Energy storage power supply load access state detection method and system

Abstract

The invention discloses an energy storage power supply load access state detection method and system, and belongs to the technical field of power grid monitoring, and the detection method specifically comprises the following steps: I, deploying an edge intelligent node and a multi-physical field sensor array at an output port of an energy storage power supply, and synchronously collecting all modal data of the energy storage power supply; iI, preprocessing the collected original modal data, constructing an energy storage power supply physical field coupling model, and analyzing relevance and causal relationship among different modal data; iII, when the load is accessed for the first time, learning and identifying each modal data feature of the new load, and performing hardware-level identity authentication; according to the method, fine sensing of the load state can be realized, introduction of a new fault risk is avoided, the recognition accuracy and robustness of an abnormal state are greatly improved, the rapid recognition capability of an unknown load is improved, and multi-node collaborative modeling and sharing are realized while sensitive information is protected.

Description

Technical Field

[0001] This invention relates to the field of power grid monitoring technology, and in particular to a method and system for detecting the load connection status of energy storage power sources. Background Technology

[0002] With the widespread application of new energy storage systems in smart grids, microgrids, electric transportation, and emergency power supply, the high-frequency dynamic interaction between power sources and loads is becoming increasingly complex. Especially in operating scenarios where multiple types of loads frequently connect and disconnect, changes in load status can induce microscopic physical anomalies such as contact resistance fluctuations, arcing, and localized heating. These phenomena are often difficult to identify directly through traditional electrical parameter detection (such as voltage and current). Furthermore, current mainstream load monitoring methods generally rely on wired sampling, power outage detection, or intrusive sensing, which have limitations such as high detection latency, high wiring costs, and poor stability, making it difficult to meet the practical needs of modern energy storage systems for non-intrusive, real-time, and high-precision status sensing.

[0003] Existing methods and systems for detecting the load access status of energy storage power sources cannot achieve precise perception of the load status, which can easily lead to new fault risks, reduce the accuracy and robustness of abnormal status identification, and fail to achieve multi-node collaborative modeling and sharing while protecting sensitive information. To address this, we propose a method and system for detecting the load access status of energy storage power sources. Summary of the Invention

[0004] The purpose of this invention is to address the deficiencies in the existing technology by proposing a method and system for detecting the load connection status of an energy storage power source.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for detecting the load connection status of an energy storage power source, the specific steps of which are as follows: Ⅰ. Deploy edge intelligent nodes and multiphysics sensor arrays at the output port of the energy storage power supply, and simultaneously collect data on each mode of the energy storage power supply; II. Preprocess the collected raw modal data and construct a physical field coupling model of the energy storage power source to analyze the correlation and causal relationship between different modal data; III. When the load is first connected, the modal data characteristics of the new load are learned and identified, and hardware-level identity authentication is performed; IV. Analyze the characteristic change trends in the initial stage of load connection, determine the load connection intention, and predict the short-term state of the energy storage power source through the state detection model. V. Train a state detection model through edge intelligent nodes to dynamically monitor load behavior and state evolution, and update the load feature fingerprint database in real time.

[0006] As a further aspect of the present invention, the multi-physics field sensor array described in step I specifically includes a micro-current magnetic field sensor, an ultrasonic sensor, a micro-vibration accelerometer, a non-contact infrared thermal imaging micro-array, and voltage and current transient detection sensors. The physical field data mentioned in step I specifically includes magnetic field strength waveform, current transient waveform, voltage transient waveform, raw ultrasonic sound pressure signal, acceleration time series, infrared thermal image pixel matrix, ambient temperature data, triaxial vibration waveform, and timestamp information. Each set of physical field data is acquired and transmitted through a non-invasive multidimensional data channel.

[0007] As a further aspect of the present invention, the specific steps for preprocessing the acquired raw modal data in step II are as follows: S1.1: Normalize the physical field data collected by each physical field sensor array, set a fixed-length sliding window according to the sampling rate and signal fluctuation period of each sensor, and then select the normalized physical field data step by step through the set sliding window based on the preset time step. S1.2: After each movement of the sliding window, calculate the mean value of the physical field data in the sliding window after the movement, and replace the center value in the original sliding window. Compensate for the part of the physical field data at the beginning or end that does not reach the length of the sliding window by means of mirror expansion, zero filling or no processing. S1.3: Determine the required frequency band based on the sensor type and target physical phenomenon, design an FIR or IIR bandpass filter based on the selected frequency band, calculate the impulse response or transfer function coefficients of the FIR or IIR bandpass filter using the least squares method, and perform convolution or iterative processing on each physical field data after the moving average filtering through the corresponding type of bandpass filter. S1.4: After the data of each physical field is filtered, the sampling time axis of each sensor is resampled, and the sampling time of the physical field data collected by each sensor is unified by time interpolation to form an aligned multimodal dataset. S1.5: Perform Z-score standardization on the data in each modal dataset. If the standardized score generated by the standardization process is higher than the preset threshold, the corresponding physical field data will be regarded as an outlier and removed.

[0008] As a further aspect of the present invention, the specific steps of constructing the physical field coupling model of the energy storage power source and analyzing the correlation and causal relationship between different modal data in step II are as follows: S2.1: Construct the time delay sequence of each physical field data, and calculate the correlation strength of two physical field data under different time lags. Set a set of sliding windows, including window length, window function and sliding step size. Initialize the starting point of the sliding window to the starting data point of the corresponding physical field data, and move the sliding window step by step. S2.2: After each movement, the window function is used to analyze the local energy changes of the corresponding physical field data in time and frequency, and to determine whether different modes have overlapping energy peaks in the same frequency band and the same time period, so as to obtain the synchronization or coupling phenomenon between the physical field data. S2.3: Divide each physical field data into multiple intervals to form a discrete sequence, and calculate the marginal probability distribution and joint distribution of each physical field data respectively. Use the probability distribution to calculate mutual information to obtain the nonlinear coupling strength between two physical fields. S2.4: Perform autoregressive modeling for each physical field data separately, and then randomly select any other physical field data to add to the autoregressive modeling process to establish a multivariate causal regression model. If the prediction error value of the multivariate causal regression model decreases beyond the preset threshold compared to the prediction error value of the autoregressive modeling, it indicates that there is causality. Then, calculate the causal strength between each physical field data based on the residual variance of the autoregressive modeling and the multivariate causal regression model. S2.5: Statistically analyze the correlation strength, synchronization or coupling phenomena, nonlinear coupling strength, and causal strength of each physical field data. Use each physical field data as a node and each dependency relationship as an edge to establish a physical field coupling network graph.

[0009] As a further aspect of the present invention, the specific steps of step V, which involves collaboratively optimizing the load status detection process through edge intelligent nodes, are as follows: S3.1: The central coordination server distributes unified initial model parameters to the analysis and prediction modules of each edge node as initial parameters for the start of training. Each edge node divides the locally collected multimodal dataset into multiple small batch data sample pairs, each data sample pair containing a feature vector and a corresponding label. S3.2: Input the established multiple sets of data sample pairs into the local model of the edge node, and perform multiple rounds of forward propagation, loss calculation, gradient backpropagation and parameter update. During the training of the local model at the edge node, monitor the local model training information in real time according to the set early stopping strategy. If the preset early stopping strategy is met, stop training; otherwise, retrain until the model loss value converges to the preset range. S3.3: After local training is completed, calculate the parameter difference between the current local model parameters and the initial model parameters, and encrypt the calculated parameter difference. Then, each edge node uploads the encrypted parameter difference to the central coordination server. S3.4: The central coordination server performs a weighted summation of the model parameter differences uploaded by all edge nodes, updates the global model parameter values ​​based on the calculation results, sends the updated global model parameters back to all edge nodes, and waits for the next round of local training or real-time detection.

[0010] A power storage power supply load access status detection system includes a sensor acquisition module, a signal processing module, a feature extraction module, a data fusion module, an identification and storage module, a matching and classification module, an analysis and prediction module, a detection and alarm module, an authentication and access module, a collaborative update module, and a visual interaction module. The sensing and acquisition module is used to collect data on electromagnetic, acoustic, thermal, and mechanical vibrations at the moment the load is connected and during operation. The signal processing module is used to perform noise reduction, filtering, normalization, time alignment and synchronous sampling on the physical field data. The feature extraction module is used to extract time-domain, frequency-domain, and time-frequency-domain feature data of the processed physical field data; The data fusion module is used to fuse heterogeneous feature data from multiple sensors; The identification and storage module is used to identify the multimodal characteristics of the unknown load that is connected for the first time, and generate a unique load feature fingerprint of the unknown load. The matching and classification module is used to compare the identified load feature fingerprint with existing features in the load feature fingerprint database to identify the load identity and determine its access status. The analysis and prediction module analyzes the characteristic change trend of the load in the early stage of access and combines historical data to predict the state changes of the load in the short term. The detection and alarm module is used to immediately trigger a local or remote alarm mechanism when an abnormal state or potential fault trend is identified, and send an alarm signal to the control center or maintenance personnel. The authentication access module constructs a physically unclonable identifier based on the multimodal transient response at the moment of load access and compares it with a preset authentication database to authenticate the load identity. The collaborative update module monitors the operation status of the analysis and prediction module in real time and performs real-time collaborative updates on the parameters of the analysis and prediction model. The visual interaction module is used to display the access status of each load, identification results, feature evolution trends and historical data reports, and provides interactive functions such as remote query, manual annotation and strategy adjustment.

[0011] As a further aspect of the present invention, the specific steps for the identification storage module to identify the multimodal characteristics of the unknown load accessed for the first time are as follows: S4.1: Collect historical load data of energy storage power sources, unify the physical field data of different units and sampling formats in the historical load data to a standard scale, and convert them into a unified embedding vector. Divide the processed historical load data into training set and validation set. S4.2: Based on the MobileNet network, a predefined state recognition model is defined, and its number of network layers, parameters of each network layer, and training period are set. The network layers include an input layer, a convolutional layer, a pooling layer, a modeling layer, a fully connected layer, and an output layer. Then, the training set is divided into multiple batches of small-batch training samples, and each training sample is input into the state recognition model in sequence. S4.3: The state recognition model receives multimodal data from each training sample, normalizes each modal data separately, and concatenates the modal data into a unified input tensor. Then, the generated input tensors are propagated forward. The convolutional layer extracts the local joint features between the modalities in the input tensor along the time axis through a one-dimensional convolutional kernel. S4.4: After each local joint feature extraction of the convolutional layer is completed, the extracted local joint features are input into the pooling layer. Then, the pooling layer performs max pooling on each local joint feature according to the preset stride to reduce the time dimension of each local joint feature. From shallow to deep layers, the convolutional layer feature extraction and pooling layer dimensionality reduction are repeatedly performed until the maximum number of convolutional layers and pooling layers is reached. S4.5: The pooling layer inputs the joint local features after the final dimensionality reduction into the modeling layer. Then, the modeling layer performs weighted aggregation of the joint local features at different times based on the multi-head attention mechanism. The fully connected layer receives the aggregated feature tensor and maps the feature tensor to a fixed-dimensional load fingerprint feature through a non-linear function. The output layer receives the load fingerprint feature output by the fully connected layer and identifies the load state through a Softmax classifier. Finally, the output layer outputs the prediction result containing the load fingerprint feature and the corresponding load state. S4.6: Calculate the loss value between the predicted result and the actual result output by the state recognition model through the cross-entropy loss function, and then use the backpropagation algorithm to pass the loss value from the output layer to the input layer of the state recognition model layer by layer, and update the parameters of each network layer through the parameter gradient descent method. S4.7: When the training rounds reach the preset training period, the validation set is input into the state recognition model to obtain the corresponding prediction results, and the corresponding loss value is calculated through the cross-entropy loss function. If the loss value is higher than the preset threshold, the state recognition model is retrained and updated until the loss value of the state recognition model on the validation set converges to the preset threshold. S4.8: Input the latest multimodal dataset into the trained state recognition model, process the multimodal dataset layer by layer through the forward propagation algorithm, and output the current load fingerprint features of the energy storage power supply and the corresponding load operating status. At the same time, establish a load fingerprint feature library, associate and store the load fingerprint features with the corresponding load operating status, and update the load fingerprint feature library when new load fingerprint features exist.

[0012] As a further aspect of the present invention, the specific steps of the analysis and prediction module in analyzing the initial characteristic change trend of load access are as follows: S5.1: Select multimodal feature data within a short time window of 100-500ms from the moment the load is connected to the energy storage power source. Use the selected multimodal feature data as the initial response sequence. Then, use the GELU activation function to map the initial response sequence into latent space features of a unified dimension. S5.2: Initialize the forward and backward LSTM networks, including the input gate, forget gate, output gate, and weights and biases of the candidate units. Then, input the latent space features into the forward and backward LSTM networks respectively. S5.3: Calculate the unit memory and output hidden state of each hidden space feature at the current time step through the forward LSTM network, and then calculate the unit memory and output hidden state of each hidden space feature at the current time step in reverse chronological order through the reverse LSTM network. S5.4: The forward and backward hidden states output by the forward LSTM network and the backward LSTM network are concatenated to obtain the comprehensive hidden state of the bidirectional LSTM model at each time step. Then, the ReLU activation function is used to perform nonlinear activation on the comprehensive hidden state at each time step. Finally, the Softmax classifier is used to determine the load state at each time step and output it. S5.5: Extract statistics or dynamic summaries from the load operation records within a preset time period before access to form a context state vector, obtain the operation background of the corresponding node in the physical field coupled network graph, then concatenate the output of the bidirectional LSTM model with the context state vector, and then perform nonlinear activation on the concatenated fusion features. S5.6: Through the final classification layer, output the predicted probability distribution of each state category corresponding to each activated fusion feature. Each state category corresponds to a specific short-term operating state. Based on the category with the maximum probability, determine the short-term state trend of the current load and then calculate the confidence value of the prediction result. S5.7: If the confidence value is higher than the preset threshold, the prediction result will be used directly as the basis for automatic diagnosis. Otherwise, it will prompt that further observation or manual confirmation is needed, and the prediction result will be sent to the detection alarm module.

[0013] As a further aspect of the present invention, the load state described in S5.4 specifically includes normal access, abnormal plugging / unplugging, test equipment, and security threat states.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts the synchronization, coupling, and causal relationships between different physical fields through methods such as time delay correlation, sliding window energy spectrum analysis, mutual information, and causal modeling, constructing a physical field coupling network graph. Subsequently, historical load data is processed uniformly, load fingerprint features are extracted using MobileNet, and a state recognition model is trained through a multimodal deep neural network for load state classification and recognition. Next, based on short-time multimodal response data, bidirectional LSTM is used to model its evolution trend, and context state vectors are fused to output short-term operating state prediction results. Finally, each edge node independently trains its local model and uploads encrypted parameter differences, which are aggregated by the central server to form a global model for collaborative optimization. This enables fine perception of load states, avoids introducing new fault risks, greatly improves the accuracy and robustness of abnormal state identification, enhances the ability to quickly identify unknown loads, and achieves multi-node collaborative modeling and sharing while protecting sensitive information. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a flowchart of a method for detecting the load connection status of an energy storage power supply proposed in this invention. Figure 2 This is a system block diagram of an energy storage power supply load access status detection system proposed in this invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0018] Example 1, referring to Figure 1 This embodiment discloses a method for detecting the load connection status of an energy storage power source. The specific steps of the detection method are as follows: Edge smart nodes and multiphysics sensor arrays are deployed at the output port of the energy storage power supply, and data from each mode of the energy storage power supply are collected synchronously.

[0019] It should be further explained that the multi-physics sensor array specifically includes a micro-current magnetic field sensor, an ultrasonic sensor, a micro-vibration accelerometer, a non-contact infrared thermal imaging micro-array, and voltage and current transient detection sensors; the physical field data specifically includes magnetic field intensity waveforms, current transient waveforms, voltage transient waveforms, raw ultrasonic sound pressure signals, acceleration time series, infrared thermal image pixel matrix, ambient temperature data, triaxial vibration waveforms, and timestamp information, and each set of physical field data is acquired and transmitted through a non-invasive multi-dimensional data channel.

[0020] The raw modal data were preprocessed, and a physical field coupling model of the energy storage power source was constructed to analyze the correlation and causal relationship between different modal data.

[0021] Specifically, the physical field data collected by each physical field sensor array is normalized. A fixed-length sliding window is set based on the sampling rate and signal fluctuation period of each sensor. Then, based on a preset time step, the normalized physical field data is gradually selected through the sliding window. After each movement of the sliding window, the mean value of the physical field data within the new sliding window is calculated and replaced with the center value of the original sliding window. Mirror expansion, zero-filling, or no processing are used to compensate for the portion of the physical field data at the beginning or end that does not reach the length of the sliding window. The required frequency band is determined based on the sensor type and the target physical phenomenon, and the design is based on the selected frequency band. The FIR or IIR bandpass filter calculates the impulse response or transfer function coefficients of the FIR or IIR bandpass filter using the least squares method. For each physical field data after moving average filtering, convolution or iterative processing is performed through the corresponding type of bandpass filter. After filtering of each physical field data, the sampling time axis of each sensor is resampled, and the sampling time of the physical field data collected by each sensor is unified by time interpolation to form an aligned multimodal dataset. Z-score standardization is performed on the data in each modal dataset. If the standardized score generated by the standardization process is higher than the preset threshold, the corresponding physical field data is regarded as an outlier and removed.

[0022] Specifically, a time-delay sequence of each physical field data is constructed, and the correlation strength of two physical field data at different time lags is calculated. A set of sliding windows is set, including window length, window function, and sliding step size. The starting point of the sliding window is initialized to the starting data point of the corresponding physical field data. The sliding window is moved step by step. After each movement, the local energy changes of the corresponding physical field data in time and frequency are analyzed using the window function to determine whether there is energy peak overlap between different modes in the same frequency band and the same time period, so as to obtain the synchronization or coupling phenomenon between the physical field data. Each physical field data is divided into multiple intervals to form a discrete sequence, and the marginal probability distribution and joint distribution of each physical field data are calculated separately. The probability distribution is then used to calculate the correlation strength between the two physical field data. The mutual information is calculated to obtain the nonlinear coupling strength between two physical fields. Autoregressive modeling is performed for each physical field data, and then any other physical field data is randomly selected to be added to the autoregressive modeling process to establish a multivariate causal regression model. If the prediction error value of the multivariate causal regression model decreases by more than a preset threshold compared to the prediction error value of the autoregressive model, it indicates that causality exists. Then, the causal strength between each physical field data is calculated based on the residual variance of the autoregressive model and the multivariate causal regression model. The correlation strength, synchronization or coupling phenomenon, nonlinear coupling strength, and causal strength dependencies between each physical field data are statistically analyzed. With each physical field data as a node and each dependency relationship as an edge, a physical field coupling network graph is established.

[0023] When a load is first connected, its modal data characteristics are learned and identified, and hardware-level identity authentication is performed.

[0024] Analyze the characteristic change trends in the initial stage of load connection to determine the load connection intention, and predict the short-term state of the energy storage power source through a state detection model.

[0025] By training a state detection model through edge intelligent nodes, the system dynamically monitors load behavior and state evolution, and updates the load feature fingerprint database in real time.

[0026] Specifically, the central coordination server distributes unified initial model parameters to the analysis and prediction modules of each edge node as initial parameters for training. Each edge node divides its locally collected multimodal dataset into multiple small-batch data sample pairs, each containing a feature vector and a corresponding label. These multiple data sample pairs are input into the edge node's local model, and multiple rounds of forward propagation, loss calculation, gradient backpropagation, and parameter updates are performed. During the training of the local model at the edge node, the local model training information is monitored in real time according to the set early stopping strategy. If the preset early stopping strategy is met, training is stopped; otherwise, training is restarted until the model loss value converges to a preset range. After local training is completed, the parameter difference between the current local model parameters and the initial model parameters is calculated, and the calculated parameter difference is encrypted. Then, each edge node uploads the encrypted parameter difference to the central coordination server. The central coordination server performs a weighted summation of the model parameter differences uploaded by all edge nodes, updates the global model parameter values ​​based on the calculation results, sends the updated global model parameters back to all edge nodes, and waits for the next round of local training or real-time detection.

[0027] Example 2, refer to Figure 2 This embodiment discloses an energy storage power supply load access status detection system, including a sensor acquisition module, a signal processing module, a feature extraction module, a data fusion module, an identification and storage module, a matching and classification module, an analysis and prediction module, a detection and alarm module, an authentication and access module, a collaborative update module, and a visual interaction module.

[0028] The sensing and acquisition module is used to collect electromagnetic, acoustic, thermal, and mechanical vibration data of the physical fields at the moment of load connection and during operation; the signal processing module is used to denoise, filter, normalize, time-series align, and synchronously sample the physical field data; the feature extraction module is used to extract time-domain, frequency-domain, and time-frequency-domain feature data of the processed physical field data; the data fusion module is used to fuse heterogeneous feature data from multiple sensors; and the identification and storage module is used to identify the multimodal features of the unknown load connected for the first time and generate a unique load feature fingerprint of the unknown load.

[0029] Specifically, historical load data of the energy storage power supply is collected, and the physical field data of different units and sampling formats in the historical load data are unified to a standard scale and transformed into a unified embedding vector. The processed historical load data is divided into training and validation sets. A predefined state recognition model based on the MobileNet network is used, and its number of network layers, parameters of each network layer, and training cycle are set. The network layers include an input layer, convolutional layer, pooling layer, modeling layer, fully connected layer, and output layer. Then, the training set is divided into multiple batches of small-batch training samples, and each training sample is sequentially input into the state recognition model. The state recognition model receives the multimodal data of each training sample and processes each modality separately. The data is normalized, and the modal data are concatenated into a unified input tensor. The generated input tensors are then propagated forward. Convolutional layers, using one-dimensional convolutional kernels, extract local joint features between modalities from the input tensor along the time axis. After each local joint feature extraction, the extracted features are input to a pooling layer. The pooling layer then performs max pooling on each local joint feature according to a preset stride, reducing the time dimension of the features. This process is repeated from shallow to deep layers, involving convolutional feature extraction and pooling dimensionality reduction, until the maximum number of convolutional and pooling layers is reached. The pooling layer then inputs the final dimensionality-reduced local joint features into the modeling layer. The modeling layer then... The multi-head attention mechanism weights and aggregates local joint features from different time points. The fully connected layer receives the aggregated feature tensor and maps it to a fixed-dimensional load fingerprint feature using a non-linear function. The output layer receives the load fingerprint feature output from the fully connected layer and identifies the load state using a softmax classifier. The output layer then outputs a prediction result containing the load fingerprint feature and the corresponding load state. The loss between the predicted result and the true result is calculated using a cross-entropy loss function. Backpropagation is then used to propagate the loss value from the output layer to the input layer of the state recognition model, and the parameters of each network layer are updated using gradient descent. When the training epochs reach a certain threshold... After a preset training period, the validation set is input into the state recognition model to obtain the corresponding prediction results. The corresponding loss value is calculated using the cross-entropy loss function. If the loss value is higher than the preset threshold, the state recognition model is retrained and updated until the loss value of the state recognition model on the validation set converges to the preset threshold. The latest multimodal dataset is then input into the trained state recognition model. The multimodal dataset is processed layer by layer using the forward propagation algorithm, and the current load fingerprint features of the energy storage power supply and the corresponding load operating status are output. At the same time, a load fingerprint feature library is established, which associates and stores the load fingerprint features with the corresponding load operating status. When a new load fingerprint feature exists, the load fingerprint feature library is updated.

[0030] The matching and classification module compares the identified load feature fingerprint with existing features in the load feature fingerprint database to identify the load identity and determine its access status; the analysis and prediction module analyzes the trend of load feature changes in the early stage of access and combines historical data to predict the state changes of the load in the short term.

[0031] Specifically, multimodal feature data within a short time window of 100-500ms from the moment the load is connected to the energy storage power supply is selected. This selected multimodal feature data is used as the initial response sequence. The GELU activation function is then used to map the initial response sequence to latent space features of a unified dimension. Forward and backward LSTM networks are initialized, including input gates, forget gates, output gates, and weights and biases for candidate units. The latent space features are then input into the forward and backward LSTM networks respectively. The forward LSTM network calculates the unit memory and output hidden states of each latent space feature at the current time step. The backward LSTM network then calculates the unit memory and output hidden states of each latent space feature at the current time step in reverse chronological order. The forward and backward hidden states output by the forward and backward LSTM networks are concatenated to obtain the comprehensive hidden state of the bidirectional LSTM model at each time step. Finally, the ReLU activation function is applied. The comprehensive hidden state at each time step is nonlinearly activated, and the load state at each time step is determined by a Softmax classifier and output. Statistics or dynamic summaries are extracted from the load operation records within a preset time period before access to form a context state vector. The operation background of the corresponding node in the physical field coupled network graph is obtained. Then, the output of the bidirectional LSTM model and the context state vector are concatenated. The concatenated fusion features are then nonlinearly activated. Through the final classification layer, the predicted probability distribution of each state category corresponding to each activated fusion feature is output. Each state category corresponds to a specific short-term operation state. The short-term state trend of the current load is determined according to the category with the maximum probability. The confidence value of the prediction result is then calculated. If the confidence value is higher than the preset threshold, the prediction result is directly used as the basis for automatic diagnosis. Otherwise, it prompts that further observation or manual confirmation is required, and the prediction result is sent to the detection alarm module.

[0032] The detection and alarm module immediately triggers local or remote alarm mechanisms and sends alarm signals to the control center or maintenance personnel when abnormal states or potential fault trends are identified. The authentication and access module constructs a physically unclonable identifier based on the multimodal transient response at the moment of load access and compares it with a preset authentication database to authenticate the load identity. The collaborative update module monitors the operating status of the analysis and prediction module in real time and performs real-time collaborative updates to the parameters of the analysis and prediction model. The visual interaction module displays the access status of each load, identification results, feature evolution trends, and historical data reports, and provides interactive functions such as remote query, manual annotation, and strategy adjustment.

Claims

1. A method for detecting the load connection status of an energy storage power source, characterized in that, The specific steps of this detection method are as follows: Ⅰ. Deploy edge intelligent nodes and multi-physics sensor arrays at the output port of the energy storage power supply, and simultaneously collect data from each physical field of the energy storage power supply; II. Preprocess the collected physical field data and construct a physical field coupling network diagram of the energy storage power source to analyze the correlation and causal relationship between different modal data; III. When the load is first connected, learn and identify the modal data features of the new load, build a load feature fingerprint database, and perform load identity authentication. IV. Analyze the characteristic change trends in the initial stage of load connection, determine the load connection intention, and predict the short-term state of the energy storage power source through the state detection model. V. Collaboratively optimize the load status detection process through edge intelligent nodes, dynamically monitor load behavior and status evolution, and update the load feature fingerprint database in real time.

2. The method for detecting the load connection status of an energy storage power source according to claim 1, characterized in that, The specific steps for preprocessing the acquired raw modal data in step II are as follows: S1.1: Normalize the physical field data collected by each physical field sensor array, set a fixed-length sliding window according to the sampling rate and signal fluctuation period of each sensor, and then select the normalized physical field data step by step through the set sliding window based on the preset time step. S1.2: After each movement of the sliding window, calculate the mean value of the physical field data in the sliding window after the movement, and replace the center value in the original sliding window. Compensate for the part of the physical field data at the beginning or end that does not reach the length of the sliding window by means of mirror expansion, zero filling or no processing. S1.3: Determine the required frequency band based on the sensor type and target physical phenomenon, design an FIR or IIR bandpass filter based on the selected frequency band, calculate the impulse response or transfer function coefficients of the FIR or IIR bandpass filter using the least squares method, and perform convolution or iterative processing on each physical field data after the moving average filtering through the corresponding type of bandpass filter. S1.4: After the data of each physical field is filtered, the sampling time axis of each sensor is resampled, and the sampling time of the physical field data collected by each sensor is unified by time interpolation to form an aligned multimodal dataset. S1.5: Perform Z-score standardization on the data in each modal dataset. If the standardized score generated by the standardization process is higher than the preset threshold, the corresponding physical field data will be regarded as an outlier and removed.

3. The method for detecting the load connection status of an energy storage power source according to claim 2, characterized in that, The specific steps for constructing the physical field coupling model of the energy storage power source and analyzing the correlation and causal relationship between different modal data, as described in Step II, are as follows: S2.1: Construct the time delay sequence of each physical field data, and calculate the correlation strength of two physical field data under different time lags. Set a set of sliding windows, including window length, window function and sliding step size. Initialize the starting point of the sliding window to the starting data point of the corresponding physical field data, and move the sliding window step by step. S2.2: After each movement, the window function is used to analyze the local energy changes of the corresponding physical field data in time and frequency, and to determine whether different modes have overlapping energy peaks in the same frequency band and the same time period, so as to obtain the synchronization or coupling phenomenon between the physical field data. S2.3: Divide each physical field data into multiple intervals to form a discrete sequence, and calculate the marginal probability distribution and joint distribution of each physical field data respectively. Use the probability distribution to calculate mutual information to obtain the nonlinear coupling strength between two physical fields. S2.4: Perform autoregressive modeling for each physical field data separately, and then randomly select any other physical field data to add to the autoregressive modeling process to establish a multivariate causal regression model. If the prediction error value of the multivariate causal regression model decreases beyond the preset threshold compared to the prediction error value of the autoregressive modeling, it indicates that there is causality. Then, calculate the causal strength between each physical field data based on the residual variance of the autoregressive modeling and the multivariate causal regression model. S2.5: Statistically analyze the correlation strength, synchronization or coupling phenomena, nonlinear coupling strength, and causal strength of each physical field data. Use each physical field data as a node and each dependency relationship as an edge to establish a physical field coupling network graph.

4. The method for detecting the load connection status of an energy storage power source according to claim 1, characterized in that, The specific steps for collaboratively optimizing the load status detection process through edge intelligent nodes as described in step V are as follows: S3.1: The central coordination server distributes unified initial model parameters to the analysis and prediction modules of each edge node as initial parameters for the start of training. Each edge node divides the locally collected multimodal dataset into multiple small batch data sample pairs, each data sample pair containing a feature vector and a corresponding label. S3.2: Input the established multiple sets of data sample pairs into the local model of the edge node, and perform multiple rounds of forward propagation, loss calculation, gradient backpropagation and parameter update. During the training of the local model at the edge node, monitor the local model training information in real time according to the set early stopping strategy. If the preset early stopping strategy is met, stop training; otherwise, retrain until the model loss value converges to the preset range. S3.3: After local training is completed, calculate the parameter difference between the current local model parameters and the initial model parameters, and encrypt the calculated parameter difference. Then, each edge node uploads the encrypted parameter difference to the central coordination server. S3.4: The central coordination server performs a weighted summation of the model parameter differences uploaded by all edge nodes, updates the global model parameter values ​​based on the calculation results, sends the updated global model parameters back to all edge nodes, and waits for the next round of local training or real-time detection.

5. A system for detecting the load connection status of an energy storage power source, used to implement the method for detecting the load connection status of an energy storage power source according to any one of claims 1-4, characterized in that, It includes a sensor acquisition module, a signal processing module, a feature extraction module, a data fusion module, a recognition and storage module, a matching and classification module, an analysis and prediction module, a detection and alarm module, an authentication and access module, a collaborative update module, and a visual interaction module; The sensing and acquisition module is used to collect data on electromagnetic, acoustic, thermal, and mechanical vibrations at the moment the load is connected and during operation. The signal processing module is used to perform noise reduction, filtering, normalization, time alignment and synchronous sampling on the physical field data. The feature extraction module is used to extract time-domain, frequency-domain, and time-frequency-domain feature data of the processed physical field data; The data fusion module is used to fuse heterogeneous feature data from multiple sensors; The identification and storage module is used to identify the multimodal characteristics of the unknown load that is connected for the first time, and generate a unique load feature fingerprint of the unknown load. The matching and classification module is used to compare the identified load feature fingerprint with existing features in the load feature fingerprint database to identify the load identity and determine its access status. The analysis and prediction module analyzes the characteristic change trend of the load in the early stage of access and combines historical data to predict the state changes of the load in the short term. The detection and alarm module is used to immediately trigger a local or remote alarm mechanism when an abnormal state or potential fault trend is identified, and send an alarm signal to the control center or maintenance personnel. The authentication access module constructs a physically unclonable identifier based on the multimodal transient response at the moment of load access and compares it with a preset authentication database to authenticate the load identity. The collaborative update module monitors the operation status of the analysis and prediction module in real time and performs real-time collaborative updates on the parameters of the analysis and prediction model. The visual interaction module is used to display the access status of each load, identification results, feature evolution trends and historical data reports, and provides interactive functions such as remote query, manual annotation and strategy adjustment.

6. The energy storage power supply load access status detection system according to claim 5, characterized in that, The specific steps for the identification storage module to identify the multimodal characteristics of the unknown load being accessed for the first time are as follows: S4.1: Collect historical load data of energy storage power sources, unify the physical field data of different units and sampling formats in the historical load data to a standard scale, and convert them into a unified embedding vector. Divide the processed historical load data into training set and validation set. S4.2: Based on the MobileNet network, a predefined state recognition model is defined, and its number of network layers, parameters of each network layer, and training period are set. The network layers include an input layer, a convolutional layer, a pooling layer, a modeling layer, a fully connected layer, and an output layer. Then, the training set is divided into multiple batches of small-batch training samples, and each training sample is input into the state recognition model in sequence. S4.3: The state recognition model receives multimodal data from each training sample, normalizes each modal data separately, and concatenates the modal data into a unified input tensor. Then, the generated input tensors are propagated forward. The convolutional layer extracts the local joint features between the modalities in the input tensor along the time axis through a one-dimensional convolutional kernel. S4.4: After each local joint feature extraction of the convolutional layer is completed, the extracted local joint features are input into the pooling layer. Then, the pooling layer performs max pooling on each local joint feature according to the preset stride to reduce the time dimension of each local joint feature. From shallow to deep layers, the convolutional layer feature extraction and pooling layer dimensionality reduction are repeatedly performed until the maximum number of convolutional layers and pooling layers is reached. S4.5: The pooling layer inputs the joint local features after the final dimensionality reduction into the modeling layer. Then, the modeling layer performs weighted aggregation of the joint local features at different times based on the multi-head attention mechanism. The fully connected layer receives the aggregated feature tensor and maps the feature tensor to a fixed-dimensional load fingerprint feature through a non-linear function. The output layer receives the load fingerprint feature output by the fully connected layer and identifies the load state through a Softmax classifier. Finally, the output layer outputs the prediction result containing the load fingerprint feature and the corresponding load state. S4.6: Calculate the loss value between the predicted result and the actual result output by the state recognition model through the cross-entropy loss function, and then use the backpropagation algorithm to pass the loss value from the output layer to the input layer of the state recognition model layer by layer, and update the parameters of each network layer through the parameter gradient descent method. S4.7: When the training rounds reach the preset training period, the validation set is input into the state recognition model to obtain the corresponding prediction results, and the corresponding loss value is calculated through the cross-entropy loss function. If the loss value is higher than the preset threshold, the state recognition model is retrained and updated until the loss value of the state recognition model on the validation set converges to the preset threshold. S4.8: Input the latest multimodal dataset into the trained state recognition model, process the multimodal dataset layer by layer through the forward propagation algorithm, and output the current load fingerprint features of the energy storage power supply and the corresponding load operating status. At the same time, establish a load fingerprint feature library, associate and store the load fingerprint features with the corresponding load operating status, and update the load fingerprint feature library when new load fingerprint features exist.

7. The energy storage power supply load connection status detection system according to claim 5, characterized in that, The specific steps of the analysis and prediction module in analyzing the initial characteristic change trend of load access are as follows: S5.1: Select multimodal feature data within a short time window of 100-500ms from the moment the load is connected to the energy storage power source. Use the selected multimodal feature data as the initial response sequence. Then, use the GELU activation function to map the initial response sequence into latent space features of a unified dimension. S5.2: Initialize the forward and backward LSTM networks, including the input gate, forget gate, output gate, and weights and biases of the candidate units. Then, input the latent space features into the forward and backward LSTM networks respectively. S5.3: Calculate the unit memory and output hidden state of each hidden space feature at the current time step through the forward LSTM network, and then calculate the unit memory and output hidden state of each hidden space feature at the current time step in reverse chronological order through the reverse LSTM network. S5.4: The forward and backward hidden states output by the forward LSTM network and the backward LSTM network are concatenated to obtain the comprehensive hidden state of the bidirectional LSTM model at each time step. Then, the ReLU activation function is used to perform nonlinear activation on the comprehensive hidden state at each time step. Finally, the Softmax classifier is used to determine the load state at each time step and output it. S5.5: Extract statistics or dynamic summaries from the load operation records within a preset time period before access to form a context state vector, obtain the operation background of the corresponding node in the physical field coupled network graph, then concatenate the output of the bidirectional LSTM model with the context state vector, and then perform nonlinear activation on the concatenated fusion features. S5.6: Through the final classification layer, output the predicted probability distribution of each state category corresponding to each activated fusion feature. Each state category corresponds to a specific short-term operating state. Based on the category with the maximum probability, determine the short-term state trend of the current load and then calculate the confidence value of the prediction result. S5.7: If the confidence value is higher than the preset threshold, the prediction result will be used directly as the basis for automatic diagnosis. Otherwise, it will prompt that further observation or manual confirmation is needed, and the prediction result will be sent to the detection alarm module.

Images