Power electronic device precision regulation device and method based on multilayer perceptron

Abstract

The application is a precision regulation device and method for power electronic devices based on multilayer perceptron, which comprises a multi-channel data acquisition module, a timing feature extraction module, a multilayer perceptron, a data correction module, a RAM module comprising a RAM control unit and a dual-port RAM unit, and a data interface module. The method first builds a multilayer perceptron, uses multi-working-condition samples of power electronic devices with multi-dimensional hardware and scheme differences to complete training; within a single time window, the to-be-corrected output, the standard output and the excitation are synchronously collected at fixed intervals to form a data pair, and statistical features such as mean and standard deviation are extracted; after the data pair and the statistical features are compounded, they are input into the network to obtain an error compensation, and the correction is output through linear operation. The application can fit the nonlinear correlation between multi-scene excitation and output error, and can consider both instantaneous accuracy and long-term stability. The device supports parameter configuration and has strong portability, and the output accuracy of the power electronic device is improved.

Description

Technical Field

[0001] This invention belongs to the field of power electronic device output accuracy optimization technology, specifically relating to a power electronic device accuracy control device and method based on a multilayer sensor. Background Technology

[0002] Various converters, inverters, active filters, static var compensators, solid-state transformers, AC / DC conversion modules, power supplies and energy storage, magnet power supplies, and power amplifiers are widely used in industrial power supply and distribution, new energy grid connection, computing centers, nuclear physics experiments, isotope devices, accelerator devices, and nuclear energy devices. During actual operation, power electronic devices are susceptible to dynamic drift in output parameters due to various excitation factors such as temperature, load conditions, circuit topology differences, component aging, and changes in heat dissipation conditions. This results in deviations in output values ​​and reduces the accuracy of the device's output.

[0003] Existing correction schemes mostly employ fixed-coefficient linear compensation, which struggles to fit the complex nonlinear relationship between excitation and output error. Adding dedicated compensation devices significantly increases hardware costs and wiring complexity. Furthermore, traditional correction methods rely solely on single-point instantaneous sampling data, with compensation parameters often being factory-fixed, failing to dynamically and adaptively adjust to power electronic devices with multi-dimensional hardware and scheme differences, or to adapt to real-time operating conditions. In scenarios involving high and low temperatures, light and heavy loads, long-term operation, and complex electromagnetic environments, output drift and deviations accumulate continuously, resulting in not only large instantaneous output deviations but also poor long-term operational stability.

[0004] Existing technologies cannot simultaneously meet the requirements of low cost, adaptability to power electronic devices with multi-dimensional hardware and solution differences, adaptive nonlinear correction under all operating conditions, and long-term operational stability. It is difficult to achieve synergistic optimization of instantaneous accuracy and long-term stability of power electronic devices. Therefore, there is an urgent need for a power electronic device accuracy optimization scheme that features adaptive nonlinear control, low hardware modification cost, adaptability to multi-dimensional hardware and solution differences, and applicability to multiple operating conditions. Summary of the Invention

[0005] This invention provides a precision control device for power electronic devices based on a multilayer perceptron, comprising: a multi-channel data acquisition module for synchronously acquiring multiple sets of data pairs within a preset time period, each set of data pairs including an output quantity to be corrected, a standard output quantity, and at least one excitation quantity acquired at the same time; the multi-channel data acquisition module integrates multiple types of data interfaces to adapt to various communication protocols and various drive signals to complete data acquisition; a time window buffering and feature extraction module for buffering the multiple sets of data pairs acquired within a single time window and extracting multiple statistical features based on the buffered multiple sets of data pairs; a multilayer perceptron for receiving the input multiple sets of data pairs and the multiple statistical features, and outputting an output quantity error compensation quantity; and a data correction module for uniformly correcting all output quantities to be corrected within the current acquisition cycle using the output quantity error compensation quantity to obtain a corrected output quantity; the device is configured with a general data interface and supports parameter configuration.

[0006] According to one embodiment of the present invention, it further includes: a RAM module for storing and outputting the data sequence of the corrected output amount in sequence; and a data interface module for outputting the corrected output amount to a back-end device.

[0007] According to one embodiment of the present invention, the RAM module includes a RAM control unit and a dual-port RAM unit; the RAM control unit is used to generate read / write control timing, allocate storage addresses, switch read / write enable signals, schedule the output of the data sequence of the corrected output, and alternately store it into the dual-port RAM unit through a ping-pong operation; the dual-port RAM unit is provided with two sets of independent read / write ports for buffering the corrected output, one set of ports receives the write data stream of the corrected output, and the other set of ports reads out its corresponding data sequence.

[0008] This invention also provides a method for precision control of power electronic devices based on a multilayer perceptron, comprising: synchronously acquiring multiple sets of data pairs within a preset time period, each set of data pairs including an output quantity to be corrected, a standard output quantity, and at least one excitation quantity acquired at the same time; buffering the multiple sets of data pairs within a single time window, and extracting multiple statistical features based on the multiple sets of data pairs; inputting the multiple sets of data pairs and the multiple statistical features into the multilayer perceptron to obtain an output quantity error compensation amount; and using the output quantity error compensation amount to uniformly correct all output quantities to be corrected within the current acquisition period to obtain a corrected output quantity.

[0009] According to one embodiment of the present invention, before inputting the multiple sets of data pairs and the multiple statistical features into the multilayer perceptron, the method further includes the step of constructing the multilayer perceptron: configuring the input layer, hidden layer, and output layer of the multilayer perceptron; the input layer is configured to receive composite input information output by a single time window, the composite input information including a sequence of original data pairs consisting of an output quantity to be corrected, a standard output quantity, and at least one excitation quantity, and multiple statistical features calculated from the sequence of original data pairs in the same time window; the hidden layer is configured to fit the nonlinear correlation between the excitation quantity and the output quantity error of a power electronic device with multi-dimensional hardware and scheme differences under changing operating conditions using a nonlinear activation function; the output layer is configured to output the output quantity error compensation quantity corresponding to the time window.

[0010] According to one embodiment of the present invention, before inputting the multiple sets of data pairs and the multiple statistical features into the multilayer perceptron, the step of training the multilayer perceptron is further included: collecting raw data of power electronic devices with multi-dimensional hardware and scheme differences under varying operating conditions for each time window, each set of raw data including the output quantity to be corrected, the standard output quantity, and at least one excitation quantity, and summing all data to form a training sample set; using the training sample set to generate composite input information and sending it into the multilayer perceptron, using the standard output quantity as the truth benchmark, and combining the composite input information to update the network parameters of the multilayer perceptron, gradually reducing the output quantity error until the multilayer perceptron converges.

[0011] According to one embodiment of the present invention, the step of extracting multiple statistical features based on the multiple sets of data pairs further includes: calculating multiple statistical parameters that reflect the amplitude of the output quantity to be corrected and characterize the change law of the output quantity to be corrected and at least one excitation quantity data based on the multiple sets of data pairs within the same time window, thereby constituting the multiple statistical features.

[0012] According to one embodiment of the present invention, the plurality of statistical characteristics include at least one of the following: mean, standard deviation, maximum value, minimum value, range, difference between the last value and the mean, and endpoint slope.

[0013] According to one embodiment of the present invention, the step of uniformly correcting all output quantities to be corrected within the current acquisition cycle using the output quantity error compensation amount to obtain the corrected output quantity further includes: performing linear operations on each output quantity to be corrected within the current acquisition cycle with the output quantity error compensation amount to obtain the corrected output quantity.

[0014] According to one embodiment of the present invention, the step of synchronously acquiring multiple sets of data pairs within a preset time period further includes: synchronously acquiring the output quantity to be corrected, the standard output quantity, and the at least one excitation quantity at the same moment according to a fixed sampling interval within a preset time period corresponding to a single time window, thereby forming multiple sets of data pairs.

[0015] This invention abandons the traditional fixed-coefficient linear correction mode. It relies on a multilayer perceptron to accurately fit the nonlinear correlation between excitation changes and output errors in power electronic devices with multidimensional hardware and solution differences under varying operating conditions. This correlation is caused by different circuit topologies, components, sensors, process structures, water cooling, controllers, control strategies, loads, and operating environments. It is adaptable to various power electronic devices with multidimensional hardware and solution differences and complex operating conditions, effectively suppressing output deviations under extreme scenarios such as high and low temperatures and full loads. It simultaneously optimizes the instantaneous output accuracy and long-term operational stability of the device, comprehensively improving the device's output accuracy. This invention has low hardware modification costs, requiring no auxiliary detection devices. Dynamic error compensation can be completed solely through data collected during normal device operation, simplifying wiring and reducing hardware material input. Simultaneously, the device is equipped with a universal data interface, offering strong configurability and portability. It can be integrated into microprocessors or programmable logic devices, and the multilayer perceptron network parameters can adapt to differences in hardware, process, heat dissipation, controllers, control strategies, loads, and operating environments. It supports dynamic interface configuration and optimization iteration, continuously ensuring the device's accurate output performance.

[0016] This invention trains a multilayer perceptron using power electronic devices with multidimensional hardware and solution differences, and samples from multiple operating conditions and scenarios. This enables the perceptron to accurately identify operating condition disturbance errors and generate error compensation in real time, achieving adaptive dynamic correction of the output. Simultaneously, it overcomes the limitations of traditional single-point sampling correction by calculating errors based on the multidimensional statistical characteristics of multiple sets of data within a single time window, significantly improving the comprehensiveness and accuracy of compensation. Combined with a built-in RAM module and a dual-port RAM ping-pong caching mechanism, it can stably cache and output high-precision data sequences in a time-division manner. The device's universal data interface enables parameter configuration, adapting to the operational needs of various main control devices. Its strong versatility and adaptability make it widely applicable to various power conversion devices and high-precision, high-stability special power supply devices, possessing extremely high engineering application value and promising prospects for widespread adoption. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a structural block diagram of the precision control device for power electronic devices based on a multilayer sensor provided by the present invention.

[0019] Figure 2 This is a flowchart illustrating the precision control method for power electronic devices based on a multilayer sensor provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the process for constructing a multilayer perceptron provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the training process of a multilayer perceptron provided by the present invention.

[0022] Figure label: 100: Precision control device for power electronic devices; 110: Multi-channel data acquisition module; 120: Time window buffer and feature extraction module; 130: Multilayer perceptron; 140: Data correction module; 150: RAM module; 151: RAM control unit; 152: Dual-port RAM unit; 160: Data interface module. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined Figures 1 to 4 The present invention describes a precision control device and method for power electronic devices based on a multilayer sensor.

[0025] This invention provides a precision control device for power electronic devices based on a multilayer sensor. Before detailing the technical solution of this invention, the term "precision" as used in the subject matter of this invention needs to be clearly explained. The precision mentioned in this application encompasses two technical aspects: the first is instantaneous precision, which refers to the deviation between the real-time output value of the device and the standard output value, characterizing the accuracy of a single output; the second is long-term stability, which refers to the ability of the device to maintain high instantaneous output precision even after long-term continuous operation, effectively suppressing output drift caused by factors such as device aging, temperature changes, load fluctuations, and heat dissipation differences. This invention simultaneously optimizes instantaneous precision and long-term stability through a multilayer sensor nonlinear compensation mechanism, thereby improving the overall output precision of the power electronic device.

[0026] Figure 1This is a structural block diagram of the precision control device for power electronic devices based on a multilayer sensor provided by the present invention. Figure 1 As shown, the precision control device 100 for power electronic devices based on a multilayer sensor provided by the present invention includes: The multi-channel data acquisition module 110 is used to synchronously acquire multiple sets of data pairs within a preset time period. Each set of data pairs includes the output quantity to be calibrated, the standard output quantity, and at least one excitation quantity acquired at the same time. The multi-channel data acquisition module integrates multiple types of data interfaces to adapt to various communication protocols and various drive signals to complete data acquisition. The time window buffering and feature extraction module 120 is used to cache multiple sets of data pairs collected within a single time window and extract multiple statistical features based on the cached sets of data pairs. The multilayer perceptron 130 is used to receive multiple sets of input data pairs and multiple statistical features, and output the output quantity error compensation quantity; The data correction module 140 is used to uniformly correct all output quantities to be corrected within the current acquisition cycle using the output quantity error compensation amount, so as to obtain the corrected output quantity. The device is equipped with a general data interface and supports parameter configuration.

[0027] Specifically, the multi-channel data acquisition module 110 performs synchronous data acquisition at fixed sampling intervals within a preset duration corresponding to a single time window. Each set of acquired data includes the output quantity to be calibrated, the standard output quantity, and the sampling signal corresponding to at least one excitation quantity acquired at the same time. The multi-channel data acquisition module integrates multiple types of data interfaces, is compatible with various industrial communication protocols and different specifications of drive signals, and can directly connect to various sensors and power circuit sampling units without additional adapter boards, reducing external accessories and simplifying the equipment wiring structure. In this embodiment, the output quantity to be calibrated is the sampling voltage obtained by electrical parameter conversion from the actual operation of the power electronic device; the standard output quantity is the theoretical standard electrical parameter value of the device under rated operating conditions and corresponding operating state, serving as a reference for error calculation; the excitation quantity is mainly temperature, characterized by the voltage division output voltage of a negative temperature coefficient thermistor. The multi-channel data acquisition module completes data acquisition based on the existing sampling circuit, without the need for additional temperature measurement hardware, simplifying the overall wiring structure.

[0028] The time window buffering and feature extraction module 120 mainly undertakes two tasks: data caching and feature calculation. First, it completely caches all multiple sets of data pairs in a single time window, including the output to be corrected, the standard output, and the excitation, output from the multi-channel data acquisition module 110, to avoid data loss or timing errors and ensure the timing matching of the three types of synchronous data. Then, it retrieves the cached data pairs and performs calculations to obtain at least one type of statistical parameter, such as mean, standard deviation, maximum value, minimum value, range, difference between the last value and the mean, and endpoint slope. It summarizes and integrates these parameters into multiple statistical features, which, together with the original data pairs, constitute composite input information, providing complete data support for the multilayer perceptron to accurately fit error patterns and identify operating condition disturbances.

[0029] The Multilayer Perceptron 130 is a neural network model used for offline architecture construction and multi-condition training. It can stably receive raw data pairs consisting of the output to be corrected, the standard output, and the excitation quantity, as well as composite input information composed of multiple statistical features. Using the standard output as a benchmark, it accurately compares the deviation characteristics between the actual output and the standard output. It uses a nonlinear activation function to fit the nonlinear correlation between the multi-dimensional differences and output errors caused by various operating condition disturbances, different circuit topologies, components, sensors, process structures, water cooling, controllers, control strategies, loads, and operating environments. Combined with the optimal network parameters learned during model training, it calculates and outputs the output error compensation amount corresponding to the time window in real time. This module has a simplified architecture and low computational load, and can be deployed on low-cost hardware such as microcontrollers and domestic FPGAs to achieve real-time inference.

[0030] The data correction module 140 receives the output error compensation amount from the multilayer sensor 130. Using the synchronously acquired standard output amount as a benchmark, it performs linear calculations on all outputs to be corrected within the current acquisition cycle, using the same compensation amount to complete the unified correction of the entire cycle's sampled data. Finally, it obtains the corrected output amount after eliminating temperature drift, operating condition deviation, and device aging bias. This correction method achieves precise calibration based on the standard output amount benchmark, eliminating the need for point-by-point calculation of compensation parameters, effectively reducing the hardware computational load, and avoiding data jitter problems caused by single-point calculations, thus ensuring the stability and accuracy of the sampling results.

[0031] This device is equipped with a universal data interface, through which staff can configure the device parameters and flexibly adjust operating parameters such as sampling period, network operation coefficient, and correction gain according to different hardware architectures, loads, and operating environments to meet the control needs of various differentiated power electronic equipment.

[0032] According to an embodiment of the present invention, the precision control device 100 for power electronic devices based on a multilayer sensor provided by the present invention further includes: RAM module 150 is used to store and output data sequences of corrected output values ​​sequentially. The data interface module 160 is used to output the corrected output to the backend device.

[0033] Specifically, the RAM module 150 and the data interface module 160 work together to cache, time-series organize, and provide diversified external outputs of the corrected output from the data correction module 140, meeting the data reading needs of different host computers and back-end devices. The RAM module 150 specifically receives the corrected output from the data correction module 140, sequentially storing the data according to the sampling time sequence to form a continuous data sequence. This achieves batch caching and persistent storage of sampled data, solving the problems of easy loss of single-point instantaneous data, timing disorder, and inability to reproduce waveforms. Simultaneously, the RAM module 150 can read the stored corrected output data sequence completely and in a time-division manner according to external commands, adapting to high-precision detection and operational analysis scenarios requiring continuous waveform data and batch time-series data.

[0034] The data interface module 160 is responsible for the real-time transmission of corrected output values. It can extract instantaneous single-point corrected output values ​​from the cached data in real time and output them externally, adapting to application scenarios that only require single-point instantaneous data, such as conventional closed-loop control and real-time sampling monitoring. The RAM module 150 focuses on batch cached output of time series, while the data interface module 160 focuses on real-time single-point fast output. The two modules have clear division of labor and work together to enable the device to adapt to both time series waveform analysis and real-time control scenarios. The output configuration is flexible and highly versatile, further improving the portability and engineering adaptability of the entire precision control device.

[0035] According to an embodiment of the present invention, the RAM module 150 includes a RAM control unit 151 and a dual-port RAM unit 152; RAM control unit 151 is used to generate read / write control timing, allocate storage addresses, switch read / write enable signals, schedule the data sequence of output correction output, and alternately store it into dual-port RAM unit 152 through ping-pong operation. The dual-port RAM unit 152 has two sets of independent read and write ports for caching correction outputs. One set of ports receives the write data stream of the correction outputs, and the other set of ports reads out the corresponding data sequence.

[0036] Specifically, the RAM control unit 151 is the logical scheduling center of the entire RAM module 150, coordinating all data read / write and caching operations of the dual-port RAM unit 152. This unit strictly follows the overall operating rhythm of the device, generating precise standard read / write control timing sequences, rationally allocating dedicated storage addresses according to the sampling timing sequence, and switching read / write enable signals in real time to precisely control the data writing and reading states. Through a ping-pong storage scheduling mechanism, this unit alternately caches the corrected output data sequence from the data correction module 140 into the dual-port RAM unit 152. Simultaneously, based on external read commands from the backend device, it can precisely retrieve the complete data sequence already cached in the dual-port RAM unit 152, completing ordered output scheduling and ensuring the regularity of timing data storage and output.

[0037] The dual-port RAM unit 152 is configured with two sets of independent, non-interfering read / write hardware ports, supporting parallel and synchronous execution of data writing and reading. One set of ports serves as the write port, connecting to the front-end data correction module 140 to continuously receive and cache the corrected output data stream in real time; the other set of ports serves as the read port, connecting to back-end control and detection equipment to stably and continuously read out the cached time-series data sequence. Based on the parallel read / write characteristics and ping-pong caching mechanism of the dual-port RAM unit 152, the device can simultaneously output historical time-series data while continuously acquiring and writing new corrected data. This effectively overcomes the limitations of traditional fixed-coefficient single linear compensation, ensuring continuous, stable, and uninterrupted transmission of the corrected output, and significantly improving the real-time performance and reliability of the entire control device.

[0038] Combination Figure 1 The device structure diagram shown includes an external parameter input interface for both the time window buffer and feature extraction module 120 and the multilayer perceptron 130. These interfaces provide on-site configurability and dynamic adjustability. Specific functions are as follows: For the multilayer perceptron 130, this invention includes a fitting effect evaluation mechanism to determine in real-time whether the model fitting accuracy meets preset requirements. When the model fitting effect is found to be below the preset accuracy threshold, configuration parameters can be input to the multilayer perceptron 130 through the external parameter input interface to dynamically adjust important hyperparameters such as network learning rate, iteration count, and loss function weights. Simultaneously, the single time window acquisition duration, fixed sampling interval, and types of statistical parameters can be flexibly configured, enabling configurable parameters throughout the data acquisition, feature extraction, and model inference process. This proactively intervenes in the model training and inference process, continuously optimizing nonlinear fitting capabilities until the multilayer perceptron's fitting effect to operating condition disturbances and output drift meets the preset accuracy requirements, ensuring the device's long-term accurate calibration capability.

[0039] For the time window buffering and feature extraction module 120, external input parameters can be used to flexibly configure the single time window acquisition duration, fixed sampling interval, and types of statistical parameters to be enabled. This allows for dynamic adjustment of data caching rules and feature extraction dimensions based on differences in on-site temperature conditions, load levels, and equipment types. By adaptively configuring the time window and feature parameters, it can adapt to the operating characteristics of different power electronic devices, further improving the error identification accuracy and correction versatility of the device under multiple scenarios and extreme operating conditions.

[0040] This invention, through its dual-module configurable external parameter design, enables the overall control device to break free from the limitations of traditional factory-fixed parameters, possessing strong adaptability and portability to various working conditions, and continuously optimizing the device's instantaneous output accuracy and long-term operational stability.

[0041] Figure 2 This is a flowchart illustrating the precision control method for power electronic devices based on a multilayer sensor provided by the present invention. Figure 2 As shown, the precision control method for power electronic devices based on multilayer perceptrons provided by the present invention includes: In step S100, multiple sets of data pairs are synchronously acquired within a preset time period. Each set of data pairs includes the output quantity to be corrected, the standard output quantity, and at least one excitation quantity acquired at the same time. In step S200, multiple sets of data pairs within a single time window are cached, and multiple statistical features are extracted based on the multiple sets of data pairs; In step S300, multiple sets of data pairs and multiple statistical features are input into the multilayer perceptron to obtain the output error compensation amount; In step S400, the output error compensation amount is used to uniformly correct all output quantities to be corrected within the current acquisition cycle to obtain the corrected output quantity.

[0042] Specifically, during step S100, a preset duration corresponds to a single time window. Multiple data pairs are synchronously collected and generated according to a fixed sampling interval. Each data pair includes the output quantity to be corrected, the standard output quantity, and the sampling signal corresponding to at least one excitation quantity collected at the same time. The output quantity to be corrected is the sampling voltage obtained after conversion of various electrical parameters. In this embodiment, taking the current signal sampling scenario as an example, it corresponds to the current sampling voltage. The standard output quantity is the theoretical standard electrical parameter value under the corresponding operating condition, serving as a benchmark for error fitting and correction. The excitation quantity may include various operating condition parameters that affect sampling accuracy, such as temperature, current, voltage, frequency, phase, and amplitude. In this embodiment, temperature is selected as the main excitation quantity, and the voltage divider output voltage of the negative temperature coefficient thermistor is used as the corresponding sampling signal to characterize the real-time temperature state. The negative temperature coefficient thermistor has the inherent electrical characteristic that its resistance changes regularly with the ambient temperature. Under fixed power supply conditions, the voltage divider value of the component changes accordingly with the temperature. Therefore, the real-time temperature of the environment in which the device is located can be indirectly characterized by collecting the voltage divider output voltage, without the need for additional independent temperature measuring devices. In practical engineering, negative temperature coefficient thermistors are placed close to the current sampling components to avoid temperature interference from other areas of the equipment. The data acquisition object of this invention is a high-precision standardized accelerator power supply operating in constant current mode. A high-precision digital multimeter is used in conjunction with a current sensor to complete the acquisition of output current data. The selected current sensor has a transformation ratio of 1333.33A / V, which has excellent cost-effectiveness and detection accuracy. The experiment strictly covers a wide temperature range of 30℃ to 80℃, and sets 11 levels of current load conditions of 1A, 5A, 10A, 15A, 20A, 25A, 30A, 35A, 40A, 45A, and 50A to achieve accurate time-series matching of the output quantity to be calibrated, the standard output quantity, and the temperature excitation quantity data, providing a high-quality sample basis for subsequent nonlinear fitting of the model.

[0043] When executing step S200, the time window buffer and feature extraction module completely buffers multiple sets of data pairs collected synchronously within the same time window to avoid temporal disorder and data loss. Then, based on the complete data pairs after temporal matching, the calculation is carried out to obtain at least one type of statistical parameter among the mean, standard deviation, maximum value, minimum value, range, difference between the last value and the mean, and endpoint slope. Multiple statistical features are integrated to generate multiple statistical features. Combined with the original data pairs, they constitute the composite input information of the multilayer perceptron, which breaks through the limitation of traditional single-point sampling that can only rely on instantaneous data correction.

[0044] Before executing step S300, offline construction and sample training of the multilayer perceptron must be completed. First, the network architecture is built, configuring the multilayer perceptron with an input layer, hidden layer, and output layer. The input layer receives composite input information corresponding to a single time window. This composite input information consists of raw data pairs comprising the output quantity to be corrected, the standard output quantity, and the excitation quantity, as well as multiple statistical features calculated within the same time window. The hidden layer uses a nonlinear activation function to fit the complex nonlinear relationship between various factors such as temperature, load fluctuations, device aging, and heat dissipation differences and the output quantity error, considering multi-dimensional hardware and scheme differences, load, operating environment, and operating condition changes. The output layer ultimately outputs the output quantity error compensation amount corresponding to the time window. During the offline training phase, measured data pairs from multiple time windows under different temperatures and load conditions are collected to construct a training sample set containing massive operating condition features. The composite information of the samples is input into the multilayer perceptron, using the standard output quantity as the true benchmark. The output deviation is iteratively compared with the composite input information, continuously updating the multilayer perceptron network parameters and constantly reducing the operating condition drift error of the output quantity to be corrected until the model converges and the fitting accuracy meets the preset threshold requirements. After offline training is complete, hyperparameters such as the network learning rate, number of iterations, and loss function weights can be dynamically adjusted via an external parameter interface to continuously optimize the model's fitting performance. Once the model converges, it enters the online inference phase, where multiple sets of data pairs and various statistical features collected in real time within the current time window are input into the multilayer perceptron to calculate the output error compensation amount for the current period in real time.

[0045] In step S400, using the synchronously acquired standard output as a benchmark, a unified linear correction calculation is performed on the error compensation amounts of all outputs to be corrected within this acquisition cycle and the outputs of the multilayer sensor. This eliminates output deviations caused by circuit topology differences, temperature drift, load fluctuations, device aging, and uneven heat dissipation, ultimately yielding a high-precision, high-stability corrected output. This method employs a single-time-window unified compensation correction logic, eliminating the need for iterative calculations on individual sampled data, significantly reducing computational overhead, while ensuring the consistency of output data throughout the entire cycle, simultaneously improving the instantaneous output accuracy and long-term operational stability of the power electronic device.

[0046] Figure 3 This is a schematic diagram of the process for constructing a multilayer perceptron provided by the present invention. Figure 3 As shown, according to an embodiment of the present invention, before inputting multiple sets of data pairs and multiple statistical features into the multilayer perceptron, the step of constructing the multilayer perceptron is further included: In step S311, the input layer, hidden layer, and output layer of the multilayer perceptron are configured; In step S312, the input layer is configured to receive composite input information from a single time window output. The composite input information includes a sequence of raw data pairs consisting of the output quantity to be corrected, the standard output quantity, and at least one excitation quantity, as well as multiple statistical features calculated from the raw data pairs sequence of the same time window. In step S313, the hidden layer is configured to fit the nonlinear correlation between the excitation quantity, the change in operating conditions, and the output error through a nonlinear activation function; In step S314, the output layer is configured to output the output error compensation amount corresponding to the time window.

[0047] Specifically, when executing step S311, the overall architecture of the multilayer perceptron is built, the hierarchical structure parameters of the input layer, hidden layer, and output layer are configured in sequence, and the neuron connection topology between adjacent layers is standardized to form a complete network basic framework.

[0048] When executing step S312, the composition of the input data of the input layer is clarified. The composite input information corresponding to a single time window includes two types of data. The first type is the sequence of original data pairs obtained synchronously, which includes the output quantity to be corrected, the standard output quantity, and the excitation quantity sampling signals. The second type is multiple statistical features obtained by calculation based on all the original data pairs in the same time window. The two types of data are integrated and sent to the input layer to complete the input.

[0049] When executing step S313, a nonlinear activation function is configured for the hidden layer. This function can comprehensively fit the complex nonlinear relationship between various variables and output errors, including different circuit topologies, components, sensors, process structures, water cooling, controllers, and control strategies, as well as the differentiated operating conditions under different loads and operating environments, for power electronic devices with multi-dimensional hardware and scheme differences. It also matches the temperature excitation quantity collected by the negative temperature coefficient thermistor in this embodiment. This allows for accurate modeling of the nonlinear mapping law between excitation quantity and output error in differentiated power electronic devices during dynamic operation. This overcomes the limitations of traditional fixed compensation models that cannot adapt to multi-dimensional differences and multi-scenario operating condition changes in devices, providing a reliable model calculation foundation for subsequent accurate generation of error compensation quantities and adaptive dynamic correction of output quantities.

[0050] When executing step S314, the output target of the output layer is limited, so that the output layer stably outputs the error compensation amount that matches the set of time series data for the current processing time window.

[0051] After completing the configuration of all layers, a multilayer perceptron adapted to the data input, nonlinear error fitting, and compensation parameter output requirements of this method can be built. In this embodiment, the multilayer perceptron adopts a three-layer simplified architecture, with only an input layer, a single hidden layer, and an output layer. The hidden layer is configured with 12 neurons. The overall model has a small number of parameters and low real-time inference computation load, and the program can be directly solidified and deployed online in low-cost embedded hardware such as microcontrollers and domestic field-programmable gate arrays (FPGAs).

[0052] Figure 4 This is a schematic diagram of the training process for a multilayer perceptron provided by the present invention. Figure 4 As shown, according to an embodiment of the present invention, before inputting multiple sets of data pairs and multiple statistical features into the multilayer perceptron, a step of training the multilayer perceptron is further included: In step S321, raw data of power electronic devices with multi-dimensional hardware and scheme differences are collected under different operating conditions for each time window. Each set of raw data includes the output quantity to be corrected, the standard output quantity, and at least one excitation quantity. All data are summarized to form a training sample set. In step S322, composite input information is generated using the training sample set and fed into the multilayer perceptron. The standard output is used as the true value benchmark, and the network parameters of the multilayer perceptron are updated in combination with the composite input information to gradually reduce the output error until the multilayer perceptron converges.

[0053] Specifically, during step S321, for power electronic devices with multi-dimensional hardware and solution differences, the system covers differentiated equipment and operating conditions with different circuit topologies, components, sensors, process structures, water cooling, controllers, control strategies, loads, and operating environments. It covers various operating circuits and actual operating conditions of the equipment, and simultaneously collects complete raw data pairs within multiple time windows corresponding to each operating condition. Each raw data pair includes the output quantity to be corrected, the excitation quantity sampling signal, and simultaneously collects a one-to-one matching standard output quantity as a truth reference. All time-series matched correlation data are summarized and integrated to build a complete training sample set. In this embodiment, the training samples cover various typical operating conditions such as low-temperature standby, rated load, and short-term overload, with a temperature acquisition range of 30℃ to 80℃. The rich operating condition samples can improve the model's generalization and fitting ability.

[0054] In step S322, statistical features are calculated based on the original data in the training sample set, and composite input information is generated. This composite input information is then imported into the multilayer perceptron that has completed its hierarchical construction. Using the standard output value inherent in the sample as the baseline truth value, the output error compensation amount of the multilayer perceptron is compared, and the internal network parameters of the multilayer perceptron are iteratively adjusted to continuously reduce the output deviation caused by various circuit and operating condition disturbances. When the fluctuation amplitude of the network parameters is lower than the preset threshold during continuous iteration, the multilayer perceptron is considered to have converged. After completing training with full-condition samples, the multilayer perceptron does not need to rely on the factory-fixed compensation parameters. Even if the device experiences aging drift during long-term operation, it can adaptively generate a suitable output error compensation amount by combining real-time collected multi-dimensional operating condition data, ensuring long-term calibration accuracy and allowing it to be directly used for subsequent online calibration calculations.

[0055] According to an embodiment of the present invention, extracting multiple statistical features based on multiple sets of data further includes: Based on multiple sets of data pairs within the same time window, various statistical parameters are calculated to reflect the amplitude of the output quantity to be corrected and to characterize the change pattern of the output quantity to be corrected and at least one excitation quantity, thus forming multiple statistical features.

[0056] Specifically, multiple sets of data pairs for time-series synchronization matching within a single time window are selected for centralized computation. Based on the time-series data of the output and excitation quantities to be corrected, which are continuously sampled within the time window, various statistical parameters are obtained. These statistical parameters include at least one of the following: mean, standard deviation, maximum value, minimum value, range, difference between the last value and the mean, and endpoint slope. The various statistical parameters obtained through screening and calculation are summarized and integrated to form multiple statistical features used to construct composite input information.

[0057] According to an embodiment of the present invention, a plurality of statistical features include at least one of the following: mean, standard deviation, maximum value, minimum value, range, difference between the last value and the mean, and endpoint slope.

[0058] Specifically, the seven statistical parameters mentioned above can comprehensively characterize the static characteristics and dynamic changes of sampled data within a single time window from multiple dimensions, including amplitude level, dispersion, fluctuation range, and rate of change. Among them, the mean reflects the average amplitude level of all sampled data within the time window; the standard deviation reflects the degree of dispersion and fluctuation of the entire set of sampled data, effectively characterizing data fluctuations caused by various operating condition disturbances; the maximum and minimum values ​​limit the upper and lower limits of the numerical values ​​of the sampled data within a single window; the range intuitively reflects the overall fluctuation range of the data through the difference between the maximum and minimum values; the difference between the final value and the mean characterizes the degree of deviation of the data at the end of the sampling period from the steady-state mean, accurately capturing the drift deviation at the end of the time period; and the slope at the endpoints describes the rate of data change at the beginning and end of the sampling interval, effectively reflecting the instantaneous trend of sudden changes in operating conditions.

[0059] In actual feature extraction, it is not necessary to use all seven types of parameters. Instead, a feature set can be constructed by adaptively selecting single or multiple statistical parameter combinations based on the operating conditions of the equipment, temperature fluctuations, load levels, and target compensation accuracy requirements. This flexibly adapts to the error fitting needs of different scenarios. This invention, through multi-dimensional statistical feature-assisted modeling, overcomes the limitations of traditional methods that rely solely on instantaneous single-point data correction. It effectively filters out random sampling noise and avoids interference from single-point abnormal data, significantly improving the model's accuracy and generalization ability in recognizing long-term device aging and slow temperature drift. It greatly enhances the multilayer perceptron's accuracy in recognizing subtle errors caused by operating condition drift, temperature disturbances, and device aging, providing comprehensive and reliable feature support for subsequent accurate fitting of error nonlinear relationships and calculation of optimal compensation amounts.

[0060] According to an embodiment of the present invention, the corrected output is obtained by uniformly correcting all outputs to be corrected within the current acquisition cycle using an output error compensation amount, further comprising: The corrected output is obtained by performing linear operations on each output quantity to be corrected within the current acquisition cycle and the output quantity error compensation quantity.

[0061] Specifically, a single time window acquisition cycle generates only one unique output error compensation value. This compensation value comprehensively fits the sampling drift deviation caused by circuit differences, temperature fluctuations, load offsets, and device aging within this cycle, accurately characterizing the overall output deviation characteristics under the current operating conditions. All output values ​​to be corrected that are synchronously acquired within the same acquisition cycle share this error compensation value for batch linear correction calculations, eliminating the need to calculate compensation parameters separately for each sampling point. During the correction process, linear correction calculations are performed on each group of output values ​​to be corrected within the time window against the output error compensation value, systematically offsetting the sampling offset deviation caused by operating condition disturbances, ultimately obtaining a high-precision, drift-free corrected output value.

[0062] This invention employs a single-window unified compensation and batch linear correction operational logic, which differs from the traditional point-by-point iterative correction method. This significantly reduces the real-time computational load on the main control chip and minimizes the resource consumption of embedded devices. It also effectively avoids parameter fluctuations and data jitter caused by point-by-point calculations, ensuring the consistency and stability of the entire cycle of sampled data. Simultaneously, the trained multilayer perceptron network parameters can be directly embedded in low-cost embedded hardware such as domestically produced FPGAs and microcontrollers. Leveraging the advantages of hardware parallel computing, it achieves high-speed real-time inference, continuously and stably outputting accurate output error compensation, meeting the high-frequency, high-precision real-time correction requirements of power electronic devices.

[0063] According to an embodiment of the present invention, the synchronous acquisition of multiple sets of data pairs within a preset time period further includes: Within a preset duration corresponding to a single time window, the output quantity to be calibrated, the standard output quantity, and at least one excitation quantity are synchronously collected at the same time according to a fixed sampling interval, forming multiple sets of data pairs.

[0064] Specifically, a single time window's acquisition duration and fixed sampling interval are pre-set. Within this single time window, synchronous acquisition is continuously performed in a uniform time sequence. At each acquisition moment, three types of time-series matching data are acquired simultaneously: the output quantity to be corrected during real-time operation of the power electronic device, the standard output quantity under the corresponding operating condition, and the excitation quantity sampling data used to characterize operating condition disturbances. In this embodiment, temperature is used as the primary excitation quantity. The real-time temperature state is characterized by acquiring the voltage divider output voltage of a negative temperature coefficient thermistor, achieving simultaneous acquisition and matching of electrical parameters and temperature conditions. The resistance value of the negative temperature coefficient thermistor changes regularly with ambient temperature. Under fixed power supply conditions, its voltage divider value has a stable correlation with temperature, accurately reflecting the real-time temperature condition of the device's sampling circuit without requiring additional independent temperature measurement hardware, thus simplifying the hardware architecture. The three types of data acquired at the same time are integrated into a complete data pair. After continuous sampling throughout the entire time window, multiple sets of time-aligned and operating condition-matched original data pairs are summarized, providing a high-precision and highly consistent raw data foundation for subsequent feature extraction and model fitting.

[0065] This solution utilizes the existing sampling loop of the device to collect the excitation signal corresponding to the temperature, eliminating the need for additional independent temperature measuring devices, effectively simplifying the equipment wiring structure and reducing overall hardware costs. Simultaneously, by extracting multi-dimensional statistical features from multiple sets of time-series data within a single time window, the dimensions of the model input features are expanded, significantly improving the multilayer perceptron's ability to identify and fit various operating condition disturbances such as temperature and load. The multilayer perceptron, fully trained with massive samples across multiple hardware and solution variations, operating conditions, and wide temperature ranges, can adapt to different circuit topologies, components, sensors, process structures, water cooling systems, controllers, control strategies, loads, and error characteristics under different operating environments. It breaks through the limitations of traditional fixed-coefficient single linear compensation and can dynamically optimize network hyperparameters based on external parameter input interfaces, adapting to long-term aging and drift of the equipment, and adaptively generating accurate output error compensation quantities.

[0066] This invention employs a single-time-window, whole-cycle unified correction method, significantly reducing the real-time computational load on the embedded main control processor and improving the real-time performance of the device. Simultaneously, it effectively offsets sampling deviations caused by temperature drift, significantly improving sampling and correction accuracy under extreme conditions such as low-temperature standby and full-load operation. This solution requires no hardware architecture modification; it can be implemented using existing hardware combined with embedded programming. It is adaptable to various power electronic device sampling circuits with multi-dimensional hardware and solution differences, and allows for online parameter adjustment via a universal data interface. It is compatible with low-cost embedded hardware deployments such as microcontrollers and domestically produced FPGAs. Leveraging the synergistic effect of multi-dimensional feature analysis and nonlinear fitting algorithms, this invention effectively suppresses sampling data fluctuations across the entire temperature range and under full load conditions, significantly improving the long-term stability and accuracy of the power electronic device's sampling output.

[0067] The device embodiments disclosed in this invention correspond one-to-one with the aforementioned accuracy control method embodiments and support each other. The working principles, execution processes and method steps of each functional module are matched and support each other. The device structure can effectively implement the various correction functions of this method. The relevant implementation principles and technical processes can be referred to each other, and will not be described in detail here.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A precision control device for power electronic devices based on a multilayer sensor, characterized in that, include: The multi-channel data acquisition module is used to synchronously acquire multiple sets of data pairs within a preset time period. Each set of data pairs includes the output quantity to be calibrated, the standard output quantity, and at least one excitation quantity acquired at the same time. The multi-channel data acquisition module integrates multiple types of data interfaces to adapt to various communication protocols and various driving signals to complete data acquisition. The time window buffering and feature extraction module is used to cache the multiple sets of data pairs collected within a single time window, and extract multiple statistical features based on the cached multiple sets of data pairs. A multilayer perceptron is used to receive the input multiple sets of data pairs and the multiple statistical features, and output the output quantity error compensation quantity; The data correction module is used to uniformly correct all output quantities to be corrected within the current acquisition cycle using the output quantity error compensation amount, so as to obtain the corrected output quantity. The device is equipped with a general data interface and supports parameter configuration.

2. The apparatus according to claim 1, characterized in that, Also includes: The RAM module is used to store and output the data sequence of the corrected output amount in sequence; The data interface module is used to output the corrected output to the backend device.

3. The apparatus according to claim 2, characterized in that, The RAM module includes a RAM control unit and a dual-port RAM unit; The RAM control unit is used to generate read / write control timing, allocate storage addresses, switch read / write enable signals, schedule the output of the data sequence of the corrected output, and alternately store it into the dual-port RAM unit through a ping-pong operation. The dual-port RAM unit has two sets of independent read / write ports for caching the correction output. One set of ports receives the write data stream of the correction output, and the other set of ports reads out its corresponding data sequence.

4. A method for precision control of power electronic devices based on multilayer perceptrons, characterized in that, include: Multiple sets of data pairs are collected synchronously within a preset time period. Each set of data pairs includes the output quantity to be calibrated, the standard output quantity, and at least one excitation quantity collected at the same time. Cache the multiple sets of data pairs within a single time window, and extract multiple statistical features based on the multiple sets of data pairs; The multiple sets of data pairs and the multiple statistical features are input into a multilayer perceptron to obtain the output error compensation amount. The output error compensation amount is used to uniformly correct all outputs to be corrected within the current acquisition cycle to obtain the corrected output amount.

5. The method according to claim 4, characterized in that, Before inputting the multiple sets of data pairs and the multiple statistical features into the multilayer perceptron, the method further includes the step of constructing the multilayer perceptron: Configure the input layer, hidden layer, and output layer of the multilayer perceptron; The input layer is configured to receive composite input information output by a single time window. The composite input information includes a sequence of raw data pairs consisting of the output quantity to be corrected, a standard output quantity, and at least one excitation quantity, as well as multiple statistical features calculated from the sequence of raw data pairs in the same time window. The hidden layer is configured to fit the nonlinear correlation between the excitation quantity and the output quantity error of a power electronic device with multi-dimensional hardware and scheme differences under changing operating conditions using a nonlinear activation function. The output layer is configured to output the error compensation amount corresponding to the time window.

6. The method according to claim 4, characterized in that, Before inputting the multiple sets of data pairs and the multiple statistical features into the multilayer perceptron, the method further includes a step of training the multilayer perceptron: Raw data of power electronic devices with multi-dimensional hardware and scheme differences are collected under varying operating conditions at different time windows. Each set of raw data includes the output quantity to be corrected, the standard output quantity, and at least one excitation quantity. All data are summarized to form a training sample set. The training sample set is used to generate composite input information, which is then fed into the multilayer perceptron. The standard output is used as the true value benchmark, and the network parameters of the multilayer perceptron are updated in combination with the composite input information to gradually reduce the output error until the multilayer perceptron converges.

7. The method according to claim 4, characterized in that, The extraction of multiple statistical features based on the multiple sets of data further includes: Based on the multiple sets of data pairs within the same time window, a variety of statistical parameters are calculated to reflect the amplitude of the output quantity to be corrected and to characterize the change pattern of the output quantity to be corrected and at least one excitation quantity, thus constituting the multiple statistical features.

8. The method according to claim 4, characterized in that, The statistical characteristics include at least one of the following: mean, standard deviation, maximum value, minimum value, range, difference between the last value and the mean, and slope of the endpoints.

9. The method according to claim 4, characterized in that, The step of uniformly correcting all output quantities to be corrected within the current acquisition cycle using the output quantity error compensation amount to obtain the corrected output quantity further includes: The corrected output quantity is obtained by performing linear operations on each output quantity to be corrected within the current acquisition cycle and the output quantity error compensation quantity.

10. The method according to claim 4, characterized in that, The simultaneous collection of multiple data pairs within a preset time period further includes: Within a preset duration corresponding to a single time window, the output quantity to be corrected, the standard output quantity, and the at least one excitation quantity are synchronously collected at the same time according to a fixed sampling interval to form multiple sets of data pairs.

Images