Parametric digital control method and system for wide power range charging module

Abstract

The application relates to the field of digital power supply control and provides a parameterized digital control method and system suitable for a wide power range charging module, the method comprising the following steps: collecting original sensor signals, constructing a slow working condition characteristic vector through adaptive temperature drift compensation, and driving a physical constraint reference prediction model to generate a second baseline behavior parameter vector; monitoring gate driving current by using gate leakage capacitance, extracting a transient characteristic value and quantifying the transient characteristic value into a ZVS quality factor; fusing the slow working condition characteristic vector and the ZVS quality factor, generating a correction vector through an adaptive prediction model, generating a control parameter vector in combination with the second baseline behavior parameter vector, and updating online according to feedback; executing hardware physical limit constraint according to a safety constraint set, generating a safety parameter and an abnormal flag bit, and writing the safety parameter and the abnormal flag bit into a hardware register after arbitration with the ZVS quality factor. The application fuses physical constraint prediction and transient sensing correction, and constructs a digital control system which considers dynamic response, drift compensation and safety redundancy.

Description

Technical Field

[0001] This invention relates to the field of digital power control technology, and in particular to a parametric digital control method and system applicable to charging modules with a wide power range. Background Technology

[0002] With the rapid development of new energy industrial vehicles, energy storage systems, and data center power supply, high-efficiency, high-power-density DC charging modules have become core power conversion units. To achieve efficient and stable operation under wide input voltage and output load ranges, digitally controlled soft-switching resonant topologies have become the mainstream technology for wide-power-range charging modules. However, traditional digital control methods often rely on fixed parameters or segmented scheduling, making it difficult to adapt to continuously changing operating conditions. This can easily lead to decreased operating efficiency, slow dynamic response, or loop oscillations. Therefore, there is an urgent need to develop a continuously adaptive parametric digital control method applicable to the entire operating range to improve the performance and operational reliability of wide-power-range charging modules.

[0003] Chinese patent application CN116632968A discloses a method for optimizing the efficiency of fast charging piles over a wide power range. This method involves a connection point switching control strategy for a multi-tap resonant inductor. The controlled objects include a fast charging device for electric vehicles based on a phase-shifted full-bridge topology, a charging power acquisition module, a multi-tap resonant inductor, a multi-relay switching module connected to the multi-tap resonant inductor, and a relay switching control module. When the fast charging device operates under different states of charge of the electric vehicle, the charging power changes. The charging power acquisition module acquires the real-time charging power and sends it to the relay switching control module. The relay switching control module generates switching signals for each relay based on the power level and finally sends these signals to the multi-relay switching module.

[0004] However, current technology still faces many challenges. When the operating point of the charging module switches from full-load heavy-load to deep light-load conditions, such as from 30kW to 100W, the energy state of its resonant cavity undergoes a fundamental change. At this time, traditional digital controllers with fixed parameters, whose control parameters are only optimized for the rated operating point, cannot adaptively match the current light-load state. Once the fixed dead time is insufficient to complete the charging and discharging of the parasitic capacitance of the switching transistor, the controller will be unable to detect and prevent the immediate loss of the zero-voltage switching (ZVS) condition, causing the switching transistor to be forced into a high-loss hard-switching state. This hard-switching state not only causes a sharp increase in switching losses and a precipitous drop in efficiency, but also the concentrated thermal stress generated on the power device will greatly increase the risk of overheating and damage, thereby jeopardizing the operational reliability of the charging module throughout its entire life cycle. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides a parameterized digital control method and system suitable for charging modules with a wide power range, the specific technical solution of which is as follows:

[0006] Parametric digital control methods applicable to wide power range charging modules include:

[0007] The original sensor signals of the charging module are collected to construct a slow-operating condition feature vector, which drives the physical constraint benchmark prediction model to generate a baseline behavior parameter vector. The baseline behavior parameter vector is parsed to generate a second baseline behavior parameter vector containing the predicted PWM parameters and the predicted loop parameters.

[0008] The gate drive current is monitored by utilizing the coupling effect of the gate-drain capacitance to capture transient noise fingerprints and generate digital noise sequences. Transient feature values ​​are extracted based on the digital noise sequences, and multi-level comparisons are performed between the transient feature values ​​and preset perfect ZVS thresholds and lost ZVS thresholds, which are then mapped to ZVS quality factors.

[0009] The slow operating condition feature vector and ZVS quality factor are fused to construct a fused operating condition vector. The fused operating condition vector is inferred through an adaptive prediction model to generate a correction vector. The second baseline behavior parameter vector is combined to generate a control parameter vector. The first adaptive prediction weight of the adaptive prediction model is updated based on the actual ZVS quality factor feedback to generate the second adaptive prediction weight.

[0010] Based on the set of safety constraints and the control parameter vector, a safety parameter vector and anomaly flag are generated. Based on the anomaly flag and the ZVS quality factor, an arbitration parameter vector is generated by selecting from the safety parameter vector or conservative fixed parameters. The arbitration parameter vector is fixed-pointed and updated to the hardware register according to the PWM synchronization trigger signal.

[0011] Furthermore, the method for generating the second baseline behavior parameter vector includes:

[0012] The original sensor signals of the charging module are collected to extract the steady-state input voltage, steady-state output current and steady-state device temperature. Based on the temperature deviation of the steady-state device temperature relative to the reference temperature, adaptive temperature drift compensation normalization processing is performed to generate input voltage features, output current features and temperature features respectively, which are then combined to generate a slow-operating condition feature vector.

[0013] The slow-operation feature vector is input into the physical constraint baseline prediction model for inference, and the baseline behavior parameter vector containing the pulse width modulation (PWM) target and the loop behavior target is predicted.

[0014] The pulse width modulation (PWM) target is extracted to generate predicted PWM parameters. At the same time, the loop behavior target is extracted and combined with the small signal model of the controlled object to form predicted loop parameters. The predicted PWM parameters and predicted loop parameters are combined to generate a second baseline behavior parameter vector.

[0015] Furthermore, the method for performing the adaptive-temperature drift compensation normalization process includes:

[0016] Using the temperature offset and the preset boundary temperature coefficient, the reference normalized lower boundary and the reference normalized upper boundary are corrected, and the dynamic normalized lower boundary function and the dynamic normalized upper boundary function are calculated respectively.

[0017] The lower and upper normalized bounds of the reference normalization are the minimum and maximum normalized boundary values ​​of the steady-state input voltage or steady-state output current calibrated at the reference temperature, respectively.

[0018] By using the dynamically normalized lower boundary function and the dynamically normalized upper boundary function, the steady-state input voltage and steady-state output current are linearly normalized to obtain the input voltage characteristics and output current characteristics.

[0019] Furthermore, the method for constructing the predicted loop parameters includes:

[0020] Construct a small-signal model of the controlled object based on the feature vectors of slow operating conditions;

[0021] Construct a transfer function that includes the proportional gain and integral gain to be solved, and define the product of the transfer function and the small-signal model of the controlled object as the open-loop transfer function;

[0022] Based on the predicted target closed-loop bandwidth and predicted target phase margin contained in the loop behavior objective, the amplification constraint and phase constraint of the open-loop transfer function are established.

[0023] Solve the equations of the increase constraint and the phase constraint simultaneously to calculate the proportional gain and integral gain, and construct the prediction loop parameters based on the proportional gain and integral gain.

[0024] Furthermore, the increase constraint condition is that at the predicted target closed-loop bandwidth, the magnitude of the open-loop transfer function is equal to 1;

[0025] The phase constraint condition is that at the predicted target closed-loop bandwidth, the sum of the phase angle of the open-loop transfer function and the phase angle of the controlled object's small-signal model is equal to the sum of the critical instability point -180° at the predicted target closed-loop bandwidth and the preset predicted target phase margin.

[0026] Furthermore, the mapping method for the ZVS quality factor includes:

[0027] Within the preset switching event sampling window, the gate drive current is sampled and digitized to extract the transient noise fingerprint in the gate drive current and generate a digital noise sequence.

[0028] Transient noise fingerprint components are extracted based on digital noise sequences. The transient noise fingerprint components are squared within the sampling window of the switch event to obtain the instantaneous energy. The instantaneous energy is then accumulated to generate transient feature values.

[0029] Based on the preset perfect ZVS threshold and lost ZVS threshold, multi-level threshold comparison is performed on transient feature values ​​to classify them into perfect zero-voltage switching ZVS, boundary zero-voltage switching ZVS, and lost zero-voltage switching ZVS, and mapped to the corresponding ZVS quality factors.

[0030] Furthermore, the method for classifying the perfect zero-voltage switch (ZVS), the boundary zero-voltage switch (ZVS), and the lost zero-voltage switch (ZVS) includes:

[0031] Set a perfect ZVS threshold and a lost ZVS threshold, and limit the perfect ZVS threshold to be less than the lost ZVS threshold;

[0032] If the transient characteristic value does not exceed the perfect ZVS threshold, the current state is determined to be a perfect zero-voltage switch ZVS, and the value of the ZVS quality factor is mapped to 0.

[0033] If the transient eigenvalue exceeds the perfect ZVS threshold but does not exceed the lost ZVS threshold, the current state is determined to be a boundary zero voltage switch ZVS, and the ZVS quality factor is mapped to a value of 1.

[0034] If the transient characteristic value exceeds the lost ZVS threshold, the current state is determined to be a lost zero-voltage switch (ZVS), and the ZVS quality factor is mapped to a value of 2.

[0035] Furthermore, the method for generating the second adaptive prediction weights includes:

[0036] Based on the slow operating condition feature vector, combined with the ZVS quality factor, the feature concatenation operation is used to fuse the information dimension in order to construct a fused operating condition vector.

[0037] The fused operating condition vector is input into the adaptive prediction model, and the first adaptive prediction weight of the adaptive prediction model is used to infer the fused operating condition vector, and the correction vector is output.

[0038] The second baseline behavior parameter vector and the correction vector are added element by element to generate the control parameter vector;

[0039] Based on the fused operating condition vector and the correction vector, the first adaptive prediction weights are updated by minimizing the loss function between the actual ZVS quality factor and the target ZVS quality factor, thereby generating the second adaptive prediction weights.

[0040] Furthermore, the method for generating the arbitration parameter vector includes:

[0041] Based on the set of security constraints defined by the minimum and maximum hardware security boundary vectors, hardware physical limit constraints are applied to the control parameter vectors through a limiting function to generate security parameter vectors and anomaly flags.

[0042] The system coordinates the monitoring of anomaly flags and ZVS quality factors to determine whether a hard or soft fault is triggered. If a fault is triggered, a conservative fixed parameter is forcibly selected as the arbitration parameter vector. If no fault is triggered, a safe parameter vector is selected as the arbitration parameter vector.

[0043] The arbitration parameter vector is fixed-point processed into a hardware control word, and the hardware control word is updated to the hardware register at the synchronization start point of the next cycle according to the pulse width modulation PWM synchronization trigger signal SyncTrigger.

[0044] A parameterized digital control system suitable for wide-power-range charging modules, which implements the aforementioned parameterized digital control method for wide-power-range charging modules, includes a baseline prediction module, a real-time sensing module, an adaptive correction module, and a safety arbitration module.

[0045] The baseline prediction module is used to collect the original sensor signals of the charging module to construct a slow operating condition feature vector, drive the physical constraint benchmark prediction model to generate a baseline behavior parameter vector, and parse the baseline behavior parameter vector to generate a second baseline behavior parameter vector containing the predicted PWM parameters and the predicted loop parameters.

[0046] The real-time sensing module is used to monitor the gate drive current by utilizing the coupling effect of the gate-drain capacitance to capture transient noise fingerprints and generate digital noise sequences. Based on the digital noise sequences, transient feature values ​​are extracted, and the transient feature values ​​are compared with preset perfect ZVS thresholds and lost ZVS thresholds at multiple levels to map them into ZVS quality factors.

[0047] The adaptive correction module is used to fuse the slow operating condition feature vector and the ZVS quality factor to construct a fused operating condition vector, to infer the fused operating condition vector through an adaptive prediction model to generate a correction vector, to combine the second baseline behavior parameter vector to generate a control parameter vector, and to update the first adaptive prediction weight of the adaptive prediction model based on the actual ZVS quality factor feedback to generate the second adaptive prediction weight.

[0048] The security arbitration module is used to generate a security parameter vector and anomaly flag based on the security constraint set and control parameter vector, select from the security parameter vector or conservative fixed parameters based on the anomaly flag and ZVS quality factor to generate an arbitration parameter vector, and fix the arbitration parameter vector and update it to the hardware register according to the PWM synchronization trigger signal.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention introduces a ZVS physical constraint loss term during the offline training phase of the physical constraint benchmark prediction model, ensuring that the dead time of the online prediction model always conforms to the ZVS physical boundary. This overcomes the problem that traditional fixed-parameter or data-driven models lose ZVS conditions, causing a sharp drop in efficiency and thermal stress on devices due to fixed parameters or blind interpolation under edge conditions such as deep light load.

[0051] This invention actively isolates and creates a warning zone for boundary zero-voltage switch ZVS by comparing the real-time extracted transient feature values ​​with preset perfect ZVS threshold and lost ZVS threshold in multiple levels. This solves the problem that traditional binary detection methods cannot detect insufficient ZVS margin caused by device aging or temperature drift before ZVS is completely lost.

[0052] This invention constructs a slow self-learning closed loop, uses the actual perceived ZVS quality factor as performance feedback, and utilizes an online self-learning algorithm to iteratively update the weights of the adaptive prediction model. This enables the model to autonomously adapt to parameter changes caused by long-term operation of the specific hardware it is deployed on, overcoming the problem that traditional lookup tables or fixed parameter models cannot compensate for individualized model mismatch caused by device aging, temperature drift, and manufacturing tolerances.

[0053] This invention utilizes the pulse width modulation (PWM) synchronous trigger signal SyncTrigger as a hardware trigger to ensure that the arbitrated control parameters are atomically updated to the hardware register only at the end of the current PWM cycle and the start of the next cycle's synchronization, thus avoiding dynamic loop oscillations and instability caused by asynchronous changes in control parameters during the switching cycle. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1This is a flowchart illustrating the principle of the parametric digital control method applicable to charging modules with a wide power range according to the present invention.

[0056] Figure 2 This is a functional block diagram of the parametric digital control system of the present invention, applicable to charging modules with a wide power range. Detailed Implementation

[0057] The technical solutions of 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.

[0058] Example 1:

[0059] Please see Figure 1 As shown, this embodiment provides a parameterized digital control method suitable for charging modules with a wide power range, including:

[0060] S1000 collects raw sensor signals from the charging module to construct a slow-operating-condition feature vector. Driving the physical constraint benchmark prediction model Generate baseline behavior parameter vector For baseline behavior parameter vector Perform analysis to generate parameters including predicted PWM parameters and predicted loop parameters The second baseline behavior parameter vector .

[0061] Specifically, this step aims to predict the charging module's original analog sensor signals within a millisecond-level low-frequency scheduling cycle using a physical constraint benchmark model based on a deeply fused resonant topology. Open-loop prediction yields a set of underlying parameters that satisfy the zero-voltage switching (ZVS) condition and maintain loop stability margin; this is the second baseline behavior parameter vector. It achieves smooth, stable, and efficient parameter adaptation across the entire power range, avoiding parameter jump and oscillation problems in traditional look-up table (LUT) schemes, and overcoming the physical uninterpretability limitations of pure data-driven models.

[0062] Further, step S1000 includes:

[0063] Step S1100: Acquire the raw sensor signals of the charging module to extract the steady-state input voltage. Steady-state output current and steady-state device temperature Based on the steady-state device temperature Relative to reference temperature The temperature offset is subjected to adaptive temperature drift compensation normalization processing to generate input voltage characteristics. Output current characteristics and temperature characteristics Generating slow-working-condition feature vectors by combination .

[0064] Specifically, this step aims to extract and construct a slow-condition feature vector with a consistent physical benchmark under different temperature conditions from the raw sensor signals of the charging module, which contain high-frequency disturbances and nonlinear drift. The raw sensor signal includes: raw input voltage. Original output current and the temperature of the original power devices Its core purpose is to utilize the temperature of the original power device. The steady-state device temperature extracted Eliminate the influence of the steady-state device temperature at the feature level in advance. The resulting sampling drift and component parameter changes provide a unified and decoupled operating condition input for subsequent steps.

[0065] In the specific implementation process, the original sensor signal is acquired through the following sensing circuit:

[0066] First, the original input voltage The high input voltage is scaled down to a low-voltage analog signal, i.e., the original input voltage, by setting a high-resistance precision voltage divider network on the DC input bus of the charging module. This scaling is compatible with the input range of the analog-to-digital converter (ADC) in the digital controller. .

[0067] Second, the original output current The data is collected by connecting a current sensor, such as a precision sampling resistor or a Hall effect-based current sensor, in series in the output path of the charging module. This sensor converts the output current into a proportional voltage or current signal, which, after signal conditioning, becomes the original output current. .

[0068] Third, the temperature of the original power devices The temperature is collected by physically mounting a temperature sensor, such as a negative temperature coefficient thermistor or an integrated temperature sensing chip, at a predetermined temperature measurement point on a key power device of the charging module, such as the main switch of an inductor-inductor-capacitor resonant topology (LLC) or its heatsink, to obtain an analog electrical signal reflecting the junction temperature or case temperature of the device, which is the original power device temperature. .

[0069] Since the acquired signals inevitably couple with switching noise, in order to filter out high-frequency transient disturbances and extract the steady-state operating point that reflects the system's thermal balance and power flow, the original analog sensor signals, i.e., the original input voltage, are processed. Original output current and the temperature of the original power devices Low-pass filtering is performed on each component, or periodic averaging is performed within a preset time window to obtain steady-state operating condition data, i.e., steady-state input voltage. Steady-state output current and steady-state device temperature .

[0070] In wide-power, wide-temperature applications where charging modules operate, not only will the characteristics of power devices change with the steady-state device temperature, but also... Changes in various components, such as input capacitance and on-resistance, as well as key components of the resonant cavity, such as resonant inductors and capacitors, and even the analog sampling circuit itself, all exhibit significant nonlinear temperature drift characteristics. If a traditional normalization method with fixed boundaries is used, this temperature drift will cause the input features of the prediction model to drift, meaning that the same physical conditions will manifest as different feature vectors at different temperatures, greatly increasing the training difficulty and generalization error of the prediction model.

[0071] To address this issue, this step employs an adaptive temperature drift compensation normalization method, designed to adjust for steady-state device temperature. The function of input voltage characteristics. Taking normalization as an example, the specific process and formula are as follows:

[0072] ;

[0073] in, This represents the input voltage characteristic, and its value is usually normalized to the [0,1] interval; This represents the dynamically normalized lower boundary function, which varies with the steady-state device temperature. The dynamically adjusted normalized lower boundary value is used to compensate for the lower boundary drift caused by temperature in the normalization calculation; This represents the dynamically normalized upper boundary function, which varies with the steady-state device temperature. The dynamically adjusted normalized upper boundary value is used to compensate for upper boundary drift caused by temperature in the normalization calculation.

[0074] The dynamically normalized lower boundary function Based on the baseline normalized lower boundary and lower boundary temperature coefficient Based on the current steady-state device temperature Relative to reference temperature The temperature offset is obtained. Wherein, the reference normalized lower boundary... At the reference temperature The steady-state input voltage obtained by subcalibration Minimum normalized boundary value.

[0075] The specific process formula is as follows:

[0076] ;

[0077] The dynamically normalized upper boundary function Based on the benchmark normalized upper boundary and upper boundary temperature coefficient Based on the current steady-state device temperature Relative to reference temperature The temperature offset is obtained. Wherein, the reference normalized upper boundary... At the reference temperature The steady-state input voltage obtained by subcalibration Maximum normalized boundary value.

[0078] The specific process formula is as follows:

[0079] ;

[0080] Similarly, for steady-state output current Adaptive normalization is performed using the corresponding boundary temperature coefficient to generate normalized output current characteristics. And the steady-state device temperature It can perform either fixed or adaptive normalization to generate normalized temperature features. .

[0081] Finally, by combining the above features, a slow-operating-condition feature vector with intrinsic temperature drift compensation is generated. .

[0082] Step S1200, convert the slow operating condition feature vector Input physical constraint benchmark prediction model Inference is performed to predict the baseline behavioral parameter vector that includes the pulse width modulation (PWM) target and the loop behavior target. .

[0083] More specifically, through physical constraint benchmark prediction models For slow operating condition feature vectors Inference is performed to predict the baseline behavior parameter vector. The baseline behavior parameter vector Includes predicted switching frequency and predicting dead zone time The pulse width modulation (PWM) target is constructed, and the closed-loop bandwidth of the predicted target is determined. Phase margin of the predicted target The target of the loop behavior.

[0084] Specifically, this step aims to leverage the slow-working-condition feature vector generated in step S100. Physically constrained benchmark prediction model using deeply fused resonant topology Structured baseline behavior parameter vectors can be predicted in an open-loop manner within a low-frequency scheduling cycle at the millisecond level. The baseline behavior parameter vector This not only ensures that the predicted pulse width modulation (PWM) satisfies the zero-voltage switching (ZVS) condition, but also provides a smooth and stable loop behavior target for subsequent steps.

[0085] In the specific implementation process, during the controller's millisecond-level low-frequency scheduling cycle, the slow-operating condition feature vector is... Input to pre-trained physical constraint benchmark prediction model Execute inference. This physical constraint benchmark prediction model It is a physically constrained neural network (PINN) or a proxy model with embedded physical mechanisms, which is different from the traditional black-box model.

[0086] Physically constrained baseline prediction model and its benchmark prediction weight The construction process, i.e., the offline training phase, is the key to this step. To avoid the prediction failure of traditional black-box models due to blind interpolation in edge cases with sparse data, baseline prediction weights are used. By minimizing the composite loss function This is obtained through training. The composite loss function... Cooperatively balancing the data fitting loss term and ZVS physical constraint loss term Among them, the data fitting loss term Used for quantifying physical constraint benchmark prediction models The error between the current predicted value and the offline-calibrated optimal target parameters; ZVS physical constraint loss term. Used for quantifying physical constraint benchmark prediction models The extent to which the current predicted value violates the ZVS physical boundary.

[0087] The ZVS physical constraint loss term This is a core mechanism designed to address ZVS failure under light load conditions. Its physical principle lies in the fact that, in resonant topologies such as inductor-inductor-capacitor (LLC), achieving zero-voltage switching (ZVS) requires ensuring that the parasitic output capacitance is completely discharged by the resonant current before the switching transistor turns on. This discharge process requires a minimum time, i.e., the minimum ZVS resonant time. Therefore, the controller is set with a predictive dead time. It must be strictly greater than or equal to the minimum ZVS resonant time. The minimum ZVS resonant time It is not a constant, but is based on the current slow-working condition feature vector. The operating conditions and fixed topology parameters of the charging module, such as the magnetizing inductance, are characterized. Resonant capacitor These are dynamic variables that are jointly determined by [various factors].

[0088] During the offline training of the model, the predicted dead time output by the model in each training step. It will be directly substituted into the ZVS physical constraint loss term The ZVS physical constraint loss term The mechanism is as follows: when predicting dead time Greater than or equal to the minimum ZVS resonant time At that time, ZVS physical constraint loss term No penalty is imposed; composite loss function The loss term will be mainly composed of data fitting. Dominantly, the model will focus on fitting the optimal experimental data; when predicting dead time... Less than the minimum ZVS resonant time At that time, ZVS physical constraint loss term A positive penalty value is generated, which will force the baseline prediction weights through the optimization algorithm. The adjustment will be made to increase the predicted dead time. The value is maintained until it returns to the predicted dead time. Greater than or equal to the minimum ZVS resonant time Within the physical security boundaries.

[0089] Physical constraint baseline prediction model during online controller execution Predict weights using offline-trained, fixed benchmarks For the input slow-working-condition feature vector Perform fast forward inference to predict the baseline behavior parameter vector. Unlike traditional explicit lookup table (LUT) methods that directly output flat, specific low-level controller parameters, baseline behavior parameter vectors... Physical constraint benchmark prediction model The prediction task is decoupled into a pulse width modulation (PWM) target and a loop behavior target. This vector It contains four structured vectors, namely the predicted switching frequency. Predicting dead time Predict the target closed-loop bandwidth and the predicted target phase margin .in, The value of is the physical constraint baseline prediction model It is assumed that the feature vector under the current slow working condition The optimal PWM switching frequency is given in Hertz (Hz). The value of is the physical constraint baseline prediction model It is assumed that the feature vector under the current slow working condition The dynamic performance index set below characterizes the loop response speed, with the unit being radians per second (rad / s). The value of is the physical constraint baseline prediction model It is assumed that the feature vector under the current slow working condition The dynamic performance index set below to characterize the stability of the loop, with the unit being degrees (°).

[0090] The predicted switching frequency and predicting dead zone time This constitutes the target of pulse width modulation (PWM). Specifically, it involves predicting the dead time. The value has been affected by the ZVS physical constraint loss term during the offline training phase. The physical constraints ensure its physical compliance with lightly loaded ZVS failures. The predicted target closed-loop bandwidth... Phase margin of the predicted target This constitutes a loop behavior objective.

[0091] Step S1300: Extract the target parameters for pulse width modulation (PWM) to generate predicted PWM. Simultaneously, the loop behavior target is extracted and combined with the small-signal model of the controlled object. To form the predicted loop parameters Combined Predictive PWM Parameters and predicted loop parameters Generate the second baseline behavior parameter vector .

[0092] More specifically, the pulse width modulation (PWM) target is unpacked to generate predictive PWM parameters that ensure zero-voltage switching (ZVS). Simultaneously, the loop behavior target is unpacked and combined with the small-signal model of the controlled object. Establish and solve the equations for the increase constraint and phase constraint, and inversely solve for the proportional gain. and integral gain , forming the prediction loop parameters The predicted PWM parameters and predicted loop parameters Merged into a second baseline behavior parameter vector .

[0093] Specifically, this step aims to transform the baseline behavioral parameter vector predicted in step S1200... Unpacking and computation are performed to generate the underlying parameter set that the digital controller can execute, namely the second baseline behavior parameter vector. The second baseline behavior parameter vector Includes predictive PWM parameters and predicted loop parameters Its core objective is to directly map the baseline behavior parameter vector. Given a physically constrained PWM target, predictive PWM parameters are generated to ensure zero-voltage switching (ZVS). Based on baseline behavior parameter vector The smooth loop behavior target provided in the algorithm is used to determine the predictive loop parameters that change smoothly in real time through deterministic analytical calculation. This avoids parameter jumps and dynamic oscillations at the source.

[0094] In the specific implementation process, this step processes the baseline behavior parameter vector within a millisecond-level scheduling cycle. The analysis and computation are performed. This process is decoupled into the following two parallel tasks, as follows:

[0095] First, unpack the PWM parameters. Baseline behavior parameter vector. The PWM target in the text is the predicted switching frequency. and predicting dead zone time It was directly unpacked and configured to predict PWM parameters. Due to the predictive dead time of this PWM target. The ZVS physical constraint loss term has been applied during the offline training phase in step S1200. The physical constraints of this task are directly applied to ensure that the controller parameters comply with the ZVS physical boundary, thereby solving the problem of light load failure.

[0096] Second, loop parameter unpacking. This task is a key mechanism for achieving dynamic stability and resolving dynamic oscillations. The controller determines the parameters based on the current operating condition, i.e., the slow-operating condition feature vector. A small-signal model of the controlled object that varies with operating conditions can be generated from internal storage or calculated in real time. The model Describe the main topology of the charging module, such as an inductor-inductor-capacitor resonant topology LLC, and its feature vector under the current slow operating condition. The dynamic transfer function.

[0097] Subsequently, the transfer function of the controller is set. The specific logical formula is as follows:

[0098] ;

[0099] in, The Laplace transform is a complex variable whose value is a complex number. It is the standard mathematical symbol used in control theory to analyze the dynamic characteristics of a system in the complex frequency domain. and These are the proportional gain and integral gain of the controller, respectively, which are the unknowns to be solved in this task. Both are predicted loop parameters. The component, proportional gain The integral gain is used to determine the proportion of the controller's response to the current error. Used to determine the strength of the controller's response to accumulated errors in order to eliminate steady-state errors.

[0100] Next, to solve for the proportional gain of the controller... and integral gain This makes the system's open-loop transfer function The predicted target closed-loop bandwidth specified in the smooth loop behavior objective. At this point, both the increase constraint and the phase constraint are satisfied, ensuring stability. The open-loop transfer function... It is a transfer function and the small signal model of the controlled object The product of.

[0101] First, the increase constraint is used to require the total open-loop gain of the system, i.e., the controller gain. and the gain of the controlled object The product of these factors is used to predict the target closed-loop bandwidth. At this point, the value must be exactly equal to 1. Wherein, The magnitude operator is used to calculate the open-loop system's closed-loop bandwidth in the predicted target. Total open-loop gain at the frequency; Represents the imaginary unit. It is the complex variable of the Laplace transform when performing frequency response analysis. The value of .

[0102] Second, phase constraints are used to require the total open-loop phase of the system, i.e., the controller phase angle. Phase angle with the controlled object The sum of the predicted target closed-loop bandwidth At this point, it must be exactly equal to "critical instability point - 180°" plus the predicted target phase margin specified in the smooth loop behavior objective. .in, The phase angle operator takes a value in degrees (°) or radians (rad) and is used to calculate the phase of a complex number.

[0103] By solving the above proportional gain including the controller in the embedded controller and integral gain The stability constraint equations for the two unknowns are used to calculate the unique closed-loop bandwidth that satisfies the predicted target. Phase margin of the predicted target proportional gain and integral gain value.

[0104] Based on the calculated proportional gain of the controller and integral gain Constructing predicted loop parameters .

[0105] Finally, the predicted PWM parameters calculated above are... and predicted loop parameters Perform vector concatenation to combine them into a unified second baseline behavior parameter vector. .

[0106] S2000 utilizes gate-drain capacitance Coupling effect monitoring gate drive current To capture transient noise fingerprints and generate digital noise sequences Based on digital noise sequence Extracting transient eigenvalues And based on transient eigenvalues Compared to the preset perfect ZVS threshold and lost ZVS threshold Perform multi-level comparisons and map them to the ZVS quality factor. .

[0107] Specifically, this step aims to perceive the actual physical execution effect of the predicted parameters output by step S1000 in real time within a microsecond-level switching cycle through a non-invasive transient noise fingerprint monitoring technology, such as predicting dead time. It outputs the quantized ZVS quality factor. As a key feedback input for subsequent steps.

[0108] Further, step S2000 includes:

[0109] Step S2100, in the preset switch event sampling window Internally, the gate drive current Sampling and digitization are performed to extract the gate drive current. Transient noise fingerprints in the data are used to generate digitized noise sequences. .

[0110] More specifically, based on gate-drain capacitance The gate drive current is monitored non-invasively using a high-frequency current sensor. To capture the drain-source voltage Transient noise fingerprints coupled by transient changes, and the gate drive current carrying the transient noise fingerprint. In the preset switch event sampling window It performs high-frequency sampling and digitization processing to generate a digitized noise sequence. .

[0111] Specifically, this step aims to employ a non-invasive transient noise fingerprinting method within a preset switching event sampling window. Internally, a transient noise fingerprint capable of reflecting the ZVS state with a high signal-to-noise ratio is captured. This transient noise fingerprint is superimposed on the gate drive current. Above this, the signal it carries is converted into a digital noise sequence. This provides a high-fidelity data foundation for real-time feature extraction in subsequent steps.

[0112] In the specific implementation process, to solve the problem that traditional ZVS detection requires direct monitoring of high voltage and high noise drain-source voltage, To address the challenges of isolation, high cost, and low signal-to-noise ratio associated with transient noise, this step employs an innovative non-invasive transient noise fingerprinting method. This method non-invasively monitors the gate drive current on the low-voltage side of the power device using a high-frequency current sensor, such as a Rogowski coil or a high-bandwidth current transformer, for example, around the gate drive path. .

[0113] The physical principle of the sampling method lies in utilizing the Miller capacitance, i.e., the gate-drain capacitance, inside the power device. As a natural sensor, it is specifically as follows:

[0114] When the prediction in step S1000 is accurate, for example, when the dead time is predicted... When accurate, the system achieves zero-voltage switching (ZVS), where the switching transistor operates at the drain-source voltage. When the voltage is approximately zero, conduction occurs; at this point, the rate of voltage change is extremely small. This is achieved through the gate-drain capacitance. The coupling current is negligible, and the gate drive current is... Maintain a smooth fingerprint without generating transient noise. The gate drive current... The value of is a variable that varies with continuous time. The changing simulated real value, usually in amperes (A); Representing a continuous-time variable, it is the time dimension describing the physical process and is used to identify the gate drive current. The instantaneous state at any physical moment.

[0115] Conversely, when the prediction in step S1000 fails, for example, when the dead time is predicted... In the event of failure, the system enters a hard-switching state, and the switching transistor operates at the drain-source voltage. The instantaneous conduction generates a high rate of voltage change. This transient voltage with a high rate of voltage change passes through the gate-drain capacitance. Coupling back to the low-voltage gate drive loop, during gate drive current A high-frequency oscillation spike with a distinct characteristic and high signal-to-noise ratio is superimposed on the signal, which is the transient noise fingerprint.

[0116] Therefore, monitoring the gate drive current It is a high signal-to-noise ratio, low-cost, and non-invasive ZVS state indirect sensing method that provides a high-fidelity signal foundation for subsequent steps.

[0117] The controller then uses an analog-to-digital converter (ADC) within a preset switching event sampling window, over a microsecond-level switching cycle. Internal gate drive current High-frequency sampling is performed to convert the noise from the analog domain to the digital domain, resulting in a digitized noise sequence. The digitized noise sequence Gate drive current An ordered set of numbers generated after discretization sampling; Represents a discrete sampling index, which is a digitized noise sequence. The serial number of each sampling point is used to identify the digitized noise sequence. The Middle The time sequence corresponding to each sampling point.

[0118] Step S2200, based on the digitized noise sequence Extract transient noise fingerprint components within the switch event sampling window. The transient noise fingerprint components are squared to obtain the instantaneous energy, and the instantaneous energy is accumulated to generate transient eigenvalues. .

[0119] More specifically, digital noise sequences are extracted using digital bandpass filters. The transient noise fingerprint component in the switch event sampling window The transient noise fingerprint components are squared to obtain the instantaneous energy, and the instantaneous energy within the window is accumulated to obtain the transient feature value. .

[0120] Specifically, this step aims to analyze the digitized noise sequence obtained in step S2100 under microsecond-level switching cycle constraints. In real time and robustly, a scalar, i.e., transient feature value, can be extracted for quantifying the intensity of transient noise fingerprints. .

[0121] In the specific implementation, to ensure that the calculation is completed within each extremely short switching cycle, the extraction calculation in this step and the sampling in step S2100 are co-deployed on an edge computing unit, such as a Field-Programmable Gate Array (FPGA) or a dedicated Digital Signal Processor (DSP), to process the digitized noise sequence. Perform pipelined real-time feature extraction.

[0122] To separate digitized noise sequences The high-frequency transient noise fingerprint and the normal low-frequency gate drive signal components contained within are fed into a digital band-pass filter (BPF). This band-pass filter only allows a preset high-frequency band that matches the physical characteristics of the transient noise fingerprint. The components within the band pass through, thereby filtering out low-frequency gate drive signals and unrelated interference from other frequency bands. The high-frequency band... The value of is a frequency range calibrated offline based on the physical characteristics of transient noise fingerprints, such as the parasitic oscillation frequency generated during hard switching.

[0123] To robustly quantify noise intensity and avoid the instability of traditional peak detection methods that are susceptible to interference from single random noise spikes, this step involves sampling within a single switching event window. That is, starting from the moment the switch is turned on, including Within a time window of each sampling point, the following processing is performed: A digital bandpass filter function is applied to the digitized noise sequence. Filtering is performed to isolate areas located in the high-frequency band. The transient noise fingerprint component is obtained; the transient noise fingerprint component after filtering is squared to obtain the instantaneous energy; the instantaneous energy of all sampling points within the sampling window is summed to extract the transient feature value. .

[0124] For example, when the charging module achieves ideal zero-voltage switching (ZVS) under a full load of 30kW, the charging and discharging process of the parasitic capacitance of the switching transistor is smooth, and the transient noise fingerprint energy is weak. The transient feature values ​​extracted in this step are... The transient characteristic value is maintained near a low baseline value. However, when the load suddenly drops to a deep light load of 100W, if the controller still uses a fixed dead time and fails to adapt in time, the resonant cavity energy will be insufficient, and the switching transistor will lose zero-voltage switching (ZVS) and enter hard-switching mode. At this time, the hard-switching process will excite violent high-frequency oscillations on the parasitic parameters of the switching transistor, resulting in a significant enhancement of the transient noise fingerprint. The transient characteristic value extracted in this step... This will result in a sudden, significant jump. This significant jump is immediately fed back to the subsequent digital control core to trigger real-time adjustments to control parameters, such as dead time or switching frequency, up to the transient characteristic value. The transient characteristic reference value corresponding to the ideal zero-voltage switch (ZVS) is brought back to near, thereby recaptured and locked into the ZVS under light load, ensuring power supply efficiency and stability over a wide power range.

[0125] Step S2300, based on the preset perfect ZVS threshold and lost ZVS threshold For transient eigenvalues Perform multi-level threshold comparisons to classify ZVS into perfect zero-voltage switch (ZVS), boundary zero-voltage switch (ZVS), and missing zero-voltage switch (ZVS), and map them to the corresponding ZVS quality factors. .

[0126] Specifically, this step aims to extract the transient feature values ​​from step S2200. Combined with the preset quantization threshold calibrated offline, i.e. the perfect ZVS threshold and lost ZVS threshold This is mapped and quantized into discrete state level values ​​with clearly defined control hierarchy meanings, i.e., the ZVS quality factor. This enables differentiated and refined responses to different zero-voltage switch (ZVS) health states.

[0127] In the specific implementation process, this step will extract transient feature values ​​in real time. Real-time multi-level comparison with two offline-calibrated preset quantization thresholds, i.e., the perfect ZVS threshold. and lost ZVS threshold And it satisfies the perfect ZVS threshold. Less than the lost ZVS threshold Wherein, the perfect ZVS threshold This is a preset lower threshold level, the value of which is usually calibrated offline based on the transient characteristic reference value measured by the charging module under ideal zero-voltage switching (ZVS) conditions; the lost ZVS threshold... It is a preset upper threshold level, and its value is always greater than the perfect ZVS threshold. Its value is usually determined offline based on the intense high-frequency oscillation energy generated by the system under specific hard-switching conditions, such as when there is a deep light load and a severely insufficient dead time.

[0128] This real-time multi-level comparison process is based on the perfect ZVS threshold. and lost ZVS threshold , will be continuous transient eigenvalues Discretized into three state level values ​​with clear engineering significance, namely the ZVS quality factor. The specific rules are as follows:

[0129] When transient eigenvalues Not exceeding the perfect ZVS threshold At that time, the system determines the ZVS quality factor. "0". This state indicates that the ZVS margin is sufficient, that is, a perfect zero-voltage switching ZVS is achieved;

[0130] When transient eigenvalues At the perfect ZVS threshold and lost ZVS threshold During this period, the system determines the ZVS quality factor. A value of "1" actively isolates a warning zone for insufficient ZVS margin, namely the boundary zero-voltage switch ZVS, enabling the control system to function even when ZVS is completely lost, i.e., the ZVS quality factor. Before jumping to "2", it detects signs of model deviation caused by device aging, temperature drift, or manufacturing tolerances, thereby initiating subsequent online corrections to achieve margin predictive maintenance of ZVS.

[0131] When transient eigenvalues Exceeding the lost ZVS threshold At that time, the system determines the ZVS quality factor. A value of "2" indicates that the ZVS has been completely lost, meaning the zero-voltage switch ZVS has been lost. In this case, the system enters a hard-switching state, and high-speed intervention must be performed immediately to ensure device safety and system stability.

[0132] S3000, fusion of slow-working-condition feature vectors and ZVS quality factor To construct a fusion working condition vector Through adaptive prediction models For fusion operating condition vector Perform inference to generate the correction vector Combined with the second baseline behavior parameter vector To generate control parameter vectors Based on actual ZVS quality factor Feedback updates adaptive prediction model First adaptive prediction weights To generate the second adaptive prediction weights .

[0133] Specifically, this step aims to leverage the slow-working-condition feature vector generated in step S1100. The second baseline behavior parameter vector output in step S1300 and the ZVS quality factor output in step S2300 By employing a predictive-corrective decoupled control paradigm, the final control parameter vector is collaboratively generated. and the updated second adaptive prediction weights .

[0134] Specifically, this step utilizes the ZVS quality factor. Real-time feedback instantly generates a correction vector. and compare it with the second baseline behavior parameter vector. The parameters are synthesized into a final control parameter vector that is ultimately applied to the system. Meanwhile, this step also utilizes online self-learning capabilities to enable the adaptive prediction model deployed in step S3200 to... It can continuously update its internal weights, from its first adaptive predicted weights. Update to the second adaptive prediction weights This actively compensates for long-term parameter drift, thereby ensuring the efficient and stable operation of the charging module across a wide power range and throughout its entire lifespan.

[0135] Further, step S3000 includes:

[0136] Step S3100, based on the slow working condition feature vector Combined with ZVS quality factor By performing feature concatenation operations, information is fused along the information dimension to construct a fused working condition vector. .

[0137] Specifically, this step aims to transform the slow-working-condition feature vector generated in step S1100. The ZVS quality factor output by step S2300 Information fusion is performed to generate a context-aware fused working condition vector. .

[0138] In the specific implementation process, this step constructs a fused operating condition vector in the digital control core. This vector, through feature concatenation, fuses two features from different sources and with different physical meanings in terms of information dimension: one is the slow-condition feature vector describing the current steady-state operating point of the system. Secondly, there is the ZVS quality factor, which reflects the transient physical switching reality of the previous cycle. That is, the physical execution effect of the prediction parameters output in step S1000.

[0139] Step S3200, fuse the working condition vector Input to adaptive prediction model Using adaptive prediction models First adaptive prediction weights For fusion operating condition vector Perform inference and output the correction vector. .

[0140] Specifically, this step aims to leverage the fused operating condition vector generated in step S3100. Utilizing a lightweight adaptive prediction model and its first adaptive prediction weights Real-time inference within microsecond-level switching cycle constraints, outputting a second baseline behavior parameter vector to compensate for the output in step S1300. Correction vector for prediction bias This enables rapid fine-tuning of control parameters and closed-loop steady-state maintenance of zero-voltage switching (ZVS).

[0141] In the specific implementation process, this step will integrate the working condition vector. This is used as input, and fed into a lightweight adaptive prediction model deployed on edge computing units. Perform high-speed inference, such as with small multilayer perceptrons, radial basis function networks, or fuzzy logic inference engines.

[0142] The adaptive prediction model The physical constraint benchmark prediction model constructed in step S1200 is architecturally similar to that built in step S1200. This achieves functional decoupling. The adaptive prediction model... Not for repeated learning of physical constraint benchmark prediction models Perform complex, full-range basic mapping tasks, i.e., slow-condition feature vectors Mapped to the second baseline behavior parameter vector Instead, it is specifically used to learn the second baseline behavior parameter vector. The prediction error is reduced, thus enabling rapid correction of dynamic deviations at the edge computing unit.

[0143] Adaptive prediction model The training objective is: when the fused working condition vector ZVS quality factor When the value is greater than 0, it means that the zero-voltage switch ZVS deviates from the ideal point, based on the fused operating condition vector. Feature vectors of medium and slow operating conditions Output the corresponding correction vector This makes the ZVS quality factor In the next cycle, the target value "0" is returned, achieving dynamic closed-loop repair of the zero-voltage switch (ZVS). The specific process formula is as follows:

[0144] ;

[0145] in, This represents the correction vector, whose values ​​are related to the second baseline behavior parameter vector. A real-number vector of the same dimension is used to represent the second baseline behavior parameter vector. Predicted switching frequency Predicting dead time Element-by-element correction of parameters; This represents the inference function of the adaptive prediction model, used to perform real-time, high-speed inference. The condition symbol is used to logically separate the fusion working condition vector. and adaptive prediction model First adaptive prediction weights That is, the first adaptive prediction weight The defined current adaptive prediction model Under the state, the fused working condition vector of the input To reason; This represents the first adaptive prediction weight, which is the adaptive prediction model. The current set of variable intrinsic parameters is used to enable the adaptive prediction model. It can learn autonomously and compensate for long-term parameter deviations caused by device aging and temperature drift.

[0146] For example, when the charging module undergoes a rapid load transition from light load to full load, such as a transition from 100W to 30kW, the second baseline behavioral parameter vector... It may not perfectly cover this dramatic dynamic transient and lags behind sudden changes in operating conditions, i.e., the second baseline behavior parameter vector. The prediction is still based on a 100W-level operating condition, resulting in insufficient dead time for a 30kW-level heavy load, thus losing the zero-voltage switching (ZVS) and causing hard switching. Through the real-time correction mechanism in this step, the system outputs a correction vector. Compensating for the corresponding deviation within a single switching cycle allows the zero-voltage switch (ZVS) to recover rapidly in subsequent switching cycles, ensuring the dynamic energy conversion efficiency and device safety margin of the high-frequency resonant circuit.

[0147] Step S3300, the second baseline behavior parameter vector and correction vector Perform element-wise addition to generate the control parameter vector. .

[0148] Specifically, this step aims to transform the second baseline behavior parameter vector predicted in step S1300 (open-loop baseline) into... The correction vector of the closed-loop real-time correction output in step S3200 Arithmetic synthesis is performed to generate the final control parameter vector applied to the power electronic hardware. .

[0149] In the specific implementation process, this step will use the second baseline behavior parameter vector and correction vector Perform vector synthesis, such as element-wise addition, to obtain the final control parameter vector. The control parameter vector This is the set of control parameters that will be actually executed in the next switching cycle, and its value is a vector of behavior parameters relative to the second baseline. Real number vectors of the same dimension are sent to the corresponding registers of the digital controller to directly drive hardware such as pulse width modulation (PWM) generators, thereby achieving closed-loop control of zero-voltage switching (ZVS).

[0150] Step S3400, based on the fused working condition vector and correction vector By minimizing the actual ZVS quality factor With the target ZVS quality factor The loss function between the first adaptive prediction weights Perform the update and generate the second adaptive prediction weights. .

[0151] More specifically, based on fused operating condition vectors and correction vector This constitutes empirical data and is combined with actual ZVS quality factors. As performance feedback, when the safety isolation mode is satisfied, an online self-learning algorithm is used to minimize the actual ZVS quality factor. and target ZVS quality factor The loss function between the first adaptive prediction weights Perform an update to generate second adaptive prediction weights. .

[0152] Specifically, this step aims to construct a slow self-learning closed loop to compensate for long-term parameter drift, utilizing the fused operating condition vector generated in step S3100 during historical runs. and the correction vector output by step S3200 The empirical data, and the actual ZVS quality factor As performance feedback, the adaptive prediction model deployed in step S3200 is improved through online self-learning. It can autonomously adapt to individual parameter mismatches caused by long-term operation of the specific hardware it deploys, and adjust its first adaptive prediction weights. Update to the optimized second adaptive prediction weights .

[0153] The actual ZVS quality factor Step S2300 applies the control parameter vector generated in step S3300 to the system. Subsequently, the ZVS state, which is actually sensed and quantified during subsequent switching cycles, takes the value of "0", "1" or "2" and is used as an adaptive prediction model. The actual results of the corrected behavior are fed back to evaluate the first adaptive prediction weights. The correction vector generated by the guidance The effectiveness.

[0154] In the specific implementation process, this step constitutes a slow self-learning closed loop to compensate for long-term parameter drift, and integrates the operating condition vector. and correction vector The empirical data, used as new training samples, is cached in a local experience replay pool. When a safe isolation mode is met, such as system idle, standby, or a specific maintenance window, this step employs transfer learning or reinforcement learning algorithms to extract data from the experience replay pool, improving the actual ZVS quality factor. Approaching the target ZVS quality factor For adaptive prediction models First adaptive prediction weights Perform mini-batch updates until they converge to the updated second adaptive prediction weights. The specific process formula is as follows:

[0155] ;

[0156] in, This represents the updated second adaptive prediction weights, whose values ​​are a set of optimized numerical values, such as the weights and biases of a neural network, used to optimize the adaptive prediction model. The corrective behavior can adapt to long-term parameter drift that has already occurred; This represents the learning rate, which is usually a small positive real number used to control the step size of weight updates in order to adjust the speed and stability of optimization. The gradient operator is indicated by its subscript relative to the first adaptive prediction weight. The partial derivative is calculated, and its value is a vector pointing towards the loss function. The direction of fastest growth is used to indicate the first adaptive prediction weight. The steepest direction to be adjusted is along the direction in which the loss decreases the fastest, since the formula subtracts the gradient operator. The loss function is a function that evaluates the error between the actual result and the ideal baseline, and is used to quantify the weights of the first adaptive prediction. The degree of poor control effect resulting from it; The target ZVS quality factor represents the ideal optimization objective of the control. Its value is always "0", representing the perfect zero-voltage switching (ZVS) state, and it is used as the loss function. An ideal benchmark for calculations.

[0157] To achieve the slow self-learning closed loop described in this step, the second adaptive prediction weights... At a preset safe time point, such as when the calculation of this step is completed and the high-speed loop of step S3200 is in a non-inference window, a higher-level control logic or scheduler overrides and replaces the first adaptive prediction weight currently used in the high-speed correction loop of step S3200. .

[0158] S4000, based on the safety constraint set and control parameter vector Generate security parameter vector and abnormal flag bits Based on the abnormal flag bit and ZVS quality factor From the security parameter vector Or conservative fixed parameters Select to generate arbitration parameter vector , arbitration parameter vector The data is fixed-point and updated to the hardware registers based on the PWM synchronous trigger signal.

[0159] Specifically, this step aims to build a highly reliable security defense based on the control parameter vector generated in step S3300. and preset safety constraint sets and conservative fixed parameters Through the hardware physical limit verification operation in step S4100 and the safety rollback arbitration operation in step S4200, a hardware control word that ensures the absolute safety of power electronic hardware is generated. .

[0160] Further, step S4000 includes:

[0161] Step S4100, based on the hardware minimum security boundary vector and hardware maximum security boundary vector Defined set of security constraints The control parameter vector is controlled by a limiting function. Apply hardware physical limits to generate a safety parameter vector. and abnormal flag bits .

[0162] Specifically, this step aims to base its implementation on a pre-defined set of security constraints. The control parameter vector generated in step S3300 Perform cycle-by-cycle hardware physical limit verification to generate a security parameter vector that ensures absolute hardware safety. It can also optionally output the exception flag. .

[0163] In the specific implementation process, to ensure the absolute reliability of the charging module, this step serves as a safety precaution by controlling the parameter vector. and a pre-defined set of security constraints Real-time physical limit verification is performed. The security constraint set... This includes the absolute physical limits of power electronic hardware, such as minimum / maximum dead time, maximum duty cycle limit, maximum / minimum switching frequency, etc.

[0164] As shown in the following process formula, this step involves the control parameter vector. Each element in the vector is subjected to hardware limiting to obtain a safety parameter vector. Optional, if the control parameter vector Significant deviation from the safety constraint set before the limit is reached. For example, if the value exceeds a preset threshold, an exception flag will be set. Among them, the abnormal flag bit The value is usually a Boolean flag, that is, 0 or 1.

[0165] ;

[0166] in, This represents the safety parameter vector after hardware-limited pruning, whose values ​​are strictly restricted to the hardware minimum safety boundary vector. and hardware maximum security boundary vector Within the closed interval; The limiting function is an element-wise limiting or clipping operation used to limit the input control parameter vector. Any element in the vector is restricted to the minimum hardware security boundary vector. and hardware maximum security boundary vector The closed interval; The minimum security boundary vector of the hardware is the set of security constraints. The defined numerical lower boundary vector takes the values ​​of a set of offline-calibrated minimum values ​​that define the absolute physical limits of the hardware, and is used as the limiting function. The lower limit criterion; The maximum security boundary vector of the hardware is the set of security constraints. The defined numerical upper boundary vector takes the values ​​of a set of offline-calibrated maximum values ​​that define the absolute physical limits of the hardware, and is used as the limiting function. The upper limit criterion.

[0167] Step S4200, Coordinate monitoring of anomaly flags and ZVS quality factor Determine whether a hard or soft fault has been triggered. If a fault is triggered, force the selection of conservative fixed parameters. As arbitration parameter vector If no fault is triggered, select the safety parameter vector. As arbitration parameter vector .

[0168] Specifically, this step aims to construct a high-priority security rollback arbitration mechanism that collaboratively monitors the security parameter vector from step S4100. and abnormal flag bits and the ZVS quality factor output in step S2300 This is to identify whether the sensing system has experienced an unforeseen serious malfunction, and in the event of a malfunction, to forcibly revert to preset conservative fixed parameters. Output arbitration parameter vector to ensure high reliability and security. .

[0169] In practice, this step implements a high-priority safety rollback arbitration mechanism, with the arbitrator continuously monitoring the anomaly flag. and collaboratively monitor the ZVS quality factor This arbitrator pursues high-performance safety parameter vectors. And preset, conservative fixed parameters that pursue high reliability A forced choice is made between them.

[0170] The arbitrator's decision is based on two independent fault signals: hard faults and soft faults, as follows: the first is a hard fault, namely an abnormal flag bit. A value of 1 indicates a control parameter vector. Exceeded the set of security constraints Secondly, there are soft faults, namely the ZVS quality factor. It continues to deteriorate and fails to converge through step S3200 within the preset time.

[0171] If any of the above fault conditions are met, the arbitrator will immediately discard the high-performance safety parameter vector for the absolute safety of the charging module. It forces the system to revert to a set of preset, safe, conservative fixed parameters. State, i.e., the arbitration parameter vector Set to conservative fixed parameters The arbitrator will only allow high-performance safety parameter vectors if it is confirmed that neither of the two aforementioned faults has occurred. Pass, i.e., arbitration parameter vector Set as a security parameter vector The conservative fixed parameters It is a set of parameters that, while potentially inefficient (e.g., forced to operate in hard switching or low power mode), ensure absolute hardware safety.

[0172] Step S4300, the arbitration parameter vector Fixed-point processing into hardware control words Based on the pulse width modulation (PWM) synchronization trigger signal SyncTrigger, at the start of the next cycle, the hardware control word... Update to hardware registers.

[0173] Specifically, this step aims to transform the arbitration parameter vector output from step S4200 into... This achieves secure, stable, and synchronous application of the arbitration parameter vector to the power electronic hardware in physical time. Its purpose is to ensure that the arbitration parameter vector is applied at the moment the pulse width modulation (PWM) synchronization trigger signal SyncTrigger arrives. Convert to the corresponding hardware control word It is updated atomically to the pulse width modulation (PWM) register to avoid dynamic loop oscillations or abnormal switching noise caused by sudden changes in parameters within a single PWM cycle.

[0174] In the specific implementation process, this step will use the floating-point format arbitration parameter vector After being processed by fixed-point conversion, it is converted into a hardware control word that can be recognized by a pulse width modulation (PWM) controller. To prevent sudden parameter changes, especially due to the safety rollback arbitration mechanism in step S4200, conservative fixed parameters are set. Arbitration parameter vector The dynamic oscillations triggered during the switching cycle are addressed in this step using a hardware triggering mechanism or a pulse width modulation (PWM) period synchronization signal, SyncTrigger, to ensure the hardware control word... The synchronization start point of the next cycle is atomically written to the hardware register only after the current pulse width modulation (PWM) cycle ends.

[0175] Example 2:

[0176] This embodiment, based on Embodiment 1, provides a parameterized digital control system suitable for charging modules with a wide power range, such as... Figure 2 As shown, it includes a baseline prediction module, a real-time perception module, an adaptive correction module, and a security arbitration module;

[0177] The baseline prediction module is used to acquire raw sensor signals from the charging module to construct a slow-operating-condition feature vector. Driving the physical constraint benchmark prediction model Generate baseline behavior parameter vector For baseline behavior parameter vector Perform analysis to generate parameters including predicted PWM parameters and predicted loop parameters The second baseline behavior parameter vector .

[0178] The real-time sensing module is used to utilize the gate-drain capacitance. Coupling effect monitoring gate drive current To capture transient noise fingerprints and generate digital noise sequences Based on digital noise sequence Extracting transient eigenvalues And based on transient eigenvalues Compared to the preset perfect ZVS threshold and lost ZVS threshold Perform multi-level comparisons and map them to the ZVS quality factor. .

[0179] The adaptive correction module is used to fuse slow-working-condition feature vectors. and ZVS quality factor To construct a fusion working condition vector Through adaptive prediction models For fusion operating condition vector Perform inference to generate the correction vector Combined with the second baseline behavior parameter vector To generate control parameter vectors Based on actual ZVS quality factor Feedback updates adaptive prediction model First adaptive prediction weights To generate the second adaptive prediction weights .

[0180] The security arbitration module is used to arbitrate based on the security constraint set. and control parameter vector Generate security parameter vector and abnormal flag bits Based on the abnormal flag bit and ZVS quality factor From the security parameter vector Or conservative fixed parameters Select to generate arbitration parameter vector , arbitration parameter vector The data is fixed-point and updated to the hardware registers based on the PWM synchronous trigger signal.

[0181] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0182] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A parameterized digital control method suitable for wide power range charging modules, characterized in that, The method comprises the following steps: The original sensor signals of the charging module are collected to construct a slow working condition feature vector, and a physical constraint reference prediction model is driven to generate a baseline behavior parameter vector. The generation method of the second baseline behavior parameter vector comprises the following steps: The original sensor signals of the charging module are collected to extract a steady-state input voltage, a steady-state output current and a steady-state device temperature, and an adaptive-temperature drift compensation normalization process is performed according to the temperature offset of the steady-state device temperature relative to the reference temperature to generate an input voltage feature, an output current feature and a temperature feature respectively, and then the slow working condition feature vector is generated by combination. The slow working condition feature vector is input into the physical constraint reference prediction model for inference to predict a baseline behavior parameter vector containing a pulse width modulation (PWM) target and a loop behavior target. The PWM target is extracted to generate a predicted PWM parameter, and the loop behavior target is extracted to combine a controlled object small signal model to form a predicted loop parameter, and then the second baseline behavior parameter vector is generated by combination of the predicted PWM parameter and the predicted loop parameter. The method for forming the predicted loop parameter comprises the following steps: A controlled object small signal model based on the slow working condition feature vector is constructed. A transfer function containing to-be-solved proportional gain and integral gain is constructed, and the product of the transfer function and the controlled object small signal model is defined as an open-loop transfer function. According to the predicted target closed-loop bandwidth and the predicted target phase margin contained in the loop behavior target, the amplitude constraint condition and the phase constraint condition of the open-loop transfer function are established. The equation set of the amplitude constraint condition and the phase constraint condition is solved to calculate the proportional gain and the integral gain, and the predicted loop parameter is formed based on the proportional gain and the integral gain. The coupling effect of the gate-drain capacitance is used to monitor the gate drive current to capture the transient noise fingerprint and generate a digitized noise sequence, the transient feature value is extracted based on the digitized noise sequence, and the transient feature value is compared with the preset perfect ZVS threshold and the lost ZVS threshold in multiple levels to map the ZVS quality factor; The slow working condition feature vector and the ZVS quality factor are fused to construct a fusion working condition vector, the fusion working condition vector is inferred by an adaptive prediction model to generate a correction vector, the second baseline behavior parameter vector is combined to generate a control parameter vector, the first adaptive prediction weight of the adaptive prediction model is updated based on the actual ZVS quality factor feedback to generate a second adaptive prediction weight; According to the safety constraint set and the control parameter vector, a safety parameter vector and an abnormal flag are generated, the abnormal flag and the ZVS quality factor are selected from the safety parameter vector or the conservative fixed parameter to generate an arbitration parameter vector, and the arbitration parameter vector is fixed-pointed and updated to a hardware register according to the PWM synchronous trigger signal.

2. The parameterized digital control method for a wide power range charging module according to claim 1, wherein, The execution method of the adaptive-temperature drift compensation normalization process comprises the following steps: The reference normalization lower boundary and the reference normalization upper boundary are corrected by using the temperature offset and the preset boundary temperature coefficient, and the dynamic normalization lower boundary function and the dynamic normalization upper boundary function are calculated respectively. The reference normalized lower boundary and the reference normalized upper boundary are respectively the minimum normalized boundary value and the maximum normalized boundary value of the steady-state input voltage or the steady-state output current calibrated at the reference temperature; The steady-state input voltage and the steady-state output current are linearly normalized by using the dynamic normalized lower boundary function and the dynamic normalized upper boundary function, so as to obtain the input voltage feature and the output current feature.

3. The parameterized digital control method for a wide power range charging module according to claim 1, wherein, The amplitude constraint condition is that the amplitude of the open-loop transfer function is equal to 1 at the predicted target closed-loop bandwidth; The phase constraint condition is that the sum of the phase angle of the open-loop transfer function and the phase angle of the small-signal model of the controlled object is equal to the sum of the critical unstable point -180° at the predicted target closed-loop bandwidth and the preset predicted target phase margin.

4. The parameterized digital control method for a wide power range charging module of claim 1, wherein, The mapping method of the ZVS quality factor comprises: In a preset switching event sampling window, the gate drive current is sampled and digitized to extract the transient noise fingerprint in the gate drive current to generate a digitized noise sequence; Based on the digitized noise sequence, the transient noise fingerprint component is extracted, the transient noise fingerprint component is squared in the switching event sampling window to obtain the instantaneous energy, and the instantaneous energy is accumulated to generate a transient feature value; According to the preset perfect ZVS threshold and the lost ZVS threshold, the transient feature value is subjected to multi-level threshold comparison to divide into perfect zero voltage switching ZVS, boundary zero voltage switching ZVS, and lost zero voltage switching ZVS, and is mapped into a corresponding ZVS quality factor.

5. The parameterized digital control method for a wide power range charging module according to claim 4, characterized in that, The division method of the perfect zero voltage switching ZVS, the boundary zero voltage switching ZVS, and the lost zero voltage switching ZVS comprises: The perfect ZVS threshold and the lost ZVS threshold are set, and the perfect ZVS threshold is limited to be smaller than the lost ZVS threshold; If the transient feature value does not exceed the perfect ZVS threshold, it is determined that the current state is perfect zero voltage switching ZVS, and the value of the ZVS quality factor is mapped to 0; If the transient feature value exceeds the perfect ZVS threshold and does not exceed the lost ZVS threshold, it is determined that the current state is boundary zero voltage switching ZVS, and the value of the ZVS quality factor is mapped to 1; If the transient feature value exceeds the lost ZVS threshold, it is determined that the current state is lost zero voltage switching ZVS, and the value of the ZVS quality factor is mapped to 2.

6. The parameterized digital control method for a wide power range charging module of claim 1, wherein, The generation method of the second adaptive prediction weight comprises: Based on the slow working condition feature vector, the ZVS quality factor is combined to perform fusion in the information dimension through a feature splicing operation to construct a fusion working condition vector; The fusion working condition vector is input into the adaptive prediction model, the first adaptive prediction weight of the adaptive prediction model is used to infer the fusion working condition vector, and a correction amount vector is output; The second baseline behavior parameter vector and the correction amount vector are added element by element to generate a control parameter vector; Based on the fusion working condition vector and the correction amount vector, the first adaptive prediction weight is updated by minimizing the loss function between the actual ZVS quality factor and the target ZVS quality factor to generate the second adaptive prediction weight.

7. The parameterized digital control method for a wide power range charging module according to claim 1, wherein, The generation method of the arbitration parameter vector comprises: According to a safety constraint set defined by a hardware minimum safety boundary vector and a hardware maximum safety boundary vector, a hardware physical limit constraint is imposed on the control parameter vector by a clipping function to generate a safety parameter vector and an abnormal flag bit; The abnormal flag bit and the ZVS quality factor are cooperatively monitored to determine whether a hard fault or a soft fault is triggered, and if the fault is triggered, a conservative fixed parameter is selected as the arbitration parameter vector, and if the fault is not triggered, the safety parameter vector is selected as the arbitration parameter vector; The arbitration parameter vector is fixed-point processed into a hardware control word, and the hardware control word is updated to a hardware register at a synchronous starting point of a next cycle according to a pulse width modulation (PWM) synchronous trigger signal SyncTrigger.

8. Parameterized digital control system for a wide power range charging module for implementing the parameterized digital control method for a wide power range charging module according to any one of claims 1 to 7, characterized in that, The system comprises a baseline prediction module, a real-time sensing module, an adaptive correction module and a safety arbitration module; The baseline prediction module is configured to collect original sensor signals of the charging module to construct a slow working condition feature vector, drive a physical constraint reference prediction model to generate a baseline behavior parameter vector, and analyze the baseline behavior parameter vector to generate a second baseline behavior parameter vector comprising a predicted PWM parameter and a predicted loop parameter; The real-time sensing module is configured to monitor a gate drive current by using a coupling effect of a gate leakage capacitor, capture a transient noise fingerprint and generate a digitized noise sequence, extract a transient feature value based on the digitized noise sequence, and perform multi-level comparison between the transient feature value and preset perfect ZVS threshold and lost ZVS threshold to map a ZVS quality factor; The adaptive correction module is configured to fuse the slow working condition feature vector and the ZVS quality factor to construct a fusion working condition vector, infer the fusion working condition vector by an adaptive prediction model to generate a correction amount vector, combine the second baseline behavior parameter vector to generate a control parameter vector, and update a first adaptive prediction weight of the adaptive prediction model based on an actual ZVS quality factor feedback to generate a second adaptive prediction weight; The safety arbitration module is configured to generate a safety parameter vector and an abnormal flag bit according to a safety constraint set and a control parameter vector, select from the safety parameter vector or a conservative fixed parameter according to the abnormal flag bit and the ZVS quality factor to generate an arbitration parameter vector, fix-point the arbitration parameter vector, and update the arbitration parameter vector to a hardware register according to a PWM synchronous trigger signal.

Images