Electromagnetic compatibility integrated protection system for electric driving system of electric vehicle
By using multi-dimensional parameter synchronous sampling and spatiotemporal graph isomorphic prediction models, combined with collaborative anti-disturbance calculation and underlying hardware reconstruction, the problem of transient high-frequency electromagnetic interference in electric vehicle electric drive systems is solved, achieving accurate prediction and efficient suppression of electromagnetic interference, and improving the operational safety and stability of the system.
Patent Information
- Application Number
- CN202610501057.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-16
- Publication Date
- 2026-05-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing electric drive systems for electric vehicles, traditional static filtering strategies or basic feedback active suppression technologies cannot cope with transient high-frequency electromagnetic interference caused by wide-bandgap semiconductor devices in the inverter, leading to drastic changes in the electromagnetic interference spectrum and causing disorder in the vehicle's power control.
A multi-dimensional parameter synchronous sampling module is used to obtain the operating condition characteristic sequence of the electric drive system. Combined with the topological parasitic parameter matrix, the interference frequency domain feature vector is predicted by the spatiotemporal graph isomorphic prediction model. The collaborative anti-interference solution module calculates the carrier band spread spectrum command and the active filter reverse compensation injection current. The underlying hardware reconfiguration execution module updates the pulse width modulation generator and active filter circuit in real time.
It achieves forward-looking prediction and multi-dimensional suppression of electromagnetic interference, improves the suppression efficiency of electromagnetic interference, and ensures the operational safety and stability of the electric drive system.
Smart Images

Figure CN122052656A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electromagnetic compatibility protection technology, and more specifically to an integrated electromagnetic compatibility protection system for an electric vehicle electric drive system. Background Technology
[0002] In the technological evolution of new energy electric vehicles, the electric drive system, as the core power output hub of the vehicle, directly determines the dynamic response and energy conversion efficiency of the entire vehicle. In the pursuit of high power density and low switching losses, related technologies have generally introduced wide bandgap semiconductor devices such as silicon carbide or gallium nitride into the inverter. These wide bandgap devices can complete the switching of large currents at the nanosecond level, and their switching frequency is often pushed to the range of tens of kilohertz or even hundreds of kilohertz, so as to reduce the size of passive filter components and increase system bandwidth. Under this specific application architecture, the electric drive system is filled with high-speed jumping voltage and current pulse signals.
[0003] However, with the rapid increase in switching speed, the high-frequency switching edges upon which the relevant technical solutions rely for executing basic pulse width modulation and energy conversion logic interact strongly with various parasitic parameters inevitably present within the electric drive system. Specifically, the stray inductance within the inverter power module, the parasitic capacitance between the power board and the heat sink, and the distributed capacitance between the motor stator windings and ground constitute an extremely complex parasitic network. High-frequency edge signals are conducted and radiated through these parasitic networks, generating highly destructive common-mode current and differential-mode voltage noise. Furthermore, this high-frequency electromagnetic interference energy spreads through the DC bus to the vehicle's low-voltage network. Based on the above operating mechanism, the relevant technologies have a very hidden underlying pain point in handling electromagnetic compatibility protection: traditional static filtering strategies or basic feedback active suppression technologies... Unable to handle nonlinear parasitic parameter drift under transient conditions; when a vehicle performs rapid acceleration, climbs hills, or experiences drastic changes in ambient temperature, the temperature and electromagnetic fields inside the electric drive system undergo severe distortion, directly causing dynamic drift in the values of parasitic capacitance and inductance. This dynamic drift causes the spectral characteristics of electromagnetic interference to change drastically within milliseconds. Related technologies only perform fixed frequency compensation or fixed cutoff frequency filtering based on static parameters calibrated at the factory. When faced with such transient spectral changes, they not only fail to effectively attenuate interference energy, but are also prone to triggering high-frequency resonance due to phase mismatch in compensation, leading to a sharp amplification of local electromagnetic energy. This transient high-frequency resonance not only penetrates the insulation layer and shortens the motor's lifespan, but also causes serious logic errors and frame loss in adjacent low-voltage vehicle networks, ultimately leading to serious consequences such as disordered vehicle power control. Summary of the Invention
[0004] To address the aforementioned technical problems, an integrated electromagnetic compatibility protection system for electric vehicle drive systems is provided, which solves the problems described above.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An integrated electromagnetic compatibility protection system for an electric vehicle's electric drive system, comprising: The multi-dimensional parameter synchronous sampling module is used to acquire the operating condition feature sequence of the target electric drive system within a preset time window, and synchronously extract the topological parasitic parameter matrix of the target electric drive system under the current physical environment. The spatiotemporal graph isomorphic prediction module is used to concatenate the operating condition feature sequence with the topological parasitic parameter matrix, input it into the pre-trained spatiotemporal graph isomorphic prediction model, and output the predicted interference frequency domain feature vector corresponding to the current control cycle. The collaborative anti-interference calculation module is used to calculate the carrier band spreading command for the underlying microcontroller and the reverse compensation injection current for the active filter circuit based on the peak frequency coordinates and common mode energy density in the predicted interference frequency domain feature vector. The underlying hardware reconfiguration execution module is used to update the carrier frequency register of the pulse width modulation generator in the target electric drive system according to the carrier frequency band spread spectrum instruction, and synchronously control the active filter circuit to output the reverse compensation injection current.
[0006] Optionally, the multi-dimensional parameter synchronous sampling module, when extracting the topological parasitic parameter matrix of the target electric drive system under the current physical environment, specifically includes: Collect the substrate junction temperature data of the power semiconductor module in the target electric drive system, as well as the real-time current carrying capacity of the DC bus. The substrate junction temperature data is compared with the preset semiconductor parasitic capacitance temperature drift curve by table lookup interpolation to determine the first dynamic parasitic capacitance value. The real-time current carrying capacity is input into a preset skin effect impedance evaluation model to calculate the second dynamic parasitic inductance value of the DC bus in different high frequency bands. The topological parasitic parameter matrix is constructed based on the first dynamic parasitic capacitance value, the second dynamic parasitic inductance value, and the pre-stored high-frequency parasitic parameters of the motor stator winding to ground.
[0007] Optionally, the spatiotemporal graph isomorphic prediction model includes a graph convolutional spatial feature extraction network and a temporal convolutional temporal feature extraction network. When the spatiotemporal graph isomorphic prediction module outputs the predicted interference frequency domain feature vector corresponding to the current control period, it specifically includes: Each physical power device in the target electric drive system is mapped to a graph node in the graph convolutional spatial feature extraction network, and the parasitic impedance value in the topological parasitic parameter matrix is used as the weight of the adjacent edge between the graph nodes to construct a physical topological isomorphic graph. The operating condition feature sequence is mapped to the corresponding graph node as the node's initial feature tensor; The graph convolutional spatial feature extraction network is used to perform multi-hop graph convolutional aggregation on the initial feature tensor of the node to obtain a spatial feature matrix that captures spatial electromagnetic coupling relationships. The spatial feature matrix is input into the temporal convolutional temporal feature extraction network, and transient high-frequency noise features that evolve over time are extracted through causal dilated convolution kernels. After fully connected mapping, the predicted interference frequency domain feature vector is output.
[0008] Optionally, the system also includes an offline pre-training module, which is configured to perform the following loss function calculation logic during the pre-training phase of the spatiotemporal graph isomorphic prediction model: Obtain the sample predicted interference frequency domain feature vector and the corresponding real interference frequency domain label vector; The mean square error between the sample predicted interference frequency domain feature vector and the real interference frequency domain label vector is calculated and used as the basic reconstruction loss. Calculate the energy leakage penalty term of the sample predicted interference frequency domain feature vector at the Nyquist folding frequency; Calculate the theoretical common-mode current constraint boundary of Kirchhoff's current law based on the topological parasitic parameter matrix, determine whether the sample predicted interference frequency domain feature vector exceeds the theoretical common-mode current constraint boundary, and if it does, calculate the physical boundary over-limit penalty term. The basic reconstruction loss, the energy leakage penalty term, and the physical boundary violation penalty term are weighted and summed to obtain the target loss function. The model network weights are then updated based on the target loss function using the backpropagation algorithm.
[0009] Optionally, when the cooperative anti-interference calculation module calculates the carrier band spreading command for the underlying microcontroller, it specifically includes: Extract all feature frequency nodes whose energy density exceeds a preset interference threshold from the predicted interference frequency domain feature vector to determine the target discrete interference spectrum; Locate the peak frequency coordinates corresponding to the maximum common-mode energy density from the target discrete interference spectral lines; The spread spectrum modulation bandwidth for the peak frequency coordinate is calculated using the following formula: the spread spectrum bandwidth is equal to the energy density value corresponding to the peak frequency coordinate divided by a preset frequency band smoothing attenuation coefficient. A pseudo-random sequence is generated based on the spread spectrum modulation broadband range, and the pseudo-random sequence is mapped to the dynamic frequency adjustment period of the pulse width modulation signal to generate the carrier band spread spectrum command.
[0010] Optionally, when the collaborative disturbance rejection calculation module calculates the reverse compensation injection current for the active filter circuit, it specifically includes: Perform an inverse discrete Fourier transform on the predicted interference frequency domain feature vector to restore the time-domain transient common-mode current waveform; Obtain the amplitude-frequency transmission attenuation coefficient and phase delay angle of the high-frequency injection transformer in the active filter circuit; The time-domain transient common-mode current waveform is inverted to obtain an inverted reference current; The amplitude of the inverting reference current is pre-amplified and compensated based on the amplitude-frequency transmission attenuation coefficient, and the phase of the inverting reference current is over-adjusted and compensated based on the phase delay angle to obtain the reverse compensation injection current.
[0011] Optionally, it also includes an extreme condition safety interception submodule, specifically including: Real-time monitoring of the phase current change rate and DC bus voltage fluctuation rate of the target electric drive system; When it is determined that the phase current change rate is greater than the preset transient change threshold and the DC bus voltage fluctuation rate exceeds the voltage safety envelope, a hardware-level anti-push-through interrupt is triggered. In response to the hardware-level anti-penetration interruption, the data bus transmission from the multi-dimensional parameter synchronous sampling module to the spatiotemporal graph isomorphic prediction module is suspended through the bus isolation device; The maximum derating modulation command is directly sent to the pulse width modulation generator based on the preset limit safety parameter library.
[0012] Optionally, when the extreme condition safety interception submodule issues the maximum derating modulation command, it specifically includes: Extract the preset dynamic dead time compensation amount and switching frequency clamping value from the extreme safety parameter library; The carrier frequency of the pulse width modulation generator is forcibly locked to the switching frequency clamp value, terminating all spread spectrum adjustments. In the dead-time control register of the pulse width modulation generator, the initial dead time is superimposed with the dynamic dead-time compensation amount; The actual dead time after superposition is used to physically isolate and extend the drive signals of the upper and lower bridge arms in order to forcibly suppress the common-mode voltage spikes caused by the maximum parasitic oscillation.
[0013] Optionally, it also includes a physical layer closed-loop feedback fine-tuning submodule, specifically including: The actual output conducted interference spectrum was collected by a high-frequency current probe placed at the front end of the high-voltage DC bus of the target electric drive system. Compare the residual sequence between the actual output conducted interference spectrum and the predicted interference frequency domain feature vector; When the energy error of a specific frequency band in the residual sequence is greater than the dynamic tolerance threshold, the operating condition features corresponding to that specific frequency band are extracted. Using the operating condition features and the actual output conducted interference spectrum as incremental learning samples, the network weights in the spatiotemporal graph convolutional spatial features of the spatiotemporal graph isomorphic prediction model are extracted and fine-tuned online using an exponential moving average strategy.
[0014] Optionally, when the multi-dimensional parameter synchronous sampling module acquires the operating condition characteristic sequence of the target electric drive system within a preset time window, it specifically includes: According to the preset oversampling frequency, the three-phase stator current, rotor electromechanical angle, direct torque control command and modulation ratio parameters of the target electric drive system are collected respectively. The three-phase stator currents are subjected to Clark and Park coordinate transformations to obtain the d-axis and q-axis current components. The d-axis and q-axis current components, the rotor electromechanical angle, the direct torque control command, and the modulation ratio parameter are spliced together according to the time dimension to form a multidimensional tensor; The multidimensional tensor is normalized to eliminate the magnitude difference between different physical dimensions, thereby obtaining the operating condition feature sequence.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses a multi-dimensional parameter synchronous sampling module to collect the operating condition characteristic sequence of the electric drive system within the same preset time window. Combined with the junction temperature of the power semiconductor substrate and the DC bus current carrying capacity, it dynamically calculates the real-time values of parasitic capacitance and parasitic inductance. It also integrates the inherent high-frequency parasitic parameters of the motor stator winding to construct a topological parasitic parameter matrix. This completely solves the problem of treating parasitic parameters as fixed values and ignoring their dynamic drift in the prior art. At the same time, it ensures the sampling synchronization of operating condition characteristics and parasitic parameters, eliminates timing deviations, and provides a high-precision data foundation that closely matches the actual operating state of the electric drive system for subsequent interference prediction.
[0016] 2. The spatiotemporal graph isomorphic prediction model of this invention integrates a graph convolutional spatial feature extraction network and a temporal convolutional temporal feature extraction network. It maps the physical power devices of the electric drive system as graph nodes and constructs a physical topological isomorphic graph with parasitic impedance as edge weights to capture spatial electromagnetic coupling relationships. At the same time, it extracts the temporal transient features of interference through causal dilated convolution kernels, realizing spatiotemporal dual-dimensional feature analysis of electromagnetic interference. Combined with an offline pre-training mechanism with multiple loss function constraints, the model prediction results take into account numerical accuracy, energy leakage suppression, and physical law constraints, which greatly improves the accuracy of predicting the frequency domain feature vector of interference and realizes the forward prediction of electromagnetic interference, providing a precise basis for the calculation of subsequent anti-interference measures.
[0017] 3. The collaborative anti-interference calculation module of the present invention calculates the carrier band spreading command and the active filter reverse compensation injection current based on the peak frequency coordinates and common-mode energy density of the predicted interference. The carrier spreading is used to smooth the spectrum of the interference peak frequency band, and the active filter outputs a reverse current with amplitude and phase compensation to cancel the common-mode interference. The two work together and each has its own focus, which breaks through the limitations of the single anti-interference method in the prior art. It realizes multi-dimensional interference suppression from the signal modulation layer and the hardware compensation layer, which greatly improves the suppression efficiency of electromagnetic interference and effectively covers the main interference frequency bands of the electric drive system.
[0018] 4. The underlying hardware reconfiguration execution module of this invention can update the carrier frequency register of the pulse width modulation generator in real time and synchronously control the output compensation current of the active filter circuit, realizing a hardware-level fast response to the anti-interference command and eliminating the delay of software adjustment; at the same time, an extreme condition safety interception submodule is added, which triggers a hardware-level anti-penetration interrupt by monitoring the phase current change rate and DC bus voltage fluctuation rate in real time, suspending data transmission and issuing a maximum derating modulation command, forcibly locking the switching frequency and extending the dead time, effectively suppressing common-mode voltage spikes under extreme conditions, avoiding power device breakdown and control module failure, and greatly improving the operational safety of the protection system and electric drive system. Attached Figure Description
[0019] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart illustrating the steps of the present invention. Detailed Implementation
[0020] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0021] Reference Figure 1-2 As shown, an integrated electromagnetic compatibility protection system for an electric vehicle's electric drive system includes: The multi-dimensional parameter synchronous sampling module is used to acquire the operating condition feature sequence of the target electric drive system within a preset time window, and synchronously extract the topological parasitic parameter matrix of the target electric drive system under the current physical environment. The spatiotemporal graph isomorphic prediction module is used to concatenate the operating condition feature sequence with the topological parasitic parameter matrix, input it into the pre-trained spatiotemporal graph isomorphic prediction model, and output the predicted interference frequency domain feature vector corresponding to the current control cycle. The collaborative anti-interference calculation module is used to calculate the carrier band spreading command for the underlying microcontroller and the reverse compensation injection current for the active filter circuit based on the peak frequency coordinates and common mode energy density in the predicted interference frequency domain feature vector. The underlying hardware reconfiguration execution module is used to update the carrier frequency register of the pulse width modulation generator in the target electric drive system according to the carrier frequency band spread spectrum instruction, and synchronously control the active filter circuit to output the reverse compensation injection current.
[0022] In some embodiments, the multi-dimensional parameter synchronous sampling module acquires the operating condition feature sequence of the target electric drive system within a preset time window, and synchronously extracts the topological parasitic parameter matrix of the target electric drive system under the current physical environment; the operating condition feature sequence is not a single scalar set, but a multi-dimensional tensor characterizing the full state of motor control; the built-in analog-to-digital converter is configured in direct memory access mode, and continuously reads the three-phase stator current, rotor mechanical absolute position, bus transient voltage, and dq axis command voltage output by the control loop in the hardware interrupt service subroutine of each PWM carrier cycle; the preset time window is set to a sliding time window containing at least 64 complete PWM cycles; the above data are concatenated and spliced in the time domain into a two-dimensional floating-point feature tensor with a shape equal to the window length and the number of physical features.
[0023] Specifically, extracting the topology parasitic parameter matrix is a core prerequisite that distinguishes this system from the conventional static filtering architecture. The matrix is an NxN complex matrix representing the impedance relationship between nodes in the network topology, where N represents the number of key physical nodes selected in the inverter hardware topology. The core element of this matrix is the dynamic complex impedance between nodes, and its calculation formula is as follows:
[0024] In the formula, The first parasitic parameter in the topological parameter matrix The node and the first Dynamic complex impedance between nodes The angular frequency of the electromagnetic interference signal. for , Dynamic parasitic inductance between nodes, for , The dynamic parasitic capacitance between nodes, the elements in this matrix not only include the dynamic equivalent series inductance of the DC bus due to the high-frequency skin effect, but also mainly include the nonlinear junction capacitance of power devices at different junction temperatures. The synchronous extraction action means that the memory generation timestamp of the parasitic parameter matrix must be strictly aligned with the acquisition timestamp of the operating condition feature sequence to ensure the absolute physical consistency between the spatial state and the temporal state.
[0025] Specifically, in the spatiotemporal graph isomorphic prediction module, the operating condition feature sequence and the topological parasitic parameter matrix are concatenated and input into the pre-trained spatiotemporal graph isomorphic prediction model. The output is the predicted interference frequency domain feature vector corresponding to the current control cycle. When the prediction model performs tensor inference, its underlying calculation calls the underlying operator library to execute tens of thousands of floating-point multiply-accumulate instructions. The output predicted interference frequency domain feature vector is a high-dimensional one-dimensional tensor with a length strictly defined by the preset standard EMI frequency band from 150kHz to 30MHz and a resolution of 10kHz. Each discrete element value in this vector represents the physical prediction value of the common-mode interference energy density that the underlying electric drive system will excite outward at the corresponding frequency point in the future control cycle.
[0026] Specifically, in the collaborative anti-interference solution module, based on the peak frequency coordinates and common-mode energy density in the predicted interference frequency domain feature vector, the carrier band spreading command for the underlying microcontroller and the reverse compensation injection current for the active filter circuit are calculated respectively. During the calculation process, the extreme value search addressing of the high-dimensional feature vector is performed. The peak frequency coordinates physically represent the frequency point at which the system will experience the strongest parasitic electromagnetic resonance in the future, while the common-mode energy density corresponds to the destructive force level of electromagnetic penetration. The carrier band spreading command for the microcontroller is essentially a set of hexadecimal parameter pairs containing the frequency jitter step size and jitter period. If a strong interference peak is predicted at 2.5MHz, the command will require the DSP peripheral to break the default fixed 10kHz PWM switching frequency and make it jump between 9.5kHz and 10.5kHz according to a specific pattern. This active spectrum shifting smooths out the concentrated electromagnetic radio frequency energy at the source. At the same time, the reverse compensation injection current is solved by amplitude and phase joint calculation based on the common-mode energy density to calculate theoretical physical compensation signals of equal magnitude and opposite direction.
[0027] Specifically, in the underlying hardware reconfiguration execution module, the carrier frequency register of the pulse width modulation generator in the target electric drive system is updated according to the carrier band spread spectrum instruction, and the active filter circuit is synchronously controlled to output the reverse compensation injection current; this is the final step to achieve a substantial leap in the system hardware state; through the microcontroller's memory addressing bus, a new 16-bit unsigned integer value is directly written to the period control register of the PWM module; this underlying write operation is subject to strict hardware time base event timing constraints to prevent unpredictable duty cycle truncation from inducing semiconductor module failure; synchronous control of the active filter circuit converts the reverse compensation injection current into a discrete voltage waveform of a high-frequency digital-to-analog converter, drives an external high-speed linear power amplifier, and directly injects the reverse current into the high-voltage DC bus through a high-frequency broadband transformer.
[0028] In some embodiments, the dynamic extraction logic for the topological parasitic parameter matrix of the multi-dimensional parameter synchronous sampling module can be expanded through an extremely fine physical data mapping mechanism; the substrate junction temperature data of the power semiconductor module and the real-time current carrying capacity of the DC bus are collected; the acquisition of the substrate junction temperature data does not rely on the coarse external heat sink temperature sensor with great thermal inertia, but directly reads the resistance value of the NTC thermistor integrated inside the power module, uses high-speed sampling to capture the on-state voltage drop of the device and substitutes it into the underlying semiconductor thermophysical equation to calculate the transient junction temperature; the bus current carrying capacity is picked up non-destructively by a high-frequency non-inductive Hall sensor.
[0029] Specifically, the substrate junction temperature data is compared with a preset semiconductor parasitic capacitance temperature drift curve by a lookup table to determine the first dynamic parasitic capacitance value. The system stores a multi-dimensional lookup table obtained through calibration on a large experimental bench. Because the width of the depletion layer inside the semiconductor device will undergo nonlinear physical displacement with absolute temperature, it directly changes the size of the inter-electrode parasitic capacitance. The interpolation comparison process not only involves linear search in one-dimensional space, but also includes quadratic calculus compensation correction for junction temperature fluctuation, so that the accuracy of the capacitance value is reduced to the picofarad level.
[0030] Specifically, the real-time current carrying capacity is input into a preset skin effect impedance evaluation model to calculate the second dynamic parasitic inductance value of the DC busbar under different high-frequency bands. Due to the skin effect and proximity effect of high-frequency AC signals, when the transient current changes drastically, the physical current density distribution inside the busbar will be squeezed violently from the center of the cross-section to the outer surface, resulting in a sharp reduction in the equivalent conductive cross-sectional area and a significant decrease in inductance. The evaluation model is constructed based on the analytical solution of Maxwell's electromagnetic field equations under specific rectangular conductor boundary conditions, and outputs the dynamic inductance value in one-dimensional vector form in real time according to the current spectrum components.
[0031] Specifically, the topological parasitic parameter matrix is constructed based on the first dynamic parasitic capacitance value, the second dynamic parasitic inductance value, and the pre-stored high-frequency parasitic parameters of the motor stator winding to ground. Discrete time-varying physical quantities are filled into the two-dimensional matrix address allocated in memory, strictly following the physical node connection relationship of the three-phase inverter circuit schematic. The off-diagonal elements of the matrix objectively characterize the parasitic coupling path strength between different spatial nodes. This construction method realizes a seamless and objective mapping from external physical parameters to internal abstract mathematical matrices.
[0032] In some embodiments, the internal network architecture of the spatiotemporal graph isomorphic prediction module is discussed, which fully demonstrates the deep isomorphism between spatial topology and algorithmic topology. Each physical power device in the target electric drive system is mapped to a graph node in the graph convolutional spatial feature extraction network, and the parasitic impedance value in the topological parasitic parameter matrix is used as the adjacency edge weight between graph nodes to construct a physical topological isomorphic graph. In the physical entity of the three-phase full-bridge inverter, there are 6 main switch nodes and DC bus positive and negative capacitors and other bus nodes. The system defines them as a vertex set V in the memory graph theoretical domain. The topological parasitic parameter matrix is directly instantiated as the adjacency matrix A of the graph. This structure fully inherits the parasitic distribution law of real hardware at the mathematical level, and the electromagnetic crosstalk paths between nodes are objectively quantified by the edge weights.
[0033] The operating condition feature sequence is mapped to the corresponding graph node as the node's initial feature tensor; for example, in the graph node address space corresponding to the U-phase upper bridge arm switch, its initial feature tensor is loaded with the device's current conduction pulse width count, bus voltage division value, and transient phase current value.
[0034] Specifically, a graph convolutional spatial feature extraction network is used to perform multi-hop graph convolution aggregation on the initial feature tensors of nodes to obtain a spatial feature matrix that captures spatial electromagnetic coupling relationships. The graph convolution operation uses a matrix multiplication acceleration library to perform degree normalization and Laplace smoothing on the adjacency matrix. In a single layer of computation, a single node not only retains its own electrical characteristics but also absorbs the electromagnetic radiation characteristics of its surrounding first-order physical neighbor nodes according to the parasitic impedance weight matrix. Through the deep multi-hop mechanism, the network can capture complex global spatial parasitic coupling effects such as bus oscillation induced by high-frequency turn-off action of phase W, and then induce common-mode spikes to phase U through DC bus inductance. The output feature matrix dimension strictly evolves into the total number of graph nodes and the number of hidden layer feature channels.
[0035] Specifically, the spatial feature matrix is input into a temporal convolutional temporal feature extraction network. A causal dilated convolution kernel is used to extract transient high-frequency noise features that evolve over time. After a fully connected layer, the predicted interference frequency domain feature vector is output. Since the spatial matrix is only a transient static slice, the temporal convolutional network uses a causal dilated convolution operator to ensure that the sliding window feature extraction strictly follows the laws of physical time elapsed, preventing data leakage at future moments. The exponential increase in the dilation factor allows the receptive field to cover a time series at the microsecond level, keenly capturing high-frequency transient noise clusters caused by the Miller plateau effect accumulated in silicon carbide devices. The final fully connected layer completes the dimensionality reduction mapping from high-dimensional features to a frequency domain energy probability distribution.
[0036] This system internal architecture modification forcibly rewrites the purely theoretical neural network structure into a customized structure adapted to the inverter's electrical topology. It is a fundamental technical modification aimed at addressing the obstacles of cross-domain application scenarios.
[0037] In some embodiments, for an architecture that includes an offline pre-training module for the model, the underlying logic for constructing its loss function is explained; the sample prediction interference frequency domain feature vector and the corresponding real interference frequency domain label vector are obtained; the real label is derived from the physical spectrum extracted by sweeping the frequency using a Rogowski coil and a high-frequency EMI receiver on a noise reduction laboratory bench.
[0038] The mean square error between the predicted interference frequency domain feature vector and the actual interference frequency domain label vector is calculated as the basic reconstruction loss; the energy deviation at each frequency point coordinate in the high-dimensional tensor is squared and summed to evaluate the overall spectral approximation.
[0039] The energy leakage penalty term of the sample predicted interference frequency domain feature vector at the Nyquist folding frequency is calculated. Due to the limitation of the ADC discrete sampling theorem, digital signal aliasing is very likely to occur when the predicted high-frequency interference component approaches half of the system sampling frequency. This penalty term is used to separate the energy of the folding frequency point through the high-pass filter operator and is given a high weight to suppress it.
[0040] The theoretical common-mode current constraint boundary of Kirchhoff's current law is calculated based on the topological parasitic parameter matrix. It is then determined whether the sample predicted interference frequency domain feature vector exceeds the theoretical common-mode current constraint boundary. If it does, a physical boundary violation penalty term is calculated. The common-mode current is essentially constrained by the physical boundary of the total parasitic impedance to ground. The system calculates this extreme envelope through a complex plane impedance network. Any algorithm prediction value that exceeds this pure physical envelope is considered illusory data and must be subject to an extreme gradient truncation penalty.
[0041] The basic reconstruction loss, energy leakage penalty, and physical boundary violation penalty are weighted and summed, and the model network weights are updated through backpropagation based on the target loss function. This customized loss function forcibly binds the mean square error of pure mathematics to Kirchhoff's physical laws in the field of electricity, ensuring that the direction of the model gradient descent is always bound within a reasonable range of the real physical world.
[0042] In some embodiments, the microscopic calculation mechanism of the spread spectrum command in the cooperative anti-interference solution module is expanded; all characteristic frequency nodes whose energy density exceeds the preset interference threshold in the predicted interference frequency domain feature vector are extracted to determine the target discrete interference spectrum; the preset interference threshold is set below the safety line of the system low-voltage control network anti-interference specification, reserving a specific margin band.
[0043] The peak frequency coordinates corresponding to the maximum common-mode energy density are located from the target discrete interference spectrum, and the core pain point frequency band with the most concentrated energy is locked using a global peak finding algorithm.
[0044] The spread spectrum modulation bandwidth for the peak frequency coordinate is calculated using a formula that satisfies the following: the spread spectrum bandwidth is equal to the energy density value corresponding to the peak frequency coordinate divided by a preset frequency band smoothing attenuation coefficient. This adaptive bandwidth calculation mechanism ensures that the more concentrated the energy is in a frequency band, the wider the range of jitter diffusion, so as to ensure that the radiated energy per unit frequency is uniformly diluted.
[0045] A pseudo-random sequence is generated based on the wide bandwidth of spread spectrum modulation. The pseudo-random sequence is mapped to the dynamic frequency adjustment period of the pulse width modulation signal to generate the carrier band spread spectrum command. The pseudo-random sequence is generated based on a linear feedback shift register to ensure the non-periodic characteristics of the jitter period and prevent the generation of second harmonics.
[0046] In some embodiments, the depth of the inverse compensation calculation in the collaborative anti-interference solution module is drilled down; the predicted interference frequency domain feature vector is subjected to discrete Fourier inverse transform to restore the time domain transient common-mode current waveform; this operation calls the radix-2 time decimation algorithm in the DSP internal coprocessor to realize the lossless reconstruction of the frequency domain tensor to the time domain continuous time step sequence.
[0047] The amplitude-frequency transmission attenuation coefficient and phase delay angle of the high-frequency injection transformer in the active filter circuit are obtained. Due to the leakage inductance and excitation current characteristics of the physical high-frequency transformer, it will inevitably produce physical attenuation and phase lag when transmitting signals in the megahertz band. The system obtains these two key correction parameters by pre-reading the factory sweep frequency characteristic curve table of the transformer.
[0048] The time-domain transient common-mode current waveform is inverted to obtain an inverted reference current with a phase difference of 180 degrees.
[0049] The amplitude of the inverting reference current is pre-amplified and compensated based on the amplitude-frequency transmission attenuation coefficient, and the phase of the inverting reference current is compensated for by advance overshoot based on the phase delay angle, thus obtaining the reverse compensation injection current. This correction logic, which has great engineering depth, overcomes the fatal engineering obstacle in related technologies where the phase difference of less than 180 degrees caused by the physical delay of the injection circuit itself leads to the superposition of resonance in the same direction.
[0050] In some embodiments, the extreme condition safety interception subsystem constitutes an indispensable boundary adaptation defense line; it monitors the phase current change rate and DC bus voltage fluctuation rate of the target electric drive system in real time.
[0051] When it is determined that the phase current change rate is greater than the preset transient change threshold, such as when a wheel locks up during emergency braking or the drive shaft breaks, and the DC bus voltage fluctuation rate exceeds the voltage safety envelope, a hardware-level anti-through-circuit interrupt is triggered. This interrupt event bypasses all software operating system scheduling layers and is directly responded to by the DSP kernel interrupt controller with the highest priority.
[0052] In response to hardware-level anti-penetration interruption, the data bus transmission from the multi-dimensional parameter synchronous sampling module to the spatiotemporal graph isomorphic prediction module is suspended through a bus isolation device. In extreme electromagnetic environments, the data bus returns only garbled text and noise. If the input continues to the model, it will inevitably lead to solution collapse. Therefore, physical layer bus isolation is implemented.
[0053] Based directly on the preset limit safety parameter library, a maximum derating modulation command is issued to the pulse width modulation generator; the preset dynamic dead time compensation amount and switching frequency clamping value are extracted from the limit safety parameter library; the carrier frequency is forcibly locked to the switching frequency clamping value, terminating all spread spectrum adjustment behavior; in the dead time control register, the initial dead time and the dynamic dead time compensation amount are superimposed; the actual dead time after superposition is used to physically isolate and extend the drive signals of the upper and lower bridge arms, forcibly increasing the interlocking time of the upper and lower tubes, so as to suppress the bridge arm shoot-through short circuit disaster that may be caused by the maximum parasitic oscillation; this dedicated anomaly interception mechanism for extreme boundary conditions resolves the reliability challenges in complex automotive-grade scenarios.
[0054] In some embodiments, to address aging drift caused by long-term service, the physical layer closed-loop feedback fine-tuning subsystem performs the following timing: acquiring the actual output conducted interference spectrum through a high-frequency through-hole current probe arranged at the front end of the high-voltage DC bus of the target electric drive system; the probe intercepts the residual interference signal flowing to the battery pack in real time through extremely small coupling inductance.
[0055] Compare the residual sequence between the actual output conducted interference spectrum and the previously predicted interference frequency domain feature vector. When the energy error of a specific high-frequency band in the residual sequence exceeds the dynamic tolerance threshold, extract the operating condition features corresponding to the timestamp of that specific frequency band.
[0056] Using the operating condition characteristics and the actual output conducted interference spectrum as incremental learning samples, the network weights in the spatiotemporal graph convolutional spatial features of the spatiotemporal graph isomorphic prediction model are extracted and fine-tuned online during idle computing power cycles through an exponential moving average strategy. This self-driven evolution based on physical residuals maintains the ultra-high prediction accuracy throughout the entire life cycle of the system.
[0057] In some embodiments, regarding the micro-control flow of seamless register switching in the underlying hardware reconfiguration execution module: after receiving the spread spectrum instruction, read the current count value and counting direction of the time base counter in the pulse width modulation generator; determine whether the counting direction is at a valley update point or a peak update point.
[0058] If it is not at a valley or peak update point, it indicates that it is currently in the middle of the PWM pulse output. If the period value is forcibly rewritten at this time, it is very easy to cut off the IGBT / SiC drive pulse that is being conducted, resulting in local current stress overload. Therefore, the carrier frequency band spread spectrum instruction is temporarily stored in the shadow register.
[0059] Until the time base counter reaches the valley or peak update point, the DSP's underlying hardware automatically triggers a hardware loading event, instantly and automatically writing the spread spectrum instruction in the shadow register into the active effective frequency register, completing the seamless switching of the carrier frequency in the next clock cycle. This meticulous timing control logic, down to the transistor level, objectively ensures that the system can maintain absolute smoothness and stability of the core power control signal even when implementing extremely aggressive high-frequency jitter.
[0060] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. An integrated electromagnetic compatibility protection system for an electric vehicle's electric drive system, characterized in that, include: The multi-dimensional parameter synchronous sampling module is used to acquire the operating condition feature sequence of the target electric drive system within a preset time window, and synchronously extract the topological parasitic parameter matrix of the target electric drive system under the current physical environment. The spatiotemporal graph isomorphic prediction module is used to concatenate the operating condition feature sequence with the topological parasitic parameter matrix, input it into the pre-trained spatiotemporal graph isomorphic prediction model, and output the predicted interference frequency domain feature vector corresponding to the current control cycle. The collaborative anti-interference calculation module is used to calculate the carrier band spreading command for the underlying microcontroller and the reverse compensation injection current for the active filter circuit based on the peak frequency coordinates and common mode energy density in the predicted interference frequency domain feature vector. The underlying hardware reconfiguration execution module is used to update the carrier frequency register of the pulse width modulation generator in the target electric drive system according to the carrier frequency band spread spectrum instruction, and synchronously control the active filter circuit to output the reverse compensation injection current.
2. The electromagnetic compatibility integrated protection system for an electric vehicle drive system according to claim 1, characterized in that, The multi-dimensional parameter synchronous sampling module, when extracting the topological parasitic parameter matrix of the target electric drive system under the current physical environment, specifically includes: Collect the substrate junction temperature data of the power semiconductor module in the target electric drive system, as well as the real-time current carrying capacity of the DC bus. The substrate junction temperature data is compared with the preset semiconductor parasitic capacitance temperature drift curve by table lookup interpolation to determine the first dynamic parasitic capacitance value. The real-time current carrying capacity is input into a preset skin effect impedance evaluation model to calculate the second dynamic parasitic inductance value of the DC bus in different high frequency bands. The topological parasitic parameter matrix is constructed based on the first dynamic parasitic capacitance value, the second dynamic parasitic inductance value, and the pre-stored high-frequency parasitic parameters of the motor stator winding to ground.
3. The electromagnetic compatibility integrated protection system for an electric vehicle drive system according to claim 1, characterized in that, The spatiotemporal graph isomorphic prediction model includes a graph convolutional spatial feature extraction network and a temporal convolutional temporal feature extraction network. When the spatiotemporal graph isomorphic prediction module outputs the predicted interference frequency domain feature vector corresponding to the current control cycle, it specifically includes: Each physical power device in the target electric drive system is mapped to a graph node in the graph convolutional spatial feature extraction network, and the parasitic impedance value in the topological parasitic parameter matrix is used as the weight of the adjacent edge between the graph nodes to construct a physical topological isomorphic graph. The operating condition feature sequence is mapped to the corresponding graph node as the node's initial feature tensor; The graph convolutional spatial feature extraction network is used to perform multi-hop graph convolutional aggregation on the initial feature tensor of the node to obtain a spatial feature matrix that captures spatial electromagnetic coupling relationships. The spatial feature matrix is input into the temporal convolutional temporal feature extraction network, and transient high-frequency noise features that evolve over time are extracted through causal dilated convolution kernels. After fully connected mapping, the predicted interference frequency domain feature vector is output.
4. The electromagnetic compatibility integrated protection system for an electric vehicle drive system according to claim 3, characterized in that, It also includes an offline model pre-training module, which is configured to perform the following loss function calculation logic during the pre-training stage of the spatiotemporal graph isomorphic prediction model: Obtain the sample predicted interference frequency domain feature vector and the corresponding real interference frequency domain label vector; The mean square error between the sample predicted interference frequency domain feature vector and the real interference frequency domain label vector is calculated and used as the basic reconstruction loss. Calculate the energy leakage penalty term of the sample predicted interference frequency domain feature vector at the Nyquist folding frequency; Calculate the theoretical common-mode current constraint boundary of Kirchhoff's current law based on the topological parasitic parameter matrix, determine whether the sample predicted interference frequency domain feature vector exceeds the theoretical common-mode current constraint boundary, and if it does, calculate the physical boundary over-limit penalty term. The basic reconstruction loss, the energy leakage penalty term, and the physical boundary violation penalty term are weighted and summed to obtain the target loss function. The model network weights are then updated based on the target loss function using the backpropagation algorithm.
5. The electromagnetic compatibility integrated protection system for an electric vehicle electric drive system according to claim 1, characterized in that, When the collaborative anti-interference calculation module calculates the carrier band spreading command for the underlying microcontroller, it specifically includes: Extract all feature frequency nodes whose energy density exceeds a preset interference threshold from the predicted interference frequency domain feature vector to determine the target discrete interference spectrum; Locate the peak frequency coordinates corresponding to the maximum common-mode energy density from the target discrete interference spectral lines; The spread spectrum modulation bandwidth for the peak frequency coordinate is calculated using the following formula: the spread spectrum bandwidth is equal to the energy density value corresponding to the peak frequency coordinate divided by a preset frequency band smoothing attenuation coefficient. A pseudo-random sequence is generated based on the spread spectrum modulation broadband range, and the pseudo-random sequence is mapped to the dynamic frequency adjustment period of the pulse width modulation signal to generate the carrier band spread spectrum command.
6. The electromagnetic compatibility integrated protection system for an electric vehicle electric drive system according to claim 1, characterized in that, When the collaborative disturbance rejection calculation module calculates the reverse compensation injection current for the active filter circuit, it specifically includes: Perform an inverse discrete Fourier transform on the predicted interference frequency domain feature vector to restore the time-domain transient common-mode current waveform; Obtain the amplitude-frequency transmission attenuation coefficient and phase delay angle of the high-frequency injection transformer in the active filter circuit; The time-domain transient common-mode current waveform is inverted to obtain an inverted reference current; The amplitude of the inverting reference current is pre-amplified and compensated based on the amplitude-frequency transmission attenuation coefficient, and the phase of the inverting reference current is over-adjusted and compensated based on the phase delay angle to obtain the reverse compensation injection current.
7. The electromagnetic compatibility integrated protection system for an electric vehicle drive system according to claim 1, characterized in that, It also includes an extreme condition safety interception submodule, specifically including: Real-time monitoring of the phase current change rate and DC bus voltage fluctuation rate of the target electric drive system; When it is determined that the phase current change rate is greater than the preset transient change threshold and the DC bus voltage fluctuation rate exceeds the voltage safety envelope, a hardware-level anti-push-through interrupt is triggered. In response to the hardware-level anti-penetration interruption, the data bus transmission from the multi-dimensional parameter synchronous sampling module to the spatiotemporal graph isomorphic prediction module is suspended through the bus isolation device; The maximum derating modulation command is directly sent to the pulse width modulation generator based on the preset limit safety parameter library.
8. The electromagnetic compatibility integrated protection system for an electric vehicle drive system according to claim 7, characterized in that, When the extreme condition safety interception submodule issues the maximum derating modulation command, it specifically includes: Extract the preset dynamic dead time compensation amount and switching frequency clamping value from the extreme safety parameter library; The carrier frequency of the pulse width modulation generator is forcibly locked to the switching frequency clamp value, terminating all spread spectrum adjustments. In the dead-time control register of the pulse width modulation generator, the initial dead time is superimposed with the dynamic dead-time compensation amount; The actual dead time after superposition is used to physically isolate and extend the drive signals of the upper and lower bridge arms in order to forcibly suppress the common-mode voltage spikes caused by the maximum parasitic oscillation.
9. The electromagnetic compatibility integrated protection system for an electric vehicle drive system according to claim 1, characterized in that, It also includes a physical layer closed-loop feedback fine-tuning submodule, specifically including: The actual output conducted interference spectrum was collected by a high-frequency current probe placed at the front end of the high-voltage DC bus of the target electric drive system. Compare the residual sequence between the actual output conducted interference spectrum and the predicted interference frequency domain feature vector; When the energy error of a specific frequency band in the residual sequence is greater than the dynamic tolerance threshold, the operating condition features corresponding to that specific frequency band are extracted. Using the operating condition features and the actual output conducted interference spectrum as incremental learning samples, the network weights in the spatiotemporal graph convolutional spatial features of the spatiotemporal graph isomorphic prediction model are extracted and fine-tuned online using an exponential moving average strategy.
10. The electromagnetic compatibility integrated protection system for an electric vehicle electric drive system according to claim 1, characterized in that, The multi-dimensional parameter synchronous sampling module, when acquiring the operating condition characteristic sequence of the target electric drive system within a preset time window, specifically includes: According to the preset oversampling frequency, the three-phase stator current, rotor electromechanical angle, direct torque control command and modulation ratio parameters of the target electric drive system are collected respectively. The three-phase stator currents are subjected to Clark and Park coordinate transformations to obtain the d-axis and q-axis current components. The d-axis and q-axis current components, the rotor electromechanical angle, the direct torque control command, and the modulation ratio parameter are spliced together according to the time dimension to form a multidimensional tensor; The multidimensional tensor is normalized to eliminate the magnitude difference between different physical dimensions, thereby obtaining the operating condition feature sequence.
