Portable charger overheating protection method based on multi-source sensing

Abstract

The invention discloses a portable charger overheating protection method based on multi-source sensing, and relates to the technical field of portable chargers, and the method comprises the following steps: S1, data collection and preprocessing; s2, introducing a probabilistic modeling architecture based on a DA-RNN model, and extracting a probabilistic hidden state vector sequence of the encoder through a probabilistic cycle unit; s3, extracting a prediction parameter sequence of a future time step of the decoder through a probabilistic cycle unit; s4, constructing probability distribution of future time step prediction parameters, and extracting a comprehensive time sequence feature vector; s5, executing security risk level assessment, and generating a risk processing strategy; and S6, deploying a risk processing strategy. The method overcomes the limitations of dependence on a single temperature threshold, response lag and lack of predictability in a traditional charger protection method, and provides an efficient and intelligent solution for safe and reliable operation of portable charging equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of portable chargers, and particularly relates to a portable charger overheat protection method based on multi-source sensing. BACKGROUND

[0002] With the widespread use of portable charging devices, the requirement for their safety is also increasing. Traditional charger overheat protection methods mostly rely on simple temperature threshold judgment or regular inspection. Although these methods can prevent overheating risks to some extent, they have significant limitations, especially in situations where chargers are used frequently and the environment is complex. Existing overheat protection technologies often ignore the dynamic changes during charger operation, making it difficult to monitor and predict potential overheating risks in real time. Therefore, improving the overheat protection capabilities of chargers, especially improving prediction accuracy and response speed, has become a core requirement for the development of intelligent chargers.

[0003] Currently, most portable charger overheat protection methods rely on fixed temperature thresholds or preset rules to determine whether overheating occurs. This approach has obvious shortcomings. The fixed threshold method usually cannot adapt to changes in the charger under different working environments, leading to false positives or negatives. In addition, methods based on simple rules cannot dynamically adjust the working state of the charger, resulting in the device failing to respond in time before the temperature threshold point during charging. Therefore, traditional overheat protection methods are difficult to meet the needs of intelligent and real-time management, lacking intelligent analysis and adaptive adjustment capabilities for device state data.

[0004] The main limitation of traditional methods is the lack of real-time and accuracy, especially when device state data exhibits high dimensionality, multi-source, nonlinearity, and other characteristics. The generalization ability of traditional methods is poor. When the temperature, current, power, and voltage of the charger and other state data change simultaneously, simple single threshold methods cannot efficiently and accurately identify overheating risks, making it difficult to discover and handle overheating problems in a timely manner, seriously affecting the safety and reliability of the device. Especially in complex environments and with large fluctuations in charger load, the prediction accuracy and response speed of traditional methods cannot meet the needs of real-time overheat protection.

[0005] In addition, traditional overheat protection methods often ignore the relevance between multi-source sensor data, especially the interaction between multi-dimensional data such as temperature, current, power, and voltage. In the complex working environment of the charger, traditional methods cannot fully utilize the time series changes and multi-dimensional characteristics of these data, resulting in a lag in the response of the protection mechanism, a high false positive rate, and difficulty in accurate fault warning in some cases. Even with machine learning or deep learning techniques, existing methods still fail to effectively exploit the deep relevance of sensor data, making it difficult to provide efficient, accurate, and real-time overheat protection solutions.

[0006] Therefore, how to provide a portable charger overheat protection method based on multi-source sensing is a problem that those skilled in the art urgently need to solve. SUMMARY

[0007] The application provides a portable charger overheat protection method based on multi-source sensing, which can more accurately predict and evaluate the thermal runaway risk in the charging process. This method not only automatically extracts deep temporal features of multi-source state data such as battery core temperature, mainboard temperature, interface temperature, voltage, current and power, but also combines time attention mechanism with probabilistic recurrent units through an improved DA-RNN model to realize probabilistic prediction and uncertainty quantification of future temperature trends. By introducing confidence upper limit evaluation and two-dimensional risk classification strategy, this method fuses the expected value and uncertainty information of the prediction, dynamically evaluates the safety risk level, and formulates differentiated protection strategies such as power reduction, fan start-stop or suspension of charging, thereby realizing intelligent protection from passive response to active prevention. This method overcomes the limitations of traditional charger protection methods that rely on a single temperature threshold, have a lagging response and lack foresight, providing an efficient and intelligent solution for the safe and reliable operation of portable charging devices.

[0008] The portable charger overheat protection method based on multi-source sensing according to the embodiments of the application comprises the following steps:

[0009] S1, real-time acquisition of charging state raw data of a portable charger by a sensor, and preprocessing of the charging state raw data to generate a charging state time series data set;

[0010] S2, inputting the charging state time series data set into an encoder of an improved DA-RNN model, introducing a probabilistic modeling architecture, and extracting a probabilistic hidden state vector sequence of the encoder through a probabilistic recurrent unit;

[0011] S3, inputting the probabilistic hidden state vector sequence of the encoder into a decoder of the improved DA-RNN model, generating a probabilistic context vector based on a time attention mechanism, and extracting a prediction parameter sequence of a future time step of the decoder through a probabilistic recurrent unit;

[0012] S4, constructing a probability distribution of the prediction parameter of the future time step according to the prediction parameter sequence of the future time step of the decoder, and extracting a comprehensive time series feature vector;

[0013] S5, safety risk level evaluation according to the comprehensive time series feature vector, and formulating a corresponding risk handling strategy according to the safety risk level;

[0014] S6, deploy the risk handling strategy to the charging management system, compare historical data and real-time data, and generate a risk handling strategy implementation effect report.

[0015] Optionally, the charging state raw data includes battery core temperature, battery mainboard temperature, output interface temperature, environment temperature, input / output voltage, input / output current, charging power, power factor, charging time, battery current percentage; and the preprocessing includes data cleaning, space-time alignment, standardization, data smoothing and denoising, sliding window segmentation, and feature extraction.

[0016] Optionally, the probabilistic recurrent unit specifically includes:

[0017] The attention-weighted context vector and the mean of the probabilistic hidden state vector of the previous moment are spliced to generate a spliced time sequence feature vector;

[0018] The spliced time sequence feature vector is input into the LSTM to obtain the calculation results of the forget gate, the input gate, the output gate, and the candidate cell state;

[0019] The spliced time sequence feature vector is linearly converted by the full connection layer and activated by the Softplus function to obtain the variance of the candidate cell state;

[0020] The mean of the probabilistic hidden state vector of the previous moment is multiplied by the calculation result of the forget gate to obtain a mean forgetting element; the candidate cell state is multiplied by the calculation result of the input gate to obtain a mean update element; and the mean forgetting element is added to the mean update element to obtain the mean of the cell state at the current moment;

[0021] The variance of the cell state of the previous moment is multiplied by the square of the calculation result of the forget gate to obtain a variance forgetting element; the variance of the candidate cell state at the current moment is multiplied by the square of the calculation result of the input gate to obtain a variance update element; and the variance forgetting element is added to the variance update element to obtain the variance of the cell state at the current moment;

[0022] The calculation result of the output gate is multiplied by the tangent value of the mean of the cell state at the current moment to obtain the mean of the probabilistic hidden state vector at the current moment;

[0023] The square of the tangent value of the mean of the cell state at the current moment is calculated as a cell state mean attenuation factor, and a cell state mean residual factor is obtained by subtracting the cell state mean attenuation factor from 1; the square of the calculation result of the output gate is multiplied by the variance of the cell state at the current moment, multiplied by the square of the cell state mean residual factor, to obtain the variance of the probabilistic hidden state vector at the current moment;

[0024] Concatenate the mean and variance of the current time probability hidden state vector to obtain the current time probability hidden state vector.

[0025] Optionally, the S2 specifically comprises:

[0026] S21, set the initial value of the mean and the initial value of the variance of the probability hidden state vector;

[0027] S22, for each time step in the charging state time series data set, based on the mean of the probability hidden state vector of the last time, using a multi-layer perception as a feedforward neural network, the energy score of all charging state time series vectors from the first time step to the current time step is calculated by forward propagation;

[0028] S23, normalize the energy score by the Softmax function to obtain the attention weight vector;

[0029] S24, based on the attention weight, the charging state time series vectors from the first time step to the current time step are weighted and summed to generate the attention weighted context vector of the current time step;

[0030] S25, input the attention weighted context vector and the mean of the probability hidden state vector of the last time step of the encoder into the probability recurrent unit to obtain the probability hidden state vector of the current time;

[0031] S26, repeat the steps of S22 to S25 in time step order until all time steps are processed, and concatenate the probability hidden state vectors of each time step in time step order to obtain the probability hidden state vector sequence of the encoder.

[0032] Optionally, the S3 specifically comprises:

[0033] S31, set the initial value of the mean and the initial value of the variance of the probability hidden state vector of the decoder;

[0034] S32, for each future time step that needs to be predicted, based on the entire probability hidden state vector sequence of the encoder and the mean of the probability hidden state vector of the last time of the decoder, using a multi-layer perception as a feedforward neural network, the energy score of each time step in the sequence is calculated by forward propagation;

[0035] S33, normalize the energy score by the Softmax function to obtain the time attention weight vector;

[0036] S34, based on the time attention weight vector, the mean of the probability hidden state vector sequence is weighted and summed to generate the time attention weighted context vector;

[0037] S35, input the time attention weighted context vector and the mean of the probabilistic hidden state vector of the previous time step of the decoder into the probabilistic recurrent unit to obtain the prediction parameter of the current prediction time step;

[0038] S36, repeat the steps of S32 to S35 for each future time step that needs to be predicted until the prediction of all future time steps is completed to obtain the prediction parameter sequence of the future time steps of the decoder.

[0039] Optionally, the S4 specifically comprises:

[0040] S41, extract the prediction parameter of each future time step from the prediction parameter sequence of the future time steps of the decoder;

[0041] S42, based on each future time step, construct a Gaussian distribution using the mean and variance corresponding to the prediction parameter;

[0042] S43, based on the Gaussian distribution, extract the expected value and variance value of each future time step;

[0043] S44, concatenate the corresponding variance value after the expected value of each time step to obtain the feature of the current time step;

[0044] S45, sequentially concatenate the features of all time steps in time order to obtain a comprehensive time sequence feature vector.

[0045] Optionally, the S5 specifically comprises:

[0046] S51, based on the comprehensive time sequence feature vector, obtain the expected value and variance value of each time step, and obtain the standard deviation of each time step by taking the square root of the variance value;

[0047] S52, set a confidence level coefficient, multiply the standard deviation by the confidence factor, and add the expected value to obtain the upper confidence limit of the current time step;

[0048] S53, count the upper confidence limits of all time steps to obtain an upper confidence limit list, count the expected values of all time steps to obtain an expected value list, and count the variance values of all time steps to obtain a variance value list;

[0049] S54, according to historical experience and expert advice, set an expected threshold and a safety threshold, if there is an element in the expected value list that exceeds the expected threshold, record it as an expected risk element; if there is an element in the upper confidence limit list that exceeds the safety threshold, record it as a confidence risk element;

[0050] S55, set the expected safety threshold, the expected danger threshold, the confidence safety threshold, the confidence danger threshold; if the value of the expected risk element is higher than the expected danger threshold, or the value of the confidence risk element is higher than the confidence danger threshold, it is confirmed that the risk level is high risk; if the value of the expected risk element is lower than the expected danger threshold and higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold and higher than the confidence safety threshold, it is confirmed that it is medium risk; otherwise, it is confirmed to be low risk;

[0051] S56, according to the safety risk level, the corresponding risk processing strategy is formulated, including reducing the charging power, starting the low-speed cooling fan, suspending the charging, and starting the full-speed cooling system.

[0052] Optionally, the S6 specifically comprises:

[0053] S61, the risk processing strategy is deployed to the charging control system, and a risk event log is recorded, including an event ID, a time of occurrence, an expected risk element and a confidence risk element, a corresponding risk processing strategy, and charging state original data;

[0054] S62, after the risk processing strategy is deployed, the charging state original data is collected in real time, compared with historical normal data, and the risk removal time, the charging time loss value, the energy loss value, and the parameter change value of the charging state original data are counted;

[0055] S63, based on the risk removal time, the charging time loss value, the energy loss value, and the parameter change value of the charging state original data, a risk processing strategy implementation effect report is generated.

[0056] The beneficial effects of the present application are:

[0057] (1) The present application introduces a probabilistic modeling architecture and designs a probabilistic recurrent unit, which fundamentally improves the depth and reliability of the overheat risk prediction of the portable charger. The hidden state of the traditional recurrent neural network is deterministic when processing time series data, which cannot reflect the inherent random noise in the system operation and the uncertainty of the model itself, resulting in overly "confident" prediction results and insufficient robustness when facing abnormal working conditions. To solve this problem, the present application innovatively designs a probabilistic recurrent unit, which expands the hidden state from a single vector to a probability distribution defined by mean and variance. This unit uses a double-track parallel structure to update the state mean while strictly following the mathematical principles of uncertainty propagation to update the state variance simultaneously. This design enables the model to not only provide the most likely expected value when predicting future temperature, but also provide a confidence interval, thereby accurately quantifying the certainty of the prediction. This enables the protection system to effectively distinguish between "stable high temperature" and "risk temperature with severe fluctuations", greatly enhancing the accuracy of the early warning and the reliability of the decision-making.

[0058] (2) The improved DA-RNN model is adopted to effectively solve the technical problems of multi-source heterogeneous data feature fusion and key information capture in the charging process. The traditional protection method usually relies on a single temperature sensor for threshold judgment, which is difficult to comprehensively reflect the complex thermal dynamic changes under the joint action of voltage, current, power and other factors. The improved DA-RNN model can adaptively focus on the historical data segments and sensor channels that have the most critical impact on temperature changes through its encoder-decoder structure and the fusion of input attention and time attention mechanism, and automatically mine the deep association and time sequence dependence between multi-source data. This method breaks through the limitations of traditional methods in information utilization, greatly improves the depth of model understanding and prediction accuracy of charging state evolution, and provides a solid technical foundation for realizing intelligent active protection. BRIEF DESCRIPTION OF DRAWINGS

[0059] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0060] Figure 1 The overall flowchart of the portable charger overheat protection method based on multi-source sensing proposed by the application;

[0061] Figure 2 The working principle flowchart of the probabilistic recurrent unit of the improved DA-RNN model of the portable charger overheat protection method based on multi-source sensing proposed by the application. DETAILED DESCRIPTION

[0062] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0063] Reference Figure 1 and Figure 2 The portable charger overheat protection method based on multi-source sensing comprises the following steps:

[0064] S1, real-time acquisition of charging state raw data of the portable charger through the sensor, and preprocessing of the charging state raw data to generate charging state time series data set;

[0065] S2, inputting the charging state time series data set into the encoder of the improved DA-RNN model, introducing a probabilistic modeling architecture, and extracting a probabilistic hidden state vector sequence of the encoder through a probabilistic recurrent unit;

[0066] S3, input the probabilistic hidden state vector sequence of the encoder into the decoder of the improved DA-RNN model, generate a probabilistic context vector based on a time attention mechanism, and extract a prediction parameter sequence of a future time step of the decoder through a probabilistic recurrent unit;

[0067] S4, construct a probability distribution of the prediction parameter of the future time step according to the prediction parameter sequence of the future time step of the decoder, and extract a comprehensive time sequence feature vector;

[0068] S5, perform safety risk level evaluation according to the comprehensive time sequence feature vector, and formulate a corresponding risk treatment strategy according to the safety risk level;

[0069] S6, deploy the risk treatment strategy to the charging management system, compare historical data and real-time data, and generate a risk treatment strategy implementation effect report.

[0070] In the embodiment, the charging state raw data includes battery core temperature, battery mainboard temperature, output interface temperature, environment temperature, input and output voltage, input and output current, charging power, power factor, charging time, and battery current percentage; and the preprocessing includes data cleaning, space-time alignment, standardization, data smoothing and denoising, sliding window segmentation, and feature extraction.

[0071] The embodiment constructs a high-quality and multi-dimensional data basis for subsequent accurate prediction through comprehensive data collection and fine preprocessing process. The charging state raw data includes battery core temperature, battery mainboard temperature, output interface temperature, environment temperature, input and output voltage, input and output current, charging power, power factor, charging time, and battery current percentage, which ensures the comprehensiveness and representativeness of the data source. The preprocessing includes data cleaning to eliminate outliers and missing values and ensure the effectiveness of the data. The preprocessing includes space-time alignment to unify the timestamps and spatial references of different sensors and ensure the consistency of the data. The preprocessing includes standardization to eliminate the influence of different physical dimensions and improve the model convergence speed. The preprocessing includes data smoothing and denoising to filter out high-frequency interference and highlight the real trend of the data. The preprocessing includes sliding window segmentation to convert continuous data into time sequence samples suitable for model input. The preprocessing includes feature extraction to mine deeper statistical and correlation features from the original data and enhance the information input of the model.

[0072] In the embodiment, the probabilistic recurrent unit specifically includes:

[0073] The context vector weighted by attention and the mean of the probabilistic hidden state vector of the previous time are spliced to generate a spliced time sequence feature vector;

[0074] The spliced time sequence feature vector is input into the LSTM to obtain calculation results of a forget gate, an input gate, an output gate, and a candidate cell state;

[0075] The spliced time sequence feature vector is linearly converted by a fully connected layer and activated by a Softplus function to obtain a variance of the candidate cell state;

[0076] The mean forgetting element is obtained by multiplying the calculation result of the forget gate by the mean of the probabilistic hidden state vector at the previous moment; the mean update element is obtained by multiplying the calculation result of the input gate by the candidate cell state; and the cell state mean at the current moment is obtained by adding the mean forgetting element to the mean update element;

[0077] The variance forgetting element is obtained by multiplying the variance of the cell state at the previous moment by the square of the calculation result of the forget gate; the variance update element is obtained by multiplying the variance of the candidate cell state at the current moment by the square of the calculation result of the input gate; and the cell state variance at the current moment is obtained by adding the variance forgetting element to the variance update element;

[0078] The cell state mean at the current moment is multiplied by the tangent value of the output gate calculation result to obtain the mean of the probabilistic hidden state vector at the current moment;

[0079] The square of the tangent value of the cell state mean at the current moment is calculated as a cell state mean decay factor, and the cell state mean residual factor is obtained by subtracting the cell state mean decay factor from 1; the square of the output gate calculation result is multiplied by the cell state variance at the current moment, multiplied by the square of the cell state mean residual factor, to obtain the variance of the probabilistic hidden state vector at the current moment;

[0080] The mean and variance of the probabilistic hidden state vector at the current moment are spliced to obtain the probabilistic hidden state vector at the current moment.

[0081] The embodiment realizes the synchronous and accurate propagation and update of the mean value of the time sequence feature and the uncertainty by constructing a double-track parallel probabilistic recurrent unit. The context vector weighted by attention and the mean value of the probabilistic hidden state vector at the last time are spliced to generate a spliced time sequence feature vector, so as to fuse the external context information and the internal historical state. The spliced time sequence feature vector is input into the LSTM to obtain the calculation results of the forget gate, the input gate, the output gate and the candidate cell state, which are used to obtain the gating signal required for state update. The spliced time sequence feature vector is linearly converted by the full connection layer and activated by the Softplus function to obtain the variance of the candidate cell state, ensuring that the variance value is always positive. The calculation result of the forget gate is multiplied by the mean value of the probabilistic hidden state vector at the last time to obtain the mean value forgetting element, the calculation result of the input gate is multiplied by the candidate cell state to obtain the mean value update element, and the mean value forgetting element is added to the mean value update element to obtain the cell state mean value at the current time, so as to complete the update of the state mean value. The variance of the cell state at the last time is multiplied by the square of the calculation result of the forget gate to obtain the variance forgetting element, the variance of the candidate cell state at the current time is multiplied by the square of the calculation result of the input gate to obtain the variance update element, and the variance forgetting element is added to the variance update element to obtain the cell state variance at the current time, so as to realize the gated propagation of the uncertainty. The calculation result of the output gate is multiplied by the tangent value of the cell state mean value at the current time to obtain the mean value of the probabilistic hidden state vector at the current time, which is used to generate the final hidden state mean value. The square of the tangent value of the cell state mean value at the current time is taken as the cell state mean value decay factor, 1 is subtracted from the cell state mean value decay factor to obtain the cell state mean value residual factor, the square of the calculation result of the output gate is multiplied by the cell state variance at the current time, and the square of the cell state mean value residual factor is multiplied by the cell state variance at the current time to obtain the variance of the probabilistic hidden state vector at the current time, so as to accurately calculate the hidden state variance after nonlinear transformation. The mean value and the variance of the probabilistic hidden state vector at the current time are spliced to obtain the probabilistic hidden state vector at the current time, so as to completely express the probabilistic state information at the current time.

[0082] In the embodiment, the S2 specifically comprises:

[0083] S21, setting the initial value of the mean value of the probabilistic hidden state vector and the initial value of the variance;

[0084] S22, for each time step in the charging state time sequence data set, based on the mean value of the probabilistic hidden state vector at the last time, using a multilayer perceptron as a feedforward neural network, calculating the energy score of all charging state time sequence vectors from the first time step to the current time step by forward propagation;

[0085] S23, normalize the energy scores by a Softmax function to obtain an attention weight vector;

[0086] S24, based on the attention weight, perform weighted summation on the charging state time series vector from the first time step to the current time step to generate an attention weighted context vector of the current time step;

[0087] S25, input the attention weighted context vector and the mean of the probabilistic hidden state vector of the previous time step of the encoder into the probabilistic recurrent unit to obtain the probabilistic hidden state vector of the current time step;

[0088] S26, repeat the steps of S22 to S25 in time step order until all time steps are processed, and concatenate the probabilistic hidden state vectors of each time step in time step order to obtain a probabilistic hidden state vector sequence of the encoder.

[0089] The embodiment introduces an input attention mechanism in the encoder, enabling the model to dynamically focus on key information in historical data, thereby generating a probabilistic hidden state sequence rich in context information. The initial values of the mean and variance of the probabilistic hidden state vector are set to provide a starting point for the iterative calculation of the encoder. For each time step in the charging state time series dataset, based on the mean of the probabilistic hidden state vector of the previous time step, a multi-layer perceptron is used as a feedforward neural network to calculate the energy scores of all charging state time series vectors from the first time step to the current time step through forward propagation, to evaluate the importance of each historical input to the current state. The energy scores are normalized by a Softmax function to obtain an attention weight vector, thereby converting the importance scores into a usable probability distribution. Based on the attention weight, the charging state time series vectors from the first time step to the current time step are weighted and summed to generate an attention weighted context vector of the current time step, thereby achieving adaptive aggregation of key historical information. The attention weighted context vector and the mean of the probabilistic hidden state vector of the previous time step of the encoder are input into the probabilistic recurrent unit to obtain the probabilistic hidden state vector of the current time step, to fuse the context information and update the internal state. The steps of S22 to S25 are repeated in time step order until all time steps are processed, and the probabilistic hidden state vectors of each time step are concatenated in time step order to obtain a probabilistic hidden state vector sequence of the encoder, thereby forming a probabilistic encoding representation of the entire input time series data.

[0090] In the embodiment, the S3 specifically comprises:

[0091] S31, set the initial values of the mean and variance of the probabilistic hidden state vector of the decoder;

[0092] S32, for each future time step that needs to be predicted, based on the entire sequence of probabilistic hidden state vectors of the encoder and the mean of the probabilistic hidden state vector of the decoder at the last time step, using a multi-layer perceptron as a feedforward neural network, the energy score of each time step in the sequence is calculated by forward propagation;

[0093] S33, the energy score is normalized by the Softmax function to obtain a time attention weight vector;

[0094] S34, based on the time attention weight vector, the mean of the probabilistic hidden state vector sequence is weighted and summed to generate a time attention weighted context vector;

[0095] S35, the time attention weighted context vector and the mean of the probabilistic hidden state vector of the decoder at the last time step are input into the probabilistic recurrent unit to obtain the prediction parameter of the current prediction time step;

[0096] S36, for each future time step that needs to be predicted, the steps of S32 to S35 are repeatedly performed until the prediction of all future time steps is completed, and the prediction parameter sequence of the future time steps of the decoder is obtained.

[0097] The embodiment introduces a time attention mechanism in the decoder, so that the model can review and focus on the key historical time of the encoder output, thereby realizing accurate probabilistic prediction of future time sequence parameters. The initial values of the mean and variance of the probabilistic hidden state vector of the decoder are set to provide a starting point for the autoregressive prediction process of the decoder. For each future time step that needs to be predicted, based on the entire sequence of probabilistic hidden state vectors of the encoder and the mean of the probabilistic hidden state vector of the decoder at the last time step, a multi-layer perceptron is used as a feedforward neural network, and the energy score of each time step in the sequence is calculated by forward propagation to evaluate the importance of each historical encoding state for future prediction. The energy score is normalized by the Softmax function to obtain a time attention weight vector, thereby converting the importance score into a usable probability distribution. Based on the time attention weight vector, the mean of the probabilistic hidden state vector sequence is weighted and summed to generate a time attention weighted context vector to adaptively extract the most relevant historical information for the current prediction. The time attention weighted context vector and the mean of the probabilistic hidden state vector of the decoder at the last time step are input into the probabilistic recurrent unit to obtain the prediction parameter of the current prediction time step, which is used to fuse the historical context and the current prediction state to generate a probabilistic output. For each future time step that needs to be predicted, the steps of S32 to S35 are repeatedly performed until the prediction of all future time steps is completed, and the prediction parameter sequence of the future time steps of the decoder is obtained, thereby forming a complete probabilistic prediction of the future.

[0098] In this embodiment, S4 specifically includes:

[0099] S41, extracting the prediction parameter of each future time step from the prediction parameter sequence of the future time step of the decoder;

[0100] S42, based on each future time step, using the mean and variance corresponding to the prediction parameter to construct a Gaussian distribution;

[0101] S43, based on the Gaussian distribution, extracting the expected value and variance value of each future time step;

[0102] S44, concatenating the variance value corresponding to the expected value of each time step to obtain the feature of the current time step;

[0103] S45, sequentially concatenating the features of all time steps in time order to obtain a comprehensive time sequence feature vector.

[0104] The embodiment converts the probabilistic prediction parameters output by the model into a structured feature vector, providing a quantitative basis for subsequent risk assessment with both prediction value and uncertainty information. From the prediction parameter sequence of the future time step of the decoder, the prediction parameter of each future time step is extracted to obtain a complete probabilistic description of each future time by the model. Based on each future time step, the mean and variance corresponding to the prediction parameter are used to construct a Gaussian distribution, thereby formalizing the prediction parameter into an explicit probability model. Based on the Gaussian distribution, the expected value and variance value of each future time step are extracted to obtain the central tendency and uncertainty measure of the prediction. Concatenate the variance value corresponding to the expected value of each time step to obtain the feature of the current time step, thereby integrating the prediction information of each time into a unified feature representation. Concatenate the features of all time steps in time order to obtain a comprehensive time sequence feature vector, and finally form an input vector containing complete prediction trend and confidence information for the decision model.

[0105] In this embodiment, S5 specifically includes:

[0106] S51, based on the comprehensive time sequence feature vector, obtaining the expected value and variance value of each time step, and obtaining the standard deviation of each time step by taking the square root of the variance value;

[0107] S52, set the confidence level coefficient, multiply the standard deviation by the confidence factor, and add the expected value to obtain the upper limit of confidence of the current time step;

[0108] S53, statistics of all time steps of the confidence upper limit to obtain the confidence upper limit list, statistics of all time steps of the expected value to obtain the expected value list, statistics of all time steps of the variance value to obtain the variance value list;

[0109] S54, set the expected threshold and the safety threshold according to historical experience and expert advice, if there is an element in the expected value list that exceeds the expected threshold, record it as an expected risk element; if there is an element in the confidence upper limit list that exceeds the safety threshold, record it as a confidence risk element;

[0110] S55, set the expected safety threshold, the expected danger threshold, the confidence safety threshold, and the confidence danger threshold; if the value of the expected risk element is higher than the expected danger threshold, or the value of the confidence risk element is higher than the confidence danger threshold, confirm that the risk level is high risk; if the value of the expected risk element is lower than the expected danger threshold and higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold and higher than the confidence safety threshold, confirm that it is medium risk; otherwise, confirm that it is low risk;

[0111] S56, according to the safety risk level, formulate the corresponding risk processing strategy, including reducing the charging power, starting the low-speed cooling fan, suspending the charging, and starting the full-speed cooling system.

[0112] The embodiment realizes accurate quantitative evaluation and intelligent decision of the overheat risk by constructing a two-dimensional risk classification model combining expected value and confidence upper limit. Based on the comprehensive time sequence feature vector, the expected value and the variance value of each time step are obtained, and the standard deviation of each time step is obtained by taking the square root of the variance value, which is used to quantify the uncertainty degree of the prediction. Set the confidence level coefficient, multiply the standard deviation by the confidence factor, and add the expected value to obtain the confidence upper limit of the current time step, thereby constructing a conservative risk boundary considering uncertainty. The confidence upper limit list is obtained by counting all time steps, the expected value list is obtained by counting all time steps, and the variance value list is obtained by counting all time steps, which provides structured data for subsequent risk screening. According to historical experience and expert advice, set the expected threshold and the safety threshold, if there is an element in the expected value list that exceeds the expected threshold, record it as an expected risk element, if there is an element in the confidence upper limit list that exceeds the safety threshold, record it as a confidence risk element, to preliminarily identify the time points with potential risks. Set the expected safety threshold, the expected danger threshold, the confidence safety threshold, and the confidence danger threshold, if the value of the expected risk element is higher than the expected danger threshold, or the value of the confidence risk element is higher than the confidence danger threshold, confirm that the risk level is high risk, if the value of the expected risk element is lower than the expected danger threshold and higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold and higher than the confidence safety threshold, confirm that it is medium risk, otherwise, confirm that it is low risk, thereby realizing fine division of the risk level. According to the safety risk level, formulate the corresponding risk processing strategy, including reducing the charging power, starting the low-speed cooling fan, suspending the charging, and starting the full-speed cooling system, and finally form a closed-loop protection from risk perception to active intervention.

[0113] In this embodiment, S6 specifically includes:

[0114] S61, deploying the risk handling strategy to the charging control system and recording the risk event log, including event ID, occurrence time, expected risk element and confidence risk element, corresponding risk handling strategy, and charging state raw data;

[0115] S62, after deploying the risk handling strategy, collecting charging state raw data in real time, comparing with historical normal data, and counting risk removal time, charging time loss value, energy loss value, and parameter change value of each item of charging state raw data;

[0116] S63, generating a risk handling strategy implementation effect report based on risk removal time, charging time loss value, energy loss value, and parameter change value of each item of charging state raw data.

[0117] The embodiment forms a closed-loop optimization system from execution to feedback by deploying the strategy and quantitatively evaluating its effect, to continuously improve the effectiveness and economy of the protection strategy. The risk handling strategy is deployed to the charging control system, and the risk event log is recorded, including event ID, occurrence time, expected risk element and confidence risk element, corresponding risk handling strategy, and charging state raw data, to realize the automatic execution of the strategy and provide complete data traceability for subsequent analysis. After deploying the risk handling strategy, charging state raw data is collected in real time, compared with historical normal data, and risk removal time, charging time loss value, energy loss value, and parameter change value of each item of charging state raw data are counted, to quantitatively evaluate the actual influence and cost after the implementation of the strategy. Based on risk removal time, charging time loss value, energy loss value, and parameter change value of each item of charging state raw data, a risk handling strategy implementation effect report is generated, to provide objective data support for the adjustment and optimization of the strategy.

[0118] Example 1:

[0119] In order to verify the effectiveness of the application in the safety protection of portable chargers, the method of the application is applied to the intelligent temperature control and protection system of a new generation of fast charging charger (hereinafter referred to as "charger H") developed by a well-known electronic technology company. In the traditional protection scheme of portable chargers, a fixed threshold judgment method based on a single temperature sensor is usually used. This method has a lagging response and cannot predict the temperature change trend, and is prone to false triggering or non-triggering due to environmental interference or load mutation, which poses a serious safety hazard. To solve the above problems, charger H decides to use the overheat protection method for portable chargers based on multiple sources of sensing proposed by the application.

[0120] In the implementation process, the charger H first uses the built-in battery core temperature sensor, mainboard temperature sensor, interface temperature sensor, environmental temperature sensor and voltage current monitoring chip to collect multi-source state data in real time during the charging process. After median filtering, missing value interpolation, outlier detection and data normalization processing, a high-quality structured charging state data sequence is formed. At the same time, the system divides the data through sliding window technology, and extracts statistical features such as mean, variance and peak value, and constructs a time series feature set that can reflect the charging dynamic process.

[0121] The charger H performs deep feature mining on multi-source time series data through the improved DA-RNN model. The encoder part uses input attention mechanism to automatically learn and focus on the historical data segments and sensor channels that have the greatest impact on the current state, effectively filtering irrelevant noise. The decoder part uses time attention mechanism to review and reference the key historical moments of the encoder output when predicting future temperature, achieving accurate capture of complex thermal dynamic changes. In particular, the probabilistic recurrent unit designed in the present application extends the hidden state from a deterministic vector to a probability distribution, so that the model can not only give the expected value when predicting future temperature, but also provide a confidence interval, thereby quantifying the uncertainty of the prediction.

[0122] The research and development team of the charger H found in testing that compared with traditional fixed threshold methods and ordinary machine learning methods, the method of the present application significantly improves the accuracy and timeliness of the overheat risk warning. The traditional method cannot predict the temperature trend, while the method of the present application can identify potential overheat risks several minutes in advance through probabilistic time series prediction and give risk levels, realizing the change from passive response to active prevention.

[0123] To further verify the actual performance of the method of the present application, the charger H compared the method of the present application with the traditional method in detail, and the specific performance data is shown in Table 1:

[0124] Table 1 Performance comparison table of overheat protection method of charger H

[0125] Indicator Conventional method Method of the present application Lifting amplitude Overheating risk early warning accuracy (%) 75.2 96.8 +21.6% Overheating risk false alarm rate (%) 12.5 2.1 -10.4% Overheating risk false negative rate (%) 12.3% 1.1% -11.2% Average early warning lead time (seconds) 5 45 +800% Single prediction processing time (milliseconds) 15 8 -46.7% High temperature protection trigger times (times / 100 charges) 8 2 -75.0% Charging interruption rate due to overheating (%) 3.5 0.4 -88.6% Average charging completion rate (%) 92.0 99.2 +7.8% Average trouble-free life of charger (hours) 1500 2500 +66.7% User safety satisfaction (%) 85.0 98.5 +13.5%

[0126] As can be seen from Table 1, after applying the method of the present application, the intelligent protection system performance of the charger H is comprehensively improved. The overheat risk early warning accuracy is improved from 75.2% of the traditional method to 96.8%, the false positive rate and the false negative rate are reduced by 10.4% and 11.2% respectively, and the reliability of the early warning is significantly improved. The average early warning time is greatly improved from 5 seconds of the traditional method to 45 seconds, which wins valuable time for the system to take protective measures. The single prediction processing time is reduced from 15 milliseconds to 8 milliseconds, which meets the real-time requirements of the embedded system. In addition, the number of high-temperature protection triggers and the charging interruption rate caused by overheating are greatly reduced, effectively improving the user experience. The average trouble-free life of the charger is also improved from 1500 hours to 2500 hours, and the product reliability is significantly enhanced. The user safety satisfaction is also greatly improved from 85.0% to 98.5%.

[0127] Through the method of the present application, the charger H successfully realizes the accurate prediction and active protection of the overheat risk, effectively avoids safety accidents caused by overheating, ensures the safety and stability of the charging process, greatly improves the intelligent level of the portable charger, significantly enhances the market competitiveness of the product, and provides strong technical support for the safety design of consumer electronic products.

[0128] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacements or changes within the technical range disclosed by the present application according to the technical solution and the inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A portable charger overheat protection method based on multi-source sensing, characterized in that, The method comprises the following steps: S1, collecting the charging state raw data of the portable charger in real time through a sensor, preprocessing the charging state raw data, and generating a charging state time series data set; S2, inputting the charging state time series data set into the encoder of the improved DA-RNN model, introducing a probabilistic modeling architecture, and extracting a probabilistic hidden state vector sequence of the encoder through a probabilistic recurrent unit; S3, inputting the probabilistic hidden state vector sequence of the encoder into the decoder of the improved DA-RNN model, generating a probabilistic context vector based on a time attention mechanism, and extracting a prediction parameter sequence of the future time step of the decoder through a probabilistic recurrent unit; S4, constructing a probability distribution of the future time step prediction parameter according to the prediction parameter sequence of the future time step of the decoder, and extracting a comprehensive time series feature vector; S5, performing safety risk level assessment according to the comprehensive time series feature vector, and formulating a corresponding risk treatment strategy according to the safety risk level; S6, deploying the risk treatment strategy to the charging management system, comparing historical data and real-time data, and generating a risk treatment strategy implementation effect report.

2. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The charging state raw data includes battery core temperature, battery mainboard temperature, output interface temperature, environmental temperature, input and output voltage, input and output current, charging power, power factor, charging time, battery current percentage; the preprocessing includes data cleaning, space-time alignment, standardization, data smoothing and denoising, sliding window segmentation, and feature extraction.

3. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The probabilistic recurrent unit specifically comprises: The context vector weighted by attention and the mean of the probabilistic hidden state vector of the previous time are spliced to generate a spliced time series feature vector; The spliced time series feature vector is input into an LSTM to obtain the calculation results of the forget gate, the input gate, the output gate, and the candidate cell state; The spliced time series feature vector is linearly converted by a fully connected layer and activated by a Softplus function to obtain the variance of the candidate cell state; The calculation result of the forget gate is multiplied by the mean of the probabilistic hidden state vector of the previous time to obtain a mean forgetting element; the calculation result of the input gate is multiplied by the candidate cell state to obtain a mean update element; and the mean forgetting element is added to the mean update element to obtain the mean of the cell state at the current time; The variance of the cell state of the previous time is multiplied by the square of the calculation result of the forget gate to obtain a variance forgetting element; the variance of the candidate cell state at the current time is multiplied by the square of the calculation result of the input gate to obtain a variance update element; and the variance forgetting element is added to the variance update element to obtain the variance of the cell state at the current time; The calculation result of the output gate is multiplied by the tangent value of the mean of the cell state at the current time to obtain the mean of the probabilistic hidden state vector at the current time; Square the tangent value of the mean value of the cell state at the current time as a cell state mean value decay factor, and obtain a cell state mean value residual factor by subtracting the cell state mean value decay factor from 1; square the output gate calculation result, multiply the cell state variance at the current time, multiply the square of the cell state mean value residual factor, and obtain the variance of the probabilistic hidden state vector at the current time; The mean value and the variance of the probabilistic hidden state vector at the current time are spliced to obtain the probabilistic hidden state vector at the current time.

4. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The S2 specifically comprises: S21, setting the initial value of the mean value and the initial value of the variance of the probabilistic hidden state vector; S22, for each time step in the charging state time series data set, based on the mean value of the probabilistic hidden state vector at the last time, using a multilayer perceptron as a feedforward neural network, calculating the energy score of all charging state time series vectors from the first time step to the current time step through forward propagation; S23, normalizing the energy score by the Softmax function to obtain an attention weight vector; S24, based on the attention weight, performing weighted summation on the charging state time series vectors from the first time step to the current time step to generate an attention weighted context vector at the current time step; S25, inputting the attention weighted context vector and the mean value of the probabilistic hidden state vector of the encoder at the last time step into the probabilistic recurrent unit to obtain the probabilistic hidden state vector at the current time; S26, repeating the steps of S22 to S25 in time step order until all time steps are processed, and splicing the probabilistic hidden state vectors of each time step in time step order to obtain the probabilistic hidden state vector sequence of the encoder.

5. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The S3 specifically comprises: S31, setting the initial value of the mean value and the initial value of the variance of the probabilistic hidden state vector of the decoder; S32, for each future time step to be predicted, based on the entire probabilistic hidden state vector sequence of the encoder and the mean value of the probabilistic hidden state vector of the decoder at the last time, using a multilayer perceptron as a feedforward neural network, calculating the energy score of each time step in the sequence through forward propagation; S33, normalizing the energy score by the Softmax function to obtain a time attention weight vector; S34, based on the time attention weight vector, performing weighted summation on the mean values in the probabilistic hidden state vector sequence to generate a time attention weighted context vector; S35, inputting the time attention weighted context vector and the mean value of the probabilistic hidden state vector of the decoder at the last time step into the probabilistic recurrent unit to obtain the prediction parameter of the current prediction time step; S36, for each future time step to be predicted, continuously repeating the steps of S32 to S35 until the prediction of all future time steps is completed, obtaining the prediction parameter sequence of the future time steps of the decoder.

6. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The S4 specifically comprises: S41, extracting the prediction parameter of each future time step from the prediction parameter sequence of the future time steps of the decoder; S42, based on each future time step, using the mean and variance corresponding to the prediction parameter, a Gaussian distribution is constructed; S43, based on the Gaussian distribution, the expected value and variance value of each future time step are extracted; S44, the variance value corresponding to the expected value of each time step is spliced to obtain the feature of the current time step; S45, the features of all time steps are spliced in time sequence to obtain a comprehensive time sequence feature vector.

7. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The S5 specifically includes: S51, based on the comprehensive time sequence feature vector, the expected value and variance value of each time step are obtained, and the standard deviation of each time step is obtained by taking the square root of the variance value; S52, set the confidence level coefficient, use the confidence factor to multiply the standard deviation, add the expected value to obtain the upper limit of the confidence of the current time step; S53, the upper limit of the confidence of all time steps is obtained to obtain a confidence upper limit list, the expected value of all time steps is counted to obtain an expected value list, and the variance value of all time steps is counted to obtain a variance value list; S54, according to historical experience and expert advice, set the expected threshold and the safety threshold, if there is an element in the expected value list that exceeds the expected threshold, it is recorded as an expected risk element; if there is an element in the confidence upper limit list that exceeds the safety threshold, it is recorded as a confidence risk element; S55, set the expected safety threshold, the expected danger threshold, the confidence safety threshold and the confidence danger threshold; if the value of the expected risk element is higher than the expected danger threshold, or the value of the confidence risk element is higher than the confidence danger threshold, the risk level is confirmed as high risk; if the value of the expected risk element is lower than the expected danger threshold and higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold and higher than the confidence safety threshold, it is confirmed as medium risk; otherwise, it is confirmed as low risk; S56, according to the safety risk level, corresponding risk processing strategies are developed, including reducing the charging power, starting the low-speed cooling fan, suspending the charging, and starting the full-speed cooling system.

8. The multi-source sensing based portable charger overheat protection method of claim 1, wherein, The S6 includes the following steps: S61, deploy the risk processing strategy to the charging control system, and record the risk event log, including event ID, occurrence time, expected risk element and confidence risk element, corresponding risk processing strategy, charging state raw data; S62, after the deployment of the risk processing strategy, the charging state raw data is collected in real time, compared with the historical normal data, and the risk removal time, the charging time loss value, the energy loss value, and the parameter change value of the charging state raw data are counted; S63, based on the risk removal time, the charging time loss value, the energy loss value, and the parameter change value of the charging state raw data, a risk processing strategy implementation effect report is generated.

Images