Multifunctional mobile power supply light control system and method based on deep learning

Abstract

The invention discloses a multifunctional mobile power supply light control system and method based on deep learning, and relates to the technical field of mobile power supply light control, and the system comprises the following modules: an electrical signal feature extraction module which is used for constructing an electrical signal feature vector sequence; the line poor contact early warning module is used for generating a line poor contact early warning signal; the protocol type identification module is used for outputting a protocol type identifier; the light control instruction analysis module is used for analyzing the light control instruction; the load sudden change prediction module is used for outputting a power prediction numerical value by improving an Autoformer model; the self-adaptive protection control module is used for generating an output voltage set value; and the data transmission and visualization module is used for uploading the state data to the cloud. According to the method, the limitation that a traditional method ignores frequency domain energy entropy characteristics, the multi-dimensional data fusion degree is low and control response lags is overcome, and an efficient solution is provided for intelligent control over light of the mobile power source.

Description

Technical Field

[0001] This invention relates to the field of mobile power bank lighting control technology, and in particular to a multifunctional mobile power bank lighting control system and method based on deep learning. Background Technology

[0002] With the continuous expansion of mobile power bank application scenarios, multi-functional mobile power bank lighting control systems face complex load environments and diverse control requirements. Existing lighting control methods typically employ simple threshold judgment or proportional-integral-derivative (PID) control algorithms. While these methods can achieve basic switching and dimming functions, they primarily rely on feedback adjustment based on current instantaneous voltage or current sampling values. This control method, based solely on instantaneous values, ignores the frequency domain energy entropy characteristics implicit in electrical signals and the deep temporal dependencies between multi-dimensional heterogeneous data such as voltage, current, and temperature. This results in control decisions failing to accurately capture the potential patterns of load power changes, thus limiting the smoothness and response accuracy of lighting adjustment. Furthermore, traditional methods struggle to effectively integrate non-numerical lighting control commands with numerical electrical monitoring data within a unified feature space. This prevents the control strategy from perceiving the dynamic coupling mechanism between control commands and load responses, often leading to power prediction errors when handling complex time-varying loads, reducing system stability and energy efficiency.

[0003] Therefore, how to provide a multifunctional mobile power supply lighting control system and method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a multifunctional mobile power bank lighting control system and method based on deep learning. It collects waveform data of the lighting link through voltage and current sensors, generates an electrical signal feature vector sequence through frequency band energy distribution calculation and information entropy construction, and establishes an index mapping association with a historical database. It collects physical layer signal waveforms, extracts waveform fingerprint vectors using complex wavelet transform to identify protocol types, constructs a directed graph based on the protocol parsing rule table to verify instruction logic, and parses lighting control instructions. It calls the associated historical data and parsed control instructions to construct a multidimensional time-series feature matrix, inputs it into an improved Autoformer model, and outputs a load power prediction value through graph embedding alignment, cascade decomposition, stochastic resonance enhancement, and sequence parameter aggregation. Based on the power prediction value, it calculates the battery output current estimate, calculates the probability of voltage reaching a dangerous threshold based on Brownian motion modeling to dynamically adjust protection parameters, and encapsulates all state data into JSON messages. Based on the priority determined by the Hasse diagram, it uploads the data to the cloud via the MQTT protocol for visualization, thus achieving a closed-loop control process from signal acquisition, feature extraction, instruction parsing, load prediction to adaptive protection and visual interaction. This invention overcomes the limitations of traditional methods, such as ignoring frequency domain energy entropy characteristics, low multidimensional data fusion, and control response lag, and provides an efficient solution for intelligent control of mobile power supply lighting.

[0005] A deep learning-based method for controlling the lighting of a multifunctional mobile power bank, according to an embodiment of the present invention, includes the following steps: S1. Collect voltage and current waveform data of the lighting link, segment it into multiple frequency band components and calculate the energy distribution density, generate an energy entropy spectrum vector sequence based on information entropy, and combine them to form an electrical signal feature vector sequence; S2. Construct the covariance matrix of the electrical signal feature vector sequence and calculate the eigenvalue spectrum. When the spectral gap narrows and the clustering structure collapses, generate a line contact failure warning signal. S3. Acquire physical layer signal waveforms, extract phase consistency matrix using complex wavelet transform and map it to waveform fingerprint vector, calculate matching probability and output protocol type identifier; S4. Retrieve the preset parsing rule table according to the protocol type identifier, map the logical fields of the instruction frame to nodes and the timing jumps to directed edges to construct a directed graph, and parse the light control instruction after verifying the Euler circuit. S5. Call the voltage data, current data and temperature data of the historical time period associated with the electrical signal feature vector sequence, and construct a multi-dimensional time series feature matrix in combination with the lighting control command. Input the improved Autoformer model, and output the load power prediction value through graph embedding, cascade decomposition, stochastic resonance enhancement and sequence parameter aggregation. S6. Calculate the battery output current estimate based on the load power prediction value, model the voltage drop process as Brownian motion, calculate the probability distribution of the voltage first touching the danger threshold, and dynamically adjust the protection cut-off voltage and output voltage set value. S7. Encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into a JSON message, construct a partial order set and use a Hasse diagram to determine the transmission priority, and upload it to the cloud for visualization via MQTT.

[0006] Optionally, S1 specifically includes: S11. The voltage waveform data and current waveform data of the lighting link are collected by the voltage sensor and the current sensor at a preset sampling frequency. The collected analog signal is converted into a digital signal sequence and the timestamp of the collection time is recorded at the same time. S12. Calculate the signal energy value of the digital signal sequence in different frequency bands as the energy distribution density of each frequency band component; S13. Calculate the information entropy based on the energy distribution density of each frequency band component, construct the energy entropy spectrum vector based on the information entropy, and combine the energy entropy spectrum vectors corresponding to voltage and current according to the time sequence to form an electrical signal feature vector sequence. S14. Extract the timestamp index of the electrical signal feature vector sequence, retrieve the original voltage data, current data and temperature data under the corresponding time window from the historical database, and establish the electrical feature vector-multidimensional historical data index mapping relationship.

[0007] Optionally, S2 specifically includes: S21. Construct the covariance matrix of the electrical signal feature vector sequence, calculate the eigenvalues ​​of the covariance matrix, and arrange the eigenvalues ​​in descending order of their numerical values ​​to form an eigenvalue spectrum; S22. Calculate the difference between adjacent feature values ​​in the feature value spectrum to obtain the feature value interval sequence, and detect the maximum interval in the feature value interval sequence as the spectral gap; S23. Monitor the trend of the spectral gap change over time. When the spectral gap value decreases to below a preset threshold, it is determined that the cluster structure has collapsed, and a poor line contact warning signal is generated.

[0008] Optionally, S3 specifically includes: S31. Acquire physical layer signal waveforms, perform complex wavelet transform on the physical layer signal waveforms, calculate phase information at different scales, and combine them to generate a phase consistency matrix. S32. Perform dimensionality reduction and vectorization mapping on the phase consistency matrix to generate a waveform fingerprint vector for representing waveform features; S33. Calculate the similarity between the waveform fingerprint vector and each standard fingerprint in the preset protocol fingerprint library, convert the similarity into a matching probability, and select the protocol type with the highest matching probability as the protocol type identifier.

[0009] Optionally, S4 specifically includes: S41. Retrieve the preset parsing rule table according to the protocol type identifier to obtain the corresponding instruction frame format definition and timing logic rules; S42. Extract logical fields as nodes according to the instruction frame format definition, determine the jump relationship between nodes as directed edges according to the timing logic rules, and construct an instruction state directed graph. S43. Search the directed graph of the instruction state for a valid path from the start node to the end node. If a valid path exists, determine that the instruction structure is complete and parse the light control instruction. If a valid path does not exist, determine that the instruction structure is abnormal.

[0010] Optionally, the improved Autoformer model specifically includes a multi-dimensional temporal embedding layer, a cascaded decomposition network layer, an autocorrelation precursor capture layer, an adaptive aggregation layer, and a mutation prediction regression layer: The multidimensional temporal embedding layer is used to receive the electrical signal feature vector sequence. Based on the mapping relationship between the electrical feature vector and the multidimensional historical data index, it retrieves the original voltage data, current data, and temperature data associated with the mapping through the timestamp of the electrical signal feature vector. The original voltage data, current data, and temperature data are concatenated along the feature dimension to construct the original multidimensional temporal feature tensor. The lighting control command is mapped to the command embedding vector as the edge weight constraint. The minimum cost connection topology is constructed with data points as graph nodes. By introducing auxiliary nodes, the connection path of heterogeneous nodes is optimized to generate an optimized graph topology. The sparse low-frequency data is aligned and completed to dense high-frequency data through graph convolution aggregation to generate a multidimensional temporal embedding tensor and map it to a high-dimensional space to obtain the model input embedding tensor. The cascaded decomposition network layer is used to embed the model input into the tensor input cascaded decomposition module. It constructs a symmetry metric by calculating the autocorrelation coefficient of local amplitude changes and phase rotation through a sliding window. It maps the periodic components that meet the symmetry threshold to group invariants and extracts the seasonal term. It extracts the trend term through the residual components of the symmetry metric. It uses the seasonal term and the trend term as inputs to the next level decomposition module to refine the separation accuracy of the symmetry features layer by layer, and outputs a fine trend tensor and a fine seasonal tensor. The autocorrelation precursor capture layer is used to inject a non-Gaussian noise sequence with a set signal-to-noise ratio into the fine seasonal tensor, calculate the peak value of the cross-correlation function between the noisy signal and the original signal, adjust the noise intensity until the peak value of the cross-correlation function reaches a maximum value, and output the precursor feature tensor after nonlinear enhancement; calculate the signal-to-noise ratio gain curve under different noise intensities for the fine trend tensor, and extract the feature component corresponding to the maximum value of the signal-to-noise ratio gain as the cumulative trend feature tensor; The adaptive aggregation layer is used to identify the slow relaxation parameters in the precursor feature tensor and the cumulative trend feature tensor as order parameters, project the trend features of the fast relaxation parameters onto the subspace generated by the precursor order parameters for dominant aggregation, and generate an aggregated feature tensor containing the dominant aggregation pattern; the aggregated feature tensor is nonlinearly normalized by calculating the potential function gradient of the order parameters to obtain the normalized aggregated feature tensor. The mutation prediction regression layer is used to flatten the normalized aggregated feature tensor into a one-dimensional feature vector, which is then input into the fully connected layer for mapping; the probability distribution of the occurrence time of load mutation is output through the Softmax activation function, and the predicted load power value is output through linear regression.

[0011] Optionally, S6 specifically includes: S61. Calculate the estimated value of battery output current based on the predicted load power value and the current battery voltage, and calculate the voltage drop value as the voltage drop amplitude by combining the battery internal resistance. S62. Based on the voltage drop amplitude, determine the Brownian motion drift coefficient, model the voltage drop process as a drifting Brownian motion, set the danger threshold as the absorption wall, and calculate the probability density function and expected time when the voltage first touches the danger threshold. S63. Calculate the risk quantification index using the probability density function of the voltage first reaching the danger threshold and the expected time. When the risk quantification index exceeds the preset safety threshold, dynamically increase the protection cut-off voltage and decrease the output voltage setting value according to the risk quantification index.

[0012] Optionally, S7 specifically includes: S71. Encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into a JSON message, and assign a corresponding service type label to the JSON message; S72. Construct a partial order set based on the service type label, calculate the comparability relationship of each element in the partial order set, draw a Hasse diagram based on the comparability relationship, and determine the transmission priority sequence. S73. Upload the JSON message to the cloud server via the MQTT protocol according to the transmission priority sequence, and drive the cloud visualization interface to display the line status and control parameters.

[0013] A multi-functional mobile power bank lighting control module based on deep learning according to an embodiment of the present invention includes the following modules: The electrical signal feature extraction module is used to collect voltage and current waveform data of the lighting link, segment it into multiple frequency band components and calculate the energy distribution density, generate an energy entropy spectrum vector sequence based on information entropy, and combine them to form an electrical signal feature vector sequence. The poor line contact early warning module is used to construct the covariance matrix of the electrical signal feature vector sequence and calculate the eigenvalue spectrum. When the spectral gap narrows and causes the clustering structure to collapse, a poor line contact early warning signal is generated. The protocol type identification module is used to acquire physical layer signal waveforms, extract the phase consistency matrix using complex wavelet transform and map it to waveform fingerprint vectors, calculate the matching probability and output the protocol type identifier; The lighting control command parsing module is used to retrieve a preset parsing rule table according to the protocol type identifier, map the logical fields of the command frame to nodes and the timing jumps to directed edges to construct a directed graph, and parse the lighting control command after verifying the Euler circuit. The load mutation prediction module is used to call the voltage data, current data and temperature data of the historical time period associated with the electrical signal feature vector sequence, and combine them with the lighting control command to construct a multi-dimensional time series feature matrix. The improved Autoformer model is input, and after graph embedding, cascade decomposition, stochastic resonance enhancement and sequence parameter aggregation, the predicted power value is output. The adaptive protection control module is used to calculate the estimated value of battery output current based on the predicted load power value, model the voltage drop process as Brownian motion, calculate the probability distribution of the voltage first touching the danger threshold, and dynamically adjust the protection cut-off voltage and output voltage set value. The data transmission and visualization module is used to encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into JSON messages, construct a partial order set and use a Hasse diagram to determine the transmission priority, and upload it to the cloud for visualization display via MQTT.

[0014] The beneficial effects of this invention are: (1) This invention achieves deep fusion of heterogeneous electrical features and numerical time-series data by constructing an index mapping association between the frequency domain energy entropy spectrum of electrical signals and multidimensional historical data. The multidimensional time-series embedding layer retrieves the original data based on the index mapping, constructs a multidimensional time-series feature tensor, uses instruction embedding vector constraints to construct a minimum-cost connection topology, and optimizes the heterogeneous node paths by introducing auxiliary nodes. Through graph convolution aggregation, it achieves alignment and completion of sparse low-frequency data to dense high-frequency data. This mechanism effectively fills the data sparsity in the feature space and provides high-quality input data containing complete spatiotemporal semantics for load mutation prediction.

[0015] (2) This invention establishes a nonlinear time series prediction system driven by physical fields by designing an improved Autoformer model that includes cascaded decomposition, stochastic resonance enhancement, and order parameter aggregation. The cascaded decomposition network layer strips away evolution patterns at different time scales, the autocorrelation precursor capture layer enhances weak precursor signals using stochastic resonance effects, and the adaptive aggregation layer identifies order parameters and performs dominant aggregation. Through multi-scale decomposition, noise-induced signal enhancement, and collaborative learning mechanisms, this system extracts key evolution laws from mixed signals, ensuring that the power prediction values ​​accurately reflect the dynamic process of load mutations. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based multifunctional mobile power bank lighting control method proposed in this invention. Figure 2 This is a flowchart illustrating the working principle of the improved Autoformer model for a multifunctional mobile power bank lighting control method based on deep learning proposed in this invention. Figure 3 This is an overall flowchart of a multifunctional mobile power supply lighting control system based on deep learning proposed in this invention. Detailed Implementation

[0017] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 and Figure 2 A deep learning-based method for controlling the lighting of a multifunctional mobile power bank includes the following steps: S1. Collect voltage and current waveform data of the lighting link, segment it into multiple frequency band components and calculate the energy distribution density, generate an energy entropy spectrum vector sequence based on information entropy, and combine them to form an electrical signal feature vector sequence; S2. Construct the covariance matrix of the electrical signal feature vector sequence and calculate the eigenvalue spectrum. When the spectral gap shrinks and the cluster structure collapses, generate a line contact failure warning signal. S3. Acquire physical layer signal waveforms, extract phase consistency matrix using complex wavelet transform and map it to waveform fingerprint vector, calculate matching probability and output protocol type identifier; S4. Retrieve the preset parsing rule table according to the protocol type identifier, map the logical fields of the instruction frame to nodes and the timing jumps to directed edges to construct a directed graph, and parse the light control instruction after verifying the Euler circuit. S5. Call the voltage, current and temperature data of the historical time period associated with the electrical signal feature vector sequence, and construct a multi-dimensional time series feature matrix in combination with the lighting control command. Input the improved Autoformer model, and output the load power prediction value through graph embedding, cascade decomposition, stochastic resonance enhancement and sequence parameter aggregation. S6. Calculate the battery output current estimate based on the load power prediction value, model the voltage drop process as Brownian motion, calculate the probability distribution of the voltage first touching the danger threshold, and dynamically adjust the protection cut-off voltage and output voltage set value. S7. Encapsulate the line poor contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into a JSON message, construct a partial order set and use a Hasse diagram to determine the transmission priority, and upload it to the cloud for visualization via MQTT.

[0019] In this embodiment, S1 specifically includes: S11. The voltage and current waveform data of the lighting link are collected by the voltage sensor and the current sensor at a preset sampling frequency of 10kHz. The collected analog signal is converted into a digital signal sequence and the timestamp of the collection time is recorded at the same time.

[0020] S12. Perform 4-level wavelet packet decomposition on the digital signal sequence to divide the signal frequency band into 16 independent sub-frequency bands and extract the reconstructed signal coefficients of each sub-frequency band; calculate the sum of squares of the reconstructed signal coefficients of each sub-frequency band to obtain the signal energy value of each sub-frequency band; calculate the ratio of the energy value of each sub-frequency band to the total energy value, and use this ratio as the energy distribution density of each frequency band component.

[0021] S13. Based on the energy distribution density of each frequency band component, the information entropy value of the frequency band energy distribution is calculated using the information entropy formula. The specific process is as follows: calculate the natural logarithm of the energy distribution density of each frequency band, multiply the energy distribution density value by the corresponding natural logarithm to obtain the intermediate term of the energy entropy of each frequency band, sum the intermediate terms of the energy entropy of all frequency bands to obtain the sum value, and take the negative of the sum value to obtain the information entropy value of the frequency band energy distribution; arrange the information entropy values ​​corresponding to each frequency band in descending order of frequency to construct the energy entropy spectrum vector; construct the energy entropy spectrum vectors of voltage signal and current signal respectively, and align and combine them according to the time series to form the electrical signal feature vector sequence.

[0022] S14. Extract the timestamp index of each data point in the electrical signal feature vector sequence and generate a query key value; traverse the historical database according to the query key value and retrieve the original voltage data, current data and temperature data under the corresponding time window; establish the mapping relationship between electrical feature vector and multidimensional historical data index, and associate and store the retrieved multidimensional historical data with the electrical signal feature vector sequence.

[0023] In this embodiment, S2 specifically includes: S21. Construct the covariance matrix of the electrical signal feature vector sequence. The specific process is to calculate the mean vector of the feature vector sequence, subtract the mean vector from each feature vector to obtain the centered feature vector matrix, calculate the product of the centered feature vector matrix and its transpose to obtain the covariance matrix, perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalue set, and arrange the values ​​in the eigenvalue set in descending order to form the eigenvalue spectrum.

[0024] S22. Calculate the difference between adjacent eigenvalues ​​in the eigenvalue spectrum, arrange the calculation results in order to construct an eigenvalue interval sequence; traverse the eigenvalue interval sequence and extract the interval with the largest value as the spectral gap.

[0025] S23. Monitor the trend of spectral gap change over time in real time, set a preset threshold of 5% of the maximum feature value, and when the spectral gap value decreases to below the preset threshold, determine that the clustering structure of the electrical signal feature vector sequence in the feature space has collapsed, and generate a line contact failure warning signal.

[0026] In this embodiment, S3 specifically includes: S31. Acquire physical layer signal waveforms through a signal acquisition probe, and perform complex wavelet transform on the physical layer signal waveforms. Specifically, select a complex Morlet wavelet as the mother wavelet, set a scale parameter sequence, perform convolution operations on the signal at different scales, extract the complex coefficients of the transform results, calculate the argument of the complex coefficients as phase information, arrange the phase information at each scale in scale order, and generate a phase consistency matrix.

[0027] S32. Perform dimensionality reduction and vectorization mapping on the phase consistency matrix. Specifically, perform singular value decomposition on the phase consistency matrix, extract the left singular vector corresponding to the largest singular value as the principal feature direction, project the phase consistency matrix onto the principal feature direction, flatten the projection result into a one-dimensional row vector, and generate a waveform fingerprint vector to represent the waveform features.

[0028] S33. Calculate the similarity between the waveform fingerprint vector and each standard fingerprint in the preset protocol fingerprint database. Specifically, calculate the Euclidean distance between the waveform fingerprint vector and the standard fingerprint vector, take the reciprocal of the Euclidean distance and normalize it, and use the normalized result as the matching probability. Traverse the preset protocol fingerprint database and select the protocol type with the highest matching probability as the protocol type identifier.

[0029] In this embodiment, S4 specifically includes: S41. Retrieve the preset parsing rule table according to the protocol type identifier to obtain the corresponding instruction frame format definition and timing logic rules.

[0030] S42. Extract logical fields as nodes according to the instruction frame format definition, determine the jump relationship between nodes as directed edges according to the timing logic rules, and construct an instruction state directed graph. The specific process is to identify the frame header, address field, control field and check field in the instruction frame format definition as state nodes; construct directed edges connecting each state node according to the field sending order and response waiting logic specified in the timing logic rules; combine all nodes and directed edges to generate an instruction state directed graph.

[0031] S43. Search for a valid path from the start node to the end node in the directed graph of the instruction state. Specifically, use a depth-first search algorithm to traverse the directed graph of the instruction state and find a connected path between the frame header node corresponding to the start node and the check node corresponding to the end node. If a connected path exists, the instruction structure is considered complete, and the light control instruction is parsed according to the path order. If no connected path exists, the instruction structure is considered abnormal.

[0032] In this embodiment, the improved Autoformer model specifically includes a multi-dimensional temporal embedding layer, a cascaded decomposition network layer, an autocorrelation precursor capture layer, an adaptive aggregation layer, and a mutation prediction regression layer: The multidimensional time-series embedding layer is used to receive electrical signal feature vector sequences. Based on the mapping relationship between electrical feature vectors and multidimensional historical data indexes, the original voltage, current, and temperature data associated with the mapping are retrieved through the timestamps of the electrical signal feature vectors. The original voltage, current, and temperature data are concatenated along the feature dimensions to construct the original multidimensional time-series feature tensor. The original multidimensional time-series feature tensor is then Z-score standardized to eliminate dimensional differences. The semantic content of lighting control commands is analyzed, and the semantic content is mapped to 64-dimensional word vectors using the Word2Vec model. L2 norm normalization is applied to the word vectors to generate command embedding vectors. The cosine similarity between the command embedding vectors and the feature vectors of each data point in the original multidimensional temporal feature tensor is calculated. The dissimilarity value is obtained by inverting the cosine similarity value and used as the edge weight of the fully connected undirected weighted graph. A heterogeneity threshold of 3.5 is set, and the standardized feature Euclidean distance between all node pairs in the fully connected undirected weighted graph is calculated. Node pairs with an Euclidean distance greater than 3.5 are selected as heterogeneous nodes. Auxiliary nodes with the same feature dimension as the original multidimensional temporal feature tensor are constructed, and auxiliary nodes are set... The feature vector of a node is the mean vector of the feature vectors of heterogeneous nodes. An auxiliary node is added to the node set, and the Euclidean distance between the auxiliary node and the heterogeneous node is calculated as the new edge weight. Triple connection paths of heterogeneous node-auxiliary node-heterogeneous node are established, and the fully connected undirected weighted graph is updated. An empty edge set is initialized, and any node is randomly selected to be added to the visited node set. The following operations are repeated: calculate the edge weights between all nodes in the visited node set and all nodes in the unvisited node set, select the unvisited node corresponding to the edge with the smallest weight, add it to the visited node set, and add that edge to the edge set, until all nodes are visited. The minimum cost connection topology is constructed, and the optimized graph topology is generated.A graph convolutional network is used to extract neighborhood features of each node in an optimized graph topology. Feature vectors of all nodes in the optimized graph topology are collected to construct a global feature matrix. Voltage and current feature subspace matrices are separated from the global feature matrix. Principal component analysis (PCA) is used to calculate the first principal component direction vectors of the voltage and current feature subspace matrices, respectively. The cosine of the angle between the two first principal component direction vectors is calculated, and the inverse cosine function is used to obtain the principal component angle. A feature alignment threshold of 0.5 radians is set. When the angle between the principal components is less than 0.5 radians, the voltage feature subspace matrix is ​​projected onto the principal component direction of the current feature subspace matrix, and feature alignment of multi-source heterogeneous data is performed. The mean of the neighborhood features is calculated to fill in missing data points and perform a completion operation to generate a multi-dimensional temporal embedding tensor. A fully connected layer with the feature number of the multi-dimensional temporal embedding tensor as the input dimension and 512 dimensions as the output dimension is constructed. The multi-dimensional temporal embedding tensor is input into this fully connected layer, and a nonlinear transformation is performed using the ReLU activation function. The transformation result is mapped to a high-dimensional space to obtain the model input embedding tensor. The model input is embedded into the first-level decomposition module of the tensor input. The sliding window length is set to 24, and local data segments are truncated along the time axis with a step size of 1. The Euclidean distance between the feature vectors of adjacent time steps in the local data segment is calculated as the amplitude change value. The amplitude change values ​​are arranged in time order to construct the first-order difference sequence of amplitude change. The cosine similarity of the feature vectors of adjacent time steps in the local data segment is calculated. The phase rotation angle value is calculated using the inverse cosine function. The phase rotation angle value is arranged in time order to construct the phase rotation angle sequence. The autocorrelation coefficient of the first-order difference sequence of amplitude change and the autocorrelation coefficient of the phase rotation angle sequence are calculated respectively. The product of the two autocorrelation coefficients is used as the local symmetry measure value. The threshold range for symmetry measurement is set to 0.8 to 1.2. Data segments with local symmetry measurement values ​​within the threshold range are extracted as candidate periodic components. The arithmetic mean of the feature vectors of all data points in the candidate periodic components is calculated to obtain the mean feature vector, which is defined as an invariant of the group. A periodic basis function with a period length equal to the length of the candidate periodic component is constructed to determine the relative position of the current time index within the period. The relative position is substituted into the sine function to calculate a scalar value that continuously varies between -1 and +1 as the basis function value. The basis function value sequence is generated by traversing all time indices. The mean feature vector and the basis function value sequence are multiplied by the outer product to generate the seasonal term kernel function. The model input embedding tensor is subtracted from the mapping result of the seasonal term kernel function to extract the initial seasonal term tensor. The difference between the embedded tensor and the initial seasonal tensor is used to obtain the initial trend tensor. The local symmetry metric of the initial trend tensor is calculated, and the values ​​less than 0.8 are used as residual components for secondary decomposition. Low-frequency trend features are extracted from the residual components and merged into the initial trend tensor, outputting the current-level trend tensor and seasonal tensor. The current-level trend tensor is used as the input signal for the next-level decomposition module. Based on the above cascaded operations, the decomposition is continuously iterated, refining the separation accuracy of symmetry features layer by layer. Decomposition stops when the variance of the trend tensor is less than a set threshold of 0.01, outputting the refined trend tensor and refined seasonal tensor. The autocorrelation precursor capture layer is used to inject a non-Gaussian noise sequence with a set signal-to-noise ratio (SNR) into the fine seasonal tensor. The initial SNR is set to -10 dB, and the noise intensity is gradually increased in a preset step size of 0.5 dB. In each iteration, the following specific operations are performed: a fast Fourier transform is performed on the noisy signal to convert the time-domain signal into frequency-domain complex coefficients. The squared modulus of the complex coefficients is calculated to obtain the energy spectrum. The energy spectrum is divided by the signal length to obtain the power spectral density function. The power spectral density function is traversed to find the spectral line with the largest amplitude as the dominant frequency component, and its power value is recorded as the signal power. The average power of all remaining spectral lines in the power spectral density function after excluding the DC component and the dominant frequency component is calculated and recorded as the background noise power. The logarithm of the ratio of signal power to background noise power is calculated to obtain the output SNR value. The changing trend of the output SNR value is continuously monitored. When the output SNR value is detected to decrease compared to the previous iteration, it is determined that the system has reached the optimal resonance state. The noise parameters of the previous iteration are locked, and the signal in this state is output as the precursor feature tensor after nonlinear enhancement. Using locked noise parameters as reference parameters, a stochastic resonance enhancement operation is performed on the fine trend tensor. The specific steps are as follows: The noise intensity in the reference parameters is set as the initial value; the signal-to-noise ratio (SNR) gain is calculated by scanning the fine trend tensor according to a preset fine-tuning step size; the root mean square (RMS) value of the original fine trend tensor is calculated as the reference signal energy, and the RMS value of the injected noise vector is calculated as the noise energy. The initial SNR is calculated using the ratio of the reference signal energy to the noise energy; subsequently, the RMS value of the processed noisy trend signal is calculated as the output signal energy, and the RMS value of the difference between the output signal and the original signal is calculated as the residual noise energy. The output SNR is calculated using the ratio of the output signal energy to the residual noise energy; the difference between the output SNR and the initial SNR is calculated, and this difference is defined as the SNR gain at the current noise intensity; the above calculation is repeated by changing the noise intensity to obtain the gain values ​​corresponding to different noise intensities, and a curve is plotted with noise intensity as the horizontal axis and SNR gain as the vertical axis; the data point corresponding to the maximum SNR gain in the curve is identified, and the trend feature component corresponding to this data point is extracted and defined as the cumulative trend feature tensor; The adaptive aggregation layer is used to calculate the decay time constant of the autocorrelation function of each feature component in the precursor feature tensor and the cumulative trend feature tensor. Specifically, for each feature component, the variance at zero time lag is calculated as the initial energy. The inner product of the signal and the delayed signal is calculated using different time lag values ​​and divided by the initial energy to obtain the autocorrelation function curve. All time points on the curve are traversed to find the time point where the autocorrelation coefficient value first becomes less than or equal to half of the coefficient value at zero lag. The time lag value corresponding to this time point is extracted as the relaxation time constant. The relaxation time constant is set to be greater than a preset relaxation threshold of 0.5 seconds. The eigenvalues ​​are identified as slow relaxation parameters and defined as order parameters, while the remaining eigenvalues ​​are identified as fast relaxation parameters. The mean-centered data of the order parameter vector is calculated, and the covariance matrix is ​​constructed by multiplying the transpose of the centered data with its own matrix. The eigenvalues ​​and eigenvectors of the covariance matrix are solved, and the eigenvector with the largest eigenvalue is selected to span the preorder parameter quantum space. The inner product of the fast relaxation parameter vector and the eigenvector is calculated, and the inner product result is used as the projection coefficient. The fast relaxation parameters are weighted and summed using the projection coefficient to realize the aggregation operation of the fast relaxation parameters dominated by the order parameters, generating an aggregated feature tensor containing the dominant adaptive pattern. The potential function model of the order parameter is constructed as follows: the sum of squares of each element in the aggregated feature tensor is calculated as the energy value of the quadratic term of the potential function; half of the sum of the fourth powers of each element is calculated as the constraint value of the fourth term; the energy value of the quadratic term is subtracted from the constraint value of the fourth term to obtain the potential function value of the current state; the partial derivative of the potential function value with respect to each element of the aggregated feature tensor is calculated, and the vector composed of all partial derivatives is used as the gradient vector of the potential function; the Euclidean norm of the gradient vector is calculated to obtain the gradient magnitude; the negative gradient magnitude of the natural constant e is calculated to obtain the nonlinear scaling factor; each element of the aggregated feature tensor is divided by the nonlinear scaling factor to achieve nonlinear normalization based on the gradient of the potential function, resulting in the normalized aggregated feature tensor. The mutation prediction regression layer is used to flatten the normalized aggregated feature tensor into a one-dimensional feature vector, which is then input into the fully connected layer for mapping. The Softmax activation function is used to output the probability distribution of the occurrence time of load mutations, and linear regression is used to output the predicted load power value.

[0033] In this embodiment, S6 specifically includes: S61. Calculate the estimated battery output current based on the predicted load power and the current battery voltage. Combine this with the battery internal resistance to calculate the voltage drop value as the voltage drop amplitude. Specifically, read the battery terminal voltage value at the current sampling moment as the current battery voltage, and read the predicted power value for the next moment output by the load power prediction model. Divide the predicted power value by the current battery voltage value to calculate the estimated battery output current. Consult the battery datasheet to obtain the battery internal resistance value, which is 0.05 ohms. Multiply the estimated battery output current value by 0.05 ohms to calculate the voltage drop value, and set this voltage drop value as the voltage drop amplitude.

[0034] S62. Based on the voltage drop amplitude, determine the drift coefficient of Brownian motion, model the voltage drop process as Brownian motion with drift, set a danger threshold as an absorption wall, and calculate the probability density function and expected time when the voltage first touches the danger threshold. Specifically, calculate the difference between the current voltage value and the preset danger threshold voltage value to obtain the safe distance value from the danger threshold; set the negative of the voltage drop amplitude as the drift coefficient value; perform statistical analysis on historical voltage fluctuation data, calculate the standard deviation of voltage fluctuation and set it as the diffusion coefficient value; calculate the square of the safe distance value divided by the square of the diffusion coefficient value to obtain the first intermediate parameter, and calculate the drift coefficient value divided by the square of the diffusion coefficient value to obtain the second intermediate parameter; set... For the time variable, calculate the first intermediate parameter divided by 2, then by the value of pi, then by the cube of the time variable value, and then take the square root to obtain the square root term value. Calculate the first intermediate parameter divided by 2, then by the time variable value to obtain the exponent numerator value. Calculate the exponent numerator value minus the second intermediate parameter multiplied by the time variable value to obtain the exponent parameter. Calculate the natural constant e raised to the power of the exponent parameter to obtain the exponent value. Multiply the square root term value by the exponent value to obtain the probability density value of the voltage first reaching the danger threshold. Set the time step to 0.1 seconds. Within the range of 0 to 1000 seconds, calculate the probability density value every 0.1 seconds. Multiply all the calculated probability density values ​​by the time step of 0.1 seconds and sum them to obtain the desired time value.

[0035] S63. Calculate the risk quantification index using the probability density function of the voltage first reaching the danger threshold and the expected time. When the risk quantification index exceeds the preset safety threshold, dynamically increase the protection cut-off voltage and decrease the output voltage setting value based on the risk quantification index. Specifically, calculate the reciprocal of the expected time value and set this reciprocal as the risk quantification index value; set the preset safety threshold to 100, corresponding to an emergency danger time limit of 10 milliseconds; compare the risk quantification index value with 100. If the risk quantification index value is greater than 100, calculate the difference between the risk quantification index value and 100 to obtain the risk deviation value; multiply the risk deviation value by the preset proportional coefficient 0.01 to obtain the voltage adjustment step value; add the voltage adjustment step value to the current protection cut-off voltage value to update it to a new protection cut-off voltage; subtract the voltage adjustment step value from the current output voltage setting value to update it to a new output voltage setting value.

[0036] In this embodiment, S7 specifically includes: S71. Encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value, and output voltage setpoint into a JSON message, and assign a corresponding service type tag to the JSON message. Specifically, read the line status monitoring results; if a fault exists, generate a value of 1 as the poor line contact warning signal; otherwise, generate a value of 0. Read the current communication protocol version number and use the version number string as the protocol type identifier. Read the value output by the load power prediction model as the load power prediction value, and read the value stored in the control module as the output voltage setpoint. Construct a key-value pair set, assigning the value 1 or 0 to the key named "alarm_". The "status" field is assigned the version number string to the "protocol_id" field, the load power prediction value to the "power_pred" field, and the output voltage setting value to the "voltage_set" field. A message string conforming to JSON format is generated through string concatenation. The preset service type mapping table is traversed. When the value of the "alarm_status" field in the message is 1, the string "emergency" is assigned as the service type label. When the value is 0, the string "normal" is assigned as the service type label.

[0037] S72. Construct a partial order set based on the service type tags, calculate the comparability relationship between each element in the partial order set, draw a Hasse diagram based on the comparability relationship, and determine the transmission priority sequence. Specifically, create an empty message set as the partial order set, insert all JSON messages to be sent as elements into the partial order set; define the partial order relationship operation rules, set the priority weight values ​​for the service type tags, set the weight value of the "emergency" tag to 1, and the weight value of the "normal" tag to 0; traverse any two message elements in the partial order set one by one, and calculate the difference in the priority weight values ​​of the service type tags for the two messages. If the difference is greater than 0, the preceding message is determined to be superior to the following message, establishing a comparability relationship. If the difference is equal to 0, the difference between the timestamp values ​​of the two messages is further calculated. If the difference between the timestamp values ​​is less than 0, the following message is determined to be superior to the preceding message, establishing a comparability relationship. All established comparability relationships are connected to form directed edges, and a Hasse graph is drawn. The in-degree value of each node element in the Hasse graph is calculated. The node with an in-degree value of 0 and that is not marked is selected as the current highest priority message. It is removed from the partial order set and added to the transmission priority sequence. This selection operation is repeated until the partial order set is empty, resulting in a complete transmission priority sequence.

[0038] S73. Upload JSON messages to the cloud server via the MQTT protocol according to the transmission priority sequence, driving the cloud-based visual interface to display the line status and control parameters. Specifically, initialize the MQTT client parameters, setting the server IP address to "192.168.1.100", the port number to "1883", and the client ID to "Device_001"; call the connection function to establish a TCP long connection, calculate the MD5 hash value of the username and password to complete authentication; extract the JSON message string one by one from the sequence header according to the transmission priority sequence; calculate the byte length of the JSON message string, write the byte length value into the remaining length field in the fixed header, and construct the MQTT data. The system calls the publish function to send an MQTT data packet to a channel with the topic "Battery / System / Status". The cloud server receives and parses the MQTT data packet, extracts the value of the "alarm_status" field, and if the value is equal to 1, calculates the screen coordinates (100, 200) and draws a flashing red circle icon at that coordinate position on the visualization interface to indicate the warning status. The system extracts the values ​​of the "power_pred" and "voltage_set" fields, plots the data points in the trend chart control of the visualization interface with the current system timestamp as the horizontal axis and the field values ​​as the vertical axis, and connects the lines to update and display the changing trend of the load power prediction value and the output voltage setting value in real time.

[0039] refer to Figure 3A multifunctional mobile power bank lighting control system based on deep learning includes the following modules: The electrical signal feature extraction module is used to collect voltage and current waveform data of the lighting link, segment it into multiple frequency band components and calculate the energy distribution density, generate an energy entropy spectrum vector sequence based on information entropy, and combine them to form an electrical signal feature vector sequence. The poor line contact early warning module is used to construct the covariance matrix of the electrical signal feature vector sequence and calculate the eigenvalue spectrum. When the spectral gap narrows and causes the cluster structure to collapse, a poor line contact early warning signal is generated. The protocol type identification module is used to acquire physical layer signal waveforms, extract the phase consistency matrix using complex wavelet transform and map it to waveform fingerprint vectors, calculate the matching probability and output the protocol type identifier; The lighting control command parsing module is used to retrieve the preset parsing rule table according to the protocol type identifier, map the logical fields of the command frame to nodes and the timing jumps to directed edges to construct a directed graph, and parse the lighting control command after verifying the Eulerian circuit. The load mutation prediction module is used to call up voltage, current and temperature data from historical time periods associated with electrical signal feature vector sequences, and construct a multi-dimensional time series feature matrix by combining lighting control commands. The improved Autoformer model is input, and after graph embedding, cascade decomposition, stochastic resonance enhancement and sequence parameter aggregation, the predicted power value is output. The adaptive protection control module is used to calculate the estimated value of battery output current based on the load power prediction value, model the voltage drop process as Brownian motion, calculate the probability distribution of the voltage first reaching the dangerous threshold, and dynamically adjust the protection cut-off voltage and output voltage set value. The data transmission and visualization module is used to encapsulate poor line contact warning signals, protocol type identifiers, load power prediction values, and output voltage setting values ​​into JSON messages, construct a partial order set, determine transmission priority using a Hasse diagram, and upload them to the cloud for visualization via MQTT.

[0040] Example 1: To verify the feasibility of this invention in intelligent control of lighting in portable power banks, the method of this invention was applied to the R&D project of a new generation of multifunctional portable power banks for a well-known outdoor energy storage equipment manufacturer (hereinafter referred to as "Company E"). In traditional portable power bank lighting control systems, simple control logic based on fixed threshold comparison or basic PWM dimming technology are typically used. These methods struggle to accurately identify complex lighting load protocols and cannot provide early warnings of poor contact based on the frequency domain characteristics of electrical signals, easily leading to battery over-discharge damage due to sudden load changes or fire hazards caused by circuit faults. To solve these problems, Company E decided to adopt the multifunctional portable power bank lighting control system and method based on deep learning proposed in this invention.

[0041] During implementation, Enterprise E first used voltage and current sensors integrated into the power bank's output to collect real-time waveform data of the lighting link. Through frequency band energy segmentation, information entropy calculation, and covariance matrix spectral gap analysis, it constructed an electrical signal feature vector sequence and successfully established an index mapping relationship between the feature vectors and historical multidimensional data. Enterprise E's technical team built a fingerprint database containing various mainstream lighting protocols (such as DMX512 and WS2812) and annotated the collected historical operating condition data with power mutations, serving as a benchmark for improving Autoformer model training and risk assessment.

[0042] Enterprise E extracts the phase consistency matrix of the physical layer signal waveform using complex wavelet transform and maps it to a fingerprint vector, achieving millisecond-level identification of protocol types. It also verifies the Eulerian circuit by constructing a directed graph of command states, completing the logical verification and accurate parsing of lighting control commands. Next, it constructs a multi-dimensional time-series feature matrix using associated historical data, inputs it into an improved Autoformer model, and through graph embedding alignment in the multi-dimensional time-series embedding layer, trend and seasonal term separation in the cascaded decomposition network layer, stochastic resonance enhancement in the autocorrelation precursor capture layer, and order parameter aggregation in the adaptive aggregation layer, outputs a high-precision load power prediction value.

[0043] In the core protection and interaction phase, this invention calculates the predicted battery output current based on power prediction values ​​and calculates the probability distribution of voltage reaching the dangerous threshold based on a Brownian motion model, achieving millisecond-level dynamic adjustment of protection parameters. Subsequently, the system encapsulates the status data into JSON messages, uses a Hasse diagram to determine transmission priority, and uploads it to the cloud via the MQTT protocol, achieving visual interaction. This system realizes closed-loop control throughout the entire process, from signal acquisition, command parsing, load prediction to adaptive protection.

[0044] During implementation, the technical team at Company E discovered that, compared to traditional mobile power bank light control methods, the method of this invention significantly improves the system's intelligence and security. Traditional methods cannot parse light commands using non-standard protocols and exhibit lag in response to sudden load changes. In contrast, the method of this invention, through deep learning models and Brownian motion risk assessment, effectively achieves accurate perception of load characteristics and proactive safety protection.

[0045] To further verify the actual performance of the method of the present invention, Company E conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1 Performance Comparison of Enterprise E-Mobile Power Bank Lighting Control System

[0046] As shown in Table 1, the performance of the mobile power bank lighting control system was comprehensively improved after applying the method of this invention. Protocol recognition accuracy increased from 85.2% with traditional methods to 99.1%, and load power prediction error decreased from 5.8 watts to 0.6 watts, significantly improving the perception accuracy of the control system and providing a reliable basis for refined energy management. The recall rate for poor line contact warnings increased dramatically from 72.5% to 98.4%, effectively avoiding the risk of electrical fires caused by poor contact. Overcurrent protection response time was shortened from 50 milliseconds to 8 milliseconds, significantly enhancing system timeliness and greatly ensuring battery safety. Furthermore, cloud data transmission packet loss rate decreased from 4.5% to 0.1%, battery cycle life loss was significantly reduced, and user interaction response satisfaction increased from 82.0% to 97.5%.

[0047] Through the method of this invention, Enterprise E has successfully achieved intelligent identification and active safety protection for the mobile power supply lighting system, effectively solving the problems of single control, slow response and large safety hazards of traditional methods, greatly improving the market competitiveness and user trust of the product, and providing strong technical support for the intelligent upgrading of outdoor energy storage equipment.

[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multifunctional mobile power bank lighting control method based on deep learning, characterized in that, Specifically, the following steps are included: S1. Collect voltage and current waveform data of the lighting link, segment it into multiple frequency band components and calculate the energy distribution density, generate an energy entropy spectrum vector sequence based on information entropy, and combine them to form an electrical signal feature vector sequence; S2. Construct the covariance matrix of the electrical signal feature vector sequence and calculate the eigenvalue spectrum. When the spectral gap narrows and the clustering structure collapses, generate a line contact failure warning signal. S3. Acquire physical layer signal waveforms, extract phase consistency matrix using complex wavelet transform and map it to waveform fingerprint vector, calculate matching probability and output protocol type identifier; S4. Retrieve the preset parsing rule table according to the protocol type identifier, map the logical fields of the instruction frame to nodes and the timing jumps to directed edges to construct a directed graph, and parse the light control instruction after verifying the Euler circuit. S5. Call the voltage data, current data and temperature data of the historical time period associated with the electrical signal feature vector sequence, and construct a multi-dimensional time series feature matrix in combination with the lighting control command. Input the improved Autoformer model, and output the load power prediction value through graph embedding, cascade decomposition, stochastic resonance enhancement and sequence parameter aggregation. S6. Calculate the battery output current estimate based on the load power prediction value, model the voltage drop process as Brownian motion, calculate the probability distribution of the voltage first touching the danger threshold, and dynamically adjust the protection cut-off voltage and output voltage set value. S7. Encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into a JSON message, construct a partial order set and use a Hasse diagram to determine the transmission priority, and upload it to the cloud for visualization via MQTT.

2. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. The voltage waveform data and current waveform data of the lighting link are collected by the voltage sensor and the current sensor at a preset sampling frequency. The collected analog signal is converted into a digital signal sequence and the timestamp of the collection time is recorded at the same time. S12. Calculate the signal energy value of the digital signal sequence in different frequency bands as the energy distribution density of each frequency band component; S13. Calculate the information entropy based on the energy distribution density of each frequency band component, construct the energy entropy spectrum vector based on the information entropy, and combine the energy entropy spectrum vectors corresponding to voltage and current according to the time sequence to form an electrical signal feature vector sequence. S14. Extract the timestamp index of the electrical signal feature vector sequence, retrieve the original voltage data, current data and temperature data under the corresponding time window from the historical database, and establish the electrical feature vector-multidimensional historical data index mapping relationship.

3. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Construct the covariance matrix of the electrical signal feature vector sequence, calculate the eigenvalues ​​of the covariance matrix, and arrange the eigenvalues ​​in descending order of their numerical values ​​to form an eigenvalue spectrum; S22. Calculate the difference between adjacent feature values ​​in the feature value spectrum to obtain the feature value interval sequence, and detect the maximum interval in the feature value interval sequence as the spectral gap; S23. Monitor the trend of the spectral gap change over time. When the spectral gap value decreases to below a preset threshold, it is determined that the cluster structure has collapsed, and a poor line contact warning signal is generated.

4. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Acquire physical layer signal waveforms, perform complex wavelet transform on the physical layer signal waveforms, calculate phase information at different scales, and combine them to generate a phase consistency matrix. S32. Perform dimensionality reduction and vectorization mapping on the phase consistency matrix to generate a waveform fingerprint vector for representing waveform features; S33. Calculate the similarity between the waveform fingerprint vector and each standard fingerprint in the preset protocol fingerprint library, convert the similarity into a matching probability, and select the protocol type with the highest matching probability as the protocol type identifier.

5. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Retrieve the preset parsing rule table according to the protocol type identifier to obtain the corresponding instruction frame format definition and timing logic rules; S42. Extract logical fields as nodes according to the instruction frame format definition, determine the jump relationship between nodes as directed edges according to the timing logic rules, and construct an instruction state directed graph. S43. Search the directed graph of the instruction state for a valid path from the start node to the end node. If a valid path exists, determine that the instruction structure is complete and parse the light control instruction. If a valid path does not exist, determine that the instruction structure is abnormal.

6. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, The improved Autoformer model specifically includes a multi-dimensional temporal embedding layer, a cascaded decomposition network layer, an autocorrelation precursor capture layer, an adaptive aggregation layer, and a mutation prediction regression layer: The multidimensional temporal embedding layer is used to receive the electrical signal feature vector sequence. Based on the mapping relationship between the electrical feature vector and the multidimensional historical data index, it retrieves the original voltage data, current data, and temperature data associated with the mapping through the timestamp of the electrical signal feature vector. The original voltage data, current data, and temperature data are concatenated along the feature dimension to construct the original multidimensional temporal feature tensor. The lighting control command is mapped to the command embedding vector as the edge weight constraint. The minimum cost connection topology is constructed with data points as graph nodes. By introducing auxiliary nodes, the connection path of heterogeneous nodes is optimized to generate an optimized graph topology. The sparse low-frequency data is aligned and completed to dense high-frequency data through graph convolution aggregation to generate a multidimensional temporal embedding tensor and map it to a high-dimensional space to obtain the model input embedding tensor. The cascaded decomposition network layer is used to embed the model input into the tensor input cascaded decomposition module. It constructs a symmetry metric by calculating the autocorrelation coefficient of local amplitude changes and phase rotation through a sliding window. It maps the periodic components that meet the symmetry threshold to group invariants and extracts the seasonal term. It extracts the trend term through the residual components of the symmetry metric. It uses the seasonal term and the trend term as inputs to the next level decomposition module to refine the separation accuracy of the symmetry features layer by layer, and outputs a fine trend tensor and a fine seasonal tensor. The autocorrelation precursor capture layer is used to inject a non-Gaussian noise sequence with a set signal-to-noise ratio into the fine seasonal tensor, calculate the peak value of the cross-correlation function between the noisy signal and the original signal, adjust the noise intensity until the peak value of the cross-correlation function reaches a maximum value, and output the precursor feature tensor after nonlinear enhancement. Calculate the signal-to-noise ratio gain curves under different noise intensities for the fine trend tensor, and extract the feature component corresponding to the maximum signal-to-noise ratio gain as the cumulative trend feature tensor; The adaptive aggregation layer is used to identify the slow relaxation parameters in the precursor feature tensor and the cumulative trend feature tensor as order parameters, project the trend features of the fast relaxation parameters onto the subspace generated by the precursor order parameters for dominant aggregation, and generate an aggregated feature tensor containing the dominant aggregation pattern; the aggregated feature tensor is nonlinearly normalized by calculating the potential function gradient of the order parameters to obtain the normalized aggregated feature tensor. The mutation prediction regression layer is used to flatten the normalized aggregated feature tensor into a one-dimensional feature vector, which is then input into the fully connected layer for mapping; the probability distribution of the occurrence time of load mutation is output through the Softmax activation function, and the predicted load power value is output through linear regression.

7. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Calculate the estimated value of battery output current based on the predicted load power value and the current battery voltage, and calculate the voltage drop value as the voltage drop amplitude by combining the battery internal resistance. S62. Based on the voltage drop amplitude, determine the Brownian motion drift coefficient, model the voltage drop process as a drifting Brownian motion, set the danger threshold as the absorption wall, and calculate the probability density function and expected time when the voltage first touches the danger threshold. S63. Calculate the risk quantification index using the probability density function of the voltage first reaching the danger threshold and the expected time. When the risk quantification index exceeds the preset safety threshold, dynamically increase the protection cut-off voltage and decrease the output voltage setting value according to the risk quantification index.

8. The method for controlling the lighting of a multifunctional mobile power bank based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into a JSON message, and assign a corresponding service type label to the JSON message; S72. Construct a partial order set based on the service type label, calculate the comparability relationship of each element in the partial order set, draw a Hasse diagram based on the comparability relationship, and determine the transmission priority sequence. S73. Upload the JSON message to the cloud server via the MQTT protocol according to the transmission priority sequence, and drive the cloud visualization interface to display the line status and control parameters.

9. A deep learning-based multifunctional mobile power bank lighting control system, executing the deep learning-based multifunctional mobile power bank lighting control method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The electrical signal feature extraction module is used to collect voltage and current waveform data of the lighting link, segment it into multiple frequency band components and calculate the energy distribution density, generate an energy entropy spectrum vector sequence based on information entropy, and combine them to form an electrical signal feature vector sequence. The poor line contact early warning module is used to construct the covariance matrix of the electrical signal feature vector sequence and calculate the eigenvalue spectrum. When the spectral gap narrows and causes the clustering structure to collapse, a poor line contact early warning signal is generated. The protocol type identification module is used to acquire physical layer signal waveforms, extract the phase consistency matrix using complex wavelet transform and map it to waveform fingerprint vectors, calculate the matching probability and output the protocol type identifier; The lighting control command parsing module is used to retrieve a preset parsing rule table according to the protocol type identifier, map the logical fields of the command frame to nodes and the timing jumps to directed edges to construct a directed graph, and parse the lighting control command after verifying the Euler circuit. The load mutation prediction module is used to call the voltage data, current data and temperature data of the historical time period associated with the electrical signal feature vector sequence, and combine them with the lighting control command to construct a multi-dimensional time series feature matrix. The improved Autoformer model is input, and after graph embedding, cascade decomposition, stochastic resonance enhancement and sequence parameter aggregation, the predicted power value is output. The adaptive protection control module is used to calculate the estimated value of battery output current based on the predicted load power value, model the voltage drop process as Brownian motion, calculate the probability distribution of the voltage first touching the danger threshold, and dynamically adjust the protection cut-off voltage and output voltage set value. The data transmission and visualization module is used to encapsulate the poor line contact warning signal, protocol type identifier, load power prediction value and output voltage setting value into JSON messages, construct a partial order set and use a Hasse diagram to determine the transmission priority, and upload it to the cloud for visualization display via MQTT.

Images