A full-function integrated processing system of vehicle body domain control fusing whole vehicle power management

Abstract

The application relates to the technical field of automobile electronics and edge computing, and discloses a vehicle body domain control full-function integrated processing system fusing whole-vehicle power management, which comprises a data sensing module, a computing scheduling module, a state prediction module and a protection execution module.The data sensing module collects multi-modal signals and constructs a high-dimensional time sequence feature tensor.The computing scheduling module performs heterogeneous computing scheduling by using a sentinel core unit and an inference core unit, and the inference core runs a fixed-point quantized time sequence convolution network.The state prediction module predicts a load current based on the time sequence convolution network, and performs residual error checking on the prediction result by using a first principle value of thermodynamics.The protection execution module generates a dynamic off threshold value from the checked prediction value and converts the dynamic off threshold value into an analog reference voltage.The system compares the real-time current with the analog reference voltage by using a hardware comparator, and triggers a hardware latch logic at an overcurrent moment.The application solves the problems of poor interpretability and inference delay of a deep learning model in an embedded edge end through deep fusion of data driving and physical model constraints.

Description

Technical Field

[0001] This invention relates to the field of automotive electronics and edge computing technology, specifically to a fully integrated processing system for vehicle domain control that integrates vehicle power management. Background Technology

[0002] As automotive electronic and electrical architecture evolves towards regional controllers, the types and number of vehicle loads are increasing dramatically, placing higher demands on the safety and real-time performance of power management. Existing protection mechanisms mostly employ fuses or are based on I / O... 2 The thermal accumulation algorithm of t, this type of passive protection relies on the accumulation of physical thermal effects. The response time is usually in the hundreds of milliseconds or even seconds. It is difficult to deal with hard short circuit faults in the milliseconds and is prone to overheating and insulation failure of the wiring harness.

[0003] Furthermore, to avoid the surge current during motor startup, traditional solutions typically set high protection thresholds, resulting in blind spots in low-current fault ranges and failing to balance startup passability and steady-state sensitivity. Although deep learning-based fault diagnosis technologies are increasingly being applied in industrial scenarios, general-purpose neural network models usually have a large number of parameters and high computational complexity, making them difficult to deploy directly on resource-constrained embedded edge computing nodes. Simultaneously, purely data-driven "black box" models lack physical common sense and are prone to nondeterministic outputs when faced with out-of-distribution (OOD) samples. Using them to drive millisecond-level low-level hardware protection logic poses significant security risks. Moreover, existing deep learning algorithms, when deployed on resource-constrained embedded chips, often struggle to balance inference accuracy, computational real-time performance, and physical interpretability, lacking effective physical constraint mechanisms to ensure the safety of model outputs.

[0004] Therefore, this invention proposes a fully integrated processing system for vehicle domain control that integrates vehicle power management to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a fully integrated processing system for vehicle power management, which integrates vehicle domain control functions. This system solves the problems of existing vehicle power management relying on heat accumulation mechanisms, which leads to delayed fault response, and the inability of simply applying neural network models to balance inference computation latency and microsecond-level hardware deterministic protection requirements at embedded edge devices.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a fully integrated processing system for vehicle domain control that integrates vehicle power management, comprising: The data sensing module is configured to collect current signals, voltage signals, and ambient temperature signals of the load circuit, and construct a time-series feature tensor based on the collected signals; The computation scheduling module includes a sentinel core unit and an inference core unit. The sentinel core unit is configured to run resident to perform deterministic logic computation, and the inference core unit is configured to run neural network inference computation. The state prediction module is configured to predict the load current prediction value at the next moment using a temporal convolutional network based on the temporal feature tensor, and to calculate the theoretical physical current value based on the real-time acquired voltage signal and the estimated loop temperature using a physical mechanism model to verify the load current prediction value. The protection execution module is configured to generate a dynamic shutdown current threshold based on the verified load current prediction value, and convert the dynamic shutdown current threshold into an analog reference voltage. The protection execution module includes a hardware comparator configured to compare the real-time acquired current signal with an analog reference voltage, and to trigger hardware latching logic to cut off power output when the amplitude of the current signal exceeds the analog reference voltage.

[0007] Preferably, before constructing the time-series feature tensor, the data sensing module is configured to calculate the equivalent temperature drift coefficient of the hybrid material circuit; the data sensing module reads a preset set of physical parameters including the resistance temperature coefficients of the connecting harness and the power device; the data sensing module performs a weighted fitting of the resistance temperature coefficients of the connecting harness and the power device based on their respective weights in the nominal impedance of the entire circuit to obtain the equivalent temperature drift coefficient, which is used to characterize the overall impedance change characteristics of the hybrid material circuit under a single ambient temperature variable.

[0008] Preferably, the computation scheduling module is configured to execute hard feedforward control logic in the transient phase; when the data sensing module detects a step change in the control intent signal, the computation scheduling module determines that the system has entered the transient control phase and temporarily bypasses the output of the inference core unit; in the transient control phase, the computation scheduling module forcibly sets the hysteresis voltage window of the hardware comparator to the maximum tolerance threshold and maintains it for a preset hold time; the hold time is determined based on the sum of the average inference time of the inference core unit and the safety margin coefficient to cover the latency blind zone of model inference.

[0009] Preferably, the computation scheduling module is configured to execute a sparse inference strategy during the steady-state control phase; the sentinel core unit monitors the sum of the absolute values ​​of the first-order differences of the current data in the time-series feature tensor in real time to calculate the fluctuation characteristic index; the computation scheduling module compares the fluctuation characteristic index with a preset wake-up threshold: if the fluctuation characteristic index is greater than the wake-up threshold, a wake-up interrupt is sent to the inference core unit to trigger a new round of inference computation; if the fluctuation characteristic index is less than or equal to the wake-up threshold, the inference core unit is controlled to remain in a sleep state and maintain the load current prediction value of the previous frame.

[0010] Preferably, the temporal convolutional network running in the state prediction module includes multiple stacked dilated causal convolutional layers; the temporal convolutional network receives temporal feature tensors as input and uses convolutional kernels with dilation factors that increase exponentially with the number of layers to extract current fluctuation features within a historical time window; after training, the temporal convolutional network undergoes fixed-point quantization to map floating-point weight parameters to fixed-point integers and deploys them in the inference kernel unit to establish a nonlinear mapping relationship between input features and future current states.

[0011] Preferably, the state prediction module is configured to execute residual verification logic based on first principles; the state prediction module runs thermodynamic iterative equations, estimates the conductor temperature based on the balance between Joule heat generated by the real-time current and heat dissipation from the environment, and calculates the theoretical physical current value by combining the equivalent temperature drift coefficient; the state prediction module calculates the Euclidean distance residual between the predicted load current value and the theoretical physical current value; if the Euclidean distance residual is less than the dynamic confidence threshold, the predicted load current value is accepted; if the Euclidean distance residual is greater than or equal to the dynamic confidence threshold, the output is corrected using the theoretical physical current value.

[0012] Preferably, the protection execution module performs envelope synthesis logic when generating the dynamic shutdown current threshold; the protection execution module calculates the standard deviation statistics of historical current data within the sliding window to quantify the current ripple characteristics; the protection execution module multiplies the predicted load current value by the proportional gain coefficient, and superimposes the standard deviation statistics calculated based on the historical current data with a fixed safety tolerance to synthesize the dynamic shutdown current threshold that closely follows the physical limits of the load, so as to cover normal load fluctuations and reserve dynamic margin.

[0013] Preferably, the protection execution module further includes a digital-to-analog converter and an SR latch; the digital-to-analog converter is connected to the inverting input of a hardware comparator and is used to convert the digital value of the dynamic shutdown current threshold into an analog reference voltage; the output of the hardware comparator is connected to the set pin of the SR latch, and the output of the SR latch is connected to the output disable pin of the power switch drive circuit; when the SR latch is set, the power switch is forcibly disconnected through a pure hardware link.

[0014] Preferably, the system further includes a diagnostic feedback module, which is configured with an active detection function; the diagnostic feedback module monitors the conductor temperature in real time after the hardware latching logic is triggered, and controls the power device to output a trial voltage pulse when the conductor temperature falls below the safe baseline; the duration of the trial voltage pulse is limited by the safe operating area of ​​the power device, and the injected energy corresponding to the duration is less than the single-pulse avalanche energy limit of the power device.

[0015] Preferably, the diagnostic feedback module is configured to execute impedance verification logic based on voltage domain residuals; during the output of the trial voltage pulse, if the hardware comparator is not triggered, the diagnostic feedback module acquires the sampled voltage and sampled current at the end of the trial voltage pulse; the diagnostic feedback module calculates the residual amplitude between the sampled voltage and the estimated voltage, where the estimated voltage is the product of the sampled current and the temperature-corrected nominal impedance; if the residual amplitude is less than the impedance verification tolerance, the previous overcurrent is determined to be a soft interference and a system reset is performed; if the residual amplitude is greater than the impedance verification tolerance, an impedance abnormality is determined and a locked state is maintained.

[0016] This invention provides a fully integrated processing system for vehicle domain control that integrates vehicle power management. It offers the following advantages: 1. This invention introduces a physical mechanism model based on the first principles of thermodynamics in parallel outside the inference channel of a temporal convolutional network. The theoretical values ​​calculated using physical equations are then used to perform real-time residual verification on the neural network's predicted output. This fusion computing architecture imposes deterministic physical constraints on the deep learning model, effectively correcting nonlinear prediction biases that may arise when the neural network faces unseen fault samples or data distribution shifts. This ensures that the output of the computational model always remains within physically interpretable safety boundaries, enhancing the credibility of artificial intelligence algorithms in automotive-grade safety-critical systems.

[0017] 2. This invention employs a heterogeneous computing power allocation strategy, deploying low-latency logic monitoring and high-performance model inference on the sentinel core unit and inference core unit, respectively. By designing a transient hard feedforward mechanism to fill the inference blind spot during model cold start, and utilizing a sparse inference strategy based on fluctuation characteristics to dynamically wake up the inference core, the system achieves on-demand allocation of computing resources while ensuring low power consumption. This architecture enables the deployment of complex temporal convolutional networks on resource-constrained embedded edge devices, and, without sacrificing response speed, utilizes deep learning models to accurately distinguish between load-starting surges and short-circuit faults.

[0018] 3. This invention constructs a high-dimensional time-series feature tensor containing current, voltage, and temperature. It utilizes dilated causal convolution to capture the long-cycle thermal accumulation characteristics of the load and synthesizes the probability predictions output by the neural network into a dynamic current envelope. By mapping this digitized dynamic envelope to an analog reference voltage in real time and applying it to a hardware comparator, this invention directly transforms the abstract algorithm inference results into microsecond-level physical layer control actions. It achieves hardware-level fault truncation without the need for central processing unit intervention, solving the technical problem of traditional software protection mechanisms failing under crash or high-load conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the all-condition threshold control strategy of the present invention; Figure 4 This is a schematic diagram comparing the hard short-circuit response of the present invention.

[0020] Among them, 100 is the data perception module; 200 is the calculation and scheduling module; 300 is the status prediction module; 400 is the protection execution module; and 500 is the diagnostic feedback module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 This invention provides a fully integrated processing system for vehicle domain control that integrates vehicle power management. The system includes: The data sensing module 100 is used to acquire physical signals of the vehicle body load circuit. The data sensing module 100 acquires the circuit current signal and the high-side switch input voltage signal through an analog-to-digital converter interface, and acquires the ambient temperature signal and control intent signal through a communication interface. The data sensing module 100 performs timing alignment on signals with different sampling frequencies to construct an input tensor containing current, voltage, temperature, and intent features.

[0023] The computation scheduling module 200 is used to allocate computational resources and generate control boundaries. When the control intent signal undergoes a step change, the computation scheduling module 200 generates a maximum tolerance comparator hysteresis voltage. Based on the fluctuation characteristics of the input data, the computation scheduling module 200 dynamically activates the inference kernel unit and sends scheduling instructions to the state prediction module 300.

[0024] The state prediction module 300 is used to perform physical calculations and model inference in parallel. The state prediction module 300 runs a physical observation algorithm, using thermodynamic equations to calculate the conductor temperature and the theoretical physical current value. The state prediction module 300 runs a convolutional network to extract the temporal features of the input tensor and outputs the predicted load current value. The state prediction module 300 calculates the residual between the theoretical physical current value and the predicted load current value.

[0025] The protection execution module 400 is used to perform circuit protection actions. The protection execution module 400 adjusts the hardware comparator reference voltage and hysteresis window width based on predicted values. The protection execution module 400 monitors the loop current and triggers hardware latching logic when the loop current exceeds a hardware limit threshold, cutting off power output and generating a fault signal.

[0026] The diagnostic feedback module 500 is used to identify the nature of the fault. In response to a fault signal, the diagnostic feedback module 500 controls the power device to conduct a trial pulse. During the trial pulse, the diagnostic feedback module 500 keeps the hardware overcurrent comparator active. Based on whether the pulse is hardware truncated, the diagnostic feedback module 500 determines whether it is a hard short-circuit fault or a soft interference, and updates the system reset state accordingly.

[0027] See attached document Figure 2 This invention provides a fully integrated processing method for vehicle domain control that integrates vehicle power management, comprising the following steps: S100 loads the physical model parameter set, synchronously acquires loop current and high-side voltage signals, performs upsampling processing on ambient temperature and control intention signals, and constructs a multimodal input tensor; S200 detects the step characteristics of the control intention signal, outputs the maximum hysteresis voltage boundary in the transient phase, performs linear decay calculation in the transition phase, and dynamically fine-tunes the voltage boundary based on the prediction results in the steady-state phase. S300 uses dilated convolutional layers to extract temporal features of the input tensor, outputs predicted load current values ​​for future times through a residual connection network, and calculates the reliability residual of the prediction results by combining theoretical physical current values. S400 uses thermodynamic equations to iteratively calculate the real-time temperature of the conductor, combines the loop impedance model to calculate the theoretical physical current value, and calculates the Euclidean distance residual between the predicted load current value and the theoretical physical current value. The S500 performs hardware latching when the loop current exceeds the hardware limit threshold, sends a probe pulse and keeps the hardware comparator active during the pulse, determines the fault type based on the pulse response characteristics and performs a reset or lockout operation.

[0028] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0029] See attached document Figure 1 In this embodiment, the data perception module 100 executes the following data processing logic, which aims to construct a high-dimensional temporal feature tensor, thereby ensuring that the heterogeneous data input to the subsequent calculation model is strictly aligned in the temporal domain and has consistent physical characteristics.

[0030] S101, Loading of physical model parameters and definition of basic impedance. During the system power-on initialization phase, the data sensing module 100 reads a preset set of physical parameters from non-volatile memory. This parameter set contains the nominal impedance of the entire loop, serving as the basis for subsequent thermodynamic calculations. Its value is a linear superposition of the load internal resistance, the wiring harness resistance, and the on-resistance of the power devices. Considering manufacturing process tolerances and initial environmental differences, the nominal impedance of the entire circuit is... Calculated using the following formula: ; In the formula, For the load at the reference temperature (Usually taken as the rated internal resistance at 25℃) For connecting wire harnesses The nominal resistance below, For high-side drive power devices in The on-resistance is measured below. In this embodiment, the above impedance parameters are all obtained based on high-precision bridge testing to eliminate initial errors caused by hardware differences.

[0031] S102, Calculation of equivalent temperature drift coefficient for hybrid materials. Given that the vehicle body load circuit contains two different materials—copper wiring harnesses and silicon-based power devices—and that their impedance drift characteristics to temperature changes differ significantly (e.g., copper exhibits a linear positive temperature coefficient, while semiconductors have nonlinear characteristics), calculating them separately and then summing them would consume unnecessary computational power. Therefore, the data sensing module 100 introduces the equivalent temperature coefficient of the circuit. This coefficient is not a physical constant of a single material, but rather a weighted fitting value based on the proportions of the wiring harness and power devices in the total impedance, used to uniformly characterize the temperature drift characteristics of hybrid circuits. (Circuit equivalent temperature coefficient) Defined by the following formula: ; In the formula, This is the temperature coefficient of resistance of copper conductors, typically ranging from 0.00393 / ℃ to 0.00400 / ℃. This represents the equivalent resistance temperature coefficient of silicon-based power devices. This weighted calculation logic ensures that the physical model can operate based on a single temperature variable even when ambient temperature fluctuates significantly. The overall impedance change of the hybrid material circuit can be accurately calculated, thereby reducing the computational complexity while ensuring accuracy.

[0032] S103, High-frequency analog signal synchronous acquisition. To obtain real-time circuit status, the data sensing module 100 controls the analog-to-digital converter (ADC) to perform multi-channel synchronous sampling. This is specifically for loop current signals. voltage signal at the high-side switch input terminal The system utilizes the synchronization group triggering function of the microcontroller's (MCU) internal ADC module, or the hardware trigger signal of the master-slave ADC, to ensure that the sample-and-hold circuits of both channels operate within the same clock cycle. This step eliminates phase deviation caused by channel switching time differences (typically several microseconds) at the hardware level. This ensures that the calculation of real-time impedance... When the numerator and denominator strictly correspond to the system state at the same physical moment, it effectively prevents false impedance fluctuation noise introduced due to phase misalignment.

[0033] S104, heterogeneous data timing alignment, upsampling, and normalization. The system collects ambient temperature signals. With control intent signal Typically transmitted via CAN or LIN bus, its update frequency (e.g., 100Hz) is much lower than the current sampling frequency (e.g., 10kHz). To address the data sparsity problem caused by the frequency mismatch of multimodal data, the data sensing module 100 performs upsampling processing. For continuously changing ambient temperature signals... A linear interpolation algorithm is used to calculate the high-frequency temperature sequence corresponding to the current sampling time based on the timestamps and values ​​of two adjacent bus data frames, in order to maintain a smooth transition of physical quantities; while for control intention signals that exhibit step changes... The zero-order hold algorithm is used to keep the state value of the current bus frame unchanged until the data of the next frame is updated.

[0034] After timing alignment is completed, to accelerate the convergence of the neural network model and prevent gradient explosion caused by differences in dimensions (such as current in amperes and voltage in volts), the data sensing module 100 performs Z-score normalization on the data of each channel: ; In the formula, For the first The raw data of each feature channel For standardized data, and These represent the mean and standard deviation of the channel within a historical statistical window. To prevent small constants with a denominator of zero (such as 1e) -6 The introduction of this constant enhances the numerical stability of the algorithm when processing constant signals.

[0035] S105, Multimodal Input Tensor Construction and Training Data Definition. The data-aware module 100 encapsulates various signals, after synchronization, alignment, and normalization, into standardized input tensors. In this embodiment, the input tensor Constructed as a dimension A three-dimensional matrix, in which This is the batch size (1 during the inference phase). Number of characteristic channels (including current) ,Voltage ,temperature and intention (4 channels in total) The length of the sliding time window (e.g., taking the past 100 sampling points).

[0036] In this embodiment, the Temporal Convolutional Network (TCN) needs to undergo supervised learning training before deployment. Its training employs an autoregressive model, utilizing… Historical time-series data prior to the given time point is used as input. The actual sampled current value at each moment is used as the supervision label. The training sample set comes from fault waveforms generated by high-precision circuit simulation software and actual load waveforms acquired from bench experiments. A loss function is defined during the training process. Mean squared error (MSE) is used to quantify the current value predicted by the model at the next time step. Compared with the true sampled value Differences between them: ; In the formula, This represents the number of training samples; This serves as the sample index. By minimizing this loss function, the model can learn the current evolution patterns of different load types under diverse operating conditions, establishing a nonlinear mapping relationship between input features and future states. The completed input tensor... It is then transmitted to the hybrid computing scheduling module 200 for subsequent feature extraction and inference calculations.

[0037] See attached document Figure 1 In this embodiment, the computing scheduling module 200 adopts the heterogeneous computing power allocation strategy and phased hybrid control logic described below, aiming to resolve the inherent contradiction between the inference latency of deep learning models and the real-time transient response of vehicle load.

[0038] S201, constructing a heterogeneous computing resource pool and a hierarchical operation mechanism. To achieve high energy efficiency in real-time control, the computing scheduling module 200 divides the central processing unit (CPU) logic resources of the vehicle domain controller into two functionally distinct computing domains: the sentinel core unit and the inference core unit. Specifically, the sentinel core unit is typically mapped to a high real-time core (such as the ARM Cortex-R series) in the MCU and is configured to run low-computational-consumption deterministic logic, including thermodynamic physical model solving and basic safety threshold comparisons (described later). As a resident monitoring node of the system, the sentinel core unit remains operational throughout the system's power-on period to ensure microsecond-level response capabilities to extremely high-risk faults such as hard short circuits. In contrast, the inference core unit integrates a neural network processing unit (NPU) or digital signal processing (DSP) instruction set, specifically designed to execute high-computational-consumption temporal convolutional network (TCN) inference tasks. Considering the power sensitivity of onboard power management, the inference core unit is in a low-power sleep state by default, switching to a high-frequency operation mode only upon receiving a specific wake-up event or scheduling instruction.

[0039] S202, hard feedforward pre-control during the transient phase. When the data sensing module 100 detects the control intention signal... When a step change from 0 to 1 occurs (i.e., the instant the load is turned on), the system determines that it is currently in a high-risk transient control phase. This is due to the inherent inference latency in deep learning models, from input data loading and convolution operations to output. (e.g., 10ms to 20ms) This delay may cause the protection logic to lag during the load inrush current. Therefore, the computation scheduling module 200 employs a pre-intervention strategy, temporarily bypassing the output of the inference core unit and directly executing the hard feedforward logic. During this period, the system forcibly adjusts the hysteresis voltage window of the hardware comparator. Set as the maximum allowable tolerance threshold And maintain this state for a preset duration. .

[0040] To ensure coverage of the entire model inference process and to reserve sufficient safety margin, the retention time is... Determined by the following formula: ; In the formula, This represents the average inference time of the model on the current hardware platform. This is a safety margin factor, typically ranging from 0.1 to 0.2, used to compensate for time skew caused by operating system scheduling jitter. Maximum tolerance threshold. The maximum heat tolerance of the load harness under short-term overload is set according to the load harness, usually 10% to 15% of the nominal operating voltage, to ensure that the system will not malfunction due to the start-up surge current during the blind zone time when the model calculation results are not yet ready.

[0041] S203, linear decay and smooth switching during the transition phase. (When the system runtime...) Exceeding the holding time Afterwards, the transition control phase begins. At this point, the TCN model has completed its first round of inference and starts outputting valid prediction results. To avoid the protection threshold from falling from a high level... Suddenly jumps to the actual target value calculated by the model This triggers control oscillations, and the scheduling module 200 uses a linear interpolation algorithm to generate a smooth decay trajectory. During this stage, the hysteresis voltage window... The calculation follows the following logic: ; In the formula, The real-time predicted boundary values ​​output by the state prediction module 300; This is the voltage decay slope coefficient, with units of V / ms. The value of is determined by the fastest allowable convergence time calibrated by the system, aiming to control the protection boundary to approximate the actual load characteristics in a smooth and gradual manner, ensuring both safety during the transition period and a smooth transfer of control. In the formula... The introduction of the function ensures that the decay process will not fall below the safe lower limit predicted by the model, preventing false triggering due to excessively rapid decay.

[0042] S204, sparse inference and dynamic fine-tuning in the steady-state phase. When the hysteresis voltage window fully converges to... After the load current reaches a relatively stable state, the system enters the steady-state control phase. During this phase, to address the energy efficiency limitations of embedded systems, the computation scheduling module 200 executes a sparse inference strategy. The system utilizes the sentinel core unit to monitor the input tensor in real time. The rate of change is used to calculate the volatility characteristic index. In this embodiment, Calculation based on the sum of the absolute values ​​of the first-order differences of the currents within the sliding window: ; In the formula, The length of the detection window (e.g., 10 sampling points); This is the historical sampled current value. The scheduling module 200 will calculate and schedule it. With preset wake-up threshold Compare. Only when When this occurs, it indicates a significant fluctuation in the load condition (such as the initial stage of stall or poor contact). At this time, the system sends a high-priority wake-up interrupt to the inference core unit, triggering a new round of TCN inference calculation to update the load. Conversely, if The inference kernel unit remains dormant, and the control boundary maintains the prediction value of the previous frame or only makes low-power linear corrections based on the physical model, thereby significantly reducing the average computational power consumption in steady state.

[0043] See attached document Figure 1 In this embodiment, the state prediction module 300 constructs a dual-channel processing architecture that combines "deep learning inference" and "physical mechanism calculation". Through the mutual verification of data-driven and first-principles calculations, it achieves high-precision prediction and interpretability verification of the vehicle body load current state.

[0044] S301, Construct a Temporal Convolutional Network (TCN) topology. To extract long-term dependency features from time-series current signals containing electromagnetic noise, the state prediction module 300 establishes a neural network model composed of multiple layers of dilated causal convolutions. Compared to traditional recurrent neural networks (RNNs) which are limited by serial computation bottlenecks, the TCN model used in this embodiment supports parallel computation and utilizes the dilation factor. With the exponential increase in the number of layers, the receptive field expands exponentially without increasing the number of parameters, enabling it to cover a sufficiently long historical time window to capture the macroscopic trend of current fluctuations.

[0045] Specifically, the network consists of an input layer, several residual blocks, and a fully connected output layer. The input layer has a receiving dimension of... The tensor, in which The number of feature channels, This refers to the length of the time window. The backbone network consists of... The residual blocks are connected in series (in this embodiment) Each residual block contains two dilated convolutional layers, a weight normalization layer, a ReLU activation function, and a Dropout regularization layer. Layer expansion factor Set as (i.e., 1, 2, 4, 8), kernel size Set to 3. Based on this configuration, the network's effective receptive field is... It covers all historical information of the input sequence: ; In the logic of convolution computation, the first... The first layer Output elements Following the rules of causal convolution, its calculation formula is expressed as follows: ; In the formula, This is the input feature map for the previous layer; For the first Each convolutional kernel weight; For bias terms; denoted as the inflation factor of the current layer. This formula shows that the output at the current time step depends only on the input at previous time steps, strictly satisfying the causality constraint. Furthermore, to address the vanishing gradient problem that arises with increasing network depth, the network introduces a residual connection structure, directly adding the input tensor to the output of the convolutional layer, ensuring the training stability of deep networks and the effectiveness of feature transfer.

[0046] S302, Network Model Lightweighting and Fixed-Point Quantization. Considering the storage space and computing power limitations of automotive-grade microcontrollers (MCUs), the state prediction module 300 executes a post-training quantization strategy after model training. This step aims to map the original 32-bit floating-point (FP32) weight parameters to 8-bit fixed-point integers (INT8), thereby significantly reducing the Flash space occupied by the model and accelerating inference using the MCU's DSP instruction set. The quantization process follows the following affine transformation formula: ; In the formula, These are the original floating-point weight values; The quantized integer value; The scaling factor is determined by the maximum absolute value of the weight tensor; This is a zero-point offset. This embodiment employs a symmetric quantization scheme, forcing... Furthermore, the truncation threshold is set to the 99.9th percentile of the weight distribution to preserve the dynamic range of the parameters to the maximum extent and ensure that the accuracy loss after quantization is less than 0.5%.

[0047] S303 is a deterministic physical observation based on thermodynamic equations. While performing deep learning inference, the state prediction module 300 runs a first-principles-based physical mechanism model in parallel. The core of this model lies in the real-time estimation of the thermal accumulation state of the conductor, as the conductor temperature directly determines the physical impedance, thus affecting the theoretical current value. The system utilizes Joule's law and the principle of heat conduction to construct iterative thermodynamic equations in the discrete-time domain: ; In the formula, The estimated conductor temperature at the current moment; This refers to the temperature state at the previous moment; For real-time sampling of current; The loop resistance is corrected based on the temperature at the previous moment; To calculate the step size (i.e., the sampling interval, such as 100 μs); The heat capacity of the wire harness is determined by the specific heat capacity of the wire material. Its quality The product determines ( ); The thermal resistance of the wiring harness to the environment depends on the thermal conductivity and surface area of ​​the insulation material. The physical meaning of this equation lies in dynamically simulating the energy balance process between the Joule heat generated by the current and the heat dissipated from the environment.

[0048] S304, Calculation of theoretical physical current value. The conductor temperature is calculated based on the iterative calculation of the above thermodynamic model. The system combines the data sensing module 100 to calculate the loop equivalent temperature drift coefficient. The theoretical physical impedance under the current operating conditions is calculated, and the synchronously acquired high-side voltage is used to deduce the impedance. Calculate the theoretical physical current value : ; In the formula, Reference temperature The nominal impedance of the entire circuit is as follows; The minimum constant (e.g., 1e) -6 This step is used to prevent calculation overflow caused by a zero denominator. The output of this step represents the current reference value that the loop should present under ideal conditions, considering only Ohm's law and without faults or nonlinear disturbances (such as motor back EMF).

[0049] S305, based on first-principles residual verification. To prevent prediction distortion or nonlinear bias (i.e., outputting unreasonable predicted values) when the deep learning model encounters unseen adversarial samples or extreme conditions, the state prediction module 300 calculates the TCN model's predicted current. Calculated value from theoretical physical current Euclidean distance residuals between : ; The residual The deviation between "data-driven prediction" and "physical mechanism constraints" was quantified. The system sets a dynamic confidence threshold. To accommodate the increased uncertainty in the physical model under high-temperature conditions, this threshold is designed to relax linearly with temperature: ; In the formula, The basic tolerance (usually taken as 15% of the theoretical current value); This is the temperature compensation coefficient. If... This indicates that the model's predicted values ​​conform to physical laws, and the system adopts the prediction results of TCN to generate fine control boundaries; if If the current condition is abnormal, it indicates that the current operating condition may be beyond the model's cognitive range or that there is a nonlinear physical abrupt change. In this case, the state prediction module 300 executes a weighted fusion strategy or a direct circuit breaker mechanism, outputting the predicted load current value to subsequent modules. Forced correction to theoretical physical current value Alternatively, preset safety boundary values ​​are used to ensure the interpretability and safety of AI algorithms in industrial control applications.

[0050] See attached document Figure 1 In this embodiment, the protection execution module 400 serves as a key bridge connecting the digital algorithm domain and the analog physical domain. It is responsible for mapping the digital prediction results generated by the state prediction module 300 into the control voltage of the physical circuit in real time, and realizing the microsecond-level shutdown of overcurrent faults through a pure hardware link, thereby ensuring system safety even under extreme conditions such as software failure or MCU crash.

[0051] S401, Synthesis of Dynamic Control Boundary and Generation of Envelope. To maximize the release of load transient performance while ensuring safety, the protection execution module 400 no longer uses a single static threshold, but instead synthesizes a time-varying current envelope that closely follows the physical limits of the load based on the prediction results. The system reads the average value of the predicted current at the next time step from the TCN model output. By combining the statistical characteristics in the input tensor, the dynamic turn-off current threshold is calculated. The logic for synthesizing this threshold is designed to cover normal load fluctuations and measurement noise, while reserving a reasonable dynamic margin.

[0052] Specifically, the dynamic shutdown current threshold Calculated using the following envelope synthesis formula: ; In the formula, This is the proportional gain coefficient (e.g., 0.1), used to compensate for linearity errors in the model when there are large dynamic changes. To establish a fixed safety tolerance, this method covers zero-point drift and quantization errors in the hardware link. Specifically, to accurately quantize current ripple characteristics and avoid false triggering due to signal noise floor, this embodiment introduces a standard deviation statistic. As a dynamically floating term, its calculation is based on the current sliding window. Historical current data within: ; In the formula, This represents the average current within the window. The confidence interval coverage factor is usually set to 3. Based on Gaussian distribution theory, this setting ensures that the dynamic boundary can cover 99.7% of random noise fluctuations, thereby achieving a precise distinction between "normal ripple release" and "abnormal mutation truncation" at the physical level.

[0053] S402, Digital-to-Analog Conversion and Physical Control Signal Generation. The calculated digital quantity... It needs to be converted into a physical voltage signal before it can act on the hardware comparator. The protection execution module 400 uses the high-speed digital-to-analog converter (DAC) inside the microcontroller to map the current threshold to an analog reference voltage. Considering the physical range limitations and resolution characteristics of DAC hardware, this mapping process needs to include saturation truncation logic and discretization calculations to prevent numerical overflow. ; ; In the formula, For theoretical voltage calculation; The value of the sampling resistor (e.g., 0.5mΩ); For operational amplifier gain (e.g., 20V / V); This is the analog reference source voltage for the DAC (typically 3.3V or 5V). This refers to the resolution bit depth of the DAC (e.g., 12 bits, which is 4095 levels). This is the integer value that will ultimately be written to the DAC data register. In the formula... The function is introduced to handle the overflow situation where the theoretically calculated value exceeds the hardware power rail, and to prevent serious faults such as the DAC output returning to zero due to numerical rollback, which would cause the protection to fail.

[0054] S403, hardware latching logic and physical shutdown. In the analog domain, the system is configured with a high-speed analog comparator. Its non-inverting input is connected to the signal-conditioned real-time current sampling voltage. The inverting input is connected to the dynamic reference voltage output by the aforementioned DAC. When a short circuit or abnormal load occurs, causing a real-time current... Instantly breaking through dynamic boundaries When this happens, the comparator output immediately flips to a high level.

[0055] This high-level signal directly triggers the set (S) pin of the subsequent SR latch. The latch's output (Q) pin is connected to the "output disable" pin of the high-side power switch driver circuit. Once the latch is set, the driver circuit's control is forcibly revoked, and the power switch immediately turns off. This shutdown path is completely independent of the CPU software loop, and its response delay depends only on the sum of the comparator toggle time and the logic gate propagation delay (typically less than 5μs). This ensures that even in extreme cases where the CPU has not yet responded or even deadlock has occurred, the hardware level can still reliably cut off the fault circuit, preventing the wiring harness from overheating and catching fire.

[0056] S404, Fault Interrupt Generation and State Transition. Simultaneously with the activation of the hardware latch, its output signal serves as an external interrupt source, triggering the microcontroller's Interrupt Service Routine (ISR). The system then transitions its operating state from "Normal Monitoring Mode" to "Hardware Latch Mode." Within the interrupt service routine, the protection execution module 400 records context data such as a current snapshot, predicted value, and ambient temperature at the moment of the fault occurrence and stores it in non-volatile memory (NVM) as a basis for subsequent fault diagnosis. At this time, unless an explicit reset command is received from the upper-level control strategy, the latch remains locked, preventing the power device from conducting again and avoiding secondary damage caused by blind retries.

[0057] See attached document Figure 1 In this embodiment, the diagnostic feedback module 500 abandons the "blind retry" strategy that relies solely on software timers in traditional solutions and innovatively introduces an "excitation-response analysis" mechanism with physical safety boundaries. This mechanism injects controlled energy trial pulses into the load circuit and accurately identifies whether the fault type is an irreversible "hard short circuit" or an occasional "soft interference" based on the transient electrical response characteristics of the circuit, thereby achieving a balance between protecting device safety and maintaining system availability.

[0058] S501, active detection pre-cooling and safety pulse generation. To prevent excessive heat buildup in the wiring harness and power devices due to repeated power-on before a fault is completely cleared, the diagnostic feedback module 500 incorporates cooling interlock logic based on a thermal model. When the system enters a hardware latch state due to overcurrent, the module will monitor the wires in real time and estimate the temperature. Only when the temperature drops back to the preset safety baseline (In this embodiment, the temperature is set to below 60°C) The system will only unlock the active detection function when the temperature is below 60°C.

[0059] After meeting thermal safety requirements, the microcontroller generates an extremely short trial voltage pulse. To ensure that even under extreme conditions of a complete short circuit at the load end (i.e., impedance close to zero), the total energy injected into the circuit is insufficient to trigger thermal failure of the power device, the duration of the pulse is... The time constraints must be strictly limited by the safe operating area (SOA) of the power device. The specific time constraints are calculated based on the following energy conservation inequality: ; In the formula, This refers to the single-pulse avalanche energy specified in the power switch datasheet. The current system power supply voltage; The estimated peak short-circuit current for the loop (usually taken as a value) ,in (the minimum parasitic resistance of the loop). This is a safety derating factor (e.g., 0.8). The physical meaning of this formula is to limit the Joule heat injected in a single probe, ensuring that it is below the critical energy for thermal breakdown of the device, thereby making the probe itself inherently safe.

[0060] S502, circuit arbitration logic under the "hardware priority principle". During the transmission of the probe pulse, this embodiment strictly implements the "hardware priority" circuit control strategy. Specifically, the microcontroller's drive signal and the hardware comparator's output signal are connected to the drive circuit through an AND gate or a dedicated "output disable" pin. Although the microcontroller attempts to output a high level to turn on the power switch, the underlying analog comparator has the highest priority "veto power".

[0061] During this phase, the system temporarily adjusts the DAC's reference voltage to the diagnostic threshold. (Typically set to 1.5 to 2 times the nominal load current). If in During the duration, the loop current The price rose sharply due to a load short circuit and exceeded the corresponding level. Upon reaching the limit, the hardware comparator will immediately toggle to a low level, directly blocking the drive signal through the logic circuit and forcibly truncating the microcontroller's probe pulse. This purely hardware-level arbitration mechanism eliminates the risks associated with software response delays, ensuring that the protection logic remains reliable and timely in the event of any software failure or MCU deadlock.

[0062] S503, hard short-circuit fault detection and locking. After the trial pulse ends, the diagnostic feedback module 500 immediately reads the status latch bit of the hardware comparator and the actual duration of the pulse. If the latch is detected to be set again, or the actual duration of the pulse is less than the preset instruction time (i.e., ... The system will determine that there is a "hard short circuit" in the loop.

[0063] The determination indicates the presence of a physical low-impedance path at the load end (such as a short circuit between the wiring harness and ground), and any attempt to apply power will result in overcurrent. Accordingly, the system enters a permanent lockout state, prohibiting subsequent automatic retry operations, and sends a "serious hardware failure" alarm frame to the vehicle's central gateway until the system is completely powered off and restarted or receives a reset command from an external maintenance tool.

[0064] S504, soft interference detection and impedance verification. If the hardware comparator is in the entire... If no trigger is detected during this period, it indicates that the loop current remains within a controllable range. At this time, the diagnostic feedback module 500 acquires the current sampling value at the end of the pulse. With synchronization voltage value By verifying the transient impedance of the load, "soft interference" caused by electromagnetic interference (EMI) or connector momentary disconnection can be eliminated.

[0065] Considering that direct division calculation of impedance under open-circuit or extremely low current conditions can lead to numerical overflow or accuracy divergence, this embodiment employs a voltage domain residual verification method based on product form to enhance the numerical stability of the algorithm: ; In the formula, the symbol This indicates that the absolute value operation is performed to ensure that the residual value can be detected as long as the voltage is too high or too low, as long as the residual value exceeds the threshold. This is the nominal impedance of the load at the reference temperature; The equivalent temperature coefficient of the circuit; This is the current estimated temperature rise. This is the impedance verification tolerance (unit: voltage V), used to cover manufacturing tolerances and aging drift. If the inequality holds, it means that the load electrical characteristics meet Ohm's law expectations, and the previous overcurrent trigger can be attributed to an occasional disturbance; if the inequality does not hold, it indicates that there may be a partial short circuit or load parameter drift, and the system will be marked as "impedance abnormal" and kept in the off state to avoid operation with the fault.

[0066] S505, closed-loop feedback and system reset. Once soft interference is confirmed and impedance verification passes, the diagnostic feedback module 500 executes a system reset strategy, sending an "interference clear" signal to the state machine, releasing the hardware latch state, and restoring normal power output. To prevent potential intermittent contact faults from escalating into fire hazards, the system maintains an "interference counter" within a sliding time window. Only when When the frequency reaches 3 times per minute, automatic reset is allowed; once this safety threshold is exceeded, the system will automatically upgrade the fault level, stop performing the reset operation, and enter the safety degradation mode, thus achieving a dynamic balance from "aggressive diagnosis" to "conservative protection".

[0067] To more intuitively illustrate the practical application effect of this invention in complex vehicle electrical environments, this embodiment selects "power window anti-pinch motor drive in intelligent door controller" as a typical application scenario. It should be noted that although this embodiment uses a power window motor as an example, the application of this invention is not limited to this. This system is also applicable to inductive loads in the vehicle body with similar electrical characteristics, such as seat adjustment motors, wiper motors, and power tailgates.

[0068] In this scenario, the load is an 80W rated power DC brushed motor. Its operation includes starting surges, mechanical commutation ripple, and nonlinear resistance fluctuations caused by aging of the sealing strip. This presents a significant challenge to traditional current protection solutions. (See attached diagram) Figure 3 The figure illustrates the current control strategy of the system of the present invention during the entire process of motor startup and steady-state operation, and provides a direct comparison with the prior art.

[0069] During the system power-on initialization phase, the data sensing module 100 first reads a preset set of physical parameters from the non-volatile memory. In this embodiment, considering the material differences between the connecting harness and the power device, the system does not simply accumulate resistance, but introduces the equivalent temperature coefficient of the hybrid material circuit. This coefficient, through a weighted fitting method, unifies the characterization of the linear temperature drift of copper wires and the nonlinear temperature drift of silicon-based power devices, ensuring that subsequent thermodynamic calculations are performed on a single scalar basis. The accuracy of the following.

[0070] When the driver operates the window lift switch, the data sensing module 100 captures the control intention signal. A step change occurs from 0 to 1. At this time, the motor current rises rapidly, forming a... Figure 3 The surge during startup is shown (between 0ms and 100ms in the figure). In existing conventional technologies, to prevent false triggering during this phase, it is usually necessary to set an extremely high fixed circuit breaker threshold (e.g., ...). Figure 3 The 40A (shown by the thick dashed line) avoided the surge, but left a huge protection blind spot in the subsequent steady-state operation phase.

[0071] In contrast, the computation scheduling module 200 in this embodiment employs a phased hybrid control strategy. Upon detecting an intention step, considering the inference latency of the deep learning model, the system utilizes a high-real-time "sentinel core unit" to directly take over control, temporarily setting the hysteresis voltage window of the hardware comparator to the maximum allowable tolerance threshold. This state continues until the preset holding time. This concludes the process, thus safely "allowing" the startup surge without relying on complex software logic.

[0072] As the motor speed stabilizes, the system smoothly transitions to the steady-state control phase (e.g., ...). Figure 3 (After 100ms). At this point, the temporal convolutional network (TCN) in the state prediction module 300 does not simply smooth the signal like a traditional filter, but plays a core role as a predictor with physical perception capabilities. By learning the frequency domain characteristics of the motor ripple, it can effectively distinguish between normal mechanical resistance fluctuations (such as friction during window lifting) and early inter-turn short-circuit faults, and combine this with the theoretical physical current value calculated by thermodynamic equations. Perform double verification. If the residuals calculated by both methods... At the confidence threshold Within the specified range, the protection execution module 400 will synthesize a dynamic shutdown current threshold that closely reflects the actual load conditions based on the predicted average. (like Figure 3 (As shown by the fine dotted line).

[0073] It is worth noting that this dynamic envelope has a safety margin of only about 3A to 4A higher than the actual load current, far lower than the fixed margin of up to 30A in traditional solutions. This means that the system can sensitively detect even minor overcurrent anomalies. To verify the protection capability of this mechanism under extreme fault conditions, a hard short-circuit simulation experiment was conducted, and the experimental results are attached. Figure 4 As shown.

[0074] In the experiment, a low-impedance short-circuit fault was artificially introduced when the system reached steady state. For example... Figure 4As shown by the solid line, the loop current rises sharply at the moment the fault occurs. Because the dynamic threshold (dotted line) of this invention is close to the normal operating current, the fault current touches the toggling threshold of the hardware comparator the instant it rises. At this moment, the protection execution module 400 immediately triggers the hardware latching logic, cutting off the power output.

[0075] In contrast. Figure 4 The dotted lines in the middle show the tradition The response curve of the algorithm or fuse. Because its threshold (40A, horizontal dashed line in the figure) is set too high, the current must rise above this value and accumulate heat for a period of time before the protection mechanism will activate. The gray shaded area in the figure intuitively reflects the extra Joule heat energy released by the traditional solution compared to this invention. Experimental data shows that the traditional solution has a longer response time, leading to softening of the wire harness insulation layer; while the response time of this embodiment is less than 1ms (as marked in the figure), and the wire harness temperature rise is limited to an extremely low range (e.g., less than 5°C), achieving a near intrinsically safe protection effect.

[0076] After the fault was disconnected, the system did not restart blindly, but instead entered the diagnostic feedback process. The diagnostic feedback module 500 first waited for the conductor temperature to drop back to the safe baseline. Subsequently, the control power device outputs a microsecond-level probe pulse. During this period, the diagnostic feedback module 500 performs real-time verification of the loop impedance using the following inequality: ; The introduction of this inequality cleverly solves the problem of accuracy divergence in the traditional voltage comparison method under low current. If the verification passes, it means that the previous overcurrent was only an occasional electromagnetic interference, and the system will automatically reset and restore the window operation, with almost no noticeable impact on the user; if the verification fails, it is determined to be a hard short circuit, and the system will remain locked and report a fault code.

[0077] In summary, this embodiment, by integrating data-driven prediction algorithms with first-principles physical models, not only resolves the contradiction between the vehicle load during startup transients and steady-state operation, but also improves the response speed of short-circuit protection by three orders of magnitude without increasing hardware costs, thereby enhancing the safety and reliability of the vehicle's electrical system.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fully integrated processing system for vehicle domain control that integrates vehicle power management, characterized in that, include: The data sensing module is configured to collect current signals, voltage signals, and ambient temperature signals of the load circuit, and construct a time-series feature tensor based on the collected signals; The computation scheduling module includes a sentinel core unit and an inference core unit. The sentinel core unit is configured to run resident to perform deterministic logic computation, and the inference core unit is configured to run neural network inference computation. The computation scheduling module is configured to execute hard feedforward control logic during the transient phase. When the data sensing module detects a step change in the control intent signal, the computation scheduling module determines that the system has entered the transient control phase and temporarily bypasses the output of the inference core unit. During the transient control phase, the computation scheduling module forcibly sets the hysteresis voltage window of the hardware comparator to the maximum tolerance threshold and maintains it for a preset hold time. The hold time is determined based on the sum of the average inference time of the inference core unit and the safety margin coefficient to cover the latency blind zone of the model inference. The state prediction module is configured to predict the load current prediction value at the next moment using a temporal convolutional network based on the temporal feature tensor, and to calculate the theoretical physical current value based on the real-time acquired voltage signal and the estimated loop temperature using a physical mechanism model to verify the load current prediction value. The protection execution module is configured to generate a dynamic shutdown current threshold based on the verified load current prediction value, and convert the dynamic shutdown current threshold into an analog reference voltage. The protection execution module includes a hardware comparator configured to compare the real-time acquired current signal with an analog reference voltage, and to trigger hardware latching logic to cut off power output when the amplitude of the current signal exceeds the analog reference voltage.

2. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, Before constructing the temporal feature tensor, the data sensing module is configured to calculate the equivalent temperature drift coefficient of the hybrid material loop. The data sensing module reads a preset set of physical parameters, including the temperature coefficient of resistance of the connecting wire harness and the temperature coefficient of resistance of the power device. The data sensing module performs a weighted fitting of the resistance temperature coefficient of the connecting harness and the resistance of the power device in the nominal impedance of the entire circuit based on the proportion of the resistance of the connecting harness and the resistance temperature coefficient of the power device, and obtains the equivalent temperature drift coefficient. The equivalent temperature drift coefficient is used to characterize the overall impedance change characteristics of the hybrid material circuit under a single ambient temperature variable.

3. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, The computation scheduling module is configured to execute a sparse inference strategy during the steady-state control phase. The sentinel core unit monitors the sum of the absolute values ​​of the first-order differences of the current data in the time-series feature tensor in real time to calculate the fluctuation feature index; The computation scheduling module compares the fluctuation characteristic index with a preset wake-up threshold: if the fluctuation characteristic index is greater than the wake-up threshold, a wake-up interrupt is sent to the inference core unit to trigger a new round of inference computation; if the fluctuation characteristic index is less than or equal to the wake-up threshold, the inference core unit is controlled to remain in a sleep state and maintain the load current prediction value of the previous frame.

4. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, The temporal convolutional network running in the state prediction module contains multiple stacked dilated causal convolutional layers; The temporal convolutional network receives temporal feature tensors as input and uses convolutional kernels with dilation factors that increase exponentially with the number of layers to extract current fluctuation features within a historical time window. After training, the temporal convolutional network undergoes fixed-point quantization to map floating-point weight parameters to fixed-point integers, which are then deployed in the inference kernel unit to establish a nonlinear mapping relationship between input features and future current states.

5. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, The state prediction module is configured to execute residual verification logic based on first principles; The state prediction module runs a thermodynamic iterative equation, estimates the conductor temperature based on the balance between the Joule heat generated by the real-time current and the heat dissipation from the environment, and calculates the theoretical physical current value by combining the equivalent temperature drift coefficient. The state prediction module calculates the Euclidean distance residual between the predicted load current value and the theoretical physical current value; if the Euclidean distance residual is less than the dynamic confidence threshold, the predicted load current value is accepted; if the Euclidean distance residual is greater than or equal to the dynamic confidence threshold, the output is corrected using the theoretical physical current value.

6. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, The protection execution module executes envelope synthesis logic when generating the dynamic shutdown current threshold. The protection execution module calculates the standard deviation statistics of historical current data within the sliding window to quantify current ripple characteristics. The protection execution module multiplies the predicted load current value by a proportional gain coefficient, and adds the standard deviation statistic calculated based on historical current data and a fixed safety tolerance to synthesize the dynamic shutdown current threshold that closely matches the physical limits of the load, so as to cover normal load fluctuations and reserve dynamic margin.

7. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, The protection execution module also includes a digital-to-analog converter and an SR latch; The digital-to-analog converter is connected to the inverting input of the hardware comparator and is used to convert the digital value of the dynamic shutdown current threshold into an analog reference voltage. The output of the hardware comparator is connected to the set pin of the SR latch, and the output of the SR latch is connected to the output disable pin of the power switch driver circuit. When the SR latch is set, the power switch is forcibly disconnected through a pure hardware link.

8. The integrated vehicle domain control system for vehicle power management as described in claim 1, characterized in that, The system also includes a diagnostic feedback module, which is equipped with an active detection function; The diagnostic feedback module monitors the conductor temperature in real time after the hardware latching logic is triggered, and controls the power device to output a trial voltage pulse when the conductor temperature drops below the safety baseline. The duration of the exploratory voltage pulse is limited by the safe operating area of ​​the power device, and the injected energy corresponding to the duration is less than the single-pulse avalanche energy limit of the power device.

9. The integrated vehicle domain control system for vehicle power management as described in claim 8, characterized in that, The diagnostic feedback module is configured to execute impedance verification logic based on voltage domain residuals; During the output of the trial voltage pulse, if the hardware comparator is not triggered, the diagnostic feedback module acquires the sampled voltage and sampled current at the end of the trial voltage pulse; The diagnostic feedback module calculates the residual amplitude between the sampled voltage and the estimated voltage, where the estimated voltage is the product of the sampled current and the temperature-corrected nominal impedance. If the residual amplitude is less than the impedance verification tolerance, the previous overcurrent is determined to be a soft interference and the system is reset. If the residual amplitude is greater than the impedance verification tolerance, the impedance is determined to be abnormal and the system remains locked.

Images