Vehicle-mounted power supply short circuit heat prediction method based on big data analysis

Abstract

The invention discloses a vehicle-mounted power supply short-circuit heat prediction method based on big data analysis, and the method comprises the following steps: collecting real-time operation data, and storing the data in a data cache, thereby obtaining a cache time sequence signal; filtering the cached time sequence signal, and extracting a basic feature sequence; constructing an RC network model based on the basic feature sequence, and converting a fractional order operator into an equivalent integer order model; performing parameter estimation on the equivalent integer order model to generate a thermal prediction model; taking the thermal prediction model and the basic feature sequence as joint input and output power loss features; inputting the power loss characteristics into a quantile regression forest model to form a heat prediction result; comparing the heat prediction result with a heat threshold value and a temperature rise threshold value, and outputting a control trigger signal; and calling parameters in a control strategy library according to the control trigger signal to generate a control strategy. According to the invention, vehicle-mounted power supply short-circuit heat prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of automotive fault prediction technology, and in particular to a method for predicting short-circuit heat in vehicle power supplies based on big data analysis. Background Technology

[0002] In the design of current automotive power systems, with the increasing number of electronic control units and the increasingly stringent requirements for the stability and safety of the vehicle's electrical architecture, accurately predicting the thermal response of power devices under abnormal conditions such as short circuits has become a key issue in thermal management and power protection design. In existing technologies, to simplify the thermal modeling process, linear integer-order RC thermal network models are typically used to estimate the thermal behavior of power devices. These models assume constant thermal capacity and thermal resistance, lack the ability to respond to the thermal memory effect within materials, and cannot accurately simulate the nonlinear thermal process under high-dynamic short-circuit conditions. Furthermore, model parameters often rely on offline identification and assume that devices operate under constant or slowly changing load conditions over long periods. This method is ineffective at identifying rapidly changing thermal abrupt changes.

[0003] In terms of building heat prediction models, most existing methods rely on training with batches of static data, lacking the ability to incrementally learn from continuously changing data in complex vehicle operating scenarios. After deployment, these methods often struggle to dynamically adjust model structure and parameters according to system status, easily leading to inaccurate predictions or delayed responses. Furthermore, for risk response mechanisms, current strategies often rely on setting fixed thresholds for simple judgments, failing to fully utilize the upper and lower bounds of the heat prediction distribution or incorporate dynamic judgments based on the heat rise rate trend. This easily results in false alarms or missed alarms, potentially delaying protection response in severe cases.

[0004] Therefore, how to provide a method for predicting short-circuit heat in vehicle power supplies based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for predicting short-circuit heat in vehicle power supplies based on big data analysis. This invention provides a method for predicting short-circuit heat in vehicle power supplies based on big data analysis, which integrates filtering processing, fractional-order modeling, online updating, and quantile regression forest to achieve accurate heat prediction and dynamic control strategy response, thereby improving the intelligence and reliability of short-circuit thermal management of power systems.

[0006] A method for predicting short-circuit heat in vehicle power supplies based on big data analysis, according to an embodiment of the present invention, includes the following steps: Collect real-time operating data of the vehicle power system and store it in the sliding data buffer to obtain the buffer timing signal; Preprocess the cached time-series signals to extract the denoised basic feature sequences; A fractional-order hot RC network model is constructed based on the denoised basic feature sequence, and the fractional-order operator is transformed into an equivalent integer-order model structure through the Oustaloup recursive approximation algorithm. The parameters of the equivalent integer-order model structure are estimated to obtain the vehicle power supply parameters, and a benchmark thermal prediction model is generated by combining the model structure. The baseline thermal prediction model and the basic feature sequence within the current sampling period are used as joint inputs. Valid samples are selected, and the model is updated incrementally online to output power loss features. The power loss characteristics are input into the quantile regression forest model to obtain the median, upper quantile and prediction interval of the current heat prediction, thus forming the heat prediction result. The heat prediction results are compared with the set heat threshold and temperature rise threshold to determine whether a heat risk flag is generated. When the heat risk flag is valid, a control trigger signal is output. Based on the control trigger signal and the current operating condition characteristics, the corresponding current limiting parameters and gating parameters are called from the control strategy library to generate a control strategy, which is then sent to the vehicle power control unit for execution.

[0007] Optionally, obtaining the buffer timing signal specifically includes: Set the sampling period, buffer length, and signal channel sequence for collecting operational data from the vehicle power system, start the data acquisition process, and receive the sampled data. Within each sampling period, power parameter data including input voltage, output voltage, output current, device temperature, ambient temperature, gate signal, and control signal are synchronously collected from the vehicle power system, and data records corresponding to the time are generated. Data records are written to the sliding data buffer in chronological order. The records are organized according to the sampling order. Control signals are divided by field and combined into a unified structure to obtain a buffered timing signal containing continuous multi-cycle data.

[0008] Optionally, the extraction of the denoised basic feature sequence specifically includes: Voltage, current, and temperature-related data are extracted from the buffered timing signal, including continuous sampled values ​​of input voltage and output voltage, continuous sampled values ​​of output current, and continuous sampled values ​​of device temperature and ambient temperature, which are then arranged in chronological order to form voltage sequence, current sequence, and temperature sequence, respectively. Empirical wavelet transform was performed on voltage, current and temperature sequences respectively, decomposing each type of original signal into several subsequences at different time scales. In each type of signal, the subsequences with the main feature changes were retained, and the high-frequency detail subsequences with strong fluctuations were removed. Hampel filtering is applied to the subsequences that retain the main feature changes. An outlier filtering method based on the sliding window method is used to identify and replace abnormal fluctuation values ​​in the signal, thereby suppressing noise. Finally, the processed subsequences are recombined to form the basic feature sequence.

[0009] Optionally, the transformation of the equivalent integer-order model structure specifically includes: The current sequence corresponding to the time axis in the basic feature sequence is used as the modeling input signal, and the device temperature sequence corresponding to the time axis is used as the modeling output signal. The length of the input and output signals is unified and the boundaries are aligned to form the equal-length input and output sample pairs required for thermal modeling, which are used to capture the causal response relationship between power consumption and temperature rise. Based on the input-output sample pairs, a fractional-order thermal RC network model of non-integer order is constructed. Multiple non-integer-order delay terms are introduced into the fractional-order thermal RC network model to describe the historical memory characteristics of heat in the conduction path, simulate the hysteresis, diffusion and material memory in the heat conduction process, and set the relationship between input and output to consist of three sub-terms, including a thermal stability constant term, a low-order fractional-order response term and a high-order fractional-order response term, so that the modeling structure has the ability to simultaneously express thermal inertia and nonlinear temperature rise trend. The low-order fractional-order response term is used to model the hysteresis effect of device temperature rise under short-cycle power excitation, and the high-order fractional-order response term is used to describe the slow temperature rise process caused by heat accumulation under long-cycle power excitation. Two fractional-order response terms are used as approximation targets. The upper and lower limits of the frequency range for fractional-order approximation are set. The lower and upper limits of the frequency range are the minimum operating frequency and the maximum thermal response frequency, respectively. At the same time, the approximation order is set to control the balance between approximation accuracy and computational complexity. Oustaloup recursive approximation expansion is used for each fractional-order term to construct a set of logarithmically uniform frequency points. The frequency domain characteristics of the fractional-order operator are discretized and fitted. The amplitude and phase frequency characteristics of the original fractional-order term are simulated using a series structure of multiple first-order filters. An equivalent integer-order model is generated within the set range. The equivalent integer-order model obtained by approximating the two fractional-order terms is substituted into the fractional-order thermal RC network model to replace the original fractional-order operators, while maintaining the consistency of the physical meaning and mathematical continuity of each term. The transformed fractional-order thermal RC network model is then associated with the input-output sample pair to form a complete equivalent integer-order model structure.

[0010] Optionally, the generation of the benchmark thermal prediction model specifically includes: The equivalent integer-order model structure is used as a fixed modeling framework. The synchronized current sequence is used as the thermal modeling input signal, and the device temperature sequence is used as the thermal modeling output signal. The input and output signals are divided into equal lengths and resampled according to a fixed sampling period to generate multiple sample windows with fixed time spans. For each output sample window, the Prony identification algorithm is used to identify parameters, construct a linear combination expression composed of the current temperature and the historical temperature values, set the number of linear combination coefficients as the model order, establish an error function for the linear combination, minimize the sum of squared temperature prediction errors for all times within the window, construct a system of linear matrix equations with a time series structure, and perform least squares estimation to obtain a parameter vector representing the dynamic characteristics of the thermal model. Each element of the parameter vector corresponds to the feedback weight of the integer order system in the equivalent thermal modeling structure. Based on the parameter vector, a characteristic equation for discrete-time variables is constructed. The root value decomposition method is used to obtain all complex solutions of the characteristic equation. Each complex solution is interpreted as a characteristic response pole. By taking the natural logarithm of these poles and dividing by the sampling period, the modal frequencies of the equivalent integer-order model structure in continuous time are obtained. Furthermore, the exponential function structure is recombined based on the modal frequencies to reconstruct a set of thermal response expressions composed of modal amplitudes and modal frequencies. These expressions are then substituted back into the equivalent integer-order model structure to form a benchmark thermal prediction model with dynamic identification capabilities.

[0011] Optionally, the output process of the power loss characteristics specifically includes: The basic feature sequence extracted within the current sampling period is organized into a fixed-dimensional input vector. The temperature rise output of the benchmark thermal prediction model in the current sampling period is used as the target variable to form a joint sample composed of input features and prediction target, which is then fed into the kernel recursive least squares model. The kernel recursive least squares model is created based on the kernel recursive least squares algorithm. In the kernel recursive least squares model, the similarity between the current sample and all historical support samples is calculated based on the Euclidean distance between the input vectors using a set nonlinear kernel mapping structure. The similarity is then converted into a mapping result in the kernel space through an exponential weighting method, and the kernel similarity matrix for the current sampling period is constructed accordingly. Based on the minimum similarity value between the current sample and the supporting samples in the kernel mapping space, the novelty of the sample is evaluated. The novelty of the sample is compared with the novelty judgment threshold preset in the system. When the novelty is greater than or equal to the threshold, it is considered a valid sample and included in the update sequence, and the internal kernel state matrix and weight coefficient vector are updated. If the novelty is lower than the threshold, the sample is ignored to control the model complexity and improve the real-time processing capability, and the output power loss feature is obtained.

[0012] Optionally, the formation of the heat prediction result specifically includes: The output power loss feature is concatenated with the basic feature vector extracted in the current sampling period. An abnormal feature group that deviates significantly from the training distribution is removed by a dynamic sample screening mechanism based on Mahalanobis distance. The minimum redundancy and maximum relevance selection method is applied to the remaining samples to obtain the target sub-feature set. The target sub-feature set is used to construct the prediction sample and input into the quantile regression forest model. The quantile regression forest model contains several regression trees. Each tree assigns input samples to corresponding leaf nodes through structural partitioning. In the leaf node where the predicted sample falls, all historical response values ​​within the current node are collected to form a response sample set. Based on the hierarchical quantile dynamic sampling mechanism, the heat response values ​​in the response sample set are resampled according to the target quantile to construct a pseudo-distribution. The quantile loss function is used to obtain the local optimal solution estimate. The local optimum estimates in all regression trees are density-sensitively fused according to their corresponding weights. The median, upper quantile and lower quantile predictions are weighted averages to construct the current central hot prediction and upper and lower confidence boundaries. At the same time, the inter-tree distribution offset index is introduced to statistically analyze the sparsity of the prediction results in the forest structure and dynamically correct the weight distribution. The median, upper quantile, and prediction interval of the heat forecast obtained at the current moment are used to form the heat forecast result. At the same time, the difference between the upper and lower boundaries of the heat forecast interval is calculated to represent the degree of uncertainty of the forecast. This value is used in the heat risk comparison and dynamic threshold determination module to support the real-time performance and accuracy of the heat risk management logic.

[0013] Optionally, the output process of the control trigger signal specifically includes: The median, upper quantile, and prediction interval of the current heat prediction are obtained from the heat prediction results to form the current heat prediction triplet. The static heat threshold and temperature rise threshold of the corresponding operating mode are called from the control system as comparison standards, and the allowable error tolerance is set as a dynamic safety boundary. The median predicted value of the heat prediction triplet is extracted as the heat estimate value, and the difference is compared with the heat threshold. The difference is compared with the preset tolerance range. When the comparison value exceeds the upper limit of the tolerance range, the current heat status is marked as a risk status to be confirmed, allowing the next step of temperature rise risk judgment logic to proceed. At the same time, the current heat offset value is recorded for trend modeling and dynamic adjustment. Based on the heat state label, the change range between the current heat estimate and the median heat value of the previous sampling period is calculated as the heat prediction increment, and the difference between the upper and lower boundaries of the current heat prediction interval is obtained as the uncertainty range. The ratio of the current heat prediction increment and the uncertainty range is calculated to evaluate the relationship between the current heat change rate and the model fluctuation. When the ratio exceeds the preset temperature rise rate threshold standard, it is determined that the current heat change has an abnormal rate and an effective heat risk indicator signal is generated. When the thermal risk indicator signal is valid, the control logic module immediately outputs a control trigger signal and pushes the signal to the vehicle power management unit to activate the corresponding emergency control strategy path. The risk trigger event number, timestamp, trigger reason and related thermal prediction data are recorded in the system event log and status buffer pool.

[0014] Optionally, the generation of the control strategy specifically includes: Receive control trigger signals and synchronously read the real-time operating characteristics of the current vehicle power system, including output power, voltage level, temperature status and control mode identifier. Input the control trigger signals and operating characteristics together as index conditions into the control strategy scheduling module to generate strategy retrieval instructions. According to the strategy retrieval instruction, the control strategy library is accessed, and the current limiting parameter template and gating parameter configuration file that match the current operating condition are retrieved from it. The current limiting parameters include the maximum allowable current value, the slope drop coefficient and the current limiting start delay time. The gating parameters include the drive voltage duty cycle, the turn-off holding time and the soft turn-off gradient. They are integrated into the original parameter set of the control strategy. The control strategy library quickly matches and retrieves the appropriate control strategy according to the current system state and the thermal risk trigger signal, so as to achieve effective intervention and system protection against short-circuit thermal risks. Based on the original parameter set and the current system state, adaptive adjustments are made, the current limit value is dynamically scaled, the gating control logic pins are prioritized, encoded into control frames and encapsulated in signal format to form control command frames. The control command frame is sent to the vehicle power control unit via the communication interface, triggering the execution of current limiting, shutdown or delay action mechanisms, and the system records the policy issuance event and execution feedback results to complete the closed-loop process of thermal risk response.

[0015] The beneficial effects of this invention are: 1. By replacing the traditional integer-order model with a fractional-order thermal RC network model, the thermal memory characteristics and nonlinear heat transfer behavior inside the power device material can be more realistically reflected, effectively reducing the prediction error caused by the oversimplification of the model. The fractional-order model is transformed into an equivalent integer-order form through the Oustaloup approximation algorithm, which not only preserves the modeling accuracy but also takes into account the computational efficiency, making it suitable for resource-constrained vehicle computing environments.

[0016] 2. By introducing the Prony identification algorithm, the model parameters are estimated efficiently and in a refined manner, making the thermal prediction model more physically consistent and structurally stable. In terms of parameter updates, the kernel recursive least squares algorithm is combined with the introduction of the novelty criterion to support the model's online incremental learning capability under continuously changing operating conditions. This solves the problem that traditional batch modeling algorithms cannot adapt to operating condition drift and ensures the long-term effectiveness and real-time performance of the prediction model.

[0017] 3. In the prediction output stage, a heat distribution modeling method based on quantile regression forest was constructed, which not only provides the median value of heat but also outputs the upper and lower confidence intervals, enhancing the ability to express thermal uncertainty and extreme risks. Combined with dynamic thermal risk judgment logic and control triggering mechanism, it realizes early perception and response to thermal over-limit trends, effectively avoiding the risk of power devices being damaged due to hysteresis response. Attached Figure Description

[0018] 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:

[0019] Figure 1 This is a flowchart of a method for predicting short-circuit heat in vehicle power supplies based on big data analysis, as proposed in this invention. Figure 2 This is a schematic diagram of the algorithm results for a vehicle power supply short-circuit heat prediction method based on big data analysis proposed in this invention. Detailed Implementation

[0020] The present 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.

[0021] refer to Figure 1-2 A method for predicting short-circuit heat in vehicle power supplies based on big data analysis includes the following steps: Collect real-time operating data of the vehicle power system and store it in the sliding data buffer to obtain the buffer timing signal; Preprocess the cached time-series signals to extract the denoised basic feature sequences; A fractional-order hot RC network model is constructed based on the denoised basic feature sequence, and the fractional-order operator is transformed into an equivalent integer-order model structure through the Oustaloup recursive approximation algorithm. The parameters of the equivalent integer-order model structure are estimated to obtain the vehicle power supply parameters, and a benchmark thermal prediction model is generated by combining the equivalent integer-order model structure. The baseline thermal prediction model and the basic feature sequence within the current sampling period are used as joint inputs. Valid samples are selected, and the model is updated incrementally online to output power loss features. The power loss characteristics are input into the quantile regression forest model to obtain the median, upper quantile and prediction interval of the current heat prediction, thus forming the heat prediction result. The heat prediction results are compared with the set heat threshold and temperature rise threshold to determine whether a heat risk flag is generated. When the heat risk flag is valid, a control trigger signal is output. Based on the control trigger signal and the current operating condition characteristics, the corresponding current limiting parameters and gating parameters are called from the control strategy library to generate a control strategy, which is then sent to the vehicle power control unit for execution.

[0022] This invention is based on multi-channel real-time operating data, utilizing empirical wavelet transform and Hampel filtering for feature denoising, combining fractional-order thermal RC modeling and Oustaloup approximation to construct a highly fitted thermal model structure, and employing the Prony algorithm for parameter identification. Quantile regression forests are used to predict the heat distribution range, define confidence intervals, and determine the thermal risk level accordingly. Upon detecting abnormal trends, control signals are immediately generated, and a control strategy library is invoked for current limiting and gating responses to ensure safe device operation.

[0023] In this embodiment, obtaining the buffer timing signal specifically includes: Set the sampling period, buffer length, and signal channel sequence for collecting operational data from the vehicle power system, start the data acquisition process, and receive the sampled data. Within each sampling period, power parameter data including input voltage, output voltage, output current, device temperature, ambient temperature, gate signal, and control signal are synchronously collected from the vehicle power system, and data records corresponding to the time are generated. Data records are written to the sliding data buffer in chronological order. The recorded content is organized according to the sampling order. Voltage, current, temperature and control signals are divided and combined into a unified structure by field to obtain a buffered timing signal containing continuous multi-cycle data.

[0024] This step establishes a systematic data acquisition and caching mechanism by setting the sampling period, buffer length, and channel order, enabling high-frequency synchronous acquisition of multi-dimensional electrical parameter data of the vehicle power system during operation. The acquired data covers key thermally relevant characteristic parameters such as input / output voltage, current, device and ambient temperature, gate and control signals, and is used to generate structured data records indexed by time. The acquired data is written to a sliding buffer in real time, ensuring signal continuity and timing consistency. Simultaneously, field partitioning and structural integration are performed according to signal type, forming a continuous, multi-period time-series data sequence. This process provides a high-quality, time-complete data foundation for subsequent signal processing and thermal modeling, enhancing the overall data support capability of the prediction system.

[0025] In this embodiment, the extraction of the denoised basic feature sequence specifically includes: Voltage, current, and temperature-related data are extracted from the buffered timing signal, including continuous sampled values ​​of input voltage and output voltage, continuous sampled values ​​of output current, and continuous sampled values ​​of device temperature and ambient temperature, which are then arranged in chronological order to form voltage sequence, current sequence, and temperature sequence, respectively. Empirical wavelet transform was performed on voltage, current and temperature sequences respectively, decomposing each type of original signal into several subsequences at different time scales. In each type of signal, the subsequences with the main feature changes were retained, and the high-frequency detail subsequences with strong fluctuations were removed. Hampel filtering is applied to the subsequences that retain the main feature changes. An outlier filtering method based on the sliding window method is used to identify and replace abnormal fluctuation values ​​in the signal, thereby suppressing noise. Finally, the processed subsequences are recombined to form the basic feature sequence.

[0026] This step involves refined preprocessing of voltage, current, and temperature data in the buffered time-series signals, sequentially constructing continuous feature sequences. The original signals are decomposed into multiple scales using empirical wavelet transform, retaining subsequences reflecting the main trends while filtering out high-frequency noise and interference. Subsequently, the Hampel filtering algorithm is used to detect and remove local outliers within a sliding window, effectively correcting abrupt signal changes. After this dual processing, noise interference in each signal sequence is significantly suppressed, preserving the main dynamic features with physical significance and thermal behavior correlation, forming stable and reliable basic feature sequences, providing high-quality data support for subsequent modeling.

[0027] In this embodiment, the transformation of the equivalent integer-order model structure specifically includes: The current sequence corresponding to the time axis in the basic feature sequence is used as the modeling input signal, and the device temperature sequence corresponding to the time axis is used as the modeling output signal. The length of the input and output signals is unified and the boundaries are aligned to form the equal-length input and output sample pairs required for thermal modeling, which are used to capture the causal response relationship between power consumption and temperature rise. Based on the input-output sample pairs, a fractional-order thermal RC network model of non-integer order is constructed. Multiple non-integer-order delay terms are introduced into the fractional-order thermal RC network model to describe the historical memory characteristics of heat in the conduction path, simulating hysteresis, diffusion, and material memory during heat conduction. The relationship between input and output is defined as consisting of three sub-terms: a thermal stability constant term, a low-order fractional-order response term, and a high-order fractional-order response term. This enables the modeling structure to simultaneously express thermal inertia and nonlinear heating trends. ; in, Indicates the number of fractional channels. Indicates the equivalent power loss input. Represents the thermal stability constant term. Indicates temperature difference. express The weights of each fractional derivative channel. Denotes the fractional derivative. Indicates order, Indicates the first A delay time, Indicates the first The weights of each delay channel, Indicates the number of delay channels. and The short-time constant channel represents the low-order fractional response term, and the long-time constant channel represents the high-order fractional response term. The low-order fractional response term is used to model the hysteresis effect of device temperature rise under short-cycle power excitation, and the high-order fractional response term is used to describe the slow temperature rise process caused by heat accumulation under long-cycle power excitation. Two fractional-order response terms are used as approximation targets. Upper and lower limits of the frequency range for fractional-order approximation are set, with the lower and upper limits representing the minimum operating frequency and the maximum thermal response frequency, respectively. An approximation order is also set to control the balance between approximation accuracy and computational complexity. Oustaloup recursive approximation expansion is applied to each fractional-order term to construct a set of logarithmically uniformly distributed frequency points. The frequency domain characteristics of the fractional-order operator are discretized and fitted. A cascaded structure of multiple first-order filters is used to simulate the amplitude-frequency and phase-frequency characteristics of the original fractional-order term, generating an equivalent integer-order model within the set range. The recursive approximation expression is: ; ; in, Denotes the fractional operator to be approximated. Represents the Laplace operator. This indicates the order of the fraction to be approximated. Indicates the order of the expansion. This represents the approximation gain constant. This represents the frequency of the poles of the denominator in the recursive approximation. This indicates the corresponding zero-point frequency, used to maintain the stability of the amplitude and phase frequency response within the target frequency band. , These represent the lower and upper frequencies of the approximation band, respectively. The equivalent integer-order model obtained by approximating the two fractional-order terms is substituted into the fractional-order thermal RC network model to replace the original fractional-order operators, while maintaining the consistency of the physical meaning and mathematical continuity of each term. The transformed fractional-order thermal RC network model is then associated with the input-output sample pair to form a complete equivalent integer-order model structure.

[0028] This step establishes a modeling foundation reflecting the relationship between power loss and temperature rise response by constructing equal-length input-output sample pairs from the basic feature sequence of current and device temperature signals. The model is based on the Oustaloup recursive approximation method, fitting the frequency domain characteristics of two fractional-order terms under set frequency range and order constraints. A first-order filter cascade structure is used to convert the fractional-order operators into an integer-order model. Finally, the approximated equivalent model is substituted into the original network structure, maintaining its physical and mathematical continuity, to form a complete equivalent integer-order thermal modeling framework.

[0029] In this embodiment, the generation of the benchmark thermal prediction model specifically includes: The equivalent integer-order model structure is used as a fixed modeling framework. The synchronized current sequence is used as the thermal modeling input signal, and the device temperature sequence is used as the thermal modeling output signal. The input and output signals are divided into equal lengths and resampled according to a fixed sampling period to generate multiple sample windows with fixed time spans. For each output sample window, the Prony identification algorithm is used to identify parameters, constructing a linear combination expression consisting of the current temperature and historical temperature values: ; in, This indicates the number of coefficients in the linear combination, i.e., the model order. Represents the feedback weights of an equivalent integer-order system. To represent the prediction error; the number of linear combination coefficients is set to the model order, and an error function is established for the linear combination: ; in, Indicates the window length. The feedback weight vector to be estimated is represented by: minimizing the sum of squared temperature prediction errors at all times within the window, constructing a system of linear matrix equations with a time series structure, and performing least squares estimation to obtain a parameter vector representing the dynamic characteristics of the thermal model. Each element of the parameter vector corresponds to the feedback weight of the integer-order system in the equivalent thermal modeling structure. Based on the parameter vector, a characteristic equation for discrete-time variables is constructed. Root value decomposition is used to obtain all complex solutions to the characteristic equation. Each complex solution is interpreted as a characteristic response pole. By taking the natural logarithm of these poles and dividing by the sampling period, the modal frequencies of the equivalent integer-order model structure in continuous time are obtained. Furthermore, the exponential function structure is recombined based on the modal frequencies to reconstruct a set of thermal response expressions composed of modal amplitudes and frequencies. The formation of these thermal response expressions involves identifying discrete poles using the Proni and least squares methods, which are then converted into continuous-domain modal parameters. The real part represents damping, i.e., the decay rate, and the imaginary part represents the angular frequency, i.e., whether oscillations exist. These modes are then used to determine the response shape. Weighted fitting is performed on the measured temperature sequence within the same window to obtain the corresponding modal amplitude as the contribution value. Thus, the overall temperature response is expressed as a superposition of several exponential or damped cosine components when conjugate modes exist. The resulting modal frequencies and modal amplitudes together constitute the thermal response expression, which is then substituted back into the equivalent integer-order model structure to form a benchmark thermal prediction model with dynamic identification capabilities.

[0030] This step effectively enhances the structural expressiveness and response sensitivity of the thermal modeling process by introducing a dynamic parameter estimation mechanism based on the Prony identification algorithm on top of the equivalent integer-order model. By establishing a feedback weight parameter vector through least-squares solving, an identification process adaptable to transient temperature rise changes in power devices is constructed. Furthermore, by combining modal decomposition and exponential function reconstruction methods, the dynamic response components in the modeling structure are clarified, enabling the thermal prediction model to possess time-varying tracking capabilities and interpretability, significantly improving the accuracy and physical consistency of thermal characteristic modeling under short-circuit conditions.

[0031] In this embodiment, the output process of the power loss characteristic specifically includes: The basic feature sequence extracted within the current sampling period is organized into a fixed-dimensional input vector. The temperature rise output of the benchmark thermal prediction model in the current sampling period is used as the target variable to form a joint sample composed of input features and prediction target, which is then fed into the kernel recursive least squares model. The kernel recursive least squares model is created based on the kernel recursive least squares algorithm. In the kernel recursive least squares model, the similarity between the current sample and all historical support samples is calculated based on the Euclidean distance between the input vectors using a set nonlinear kernel mapping structure. The similarity is then converted into a mapping result in the kernel space through an exponential weighting method, and the kernel similarity matrix for the current sampling period is constructed accordingly. Based on the minimum similarity value between the current sample and the supporting samples in the kernel mapping space, the novelty of the sample is evaluated. The novelty of the sample is compared with the preset novelty judgment threshold in the system. When the novelty is greater than or equal to the threshold, it is considered a valid sample and included in the update sequence. The sample is added to the dictionary and the internal kernel state matrix and weight coefficient vector are updated synchronously in a block update manner. This absorbs new operating condition information while maintaining model stability and updating the internal kernel state matrix and weight coefficient vector. If the novelty is lower than the threshold, the dictionary is not expanded. Only the weights are adjusted by rank-one update to control the model size and ensure real-time performance. The weight coefficient vector represents the contribution of each dictionary sample in the current prediction, that is, the linear combination coefficient that maps the input to the equivalent power loss. The internal kernel state matrix is ​​used to quickly complete the incremental solution and numerical stability control to obtain the output power loss characteristics.

[0032] This step, by introducing a kernel recursive least squares model and combining it with a novelty criterion, enables online adaptive updating of the thermal prediction model, effectively enhancing the system's learning ability to handle non-stationary operating conditions. By dynamically filtering support samples through a novelty threshold, not only is the accumulation of redundant data avoided and computational efficiency improved, but the impact of outlier samples on the model's robustness is also suppressed.

[0033] In this embodiment, the formation of the heat prediction result specifically includes: The output power loss feature is concatenated with the basic feature vector extracted in the current sampling period. The concatenation process first aligns the basic features and power loss features in time and resamples them uniformly. The standardization is completed according to the statistics of the training period. Then, two vector segments are generated according to a fixed list of fields with version numbers, and they are strictly concatenated into a single input vector in a predetermined order. At the same time, the corresponding feature name list is output for traceability and playback. A dynamic sample screening mechanism based on Mahalanobis distance is used to remove abnormal feature groups that deviate significantly from the training distribution. The minimum redundancy and maximum relevance selection method is performed on the remaining samples to obtain the target sub-feature set. The target sub-feature set is used to construct prediction samples and input into the quantile regression forest model. The quantile regression forest model contains several regression trees. Each tree assigns input samples to corresponding leaf nodes through structural partitioning. Within the leaf node where the predicted sample falls, all historical response values ​​are collected to form a response sample set. Based on a hierarchical quantile dynamic sampling mechanism, the heat response values ​​within the response sample set are resampled according to the target quantile to construct a pseudo-distribution. Finally, the quantile loss function is used to obtain estimates of local optima. ; in, Indicates in The first moment In a regression tree, for input features When performing regression, for a specified quantile The predicted value is the estimate of the local optimum obtained by minimizing the quantile loss function. Let a real number be used to search for the optimal predicted value during the optimization process. Indicates the first Among the returning trees, The leaf node number where it falls. This represents the function for finding the minimum value. Indicates the first The first of the trees A leaf node The historical response values ​​corresponding to the samples; This represents a linear loss function that applies asymmetric weights to the error, imposing a greater penalty when the error is negative, and is used to estimate the heat value at different quantile locations. The local optimum estimates in all regression trees are density-sensitively fused according to their corresponding weights. The corresponding weights are the number of samples that each regression tree routes to its leaf node, and the number of samples accumulated in the leaf node during the training period is taken as the local density indicator, which is recorded as the leaf sample number. The leaf sample number of all trees is normalized. The median prediction value, upper quantile prediction value and lower quantile prediction value are calculated separately and weighted averages are calculated to construct the current central hot prediction value and upper and lower confidence boundaries. At the same time, the inter-tree distribution offset index is introduced to statistically analyze the sparsity of the prediction results in the forest structure and dynamically correct the weight distribution. The median, upper quantile, and prediction interval of the heat forecast obtained at the current moment are used to form the heat forecast result. At the same time, the difference between the upper and lower boundaries of the heat forecast interval is calculated to represent the degree of uncertainty of the forecast. This value is used in the heat risk comparison and dynamic threshold determination module to support the real-time performance and accuracy of the heat risk management logic.

[0034] This step employs Mahalanobis distance combined with a minimum redundancy maximum correlation strategy to ensure that the input features possess high discriminative power and low redundancy, thereby improving the model's generalization ability. Hierarchical sampling and quantile loss optimization mechanisms are introduced into the quantile regression forest to accurately estimate the median and upper and lower confidence intervals of the heat prediction. Simultaneously, the interval width is used to characterize the strength of uncertainty, providing a quantitative basis for subsequent heat risk threshold comparison and enhancing the overall system's dynamic perception capability and response accuracy.

[0035] In this embodiment, the output process of the control trigger signal specifically includes: The median, upper quantile, and prediction interval of the current heat prediction are obtained from the heat prediction results to form the current heat prediction triplet. The static heat threshold and temperature rise threshold of the corresponding operating mode are called from the control system as comparison standards, and the allowable error tolerance is set as a dynamic safety boundary. The median predicted value of the heat prediction triplet is extracted as the heat estimate value, and the difference is compared with the heat threshold. The difference is compared with the preset tolerance range. When the comparison value exceeds the upper limit of the tolerance range, the current heat status is marked as a risk status to be confirmed, allowing the next step of temperature rise risk judgment logic to proceed. At the same time, the current heat offset value is recorded for trend modeling and dynamic adjustment. Based on the heat state marker, the change range between the current heat estimate and the median heat value of the previous sampling period is calculated as the heat prediction increment, and the difference between the upper and lower boundaries of the current heat prediction interval is obtained as the uncertainty range. The current heat prediction increment and the uncertainty range are compared. When the ratio exceeds the preset temperature rise rate threshold standard, it is determined that the current heat change has an abnormal rate, and an effective heat risk indicator signal is generated. When the thermal risk indicator signal is valid, the control logic module immediately outputs a control trigger signal and pushes the signal to the vehicle power management unit to activate the corresponding emergency control strategy path. The risk trigger event number, timestamp, trigger reason and related thermal prediction data are recorded in the system event log and status buffer pool.

[0036] This step establishes a dynamic risk identification mechanism based on a heat prediction triplet, integrating static threshold judgment and incremental rate assessment to achieve sensitive detection of short-term thermal anomalies. Preliminary risk is determined by the difference between the median and threshold values, and the rate of heat rise is assessed by combining the predicted increment with the uncertainty ratio, thus improving the accuracy and real-time performance of thermal risk identification. Once an anomaly is confirmed, the system immediately triggers a control signal and coordinates with the power management unit to respond. Simultaneously, risk events are numbered, recorded, and logged, constructing a closed-loop thermal safety response system.

[0037] In this embodiment, the generation of the control strategy specifically includes: Receive control trigger signals and synchronously read the real-time operating characteristics of the current vehicle power system, including output power, voltage level, temperature status and control mode identifier. Input the control trigger signals and operating characteristics together as index conditions into the control strategy scheduling module to generate strategy retrieval instructions. According to the strategy retrieval instruction, the control strategy library is accessed, and the current limiting parameter template and gating parameter configuration file that match the current operating condition are retrieved from it. The current limiting parameters include the maximum allowable current value, the slope drop coefficient and the current limiting start delay time. The gating parameters include the drive voltage duty cycle, the turn-off holding time and the soft turn-off gradient. They are integrated into the original parameter set of the control strategy. The control strategy library quickly matches and retrieves the appropriate control strategy according to the current system state and the thermal risk trigger signal, so as to achieve effective intervention and system protection against short-circuit thermal risks. Based on the original parameter set and the current system state, adaptive adjustments are made, the current limit value is dynamically scaled, the gating control logic pins are prioritized, encoded into control frames and encapsulated in signal format to form control command frames. The control command frame is sent to the vehicle power control unit via the communication interface, triggering the execution of current limiting, shutdown or delay action mechanisms, and the system records the policy issuance event and execution feedback results to complete the closed-loop process of thermal risk response.

[0038] This step dynamically generates control command frames through adaptive parameter adjustment and control priority sequencing, enabling multi-path intervention such as current limiting and soft shutdown. Control commands are rapidly sent to the power control unit via the communication interface, ensuring the immediacy and effectiveness of thermal risk response. Simultaneously, the execution process is recorded to form a closed-loop management system, enhancing the intelligent protection capabilities and operational stability of the vehicle power system.

[0039] Example 1: To verify the feasibility of this invention in practice, it was applied to a power supply short-circuit protection test project for a domestically produced new energy vehicle. The test site was located at a smart car R&D and testing base. This testing site possesses a comprehensive vehicle thermal control environment simulation system, capable of comprehensive control over different operating conditions, voltage levels, load types, and external ambient temperatures. The objective of this embodiment is to verify the predictive ability, response speed, and accuracy and stability of the control strategy of the "on-board power supply short-circuit heat prediction method based on big data analysis" proposed in this invention in a real vehicle environment, assessing its ability to predict thermal risks, response speed, and control strategy.

[0040] During testing, a prototype vehicle equipped with a permanent magnet synchronous motor drive system and a silicon-based MOSFET main control device was selected as the test subject. This vehicle used a traditional thermal RC model in its power system short-circuit protection design, which suffered from response delay and large prediction errors, and repeatedly failed to respond in time during internal self-tests, triggering thermal risks. To address this issue, researchers, without altering the original hardware structure, incorporated the method of this invention through software embedding to enhance its thermal management control strategy.

[0041] The specific application process is as follows: Data such as input voltage, output current, device temperature, and gate control signals of the vehicle under various typical operating conditions are collected. A sliding data buffer system is established to continuously track the temperature rise process. After removing measurement anomalies introduced by operating condition disturbances through empirical wavelet and Hampel filtering, a high-reliability feature sequence is constructed. A fractional-order thermal RC model is built based on the feature data, and integer-order thermal modeling parameters are obtained through Oustaloup recursive approximation and Prony identification. Then, the kernel recursive least squares method is combined to achieve online updates during the thermal modeling process, correcting modeling errors under abrupt operating conditions.

[0042] Subsequently, a thermal prediction interval model was constructed using quantile regression forest, and dynamic judgment logic was set in the vehicle controller to judge the deviation and rate of change of the median calorific value from the thermal threshold. When the system detects that the thermal rise trend exceeds the prediction tolerance, the control strategy scheduling module is immediately triggered to extract current limiting and gating parameters that match the current load and temperature state from the control strategy library, and encapsulate the control strategy and send it to the power supply main control unit within 20 milliseconds to execute fast shutdown and soft current limiting control.

[0043] Table 1. Measured data on thermal risk identification and response performance of vehicle power systems based on the method of this invention. ; As shown in Table 1, based on the test results of eight typical operating conditions, the average relative error between the median predicted heat value and the measured peak temperature rise is controlled within 3.1%, with the lowest error being 0.33% and the highest being 3.05%. The overall prediction accuracy is significantly better than that of the traditional thermal RC model. Regarding thermal risk identification, the system's average identification time under high-risk conditions is 38.4 milliseconds, maintaining good real-time performance without affecting control decisions. Simultaneously, the average response time of the control strategy is 15.7 milliseconds, with the fastest response time being 12 milliseconds and the slowest not exceeding 19 milliseconds, meeting the stringent time requirements of the vehicle system and reflecting stable and efficient command transmission between the strategy scheduling module and the power management unit.

[0044] 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 vehicle-mounted power supply short circuit heat prediction method based on big data analysis, characterized in that, The method comprises the following steps: Collecting real-time running data of the vehicle-mounted power supply system and storing the data in a sliding data buffer to obtain a buffered time sequence signal; Preprocessing the buffered time sequence signal to extract a basic feature sequence after noise removal; Based on the basic feature sequence after noise removal, a fractional order thermal RC network model is constructed, and the fractional order operator is converted into an equivalent integer order model structure through an Oustaloup recursive approximation algorithm; Parameters of the vehicle-mounted power supply are obtained by parameter estimation on the equivalent integer order model structure, and a benchmark thermal prediction model is generated in combination with the model structure; The benchmark thermal prediction model and the basic feature sequence in the current sampling period are taken as joint inputs, and effective samples are screened to perform online incremental updating on the model, and a power loss feature is output; The power loss feature is input into a quantile regression forest model to obtain a current time thermal prediction median value, an upper quantile value and a prediction interval, and a thermal prediction result is formed; The thermal prediction result is compared with a set thermal threshold and temperature rise threshold to determine whether to generate a thermal risk flag, and a control trigger signal is output when the thermal risk flag is valid; According to the control trigger signal and the current working condition feature, corresponding current limiting parameters and gating parameters are called from a control strategy library to generate a control strategy, and the control strategy is sent to a vehicle-mounted power supply control unit for execution.

2. The method of claim 1, wherein the method is characterized by, The buffered time sequence signal is obtained specifically as follows: A sampling period for collecting vehicle-mounted power supply system running data is set, a data collection process is started, and sampling data is received; In each sampling period, power supply parameter data is synchronously collected from the vehicle-mounted power supply system to generate time corresponding data records; The data records are written into a sliding data buffer in chronological order, the record contents are arranged according to the sampling order, control signals are divided and combined into a unified structure according to fields, and a buffered time sequence signal containing continuous multi-period data is obtained.

3. The method of claim 1, wherein the method is characterized by: The basic feature sequence after noise removal is extracted specifically as follows: Voltage, current and temperature related data are extracted from the buffered time sequence signal and respectively grouped into voltage sequences, current sequences and temperature sequences in chronological order; Empirical wavelet transform processing is performed on the voltage sequences, current sequences and temperature sequences respectively, and subsequences with main characteristic changes are retained in each type of signal, and high frequency detail subsequences with strong volatility are removed; Hampel filtering processing is performed on the subsequences with main characteristic changes to identify and replace abnormal fluctuation values in the signal, and the processed subsequences are recombined to form a basic feature sequence.

4. The method of claim 1, wherein the method is characterized by: The conversion of the equivalent integer order model structure specifically includes: The current sequence corresponding to the time axis in the basic feature sequence is taken as the modeling input signal, and the device temperature sequence corresponding to the time axis is taken as the modeling output signal, the input and output signals are unified in length and aligned in boundary to form an equal-length input and output sample pair required for thermal modeling; Based on the input-output sample pair, a fractional order thermal RC network model of non-integer order is constructed, a plurality of non-integer order delay terms are introduced into the fractional order thermal RC network model to simulate the hysteresis, diffusion and material memory in the heat conduction process, and the relationship between the input and the output is set to be composed of three sub-terms, including a thermal steady constant term, a low-order fractional order response term and a high-order fractional order response term; Two fractional order response terms are set as approximation targets, the upper and lower limits of the frequency interval of fractional order approximation are set, which are the lowest working frequency and the highest thermal response frequency respectively, and the approximation order is also set, and the Oustaloup recursive approximation expansion is used for each order fractional order term, a set of logarithmically uniformly distributed frequency points are constructed, the frequency domain characteristics of the fractional order operator are discretely fitted, and the amplitude-frequency and phase-frequency characteristics of the original fractional order term are simulated using a series structure of a plurality of first-order filters, and an equivalent integer order model is generated within the set range; The equivalent integer order models obtained after the approximation of the two fractional order terms are substituted into the fractional order thermal RC network model to replace the original fractional order operator, and the converted fractional order thermal RC network model is associated with the input-output sample pair to form a complete equivalent integer order model structure.

5. The method of claim 1, wherein the method is characterized by: The generation of the reference thermal prediction model specifically includes: The equivalent integer order model structure is used as a fixed modeling framework, the input and output signals are equally divided and resampled according to the fixed sampling period, and a plurality of sample windows with fixed time span are generated; For each output sample window, a Prony identification algorithm is used for parameter identification, a linear combination expression composed of the current time temperature and the historical time temperature value is constructed, a linear matrix equation with time sequence structure is constructed, and least square estimation is performed to obtain a parameter vector representing the dynamic characteristics of the thermal model; Based on the parameter vector, a characteristic equation about the discrete time variable is constructed, all complex solutions of the characteristic equation are obtained using root value decomposition method, a thermal response expression composed of modal amplitude and modal frequency is reconstructed, and the reference thermal prediction model with dynamic identification capability is formed by substituting the expression into the equivalent integer order model structure.

6. The method of claim 1, wherein the method is based on big data analysis. The output process of the power loss feature specifically includes: The basic feature sequence is arranged into a fixed-dimension input vector, the temperature rise output of the reference thermal prediction model is taken as a target variable, a joint sample is formed, and is sent into a kernel recursive least square model, which is created based on a kernel recursive least square algorithm; In the kernel recursive least square model, the similarity between the current sample and all historical support samples is calculated according to the Euclidean distance between the input vectors, the similarity is converted into a mapping result in the kernel space through an exponential weighting method, and a kernel similarity matrix of the current sampling period is constructed accordingly; Based on the minimum similarity value between the current sample and the support sample in the kernel mapping space, the novelty of the sample is evaluated, the novelty of the sample is compared with a preset novelty judgment threshold in the system, the internal kernel state matrix and the weight coefficient vector are updated, and the output power loss feature is obtained.

7. The method of claim 1, wherein the method is based on big data analysis. The formation of the heat prediction result specifically includes: The output power loss feature is spliced with the basic feature vector extracted in the current sampling period, a dynamic sample screening mechanism based on Mahalanobis distance is used to remove abnormal feature groups deviating greatly from the training distribution, and a minimum redundancy maximum correlation selection method is performed on the remaining samples to obtain a target sub-feature set, and the target sub-feature set is used to construct a prediction sample, which is input into a quantile regression forest model; The quantile regression forest model includes a plurality of regression trees, each tree divides the input sample into a corresponding leaf node through structural division, and in the leaf node where the prediction sample falls, all historical response values in the current node are collected to form a response sample set, and based on a hierarchical quantile dynamic sampling mechanism, the heat response values in the response sample set are resampled according to the target quantile to construct a pseudo-distribution, and a local optimal solution estimate value is obtained by using a quantile loss function; The local optimal solution estimate values in all regression trees are density-sensitive fused according to the corresponding weights, the weighted average of the median prediction value, the upper quantile prediction value and the lower quantile prediction value is calculated, and the central heat prediction value and the upper and lower confidence boundaries at the current time are constructed, and a distribution deviation index between trees is introduced to statistically analyze the sparsity of the prediction result in the forest structure and dynamically correct the weight distribution; The heat prediction median value, the upper quantile value and the prediction interval obtained at the current time are combined to form the heat prediction result.

8. The method of claim 1, wherein the method is based on big data analysis. The output process of the control trigger signal specifically includes: The heat prediction median value, the upper quantile value and the prediction interval at the current time are obtained from the heat prediction result to form a heat prediction triplet at the current time; The median prediction value of the heat prediction triplet is extracted as the heat estimate value, and a difference value is obtained by comparing the heat threshold value, and the difference value is compared with the preset tolerance interval, and when the comparison value exceeds the upper limit of the tolerance interval, the current heat state is marked as a to-be-confirmed risk state; Based on the heat state marking, the change amplitude between the current heat estimate value and the median heat value in the last sampling period is calculated as the heat prediction increment, and the upper and lower boundary difference of the current heat prediction interval is obtained as the uncertainty amplitude, and the current heat prediction increment and the uncertainty amplitude are operated by ratio, and when the ratio exceeds the preset temperature rise rate threshold, an effective heat risk flag signal is generated; When the heat risk flag signal is in the effective state, the control trigger signal is immediately output by the control logic module.

9. The method of claim 1, wherein the method is based on big data analysis. The generation of the control strategy specifically includes: The control trigger signal is received, and the real-time working condition characteristics of the current vehicle-mounted power supply system are read synchronously, the control trigger signal and the working condition characteristics are jointly input into the control strategy scheduling module as index conditions to generate a strategy retrieval instruction; According to the strategy retrieval instruction, the control strategy library is accessed to retrieve the current working condition state matched current-limiting parameter template and gate parameter configuration file, and the control strategy original parameter set is integrated, the control strategy library quickly matches and retrieves the appropriate control strategy according to the current system state and the heat risk trigger signal; According to the original parameter set, adaptive adjustment is performed combined with the current system state, the current limit value is dynamically scaled, the gate control logic pin is prioritized, and the control frame is encoded and signal format packaged to form the control instruction frame; The control instruction frame is sent to the vehicle-mounted power supply control unit via the communication interface, and the strategy issuing event and the execution feedback result are recorded in the system, so as to complete the heat risk response closed loop process.

Images