Cloud edge collaboration-based low-ripple high-voltage power supply ratio optimization management method

Abstract

The application discloses a low-ripple high-voltage power supply ratio optimization management method based on cloud-edge collaboration, and relates to the technical field of high-voltage power supplies.The steps of the method are as follows:Step one: constructing an edge computing node on a high-voltage power supply device containing a linear voltage regulating unit and a high-frequency switching unit to collect sensing data, including voltage, current, temperature and ripple spectrum characteristics, and periodically uploading the data to the cloud after completing feature extraction and desensitization processing locally to form multi-device desensitization data;Step two: performing bimodal collaborative control on the edge computing node according to the sensing data to generate real-time control signals acting on the linear voltage regulating unit and the high-frequency switching unit respectively;wherein the bimodal collaborative control includes steady-state conditions and dynamic conditions, in the steady-state conditions, the linear voltage regulating unit is preferentially used for fine adjustment, and in the dynamic conditions, the high-frequency switching unit actively pre-adjusts the intermediate bus voltage, greatly reducing operation and maintenance costs and prolonging the service life.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of high-voltage power supplies, in particular to a low-ripple high-voltage power supply ratio optimization management method based on cloud edge collaboration. BACKGROUND

[0002] A high-precision low-ripple power supply is a high-performance power supply device capable of providing extremely stable and pure DC output; the deviation between the set value and the actual value of the output voltage or current is extremely small, usually reaching several ten-thousandths or even higher; meanwhile, the amplitude of the AC noise superimposed in the output is extremely low, and is usually measured by a millionth root value, for example, lower than 15ppm; under long-time operation, temperature change or load fluctuation, the output can still maintain high consistency, and the temperature drift coefficient can be as low as 10ppm / ℃ or lower; such a power supply is widely used in high-end fields that are extremely sensitive to power quality.

[0003] The existing high-precision low-ripple power supply generally adopts a hybrid architecture of linear voltage regulation and switching pre-regulation, uses a high-frequency switching power supply as a front stage to provide a rough but efficient voltage reference, and then uses a rear-stage linear power supply for fine adjustment to balance efficiency and ripple performance; however, this scheme has inherent limitations, making its performance improvement encounter a bottleneck; the key control parameters are usually set according to the target working condition when the product is shipped, and cannot be adaptively adjusted with the load dynamics, environmental temperature changes or device aging, resulting in a sharp decrease in ripple suppression capability under non-nominal working conditions; the power supply unit can only sense its own state, lacks global awareness of historical running data, behaviors of the same type of device group and external environmental disturbances, and cannot make forward-looking predictions and compensation; although low-temperature drift devices and software are used for temperature compensation, the compensation model is usually based on a lookup table method or simple linear and polynomial fitting; the ripple is not only derived from the inherent switching noise of the switching power supply, but also affected by the coupling of multiple factors such as thermal noise of the linear adjustment tube, noise of the reference voltage source, feedback loop bandwidth limitation and ground loop interference; in view of the above technical defects, a solution is proposed. SUMMARY

[0004] To achieve the above purpose, the application is implemented by the following technical solutions:

[0005] The low-ripple high-voltage power supply ratio optimization management method based on cloud edge collaboration comprises the following steps:

[0006] Step 1: In a high-voltage power supply device comprising a linear voltage regulation unit and a high-frequency switching unit, an edge computing node is constructed for collecting sensing data, including voltage, current, temperature and ripple spectrum characteristics, and after local feature extraction and desensitization processing, the data are periodically uploaded to the cloud to form multi-device desensitization data;

[0007] Step two: perform bimodal collaborative control on the edge computing node based on the perception data, generate real-time control signals acting on the linear voltage regulation unit and high-frequency switching unit respectively;

[0008] Step three: based on the multi-device desensitization data, construct a general temperature drift model through federated learning, and identify the dominant ripple source through independent component analysis, generate corresponding cloud optimization temperature drift compensation function and control parameter fine-tuning for each edge computing node;

[0009] Step four: send the cloud optimization temperature drift compensation function and control parameter fine-tuning to the corresponding edge computing node, and atomically replace the original local control strategy.

[0010] Preferably, the process of constructing the edge computing node is as follows:

[0011] By deploying a multi-physical quantity fusion perception network, the linear voltage regulation unit and high-frequency switching unit of the high-voltage power supply are perceived and data is collected, and the isolated signal conditioning circuit built-in the edge gateway is used for preprocessing, and after conversion to digital signal, the edge computing node is constructed.

[0012] Preferably, the process of completing feature extraction and desensitization locally is as follows:

[0013] Wavelet packet decomposition algorithm is used to perform multi-scale decomposition on the perception data, extract energy features and harmonic components of each frequency band; based on three-dimensional temperature data collected by distributed temperature sensors, a temperature field distribution model is constructed using gradient descent fitting method; perform fast Fourier transform on the ripple time domain signal to convert it to frequency spectrum; through principal component analysis, the multi-dimensional features of voltage, current, temperature and ripple spectrum are fused and reduced dimension, generating a low-dimensional comprehensive feature vector;

[0014] Sensitive field identification is performed on the perception data to obtain desensitized information categories, including device identity, absolute value and scene association; ε-differential privacy technology is used to determine the privacy budget ε, calculate the noise intensity, and perform multi-dimensional effectiveness verification on the perception data after injecting noise.

[0015] Preferably, the process of forming multi-device desensitization data is as follows:

[0016] After the perception data completed feature extraction and desensitization locally is encrypted by AES-256, it is uploaded to the cloud through the 5G link. The cloud deploys multi-device data aggregation services, receives desensitized data uploaded by each edge node, aligns the data based on timestamp and feature dimension, stores it by device type and running scenario, and constructs a desensitized data set covering multiple high-voltage power supply devices.

[0017] Preferably, the process of performing bimodal collaborative control is as follows:

[0018] Dual-mode cooperative control, including steady-state working condition and dynamic working condition, in the steady-state working condition, the linear voltage regulating unit is preferentially used for fine adjustment; in the dynamic working condition, the intermediate bus voltage is actively pre-adjusted by the high-frequency switching unit.

[0019] Preferably, the process of constructing a universal temperature drift model is:

[0020] The cloud receives desensitized data uploaded from each edge node, stores it in groups according to devices, forms a structured database, deploys a universal temperature drift base model as a starting point of a global model, each device trains a model based on its own historical desensitized data on an edge computing node, calculates model parameter updates, the cloud collects model updates of all devices, and fuses them into a new global model through weighted averaging, repeats iteration until the model converges, obtains a global model through federated learning training, and fine-tunes the universal model for individual differences of each device using the historical desensitized data of the device to generate a corresponding adapter.

[0021] Preferably, the process of analyzing and identifying the dominant ripple source is:

[0022] S301: The cloud receives desensitized ripple spectrum feature data uploaded from all edge nodes;

[0023] S302: Blind source separation and component decoupling through independent component analysis (ICA);

[0024] S303: Accurate identification and engineering classification of the dominant ripple mode, and component characteristic quantitative analysis;

[0025] S304: Establishing an accurate mapping relationship between the dominant mode and system control parameters;

[0026] S305: Generation and packaging of suppression strategies.

[0027] Preferably, the process of generating a corresponding cloud optimization temperature drift compensation function and control parameter fine-tuning amount is:

[0028] Based on the multi-device desensitized data uploaded by all edge nodes, a universal temperature drift base model is trained using federated learning technology; on the basis of the universal model, the cloud performs localized adaptation for each specific edge computing node; the cloud generates a specific control parameter fine-tuning amount for each device based on the dominant ripple source information identified through independent component analysis, and fuses the localized temperature drift compensation function and the control parameter fine-tuning amount to form an optimization strategy package.

[0029] Preferably, the process of atomizing and replacing the original local control strategy is:

[0030] The packaged optimization strategy package is sent to the target edge computing node through an encrypted communication channel, the edge node receives the strategy package, immediately performs integrity verification, and after verification, performs atomic replacement, and after the strategy switching is completed, the edge node immediately starts the validation procedure.

[0031] The beneficial effects of the present application are as follows:

[0032] (1) The local temperature drift compensation model constructed by cloud federated learning fuses general rules and device individual differences, effectively suppresses the nonlinear temperature drift that cannot be covered by traditional lookup table method, simultaneously identifies the dominant noise source from mixed ripple by independent component analysis, and generates a targeted control parameter fine-tuning strategy, and the suppression effect is more significant under dynamic working conditions, solving the problem of low suppression efficiency caused by ripple source coupling.

[0033] (2) The present application realizes edge-side dual-mode collaborative control, and in the steady state, the linear unit is finely adjusted to ensure low noise, and in the dynamic state, the high-frequency switching unit actively pre-adjusts the intermediate bus voltage, reduces the instantaneous burden of the linear unit, shortens the recovery time of the load step response, and avoids overheating due to excessive pressure difference, taking into account energy efficiency, response speed and device life, overcoming the inherent limitations of traditional hybrid architecture static division, and relying on cloud-edge collaborative architecture, the system can periodically upload desensitization data, receive cloud optimization strategies, and realize non-inductive update through atomic replacement, so that the power supply has online learning and adaptive evolution ability, which can effectively cope with long-term disturbances such as component aging and environmental changes, ensuring that the device is always running in the optimal state throughout its life cycle, greatly reducing operation and maintenance costs and extending service life. BRIEF DESCRIPTION OF DRAWINGS

[0034] The present application will be further described below with reference to the accompanying drawings;

[0035] Figure 1 is the method of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0037] Please refer to Figure 1 The present embodiment provides a low-ripple high-voltage power supply ratio optimization management method based on cloud-edge collaboration, including the following steps:

[0038] Step one: build an edge computing node on the high-voltage power supply device containing a linear voltage regulation unit and a high-frequency switching unit to collect sensing data including voltage, current, temperature, and ripple spectrum characteristics, and perform feature extraction and desensitization locally before uploading periodically to the cloud to form multi-device desensitized data.

[0039] High-voltage power supply device:

[0040] Add an edge computing hardware module on the control mainboard or independent expansion board of the high-voltage power supply device, which contains an embedded processor, a high-precision analog-to-digital converter, a high-speed digital isolation interface, an industrial-grade communication module, and a local storage unit, etc.

[0041] Among them, the embedded processor is used to run real-time control algorithms and lightweight AI models; the analog-to-digital converter is used to collect output voltage, current, and key temperature point signals; the digital isolation interface is securely connected to the PWM drive signal of the high-frequency switching unit and the feedback loop of the linear voltage regulation unit; the communication module is used to establish an encrypted data channel with the cloud; the local storage unit is used to cache running logs, model parameters, and desensitized data; the hardware module is interconnected with the original digital controller of the power supply, such as DSP or MCU, through standard buses such as SPI, I 2 C, CAN or PCIe, forming a collaborative control architecture.

[0042] Meanwhile, deploy an edge computing software stack on the hardware, including: a real-time operating system, a data acquisition service, a dual-mode collaborative control engine, a data preprocessing and desensitization module, and an OTA update agent; among them, the real-time operating system is used to ensure the determinism of control tasks; the data acquisition service periodically reads sensor data and synchronously captures ripple spectrum through fast Fourier transform or wavelet analysis; the dual-mode collaborative control engine automatically switches between steady-state / dynamic modes according to the working condition, generates control instructions that act on the linear voltage regulation unit, such as adjusting the operational amplifier reference voltage, and the high-frequency switching unit, such as modifying the PWM duty cycle; the data preprocessing and desensitization module extracts features from raw data, such as mean, variance, frequency band energy, removes device unique identifier, and encrypts before uploading; the OTA update agent receives and verifies the temperature drift compensation function and parameter fine-tuning amount issued by the cloud, and performs atomic strategy replacement.

[0043] It should be noted that the edge computing node does not replace the original power supply controller, but works collaboratively with it. On the millisecond time scale, the edge node directly outputs control signals to the power unit for high-performance regulation; on the second or minute time scale, desensitized data is uploaded to the cloud and optimization strategies are received; the original protection logic, such as overcurrent, overvoltage, and overheating, is still dominated by the original controller, and the edge node only performs performance optimization within its safety boundary.

[0044] The process of constructing the edge computing node for collecting sensing data is:

[0045] For the linear voltage regulation unit and the high-frequency switching unit of the high-voltage power supply, a multi-physical quantity fusion sensing network is deployed to collect electrical parameters, monitor temperature, and identify ripple spectrum; the electrical parameter collection is performed by connecting a high-precision differential voltage sensor in parallel to the output bus, with a range of 0-10 kV, an accuracy of 0.001% FS, and a bandwidth of 1 MHz and a linearity of 0.1% by connecting a Hall current sensor in series, to capture the instantaneous value and dynamic waveform of voltage or current in real time; the temperature monitoring is performed by using a distributed platinum resistance, PT1000, with a resolution of 0.01℃, attached to the power MOSFET junction temperature of the linear voltage regulation unit, the transformer winding of the high-frequency switching unit, and the device environment cabin, to form a three-dimensional temperature field monitoring; the ripple spectrum identification is performed by embedding a miniature spectrum analysis module, collecting the ripple time-domain signal at a sampling rate of 20 MHz and a resolution of 14 bits through a high-speed ADC, and analyzing the frequency-domain characteristics of the ripple in real time through an FPGA hardware-accelerated FFT algorithm, with a frequency resolution of 1 Hz and a frequency range of 100 Hz-10 MHz; all sensing signals are preprocessed through an isolation signal conditioning circuit, i.e., optical coupling isolation and differential amplification, built-in the edge gateway, and converted into digital signals to construct the edge computing node.

[0046] The process of feature extraction is:

[0047] The voltage and current feature extraction is performed by the edge computing node on the collected voltage and current time-domain signals, using a wavelet packet decomposition algorithm for multi-scale decomposition, with 5 layers of decomposition, to extract the energy features and harmonic components of each frequency band; for example, for the output voltage of the linear voltage regulation unit, the decomposition can identify the fundamental component, low-order harmonics such as the 3rd and 5th harmonics, and high-frequency ripple components, and by calculating the energy proportion of each component, a feature vector of the voltage and current is formed, such as the harmonic distortion rate and the fundamental amplitude stability.

[0048] The temperature field feature extraction is based on the three-dimensional temperature data collected by the distributed temperature sensor, using a gradient descent fitting method to construct a temperature field distribution model; taking the power MOSFET junction temperature of the linear voltage regulation unit and the transformer winding temperature of the high-frequency switching unit as the core nodes, the temperature variation gradient with time and load is fitted, and the thermal stress coefficient such as the temperature change rate and the thermal coupling coefficient is extracted, to form a temperature feature matrix for subsequent basic analysis of temperature drift compensation.

[0049] The ripple spectrum feature extraction is performed by performing a fast Fourier transform (FFT) on the ripple time-domain signal to convert it into a frequency spectrum; the frequency resolution is set to 1 Hz, covering a frequency range of 100 Hz-10 MHz, to identify the frequency components of the dominant ripple source, such as the switching frequency harmonics of the high-frequency switching unit and the 50 Hz fundamental wave of the power grid interference, and to calculate the amplitude proportion of each frequency component to generate a ripple spectrum feature vector.

[0050] Feature fusion and dimension reduction fuse and reduce the multi-dimensional features of the above voltage, current, temperature, and ripple spectrum through principal component analysis (PCA). The principal components with an accumulated contribution rate of ≥95% are retained to generate a low-dimensional comprehensive feature vector, which reduces data redundancy and retains key information.

[0051] The process of desensitization is as follows:

[0052] The edge computing node first identifies the sensitive fields of the collected raw data. Through device identity such as the unique number of high-voltage power supply equipment, production serial number, absolute numerical value such as the absolute amplitude of output voltage, such as specific kV value, absolute reading of temperature sensor, such as specific ℃ value, and scene association class, through the implicit sensitive information strongly associated with the device geographic location and production batch, the sensitive field rule library based on the pre-set high-voltage power supply industry data privacy specification is automatically marked to mark the above-mentioned raw data that needs to be desensitized.

[0053] The privacy budget ε is determined by differential privacy noise injection, which is 0.1, meeting the safety requirements of high-voltage power supply industry data sharing; for different types of sensitive data, select the appropriate noise distribution, such as Laplace distribution for numerical data, and exponential distribution for categorical data;

[0054] According to the formula wherein, is the maximum difference of the data characteristic function on adjacent data sets; f is the data characteristic function, is the adjacent data set, combined with the value of ε, to determine the noise scale parameter to generate noise that meets the distribution and inject it into the sensitive data; for example, for the absolute amplitude of output voltage, assuming the range is 0-10kV, inject Laplace noise, the noise mean is 0, and the scale b=(10000V) / 0.1=100000V, to ensure safety protection while retaining the relative change characteristics of the data.

[0055] Perform multi-dimensional effectiveness verification on the data after injecting noise, verify the key features such as the principal component proportion of voltage ripple spectrum and temperature field gradient, and the deviation before desensitization is ≤5%; input the desensitized data into the cloud pre-trained temperature drift model to verify that the prediction error increase is ≤10%; through the indistinguishability of adjacent data sets of differential privacy, the original sensitive information is not able to be deduced from the desensitized data.

[0056] Get multi-device desensitized data:

[0057] A hierarchical periodic upload strategy is adopted. The locally collected data, including voltage harmonic features, temperature field gradients, and ripple spectrum principal components, which have undergone feature extraction and desensitization, are encrypted with AES-256 every second and uploaded to the cloud via 5G or Ethernet links. If the network is interrupted, the node has a built-in 1GB cache queue to temporarily store the data. After the network is restored, the data will be automatically resumed. The cloud-based multi-device data aggregation service receives the desensitized data uploaded by each edge node, aligns the data based on timestamps and feature dimensions, and stores it according to device type and operating scenario to build a desensitized dataset covering multiple high-voltage power supply devices.

[0058] Step 2: Based on the perceived data, perform dual-modal collaborative control on the edge computing node to generate real-time control signals that act on the linear voltage regulation unit and the high-frequency switching unit respectively; wherein, the dual-modal collaborative control includes steady-state operating conditions and dynamic operating conditions. Under steady-state operating conditions, the linear voltage regulation unit is given priority for fine adjustment; under dynamic operating conditions, the high-frequency switching unit actively pre-adjusts the intermediate bus voltage.

[0059] The real-time operating condition determination process is completed locally on the edge computing node without cloud intervention, ensuring a response speed of milliseconds. The edge computing node collects sensing data at fixed time intervals, such as every 1 millisecond, and calculates the rate of change of the output current. If the rate of change is less than a preset first threshold, such as 0.1 mA per millisecond, it is initially determined to be a steady-state condition. At the same time, the change amplitude of the voltage setpoint is detected. If the jump amplitude within several consecutive cycles is less than a second preset threshold, such as 1 volt, it further supports steady-state determination. In addition, the rate of change of the heat sink temperature is checked. If the change is gradual and lower than a third preset threshold, pseudo-dynamics caused by thermal disturbances are excluded, and it is finally confirmed to be a steady-state condition. Otherwise, it is determined to be a dynamic condition, including scenarios such as sudden load increase or decrease, step change of voltage setpoint, and grid dip recovery.

[0060] During the control process under steady-state conditions, the system executes a high-precision ripple suppression mode:

[0061] First, lock the high-frequency switching unit, fixing its operating frequency and duty cycle to the optimal values ​​derived from cloud optimization, for example, a frequency of 100 kHz and a duty cycle of... This allows the output to reach an extremely stable intermediate bus voltage. Then, the fine closed-loop control of the linear voltage regulator unit is initiated. The actual output voltage is acquired using a high-resolution analog-to-digital converter, and the error between it and the set voltage is calculated. Based on this error, an enhanced proportional, integral, and derivative control algorithm is executed, which incorporates a temperature drift compensation function sent from the cloud as an additional input term. This generates a drive signal and sends it to the drive circuit of the linear voltage regulator unit, such as the op-amp reference terminal or the MOSFET gate, to achieve fine adjustment at the microvolt level.

[0062] In this mode, most of the ripple, especially low frequency drift and thermal noise, is effectively absorbed by the linear unit, making the output ripple to the minimum level.

[0063] The entire discrimination and control process is autonomously completed on the edge side, responds quickly, and is not affected by network delay, ensuring high-precision, low-ripple performance of the high-voltage power supply under various working conditions.

[0064] Control process under dynamic working condition:

[0065] When it is determined to be dynamic, the fast response pre-adjustment mode is executed:

[0066] The high-frequency switching unit is activated for active adjustment, and the edge node calls the locally stored dynamic prediction model, which is periodically updated by the cloud, and the model describes the relationship between load changes and the required bus voltage; according to the current output current and its trend, the optimal intermediate bus voltage required at the next moment is predicted ; the PWM duty cycle of the high-frequency switching unit is adjusted in real time to quickly approach ; the linear voltage regulation unit operates under reduced load, and the linear unit still works in the linear region, but the input voltage difference is actively maintained in a small range, wherein is the voltage ultimately output by the system to the load, for example, 10-50V, its power consumption and thermal stress are significantly reduced; saturation, delay or overheat protection caused by sudden increase of voltage difference in traditional scheme is avoided; the system response speed is improved by more than 50%, the output voltage overshoot or undershoot is reduced, the recovery time is shortened, and the life of the linear unit is protected.

[0067] Mode seamless switching and smooth transition, mode switching uses hysteresis judgment and smooth interpolation mechanism to avoid sudden changes in control signals due to critical jitter;

[0068] The control signal is filtered before output to ensure that the driving waveform has no burr; the entire process is completed under the scheduling of the real-time operating system of the edge computing node, and the control delay is ≤100 microseconds; the specific process of dual-mode collaborative control is based on real-time state perception on the edge side, and the optimal control strategy is automatically selected through local intelligent discrimination:

[0069] In steady state, let the linear unit with high precision but low efficiency be finely carved; in dynamic state, let the switching unit with high efficiency but ripple pave the way in advance; through division of labor and complementary advantages, high precision, low ripple and fast response performance under all working conditions are jointly ensured; the process is completely autonomous on the edge side, which is the core execution mechanism for realizing physical performance breakthrough of the invention.

[0070] Step three: based on the multi-device desensitization data, a universal temperature drift model is constructed through federated learning, and the dominant ripple source is identified by independent component analysis, and the corresponding cloud optimization temperature drift compensation function and control parameter fine-tuning amount are generated for each edge computing node;

[0071] Construct a universal temperature drift model:

[0072] The cloud receives desensitization data packets uploaded from each edge node, including temperature, voltage, and timestamp features, and removes sensitive information such as device ID and user identification, and adds noise through differential privacy technology. Store the data by device to form a structured database; deploy a universal temperature drift base model, such as a lightweight neural network or a high-order polynomial model, as a starting point for the global model; each device trains the model locally based on its historical desensitization data on the edge computing node, calculates model parameter updates such as gradients or weight fine-tuning amounts, and only uploads model update parameters, not raw data, to ensure data security; the cloud collects model updates from all devices and fuses them into a new global model through weighted averaging, with weights allocated based on device data volume or reliability; repeat the process, for example, 10 iterations, until the model converges, such as when the validation loss is below a threshold; throughout the process, no raw data is accessed, and the robustness of the model is improved using group data.

[0073] The global model trained through federated learning can accurately describe common temperature drift patterns, such as the average trend of voltage drift when the temperature rises by 1°C; for individual differences of each device, such as component batch deviation and minor assembly differences, the cloud uses the device's historical desensitization data to fine-tune the universal model and generate a dedicated adapter, for example, adding a temperature compensation coefficient offset for a device. The combination of the universal model and the localized adapter serves as the cloud's optimized temperature drift compensation function; the cloud verifies the model's accuracy in a simulated environment and compares it with historical data to ensure that the temperature drift compensation error is less than 5ppm. The structure parameters of the universal model and the fine-tuning amount of the localized adapter are packaged and delivered to the corresponding edge node through a secure channel.

[0074] Identify the dominant ripple source using independent component analysis:

[0075] From mixed and heterogeneous multi-device ripple spectrum data, accurately separate dominant noise patterns with engineering significance, such as high-frequency switching noise dominant mode, thermal noise dominant mode, and power grid resonance dominant mode, and establish their quantitative association with system operating status, thereby providing a data foundation for generating highly targeted suppression strategies;

[0076] S301: The cloud receives the desensitized ripple spectrum feature data packets uploaded from all edge nodes, each packet containing: time series ripple voltage, sampling rate ≥ 100 kHz, covering 0-100 kHz frequency band; synchronously labeled system state label, such as load current rate of change, radiator temperature gradient, power grid voltage fluctuation amplitude; desensitized device unique identifier, only retaining device type, batch, etc. aggregate information; wavelet denoising is performed on the time domain ripple signal, and the threshold is set to signal-to-noise ratio > 20 dB to ensure the reliability of the spectrum analysis; based on the system clock synchronization mechanism, the ripple data of different devices are intercepted according to the same time window, such as 50 ms, to eliminate the clock drift between devices; the power spectrum density (PSD) of each frequency band, such as 1-50 kHz and 50-100 kHz, is normalized by Z-score to eliminate the absolute amplitude difference between devices; and a multi-device ripple feature matrix is constructed. wherein N is the number of data samples, and F is the number of frequency band features, such as 100 frequency points.

[0077] S302: Blind source separation and component decoupling of independent component analysis (ICA); ICA algorithm selection and initialization, FastICA algorithm is selected based on negative entropy maximization, which has high separation efficiency for non-Gaussian noise and low calculation complexity, and meets the real-time requirement; after the input matrix X is centrally processed with zero mean, the mixing matrix A is initialized; the independent component s = WX is solved by iterative optimization, wherein W is the separation matrix; the output s contains K statistically independent components, K is determined by data characteristics, and K is usually taken as 3-5, each component corresponds to a potential noise source; the time-frequency features and component contribution degree of the output independent component are output, wherein the time domain waveform and spectral distribution of each component , such as dominant frequency band and energy concentration degree; the component contribution degree is determined by calculating the variance proportion of each component, such as component 1 contribution 65% and component 2 contribution 20%.

[0078] S303: Accurate identification and engineering classification of dominant ripple mode, component characteristic quantitative analysis is performed, and the component characteristic quantitative analysis table is as follows:

[0079]

[0080] In the mode determination logic, the dominant mode is screened by retaining the components with contribution degree > 50% and correlation threshold > 0.5, such as switch noise component contribution degree 75% and load correlation 0.85, which is determined as the switch noise dominant mode; according to the feature matrix, the known noise source type of the project is matched to generate the dominant ripple mode label, such as mode A high-frequency switch noise dominant, and the dominant ripple mode list containing mode label, core feature and contribution degree is output; the ICA output is mapped to the noise source that can be understood by the project through the quantitative index, avoiding subjective judgment and ensuring the objectivity and reliability of the mode recognition.

[0081] S304: Establishing the precise mapping relationship between the dominant mode and the system control parameters; for each identified dominant mode, the system uses historical data to build a regression model that can quantitatively describe how the ripple intensity of the mode changes with the changes in key control parameters; for example, for the switch noise dominant mode, the model reveals the specific dependence between its amplitude and the switching frequency, dead time, and linear cell input-output voltage difference. This model is strictly cross-validated to ensure its prediction accuracy meets the predetermined standard. This mapping relationship is the bridge connecting noise identification and control optimization, converting noise characteristics into adjustable hardware parameters.

[0082] S305: Generation and packaging of suppression strategies; when the cloud monitors that the real-time ripple characteristics of a device match a dominant mode, it immediately invokes the corresponding mapping relationship to calculate a set of optimal control parameter fine-tuning values; for example, if excessive switch noise is identified, a specific adjustment value to reduce the switching frequency or optimize the dead time is generated; the fine-tuning value, along with the mode label and confidence information, is packaged into a complete strategy package, ready to be sent to the corresponding edge node.

[0083] Through the steps of data preprocessing, blind source separation, mode recognition, parameter mapping, and strategy generation, complex ripple problems are transformed into quantifiable, traceable, and intervenable engineering tasks. The precise identification of ripple sources directly outputs executable optimization instructions.

[0084] The process of generating corresponding cloud optimization temperature drift compensation functions and control parameter fine-tuning values is as follows:

[0085] Under the premise of ensuring data privacy, a set of high-precision, adaptive temperature drift compensation schemes and control parameter adjustment instructions are tailored for each high-voltage power supply device, achieving coordinated improvement of the overall system performance. Unlike simply deploying a global model directly, this is achieved through a two-level architecture of a general model and localized adaptation.

[0086] Building a general temperature drift base model:

[0087] Based on the multi-device desensitized data uploaded by all edge nodes, a general temperature drift base model is trained using federated learning technology, capturing the common laws of voltage drift caused by temperature changes in high-voltage power supply systems; for example, it can accurately describe the general drift trend of output voltage when the ambient temperature rises from twenty degrees Celsius to fifty degrees Celsius under typical working conditions. A lightweight nonlinear function is used, with key state variables such as radiator temperature, load current, and working time as inputs, and the predicted temperature drift offset as output. Federated learning is used, and the entire training process does not require access to any single device's original sensitive data, only by aggregating the model update parameters calculated locally by each device to complete, fully ensuring user data security.

[0088] Extracting device individual characteristics and generating a localized adapter: On the basis of the general model, the cloud performs localization adaptation for each specific edge computing node, analyzes its unique behavior characteristics by calling the desensitized data uploaded by the device in history, and the characteristics may be caused by differences in component batches, slight deviations in assembly process, aging characteristics caused by long-term use, or unique load patterns in specific application scenarios; Based on individual characteristics, the cloud trains a lightweight localized adapter, that is, a small and targeted correction is added to the output of the general model; For example, a device with a reference voltage source batch problem shows greater positive drift at high temperature than the average level, and its localized adapter will generate a negative compensation offset to accurately offset this unique deviation.

[0089] Correlation dominant ripple pattern generates control parameter fine tuning amount:

[0090] The cloud generates specific control parameter fine tuning amount for each device based on the dominant ripple source information identified by independent component analysis in the previous step, and checks whether the recent ripple characteristics of the device are highly related to some undesirable mode, such as excessive switch noise. If highly related, a set of optimal adjustment instructions is calculated according to the established pattern and parameter mapping relationship.

[0091] The cloud fuses the above two results, i.e. the combination of the general model and the localized adapter, and the control parameter fine tuning amount, to form a complete optimization strategy package for the edge node. Before being officially issued, the system will simulate and verify the strategy in the digital twin environment of the cloud. By loading the historical running data of the device, the output effect after applying the new strategy is simulated, the expected income in terms of temperature drift suppression and ripple reduction is evaluated, and it is ensured that no new stability risk is introduced. Only when the verification result meets the preset performance threshold, such as temperature drift error less than 10ppm and ripple reduction more than 20%, the strategy package will be marked as valid and enter the delivery queue.

[0092] Strategy packaging and targeted delivery preparation:

[0093] The effective strategy package is packaged into a standardized data structure, which explicitly includes: version number and core parameters of the general temperature drift base model; The device-specific localized adapter coefficient; The fine tuning instructions of each control parameter and its applicable conditions; The effective period and rollback mechanism of the strategy; The packaged package is strictly bound with the unique identifier of the target edge computing node to ensure that the strategy can be accurately and safely sent to the specified device.

[0094] Step four: send the cloud optimized temperature drift compensation function and control parameter fine tuning amount to the corresponding edge computing node, and replace the local original control strategy atomically;

[0095] The encapsulated optimization strategy package is sent to the target edge computing node through an encrypted communication channel, such as a secure transmission protocol based on TLS1.3; before receiving, the edge node verifies the cloud's identity certificate and checks whether the device unique identifier contained in the strategy package matches itself. Only when both the identity and the target are verified, the data package will be accepted and temporarily stored in a protected temporary storage area to prevent illegal injection or misdelivery;

[0096] After the edge node receives the strategy package, it immediately performs integrity checking, verifies the strategy package has not been tampered with using digital signature; checks whether the strategy version number is higher than the current local version to avoid repeated or downgraded updates; verifies whether the temperature drift compensation function and control parameter fine-tuning amount in the strategy are within the preset safety boundary, for example, the switch frequency adjustment range does not exceed ±10%, the temperature drift compensation offset does not exceed the maximum allowed value; if any of the checks fails, the strategy package will be discarded and an error log will be reported to the cloud;

[0097] After the check passes, through the atomic replacement stage, the entire replacement operation is considered as an indivisible whole, the real-time control engine of the edge computing node temporarily freezes the strategy reading, but the underlying hardware control loop, such as the basic PID stabilizer, continues to run, ensuring that the power output is not affected, the new temperature drift compensation function and control parameter fine-tuning amount are loaded into a separate backup memory area, through an atomic instruction such as CAS operation, the strategy reference pointer of the control engine is instantly switched from the old strategy memory address to the new strategy memory address, and after confirming that the new strategy has taken effect, the memory occupied by the old strategy is released.

[0098] After the strategy switching is completed, the edge node immediately starts the validation program, and continuously monitors key indicators such as output voltage stability, ripple level, and temperature drift residual error in the next several control periods, such as 100 milliseconds; compares the measured performance with the expected effect promised in the strategy package; if the verification result meets the expectation, it sends an update success confirmation to the cloud and marks the new strategy as the current active strategy; if an abnormality occurs, such as output fluctuation exceeding the limit, control divergence, etc., the system will automatically trigger the rollback mechanism within milliseconds: immediately switch the strategy pointer back to the old version and report the fault details; whether the update is successful or not, the edge node will generate detailed audit logs, record the strategy version, switching time, verification result and system state snapshot, and upload them to the cloud in the next communication window period.

[0099] The above formulas are obtained by collecting a large amount of data for software simulation and selecting a formula close to the true value, and the coefficients in the formulas are set by the person skilled in the art according to the actual situation. The above are only preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes within the technical range disclosed by the present application according to the technical solution and the inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration, characterized in that, The steps of this method include: Step 1: On a high-voltage power supply device containing a linear voltage regulation unit and a high-frequency switching unit, an edge computing node is built to collect sensing data, including voltage, current, temperature and ripple spectrum characteristics. After feature extraction and desensitization processing is completed locally, the data is periodically uploaded to the cloud to form desensitized data from multiple devices. Step 2: Based on the perceived data, perform dual-modal collaborative control on the edge computing node to generate real-time control signals that act on the linear voltage regulation unit and the high-frequency switching unit respectively; Step 3: Based on the de-identified data from the multiple devices, a general temperature drift model is constructed through federated learning, and the dominant ripple source is identified by independent component analysis. Corresponding cloud-optimized temperature drift compensation functions and control parameter fine-tuning amounts are generated for each edge computing node. Step 4: Send the cloud-optimized temperature drift compensation function and control parameter fine-tuning amount to the corresponding edge computing node, and atomically replace the original local control strategy.

2. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 1, characterized in that, The process of constructing edge computing nodes is as follows: By deploying a multi-physical quantity fusion sensing network, sensing data is collected from the linear voltage regulation unit and high-frequency switching unit of the high-voltage power supply. The data is then pre-processed using the isolated signal conditioning circuit built into the edge gateway, converted into digital signals, and used to construct edge computing nodes.

3. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 1, characterized in that, The process of completing feature extraction and desensitization locally is as follows: Wavelet packet decomposition algorithm is used to decompose the sensing data at multiple scales and extract the energy features and harmonic components of each frequency band; based on the three-dimensional temperature data collected by distributed temperature sensors, a temperature field distribution model is constructed using gradient descent fitting method; fast Fourier transform is performed on the ripple time-domain signal to convert it into a frequency domain spectrum; principal component analysis is used to fuse and reduce the dimensionality of multi-dimensional features of voltage, current, temperature and ripple spectrum to generate a low-dimensional comprehensive feature vector; Sensitive fields are identified in the perceived data to obtain the de-identified information categories, including device identity, absolute value, and scene association. E-differential privacy technology is used to determine the privacy budget ε, calculate the noise intensity, and perform multi-dimensional validity verification on the perceived data after noise injection.

4. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 3, characterized in that, The process of generating multi-device de-identified data is as follows: The sensory data, which has undergone feature extraction and desensitization locally, is encrypted using AES-256 and uploaded to the cloud via a 5G link. The cloud deploys a multi-device data aggregation service, which receives the desensitized data uploaded from each edge node, aligns the data based on timestamps and feature dimensions, and stores it according to device type and operating scenario to build a desensitized dataset covering multiple high-voltage power supply devices.

5. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 1, characterized in that, The process of performing dual-modal cooperative control is as follows: Dual-mode coordinated control includes steady-state and dynamic operating conditions. Under steady-state conditions, the linear voltage regulation unit is given priority for adjustment; under dynamic operating conditions, the high-frequency switching unit actively pre-adjusts the intermediate bus voltage.

6. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 1, characterized in that, The process of constructing the general temperature drift model is as follows: The cloud receives anonymized data uploaded from various edge nodes, groups and stores it by device to form a structured database; it deploys a basic temperature drift model as the starting point for the global model. Each high-voltage power supply device trains its model locally on the edge computing node based on its own historical anonymized data, calculates model parameter updates, and collects model updates from all devices in the cloud. The cloud then merges these updates into a new global model through weighted averaging. This process is repeated iteratively until the model converges. The global model obtained through federated learning is then fine-tuned in the cloud using the historical anonymized data of each device to generate the corresponding adapter, taking into account the individual differences of each device.

7. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 6, characterized in that, The process of analyzing and identifying the dominant ripple source is as follows: S301: The cloud receives desensitized ripple spectrum characteristic data uploaded from all edge nodes; S302: Blind source separation and component decoupling in independent component analysis (ICA); S303: Identification and engineering classification of dominant ripple modes, and quantitative analysis of component characteristics; S304: Establish the mapping relationship between the dominant components of higher-order harmonics and system control parameters; S305: Generation and encapsulation of suppression strategies.

8. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 7, characterized in that, The process of generating the corresponding cloud-optimized temperature drift compensation function and control parameter fine-tuning amount is as follows: Based on the anonymized data from multiple devices uploaded by all edge nodes, a general temperature drift model is trained using federated learning technology. On the basis of the fine-tuned temperature drift model, the cloud performs localized adaptation for each edge computing node. The cloud-based system identifies the dominant ripple source information through independent component analysis, generates control parameter fine-tuning values ​​for each device, and integrates the localized temperature drift compensation function and control parameter fine-tuning values ​​to form an optimization strategy package.

9. The low-ripple high-voltage power ratio optimization management method based on cloud-edge collaboration according to claim 8, characterized in that, The process of atomically replacing the original local control strategy is as follows: The encapsulated optimization strategy package is sent to the target edge computing node through an encrypted communication channel. After receiving the optimization strategy package, the target edge computing node immediately performs an integrity check. If the check passes, it performs an atomic replacement. After the strategy switch is completed, the edge node immediately starts the effectiveness verification procedure.

Images