Storage battery remote capacity checking system and capacity checking method

Abstract

The invention discloses a storage battery remote capacity checking system and method, and the method comprises the steps: controlling a storage battery to discharge when capacity checking equipment in the capacity checking system receives a capacity checking instruction, collecting multi-modal time sequence data, and transmitting the multi-modal time sequence data to a cloud platform, and enabling the cloud platform to upload the multi-modal time sequence data to a block chain; and inputting the multi-modal time series data into a hybrid learning model comprising a feature extraction network and a gradient boosting decision tree sub-model to obtain estimated capacity and confidence, and uploading the estimated capacity and confidence to a block chain, on one hand, the multi-modal time series data and a capacity checking result are uploaded to the block chain, so that the capacity checking data and the result can be prevented from being tampered; on the one hand, an effective anti-counterfeiting and tracing mechanism is provided, auditing and compliance requirements can be met, on the other hand, a mixed learning model of a feature extraction network and a gradient boosting decision tree sub-model is adopted, accurate features can be extracted for various complex working conditions, the stability and precision of capacity prediction can be improved through the gradient boosting decision tree sub-model, and the accuracy of capacity prediction is improved. And the generalization ability is better.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery capacity detection, and particularly relates to a storage battery remote capacity checking system and a capacity checking method. BACKGROUND

[0002] The storage battery is the core backup energy of the communication power supply system, and its capacity state is directly related to the reliability of the communication network. Regularly checking the capacity of the storage battery is an important means to evaluate the health of the storage battery.

[0003] The prior art checks the capacity of the remote storage battery through a remote capacity checking system. Specifically, the remote capacity checking system sends a capacity checking instruction to the remote storage battery, collects original data such as voltage, current, internal resistance and temperature during the capacity checking process, and inputs the original data into a machine learning model to estimate the capacity of the storage battery.

[0004] However, the prior art has obvious defects:

[0005] (1) The data collected during the capacity checking process and the capacity checking result are easy to be tampered with or lost, and there is a lack of effective anti-fake and traceability mechanism, which is difficult to meet the increasingly strict audit and compliance requirements;

[0006] (2) The machine learning model is a single machine learning model, and when facing complex working conditions such as temperature changes and load fluctuations, the stability and accuracy of the capacity prediction are significantly reduced, and the generalization ability is insufficient. SUMMARY

[0007] The present application provides a storage battery remote capacity checking system and a capacity checking method to ensure the credibility of the capacity checking data and improve the stability and accuracy of the capacity prediction.

[0008] In a first aspect, the present application provides a storage battery remote capacity checking system, comprising a capacity checking device and a cloud platform in communication with the capacity checking device, the capacity checking device being electrically connected with a storage battery;

[0009] The capacity checking device is configured to control the storage battery to discharge when receiving a capacity checking instruction, and collect multi-modal time series data during the discharging process, and send the multi-modal time series data to the cloud platform.

[0010] The cloud platform is configured to receive the multi-modal time series data, upload the multi-modal time series data to a blockchain, and input the multi-modal time series data into a hybrid learning model, so as to extract a feature vector from the multi-modal time series data through a feature extraction network in the hybrid learning model, and perform regression prediction on the feature vector through a gradient boosting decision tree sub-model to obtain an estimated capacity and a confidence of the storage battery, and upload the estimated capacity and the confidence to the blockchain.

[0011] Optionally, the nuclear capacity device comprises a power supply, a programmable load, a multi-modal sensing module, a controller and a communication module, the programmable load, the multi-modal sensing module and the communication module are connected with the controller, and the programmable load is connected with the battery;

[0012] The power supply is used to power the multi-modal sensing module, the controller and the communication module;

[0013] The controller is used to control the programmable load to start when receiving the nuclear capacity instruction, and control the battery to discharge the programmable load according to a preset discharge curve;

[0014] The multi-modal sensing module is used to collect multi-dimensional raw data in the battery discharge process and send the multi-dimensional raw data to the controller;

[0015] The controller is used to preprocess the multi-dimensional raw data, and send the multi-modal time series data obtained after preprocessing to the cloud platform through the communication module.

[0016] Optionally, the multi-modal sensing module is specifically used to collect battery voltage, discharge current, battery internal resistance, battery temperature, environmental temperature and environmental humidity in the battery discharge process to obtain multi-dimensional raw data according to a preset frequency;

[0017] The controller is specifically used to clean and normalize the multi-dimensional raw data, and divide the normalized data into a sliding window to obtain multi-modal time series data.

[0018] Optionally, the nuclear capacity device further comprises a distributed ledger module, which is used to calculate a hash value of the multi-modal time series data;

[0019] The controller is used to send the multi-modal time series data and the hash value to the cloud platform through the communication module.

[0020] Optionally, the cloud platform comprises a hybrid learning model and a distributed ledger;

[0021] The distributed ledger is used to write the multi-modal time series data and the hash value through a blockchain consensus mechanism;

[0022] The hybrid learning model comprises a feature extraction network and a gradient boosting decision tree submodel, and the feature extraction network is used to extract a feature vector from the multi-modal time series data;

[0023] The gradient boosting decision tree submodel is used to regress and predict the feature vector to obtain an estimated capacity and a confidence of the battery.

[0024] Optionally, the feature extraction network comprises an input layer, a plurality of convolutional pooling structures connected in sequence, a feature flattening layer, and a fully connected layer.

[0025] The input layer is configured to receive the multi-modal time series data.

[0026] The convolutional pooling structure comprises a one-dimensional convolutional layer, an activation function, and a pooling layer. In each convolutional pooling layer, the one-dimensional convolutional layer is configured to perform one-dimensional convolutional processing on the input data using different scales of one-dimensional convolutional kernels to obtain a first feature map.

[0027] The activation function is configured to perform nonlinear transformation on the first feature map to obtain a second feature map.

[0028] The pooling layer is configured to perform pooling processing on the second feature map to output a third feature map.

[0029] The feature flattening layer is configured to flatten the multi-dimensional feature map output by the convolutional pooling structure into a one-dimensional feature vector.

[0030] The fully connected layer is configured to globally integrate and nonlinearly transform the one-dimensional feature vector to obtain a comprehensive feature vector.

[0031] Optionally, the cloud platform further comprises a model optimization module configured to retrain the hybrid learning model using the multi-modal time series data and the actual capacity of the battery.

[0032] Optionally, the cloud platform is further configured to send the confidence to the core capacity device.

[0033] The core capacity device is further configured to adjust the programmable load based on the confidence.

[0034] Optionally, the cloud platform is further configured to generate a core capacity report using the multi-modal time series data, the estimated capacity, the confidence, and the transaction ID returned by the blockchain, and upload the core capacity report to the blockchain.

[0035] In a second aspect, the embodiments of the present application provide a battery remote core capacity method, applied to the battery remote core capacity system of any one of the first aspect, comprising:

[0036] Discharge the battery when receiving a core capacity instruction.

[0037] Collect multi-modal time series data during discharging.

[0038] Upload the multi-modal time series data to the blockchain.

[0039] Input the multi-modal time series data into a hybrid learning model.

[0040] The feature extraction network extracts a feature vector from the multi-modal time series data in the mixed learning model, and the gradient boosting decision tree sub-model performs regression prediction on the feature vector to obtain the estimated capacity and confidence of the battery, and the estimated capacity and the confidence are uploaded to the blockchain.

[0041] The battery remote capacity determination system includes a capacity determination device and a cloud platform in communication with the capacity determination device. The capacity determination device is electrically connected to the battery. When receiving a capacity determination instruction, the capacity determination device controls the battery to discharge and collects multi-modal time series data and sends the data to the cloud platform. The cloud platform uploads the multi-modal time series data to the blockchain and inputs the data into a mixed learning model including a feature extraction network and a gradient boosting decision tree sub-model to obtain the estimated capacity and confidence of the battery and uploads them to the blockchain. On the one hand, the multi-modal time series data and the capacity determination result are uploaded to the blockchain, which can prevent the capacity determination data and the capacity determination result from being tampered with or lost, providing an effective anti-fake and traceability mechanism that can meet increasingly stringent audit and compliance requirements. On the other hand, the mixed learning model using the feature extraction network and the gradient boosting decision tree sub-model can extract accurate features in various complex working conditions, and the gradient boosting decision tree sub-model can improve the stability and accuracy of capacity prediction and has better generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A schematic diagram of a battery remote capacity determination system provided by an embodiment of the present application is shown in the figure.

[0043] Figure 2 A schematic diagram of a battery remote capacity determination system provided by another embodiment of the present application is shown in the figure.

[0044] Figure 3 A schematic diagram of a mixed learning model provided by an embodiment of the present application is shown in the figure.

[0045] Figure 4 A process diagram of a battery capacity determination example in an embodiment of the present application is shown in the figure.

[0046] Figure 5 A flowchart of a battery remote capacity determination method provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0047] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present application, but not to limit the present application. In addition, it should be noted that only parts related to the present application are shown in the drawings for ease of description.

[0048] Figure 1 A structural schematic diagram of a battery remote capacity determination system provided by an embodiment of the present application is shown in the figure.Figure 1 As shown, the battery remote state-of-health system of the embodiment of the present application is used for remotely controlling the battery in the communication network to perform state-of-health, specifically, the battery remote state-of-health system comprises a state-of-health device and a cloud platform in communication with the state-of-health device, the state-of-health device is deployed at the battery end and electrically connected with the battery, wherein the state-of-health device can be a device for controlling the discharge of the battery, collecting data during discharge and interacting with the cloud platform, the cloud platform is used for receiving the data collected by the state-of-health device and estimating the capacity and confidence of the battery, and multiple state-of-health devices can communicate with the same cloud platform.

[0049] Specifically, the state-of-health device is used for controlling the discharge of the battery when receiving a state-of-health instruction, and collecting multi-modal time series data during the discharge process, and sending the multi-modal time series data to the cloud platform, for example, the state-of-health instruction can be sent to the state-of-health device through the cloud platform, or the state-of-health device generates the state-of-health instruction when detecting a preset event, for example, the state-of-health device generates the state-of-health instruction when detecting a preset time, and the embodiment does not limit the event triggering the generation of the state-of-health instruction.

[0050] The multi-modal time series data can be data such as battery voltage sequence, discharge current sequence, battery internal resistance sequence, battery temperature sequence, etc. collected over time during the discharge of the battery.

[0051] The cloud platform is used for receiving the multi-modal time series data, uploading the multi-modal time series data to the blockchain, and inputting the multi-modal time series data into the hybrid learning model, so as to extract a feature vector from the multi-modal time series data through a feature extraction network in the hybrid learning model, and obtain an estimated capacity and confidence of the battery through a gradient boosting decision tree sub-model for regression prediction of the feature vector, and upload the estimated capacity and confidence to the blockchain.

[0052] Specifically, the cloud platform can be connected with the blockchain network, when receiving the multi-modal time series data collected by the battery capacity determination device, the multi-modal time series data can be chained in the blockchain network through the consensus mechanism of the blockchain to prevent the multi-modal time series data from being tampered with, in addition, the cloud platform can deploy a hybrid learning model including a feature extraction network and a gradient boosting decision tree (GBDT, Gradient Boosting Decision Tree) sub-model, wherein the feature extraction network can be a CNN (Convolutional Neural Network), the feature extraction network first extracts feature maps of each multi-modal time series data by using one-dimensional convolution layers of different scales, then the feature maps are nonlinearly changed by an activation function and are pooled by a pooling layer, after repeatedly passing through the convolution and pooling operations, the features of each multi-modal time series data are flattened and connected to obtain a comprehensive feature vector, the comprehensive feature vector is input into the gradient boosting decision tree sub-model for decision tree regression to obtain the estimated capacity and confidence of the battery, and the estimated capacity and confidence are uploaded to the blockchain.

[0053] The battery remote capacity determination system of the present application can upload multi-modal time series data and capacity determination results to the blockchain, which can prevent the capacity determination data and capacity determination results from being tampered with or lost, and provide an effective anti-fake and traceability mechanism, which can meet the increasingly stringent audit and compliance requirements, on the other hand, the hybrid learning model using the feature extraction network and the gradient boosting decision tree sub-model can extract accurate features in various complex working conditions, and the gradient boosting decision tree sub-model can improve the stability and accuracy of capacity prediction and has better generalization ability.

[0054] Figure 2 Fig. 1 shows a schematic diagram of a battery remote capacity determination system according to another embodiment of the present application, as shown in Figure 2 As shown, the capacity determination device includes a power supply, a programmable load, a multi-modal sensing module, a controller and a communication module, the programmable load, the multi-modal sensing module and the communication module are connected with the controller, and the programmable load is connected with the battery.

[0055] The power supply is used to supply power to the multi-mode perception module, the controller and the communication module. Specifically, the power supply can be a power supply that takes mains or a storage battery as input and outputs direct current required by the system after a DC-DC module. The power supply can include an AC-DC module connected to the mains to convert alternating mains power into a basic direct current power source for the system to work. It also includes a DC-DC module to switch to a storage battery when the mains power is off or fails, and converts the power taken from the storage battery through the DC-DC module into direct current required by the system. The power supply has a DC-DC module with boost / buck function to support remote control of discharge voltage and current, meet the capacity requirements of different voltage level (such as 24V, 36V, 48V, etc.) battery packs, and provide discharge energy loop for programmable load.

[0056] The programmable load can be composed of a high-precision adjustable resistance network and a power inductor, controlled by the controller to realize three configurable discharge modes of constant current (CC), constant power (CP) and step, and can automatically match the discharge curve required by the capacity standard of IEC, GB / T, etc.

[0057] The multi-mode perception module can integrate various sensors, such as current, power, temperature, etc. sensors, for collecting battery voltage, discharge current, battery internal resistance, battery temperature, ambient temperature and ambient humidity during the discharge process of the storage battery to obtain multi-dimensional data according to a preset sampling frequency.

[0058] The controller can be various models of processors or MCUs. The controller is used to control the programmable load to start when receiving the capacity instruction, and control the storage battery to discharge the programmable load according to the preset discharge curve. The multi-mode perception module is used to collect multi-dimensional raw data during the discharge process of the storage battery and send it to the controller. The controller is used to preprocess the multi-dimensional raw data, and send the multi-modal time series data obtained after preprocessing to the cloud platform through the communication module.

[0059] Specifically, during the discharge process of the storage battery, the multi-mode perception module is specifically used to collect battery voltage, discharge current, battery internal resistance, battery temperature, ambient temperature and ambient humidity of the storage battery to obtain multi-dimensional raw data according to a preset frequency. For example, the voltage between the positive and negative terminals of the storage battery can be collected by a voltage sensor as the battery voltage, the current of the positive terminal of the storage battery can be measured by a Hall sensor as the discharge current, the internal resistance of the battery can be tested by an internal resistance testing unit (such as an internal resistance testing unit implementing AC injection method or DC pulse method), and the surface temperature of the battery can be measured by a temperature sensor, and the temperature and humidity of the environment where the storage battery is located can be measured by a temperature and humidity sensor. The sensor data collected by the multi-mode perception module is output to the controller after ADC conversion. The controller preprocesses the multi-dimensional raw data, and divides the normalized data into a sliding window to obtain multi-modal time series data.

[0060] For example, the controller receives raw time-series data streams collected from the multi-mode sensing module, which may include multi-dimensional sequence data such as battery voltage, discharge current, battery internal resistance, and temperature. For each dimension of sequence data, outlier removal and missing value filling can be performed. For example, based on statistical principles (such as the Z-score method) or by setting physical thresholds (such as voltage should not be negative), illegal data points caused by sensor noise or interference in the multi-dimensional sequence data can be identified and removed. For a small amount of missing data caused by transmission packet loss, a linear interpolation algorithm of neighboring data points can be used to fill the missing data, ensuring the continuity of the time sequence.

[0061] Furthermore, the Min-Max normalization algorithm can be used to linearly map each dimension of the sequence data (such as voltage, current, and temperature) with different dimensions and numerical ranges to the interval [0, 1]. The expression for the Min-Max normalization algorithm is as follows:

[0062] X norm =(X−X min ) / (X max -X min );

[0063] Among them, X norm X is the normalized value of the sequence data for each dimension, where X is the original value of the sequence data for each dimension. min It is the minimum value of the sequence data in each dimension, X. max It is the maximum value of the sequence data in each dimension.

[0064] Normalization can eliminate the magnitude difference between data of different dimensions, prevent certain features from dominating the hybrid learning model due to their large values, accelerate the convergence of the subsequent hybrid learning model, and improve the efficiency of capacity estimation.

[0065] After normalizing the time series data for each dimension, it can be segmented according to a preset fixed window length L and sliding step size S to generate a series of continuous, fixed-length data sample segments. This transforms the continuous kernel process into multiple independent analysis samples that can be processed by the hybrid learning model, thereby capturing the local features and dynamic changes of the data in the time dimension.

[0066] like Figure 2As shown, the nuclear capacity device further includes a distributed ledger module for calculating a hash value of the multi-modal time series data, and the controller is configured to send the multi-modal time series data and the hash value to the cloud platform through the communication module. The distributed ledger module at the nuclear capacity device operates as a lightweight node of a consortium chain or a side chain. The hash value of the multi-modal time series data calculated at the nuclear capacity device is uploaded to the cloud platform, which can prevent the collected data from being tampered with during collection and transmission, and provide a trusted certificate for data traceability.

[0067] In one embodiment, the communication module can be a mainstream low-power wireless communication module such as NB-IoT, LTE-Cat-M1, 5G, LoRaWAN, etc. It has network environment perception capability and can automatically select the optimal communication mode according to the field signal strength and quality to ensure reliable data transmission.

[0068] As shown in Figure 2 As shown, the cloud platform includes a hybrid learning model and a distributed ledger. The distributed ledger is configured to write the multi-modal time series data and the hash value to the chain through a blockchain consensus mechanism. The hybrid learning model includes a feature extraction network and a gradient boosting decision tree sub-model. The feature extraction network is configured to extract a feature vector from the multi-modal time series data. The gradient boosting decision tree sub-model is configured to perform regression prediction on the feature vector to obtain an estimated capacity and a confidence level of the battery.

[0069] In one embodiment, the feature extraction network can be a CNN, as shown in Figure 3 As shown, the feature extraction network includes an input layer, a plurality of sequentially connected convolutional pooling structures, a feature flattening layer, and a fully connected layer. The input layer is configured to receive the multi-modal time series data. The convolutional pooling structure includes a one-dimensional convolutional layer, an activation function, and a pooling layer. In each convolutional pooling layer, the one-dimensional convolutional layer in different convolutional pooling structures is configured to perform convolutional processing on the input data using one-dimensional convolutional kernels of different scales to obtain a first feature map. The activation function is configured to perform nonlinear transformation on the first feature map to obtain a second feature map. The pooling layer is configured to perform pooling processing on the second feature map to output a third feature map. The feature flattening layer is configured to flatten the multi-dimensional feature map output by the convolutional pooling structure into a one-dimensional feature vector. The fully connected layer is configured to globally integrate and nonlinearly transform the one-dimensional feature vector to obtain a comprehensive feature vector.

[0070] Specifically, in each convolutional pooling structure, the convolutional layer (Conv1D) uses one-dimensional convolutional kernels of different scales to slide over the time dimension of each dimension of time series data, extracting local time series features of the dimension of time series data, such as extracting features at different time scales. For example, a smaller size convolutional kernel (such as 3-5 time points) can be used for time sliding window to calculate the mean, variance, maximum value, minimum value, etc. of the data points in the time slot to extract short-term, high-frequency local features, such as:

[0071] For the voltage sequence, the instantaneous drop or spike of voltage can be extracted, reflecting the dynamic response when the load is suddenly added or reduced;

[0072] For the current sequence, the fluctuation details and small pulses of current can be extracted, revealing the instantaneous changes during the charging and discharging process of the battery;

[0073] For the temperature sequence, the short-term fluctuation pattern of temperature can be extracted, reflecting the rapid influence of internal chemical reactions or external environment of the battery.

[0074] Then use a larger size of convolution kernel (such as 10-20 time points) to extract long-term, trend macro features, such as:

[0075] For the voltage sequence, the stable downward trend of voltage in the entire discharge period can be extracted, which is the core basis for evaluating the capacity of the battery;

[0076] For the current sequence, the smoothness of current in the constant current discharge stage, or the platform switching characteristics in the step discharge can be extracted;

[0077] For the temperature sequence, the overall rising slope of the battery temperature during the entire discharge process can be extracted, which is a key indicator for evaluating the internal resistance and health status of the battery.

[0078] After the convolution operation by the convolution layer, the output is no longer time series data, but is converted into a series of feature maps, that is, in one-dimensional convolution operation, each one-dimensional convolution kernel scans the time series data and outputs a one-dimensional vector representing the activation intensity of the pattern (such as voltage slow drop, current pulse, etc.) corresponding to the convolution kernel in the entire sequence. The one-dimensional vector constitutes a feature map, and each feature map represents the response intensity of a certain scale of convolution kernel on the entire time series sequence. The feature maps together constitute the high-level, abstract feature encoding of the time series data, capturing the most effective, cross-dimensional time series rules for battery capacity prediction, and providing extremely discriminative input for the subsequent GBDT model.

[0079] After the convolution layer performs convolution operation on each dimension of the time series data to obtain a feature map, the activation function f(x) = max(0, x) performs nonlinear transformation on the feature map. Specifically, the activation function f(x) independently judges and maps each value in the feature map output by the convolution layer. If the input value x > 0, the output is x (directly passed through), and if the input value x≤0, the output is 0 (zeroed). The nonlinear transformation of the feature map by the activation function enables the feature extraction network to approximate any complex function through the stacking of layers, thereby fitting the nonlinear variation law in the voltage and current curves (such as the slow decline of the voltage on the battery discharge platform and the sudden drop of the voltage at the end of the period), improving the expression ability of the hybrid learning model. In addition, during the training process, when the input is positive, the derivative of the activation function is always 1, ensuring that the error does not decay when passing through the activation function during backpropagation, solving the gradient vanishing problem and effectively passing to deeper network layers, so that the deep network (such as a multi-layer CNN) can be effectively trained to learn deeper features.

[0080] After the activation function, the second feature map is output. The pooling layer is used to perform pooling processing on the second feature map. Specifically, the pooling layer reduces the dimension and enhances the features of the feature map output by the previous convolution layer. Specifically, the feature map is down-sampled to retain significant features and reduce the number of parameters. In one example, the pooling layer uses a maximum value pooling algorithm, which scans the feature map through a fixed-length sliding window (such as a length of 2). Only the feature point with the maximum activation value is retained in each window, and the remaining nodes are suppressed. The strong-weak selection pooling mechanism effectively retains the most significant feature response. For example, as long as there is a strong voltage drop feature signal in a local time period, the signal can be accurately transmitted to the next layer, thereby filtering out background noise and minor fluctuations. At the same time, the down-sampling operation directly reduces the dimension and total amount of data of the feature map (for example, the sequence length is halved), significantly reducing the number of parameters that need to be learned by the subsequent network layer (such as a fully connected layer). This not only reduces the computational complexity and hardware resource consumption, but also helps to suppress overfitting of the hybrid learning model and improve the generalization ability.

[0081] The sliding window scanning of the pooling operation refers to the shift of a specific pattern (such as an internal resistance peak) before and after the time axis in the time series context. The maximum value pooling assigns invariance to the hybrid learning model by focusing on the strongest signal in the local area rather than its exact position. As long as the key feature appears within the pooling window, the output after pooling remains stable regardless of the specific time point, making the hybrid learning model insensitive to timing deviations caused by sampling jitter or minor fluctuations in working conditions during core capacity processes. This significantly improves the robustness and reliability of the hybrid learning model in different core capacity tasks.

[0082] As Figure 3As shown, the hybrid learning model adopts a hierarchical design, and a feature extraction channel from local details to global state is constructed by stacking multiple convolutional pooling structures. The input of each group of convolutional pooling structures is the down-sampled feature map output by the previous level of the pooling layer. The benefits of layer-by-layer processing are as follows:

[0083] (1) Hierarchical feature abstraction is realized:

[0084] With the deepening of the network level, the feature representation undergoes an evolution process from local details to combined patterns and then to global state. The specific embodiment in the kernel capacity scenario is as follows (taking three convolutional pooling structures as an example):

[0085] Shallow features (features output by the first convolutional pooling structure): The first layer of convolutional kernels extracts basic local time sequence features from the original time sequence data, for example, instantaneous voltage drop, microsecond-level current pulse, and short-time temperature fluctuation features, which reflect the microscopic dynamics of load mutation, sensor noise, or battery internal chemical reaction.

[0086] Middle-level features (features output by the second convolutional pooling structure): The convolutional kernels of this layer combine multiple shallow features into more discriminative composite features on a wider time scale, for example, multiple small voltage drops are combined to identify a stable discharge platform lasting for several seconds, or the current and temperature changes are associated to identify the typical collaborative features of the constant power discharge phase. This level of features directly reflects the steady-state performance of the battery under different discharge standards.

[0087] Deep features (features output by the third convolutional pooling structure): Further integrate all middle-level features to form global and decisive high-level feature representation, for example, the overall shape and decay slope of the entire discharge curve, and the collaborative evolution relationship between voltage and internal resistance in the entire process. These features macroscopically encode the state of health (SOH) and the remaining capacity (SOC) of the battery, and can effectively distinguish between normal aging and potential faults such as internal resistance increase and active material shedding.

[0088] (2) Improve model performance and robustness:

[0089] Through multiple convolutional pooling structures, each layer can learn more complex patterns based on the features extracted by the previous layer, thereby realizing the step-by-step refinement and abstraction of features. The translation invariance brought by the pooling layer makes the hybrid learning model not sensitive to time sequence jitter or slight changes in working conditions during the kernel capacity process, significantly enhancing the generalization ability and robustness of the hybrid learning model in actual complex applications.

[0090] In this embodiment, as shown in FIG. 1, the hybrid learning model is composed of three convolutional pooling structures, and each convolutional pooling structure is composed of a convolutional layer and a pooling layer. Figure 3As shown, the feature flattening layer is located between the last convolutional pooling structure and the fully connected layer, and undertakes the key data structure conversion function. The feature flattening layer reconstructs the multi-dimensional feature tensor output by the pooling layer into a one-dimensional feature vector suitable for processing by the fully connected layer. The flattening operation can refer to merging the features of multiple dimensions after a specified dimension (such as the first dimension) into a continuous one-dimensional vector starting from the specified dimension. The one-dimensional feature vector after flattening represents:

[0091] (1) Transient response feature: representing the transient fluctuation mode of voltage / current, reflecting the response characteristics of the battery to load mutation and the transient change of internal impedance;

[0092] (2) Steady-state output feature: corresponding to the voltage platform stability in the constant current discharge phase, reflecting the ability of the battery to maintain the rated output, and directly related to the utilization efficiency of active materials;

[0093] (3) Capacity attenuation feature: encoding the attenuation pattern of the voltage curve during discharge, revealing the nonlinear variation law of capacity with discharge depth

[0094] (4) Thermoelectric coupling feature: recording the cooperative variation law of temperature-voltage-current, representing the interaction between the internal electrochemical process and the thermodynamic characteristics of the battery

[0095] (5) Abnormal diagnosis feature: capturing the distortion pattern in the voltage curve, corresponding to the macroscopic manifestations of defects such as dendrite growth and separator aging in the battery.

[0096] The one-dimensional feature vector after flattening establishes a complete mapping from external characteristics to internal state by fusing physical process information of multiple time scales, providing a comprehensive judgment basis for capacity prediction beyond traditional parameter analysis.

[0097] The fully connected layer performs global integration and nonlinear transformation on the one-dimensional feature vector output by the flattening layer, with the specific process as follows:

[0098] First, feature integration mechanism:

[0099] Through a learnable weight matrix, all elements in the input one-dimensional feature vector are combined by weighting. Each neuron receives all inputs from the previous layer, realizing cross-channel and cross-time feature interaction, and performing nonlinear transformation to mine deep correlation relationships between features;

[0100] Second, hierarchical integration of local features:

[0101] Each dimension feature of the flattened layer output corresponds to a time sequence feature (such as voltage drop and current pulse) extracted by a different convolution kernel, a discharge feature of different time scales (instantaneous response, steady state platform, and decay trend), and a multi-parameter coupling relationship (coordinated change of electrical and thermal characteristics). Through the multi-layer fully connected structure, a layer-by-layer mapping from specific physical features to abstract battery state is realized.

[0102] Finally, the high-level feature represents:

[0103] Each dimension feature represents a comprehensive battery state feature, which integrates multi-source information to form a distributed representation of key indicators such as battery health state and remaining capacity. The fully connected layer essentially constructs a complex mapping relationship from specific measurement features to abstract battery state indicators, which is a key conversion interface for precise capacity prediction. The comprehensive feature vector output by the fully connected layer includes the following core feature dimensions:

[0104] Capacity decay feature: reflects the nonlinear trajectory of the capacity decay of the storage battery during the entire discharge period;

[0105] Internal resistance evolution feature: represents the dynamic characteristics of the internal resistance of the storage battery changing with the discharge depth;

[0106] Thermal-electric coupling feature: describes the coordinated change law of temperature and electrical parameters (voltage and current);

[0107] Dynamic response feature: reflects the transient response capability of the storage battery under different load conditions;

[0108] Steady-state characteristic feature: represents the output stability of the storage battery during constant current / constant power discharge phase.

[0109] The gradient boosting decision tree sub-model of the present embodiment is used for regression prediction of the feature vector to obtain the estimated capacity and confidence of the storage battery. After inputting the comprehensive feature vector output by the fully connected layer into the gradient boosting decision tree sub-model, the gradient boosting decision tree sub-model analyzes the split points of each decision tree to evaluate the contribution of the above comprehensive feature vector to capacity prediction. For example, the frequency of each feature being used as a split node in all trees can be calculated. Based on the information gain brought by feature splitting, the contribution of each feature dimension to reducing prediction error is quantified, and the feature importance ranking is output to identify key factors affecting capacity prediction, such as capacity decay feature and internal resistance evolution feature, which usually have higher weights. Of course, the weights of each dimension feature can also be preset.

[0110] In the gradient boosting tree regression stage, the gradient boosting decision tree sub-model includes multiple decision trees. After inputting the comprehensive feature vector, each decision tree outputs an estimated capacity, which is calculated by the following formula:

[0111] F(x) = F_0(x) + ν·Σγ_jm;

[0112] wherein F_0(x) is an initial estimated capacity, ν is a preset learning rate, γ_jm is an estimated capacity of each decision tree, and x is an input comprehensive feature vector, the estimated capacity can be expressed in Ah or percentage (%), the gradient boosting decision tree sub-model regression decision principle can refer to the existing GBDT related technology, and will not be described in detail here.

[0113] In one embodiment, the standard deviation of the estimated capacity output by each decision tree can be calculated, and the confidence of the final output estimated capacity is calculated based on the standard deviation, that is, the confidence is equal to 1 / (1+standard deviation), the higher the confidence, the better the consistency of the output of each decision tree in the gradient boosting decision tree sub-model.

[0114] The hybrid learning model of the embodiment can be trained in the following way:

[0115] S1, obtaining a training data set, the training data set including multi-modal time series data samples collected by the historical capacity, and the label is the real capacity of the battery after the capacity is collected in the laboratory;

[0116] S2, constructing a hybrid learning model including CNN and gradient boosting decision tree sub-model, and initializing the model parameters;

[0117] S3, randomly extracting a batch of multi-modal time series data samples and inputting them into CNN, and outputting the predicted capacity from the gradient boosting decision tree sub-model;

[0118] S4, calculating the loss value such as mean square error loss by using the predicted capacity and the labeled real capacity;

[0119] S5, determining whether the stop training condition is met, such as whether the number of iteration training reaches the preset number or the loss value is less than the preset value; if yes, determining that the training is ended, and if no, executing S6;

[0120] S6, adjusting the gradient boosting decision tree sub-model and CNN through the loss value, and returning to S3.

[0121] Exemplarily, in the training process, the loss function gradient can be calculated by the prediction error (predicted value and true value) of the gradient boosting decision tree sub-model, the gradient is propagated to the CNN along the feature vector, the CNN adjusts the convolution kernel weight and bias parameters of the CNN based on the propagated gradient using an optimization algorithm (such as Adam), optimizes the feature extraction capability, and the gradient boosting decision tree sub-model itself is adjusted by incremental training, and the construction of new trees continues to fit the residual error of the current integrated gradient boosting decision tree sub-model, and the CNN forms an end-to-end joint optimization.

[0122] In another embodiment, the CNN and the gradient boosting decision tree sub-model can also use a collaborative but separate optimization mechanism. The CNN is responsible for extracting high-level features from time series data, and its feature extraction capability can be optimized in two steps: in the pre-training phase, the initial weights are optimized using supervised auxiliary tasks, such as supervised pre-training of the CNN using time series data labeled with feature vectors, so that the CNN has preliminary feature extraction capability. In the core training phase, the trained gradient boosting decision tree sub-model is used as a reliable supervision source to adjust the parameters of the CNN by minimizing the difference between the predicted value and the true value of the gradient boosting decision tree sub-model in a gradient backpropagation manner. At the same time, the gradient boosting decision tree sub-model itself is optimized in a forward step-by-step algorithm, and new decision trees are constructed to fit the prediction residual error of the current integrated gradient boosting decision tree sub-model, and the adjustment is reflected in the optimization of the tree structure, number, and depth of the super parameters. This training method enables both feature extraction and regression prediction to be optimized independently and collaboratively in a cascading manner, ultimately achieving end-to-end accurate capacity prediction.

[0123] In the hybrid learning model of the present embodiment, the CNN is good at extracting complex time series patterns, and the gradient boosting decision tree sub-model (GBDT) is good at accurate regression prediction, and can automatically adapt to capacity estimation under different battery types and different working conditions, and the prediction error is corrected through the error correction mechanism of the gradient boosting decision tree sub-model (GBDT), and the accuracy of the estimated capacity is improved.

[0124] In one embodiment, the cloud platform further includes a model optimization module for retraining the hybrid learning model using multi-modal time series data and the true capacity of the battery. Specifically, the system periodically collects new core capacity data of the battery and its corresponding laboratory verified capacity as training samples, and when the samples accumulate to a preset size, triggers the model incremental update process for retraining, and adjusts the parameters of the CNN and the gradient boosting decision tree sub-model during retraining to ensure that the hybrid learning model continuously adapts to the battery aging trend.

[0125] In another embodiment, the cloud platform is further configured to send the confidence to the capacity determination device, and the capacity determination device is further configured to adjust the programmable load based on the confidence, for example, if the confidence of the prediction of the hybrid learning model is low (e.g., the confidence interval is too wide), the controller can control the programmable load to reduce the load to reduce the discharge current, to collect data in a more gentle and stable manner after discharging, to avoid introducing additional noise by dynamic working conditions, and to improve the accuracy of subsequent predictions; if the hybrid learning model predicts with high confidence that the battery capacity has approached depletion, the controller can control to enter the fine measurement stage of the discharge end period or prepare to terminate the discharge to protect the battery, and if the hybrid learning model determines that the difference obtained by discharging the capacity of the current discharge curve is too large, the discharge mode can be automatically switched (e.g., between constant current and constant power) to match the optimal capacity determination standard.

[0126] In one embodiment, the cloud platform is further configured to generate a capacity determination report using the multi-modal time series data, the estimated capacity, the confidence, and the transaction ID returned by the blockchain, and upload the capacity determination report to the blockchain, for example, the capacity determination report is generated using the capacity determination process data (the collected multi-modal time series data, the discharge curve, etc.), the model prediction result, the confidence interval, and the corresponding transaction ID of the blockchain, wherein the transaction ID of the blockchain is a globally unique and tamper-proof transaction ID (Transaction ID) generated when the multi-modal time series data is uploaded to the blockchain, this transaction ID serves as the digital fingerprint of the capacity determination data block, and is automatically generated and returned by the blockchain network after the transaction is successfully uploaded to the blockchain, for subsequent query and verification. After the capacity determination report is calculated, it is written as a new evidence to the distributed ledger, forming a fully trusted and traceable closed loop from the process data to the final conclusion.

[0127] As shown in Figure 4 FIG. 1 shows a process schematic diagram of a battery capacity determination example according to an embodiment of the present application, and Figure 4 As shown in FIG. 1, after the capacity determination task is started, the cloud platform or the local operation and maintenance system issues a capacity determination instruction to the capacity determination device, the controller in the capacity determination device receives the capacity determination instruction, initializes the capacity determination preparation, and starts the programmable load to discharge the battery according to the preset discharge curve. The multi-modal sensing module synchronously collects raw data, which includes multi-dimensional time series data such as battery voltage, discharge current, battery internal resistance, battery temperature, etc. In this process, the capacity determination device submits the hash values of the key data (such as the discharge start timestamp, the cutoff voltage, the cumulative ampere-hour number, etc.) to the distributed ledger module in the capacity determination locally in real time. The distributed ledger module packs the hash values into a transaction, writes it to the blockchain network through a consensus mechanism, and generates a globally unique and tamper-proof transaction ID (Transaction ID). This transaction ID serves as the digital fingerprint of the capacity determination data block, and is automatically generated and returned to the capacity determination device by the blockchain network after the transaction is successfully uploaded to the blockchain, for subsequent query and verification.

[0128] The collected raw data is pre-processed, normalized and divided by sliding window, and then sent to the cloud platform through the communication module. In the cloud platform, the data is input into the CNN network. After passing through multiple convolutional pooling structures composed of convolutional layers, activation functions and pooling layers, the features are output to the flattening layer. After processing by the full connection layer, the CNN feature vector is output to the GBDT model. In the GBDT model, the features are calculated by multiple decision trees to calculate the residual error, and finally the capacity weighting calculation is performed to obtain the estimated capacity and confidence. When the battery voltage drops to the preset termination voltage, or the cumulative discharge capacity reaches the preset percentage of capacity, or the temperature, voltage and current are greater than the preset safety threshold, the capacity is stopped, the battery stops discharging, and the capacity process data, model prediction results, confidence interval and corresponding blockchain transaction ID are used to generate a capacity report. After calculating the hash of the capacity report, it is written into the distributed ledger to form a fully trusted and traceable closed loop from the process data to the final conclusion.

[0129] Among them, the controller and communication module of the capacity device can use an industrial-grade ARM Cortex-A series processor running an embedded Linux system. The distributed ledger module can use the lightweight SDK integration of Hyperledger Fabric or FISCO BCOS. The voltage / current sampling accuracy of the multi-modal perception module should be better than 0.1%, and the temperature sampling accuracy should be better than ±0.5℃. The programmable load can use high-power MOSFET or IGBT devices to ensure that it can withstand the maximum discharge current and be accurately controlled.

[0130] The hybrid learning model can be implemented in Python and built based on TensorFlow / PyTorch (for CNN) and Scikit-learn / XGBoost (for GBDT) frameworks. The hybrid learning model is trained and verified on the cloud platform using historical capacity data and deployed in the form of a service API. The smart contract of the distributed ledger is responsible for the hashing, storage and verification logic of the capacity data.

[0131] Taking the capacity of a certain base station's 48V battery pack as an example, the specific process can include the following steps:

[0132] 1. The operation and maintenance personnel set the capacity parameters on the cloud platform (such as using GB / T standard, constant power discharge).

[0133] 2. The capacity instruction is issued to the capacity device through the 5G network.

[0134] 3. The capacity device starts discharging according to the settings and collects data in real time. The data is hashed and uploaded to the chain every 10 seconds.

[0135] 4. Cloud platform real-time analysis of mixed learning model data, predicts the current capacity is 95% of the rated capacity, in good condition.

[0136] 5. Discharge ends, generates a PDF report containing discharge curve, capacity value and corresponding blockchain ID, which can be scanned to verify authenticity.

[0137] Compared with the prior art, the beneficial effects of the present application are:

[0138] 1. Trust: Through the distributed ledger, the authenticity and traceability of the core capacity data are fundamentally guaranteed, providing a trusted data foundation for battery life assessment and operation and maintenance decisions.

[0139] 2. Intelligent: CNN+GBDT hybrid learning model, which fully utilizes the advantages of CNN in feature extraction and GBDT in regression prediction, significantly improves the accuracy and robustness of capacity prediction under complex and variable conditions.

[0140] 3. Automation and flexibility: Achieve full-automatic closed-loop operation from task assignment to report generation, programmable load supports multiple discharge curves, and can adapt to different standards and battery types.

[0141] 4. High reliability and high adaptability: Multiple protocol communication ensures connectivity in various network environments, and the dual security protection system combining hardware and software greatly improves the overall reliability of the system.

[0142] Figure 5 A flowchart of a storage battery remote capacity verification method provided by the embodiment of the present application, the storage battery remote capacity verification method of the embodiment of the present application is applicable to the case of a storage battery remote capacity verification system, and the method can be executed by the storage battery remote capacity verification system, as shown in Figure 5 The storage battery remote capacity verification method of the embodiment of the present application can specifically include the following steps:

[0143] S501, control the storage battery to discharge when receiving the capacity verification instruction.

[0144] As shown in Figure 1 The storage battery remote capacity verification system of the embodiment of the present application is used for remotely controlling the storage battery in a communication network to verify capacity, specifically, the storage battery remote capacity verification system includes a capacity verification device and a cloud platform in communication with the capacity verification device, and the capacity verification device is disposed at the storage battery end and electrically connected with the storage battery.

[0145] Specifically, the capacity verification device is used to control the storage battery to discharge when receiving the capacity verification instruction, as shown in Figure 2As shown, the capacity-enhancing device includes a power supply, a programmable load, a multi-mode sensing module, a controller, and a communication module. The programmable load, the multi-mode sensing module, and the communication module are all connected to the controller. The programmable load is connected to the battery. When the controller receives a capacity-enhancing command, it controls the programmable load to start and controls the battery to discharge the programmable load according to a preset discharge curve.

[0146] S502. Collect multi-mode timing data during the discharge process.

[0147] like Figure 2 As shown, the multi-mode sensing module can integrate various sensors, such as current, power, and temperature sensors, to collect multi-dimensional data such as battery voltage, discharge current, battery internal resistance, battery temperature, ambient temperature, and ambient humidity during battery discharge according to a preset sampling frequency.

[0148] S503. Upload multimodal time series data to the blockchain.

[0149] The sensor data collected by the multi-modal sensing module is converted by ADC and output to the controller. The controller cleans and normalizes the multi-dimensional raw data, and then divides the normalized data into sliding windows to obtain multi-modal time series data.

[0150] like Figure 2 As shown, the core capacity device also includes a distributed ledger module, which is used to calculate hash values ​​for multimodal time-series data. The controller is used to send the multimodal time-series data and hash values ​​to the cloud platform through the communication module. The distributed ledger module runs as a lightweight node of the consortium blockchain or sidechain. Calculating hash values ​​for multimodal time-series data at the core capacity device and then uploading them to the cloud platform can prevent the collected data from being tampered with during collection and transmission, and provide trusted credentials for data traceability.

[0151] S504. Input multimodal time series data into the hybrid learning model.

[0152] like Figure 3 As shown, the hybrid learning model in this embodiment includes a feature extraction network and a gradient boosting decision tree sub-model. The feature extraction network is used to extract feature vectors from multimodal time series data, and the gradient boosting decision tree sub-model is used to perform regression prediction on the feature vectors to obtain the estimated capacity and confidence of the battery. Multimodal time series data can be input into the feature extraction network.

[0153] S505. In the hybrid learning model, feature vectors are extracted from multimodal time series data through a feature extraction network, and the estimated capacity and confidence level of the battery are obtained by regression prediction of the feature vectors through a gradient boosting decision tree sub-model. The estimated capacity and confidence level are then uploaded to the blockchain.

[0154] In one embodiment, the feature extraction network can be a CNN, such as Figure 3 As shown, the feature extraction network includes an input layer, multiple sequentially connected convolutional pooling structures, a feature flattening layer, and a fully connected layer. The input layer receives multimodal temporal data. The convolutional pooling structure includes a one-dimensional convolutional layer, an activation function, and a pooling layer. In each convolutional pooling layer, the one-dimensional convolutional layer uses one-dimensional convolutional kernels of different scales to convolve the input data to obtain a first feature map. The activation function performs a nonlinear transformation on the first feature map to obtain a second feature map. The pooling layer performs pooling processing on the second feature map to output a third feature map. The feature flattening layer flattens the multi-dimensional feature map output by the convolutional pooling structure into a one-dimensional feature vector. The fully connected layer performs global integration and nonlinear transformation on the one-dimensional feature vector to obtain a comprehensive feature vector.

[0155] In the gradient boosting tree regression stage, the gradient boosting decision tree sub-model consists of multiple decision trees. After inputting the comprehensive feature vector, each decision tree outputs an estimated capacity, which is calculated using the following formula:

[0156] F(x) = F_0(x) + ν·Σγ_jm;

[0157] Where F_0(x) is the initial estimated capacity, ν is the preset learning rate, γ_jm is the estimated capacity of each decision tree, and x is the input integrated feature vector. The estimated capacity can be expressed in ampere-hours (Ah) or as a percentage (%) of the rated capacity.

[0158] In one embodiment, the standard deviation of the estimated capacity output by each decision tree can be calculated, and the confidence level of the final output estimated capacity can be calculated based on the standard deviation. That is, the confidence level is equal to 1 / (1+standard deviation). The higher the confidence level, the better the consistency of the output of each decision tree in the gradient boosting decision tree sub-model.

[0159] It should be noted that, for the method embodiment, since it is basically similar to the battery remote capacity verification system embodiment, the description is relatively simple, and relevant parts can be referred to the description of the battery remote capacity verification system embodiment.

[0160] The remote capacity verification method for batteries provided in this embodiment of the invention is applied to the remote capacity verification system for batteries provided in this embodiment of the invention, thereby enabling the remote capacity verification method for batteries to have corresponding beneficial effects.

[0161] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the remote battery capacity verification method described in the various embodiments of the present invention.

[0162] It is worth noting that in the above embodiments of the remote capacity verification system for batteries, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0163] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A remote nuclear containment system for a battery, comprising: The system comprises a nuclear capacity device and a cloud platform in communication with the nuclear capacity device, and the nuclear capacity device is electrically connected with a storage battery; The nuclear capacity device is configured to control the storage battery to discharge when receiving a nuclear capacity instruction, collect multi-modal time series data during the discharging process, and send the multi-modal time series data to the cloud platform; The cloud platform is configured to receive the multi-modal time series data, upload the multi-modal time series data to a block chain, and input the multi-modal time series data into a hybrid learning model to extract a feature vector of the multi-modal time series data through a feature extraction network in the hybrid learning model, and perform regression prediction on the feature vector through a gradient boosting decision tree sub-model to obtain an estimated capacity and a confidence of the storage battery, and upload the estimated capacity and the confidence to the block chain.

2. The battery remote containment system of claim 1, wherein, The nuclear capacity device comprises a power supply, a programmable load, a multi-modal sensing module, a controller, and a communication module, the programmable load, the multi-modal sensing module, and the communication module are connected with the controller, and the programmable load is connected with the storage battery; The power supply is configured to supply power to the multi-modal sensing module, the controller, and the communication module; The controller is configured to control the programmable load to start when receiving a nuclear capacity instruction, and control the storage battery to discharge the programmable load according to a preset discharging curve; The multi-modal sensing module is configured to collect multi-dimensional raw data during the discharging process of the storage battery and send the multi-dimensional raw data to the controller; The controller is configured to preprocess the multi-dimensional raw data, and send the multi-modal time series data obtained after the preprocessing to the cloud platform through the communication module.

3. The remote storage battery capacity determination system according to claim 2, wherein The multi-modal sensing module is specifically configured to collect battery voltage, discharging current, battery internal resistance, battery temperature, ambient temperature, and ambient humidity during the discharging process of the storage battery at a preset frequency to obtain multi-dimensional raw data; The controller is specifically configured to clean and normalize the multi-dimensional raw data, and divide the normalized data into a sliding window to obtain multi-modal time series data.

4. The battery remote containment system of claim 2, wherein, The nuclear capacity device further comprises a distributed ledger module configured to calculate a hash value of the multi-modal time series data; The controller is configured to send the multi-modal time series data and the hash value to the cloud platform through the communication module.

5. The battery remote containment system of claim 4, wherein, The cloud platform comprises a hybrid learning model and a distributed ledger; The distributed ledger is configured to write the multi-modal time series data and the hash value through a block chain consensus mechanism; The hybrid learning model comprises a feature extraction network and a gradient boosting decision tree sub-model, and the feature extraction network is configured to extract a feature vector from the multi-modal time series data; The gradient boosting decision tree sub-model is configured to perform regression prediction on the feature vector to obtain an estimated capacity and a confidence of the storage battery.

6. The battery remote containment system of claim 5, wherein, The feature extraction network comprises an input layer, a plurality of convolutional pooling structures connected in sequence, a feature flattening layer, and a fully connected layer; The input layer is configured to receive the multi-modal time series data; The convolutional pooling structure comprises a one-dimensional convolution layer, an activation function and a pooling layer, in each of the convolutional pooling layers, the one-dimensional convolution layer is configured to perform one-dimensional convolution processing on input data by using one-dimensional convolution kernels of different scales to obtain a first feature map; The activation function is configured to perform nonlinear transformation on the first feature map to obtain a second feature map; The pooling layer is configured to perform pooling processing on the second feature map to output a third feature map; The feature flattening layer is configured to flatten the multi-dimensional feature map output by the convolutional pooling structure into a one-dimensional feature vector; The fully connected layer is configured to globally integrate and nonlinearly transform the one-dimensional feature vector to obtain a comprehensive feature vector.

7. The battery remote containment system of claim 6, wherein, The cloud platform further comprises a model optimization module configured to retrain the hybrid learning model by using the multi-modal time series data and the real capacity of the battery.

8. The battery remote containment system according to any of claims 2-7, wherein, The cloud platform is further configured to send the confidence to the core capacity device; The core capacity device is further configured to adjust the programmable load based on the confidence.

9. The battery remote containment system according to any of claims 2-7, wherein, The cloud platform is further configured to generate a core capacity report by using the multi-modal time series data, the estimated capacity, the confidence and the transaction ID returned by the blockchain, and upload the core capacity report to the blockchain.

10. A method for remotely verifying the integrity of a battery, comprising: The application is applied to the battery remote core capacity system of any one of claims 1-9, comprising: controlling the battery to discharge when receiving the core capacity instruction; collecting multi-modal time series data during discharging; uploading the multi-modal time series data to the blockchain; inputting the multi-modal time series data into the hybrid learning model; extracting a feature vector from the multi-modal time series data by a feature extraction network in the hybrid learning model, and performing regression prediction on the feature vector by a gradient boosting decision tree sub-model to obtain an estimated capacity and a confidence of the battery, and uploading the estimated capacity and the confidence to the blockchain.

Images