High-precision power supply control method and device based on digital signal processor

Abstract

The invention relates to the technical field of power supplies, and discloses a high-precision power supply control method based on a digital signal processor, which comprises the following steps of: acquiring data such as voltage, current, temperature, load parameters and the like of a power supply at a high frequency of kHz-level sampling frequency by using a high-precision sensor, transmitting the data to a DSP (Digital Signal Processor), and removing noise interference by applying wavelet transform. Sample quality is improved through high-frequency data acquisition and wavelet denoising, nonlinear and time-varying characteristics of a power supply system are identified by using a lightweight CNN and an LSTM, a reinforcement learning dynamic optimization control strategy is combined, self-adaptive switching between traditional control and an intelligent algorithm is realized through working condition identification, and the power supply system control method based on the LSTM is realized. The problems that noise interference influences data quality, system characteristic recognition is inaccurate, complex working condition control strategy optimization is insufficient, and output precision is unstable due to the fact that a control mode cannot be switched in a self-adaptive mode in an existing power supply can be effectively solved, and control precision, efficiency and device loss can be considered. And the stable operation requirement of the high-precision power supply under multiple working conditions is met.

Description

Technical Field

[0001] This invention relates to the field of power supply technology, specifically to a high-precision power supply control method and apparatus based on a digital signal processor. Background Technology

[0002] A power source is a device that converts other forms of energy into electrical energy to supply power to circuits or equipment. It is a core component of power supply and its energy conversion forms are diverse. For example, generators convert mechanical energy into electrical energy, and batteries convert chemical energy into electrical energy. According to the output type, power sources are classified as AC power sources and DC power sources. According to the application, they are classified as industrial power sources and household power sources.

[0003] Existing power supplies have problems such as difficulty in accurately identifying the nonlinear and time-varying characteristics of the system, insufficient optimization of control strategies under complex operating conditions, and instability in output accuracy due to the inability to adaptively switch control modes. In addition, traditional control methods are slow to respond when the load changes, and it is difficult to balance control accuracy, efficiency and device losses, so they cannot meet the stable operation requirements of high-precision power supplies under multiple operating conditions.

[0004] In practical applications, existing power supplies struggle to accurately identify their inherent nonlinear and time-varying characteristics during operation. This makes accurately grasping the system state extremely difficult. When faced with complex and ever-changing operating conditions, existing control strategies are clearly insufficiently optimized, unable to effectively cope with various emergencies, resulting in poor control performance. Furthermore, they lack adaptive switching capabilities and cannot flexibly adjust according to changes in actual operating conditions, leading to fluctuations in power supply output accuracy and difficulty in ensuring stability. In addition, traditional control methods exhibit significant lag in response speed when faced with sudden load changes, failing to make rapid adjustments. This not only affects control accuracy and efficiency but also increases device losses. In summary, traditional power supply systems cannot meet the urgent need for high-precision power supplies to operate stably and efficiently under various operating conditions.

[0005] Therefore, we propose a high-precision power control method and device based on a digital signal processor. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a high-precision power control method and apparatus based on a digital signal processor, solving the problems mentioned in the background section.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the present invention provides a high-precision power control method based on a digital signal processor, comprising the following steps:

[0010] Step 1: Using a high-precision sensor, high-frequency data such as voltage, current, temperature, and load parameters of the power supply are collected at a sampling frequency of kHz. The data is then transmitted to the DSP. Wavelet transform is first used to remove noise interference, and then normalization is performed to standardize the data format, providing high-quality samples for subsequent deep learning model training.

[0011] Step 2: Deploy lightweight CNN and LSTM in the DSP, train the model offline using a large amount of historical power supply operating data, adjust the network weights, so that the CNN can extract the spatial features of the power supply operating data and capture local patterns; let the LSTM learn the time variation law of the power supply parameters and predict the future operating trend, so as to accurately identify the nonlinear characteristics and time-varying laws of the power supply system.

[0012] Step 3: During actual operation of the power supply, the DSP inputs the pre-processed data into the trained deep learning model in real time. The model quickly completes online inference and outputs a deep feature vector of the power supply's operating state containing complex correlation information, providing a basis for subsequent control strategy optimization.

[0013] Step 4: Combine the feature vectors output by the deep learning model with the real-time power supply parameters to construct the state space of reinforcement learning. Set the adjustable range of the power supply control parameters as the action space of reinforcement learning. Combine the power supply control objective with the output accuracy, efficiency and stability to design the reward function.

[0014] Step 5: In the DSP, a deep Q-network (DQN) is used to perform online training and decision-making tasks of reinforcement learning. The algorithm selects control actions by exploring and using a balance strategy based on the current state. After the action is executed, the network parameters are updated according to the reward feedback. The algorithm continuously learns the optimal control strategy under different operating conditions, dynamically adjusts the power supply parameters, and adapts to time-varying operating conditions.

[0015] Step Six: Through operating condition identification and control mode switching, under steady-state operating conditions, the power supply can quickly stabilize the power output using traditional PID control or model predictive control (MPC). When the DSP detects complex operating conditions, it uses intelligent algorithms to generate a better control strategy to ensure that the power supply maintains high-precision output under various operating conditions.

[0016] Preferably, the wavelet transform performs time-frequency decomposition on the noisy power supply signal through the wavelet function, extracts signal features at different scales and locations, filters out high-frequency noise, and retains the true voltage and current change trends. The normalization process maps power supply parameters of different dimensions to the interval [-1,1] or [0,1], reduces the training deviation of deep learning models caused by differences in numerical range, and improves the convergence speed of the algorithm.

[0017] The denoising formula for wavelet transform is:

[0018] Where f(t) is the original sampled signal, a time series of voltage and current. ψ(t) represents the wavelet transform coefficients, which characterize the signal intensity at scale a and position b. a is the scale parameter; when a>0, the signal details are magnified, and when a<0, the signal is magnified. This parameter is used to adjust the frequency resolution. b is the translation parameter, which controls the position of the wavelet function on the time axis. This parameter is used to locate the signal features. ψ(t) is the mother wavelet function.

[0019] The formula for normalization is:

[0020] in Here are the original data points, μ is the data uniformity, and σ is the standardized deviation of the dataset. This is the normalized data.

[0021] Preferably, the lightweight CNN uses CNN convolutional layer operations to extract local features from the power signal through convolutional kernel w, capturing spatial correlations such as high-frequency noise patterns and harmonic components, providing basic features for subsequent working condition identification;

[0022] CNN convolutional layer operation formula:

[0023] Where x i,j Given the pixel value at position (i,j) of the input feature map, which corresponds to a sampling point of the signal at a certain moment, represents the weights of the M×N convolution kernel, and b is the bias term, which adjusts the baseline value of the output features. This is for outputting feature maps.

[0024] Preferably, the LSTM controls the time-series memory of power supply parameters through forget gates, input gates, and output gates to learn dynamic characteristics such as load changes and temperature drift, and predict future operating conditions.

[0025] LSTM cell state update formula: i t =σ(W i ·[h t-1 ,x t ]+b i )

[0026] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0027] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0028]

[0029] h t =o t ⊙tanh(C t )

[0030] Where i t f t o t These are the activation values ​​for the input gate, forget gate, and output gate, respectively, ranging from 0 to 1, controlling the flow of information. Input for the current time, h t-1 C is the hidden state from the previous moment. t The current cell state is represented by σ, which is the Sigmoid activation function that maps the input to [0,1] for gating decisions. W is the weight matrix, b is the bias vector, and ⊙ represents element-wise multiplication. This is the candidate unit state.

[0031] Preferably, the deep learning model inputs the pre-processed power data in real time into the trained deep neural network, and generates feature vectors through forward propagation to characterize the nonlinear characteristics of the current operating state of the power supply, load type, interference intensity, etc.

[0032] The formula is: z = f DNN (x norm ;θ)

[0033] Where z is the output feature vector, f DNN For deep neural network mapping functions, The normalized output includes multidimensional data such as voltage, current, and temperature. θ is the set of model parameters, including the weights and biases of the CNN / LSTM, which are determined through offline training.

[0034] Preferably, the state space of the reinforcement learning combines deep learning features with real-time error, current change rate, and temperature into a state vector to comprehensively describe the current operating condition of the power supply and provide a basis for reinforcement learning decision-making. The reward function balances control accuracy, efficiency, and control smoothness. The smaller the voltage error and the higher the efficiency, the higher the reward. At the same time, it penalizes sudden changes in control quantity to reduce the loss of switching devices.

[0035] State space construction formula:

[0036] Where s t Let z be the state vector at time t. t e is the feature vector extracted by deep learning. t For the output voltage error, e t =V ref -V out Vref For reference value, V out This is the actual value. Rate of change of current T t This refers to the power supply temperature.

[0037] Reward function formula:

[0038] Where r t Let be the reward value at time t, and α, β, γ be the weighting coefficients, |e t | represents the absolute value of the power supply error, η t For power efficiency, |Δu t | represents the absolute value of the control variable | Δu t |=|u t -u t-1 |

[0039] Preferably, the Deep Q-Network (DQN) iteratively updates the Q-value function using the Bellman equation, and learns the state s. t Next, execute action a t The long-term reward expectation is used to optimize the control strategy;

[0040] The formula is: Q(s) t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t-1 ,a')-Q(s t ,a t )]

[0041] Where Q(s) t ,a t ) represents state s t Next, execute action a t The Q-value, α is the learning rate, r t For immediate rewards, γ is the discount factor, and max is the maximum value. a' Q(s t-1 ,a') represents the next state s t-1 Find the optimal action Q value.

[0042] Preferably, the working condition identification and control mode switching determines whether the power supply is in a steady state by setting an error threshold and a current change rate threshold. When a sudden load change causes the error or change rate to exceed the threshold, the system switches to intelligent control mode; otherwise, traditional PID control is used.

[0043] The formula is:

[0044] Where e thFor voltage error threshold, This is the threshold for current change.

[0045] A high-precision power control device based on a digital signal processor includes a voltage sensor, a current sensor, a temperature sensor, an ADC module, a DSP chip, and a DAC module, all of which are mounted on the power supply.

[0046] Preferably, the DSP chip is connected to the auxiliary circuit of the power supply through GPIO pins to monitor and control the power supply's operating status;

[0047] The ADC module is connected to the SPI peripheral of the DSP chip via the SPI interface;

[0048] The DAC module is connected to the SPI peripheral of the DSP chip via the SPI interface;

[0049] The voltage sensor is connected to the power output terminal in parallel. The positive input terminal is connected to the positive output terminal of the power supply, the negative input terminal is connected to the negative output terminal of the power supply, and the output terminal is connected to the analog input channel of the ADC module.

[0050] The current sensor is connected in series to the main power circuit, which disconnects the path of the power output current. The primary side of the current sensor is connected in series to the circuit to ensure that all the power output current passes through the sensor. The analog signal output by the secondary side of the sensor, which is proportional to the current, is connected to the corresponding analog input channel of the ADC module.

[0051] The temperature sensor is installed near the key heat-generating components inside the power supply. The sensor's data pin is connected to the GPIO pin of the DSP to transmit temperature data. All of the above are connected to the power supply.

[0052] In summary, the technical effects and advantages of this invention are as follows:

[0053] 1. In this invention, high-frequency data acquisition and wavelet denoising are used to improve sample quality. Lightweight CNN and LSTM are used to identify the nonlinear and time-varying characteristics of the power supply system. Reinforcement learning is combined to dynamically optimize the control strategy. Then, the adaptive switching between traditional control and intelligent algorithms is achieved through operating condition identification. This can effectively solve the problems of noise interference affecting data quality, inaccurate identification of system characteristics, insufficient optimization of control strategies for complex operating conditions, and unstable output accuracy caused by the inability to adaptively switch control modes in existing power supplies. It can balance control accuracy, efficiency and device loss, and meet the stable operation requirements of high-precision power supplies under multiple operating conditions. Attached Figure Description

[0054] Figure 1 This is a flowchart of the high-precision power control method based on a digital signal processor according to the present invention. Detailed Implementation

[0055] 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.

[0056] A high-precision power supply control device based on a digital signal processor includes a voltage sensor, a current sensor, a temperature sensor, an ADC module, a DSP chip, and a DAC module, all mounted on the power supply. The DSP chip is connected to the power supply's auxiliary circuit via GPIO pins to monitor and control the power supply's operating status. The ADC module is connected to the DSP chip's SPI peripheral via an SPI interface. The DAC module is also connected to the DSP chip's SPI peripheral via an SPI interface. The voltage sensor is connected to the power supply output in parallel, with its positive input connected to the positive output terminal and its negative input connected to the negative output terminal. The output terminal is connected to the analog input channel of the ADC module. The current sensor is connected to the main power supply circuit in series, disconnecting the path of the power supply's output current. The primary side of the current sensor is connected in series to ensure that all the power supply's output current passes through the sensor. The analog signal output from the secondary side of the sensor, proportional to the current, is connected to the corresponding analog input channel of the ADC module. The temperature sensor is installed near key heat-generating components inside the power supply, and its data pin is connected to the GPIO pins of the DSP to transmit temperature data. All of these components are connected to the power supply.

[0057] A high-precision power supply control method based on a digital signal processor includes the following steps:

[0058] Step 1: Using a high-precision sensor, high-frequency data such as voltage, current, temperature, and load parameters of the power supply are collected at a sampling frequency of kHz. The data is then transmitted to the DSP. Wavelet transform is first used to remove noise interference, and then normalization is performed to standardize the data format, providing high-quality samples for subsequent deep learning model training.

[0059] Step 2: Deploy lightweight CNN and LSTM in the DSP, train the model offline using a large amount of historical power supply operating data, adjust the network weights, so that the CNN can extract the spatial features of the power supply operating data and capture local patterns; let the LSTM learn the time variation law of the power supply parameters and predict the future operating trend, so as to accurately identify the nonlinear characteristics and time-varying laws of the power supply system.

[0060] Step 3: During actual operation of the power supply, the DSP inputs the pre-processed data into the trained deep learning model in real time. The model quickly completes online inference and outputs a deep feature vector of the power supply's operating state containing complex correlation information, providing a basis for subsequent control strategy optimization.

[0061] Step 4: Combine the feature vectors output by the deep learning model with the real-time power supply parameters to construct the state space of reinforcement learning. Set the adjustable range of the power supply control parameters as the action space of reinforcement learning. Combine the power supply control objective with the output accuracy, efficiency and stability to design the reward function.

[0062] Step 5: In the DSP, a deep Q-network (DQN) is used to perform online training and decision-making tasks of reinforcement learning. The algorithm selects control actions by exploring and using a balance strategy based on the current state. After the action is executed, the network parameters are updated according to the reward feedback. The algorithm continuously learns the optimal control strategy under different operating conditions, dynamically adjusts the power supply parameters, and adapts to time-varying operating conditions.

[0063] Step Six: Through operating condition identification and control mode switching, under steady-state operating conditions, the power supply can quickly stabilize the power output using traditional PID control or model predictive control (MPC). When the DSP detects complex operating conditions, it uses intelligent algorithms to generate a better control strategy to ensure that the power supply maintains high-precision output under various operating conditions.

[0064] Among them, wavelet transform performs time-frequency decomposition on noisy power supply signals through wavelet functions, extracts signal features at different scales and locations, filters out high-frequency noise, retains the true voltage and current change trends, and normalization processes map power supply parameters of different dimensions to the [-1,1] or [0,1] interval, reducing the training bias of deep learning models caused by differences in numerical ranges and improving the convergence speed of the algorithm.

[0065] The denoising formula for wavelet transform is:

[0066] Where f(x) is the original sampled signal, a time series of voltage and current. ψ(t) represents the wavelet transform coefficients, which characterize the signal intensity at scale a and position b. a is the scale parameter; when a>0, the signal details are magnified, and when a<0, the signal is magnified. This parameter is used to adjust the frequency resolution. b is the translation parameter, which controls the position of the wavelet function on the time axis. This parameter is used to locate the signal features. ψ(t) is the mother wavelet function.

[0067] The formula for normalization is:

[0068] in Here are the original data points, μ is the data uniformity, and σ is the standardized deviation of the dataset. This is the normalized data.

[0069] Among them, the lightweight CNN uses CNN convolutional layer operations to extract local features of the power signal through convolutional kernel w, capturing spatial correlations such as high-frequency noise patterns and harmonic components, providing basic features for subsequent working condition identification.

[0070] CNN convolutional layer operation formula:

[0071] in Given the pixel value at position (i,j) of the input feature map, which corresponds to a sampling point of the signal at a certain moment, represents the weights of the M×N convolution kernel, and b is the bias term, which adjusts the baseline value of the output features. This is for outputting feature maps.

[0072] LSTM uses forget gates, input gates, and output gates to control the time-series memory of power supply parameters, learn dynamic characteristics such as load changes and temperature drift, and predict future operating conditions.

[0073] LSTM cell state update formula: i t =σ(W i ·[h t-1 ,x t ]+b i )

[0074] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0075] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0076]

[0077] h t =o t ⊙tanh(C t )

[0078] Where i t f t o t These are the activation values ​​for the input gate, forget gate, and output gate, respectively, ranging from 0 to 1, controlling the flow of information. t Input for the current time, h t-1 C is the hidden state from the previous moment. tThe current cell state is represented by σ, which is the Sigmoid activation function that maps the input to [0,1] for gating decisions. W is the weight matrix, b is the bias vector, and ⊙ represents element-wise multiplication. This is the candidate unit state.

[0079] The deep learning model inputs the pre-processed power data into the trained deep neural network and generates feature vectors through forward propagation to characterize the nonlinear characteristics of the current operating state of the power supply, load type, interference intensity, etc.

[0080] The formula is: z = f DNN (x norm ;θ)

[0081] Where z is the output feature vector, f DNN For deep neural network mapping functions, The normalized output includes multidimensional data such as voltage, current, and temperature. θ is the set of model parameters, including the weights and biases of the CNN / LSTM, which are determined through offline training.

[0082] The reinforcement learning state space combines deep learning features with real-time error, current change rate, and temperature into a state vector, which comprehensively describes the current operating condition of the power supply and provides a basis for reinforcement learning decision-making. The reward function balances control accuracy, efficiency, and control smoothness. The smaller the voltage error and the higher the efficiency, the higher the reward. At the same time, it punishes sudden changes in control quantity to reduce the loss of switching devices.

[0083] State space construction formula:

[0084] Where s t Let z be the state vector at time t. t e is the feature vector extracted by deep learning. t For the output voltage error, e t =V ref -V out V ref For reference value, V out This is the actual value. Rate of change of current T t This refers to the power supply temperature.

[0085] Reward function formula:

[0086] Where r t Let be the reward value at time t, and α, β, γ be the weighting coefficients, |e t | represents the absolute value of the power supply error, η t For power efficiency, |Δu t | represents the absolute value of the control variable | Δut |=|u t -u t-1 |

[0087] Among them, the Deep Q-Network (DQN) iteratively updates the Q-value function through the Bellman equation and learns the state s. t Next, execute action a t The long-term reward expectation is used to optimize the control strategy;

[0088] The formula is: Q(s) t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t-1 ,a')-Q(s t ,a t )]

[0089] Where Q(s) t ,a t ) represents state s t Next, execute action a t The Q-value, α is the learning rate, r t For immediate rewards, γ is the discount factor, and max is the maximum value. a' Q(s t-1 ,a') represents the next state s t-1 Find the optimal action Q value.

[0090] Among them, the working condition identification and control mode switching determines whether the power supply is in a steady state by setting error threshold and current change rate threshold. When the load change causes the error or change rate to exceed the threshold, it switches to intelligent control mode; otherwise, traditional PID control is used.

[0091] The formula is:

[0092] Where e th For voltage error threshold, This is the threshold for current change.

[0093] Example 1

[0094] This embodiment provides a high-precision power control method based on a digital signal processor for data acquisition and processing. The specific implementation details are as follows:

[0095] Objective: To collect parameters such as voltage, current, and temperature of the power supply at kHz frequency using high-precision sensors, and to provide noise-free, standardized, high-quality data samples for subsequent model training through wavelet transform denoising and normalization.

[0096] Hardware: Voltage sensor (e.g., TE Connectivity TPS2510, measurement range 0-100V, accuracy up to ±0.1%), current sensor (e.g., LEM LA55-P, measurement range 0-50A, accuracy ±0.5%), temperature sensor (Maxim Integrated DS18B20, measurement range -55℃ to +125℃, accuracy ±0.5℃), ADC module (Analog Devices AD7606 16-bit, 8-channel synchronous sampling ADC, sampling rate up to 200kSPS, meeting the high-frequency acquisition requirements of multiple power parameters), DSP chip (TI TMS320C6678). The outputs of the voltage, current, and temperature sensors are respectively connected to the analog input channels of the ADC module. The ADC module is connected to the SPI peripheral of the DSP chip through the SPI interface. The clock line (SCK), master output slave input line (MOSI), master input slave output line (MISO), and chip select line (CS) of the SPI interface are connected accordingly to ensure synchronous and stable data transmission.

[0097] Software: Integrates open-source wavelet transform for denoising, and normalization functions to standardize the collected data and map it to a specific range (such as [0,1]), providing standard format data for subsequent deep learning models.

[0098] Implementation results: Noise filtering efficiency is improved by more than 90%, the true signal trend (such as the detailed features of voltage fluctuations and current changes) is preserved, and the signal-to-noise ratio of the data is significantly improved; parameters of different dimensions (such as voltage in V level and temperature in °C level) are uniformly mapped to the range of [-1,1] or [0,1], the model training convergence speed is improved by 50%, the number of parameter optimization iterations is reduced by 30%, and a data foundation is laid for accurate modeling.

[0099] Example 2

[0100] This embodiment provides a high-precision power supply control method based on a digital signal processor, which is used to accurately identify the nonlinear and time-varying characteristics of a power supply system. The specific implementation details are as follows:

[0101] Implementation objective: To deploy lightweight CNN and LSTM in a DSP, train the model using historical operating data, extract spatial features of power supply operating data using CNN, and capture local patterns; and enable LSTM to learn the time variation law of power supply parameters and predict future operating conditions, so as to accurately identify the nonlinear characteristics and time-varying laws of the power supply system.

[0102] Hardware: A DSP chip, leveraging its multi-core parallel computing capabilities, accelerates deep learning model training. It connects to a Samsung K9F2G08U0B NAND flash memory chip with a capacity of 2Gb to store large amounts of historical power supply operation data for model training. It connects to the DSP via SPI or an 8-bit parallel interface to meet high-speed data read / write requirements.

[0103] Software: Lightweight Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models.

[0104] Implementation results: CNN achieves a 98% accuracy rate in recognizing high-frequency noise patterns and has a harmonic component extraction error of less than 2%, providing basic features for operating condition identification; LSTM has an error rate of less than 5% in predicting operating condition trends in the next 50ms to 100ms, which can predict load changes or temperature anomalies in advance, improving the model's dynamic adaptability to nonlinear systems by more than 80%.

[0105] Example 3

[0106] This embodiment provides a high-precision power control method based on a digital signal processor, which provides real-time decision-making basis for control strategy optimization. The specific implementation details are as follows:

[0107] Implementation objective: During actual power supply operation, the DSP inputs preprocessed data into the trained model in real time, and generates a deep feature vector containing complex correlation information through forward propagation, providing real-time decision-making basis for control strategy optimization.

[0108] Hardware: DSP chip, utilizing the DSP's internal high-speed L2 cache (TMS320C6678's 1MB L2 cache) to store frequently accessed model parameters and intermediate calculation results, reducing data access latency and improving inference speed.

[0109] Implementation results: The feature vector generation delay is less than 1ms, enabling real-time characterization of the power supply's operating status; the feature dimension compression ratio reaches 10:1, retaining more than 95% of the effective information while reducing subsequent computational complexity, creating conditions for rapid control decisions.

[0110] Example 4

[0111] This embodiment provides a high-precision power control method based on a digital signal processor, which provides a complete decision-making framework for reinforcement learning. The specific implementation details are as follows:

[0112] Objective: To combine deep learning features with parameters such as real-time error, current change rate, and temperature into a state vector, define the adjustable range of control parameters as the action space, and design a reward function that balances accuracy, efficiency, and stability, thereby providing a complete decision-making framework for reinforcement learning.

[0113] Hardware: The DSP chip TMS320C6678 is used to execute reinforcement learning algorithms and calculate parameters related to the state space, action space, and reward function.

[0114] Implementation results: The state vector provides a 99% complete description of the operating conditions, covering all scenarios such as voltage deviation, load mutation, and temperature drift; the reward function dynamically balances the control target through weight coefficients, increasing the reward value by 20% when the voltage error is <10mV, and increasing the reward value by 15% for every 1% increase in efficiency, while penalizing sudden changes in control quantity reduces device losses by 25%.

[0115] Example 5

[0116] This embodiment provides a high-precision power supply control method based on a digital signal processor, which is used to dynamically adjust power supply parameters and achieve adaptive response to time-varying operating conditions. The specific implementation details are as follows:

[0117] Objective: To utilize a deep Q-network (DQN) to perform reinforcement learning tasks, iteratively update the Q-value function through the Bellman equation, learn the optimal control strategy under different operating conditions, dynamically adjust power supply parameters, and achieve adaptive response to time-varying operating conditions.

[0118] Hardware: The TMS320C6678 DSP chip serves as the core processor, running the Deep Q-Network (DQN) algorithm to achieve reinforcement learning policy optimization and decision-making.

[0119] Implementation results: The control strategy optimization iteration cycle is less than 5ms, which can respond to load changes in real time (such as when the load is suddenly increased by 50%, the voltage recovery time is shortened from 50ms in the traditional method to 15ms); the control accuracy under multiple operating conditions is improved by 30% to 50%, for example, the voltage fluctuation range in steady state is reduced from ±50mV to ±10mV, and the current overshoot in dynamic state is reduced from 20% to less than 8%.

[0120] Example 6

[0121] This embodiment provides a high-precision power control method based on a digital signal processor for operating condition identification and control mode switching. The specific implementation details are as follows:

[0122] Implementation objective: By setting error thresholds and current change rate thresholds, the power supply operating status can be judged in real time: in steady state, traditional PID or MPC control is used to quickly and stably output; in complex operating conditions (such as sudden load changes or strong interference), intelligent algorithm control is switched to ensure high-precision output under all operating conditions.

[0123] Hardware: The DSP chip TMS320C6678 coordinates the switching between traditional control algorithms and deep learning-reinforcement learning control modes, and selects the appropriate control strategy based on the operating condition judgment result. The DAC module (TI's DAC8562, a 16-bit dual-channel DAC with a setup time of 5μs) converts the digital control signals output by the DSP into analog signals to drive the power devices. Its input is connected to the DSP's SPI interface, and its output analog signal is connected to the power device drive circuit.

[0124] Implementation results: The accuracy of operating condition identification reaches 99.5%, the switching delay is <2ms, reducing the limitations of traditional single control mode; the response speed of PID control in steady state is improved by 40%, the intelligent algorithm improves the output accuracy by 60% under complex operating conditions, the power supply efficiency is improved by an average of 12% under all operating conditions, and the stability index (such as MTBF) is extended by 30%.

[0125] Comparative Example

[0126] Example 1: Using a high-precision sensor, high-frequency data such as voltage, current, temperature, and load parameters of the power supply are collected at a sampling frequency of kHz. The data is then transmitted to the DSP. Wavelet transform is first used to remove noise interference, and then normalization is performed to standardize the data format, providing high-quality samples for subsequent deep learning model training.

[0127] Comparative example: Traditional analog filtering (RC circuit) combined with simple linear normalization is used. An RC filter circuit with a fixed cutoff frequency is used to remove noise. It is sampled by a 12-bit ADC (such as AD7689). The data is mapped to [0,1] only through Min-Max normalization, without considering the statistical characteristics of the signal.

[0128] Key differences: This embodiment achieves time-frequency domain adaptive denoising through wavelet transform, paired with a 16-bit AD7606 (200kSPS sampling rate), achieving a noise filtering efficiency of over 90% and improving the data signal-to-noise ratio by 40dB; while the proportional RC filter cannot adapt to time-varying noise, the 12-bit ADC has large quantization error, and there is still data distribution deviation after normalization, resulting in a 50% slower model training convergence speed.

[0129] Example 2: Deploy lightweight CNN and LSTM in DSP, train the model offline using a large amount of historical power supply operating data, adjust the network weights, so that CNN can extract the spatial features of power supply operating data and capture local patterns; let LSTM learn the time variation law of power supply parameters and predict future operating conditions to accurately identify the nonlinear characteristics and time-varying laws of the power supply system.

[0130] Comparative example: Traditionally, a single fully connected BP neural network is used, without deploying lightweight CNN / LSTM.

[0131] Key differences: This invention uses CNN to extract spatial features such as voltage ripple (98% accuracy) and LSTM to capture temporal trends such as temperature drift (prediction error <5% in the next 100ms). It combines multi-core DSP and 2Gb NAND flash memory to achieve batch training, and the model parameters are compressed 10 times after quantization. In contrast, the BP network cannot handle spatiotemporal correlation features, the training time increases by 5 times, and the recognition error for sudden load changes exceeds 20%.

[0132] Example 3: When the power supply is actually running, the DSP inputs the pre-processed data into the trained deep learning model in real time. The model quickly completes online inference and outputs a deep feature vector of the power supply's operating state containing complex correlation information, providing a basis for subsequent control strategy optimization.

[0133] Comparative example: Inference is performed on a cloud server (not local real-time processing). Data is uploaded to the cloud via Wi-Fi and calculated using a TensorFlow model. The network transmission and cloud processing latency exceeds 100ms.

[0134] Key differences: This embodiment utilizes the 1MB L2 cache inside the DSP to accelerate model inference, with a feature vector generation latency of <1ms, directly synchronized with the 10kHz control cycle; the other embodiment relies on network communication, is only suitable for non-real-time scenarios, cannot meet the millisecond-level response requirements of industrial power supplies, and has cloud traffic costs and network interruption risks.

[0135] Example 4: Combine the feature vectors output by the deep learning model with the real-time power supply parameters to construct the state space of reinforcement learning. Set the adjustable range of the power supply control parameters as the action space of reinforcement learning. Combine the power supply control objective with the output accuracy, efficiency and stability to design the reward function.

[0136] Comparative example: Fixed parameter PID control is used, without the introduction of reinforcement learning, and the parameters are manually adjusted. The control target is only voltage error, and power efficiency and device losses are not considered.

[0137] Key differences: This invention constructs a 4-dimensional state space (deep learning features + voltage error + current change rate + temperature), and dynamically balances the reward function's accuracy (20% increase in reward when voltage error < 10mV), efficiency (15% increase in reward for every 1% increase), and control smoothness (penalizing sudden changes reduces losses by 25%). The comparative example shows a voltage recovery time of 50ms when the load suddenly increases by 50%, and a significant decrease in efficiency under light load, making adaptive optimization impossible.

[0138] Example 5: A deep Q-network (DQN) is used in a DSP to perform online training and decision-making tasks of reinforcement learning. The algorithm selects control actions by exploring and using a balancing strategy based on the current state. After executing the action, the network parameters are updated based on the reward feedback. The algorithm continuously learns the optimal control strategy under different operating conditions, dynamically adjusts the power supply parameters, and adapts to time-varying operating conditions.

[0139] Comparative example: The table-based Q-Learning algorithm is used without deep networks. The state space is discrete in 10×10×10 dimensions and stored in a lookup table. When running on an STM32F4 MCU, it cannot handle high-dimensional continuous states.

[0140] Key difference: This invention uses a Deep Q-Network (DQN) to approximate the Q-value function with a CNN, supporting real-time decision-making in a 4-dimensional continuous state space (control cycle of 5ms), and can adapt to temperature changes from -40℃ to +85℃ without retraining; the discrete state space of the scale cannot represent complex working conditions (such as nonlinear loads), new working conditions require offline training for several hours, and the decision delay exceeds 50ms, which cannot match the switching frequency of high-frequency power supplies.

[0141] Example 6: By identifying operating conditions and switching control modes, the power supply can quickly stabilize its output under steady-state conditions using traditional PID control or model predictive control (MPC). When the DSP detects complex operating conditions, it uses intelligent algorithms to generate a better control strategy, ensuring that the power supply maintains high-precision output under various operating conditions.

[0142] Comparative example: A single control mode is used (intelligent algorithm only or PID only), without operating condition switching, always running deep learning control, or always using PID control, without setting an error threshold (such as ±0.5% of rated voltage) to determine the operating condition.

[0143] Key difference: This invention achieves adaptive switching of control mode through working condition identification: the PID control response speed is improved by 40% in steady state, and the intelligent algorithm is switched to improve the output accuracy by 60% in complex working conditions, while the DSP computing power consumption is reduced by 30%; in contrast, if only the intelligent algorithm is used, it will lead to a waste of steady-state computing power (energy consumption increases by 50%), and if only PID is used, the overshoot will be greater than 30% when the load changes suddenly, and it is impossible to balance efficiency and stability.

[0144] In summary, the above embodiments, through a closed-loop design of "data preprocessing - feature learning - intelligent decision-making - mode switching", improve the power control accuracy from ±1% of the traditional solution to below ±0.1%, increase the dynamic response speed by 3-5 times, optimize the efficiency of all operating conditions by 10% to 20%, and reduce device losses and computing costs.

[0145] All electrical components mentioned in this article are connected to an external main controller and 220V AC mains power, and the main controller can be a conventional known device such as a computer that can control it.

[0146] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-precision power supply control method based on a digital signal processor, characterized in that, Includes the following steps: S1: Using high-precision sensors, high-frequency data such as voltage, current, temperature, and load parameters of the power supply are collected at a sampling frequency of kHz. The data is then transmitted to the DSP. Wavelet transform is first used to remove noise interference, and then normalization is used to standardize the data format, providing high-quality samples for subsequent deep learning model training. S2: Deploy lightweight CNN and LSTM in DSP, train the model offline using a large amount of historical power supply operating data, adjust the network weights, so that CNN can extract the spatial features of power supply operating data and capture local patterns; let LSTM learn the time variation law of power supply parameters and predict future operating conditions to accurately identify the nonlinear characteristics and time-varying laws of power supply system. S3: When the power supply is actually running, the DSP inputs the pre-processed data into the trained deep learning model in real time. The model quickly completes online inference and outputs a deep feature vector of the power supply's operating status containing complex correlation information, providing a basis for subsequent control strategy optimization. S4: Combine the feature vectors output by the deep learning model with the real-time power supply parameters to construct the state space of reinforcement learning. Set the adjustable range of the power supply control parameters as the action space of reinforcement learning. Combine the power supply control objective with the output accuracy, efficiency and stability to design the reward function. S5: A deep Q-network (DQN) is used in the DSP to perform online training and decision-making tasks of reinforcement learning. The algorithm selects control actions by exploring and using a balance strategy based on the current state. After the action is executed, the policy network parameters are updated according to the reward feedback. The algorithm continuously learns the optimal control strategy under different operating conditions, dynamically adjusts the power supply parameters, and adapts to time-varying operating conditions. S6: Through operating condition identification and control mode switching, the power supply can quickly stabilize the power output under steady-state conditions using traditional PID control or model predictive control (MPC). When the DSP detects complex operating conditions, it uses intelligent algorithms to generate a better control strategy to ensure that the power supply maintains high-precision output under various operating conditions.

2. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The wavelet transform performs time-frequency decomposition on the noisy power supply signal through the wavelet function, extracts signal features at different scales and locations, filters out high-frequency noise, and retains the true voltage and current change trends. The normalization process maps power supply parameters of different dimensions to the interval [-1,1] or [0,1], reduces the training bias of deep learning models caused by differences in numerical range, and improves the convergence speed of the algorithm. The denoising formula for wavelet transform is: Where f(t) is the original sampled signal, a time series of voltage and current. ψ(t) represents the wavelet transform coefficients, characterizing the signal's intensity at scale a and position b. 'a' is the scale parameter; a>0 amplifies signal details, while a<0 shrinks them, used to adjust frequency resolution. 'b' is the translation parameter, controlling the wavelet function's position on the time axis, used to locate signal features. 'ψ(t)' is the mother wavelet function.

3. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The lightweight CNN uses CNN convolutional layer operations to extract local features from the power signal through the convolution kernel w, capturing spatial correlations such as high-frequency noise patterns and harmonic components, providing basic features for subsequent operating condition identification.

4. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The LSTM uses forget gates, input gates, and output gates to control the time-series memory of power supply parameters, learn dynamic characteristics such as load changes and temperature drift, and predict future operating conditions.

5. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The deep learning model inputs the pre-processed power data into the trained deep neural network and generates feature vectors through forward propagation to characterize the nonlinear characteristics of the current operating state of the power supply, load type, interference intensity, etc.

6. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The state space of the reinforcement learning combines deep learning features with real-time error, current change rate, and temperature into a state vector, which comprehensively describes the current operating condition of the power supply and provides a basis for reinforcement learning decisions. The reward function balances control accuracy, efficiency, and control smoothness. The smaller the voltage error and the higher the efficiency, the higher the reward. At the same time, it penalizes sudden changes in control quantity to reduce the loss of switching devices.

7. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The Deep Q-Network (DQN) iteratively updates the Q-value function using the Bellman equation and learns the state s. t Next, execute action a t The long-term reward expectation is used to optimize the control strategy.

8. The high-precision power control method based on a digital signal processor according to claim 1, characterized in that: The operating condition identification and control mode switching determines whether the power supply is in a steady state by setting an error threshold and a current change rate threshold. When a sudden load change causes the error or change rate to exceed the threshold, the system switches to intelligent control mode; otherwise, traditional PID control is used.

9. A high-precision power control device based on a digital signal processor, characterized in that: The voltage sensor, current sensor, temperature sensor, ADC module, DSP chip, and DAC module are all located on the power supply.

10. The high-precision power control device based on a digital signal processor according to claim 9, characterized in that: The DSP chip is connected to the auxiliary circuit of the power supply through GPIO pins to monitor and control the power supply's operating status. The ADC module is connected to the SPI peripheral of the DSP chip via the SPI interface; The DAC module is connected to the SPI peripheral of the DSP chip via the SPI interface; The voltage sensor is connected to the power output terminal in parallel. The positive input terminal is connected to the positive output terminal of the power supply, the negative input terminal is connected to the negative output terminal of the power supply, and the output terminal is connected to the analog input channel of the ADC module. The current sensor is connected in series to the main power circuit, which disconnects the path of the power output current. The primary side of the current sensor is connected in series to the circuit to ensure that all the power output current passes through the sensor. The analog signal output by the secondary side of the sensor, which is proportional to the current, is connected to the corresponding analog input channel of the ADC module. The temperature sensor is installed near the key heat-generating components inside the power supply. The sensor's data pin is connected to the GPIO pin of the DSP to transmit temperature data. All of the above are connected to the power supply.

Images