.5C slow charging, and only uploading enhanced sensitive feature data for further analysis in the cloud, while avoiding excessive bandwidth consumption.

[0105] For example, if a level 2 anomaly is identified, such as a sudden change in electrolyte concentration, it indicates that a minor side reaction may have occurred inside the battery. In this case, the system control edge uploads compressed, complete data, including detailed waveform records, and triggers the cloud to generate maintenance recommendations, such as suggesting that the user conduct an on-site inspection within the next 72 hours.

[0106] For example, if a level three anomaly is identified, such as a sharp rise in temperature or a sudden increase in internal resistance, it indicates a possible precursor to an internal short circuit or thermal runaway. In this case, safety is paramount. The system control edge immediately stops charging and discharging operations, disconnects the main circuit relay, triggers audible and visual warnings, and simultaneously sends a high-priority alarm signal to the cloud-based operations and maintenance center.

[0107] In summary, this embodiment constructs a complete, rigorous, and efficient battery health management system. This system not only solves the challenges of multi-source data fusion and edge computing, but also achieves high-precision prediction and safety management of battery life through deep integration of mechanisms and data, demonstrating significant technical advantages and application value.

[0108] Example 2:

[0109] This embodiment, based on Embodiment 1, further refines the feature engineering unit in the core algorithm layer. This embodiment focuses on addressing and resolving potential conflicts between static feature selection strategies and dynamic fault evolution characteristics during long-cycle battery management.

[0110] In Example 1, the system categorizes features based on their importance scores, with low-importance features being directly filtered out. This strategy significantly saves computing power and bandwidth resources under steady-state conditions. However, throughout the battery's lifespan, some features that normally behave stably and are considered redundant may exhibit sudden abnormal fluctuations in specific early fault stages, such as the initial stages of micro-short circuits or sensor zero-point drift. If the system mechanically executes the filtering strategy, these critical fault precursor signals will be discarded, creating blind spots in safety monitoring. To resolve this conflict between static feature reduction and dynamic abnormal evolution, this embodiment introduces a secondary monitoring subunit within the feature engineering unit.

[0111] Specifically, the feature engineering unit also includes a secondary monitoring subunit. This subunit is not intended to replace the original main feature processing flow, but rather to run in parallel as an independent, low-power daemon process. Its primary task is to establish a bypass monitoring channel to take over the data marked as features to be filtered by the feature engineering unit.

[0112] For example, in the feature processing flow described in Embodiment 1, when the feature importance score calculated by the self-supervised contrastive learning module is less than a second preset threshold, such as 0.3, the corresponding feature data would normally be labeled as discarded and directly blocked from subsequent processing. However, in this embodiment, these discarded data streams, which are determined to be of low value, do not actually disappear, but are redirected to a bypass monitoring channel managed by a secondary monitoring subunit.

[0113] It is also important to note that the bypass monitoring channel is designed with minimalism in mind to ensure that its operation does not significantly burden the computing power of the edge controller. This channel does not deploy any complex neural network models or high-order statistical models; it exists solely as a first-in, first-out (FIFO) data buffer. Data resides in this channel for an extremely short time, used only for a single, instantaneous morphological scan. This design ensures that while maintaining the high performance advantages described in Example 1, the system adds a safety fallback mechanism, implementing cache degradation verification for filtered data.

[0114] Specifically, the secondary monitoring subunit performs a linear scan of the data in the bypass monitoring channel and calculates the instantaneous fluctuation index. Here, "linear scan" refers to an algorithm with a time complexity strictly controlled at the O(M) level, where M is the number of data points. Unlike deep learning algorithms in the main channel that involve matrix operations and nonlinear activation functions, linear scan only involves basic addition, subtraction, multiplication, and division operations, and can complete the calculation using the basic computing power of the edge MCU (microcontroller unit) or even idle periods during DMA (direct memory access) transfers.

[0115] For example, to quantify the degree of sudden anomalies in data, this embodiment defines an index with clear physical and statistical significance: the transient volatility index. This index aims to capture outliers that deviate from the normal distribution. The transient volatility index is defined as the ratio of the absolute value of the difference between the current sample value and the historical sliding window mean to the historical sliding window standard deviation.

[0116] Specifically, this embodiment provides the following rigorous mathematical definition for the above calculation logic:

[0117] Suppose a certain feature dimension in the bypass monitoring channel is at the current time. The sampling point values ​​are The system maintains a length of A history sliding window that contains data from time [time]. At the time Historical data. Based on this historical sliding window, the system maintains and updates two statistics in real time: the historical sliding window mean. and historical sliding window standard deviation Then this feature at time... Instantaneous fluctuation index The calculation formula can be expressed as:

[0118] ;

[0119] in, This represents the degree of deviation of the current data point from the historical average level, i.e., the absolute deviation; To prevent extremely small positive numbers with a denominator of zero; and the denominator This, acting as a normalization factor, eliminates the influence of different physical dimensions, such as voltage in volts and temperature in degrees Celsius, and their inherent fluctuations. This means that regardless of the original baseline value of the characteristic, the instantaneous fluctuation index... It is always a dimensionless scalar that measures the degree of anomaly. From a statistical perspective, this index actually calculates the absolute value of the Z-score (standard score) of the current data point. If the data follows a normal distribution, this index directly reflects the probability that the current data point appears at the tail of the probability distribution.

[0120] It is also important to note that in practical engineering implementation, to avoid redundant calculations caused by recalculating the mean and standard deviation every time a sample is taken, this system employs a recursive update algorithm. That is, utilizing... Time-based statistics and newly entered data points and the removed data points Quickly updated using incremental calculation formula and This optimization further reduces the computing power consumption of the secondary monitoring subunit, enabling it to complete concurrent scans of dozens or hundreds of low-importance features within microseconds.

[0121] After calculating the instantaneous fluctuation index, the system enters the logical decision stage. Specifically, the secondary monitoring subunit compares the instantaneous fluctuation index with a preset wake-up threshold. The preset wake-up threshold is a constant set according to the system's sensitivity requirements for anomalies, usually denoted as [missing information]. According to the 3-Sigma principle in statistics, under a normal distribution, the probability of a value falling outside the mean plus or minus three standard deviations is only 0.27%. Therefore, this threshold... It is usually set to a value between 3 and 5.

[0122] For example, if the calculated Less than or equal to the wake-up threshold The system determines that the current fluctuation is a normal noise disturbance, and this feature is still judged to be of low importance. The data is then covered or discarded in the bypass monitoring channel without occupying subsequent resources.

[0123] Specifically, if the instantaneous fluctuation index is greater than the wake-up threshold, it indicates that the feature has undergone a statistically significant mutation, meaning that a previously considered silent redundant feature has suddenly become active. This mutation is highly likely a signal of an early fault. For example, an edge-assist sensor that has been operating at a constant temperature may suddenly experience a 5-degree Celsius jump. Although its absolute value still does not reach the high-temperature alarm line, its… The value may have already reached 10 or higher. At this point, the secondary monitoring unit will determine that the feature is no longer suitable for filtering, thereby triggering the priority reset process.

[0124] The priority reset process is the core means of resolving conflict-related technical problems in this embodiment, and the process includes the following actions:

[0125] A priority transition instruction is generated to force the feature importance score of the feature to be filtered to be locked at the maximum value, and to switch it from the bypass monitoring channel to the channel for processing high-importance features.

[0126] Specifically, upon receiving a priority transition instruction, the system's internal scheduler immediately ignores the low score originally output by the self-supervised contrastive learning module, such as 0.1, and forcibly overwrites the real-time importance score of the feature to 1.0, the maximum value, through a hardware interrupt or a high-priority software callback. This operation has the highest logical priority and is known as software interrupt locking.

[0127] With the score locked, the data stream's routing path changes instantaneously. The data stream for this feature no longer enters the bypass monitoring channel but is immediately redirected to the high-fidelity lossless compression channel described in Example 1. This means that from this moment on, this feature will enjoy the same treatment as core critical features; the system will utilize the high-fidelity compression algorithm mentioned in Example 1 to fully sample, preserve, and transmit it with high precision. This mechanism ensures that the complete waveform data of the moment the fault occurs and its subsequent development can be fully recorded and uploaded to the cloud, providing valuable underlying record data for subsequent fault tracing.

[0128] like Figure 6 This demonstrates how the secondary monitoring subunit captures ignored early fault signals and dynamically reverses priorities with low computational power consumption. The horizontal axis in the figure represents the sampling point sequence of the feature data, the left vertical axis corresponds to the real-time waveform of the feature to be filtered (represented by the blue solid line), and the right vertical axis corresponds to the dynamic feature importance score (represented by the red dashed line).

[0129] Within the first 250 sampling points of the simulation, this feature, due to its extremely small fluctuations and high correlation with the main sensor, was judged as low-value redundant information by the self-supervised contrastive learning module. Its feature importance score remained at an extremely low level of around 0.05, and it was filtered out and discarded.

[0130] When the sampling point reached 250, a small but statistically significant sudden shift appeared in the waveform, causing the characteristic value to jump from around 25 degrees Celsius to about 26.2 degrees Celsius. Although this value was far from reaching the system's global alarm line, the secondary monitoring subunit immediately initiated the priority reset process because its instantaneous fluctuation index quickly exceeded the preset wake-up threshold.

[0131] Observing the waveform change of the red dashed line in the figure at this moment, it can be found that the feature importance score jumps from an extremely low value and locks to the maximum value of 1.0. The drive system redirects the data stream from the bypass monitoring channel to the high-fidelity lossless processing channel.

[0132] This simulation result proves that the shadow monitoring mechanism described in this invention can accurately identify masked early micro-short circuits or sensor drift signals through the bypass wake-up function without increasing the normal computing power burden, thus ensuring the battery management system's full-coverage capture capability and safety fallback capability when dealing with dynamic evolution anomalies.

[0133] Optionally, to prevent frequent false triggers caused by single noise spikes, the priority reset process also introduces a time window locking mechanism. Once a priority transition is triggered, the high importance state of the feature will be maintained for a certain period of time, such as 60 seconds. During the locking window period, even if the instantaneous fluctuation index falls below the threshold, the feature will still be treated as a high importance feature until the locking timer expires, at which point the system will return it to the self-supervised contrastive learning module for regular evaluation.

[0134] The priority reset process not only solves the current fault detection problem, but also aims to fundamentally optimize the feature evaluation model. Specifically, while triggering the priority transition, the secondary monitoring subunit simultaneously generates a model correction signal, driving the self-supervised contrastive learning module to fine-tune the calculation parameters of the feature importance score online.

[0135] It's also important to note that in traditional open-loop systems, if the model misclassifies an important feature as unimportant, the model itself is unaware of this. However, the secondary monitoring unit in this embodiment essentially acts as a supervisor or discriminator. When the secondary monitoring unit detects that a feature initially classified as low-scoring by the model has actually experienced a significant fluctuation, this essentially constitutes a falsification of the model's original judgment.

[0136] Specifically, the model correction signal includes the feature ID of the mutation, the timestamp of the mutation, and the instantaneous fluctuation index at that time. After receiving this signal, the self-supervised contrastive learning module incorporates it as a high-weight negative sample or penalty term into the loss function. In Example 1, the goal of self-supervised learning is to maximize the mutual information between different views. In the correction process of this example, the algorithm introduces a correction term based on an attention mechanism. The system backpropagates the error gradient and adjusts the weight parameters of the corresponding feature channels in the neural network. Its physical meaning is to tell the model: your belief that this feature is unimportant is incorrect; it has just undergone a critical change, please correct your perception.

[0137] Through this online fine-tuning mechanism, the model gradually learns to identify features with potential mutation risks. In subsequent similar scenarios, the self-supervised contrastive learning module may proactively assign a higher importance score to the feature, thus avoiding the re-intervention of the secondary monitoring subunit. This mechanism enables the feature engineering unit to self-evolve, dynamically adjusting its feature selection strategy as battery aging characteristics change, always maintaining a keen awareness of key information.

[0138] To illustrate the technical effects of this embodiment more intuitively, a typical application scenario example is given below:

[0139] Assume an auxiliary temperature sensor is placed in a corner inside the battery pack. During long-term normal operation, the sensor's readings consistently follow changes in ambient temperature and are highly correlated with the main temperature sensor. Therefore, the self-supervised model of the feature engineering unit classifies it as a redundant feature, assigns it an extremely low importance score (e.g., 0.05), and includes its data in the filter list.

[0140] Without the secondary monitoring subunit of this embodiment, the sensor's data would be directly discarded. However, one day, a minor internal short circuit occurred in the battery cell near the sensor, causing the local temperature to rise abnormally by 3 degrees Celsius within 10 seconds. Since the ambient temperature remained unchanged and the temperature rise did not reach the BMS's absolute high-temperature alarm threshold, such as 55 degrees Celsius, a conventional system would not trigger an alarm. Simultaneously, due to data filtering, crucial early thermal runaway signals were missed.

[0141] However, in the system of this embodiment, although It was sent to the bypass monitoring channel, but the secondary monitoring subunit immediately performed a linear scan on it. Assuming historical average... 25 degrees Celsius, standard deviation The temperature was 0.2 degrees Celsius, indicating that the fluctuations were minimal before, while the sampled value at the moment of the abrupt change was much higher. It is 28 degrees Celsius.

[0142] According to the formula, the instantaneous fluctuation index is calculated. This value is much larger than the preset wake-up threshold, for example... The system immediately detected the anomaly and triggered a priority jump instruction. The importance score was instantly locked at 1.0, and subsequent high-frequency sampling data was uploaded to the cloud in its entirety via a high-fidelity channel. After analyzing this data, the large model in the cloud diagnosed an early internal short circuit and sent a warning to the user. Simultaneously, a model correction signal was fed back to the edge, improving the feature extraction network's performance in subsequent evaluations. The basic weights enable the system to improve itself.

[0143] In summary, this embodiment successfully established a low-cost, high-response parallel bypass monitoring subsystem by constructing a secondary monitoring subunit that includes a bypass monitoring channel, a linear scanning algorithm, instantaneous fluctuation index calculation, and a priority reset mechanism. It effectively compensates for potential cognitive blind spots in the static feature reduction process of deep learning models, resolves the contradiction between static efficiency and dynamic safety, and ensures that the battery health management system can maintain full coverage of critical information under any extreme or sudden operating conditions. This demonstrates extremely high engineering practical value and safety significance.

[0144] Example 3:

[0145] Building upon Embodiments 1 and 2, this embodiment further focuses on the in-depth optimization of the internal logic of the dynamic collaborative prediction unit in the core algorithm layer. This embodiment aims to address a highly challenging conflicting technical problem faced by battery health management systems in actual operation: the conflict between statistical high confidence and physical inconsistency.

[0146] In Example 1, the system established a data reliability score evaluation system based on data integrity, signal-to-noise ratio, and inter-domain KL divergence. In most cases, a high reliability score does indeed represent high-quality data, and it is reasonable for the system to increase the weight of the data-driven model layer accordingly. However, in certain extreme sensor failure modes, such as when a voltage sensor experiences slow linear drift or a fixed zero-point shift, the collected data may still appear perfect in terms of statistical indicators: no data loss (high integrity), smooth waveform without glitches (high signal-to-noise ratio), and the numerical distribution remains within the normal range (low KL divergence). In this case, if the system blindly trusts the statistical score, it will significantly increase the weight of the data-driven model. As a black-box model, the data-driven model is easily misled by data that conforms to statistical distribution but violates physical common sense, thus outputting incorrect lifetime prediction results.

[0147] In order to effectively filter out such highly concealed misleading data, this embodiment introduces a consistency verification subunit in the dynamic collaborative prediction unit and establishes a circuit breaker mechanism based on physical mechanisms.

[0148] Specifically, the dynamic collaborative prediction unit also includes a consistency verification subunit, which acts as a safety valve deployed at the front end of the weight allocation decision module. Its core function is to handle the conflict between statistical high confidence and physical inconsistency, ensuring that any data entering the high-weight decision path is not only statistically sound but also physically real.

[0149] It should also be noted that the activation of this subunit is not indiscriminate, but has explicit triggering conditions. This consistency verification subunit is configured to perform the following operation: in response to the data confidence score being greater than or equal to the high confidence threshold, initiate the physical consistency verification process in parallel.

[0150] For example, when the data credibility score calculation module described in Embodiment 1 calculates the current data stream's score to be 0.9, and assuming the high credibility threshold is 0.8, the system's original logic is to increase the weight of the data-driven model layer to a dominant position. However, in this embodiment, this high score triggers the system's alert mechanism. The system believes that the more seemingly perfect the data, the more rigorous the physical laws must be tested. Therefore, before adjusting the weights, the system will divert a data stream to the consistency verification subunit to begin performing physical verification.

[0151] Specifically, the core of physical verification lies in constructing a reliable physical mechanism reference system. The system calculates the theoretical terminal voltage based on a pre-defined benchmark equivalent circuit model, utilizing real-time acquired current and temperature data. This benchmark equivalent circuit model typically employs a simple yet physically meaningful second-order RC (resistance-capacitance) model or a Thevenin model. The parameters of this model, such as ohmic internal resistance, polarization resistance, and polarization capacitance, are fixed based on the battery's factory calibration data or the most recent, confirmed calibration result, and do not change with short-term data fluctuations.

[0152] During the verification process, the system inputs real-time current data (as excitation input) and temperature data (as environmental parameters) into the benchmark equivalent circuit model. Based on Ohm's law, Kirchhoff's laws, and the principle of battery polarization, the model deduces the theoretical terminal voltage that the battery should exhibit under the current and temperature. This value represents the expected state of the battery voltage under the constraints of physical laws. Simultaneously, the voltage sensor readings acquired by the hardware adaptation layer represent the actual state of the battery voltage, i.e., the measured terminal voltage.

[0153] To quantify the gap between what is and what ought to be, this embodiment introduces a core evaluation index. Specifically, the consistency verification subunit calculates the physical residual index, which is defined as the ratio of the root mean square error of the difference between the measured terminal voltage and the theoretical terminal voltage within a preset time window to the nominal voltage.

[0154] Specifically, let the current time be... The system sets a preset time window length for verification. For example, the past 100 sampling points. At any point within the time window. The measured terminal voltage collected by the sensor is recorded as follows: At the same time The theoretical terminal voltage calculated from the reference equivalent circuit model is denoted as The nominal voltage of a battery, also known as its rated voltage, is denoted as... This is a fixed constant determined by the battery specifications. Therefore, the physical residual index... The mathematical expression is:

[0155] ;

[0156] In the formula: The root mean square error (RMSE) reflects the average deviation between the measured voltage curve and the theoretical voltage curve over a period of time; the denominator is... This serves as a normalization function, making the calculated... By making it a dimensionless percentage value, the differences between battery systems of different voltage levels are eliminated, making it easier to set a unified threshold standard.

[0157] Optionally, the physical meaning of this physical residual index is as follows: if the sensor is working properly and the internal structure of the battery has not suffered severe physical damage, the measured voltage should fluctuate closely around the theoretical voltage. It should approach zero. Conversely, if The significant increase indicates a systematic bias between the two that cannot be explained by measurement noise.

[0158] After obtaining the physical residual index, the system enters the crucial decision-making stage. Specifically, the physical residual index is compared with a preset physical safety boundary threshold. The preset physical safety boundary threshold... This is a red line set based on the physical characteristics of the battery, for example, set to 0.05, or 5%. This threshold represents the maximum tolerance for deviation between the physical model and sensor errors.

[0159] For example, if the calculated Less than or equal to This indicates that the measured data conforms to physical laws, and the previously calculated high data credibility score is real and valid. The system will continue to execute the logic described in Example 1 to increase the weight of the data-driven model.

[0160] Specifically, if the physical residual index is greater than the physical safety boundary threshold, the current data is determined to have a physical violation. This situation is extremely dangerous, as it means that although the data appears statistically perfect, such as having a high signal-to-noise ratio, it has violated fundamental physical constraints between voltage, current, and impedance, such as Ohm's law. In this case, the system must intervene decisively and execute a policy-driven forced coverage operation.

[0161] The strategy's forced coverage operation comprises a series of sequential actions designed to instantly sever the control of erroneous data over the prediction results:

[0162] Action Level 1: Ignore the weight adjustment strategy based on the data credibility score. This is a direct manifestation of the circuit breaker mechanism in this embodiment. No matter how high the previous data credibility score was, even a perfect score of 1.0, once physical verification fails, the score immediately becomes invalid and is no longer used as a basis for weight allocation. This is equivalent to establishing physical consistency as a veto power over statistical credibility.

[0163] Action Level Two: The system forcibly locks the weights of the mechanistic model layer to a preset dominant weight value and reduces the weights of the data-driven model layer to a value complementary to the dominant weight value. Specifically, the system generates a circuit breaker command to directly modify the weight parameters of the model fusion unit. The preset dominant weight value is usually set to a relatively high value, such as 0.9 or 0.95. Assuming the dominant weight value is set to... At this point, the weight of the mechanistic model layer is forcibly locked at 0.9, while the weight of the data-driven model layer is suppressed to 1-0.9=0.1. The logic behind this operation is that when physical violations occur in the data, the output of the data-driven model (which relies entirely on the data input) is highly likely to be incorrect; while the mechanistic model (which relies on physical equations), although it may have errors due to inaccurate parameters, at least guarantees that the output is correct in terms of physical trends. For example, it will never draw conclusions that contradict physical common sense, such as voltage increases during discharge. By allowing the mechanistic model to dominate the prediction, the system achieves a degraded operation or safety fallback during data failures, avoiding giving outrageous lifetime prediction values.

[0164] The third layer of the policy-driven coverage operation involves the handling of source data, aiming to prevent anomalous data from causing bias disturbances to the long-term parameters of the model. Specifically, the system simultaneously generates physical anomaly labels, prohibiting the standardized data stream within the corresponding time window from participating in subsequent updates of the data-driven model layer.

[0165] It's also important to note that in Examples 1 and 2, the system typically uses historical data for online fine-tuning or federated learning training of the model. If data with high statistical scores but physical violations is mistakenly included in the training set as high-quality samples, it will cause persistent damage to the parameters of the data-driven model; this phenomenon is known as data poisoning. Therefore, after determining a physical violation, the consistency verification subunit tags the data within that time window with a physically invalid electronic tag. When the edge-cloud collaborative scheduling module executes the model update task, it scans all data tags, and once the tag is found, the data segment is directly removed. This ensures that the evolution of the data-driven model is always based on physically real data.

[0166] To more clearly demonstrate the beneficial effects of this embodiment, a typical voltage sensor drift scenario is described below:

[0167] Suppose a hardware failure occurs in the voltage acquisition circuit of a certain battery cell, causing the acquired voltage value to be 0.2V higher than the actual value (fixed deviation). The acquired data waveform is still very smooth, and there is no packet loss. In the logic of Example 1, the data integrity is good and the signal-to-noise ratio is high, so the calculated data reliability score may be as high as 0.95. Without this example, the system would significantly increase the weight of the data-driven model. Due to the higher voltage, the data-driven model might misjudge the current SOC as high, thus incorrectly assessing the SOH, or even masking the fact that the battery has been over-discharged.

[0168] However, the process changes after the consistency verification subunit of this embodiment is introduced:

[0169] 1. The system detected a credibility score. This triggers a physical verification.

[0170] 2. The system reads the current discharge current (e.g., 50A) and temperature ( Substituting into the reference equivalent circuit model, the theoretical voltage should be calculated as follows: ;

[0171] 3. The system reads the measured voltage. Due to sensor drift, the reading is... ;

[0172] 4. The system calculates the physical residual index. Assuming the nominal voltage is... The difference is A simple estimate suggests an error ratio of approximately... ;

[0173] 5. System comparison threshold: Physical security boundary thresholds, such as ;

[0174] 6. The system detects a physical violation and triggers a circuit breaker;

[0175] 7. Despite a high confidence score, the system still forcibly locks the weights of the mechanistic model to [value missing]. Although the input voltage of the mechanistic model is incorrect, its internal state observer, such as the extended Kalman filter, will find that the voltage observation value and the state estimate cannot converge, thus limiting the error correction range of SOC. Alternatively, the mechanistic model can directly give a relatively reliable SOC estimate based on the current integral.

[0176] 8. This paragraph Fake data is flagged as anomalous and strictly prohibited from being used for nighttime model self-training, thus protecting the long-term intelligence of the system.

[0177] In summary, this embodiment introduces a consistency verification subunit, adding a layer of highest-priority verification authority based on physical consistency on top of the data-driven statistical evaluation system. It successfully solves the problem of inaccurate predictions caused by blindly trusting high-quality data when sensors experience deceptive malfunctions, significantly improving the robustness and safety of the battery management system under extreme and complex operating conditions.

[0178] Example 4:

[0179] like Figure 7 As shown, this embodiment discloses a battery life prediction and health management method based on big data analysis. This method is logically consistent with the system architecture described in Embodiments 1 to 3 above. Through the orderly execution of the algorithm flow, it achieves closed-loop control from physical layer data perception to cloud-based decision feedback. This method specifically includes the following steps:

[0180] First, a data acquisition and standardization preprocessing step is performed. Multi-source data from the battery is acquired through a standardized interface, and this multi-source data undergoes protocol conversion to generate a standardized data stream. In this step, the standardized interface is responsible for connecting hardware devices with different communication protocols, such as using Controller Area Network (CAN) protocols or industrial communication protocols to acquire real-time data from the battery pack. The multi-source data not only covers the external environmental data and traditional operating data mentioned in Example 1, but more importantly, it includes internal chemical state data acquired by micro-sensors implanted inside the battery. During the generation of the standardized data stream, the system first uses timestamp alignment technology to synchronize the three types of data in the time domain. Then, it calls a two-dimensional deviation correction model, using the state of charge parameter, ambient temperature parameter, and electrolyte concentration parameter as input variables to calculate the temperature sensor deviation value. This deviation value is used to calibrate the original measurement data online, thereby eliminating physical drift caused by long-term sensor operation and ensuring that the standardized data stream entering the algorithm layer has extremely high physical authenticity.

[0181] Secondly, the edge-cloud collaborative scheduling and task triggering steps are executed. The resource availability status of the edge devices is monitored in real time, and the model update strategy is dynamically adjusted based on this status, triggering the battery life prediction task. During this process, the edge device's resource monitor continuously polls the CPU's occupancy and remaining memory space, generating a comprehensive resource availability score. When the score indicates sufficient computing power at the edge device, the edge-cloud collaborative scheduling module executes an aggressive model update strategy, employing the federated learning mechanism described in Example 1. This controls the edge device to train a lightweight local model based on local data and extract knowledge representations, which are then uploaded to the cloud for aggregation after homomorphic encryption. If resources are limited, a conservative strategy is executed, triggering only the core algorithm layer to perform basic prediction tasks. Furthermore, the global teacher model trained by the cloud through aggregating collective intelligence generates a simplified lightweight student model using model distillation technology and distributes it to the edge device. This dynamic adjustment mechanism ensures high availability of the prediction task under different hardware loads.

[0182] Next, the health status index calculation step is performed. Based on the standardized data stream, the battery life prediction task is executed to calculate the battery health status index. This step deeply integrates the microscopic sensor data calculation model described in Embodiment 1 in terms of computational logic.

[0183] Specifically, the system acquires the difference between the real-time electrolyte concentration collected by each microsensor and the initial electrolyte concentration. This difference directly quantifies the degree of chemical consumption within the battery. The difference is then multiplied by a preset electrolyte concentration weighting coefficient to obtain a first product. Simultaneously, the system acquires the ion migration rate collected by the sensors, which represents the kinetic performance of the electrochemical reactions within the battery. This value is multiplied by a preset ion migration rate weighting coefficient to obtain a second product. By summing the first and second products from all sensors, the microscopic physical baseline parameters of the synthesized mechanism model layer are finally obtained.

[0184] In this step, the microscopic physical benchmark parameter is rigorously defined as an intrinsic dimensionality-reduced index for quantifying the electrochemical activity within the battery. This parameter not only reflects the degree of irreversible consumption of internal chemical substances (electrolytes) but also maps the degree of constraint on the liquid-phase ion transport dynamics within the solid state. This process successfully transforms the microscopic degradation mechanisms, which previously required destructive disassembly or expensive laboratory equipment for observation, into quantifiable, real-time physical inputs online.

[0185] More importantly, this microscopic physical baseline parameter provides a highly interpretable mechanistic constraint foundation for subsequent dynamic model fusion, thus forming a complete logical closed loop. Specifically, this parameter is not directly equated to the final macroscopic health score, but is directly fed into the mechanistic model layer of the two-layer hybrid architecture constructed by this method as a state update parameter, used to correct the core parameters inside the baseline equivalent circuit model or electrochemical model in real time. For example, the baseline drift of the ohmic internal resistance or the attenuation baseline of the polarization capacitance is dynamically adjusted based on this baseline parameter.

[0186] Subsequently, during the execution of the core logic of dynamic collaborative prediction, the mechanism model layer, after correction by microscopic physical baseline parameters, outputs a baseline prediction trajectory with strict physical conservation constraints. Simultaneously, the data-driven model layer outputs a fitted prediction trajectory for complex nonlinear operating conditions based on massive external operational characteristics. Finally, the system uses the data credibility score obtained from multidimensional calculations to dynamically allocate and synthesize weights between these two prediction trajectories. This deep closed-loop fusion mechanism completely solves the logical disconnect between the simple linear weighted formula and the aforementioned edge-cloud dual-layer collaborative architecture, ensuring that this invention not only achieves a leap from microscopic perception to macroscopic prediction in data flow but also achieves deep logical self-consistency between physical laws and statistical learning in model mechanism.

[0187] This process transforms previously difficult-to-observe microscopic decay mechanisms into quantifiable macroscopic health scores, offering greater physical interpretability compared to traditional methods.

[0188] Based on this, the core logic of dynamic collaborative prediction is executed. Specifically, the step of performing the battery life prediction task to calculate battery health status indicators includes: real-time calculation of the data reliability score of the standardized data stream. This score is obtained by weighted summation of data integrity indicators, signal-to-noise ratio indicators, and inter-domain KL divergence indicators. Based on the data reliability score, the weight allocation of the mechanistic model layer and the data-driven model layer is dynamically adjusted to synthesize the final battery life prediction result. When the data reliability score is greater than or equal to a preset high reliability threshold, the system increases the weight of the data-driven model layer, utilizing its powerful nonlinear fitting capability to capture subtle features of battery aging.

[0189] At this point, to prevent statistical fraud, the system will synchronously activate the consistency verification subunit as described in Example 3, calculate the theoretical terminal voltage using the benchmark equivalent circuit model, and compare it with the measured terminal voltage to generate a physical residual index. If the physical residual index exceeds the physical safety boundary threshold, a physical violation is determined, and the weight of the mechanism model layer is forcibly locked as the dominant weight value, realizing the veto power of physical laws over statistical scoring.

[0190] Finally, the model adaptive fine-tuning and anomaly response steps are executed. In response to the data confidence score falling below a preset fine-tuning trigger threshold, a transfer learning-based model fine-tuning process is automatically triggered. At this point, the system recognizes that the current operating condition has deviated from the original training distribution. By introducing a small amount of target domain data, the data-driven model is rapidly fine-tuned to readjust to the new electrochemical environment. Simultaneously, the secondary monitoring subunit in the feature engineering unit continuously runs the bypass monitoring channel, performing a linear scan of low-importance features marked for filtering. Once the instantaneous fluctuation index exceeds the wake-up threshold, a priority reset process is immediately triggered, locking the feature importance score to its maximum value and switching to the lossless channel, thereby capturing weak signals of early faults.

[0191] In summary, the method described in this embodiment, through the coordinated operation of multiple logical steps, not only achieves high-quality processing of multi-source battery data, but also resolves the conflict between statistical high confidence and physical inconsistency through deep integration of data confidence scores and physical consistency verification. This method significantly improves the robustness and safety of battery life prediction under all operating conditions while ensuring edge computing efficiency, truly achieving accurate perception of the battery's safety status throughout its entire lifecycle.

[0192] Example 5:

[0193] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0194] like Figure 8 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0195] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0196] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0197] The memory 103 stores a computer program corresponding to the battery life prediction and health management method based on big data analysis in the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0198] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 8 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0199] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A battery life prediction and health management system based on big data analysis, characterized in that, It includes a hardware adaptation layer, an edge-cloud collaborative scheduling module, and a core algorithm layer; The hardware adaptation layer is used to collect multi-source data from the battery through a standardized interface, and to perform protocol conversion on the multi-source data to generate a standardized data stream. The edge-cloud collaborative scheduling module is used to monitor the resource availability status of the edge in real time, dynamically adjust the model update strategy based on the resource availability status, and trigger the core algorithm layer to execute the prediction task. The model update strategy is dynamically adjusted based on the resource availability status, specifically including: when the resource availability status is high, large-scale model parameter downloads or complex gradient uploads are allowed; when the resource availability status is low, only a small number of key features are allowed to be transmitted or model updates are paused, prioritizing real-time prediction tasks of the core algorithm layer; the model update strategy includes federated learning and model distillation mechanisms, and the edge-cloud collaborative scheduling module is configured to perform the following operations: control the edge to train a lightweight local model based on local data and generate knowledge representations of intermediate layer feature maps; upload the knowledge representations to the cloud after homomorphic encryption; control the cloud to aggregate all the knowledge representations uploaded by the edge to train a global teacher model, and generate a lightweight student model through model distillation, and distribute the lightweight student model to the edge; The core algorithm layer is used to calculate battery health status indicators based on the standardized data stream through a dynamic collaborative prediction unit. The operation of the dynamic collaborative prediction unit includes: calculating the data credibility score of the standardized data stream in real time; dynamically adjusting the weight allocation of the mechanism model layer and the data-driven model layer according to the data credibility score, and synthesizing the final battery life prediction result; and automatically triggering the model fine-tuning process based on transfer learning in response to the data credibility score being lower than a preset fine-tuning trigger threshold. The process of generating the data credibility score includes: calculating the data integrity index, signal-to-noise ratio index, and inter-domain KL divergence index of the standardized data stream respectively; performing a weighted summation of the data integrity index, the signal-to-noise ratio index, and the inter-domain KL divergence index to obtain the data credibility score; the dynamic adjustment of the weight allocation of the mechanism model layer and the data-driven model layer specifically includes: when the data credibility score is greater than or equal to a preset high credibility threshold, reducing the weight of the mechanism model layer and increasing the weight of the data-driven model layer.

2. The system according to claim 1, characterized in that, The multi-source data includes internal chemical state data, external environmental data, and conventional operating data collected by micro-sensors; The process of data fusion processing performed by the hardware adaptation layer includes: using timestamp alignment technology to perform time-dimension consistency fusion of the internal chemical state data, the external environment data, and the traditional operating data; A two-dimensional deviation correction model is constructed. The temperature sensor deviation value is calculated using the state of charge parameter, ambient temperature parameter, and electrolyte concentration parameter. The original measurement data is then calibrated online using the temperature sensor deviation value.

3. The system according to claim 2, characterized in that, The logical definition for calculating the battery health status index in the core algorithm layer is as follows: The difference between the real-time electrolyte concentration collected by each of the micro sensors and the initial electrolyte concentration is obtained, and the difference is multiplied by a preset electrolyte concentration weighting coefficient to obtain the first product; The ion migration velocity collected by each of the microsensors is obtained, and the ion migration velocity is multiplied by a preset ion migration velocity weighting coefficient to obtain a second product; The first product and the second product corresponding to all the microsensors are summed to obtain the micro-electrochemical health factor. The external operating characteristics of the battery are obtained, including at least one of the following: ampere-hour throughput, equivalent full cycle count, characteristic peak area and position offset, and macroscopic internal resistance growth rate. The microscopic electrochemical health factors are fed into the mechanism model layer as state update parameters to output a baseline predicted trajectory. The data-driven model layer outputs a fitted predicted trajectory based on the external operating characteristics. The dynamic collaborative prediction unit dynamically assigns weights and synthesizes the baseline predicted trajectory and the fitted predicted trajectory to obtain the final battery health state index.

4. The system according to claim 1, characterized in that, The core algorithm layer also includes a feature engineering unit, which is used to perform feature processing for computing power grading, including: The feature importance score for each feature dimension in the standardized data stream is calculated using a self-supervised contrastive learning module. Obtain the current computing power level of the edge controller, and select the corresponding feature compression scheme based on the current computing power level; If the feature importance score is greater than or equal to a first preset threshold, the feature is processed using a high-fidelity lossless compression scheme; If the feature importance score is between the first preset threshold and the second preset threshold, the feature is processed using a standard compression scheme with an autoencoder; If the feature importance score is less than the second preset threshold, then the feature is marked as a feature to be filtered and a filtering removal operation is performed.

5. The system according to claim 1, characterized in that, The core algorithm layer also integrates a model predictive control multi-objective optimization module, which is used to construct a cost function that includes a temperature tracking error term, an actuator energy consumption term, a health state decay rate term, and a control quantity change term. The model predictive control multi-objective optimization module, based on the cost function, solves for the optimal control strategy to minimize the total cost value of the cost function, while satisfying the battery temperature state constraints and charge / discharge input constraints.

6. The system according to claim 1, characterized in that, The core algorithm layer also includes a multi-scale anomaly response unit for executing a hierarchical response process: Real-time monitoring of battery health status change rate, electrolyte concentration change, and internal resistance growth rate; If it is determined to be a Level 1 anomaly, the control edge will adjust the charging rate and only upload the enhanced sensitive feature data; If the anomaly is determined to be level 2, control the edge device to upload the compressed complete data and trigger the cloud to generate maintenance suggestions; If the anomaly is determined to be Level 3, the control edge terminal will immediately stop the charging and discharging operation and trigger an audible and visual warning.

7. The system according to claim 4, characterized in that, The feature engineering unit further includes a secondary monitoring subunit, which is configured to perform the following operations: Establish a bypass monitoring channel to take over the data marked as features to be filtered by the feature engineering unit; The data in the bypass monitoring channel is linearly scanned, and the instantaneous fluctuation index is calculated. The instantaneous fluctuation index is defined as the ratio of the absolute value of the difference between the current sampling point value and the historical sliding window mean to the historical sliding window standard deviation. The instantaneous fluctuation index is compared with a preset wake-up threshold; If the instantaneous fluctuation index is greater than the wake-up threshold, a priority reset process is triggered: Generate a priority transition instruction to force the feature importance score of the feature to be filtered to be locked at the maximum value, and switch it from the bypass monitoring channel to the channel for processing high-importance features; Simultaneously, a model correction signal is generated to drive the self-supervised contrastive learning module to fine-tune the calculation parameters of the feature importance score online.

8. The system according to claim 1, characterized in that, The dynamic collaborative prediction unit further includes a consistency verification subunit, which is configured to perform the following operations: In response to the data confidence score being greater than or equal to the high confidence threshold, a physical consistency verification process is initiated in parallel. Based on a preset reference equivalent circuit model, the theoretical terminal voltage is calculated using real-time acquired current and temperature data. Calculate the physical residual index, which is defined as the ratio of the root mean square error of the difference between the measured terminal voltage and the theoretical terminal voltage within a preset time window to the nominal voltage. The physical residual index is compared with a preset physical safety boundary threshold. If the physical residual index is greater than the physical security boundary threshold, then the current data is determined to have a physical violation, and a forced overwrite operation is executed. Ignore the weight adjustment strategy based on the data credibility score, forcibly lock the weight of the mechanism model layer to a preset dominant weight value, and reduce the weight of the data-driven model layer to a value that is complementary to the dominant weight value; Simultaneously, physical anomaly labels are generated to prevent standardized data streams within the corresponding time window from participating in subsequent updates of the data-driven model layer.

Images