Multi-source sensing based portable charger overheat protection method

Abstract

The application discloses a portable charger overheat protection method based on multi-source sensing, and relates to the technical field of portable chargers, and comprises the following steps: S1, data acquisition and preprocessing; S2, introducing a probabilistic modeling architecture based on a DA-RNN model, extracting a probabilistic hidden state vector sequence of an encoder through a probabilistic recurrent unit; S3, extracting a predicted parameter sequence of a future time step of a decoder through the probabilistic recurrent unit; S4, constructing a probability distribution of the predicted parameter of the future time step and extracting a comprehensive time sequence feature vector; S5, performing safety risk level assessment and generating a risk treatment strategy; and S6, deploying the risk treatment strategy. The method overcomes the limitations of traditional charger protection methods, such as dependence on a single temperature threshold, response lag and lack of foresight, and provides an efficient and intelligent solution for the safe and reliable operation of portable charging equipment.

Description

Technical Field

[0001] This invention relates to the field of portable charger technology, and more particularly to an overheat protection method for portable chargers based on multi-source sensing. Background Technology

[0002] With the widespread use of portable charging devices, the requirements for their safety are also increasing. Traditional overheat protection methods for chargers mostly rely on simple temperature threshold judgments or periodic checks. While these methods can prevent overheating risks to some extent, they have significant limitations, especially in situations where chargers are used frequently and in complex environments. Existing overheat protection technologies often overlook 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 enhancing prediction accuracy and response speed, has become a core requirement for the development of smart chargers.

[0003] Currently, most portable charger overheat protection methods rely on fixed temperature thresholds or preset rules to determine whether overheating has occurred, which has significant shortcomings. Fixed threshold methods typically cannot cope with changes in charger operating environments, easily leading to false alarms or missed alarms. Furthermore, methods based on simple rules cannot dynamically adjust the charger's operating state, causing the device to fail to respond in time before reaching the temperature critical point during charging. Therefore, traditional overheat protection methods are ill-suited to the needs of intelligent, real-time management, lacking the ability to intelligently analyze and adaptively adjust device status data.

[0004] The main limitations of traditional methods lie in their lack of real-time performance and accuracy, especially when device status data exhibits high dimensionality, multi-source nature, and nonlinearity, where traditional methods demonstrate poor generalization ability. When multiple status data of the charger, such as temperature, current, power, and voltage, change simultaneously, simple single-threshold methods cannot efficiently and accurately identify overheating risks, making it difficult to detect and handle overheating problems in a timely manner, seriously affecting the safety and reliability of the equipment. Especially in complex environments and situations with large fluctuations in charger load, the prediction accuracy and response speed of traditional methods cannot meet the requirements of real-time overheat protection.

[0005] Furthermore, traditional overheat protection methods often overlook the correlations between multi-source sensor data, especially the interactions between multi-dimensional data such as temperature, current, power, and voltage. In the complex operating environment of chargers, traditional methods cannot fully utilize the temporal variations and multi-dimensional characteristics of this data, resulting in delayed response of protection mechanisms, high false alarm rates, and difficulty in providing accurate fault warnings in certain situations. Even with the adoption of machine learning or deep learning techniques, existing methods still fail to effectively uncover the deep correlations in sensor data, and cannot provide efficient, accurate, and real-time overheat protection solutions.

[0006] Therefore, how to provide overheat protection for portable chargers based on multi-source sensing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] This invention proposes a multi-source sensing-based overheat protection method for portable chargers, enabling more accurate prediction and assessment of thermal runaway risks during charging. This method not only automatically extracts deep temporal features from multi-source state data such as battery cell temperature, motherboard temperature, interface temperature, voltage, current, and power, but also combines a time attention mechanism with a probabilistic recurrent unit through an improved DA-RNN model to achieve probabilistic prediction and uncertainty quantification of future temperature trends. By introducing a confidence upper limit assessment and a two-dimensional risk grading strategy, this method fuses the predicted expected value with uncertainty information to dynamically assess the safety risk level and formulate differentiated protection strategies, such as power reduction, fan start / stop, or charging suspension, thereby achieving intelligent protection from passive response to proactive prevention. This method overcomes the limitations of traditional charger protection methods, which rely on a single temperature threshold, have lag in response, and lack predictability, providing an efficient and intelligent solution for the safe and reliable operation of portable charging devices.

[0008] The overheat protection method for portable chargers based on multi-source sensing according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect raw charging status data of the portable charger in real time through sensors, and preprocess the raw charging status data to generate a charging status time series dataset.

[0010] S2. Input the charging state time series dataset into the encoder of the improved DA-RNN model, introduce a probabilistic modeling architecture, and extract the probabilistic hidden state vector sequence of the encoder through probabilistic recurrent units.

[0011] S3. Input the sequence of probabilistic hidden state vectors of the encoder into the decoder of the improved DA-RNN model, generate probabilistic context vectors based on the time attention mechanism, and extract the sequence of predicted parameters for future time steps of the decoder through probabilistic recurrent units.

[0012] S4. Based on the predicted parameter sequence of the future time step of the decoder, construct the probability distribution of the predicted parameters of the future time step and extract the comprehensive temporal feature vector.

[0013] S5. Based on the comprehensive time-series feature vector, conduct a security risk level assessment and formulate corresponding risk handling strategies according to the security risk level.

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

[0015] Optionally, the raw charging status data includes battery cell temperature, battery motherboard temperature, output interface temperature, ambient temperature, input and output voltage, input and output current, charging power, power factor, charging time, and battery current percentage; the preprocessing includes data cleaning, spatiotemporal 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 from the previous time step are concatenated to generate a concatenated temporal feature vector.

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

[0019] The concatenated temporal feature vectors are processed by linear transformation of a fully connected layer and activation by the Softplus function to obtain the variance of the candidate cell states.

[0020] The mean forgotten element is obtained by multiplying the result of the forget gate by the mean of the probabilistic hidden state vector at the previous time step; the mean updated element is obtained by multiplying the result of the input gate by the candidate cell state; and the mean forgotten element is added to the mean updated element to obtain the mean cell state mean at the current time step.

[0021] The variance of the cell state at the previous time step 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 step is multiplied by the square of the calculation result of the input gate to obtain the variance update element; the variance forgetting element is added to the variance update element to obtain the variance of the cell state at the current time step.

[0022] The mean of the probabilistic hidden state vector at the current time is obtained by multiplying the result calculated by the output gate by the tangent of the mean of the cell state at the current time.

[0023] Calculate the square of the tangent of the cell state mean at the current time step, and use it as the cell state mean decay factor. Subtract the cell state mean decay factor from 1 to obtain the cell state mean residual factor. Use the square of the result calculated by the output gate, multiply it by the cell state variance at the current time step, and multiply it by the square of the cell state mean residual factor to obtain the variance of the probabilistic hidden state vector at the current time step.

[0024] The mean and variance of the probabilistic hidden state vector at the current time are concatenated to obtain the probabilistic hidden state vector at the current time.

[0025] Optionally, S2 specifically includes:

[0026] S21. Set the initial values ​​for the mean and variance of the probabilistic hidden state vector;

[0027] S22. 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, use a multilayer perceptron as a feedforward neural network to calculate the energy score of all charging state time series vectors from the first time step to the current time step through forward propagation.

[0028] S23. Normalize the energy score using the Softmax function to obtain the attention weight vector;

[0029] S24. Based on the attention weights, the charging state time sequence vector from the first time step to the current time step is weighted and summed to generate the attention-weighted context vector of the current time step.

[0030] S25. Input the mean of the attention-weighted context vector and the probabilistic hidden state vector of the encoder at the previous time step into the probabilistic recurrent unit to obtain the probabilistic hidden state vector at the current time step.

[0031] S26. Repeat steps S22 to S25 in the order of time steps until all time steps have been processed. Concatenate the probabilistic hidden state vectors of each time step in the order of time steps to obtain the sequence of probabilistic hidden state vectors of the encoder.

[0032] Optionally, S3 specifically includes:

[0033] S31. Set the initial values ​​for the mean and variance of the decoder's probabilistic hidden state vector;

[0034] S32. For each future time step that needs to be predicted, based on the mean of the entire probabilistic hidden state vector sequence of the encoder and the probabilistic hidden state vector of the decoder at the previous time step, a multilayer perceptron is used as a feedforward neural network to calculate the energy score of each time step in the sequence through forward propagation.

[0035] S33. Normalize the energy score using the Softmax function to obtain the temporal attention weight vector;

[0036] S34. Based on the temporal attention weight vector, the mean of the probabilistic hidden state vector sequence is weighted and summed to generate a temporal attention weighted context vector.

[0037] S35. Input the mean of the temporal attention-weighted context vector and the probabilistic hidden state vector of the decoder at the previous time step into the probabilistic recurrent unit to obtain the prediction parameters for the current prediction time step.

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

[0039] Optionally, S4 specifically includes:

[0040] S41. Extract the prediction parameters for each future time step from the prediction parameter sequence of the decoder.

[0041] S42. Based on each future time step, construct a Gaussian distribution using the mean and variance of the prediction parameters;

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

[0043] S44. Concatenate the corresponding variance value after the expected value at each time step to obtain the feature of the current time step;

[0044] S45. Concatenate the features of all time steps in chronological order to obtain the comprehensive temporal feature vector.

[0045] Optionally, S5 specifically includes:

[0046] S51. Based on the comprehensive time series 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.

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

[0048] S53. Calculate the confidence upper limit for all time steps to obtain a confidence upper limit list, calculate the expected value for all time steps to obtain an expected value list, and calculate the variance value for all time steps to obtain a variance value list.

[0049] S54. Based on historical experience and expert advice, set expected thresholds and safety thresholds. If there are elements in the expected value list that exceed the expected threshold, they are recorded as expected risk elements; if there are elements in the confidence limit list that exceed the safety threshold, they are recorded as confidence risk elements.

[0050] S55. Set the expected safety threshold, expected danger threshold, confidence safety threshold, and 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, then the risk level is confirmed as high risk; if the value of the expected risk element is lower than the expected danger threshold but higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold but higher than the confidence safety threshold, then it is confirmed as medium risk; otherwise, it is confirmed as low risk.

[0051] S56. Based on the safety risk level, formulate corresponding risk handling strategies, including reducing charging power, starting low-speed cooling fans, pausing charging, and starting the full-speed cooling system.

[0052] Optionally, S6 specifically includes:

[0053] S61. Deploy the risk handling strategy to the charging control system and record the risk event log, including the event ID, occurrence time, expected risk element and confidence risk element, corresponding risk handling strategy, and original charging status data.

[0054] S62. After the risk handling strategy is deployed, the original charging status data is collected in real time, compared with historical normal data, and the risk resolution time, charging time loss value, energy loss value, and various parameter changes of the original charging status data are statistically analyzed.

[0055] S63. Based on the risk resolution time, charging time loss value, energy loss value, and changes in various parameters of the original charging status data, generate a risk management strategy implementation effect report.

[0056] The beneficial effects of this invention are:

[0057] (1) This invention fundamentally improves the depth and reliability of overheat risk prediction for portable chargers by introducing a probabilistic modeling architecture and designing a probabilistic recurrent unit. Traditional recurrent neural networks, when processing time-series data, have deterministic hidden states that cannot reflect the inherent random noise and model uncertainty in system operation, leading to overly "confident" predictions and insufficient robustness in the face of abnormal operating conditions. To solve this problem, this invention innovatively designs a probabilistic recurrent unit, expanding the hidden state from a single vector into a probability distribution defined by mean and variance. This unit, through a dual-track parallel structure, updates the state mean using a gating mechanism while strictly adhering to the mathematical principle of uncertainty propagation, synchronously updating the state variance. This design enables the model to not only provide the most likely expected value when predicting future temperatures but also a confidence interval, thereby accurately quantifying the determinism of the prediction. This allows the protection system to effectively distinguish between "stable high temperatures" and "violently fluctuating risk temperatures," greatly enhancing the accuracy of early warnings and the reliability of decision-making.

[0058] (2) This invention effectively solves the technical challenge of feature fusion and key information capture of multi-source heterogeneous data during charging by employing an improved DA-RNN model. Traditional protection methods typically rely on a single temperature sensor for threshold judgment, which is insufficient to fully reflect the complex thermal dynamic changes caused by the combined effects of multiple factors such as voltage, current, and power. The improved DA-RNN model, through its encoder-decoder structure and the integration of input attention and time attention mechanisms, can adaptively focus on the most critical historical data segments and sensor channels that affect temperature changes, automatically mining the deep correlations and temporal dependencies between multi-source data. This method overcomes the limitations of traditional methods in information utilization, significantly improves the model's understanding of the evolution of charging states and its prediction accuracy, and provides a solid technical foundation for achieving intelligent active protection. Attached Figure Description

[0059] 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:

[0060] Figure 1 This is an overall flowchart of the portable charger overheat protection method based on multi-source sensing proposed in this invention;

[0061] Figure 2 This is a flowchart illustrating the working principle of the probabilistic recurrent unit of the improved DA-RNN model for overheat protection of portable chargers based on multi-source sensing proposed in this invention. Detailed Implementation

[0062] 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.

[0063] refer to Figure 1 and Figure 2 A method for overheat protection of portable chargers based on multi-source sensing includes the following steps:

[0064] S1. Collect raw charging status data of the portable charger in real time through sensors, and preprocess the raw charging status data to generate a charging status time series dataset.

[0065] S2. Input the charging state time series dataset into the encoder of the improved DA-RNN model, introduce a probabilistic modeling architecture, and extract the probabilistic hidden state vector sequence of the encoder through probabilistic recurrent units.

[0066] S3. Input the sequence of probabilistic hidden state vectors of the encoder into the decoder of the improved DA-RNN model, generate probabilistic context vectors based on the time attention mechanism, and extract the sequence of predicted parameters for future time steps of the decoder through probabilistic recurrent units.

[0067] S4. Based on the predicted parameter sequence of the future time step of the decoder, construct the probability distribution of the predicted parameters of the future time step and extract the comprehensive temporal feature vector.

[0068] S5. Based on the comprehensive time-series feature vector, conduct a security risk level assessment and formulate corresponding risk handling strategies according to the security risk level.

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

[0070] In this embodiment, the raw charging status data includes battery cell temperature, battery motherboard temperature, output interface temperature, ambient temperature, input and output voltage, input and output current, charging power, power factor, charging time, and battery current percentage; the preprocessing includes data cleaning, spatiotemporal alignment, standardization, data smoothing and denoising, sliding window segmentation, and feature extraction.

[0071] This implementation method, through comprehensive data acquisition and a refined preprocessing process, constructs a high-quality, multi-dimensional data foundation for subsequent accurate predictions. The raw charging status data includes battery cell temperature, battery motherboard temperature, output interface temperature, ambient temperature, input / output voltage, input / output current, charging power, power factor, charging time, and battery current percentage, ensuring the comprehensiveness and representativeness of the data source. The preprocessing includes data cleaning to remove outliers and missing values, ensuring data validity. Preprocessing includes spatiotemporal alignment to unify timestamps and spatial references from different sensors, ensuring data consistency. Preprocessing includes standardization to eliminate the influence of different physical dimensions and improve model convergence speed. Preprocessing includes data smoothing and denoising to filter out high-frequency interference and highlight the true trend of the data. Preprocessing includes sliding window segmentation to transform continuous data into time-series samples suitable for model input. Preprocessing includes feature extraction to mine deeper statistical and correlation features from the raw data, enhancing the model's information input.

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

[0073] The attention-weighted context vector and the mean of the probabilistic hidden state vector from the previous time step are concatenated to generate a concatenated temporal feature vector.

[0074] The concatenated temporal feature vector is input into the LSTM to obtain the calculation results of the forget gate, input gate, output gate and candidate cell state;

[0075] The concatenated temporal feature vectors are processed by linear transformation of a fully connected layer and activation by the Softplus function to obtain the variance of the candidate cell states.

[0076] The mean forgotten element is obtained by multiplying the result of the forget gate by the mean of the probabilistic hidden state vector at the previous time step; the mean updated element is obtained by multiplying the result of the input gate by the candidate cell state; and the mean forgotten element is added to the mean updated element to obtain the mean cell state mean at the current time step.

[0077] The variance of the cell state at the previous time step 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 step is multiplied by the square of the calculation result of the input gate to obtain the variance update element; the variance forgetting element is added to the variance update element to obtain the variance of the cell state at the current time step.

[0078] The mean of the probabilistic hidden state vector at the current time is obtained by multiplying the result calculated by the output gate by the tangent of the mean of the cell state at the current time.

[0079] Calculate the square of the tangent of the cell state mean at the current time step, and use it as the cell state mean decay factor. Subtract the cell state mean decay factor from 1 to obtain the cell state mean residual factor. Use the square of the result calculated by the output gate, multiply it by the cell state variance at the current time step, and multiply it by the square of the cell state mean residual factor to obtain the variance of the probabilistic hidden state vector at the current time step.

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

[0081] This implementation constructs a dual-track parallel probabilistic recurrent unit to achieve synchronous and accurate propagation and updating of the temporal feature mean and uncertainty. The attention-weighted context vector and the mean of the previous time step's probabilistic hidden state vector are concatenated to generate a concatenated temporal feature vector, thus fusing external context information with internal historical states. The concatenated temporal feature vector is input into an LSTM to obtain the calculation results of the forget gate, input gate, output gate, and candidate cell states, which are used to obtain the gating signals required for state updates. The concatenated temporal feature vector is then processed through a fully connected layer linear transformation and Softplus activation function to obtain the variance of the candidate cell states, ensuring that the variance value is always positive. The calculation result of the forget gate is multiplied by the mean of the previous time step's probabilistic hidden state vector to obtain the mean forgetting element. The calculation result of the input gate is multiplied by the candidate cell state to obtain the mean update element. The mean forgetting element is added to the mean update element to obtain the mean cell state mean at the current time step, thus completing the update of the state mean. The variance of the cell state at the previous time step is multiplied by the square of the result calculated by the forget gate to obtain the variance forgetting element. The variance of the candidate cell state at the current time step is multiplied by the square of the result calculated by the input gate to obtain the variance update element. The variance forgetting element is added to the variance update element to obtain the variance of the cell state at the current time step, thus achieving gating propagation of uncertainty. The result calculated by the output gate is multiplied by the tangent of the mean of the cell state at the current time step to obtain the mean of the probabilistic hidden state vector at the current time step, which is used to generate the final hidden state mean. The square of the tangent of the mean of the cell state at the current time step is calculated as the cell state mean decay factor. The cell state mean decay factor is obtained by subtracting the cell state mean decay factor from 1. The square of the result calculated by the output gate is multiplied by the variance of the cell state at the current time step, and then multiplied by the square of the cell state mean residual factor to obtain the variance of the probabilistic hidden state vector at the current time step. This is used to accurately calculate the hidden state variance after nonlinear transformation. The mean and variance of the probabilistic hidden state vector at the current time are concatenated to obtain the probabilistic hidden state vector at the current time, thus fully expressing the probabilistic state information at the current time.

[0082] In this embodiment, S2 specifically includes:

[0083] S21. Set the initial values ​​for the mean and variance of the probabilistic hidden state vector;

[0084] S22. 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, use a multilayer perceptron as a feedforward neural network to calculate the energy score of all charging state time series vectors from the first time step to the current time step through forward propagation.

[0085] S23. Normalize the energy score using the Softmax function to obtain the attention weight vector;

[0086] S24. Based on the attention weights, the charging state time sequence vector from the first time step to the current time step is weighted and summed to generate the attention-weighted context vector of the current time step.

[0087] S25. Input the mean of the attention-weighted context vector and the probabilistic hidden state vector of the encoder at the previous time step into the probabilistic recurrent unit to obtain the probabilistic hidden state vector at the current time step.

[0088] S26. Repeat steps S22 to S25 in the order of time steps until all time steps have been processed. Concatenate the probabilistic hidden state vectors of each time step in the order of time steps to obtain the sequence of probabilistic hidden state vectors of the encoder.

[0089] This implementation introduces an input attention mechanism into the encoder, enabling the model to dynamically focus on key information in historical data, thereby generating a probabilistic hidden state sequence rich in contextual information. Initial values ​​for the mean and variance of the probabilistic hidden state vectors are set to provide a starting point for the encoder's iterative computation. For each time step in the charging state time series dataset, based on the mean of the probabilistic hidden state vectors from the previous time step, a multilayer perceptron is used as a feedforward neural network to calculate the energy score of all charging state time series vectors from the first time step to the current time step through forward propagation, used to evaluate the importance of each historical input to the current state. The energy score is normalized using a Softmax function to obtain an attention weight vector, thus transforming the importance score into a usable probability distribution. The charging state time series vectors from the first time step to the current time step are weighted and summed based on the attention weights to generate the attention-weighted context vector for the current time step, thereby achieving adaptive aggregation of key historical information. The mean of the attention-weighted context vector and the probabilistic hidden state vector of the encoder from the previous time step is input into the probabilistic recurrent unit to obtain the probabilistic hidden state vector at the current time step. This vector is used to fuse context information and update the internal state. Steps S22 to S25 are repeated sequentially according to the time steps until all time steps have been processed. The probabilistic hidden state vectors of each time step are then concatenated sequentially to obtain the sequence of probabilistic hidden state vectors of the encoder, thus forming a probabilistic encoded representation of the entire input temporal data.

[0090] In this embodiment, S3 specifically includes:

[0091] S31. Set the initial values ​​for the mean and variance of the decoder's probabilistic hidden state vector;

[0092] S32. For each future time step that needs to be predicted, based on the mean of the entire probabilistic hidden state vector sequence of the encoder and the probabilistic hidden state vector of the decoder at the previous time step, a multilayer perceptron is used as a feedforward neural network to calculate the energy score of each time step in the sequence through forward propagation.

[0093] S33. Normalize the energy score using the Softmax function to obtain the temporal attention weight vector;

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

[0095] S35. Input the mean of the temporal attention-weighted context vector and the probabilistic hidden state vector of the decoder at the previous time step into the probabilistic recurrent unit to obtain the prediction parameters for the current prediction time step.

[0096] S36. For each future time step that needs to be predicted, repeat steps S32 to S35 until the prediction of all future time steps has been completed, and obtain the prediction parameter sequence of the future time steps of the decoder.

[0097] This implementation introduces a temporal attention mechanism into the decoder, enabling the model to review and focus on key historical moments in the encoder output, thereby achieving accurate probabilistic predictions of future time-series parameters. Initial values ​​for the mean and variance of the decoder's probabilistic hidden state vector are set, providing a starting point for the decoder's autoregressive prediction process. For each future time step to be predicted, based on the entire sequence of the encoder's probabilistic hidden state vectors and the mean of the decoder's probabilistic hidden state vectors from the previous time step, a multilayer perceptron is used as a feedforward neural network. Energy scores for each time step in the sequence are calculated through forward propagation to evaluate the importance of each historical encoded state to the future prediction. The energy scores are normalized using a Softmax function to obtain a temporal attention weight vector, thus transforming the importance scores into a usable probability distribution. Based on the temporal attention weight vector, the mean of the probabilistic hidden state vector sequence is weighted and summed to generate a temporally attention-weighted context vector, which adaptively extracts the historical information most relevant to the current prediction. The mean of the temporally attention-weighted context vector and the probabilistic hidden state vector of the decoder from the previous time step is input into the probabilistic recurrent unit to obtain the prediction parameters for the current prediction time step. These parameters are then used to fuse the historical context and the current prediction state to generate the probabilistic output. For each future time step that needs to be predicted, steps S32 to S35 are repeated until the predictions for all future time steps have been completed, resulting in the sequence of prediction parameters for the decoder's future time steps, thus forming a complete probabilistic prediction of the future.

[0098] In this embodiment, S4 specifically includes:

[0099] S41. Extract the prediction parameters for each future time step from the prediction parameter sequence of the decoder.

[0100] S42. Based on each future time step, construct a Gaussian distribution using the mean and variance of the prediction parameters;

[0101] S43. Based on the Gaussian distribution, extract the expected value and variance of each future time step;

[0102] S44. Concatenate the corresponding variance value after the expected value at each time step to obtain the feature of the current time step;

[0103] S45. Concatenate the features of all time steps in chronological order to obtain the comprehensive temporal feature vector.

[0104] This implementation transforms the probabilistic prediction parameters output by the model into a structured feature vector, providing a quantitative basis for subsequent risk assessment that combines predicted values ​​and uncertainty information. From the sequence of prediction parameters for each future time step in the decoder, prediction parameters for each future time step are extracted to obtain a complete probabilistic description of the model for each future moment. Based on each future time step, a Gaussian distribution is constructed using the mean and variance of the prediction parameters, thus formalizing the prediction parameters into a clear probabilistic model. Based on the Gaussian distribution, the expected value and variance of each future time step are extracted to obtain the central trend and uncertainty measure of the prediction. The corresponding variance value is concatenated after the expected value of each time step to obtain the feature of the current time step, thereby integrating the prediction information of each moment into a unified feature representation. The features of all time steps are concatenated sequentially to obtain a comprehensive temporal feature vector, ultimately forming an input vector containing complete future prediction trends and confidence information, which can be directly used by the decision-making model.

[0105] In this embodiment, S5 specifically includes:

[0106] S51. Based on the comprehensive time series 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.

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

[0108] S53. Calculate the confidence upper limit for all time steps to obtain a confidence upper limit list, calculate the expected value for all time steps to obtain an expected value list, and calculate the variance value for all time steps to obtain a variance value list.

[0109] S54. Based on historical experience and expert advice, set expected thresholds and safety thresholds. If there are elements in the expected value list that exceed the expected threshold, they are recorded as expected risk elements; if there are elements in the confidence limit list that exceed the safety threshold, they are recorded as confidence risk elements.

[0110] S55. Set the expected safety threshold, expected danger threshold, confidence safety threshold, and 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, then the risk level is confirmed as high risk; if the value of the expected risk element is lower than the expected danger threshold but higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold but higher than the confidence safety threshold, then it is confirmed as medium risk; otherwise, it is confirmed as low risk.

[0111] S56. Based on the safety risk level, formulate corresponding risk handling strategies, including reducing charging power, starting low-speed cooling fans, pausing charging, and starting the full-speed cooling system.

[0112] This implementation method constructs a two-dimensional risk grading model combining expected value and confidence upper limit, achieving accurate quantitative assessment and intelligent decision-making for overheating risk. Based on the comprehensive time series feature vector, the expected value and variance value of each time step are obtained. The standard deviation of each time step is obtained by taking the square root of the variance value, which is used to quantify the degree of uncertainty in the prediction. A confidence level coefficient is set, and the confidence factor is multiplied by the standard deviation and added to the expected value to obtain the confidence upper limit of the current time step, thus constructing a conservative risk boundary that considers uncertainty. A confidence upper limit list is obtained by statistically analyzing the confidence upper limits of all time steps, an expected value list is obtained by statistically analyzing the expected values ​​of all time steps, and a variance value list is obtained by statistically analyzing the variance values ​​of all time steps, providing structured data for subsequent risk screening. Based on historical experience and expert advice, expected thresholds and safety thresholds are set. If there are elements in the expected value list that exceed the expected threshold, they are recorded as expected risk elements; if there are elements in the confidence upper limit list that exceed the safety threshold, they are recorded as confidence risk elements, used to initially identify time points with potential risks. The system sets expected safety thresholds, expected danger thresholds, confidence safety thresholds, and confidence danger thresholds. If the value of an expected risk element is higher than the expected danger threshold, or the value of a confidence risk element is higher than the confidence danger threshold, the risk level is confirmed as high risk. If the value of an expected risk element is lower than the expected danger threshold but higher than the expected safety threshold, or the value of a confidence risk element is lower than the confidence danger threshold but higher than the confidence safety threshold, the risk level is confirmed as medium risk. Otherwise, the risk level is confirmed as low risk, thus achieving a refined classification of risk levels. Based on the safety risk level, corresponding risk handling strategies are formulated, including reducing charging power, starting low-speed cooling fans, suspending charging, and starting the full-speed cooling system, ultimately forming a closed-loop protection from risk perception to proactive intervention.

[0113] In this embodiment, S6 specifically includes:

[0114] S61. Deploy the risk handling strategy to the charging control system and record the risk event log, including the event ID, occurrence time, expected risk element and confidence risk element, corresponding risk handling strategy, and original charging status data.

[0115] S62. After the risk handling strategy is deployed, the original charging status data is collected in real time, compared with historical normal data, and the risk resolution time, charging time loss value, energy loss value, and various parameter changes of the original charging status data are statistically analyzed.

[0116] S63. Based on the risk resolution time, charging time loss value, energy loss value, and changes in various parameters of the original charging status data, generate a risk management strategy implementation effect report.

[0117] This implementation method, by deploying strategies and quantitatively evaluating their effects, forms a closed-loop optimization system from execution to feedback, continuously improving the effectiveness and cost-effectiveness of protection strategies. Risk handling strategies are deployed to the charging control system, and risk event logs are recorded, including event ID, occurrence time, expected risk elements and confidence risk elements, corresponding risk handling strategies, and raw charging status data. This enables automated strategy execution and provides complete data traceability for subsequent analysis. After the risk handling strategies are deployed, raw charging status data is collected in real time, compared with historical normal data, and statistically analyzed for risk resolution time, charging time loss, energy loss, and changes in various parameters of the raw charging status data. This quantitatively evaluates the actual impact and cost of strategy implementation. Based on the risk resolution time, charging time loss, energy loss, and changes in various parameters of the raw charging status data, a risk handling strategy implementation effect report is generated, providing objective data support for strategy adjustment and optimization.

[0118] Example 1:

[0119] To verify the effectiveness of this invention in the safety protection of portable chargers, the method of this invention was applied to the intelligent temperature control and protection system of a new generation fast charger (hereinafter referred to as "Charger H") developed by a well-known electronics technology company. Traditional portable charger protection solutions typically employ a fixed threshold judgment method based on a single temperature sensor. This method has a lag in response, cannot predict temperature change trends, and is prone to false triggering or failure to trigger due to environmental interference or sudden load changes, posing serious safety hazards. To solve the above problems, Charger H adopted the multi-source sensing-based portable charger overheat protection method proposed in this invention.

[0120] During implementation, charger H first utilizes built-in battery cell temperature sensors, motherboard temperature sensors, interface temperature sensors, ambient temperature sensors, and voltage and current monitoring chips to collect multi-source status data in real time during the charging process. After median filtering, missing value interpolation, outlier detection, and data normalization, a high-quality structured charging status data sequence is formed. Simultaneously, the system uses sliding window technology to segment the data and extract statistical features such as mean, variance, and peak value, constructing a time-series feature set that reflects the dynamic charging process.

[0121] Charger H employs an improved DA-RNN model for deep feature mining of multi-source time-series data. Its encoder utilizes an input attention mechanism to automatically learn and focus on historical data segments and sensor channels that have the greatest impact on the current state, effectively filtering out irrelevant noise. The decoder employs a temporal attention mechanism, which, when predicting future temperatures, can review and refer to key historical moments in the encoder output, achieving accurate capture of complex thermal dynamics. In particular, the probabilistic recurrent unit designed in this invention expands the hidden state from a deterministic vector to a probability distribution, enabling the model to provide not only an expected value but also a confidence interval when predicting future temperatures, thereby quantifying the uncertainty of the prediction.

[0122] The R&D team behind Charger H discovered during testing that, compared to traditional fixed threshold methods and ordinary machine learning methods, the method of this invention significantly improves the accuracy and timeliness of overheat risk warnings. Traditional methods cannot predict temperature trends, while the method of this invention, through probabilistic time-series prediction, can identify potential overheat risks minutes in advance and provide a risk level, achieving a shift from passive response to proactive prevention.

[0123] To further verify the actual performance of the method of the present invention, charger H 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:

[0124] Table 1. Performance Comparison of Overheat Protection Methods for Charger H

[0125] Overheat risk warning accuracy rate (%) 75.2 96.8 +21.6% False alarm rate for overheating risk (%) 12.5 2.1 -10.4% Overheating risk underreporting rate (%) 12.3% 1.1% -11.2% Average warning lead time (seconds) 5 45 +800% Time taken for a single prediction process (milliseconds) 15 8 -46.7% High temperature protection trigger count (times / 100 charge cycles) 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% Mean time between failures (MTBF) of charger (hours) 1500 2500 +66.7% User security satisfaction (%) 85.0 98.5 +13.5%

[0126] As shown in Table 1, the performance of the intelligent protection system of charger H was comprehensively improved after applying the method of this invention. The accuracy of overheat risk warning increased from 75.2% in the traditional method to 96.8%, while the false alarm rate and missed alarm rate decreased by 10.4% and 11.2% respectively, significantly improving the reliability of the warning. The average warning lead time increased dramatically from 5 seconds in the traditional method to 45 seconds, gaining valuable time for the system to take protective measures. The processing time for a single prediction decreased from 15 milliseconds to 8 milliseconds, meeting the real-time requirements of embedded systems. In addition, the number of high-temperature protection triggers and the charging interruption rate due to overheating decreased significantly, effectively improving the user experience. The charger's mean time between failures (MTBF) also increased from 1500 hours to 2500 hours, significantly enhancing product reliability. User safety satisfaction also increased significantly, from 85.0% to 98.5%.

[0127] Through the method of this invention, charger H successfully achieves accurate prediction and active protection against overheating risks, effectively avoiding safety accidents caused by overheating, ensuring the safety and stability of the charging process, greatly improving the intelligence level of portable chargers, significantly enhancing the market competitiveness of products, and providing strong technical support for the safety design of consumer electronics products.

[0128] 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 portable charger overheat protection method based on multi-source sensing, characterized in that, Includes the following steps: S1. Collect raw charging status data of the portable charger in real time through sensors, and preprocess the raw charging status data to generate a charging status time series dataset. S2. Input the charging state time series dataset into the encoder of the improved DA-RNN model, introduce a probabilistic modeling architecture, and extract the probabilistic hidden state vector sequence of the encoder through probabilistic recurrent units. The probabilistic loop unit specifically includes: The attention-weighted context vector and the mean of the probabilistic hidden state vector from the previous time step are concatenated to generate a concatenated temporal feature vector. The concatenated temporal feature vector is input into the LSTM to obtain the calculation results of the forget gate, input gate, output gate and candidate cell state; The concatenated temporal feature vectors are processed by linear transformation of a fully connected layer and activation by the Softplus function to obtain the variance of the candidate cell states. The mean forgotten element is obtained by multiplying the result of the forget gate by the mean of the probabilistic hidden state vector at the previous time step; the mean updated element is obtained by multiplying the result of the input gate by the candidate cell state; and the mean forgotten element is added to the mean updated element to obtain the mean cell state mean at the current time step. The variance of the cell state at the previous time step 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 step is multiplied by the square of the calculation result of the input gate to obtain the variance update element; the variance forgetting element is added to the variance update element to obtain the variance of the cell state at the current time step. The mean of the probabilistic hidden state vector at the current time is obtained by multiplying the result calculated by the output gate by the tangent of the mean of the cell state at the current time. Calculate the square of the tangent of the cell state mean at the current time step, and use it as the cell state mean decay factor. Subtract the cell state mean decay factor from 1 to obtain the cell state mean residual factor. Use the square of the result calculated by the output gate, multiply it by the cell state variance at the current time step, and multiply it by the square of the cell state mean residual factor to obtain the variance of the probabilistic hidden state vector at the current time step. The mean and variance of the probabilistic hidden state vector at the current time are concatenated to obtain the probabilistic hidden state vector at the current time. S3. Input the sequence of probabilistic hidden state vectors of the encoder into the decoder of the improved DA-RNN model, generate probabilistic context vectors based on the time attention mechanism, and extract the sequence of predicted parameters for future time steps of the decoder through probabilistic recurrent units. S4. Based on the predicted parameter sequence of the future time step of the decoder, construct the probability distribution of the predicted parameters of the future time step and extract the comprehensive temporal feature vector. S5. Based on the comprehensive time-series feature vector, conduct a security risk level assessment and formulate corresponding risk handling strategies according to the security risk level. S6. Deploy the risk management strategy into the charging management system, compare historical data and real-time data, and generate a risk management strategy implementation effect report.

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

3. The overheat protection method for portable chargers based on multi-source sensing according to claim 1, characterized in that, S2 specifically includes: S21. Set the initial values ​​for the mean and variance of the probabilistic hidden state vector; S22. 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, use a multilayer perceptron as a feedforward neural network to calculate the energy score of all charging state time series vectors from the first time step to the current time step through forward propagation. S23. Normalize the energy score using the Softmax function to obtain the attention weight vector; S24. Based on the attention weights, the charging state time sequence vector from the first time step to the current time step is weighted and summed to generate the attention-weighted context vector of the current time step. S25. Input the mean of the attention-weighted context vector and the probabilistic hidden state vector of the encoder at the previous time step into the probabilistic recurrent unit to obtain the probabilistic hidden state vector at the current time step. S26. Repeat steps S22 to S25 in the order of time steps until all time steps have been processed. Concatenate the probabilistic hidden state vectors of each time step in the order of time steps to obtain the sequence of probabilistic hidden state vectors of the encoder.

4. The method for overheat protection of a portable charger based on multi-source sensing according to claim 1, characterized in that, S3 specifically includes: S31. Set the initial values ​​for the mean and variance of the decoder's probabilistic hidden state vector; S32. For each future time step that needs to be predicted, based on the mean of the entire probabilistic hidden state vector sequence of the encoder and the probabilistic hidden state vector of the decoder at the previous time step, a multilayer perceptron is used as a feedforward neural network to calculate the energy score of each time step in the sequence through forward propagation. S33. Normalize the energy score using the Softmax function to obtain the temporal attention weight vector; S34. Based on the temporal attention weight vector, the mean of the probabilistic hidden state vector sequence is weighted and summed to generate a temporal attention weighted context vector. S35. Input the mean of the temporal attention-weighted context vector and the probabilistic hidden state vector of the decoder at the previous time step into the probabilistic recurrent unit to obtain the prediction parameters for the current prediction time step. S36. For each future time step that needs to be predicted, repeat steps S32 to S35 until the prediction of all future time steps has been completed, and obtain the prediction parameter sequence of the future time steps of the decoder.

5. The method for overheat protection of a portable charger based on multi-source sensing according to claim 1, characterized in that, S4 specifically includes: S41. Extract the prediction parameters for each future time step from the prediction parameter sequence of the decoder. S42. Based on each future time step, construct a Gaussian distribution using the mean and variance of the prediction parameters; S43. Based on the Gaussian distribution, extract the expected value and variance value for each future time step; S44. Concatenate the corresponding variance value after the expected value at each time step to obtain the feature of the current time step; S45. Concatenate the features of all time steps in chronological order to obtain the comprehensive temporal feature vector.

6. The method for overheat protection of a portable charger based on multi-source sensing according to claim 1, characterized in that, S5 specifically includes: S51. Based on the comprehensive time series 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, multiply the confidence factor by the standard deviation, and add the expected value to obtain the upper confidence limit for the current time step; S53. Calculate the confidence upper limit for all time steps to obtain a confidence upper limit list, calculate the expected value for all time steps to obtain an expected value list, and calculate the variance value for all time steps to obtain a variance value list. S54. Based on historical experience and expert advice, set expected thresholds and safety thresholds. If there are elements in the expected value list that exceed the expected threshold, they are recorded as expected risk elements; if there are elements in the confidence limit list that exceed the safety threshold, they are recorded as confidence risk elements. S55. Set the expected safety threshold, expected danger threshold, confidence safety threshold, and 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, then the risk level is confirmed as high risk; if the value of the expected risk element is lower than the expected danger threshold but higher than the expected safety threshold, or the value of the confidence risk element is lower than the confidence danger threshold but higher than the confidence safety threshold, then it is confirmed as medium risk; otherwise, it is confirmed as low risk. S56. Based on the safety risk level, formulate corresponding risk handling strategies, including reducing charging power, starting low-speed cooling fans, pausing charging, and starting the full-speed cooling system.

7. The method for overheat protection of a portable charger based on multi-source sensing according to claim 1, characterized in that, S6 includes the following steps: S61. Deploy the risk handling strategy to the charging control system and record the risk event log, including the event ID, occurrence time, expected risk element and confidence risk element, corresponding risk handling strategy, and original charging status data. S62. After the risk handling strategy is deployed, the original charging status data is collected in real time, compared with historical normal data, and the risk resolution time, charging time loss value, energy loss value, and various parameter changes of the original charging status data are statistically analyzed. S63. Based on the risk resolution time, charging time loss value, energy loss value, and changes in various parameters of the original charging status data, generate a risk management strategy implementation effect report.

Images