Integrated battery state monitoring and assessment system based on multi-modal sensing

Abstract

The application relates to the technical field of battery state monitoring, and discloses an integrated battery state monitoring and evaluation system based on multi-modal sensing. The system comprises the following steps: collecting battery signals through multi-source heterogeneous sensors, forming a standardized fusion data sequence after processing, and extracting a feature vector according to a state classification atlas. The core lies in dynamically adapting a priori evaluation network by using a real-time feature vector, generating a personalized evaluation model, and performing forward deduction compensation on the state evaluation result after verification by analyzing signal gradient and feature evolution mode. The system realizes online optimization of the evaluation model for specific batteries, improves the accuracy of state determination, and enhances the timeliness of risk early warning through a forward-looking compensation mechanism.

Description

Technical Field

[0001] This invention relates to the field of battery state monitoring technology, specifically to an integrated battery state monitoring and evaluation system based on multimodal sensing. Background Technology

[0002] Currently, battery status monitoring mainly relies on deploying several sensors to collect signals such as voltage and temperature, and then analyzing them using pre-trained fixed evaluation models. These general models are typically trained on standard operating conditions or typical battery sample data, and their parameters and structures are static. At the evaluation output level, most existing technologies analyze and judge based on sensor data from the current moment or a recent historical window, outputting diagnostic conclusions or warning information on the battery's current health status, safety status, or lifespan status.

[0003] This static, generalized evaluation model is ill-suited to the inherent differences in manufacturing processes and material degradation characteristics among individual batteries. It also fails to adequately address the dynamic changes in the aging process of the same battery throughout its lifespan, leading to challenges in the accuracy and robustness of the evaluation results in complex real-world applications. Evaluation models based solely on historical and current data are inherently lagging; their output reflects the current state and fails to anticipate rapid changes or potential risks that may occur in the battery's condition.

[0004] Existing battery condition monitoring technologies need to address the issue of the dynamic adaptability of assessment models to different individual batteries and different operating stages, as well as overcome the limitations of lagging assessment conclusions and lack of foresight, in order to achieve more accurate and timely condition perception and risk warning. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated battery state monitoring and evaluation system based on multimodal sensing to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an integrated battery state monitoring and evaluation system based on multimodal sensing, the system comprising: The data acquisition module is used to acquire raw sensing signals synchronously generated by multiple heterogeneous sensors deployed on the battery entity; The data processing module is used to perform spatiotemporal fusion and cleaning on the original sensing signals to form a standardized fused sensing data sequence. The feature mapping module is used to perform feature deconstruction and state mapping on the fused sensing data sequence based on a preset battery state classification map, and generate a set of state feature vectors with time stamps. The network adaptation module is used to dynamically adapt the prior state evaluation decision network using the set of state feature vectors to generate the adapted state evaluation decision network. The evaluation and reasoning module is used to input the fused sensing data sequence and the set of state feature vectors into the adapted state evaluation and decision network, perform parallel evaluation and reasoning, and output a preliminary state evaluation and decision set. The timing verification module is used to incorporate historical state evolution trajectories to perform timing consistency verification and correction on the preliminary state assessment decision set, forming a verified baseline state assessment decision. The compensation output module is used to perform forward extrapolation compensation on the verified baseline state evaluation decision based on the changing gradient of the real-time sensing signal and the evolution pattern of the state feature vector set, and generate an optimized state evaluation decision.

[0007] Preferably, the step of performing spatiotemporal fusion and cleaning on the original sensing signals to form a standardized fused sensing data sequence includes: Receives asynchronous raw data streams from temperature, voltage, and current sensors; Apply a uniform timestamp to each of the asynchronous raw data streams; Based on the unified timestamp marker, all asynchronous raw data streams are time-aligned and interpolated to generate time-synchronized multi-channel data streams. Anomaly detection and removal are performed on each channel of the time-synchronized multi-channel data stream; The data from each channel after the outlier removal was performed to normalize the dimensions. The multi-channel data that has undergone dimension normalization is packaged in chronological order to form the standardized fused sensing data sequence.

[0008] Preferably, the step of performing feature deconstruction and state mapping on the fused sensing data sequence based on a preset battery state classification map to generate a set of time-stamped state feature vectors includes: Load the preset battery state classification map, which defines the correspondence rules between various battery states and sensor data patterns; A data window of a fixed time length is slidably extracted from the fused sensing data sequence; For each data window, statistical features, frequency domain features, and trend features within the window are extracted according to the corresponding rules. All features extracted for a data window are concatenated to form a state feature vector, and the state feature vector is marked with the central time of the corresponding data window. The fused sensing data sequence is traversed to generate a set of time-stamped state feature vectors corresponding to each data window.

[0009] Preferably, the step of dynamically adapting the prior state evaluation decision network using the set of state feature vectors to generate an adapted state evaluation decision network includes: Invoke the pre-trained prior state evaluation decision network, which contains updatable network parameters; Select a subset of recently generated vectors from the set of state feature vectors as the adaptation sample set; Using the evaluation feedback signal recorded at the previous moment as a supervision signal, the difference between the output generated by the adaptation sample set through the prior state evaluation decision network and the supervision signal is calculated; Based on the differences, the updatable network parameters are fine-tuned using a backpropagation mechanism. When the difference is less than a preset threshold, fine-tuning stops, and the network at this point is saved as the adapted state evaluation decision network.

[0010] Preferably, the step of inputting the fused sensing data sequence and the set of state feature vectors into the adapted state evaluation decision network, performing parallel evaluation inference, and outputting a preliminary state evaluation decision set includes: The fused sensing data sequence is divided into several logical data segments; Pair the state feature vector corresponding to each logical data segment in the state feature vector set with the logical data segment; Each pair of paired data is simultaneously input into multiple parallel inference branches of the adapted state evaluation decision network; Each parallel inference branch independently processes the input data pair and outputs a candidate state evaluation decision; The candidate state evaluation decisions output by all parallel inference branches are aggregated to form the preliminary state evaluation decision set.

[0011] Preferably, the step of introducing historical state evolution trajectories to perform temporal consistency verification and correction on the preliminary state assessment decision set to form a verified baseline state assessment decision includes: The historical state evolution trajectory is retrieved from the system database. The historical state evolution trajectory records the sequence of changes in battery state assessment decisions over time. Analyze the statistical regularities and constraints of state changes in the historical state evolution trajectory; Each candidate decision in the preliminary state assessment decision set is placed at the end of the historical state evolution trajectory, and simulated as the next state. Determine whether the next state obtained from the simulation violates the stated statistical laws and constraints; If the conditions are violated, the candidate decision shall be modified according to the statistical regularity and constraints described above; After all candidate decisions have been verified and possibly corrected, the decision with the highest confidence level is selected as the verified baseline state evaluation decision.

[0012] Preferably, the step of performing forward extrapolation compensation on the verified baseline state evaluation decision based on the changing gradient of the real-time sensing signal and the evolution pattern of the state feature vector set to generate an optimized state evaluation decision includes: Calculate the gradient of the real-time sensing signal at the current moment relative to the previous moment; Analyze the evolution pattern of the state feature vector set within a recent time window to identify the dominant evolution trend; A short-term state extrapolation model is established, starting with the verified baseline state assessment decision and using the change gradient and the dominant evolution trend as extrapolation parameters. Run the short-term state inference model to predict possible state assessment points at several future time points; Based on the predicted possible state assessment points, the compensation amount for the decision at the current moment is calculated in reverse. The compensation amount is superimposed on the verified baseline state evaluation decision to generate the optimized state evaluation decision.

[0013] Preferably, the step of establishing a short-term state extrapolation model, starting with the verified baseline state assessment decision and using the change gradient and the dominant evolution trend as extrapolation parameters, includes: The verified baseline state evaluation decision is used as the initial state vector for deduction, and the state space for a finite number of consecutive future time steps is defined. Set a fixed time step and a total number of steps for the forward inference, where the total number of steps defines the time span of the inference; The instantaneous change term of the state transition is constructed based on the change gradient, and the trend term of the state transition is constructed based on the dominant evolution trend; The instantaneous change term and the trend term are weighted and fused to form a single-step state transition function; Starting with the initial state vector of the deduction, and with the single-step state transition function as the core, the predicted state vector of each subsequent time step is iteratively calculated to obtain a predicted state sequence. The predicted state sequence is smoothed, and the overall trend information of state changes is extracted from it as the inference output. Based on the predicted trend described in the inference output, a compensation amount for adjusting the current decision is calculated in reverse. The compensation amount is superimposed and fused with the verified baseline state evaluation decision to generate the final optimized state evaluation decision for output.

[0014] Preferably, the construction and updating of the preset battery state classification map is achieved through the following steps: Collect a large amount of multimodal sensing data and corresponding known state labels of the same type of battery throughout its entire life cycle; Cluster analysis was performed on the collected data to identify the typical characteristics and boundaries of sensor data patterns under different state labels; Based on the cluster analysis results, a state region partitioning diagram with key sensing features as coordinate axes is drawn. The state region division map, the typical feature descriptions of each region, and the transition probability relationships between state labels are encoded into a structured knowledge base to form an initial battery state classification map. During system operation, when the evaluation feedback signal deviates significantly from the prediction made based on the battery state classification map, the map update mechanism is triggered to iteratively correct the boundary and transition probability relationship in the battery state classification map using newly generated data.

[0015] Preferably, the system further includes: driving the actuator to complete a specified monitoring action on the battery based on the optimized state assessment decision, and recording the corresponding assessment feedback signal, including: The optimized state assessment decision is analyzed to obtain at least one specific monitoring action instruction and expected parameter range; The monitoring action command is sent to the corresponding physical actuator, which includes a load controller, a thermal management controller, or a data acquisition card. The physical actuator is controlled to apply specific electrical loads or temperature environments to the battery or initiate high-precision data acquisition according to instructions. Receive newly acquired sensing data from the plurality of heterogeneous sensors during the execution of the monitoring action; Extract key indicators related to the expected parameter range from the newly acquired sensor data, and store the key indicators as the evaluation feedback signal in the system database.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By dynamically adapting the prior network using a set of state feature vectors extracted in real time from the target battery, the generated evaluation network's parameters and structure can be optimized and adjusted to suit the battery's current actual characteristics. This transforms the evaluation model from a general, fixed paradigm into a specialized analyzer highly matched to the instantaneous state of a specific battery. It overcomes the model mismatch problem caused by individual battery differences and time-varying state characteristics, improving the individualized accuracy and scenario adaptability of state evaluation, and making the evaluation results more realistically reflect the actual condition of the monitored object.

[0017] The forward extrapolation compensation of benchmark assessment decisions based on real-time sensor signal change gradients and state feature vector evolution patterns essentially fuses information representing the instantaneous rate of state change with the continuous evolution trend, and uses this information to predict the state's trajectory in the short term. By incorporating potential future evolutions into the current assessment, the final optimized decision not only describes the current state but also includes predictive judgments about the direction and rate of state development. This overcomes the lag in traditional assessment methods, enabling a shift from post-event diagnosis to pre-event early warning, and providing a valuable time window for risk intervention and operational decision-making. Attached Figure Description

[0018] Figure 1 This is a timing diagram of the integrated battery state monitoring and evaluation system based on multimodal sensing described in this invention. Figure 2 A flowchart for data spatiotemporal fusion and cleaning; Figure 3 The flowchart is for feature deconstruction and state mapping; Figure 4 This is a graph showing the changes in battery state assessment decisions and compensation quantities under short-term state simulation. Figure 5 A heatmap of confidence scores for parallel inference branches at each stage of battery state monitoring. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides an integrated battery state monitoring and evaluation system based on multimodal sensing. The system includes: a data acquisition module that acquires raw sensing signals synchronously generated by multiple heterogeneous sensors deployed on the battery entity; a data processing module that performs spatiotemporal fusion and cleaning of the raw sensing signals to form a standardized fused sensing data sequence; a feature mapping module that performs feature deconstruction and state mapping on the fused sensing data sequence according to a preset battery state classification map to generate a set of time-stamped state feature vectors; a network adaptation module that dynamically adapts the prior state evaluation decision network using the set of state feature vectors to generate an adapted state evaluation decision network; an evaluation inference module that inputs the fused sensing data sequence and the set of state feature vectors into the adapted state evaluation decision network, performs parallel evaluation inference, and outputs a preliminary state evaluation decision set; and a time-series verification module that incorporates historical state evolution trajectories to perform time-series consistency verification and correction on the preliminary state evaluation decision set to form a verified baseline state evaluation decision. The compensation output module performs forward extrapolation compensation on the verified baseline state assessment decision based on the changing gradient of the real-time sensor signal and the evolution pattern of the state feature vector set, generating an optimized state assessment decision.

[0021] Example 1: See Figure 2 The data acquisition module receives asynchronous raw data streams from temperature, voltage, and current sensors. A unified timestamp is applied to each asynchronous raw data stream. Based on the unified timestamp, all asynchronous raw data streams undergo time-series alignment and interpolation to generate a time-synchronized multi-channel data stream. Outlier detection and removal are performed on each channel of the time-synchronized multi-channel data stream. The data from each channel after outlier removal is then normalized. The normalized multi-channel data is then encapsulated in chronological order to form a standardized fused sensor data sequence.

[0022] In practice, the data processing module receives asynchronous raw data streams from temperature, voltage, and current sensors. These asynchronous raw data streams each have their own independent sampling clock and data generation cycle. The data processing module applies a unified timestamp to each asynchronous raw data stream, derived from the system's high-precision global clock. Based on the unified timestamp, the data processing module performs timing alignment and interpolation on all asynchronous raw data streams. This process is achieved by resampling each data stream onto a common, equally spaced high-frequency time grid, generating time-synchronized multi-channel data streams.

[0023] In practice, outlier detection and removal are performed on each channel of the time-synchronized multi-channel data stream. Outlier detection employs a sliding window-based statistical thresholding method, identifying and marking data points within the window that exceed three standard deviations of the mean. Linear interpolation is then used to replace the marked outlier data. After outlier removal, each channel's data undergoes dimensional normalization, mapping sensor readings of different physical dimensions and magnitudes to a unified numerical range.

[0024] In some embodiments, the dimensional normalization process employs a linear transformation method, and the transformation formula is as follows:

[0025] in: Represents a data point in the original sensor data sequence. This represents the arithmetic mean of the channel data over a calibration period. This represents the standard deviation of the channel data within the same calibration period. This represents the normalized data value. It can be understood that the calibration cycle is preset based on the specific application scenario of the battery. The multi-channel data, after dimensional normalization, is packaged in chronological order to form a standardized fusion sensor data sequence. This standardized fusion sensor data sequence is a two-dimensional array, where the rows correspond to time points and the columns correspond to sensor channels.

[0026] Optionally, a uniform timestamp is applied immediately upon data entering the processing queue, with the tagging operation performed by a hardware interrupt service routine. It is understood that the frequency of the common, equally spaced high-frequency time grid is no less than twice the highest sampling frequency in all input asynchronous raw data streams. In some embodiments, the linear interpolation method utilizes adjacent valid data points before and after outliers for calculation. Optionally, the arithmetic mean and standard deviation are calculated based on sample data collected under known steady-state operating conditions of the battery.

[0027] Example 2: See Figure 3The feature mapping module loads a pre-defined battery state classification map, which defines the correspondence rules between various battery states and sensor data patterns. A fixed-time window is slidably extracted from the fused sensor data sequence. For each data window, statistical features, frequency domain features, and trend features are extracted according to the correspondence rules. All features extracted for a data window are concatenated to form a state feature vector, and the center time of the corresponding data window is marked on the state feature vector. The fused sensor data sequence is traversed to generate a set of time-stamped state feature vectors corresponding to each data window. The construction and updating of the battery state classification map is achieved through the following steps: collecting a large amount of multimodal sensor data of the same type of battery throughout its entire life cycle and corresponding known state labels; performing cluster analysis on the collected data to identify the typical features and boundaries of sensor data patterns under different state labels; drawing a state region partitioning map with key sensor features as coordinate axes based on the cluster analysis results; and encoding the state region partitioning map, the typical feature descriptions of each region, and the transition probability relationships between state labels into a structured knowledge base to form the initial battery state classification map. During system operation, when the evaluation feedback signal deviates significantly from the prediction made based on the battery state classification map, the map update mechanism is triggered, and the boundary and transition probability relationship in the battery state classification map is iteratively corrected using the newly generated data.

[0028] In practical implementation, the feature mapping module loads a preset battery state classification map. This map defines the correspondence rules between various battery states and sensor data patterns, stored in a list format. Each entry is associated with a state identifier and a set of feature threshold conditions. The feature mapping module slides a fixed-time-length data window from the standardized fused sensor data sequence, with a sliding step size smaller than the data window length to achieve temporal overlap sampling. For each data window, statistical features, frequency domain features, and trend features are extracted according to the corresponding rules. Statistical features include the arithmetic mean, standard deviation, maximum, and minimum values ​​of the sensor channel readings within the data window. Frequency domain features are obtained by calculating the frequency components of each channel data within the data window using Fast Fourier Transform to extract energy spectrum features. Trend features are obtained by fitting the slope of each channel data within the data window over time using the least squares method. All features extracted for a data window are concatenated to form a state feature vector, which is a one-dimensional array. The state feature vector is marked with the center time of the corresponding data window. The feature mapping module traverses the fused sensor data sequence to generate a set of time-stamped state feature vectors corresponding to each data window.

[0029] In some embodiments, the calculation of the energy spectral characteristics of the frequency domain features involves summing the energy of the transformed frequency domain components, as shown in the following formula:

[0030] in: This represents the total energy characteristic value of the current data window in the frequency domain. The first term obtained by the Fast Fourier Transform is... The complex amplitudes of each frequency domain component are summed, covering all effective frequency indices from the DC component to the Nyquist frequency. It can be understood that the number of points in the Fast Fourier Transform is equal to the number of sampling points corresponding to the data window length. Optionally, the slope fitting object for the trend feature can be the original data or the data after smoothing and filtering.

[0031] In practical implementation, the construction and updating of the preset battery state classification map is achieved through the following steps. The construction process collects a large amount of multimodal sensor data and corresponding known state labels of the same type of battery throughout its entire life cycle. These known state labels are provided by laboratory standard cyclic testing protocols or disassembly verification results. Cluster analysis is performed on the collected data, using K-means or hierarchical clustering algorithms to identify typical features and boundaries of sensor data patterns under different state labels. Based on the cluster analysis results, a state region partitioning map is drawn with key sensor features as coordinate axes. The key sensor features are selected from statistical features, frequency domain features, and trend features, choosing the top few dimensions with the highest discriminative power. The state region partitioning map, the typical feature descriptions of each region, and the transition probability relationships between state labels are encoded into a structured knowledge base, forming the initial battery state classification map. It can be understood that the transition probability relationships are obtained statistically by analyzing continuous state label sequences over time. In some embodiments, before cluster analysis, the sensor data needs to undergo standardization and feature extraction preprocessing consistent with the feature mapping module. Optionally, the map update mechanism is triggered when the prediction deviation exceeds the set tolerance multiple times in a row. Using newly generated data with high-confidence state labels, the cluster boundaries and state transition probability matrices in the battery state classification map are numerically corrected through an incremental learning algorithm.

[0032] Example 3: The network adaptation module calls the pre-trained prior state evaluation decision network, which contains updatable network parameters. A subset of recently generated vectors from the state feature vector set is selected as the adaptation sample set. The evaluation feedback signal recorded at the previous time step is used as the supervision signal, and the difference between the output generated by the adaptation sample set through the prior state evaluation decision network and the supervision signal is calculated. The updatable network parameters are fine-tuned based on the difference using a backpropagation mechanism. When the difference is less than a preset threshold, fine-tuning stops, and the network at this point is saved as the adapted state evaluation decision network. The evaluation inference module divides the fused sensor data sequence into several logical data segments. The state feature vectors corresponding to each logical data segment in the state feature vector set are paired with the logical data segments. Each paired data pair is simultaneously input into multiple parallel inference branches of the adapted state evaluation decision network. Each parallel inference branch independently processes the input data pair and outputs a candidate state evaluation decision. The candidate state evaluation decisions output by all parallel inference branches are collected to form a preliminary state evaluation decision set.

[0033] In practice, the network adaptation module calls a pre-trained prior state evaluation decision network. This prior network contains updatable network parameters, which are weights and biases learned internally using gradient descent. The module selects recently generated vectors from the state feature vector set as the adaptation sample set. "Recently generated" is determined by the time stamp falling within a specific time range prior to the current moment. The evaluation feedback signal recorded in the previous moment is used as the supervision signal. The difference between the output generated by the prior state evaluation decision network and the supervision signal is calculated. This difference is typically calculated using a loss function, whose output is a scalar value measuring the degree of mismatch between the network prediction and the supervision signal. Based on the difference, the updatable network parameters are fine-tuned using a backpropagation mechanism. Backpropagation calculates the gradient of the loss function with respect to each updatable network parameter according to the chain rule and adjusts the parameter values ​​in the opposite direction of the gradient. Fine-tuning stops when the difference is less than a preset threshold, which is a predefined numerical tolerance. The network at this point is then saved as the adapted state evaluation decision network.

[0034] In some embodiments, the loss function used to calculate the difference is in the form of mean squared error, and its formula is expressed as:

[0035] in: The calculated difference represents the loss. Represents the number of state feature vectors in the adapted sample set, and represents the number of features related to the first... The historical evaluation feedback signals obtained from the system database corresponding to each state feature vector The prior state assessment decision network represents the first... The output predicted value of each state feature vector.

[0036] In practical implementation, the evaluation inference module divides the fused sensor data sequence into several logical data segments, based on a fixed time length or a fixed number of data points. The evaluation inference module pairs the state feature vector corresponding to each logical data segment with the logical data segment, based on the alignment of the time stamp of the state feature vector with the center point of the time interval of the logical data segment. Each paired data pair is simultaneously input into multiple parallel inference branches of the adapted state evaluation decision network. These parallel inference branches have the same network structure but are initialized with different internal parameters. Each parallel inference branch independently processes the input data pair and outputs a candidate state evaluation decision, represented as a vector. Each element in the vector corresponds to a quantified evaluation value of a battery state attribute. The candidate state evaluation decisions output by all parallel inference branches are aggregated to form a preliminary state evaluation decision set. This preliminary state evaluation decision set is a three-dimensional array, where the first dimension corresponds to different parallel inference branches, the second dimension corresponds to different logical data segments, and the third dimension corresponds to the elements of the decision vector. In some embodiments, the number of parallel inference branches is a pre-defined constant. It is understood that temporal overlap is allowed in the division of logical data segments. Optionally, multiple parallel inference branches of the adapted state evaluation decision network are forced to have diversity constraints during the training phase to ensure that they can process the input data from slightly different perspectives.

[0037] Example 4: The timing verification module retrieves historical state evolution trajectories from the system database. These trajectories record the sequence of battery state assessment decisions over time. The module analyzes the statistical patterns and constraints of state changes within these historical trajectories. Each candidate decision from the initial state assessment decision set is placed at the end of the historical state evolution trajectory, simulating it as the next state. The module determines whether the simulated next state violates statistical patterns and constraints. If so, the candidate decisions are corrected based on these patterns and constraints. After all candidate decisions have been verified and potentially corrected, the decision with the highest confidence level is selected as the verified baseline state assessment decision. The compensation output module calculates the gradient of the real-time sensor signal relative to the previous moment. It analyzes the evolution pattern of the state feature vector set within the most recent time window to identify the dominant evolution trend.

[0038] A short-term state inference model is established, starting with a validated baseline state assessment decision and using the gradient of change and the dominant evolution trend as inference parameters. The short-term state inference model is run to predict possible state assessment points at several future time steps. Based on the predicted possible state assessment points, the compensation amount for the current time step decision is calculated in reverse. The compensation amount is superimposed with the validated baseline state assessment decision to generate an optimized state assessment decision. The establishment process of the short-term state inference model is as follows: The validated baseline state assessment decision is used as the initial state vector for inference, and the state space for a finite number of consecutive future time steps is defined. A fixed time step and a total number of steps are set for the forward inference, with the total number of steps defining the time span of the inference. Instantaneous change terms for state transitions are constructed based on the gradient of change, and trend terms for state transitions are constructed based on the dominant evolution trend. The instantaneous change terms and trend terms are weighted and fused to form a single-step state transition function. Starting with the initial state vector and using the single-step state transition function as the core, the predicted state vector for each subsequent time step is iteratively calculated to obtain a predicted state sequence. The predicted state sequence is smoothed, and the overall trend information of state changes is extracted as the inference output. Based on the predicted trend described by the inference output, a compensation amount is calculated in reverse to adjust the current decision. The compensation amount is superimposed and fused with the validated baseline state evaluation decision to generate the final optimized state evaluation decision used for output.

[0039] In practical implementation, the timing verification module retrieves historical state evolution trajectories from the system database. These trajectories record the sequence of battery state assessment decisions over time. The module analyzes the statistical patterns and constraints of state changes within these historical trajectories. The statistical patterns include the mean and variance of changes in different state attributes between adjacent time points, while the constraints define the valid range of state attribute values ​​and the maximum allowable change between adjacent time points. The timing verification module places each candidate decision from the initial state assessment decision set at the end of the historical state evolution trajectory and simulates it as the next state. It then determines whether the simulated next state violates the statistical patterns and constraints by checking whether the state attribute values ​​exceed the valid range and whether their change relative to the previous historical state exceeds the maximum allowable change. If violations are found, the candidate decisions are corrected according to the statistical patterns and constraints. The correction method involves trimming out-of-bounds attribute values ​​to the boundary and proportionally reducing the excessive change to the allowable range. After all candidate decisions have been verified and possibly corrected, the decision with the highest confidence is selected as the verified baseline state evaluation decision. The confidence level is assigned by the historical accuracy metric of the parallel inference branch that generated the candidate decision on the training set.

[0040] In some embodiments, referring to Table 1, historical state evolution trajectories are stored in the form of a time series database. Table 1 shows a portion of the state evaluation decision sequences recorded in the historical state evolution trajectories. The time series verification module analyzes these sequences to obtain statistical regularities and constraints.

[0041] Table 1: Example fragments of historical state evolution trajectory

[0042] In practical implementation, the compensation output module calculates the gradient of the real-time sensor signal at the current moment relative to the previous moment. The gradient is the difference vector of data from each sensor channel within a unit of time. The compensation output module analyzes the evolution pattern of the set of state feature vectors within a recent time window, identifying the dominant evolution trend. This dominant evolution trend is obtained by calculating the linear regression slope of the state feature vectors along each dimension within the time window. The compensation output module establishes a short-term state extrapolation model, starting with a validated baseline state assessment decision and using the gradient and dominant evolution trend as extrapolation parameters. The short-term state extrapolation model is run to predict possible state assessment points in the next few moments. Based on the predicted possible state assessment points, the compensation amount for the current moment's decision is calculated in reverse. The compensation amount is then superimposed with the validated baseline state assessment decision to generate an optimized state assessment decision. It can be understood that the number of future moments is a pre-set fixed value.

[0043] See Figure 4 In the short-term state extrapolation and optimization decision generation, the changing trends of the baseline state assessment decision, the optimized state assessment decision, and the smoothed optimization decision are presented, along with the distribution of compensation amounts at corresponding time points. Specifically, the horizontal axis in the figure represents the prediction time point (from the current time to t+5), the left vertical axis represents the state assessment value, and the right vertical axis represents the compensation amount. The baseline state assessment decision (purple curve) linearly decreases from the current 92.0 to 91.0 at t+5, reflecting the natural evolution of the baseline decision over time after time-series verification. The optimized state assessment decision (orange solid line), based on the baseline decision and superimposed with compensation amounts, shows a smoother downward trend (from the current 92.0 to 90.6 at t+5). The smoothed optimization decision (orange dashed line) further smooths the optimization decision over time, weakening fluctuations to improve stability. The blue area on the right represents the compensation amount, whose absolute value gradually increases as the prediction time progresses (from nearly 0.0 to -0.6). This reflects the forward extrapolation compensation logic based on the real-time sensor signal change gradient and the state feature evolution trend: the dynamic adjustment of the compensation amount is to correct the short-term deviation of the baseline decision, making the optimized decision more in line with the dominant evolution trend of the state. At the parameter configuration level, the prediction time span covers five future moments, and the compensation amount is calculated based on the weighted fusion result of the linear regression slope of the state feature vector (dominant evolution trend) and the sensor signal difference vector (change gradient).

[0044] Example 5: Based on the optimized state assessment decision, the system drives the actuator to complete the specified monitoring actions on the battery and records the corresponding assessment feedback signals. The optimized state assessment decision is parsed to obtain at least one specific monitoring action command and expected parameter range. The monitoring action command is sent to the corresponding physical actuator, which may include a load controller, thermal management controller, or data acquisition card. The physical actuator is controlled to apply electrical load, temperature environment, or initiate high-precision data acquisition on the battery according to the command. New sensor data acquired by multiple heterogeneous sensors during the execution of the monitoring action is received. Key indicators related to the expected parameter range are extracted from the newly acquired sensor data, and these key indicators are stored as assessment feedback signals in the system database.

[0045] In practical implementation, the system drives the actuators to complete specified monitoring actions on the battery based on the optimized state assessment decision, and records the corresponding assessment feedback signals. The optimized state assessment decision is parsed to obtain at least one specific monitoring action instruction and expected parameter range. The monitoring action instruction exists in the form of a structured data packet, which includes fields such as instruction type, target parameters, execution intensity, and duration. The monitoring action instruction is sent to the corresponding physical actuator, which may include a load controller, thermal management controller, or data acquisition card. Instruction transmission is completed via wired or wireless communication bus. The physical actuators are controlled to apply electrical load, temperature environment, or initiate high-precision data acquisition on the battery according to the instructions. For example, the load controller adjusts the load resistance of the battery circuit to simulate a specific discharge rate, the thermal management controller adjusts the temperature control plate in contact with the battery, and the data acquisition card switches to a higher sampling frequency mode. During the execution of the monitoring action, newly acquired sensor data from multiple heterogeneous sensors is received. The newly acquired sensor data is time-stamped with the action execution stage during transmission. Key indicators related to the expected parameter range are extracted from the newly acquired sensor data. The extraction method for these key indicators depends on the command type and may involve calculating the voltage plateau hold-up, average temperature rise rate, or the integral area of ​​the current impulse response curve over a specific time period. These key indicators are stored as evaluation feedback signals in the system database. The evaluation feedback signals, along with the optimization state evaluation decisions that triggered them, the monitoring action commands, and their parameters, form a linked record.

[0046] The calculation of the evaluation feedback signal may require fusing data from multiple sensors, and its formula is expressed as follows:

[0047] in: The scalar value representing the calculated evaluation feedback signal. This represents the number of sensor channels involved in the calculation of the evaluation feedback signal. and These represent the start and end times of the monitored action command execution, respectively. Representing the Each sensor channel at time The newly acquired sensor data, after preprocessing, is integrated to quantify the cumulative effect of the sensor signal over the entire monitoring process. This integration is implemented using numerical integration methods in discrete data systems. Optionally, the storage format for the evaluation feedback signal includes a timestamp, signal value, signal type identifier, and associated decision and action identifiers.

[0048] In practical implementation, the expected parameter range of the monitored action command defines the theoretically reasonable range of the evaluation feedback signal. It can be understood that the system database establishes a dedicated storage table structure for the evaluation feedback signal and its associated information. In some embodiments, when the evaluation feedback signal exceeds the expected parameter range, a low-level warning flag is triggered, which is associated with the original optimization state evaluation decision record. Optionally, after completing the command action, the physical actuator sends an action completion confirmation signal to the system master control unit. Upon receiving the confirmation signal, the system master control unit then begins receiving and processing newly acquired sensor data to generate the evaluation feedback signal.

[0049] See Figure 5 In the confidence analysis of parallel inference branches for battery state monitoring, the heatmap visually presents the distribution characteristics of decision confidence of the four parallel inference branches at different monitoring stages. Specifically, branch 1 (voltage-dominated) has a confidence of 0.98 in the initial monitoring stage, making it the core decision branch in this stage. However, as the process progresses (load adjustment, temperature control adjustment, etc.), its confidence decreases in a stepwise manner. Branch 2 (temperature-dominated) has a confidence of 0.98 in the temperature control adjustment stage, reflecting the decision-making dominance of the temperature parameter in this stage. The confidence of branch 3 (current-dominated) maintains a relatively stable and slow downward trend throughout the entire stage, indicating that the decision contribution of the current parameter decreases steadily as the process progresses. The confidence of branch 4 (comprehensive features) remains in a high range (0.93-0.97), reflecting the stable decision support role of multi-feature fusion inference in each stage. The distribution pattern of confidence levels closely aligns with the phased requirements of battery state monitoring: the stability of the voltage signal in the initial phase makes it a high-confidence decision-making basis; the dynamic variability of temperature parameters in the temperature control phase strengthens the decision weight of the corresponding branch; and the high confidence of the comprehensive feature branch reflects the inference strategy of "dynamic adaptation of single parameters and backup of multi-parameter fusion" in the multimodal sensing system. The color gradient of the heatmap (from dark to light corresponding to 0.98-0.88) clearly quantifies the differences in decision reliability of each branch at different stages, providing a visual basis for the dynamic weight allocation of parallel inference branches.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An integrated battery state monitoring and evaluation system based on multimodal sensing, characterized in that, The system includes: The data acquisition module is used to acquire raw sensing signals synchronously generated by multiple heterogeneous sensors deployed on the battery entity; The data processing module is used to perform spatiotemporal fusion and cleaning on the original sensing signals to form a standardized fused sensing data sequence. The feature mapping module is used to perform feature deconstruction and state mapping on the fused sensing data sequence based on a preset battery state classification map, and generate a set of state feature vectors with time stamps. The network adaptation module is used to dynamically adapt the prior state evaluation decision network using the set of state feature vectors to generate the adapted state evaluation decision network. The evaluation and reasoning module is used to input the fused sensing data sequence and the set of state feature vectors into the adapted state evaluation and decision network, perform parallel evaluation and reasoning, and output a preliminary state evaluation and decision set. The timing verification module is used to incorporate historical state evolution trajectories to perform timing consistency verification and correction on the preliminary state assessment decision set, forming a verified baseline state assessment decision. The compensation output module is used to perform forward extrapolation compensation on the verified baseline state evaluation decision based on the changing gradient of the real-time sensing signal and the evolution pattern of the state feature vector set, and generate an optimized state evaluation decision.

2. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 1, characterized in that, The process of spatiotemporal fusion and cleaning of the original sensing signals to form a standardized fused sensing data sequence includes: Receives asynchronous raw data streams from temperature, voltage, and current sensors; Apply a uniform timestamp to each of the asynchronous raw data streams; Based on the unified timestamp marker, all asynchronous raw data streams are time-aligned and interpolated to generate time-synchronized multi-channel data streams. Anomaly detection and removal are performed on each channel of the time-synchronized multi-channel data stream; The data from each channel after the outlier removal was performed to normalize the dimensions. The multi-channel data that has undergone dimension normalization is packaged in chronological order to form the standardized fused sensing data sequence.

3. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 1, characterized in that, The step involves deconstructing and mapping the fused sensor data sequence based on a preset battery state classification map to generate a set of time-stamped state feature vectors, including: Load the preset battery state classification map, which defines the correspondence rules between various battery states and sensor data patterns; A data window of a fixed time length is slidably extracted from the fused sensing data sequence; For each data window, statistical features, frequency domain features, and trend features within the window are extracted according to the corresponding rules. All features extracted for a data window are concatenated to form a state feature vector, and the state feature vector is marked with the central time of the corresponding data window. The fused sensing data sequence is traversed to generate a set of time-stamped state feature vectors corresponding to each data window.

4. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 3, characterized in that, The step of dynamically adapting the prior state evaluation decision network using the set of state feature vectors to generate an adapted state evaluation decision network includes: Invoke the pre-trained prior state evaluation decision network, which contains updatable network parameters; Select a subset of recently generated vectors from the set of state feature vectors as the adaptation sample set; Using the evaluation feedback signal recorded at the previous moment as a supervision signal, the difference between the output generated by the adaptation sample set through the prior state evaluation decision network and the supervision signal is calculated; Based on the differences, the updatable network parameters are fine-tuned using a backpropagation mechanism. When the difference is less than a preset threshold, fine-tuning stops, and the network at this point is saved as the adapted state evaluation decision network.

5. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 1, characterized in that, The process of inputting the fused sensing data sequence and the set of state feature vectors into the adapted state evaluation decision network, performing parallel evaluation inference, and outputting a preliminary state evaluation decision set includes: The fused sensing data sequence is divided into several logical data segments; Pair the state feature vector corresponding to each logical data segment in the state feature vector set with the logical data segment; Each pair of paired data is simultaneously input into multiple parallel inference branches of the adapted state evaluation decision network; Each parallel inference branch independently processes the input data pair and outputs a candidate state evaluation decision; The candidate state evaluation decisions output by all parallel inference branches are aggregated to form the preliminary state evaluation decision set.

6. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 1, characterized in that, The process of introducing historical state evolution trajectories to perform temporal consistency verification and correction on the preliminary state assessment decision set, forming a verified baseline state assessment decision, includes: The historical state evolution trajectory is retrieved from the system database. The historical state evolution trajectory records the sequence of changes in battery state assessment decisions over time. Analyze the statistical regularities and constraints of state changes in the historical state evolution trajectory; Each candidate decision in the preliminary state assessment decision set is placed at the end of the historical state evolution trajectory, and simulated as the next state. Determine whether the next state obtained from the simulation violates the stated statistical laws and constraints; If the conditions are violated, the candidate decision shall be modified according to the statistical regularity and constraints described above; After all candidate decisions have been verified and possibly corrected, the decision with the highest confidence level is selected as the verified baseline state evaluation decision.

7. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 6, characterized in that, The method of forward extrapolating and compensating the verified baseline state assessment decision based on the changing gradient of the real-time sensing signal and the evolution pattern of the state feature vector set to generate an optimized state assessment decision includes: Calculate the gradient of the real-time sensing signal at the current moment relative to the previous moment; Analyze the evolution pattern of the state feature vector set within a recent time window to identify the dominant evolution trend; A short-term state extrapolation model is established, starting with the verified baseline state assessment decision and using the change gradient and the dominant evolution trend as extrapolation parameters. Run the short-term state inference model to predict possible state assessment points at several future time points; Based on the predicted possible state assessment points, the compensation amount for the decision at the current moment is calculated in reverse. The compensation amount is superimposed on the verified baseline state evaluation decision to generate the optimized state evaluation decision.

8. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 7, characterized in that, The establishment of a short-term state extrapolation model, starting with the verified baseline state assessment decision and using the change gradient and the dominant evolution trend as extrapolation parameters, includes: The verified baseline state evaluation decision is used as the initial state vector for deduction, and the state space for a finite number of consecutive future time steps is defined. Set a fixed time step and a total number of steps for the forward inference, where the total number of steps defines the time span of the inference; The instantaneous change term of the state transition is constructed based on the change gradient, and the trend term of the state transition is constructed based on the dominant evolution trend; The instantaneous change term and the trend term are weighted and fused to form a single-step state transition function; Starting with the initial state vector of the deduction, and with the single-step state transition function as the core, the predicted state vector of each subsequent time step is iteratively calculated to obtain a predicted state sequence. The predicted state sequence is smoothed, and the overall trend information of state changes is extracted from it as the inference output. Based on the predicted trend described in the inference output, a compensation amount for adjusting the current decision is calculated in reverse. The compensation amount is superimposed and fused with the verified baseline state evaluation decision to generate the final optimized state evaluation decision for output.

9. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 1, characterized in that, The construction and updating of the preset battery state classification map are achieved through the following steps: Collect a large amount of multimodal sensing data and corresponding known state labels of the same type of battery throughout its entire life cycle; Cluster analysis was performed on the collected data to identify the typical characteristics and boundaries of sensor data patterns under different state labels; Based on the cluster analysis results, a state region partitioning diagram with key sensing features as coordinate axes is drawn. The state region division map, the typical feature descriptions of each region, and the transition probability relationships between state labels are encoded into a structured knowledge base to form an initial battery state classification map. During system operation, when the evaluation feedback signal deviates significantly from the prediction made based on the battery state classification map, the map update mechanism is triggered to iteratively correct the boundary and transition probability relationship in the battery state classification map using newly generated data.

10. The integrated battery state monitoring and evaluation system based on multimodal sensing as described in claim 1, characterized in that, The system further includes: driving the actuator to complete specified monitoring actions on the battery based on the optimized state assessment decision, and recording corresponding assessment feedback signals, including: The optimized state assessment decision is analyzed to obtain at least one specific monitoring action instruction and expected parameter range; The monitoring action command is sent to the corresponding physical actuator, which includes a load controller, a thermal management controller, or a data acquisition card. The physical actuator is controlled to apply specific electrical loads or temperature environments to the battery or initiate high-precision data acquisition according to instructions. Receive newly acquired sensing data from the plurality of heterogeneous sensors during the execution of the monitoring action; Extract key indicators related to the expected parameter range from the newly acquired sensor data, and store the key indicators as the evaluation feedback signal in the system database.

Images