An isolated power supply startup and delayed shutdown control circuit and method
Through multi-dimensional data acquisition and machine learning algorithms, intelligent power switch control of isolated power supplies is realized, solving the problems of low efficiency in load prediction, abnormal detection and fault diagnosis in the existing technology, and improving the accuracy and reliability of control , and realize delayed shutdown to ensure safe power outage of the equipment.
Patent Information
- Application Number
- CN202411159106.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-22
AI Technical Summary
The existing isolated power switch control method is difficult to comprehensively and accurately reflect the dynamic characteristics of the system, the load prediction accuracy is insufficient, the abnormal detection and fault diagnosis efficiency is low, the control strategy lacks adaptability, and it is difficult to achieve energy efficiency, safety and system performance. The best balance is achieved between the two, and delayed shutdown control cannot be achieved.
Using multi-dimensional data acquisition and advanced signal processing, using long and short-term memory network models for load prediction, combining multiple machine learning algorithms to achieve abnormal detection and fault diagnosis, and realizing the intelligent power switch process through dynamic safety margin calculation and multi-objective optimization Management, configure delay time to achieve delayed shutdown.
Improves the accuracy, efficiency and reliability of the switch-off control of isolated circuits, can achieve safe power outage without damaging the equipment, enhances processing efficiency and interpretability, and realizes parallelization of load prediction and fault prediction. Improves the security and reliability of the system.
Smart Images

Figure CN118677236B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of electric energy management and control, and in particular to an isolated power supply startup and delayed shutdown control circuit and method. Background Art
[0002] With the continuous increase in electricity demand and the increasing complexity of the power grid, efficient and reliable power management has become a key factor in ensuring the stable operation of the system. With its excellent electrical isolation performance, isolated circuits play an irreplaceable role in protecting sensitive equipment and improving system safety. However, the switching process of such circuits involves complex electrical and thermodynamic changes. If not properly controlled, it may cause voltage transients, current shocks, and thermal stress, which seriously affect the reliability and life of the system.
[0003] Although existing research on isolated circuits has made some progress, it still faces the following major technical problems in practical applications: Existing data acquisition and processing methods often fail to fully and accurately reflect the dynamic characteristics of the system, resulting in a lack of sufficient information support for control decisions. The accuracy of load prediction is insufficient, especially when facing complex and changeable load conditions, it is difficult to accurately predict future load changes, which affects the formulation of control strategies. The efficiency and accuracy of anomaly detection and fault diagnosis need to be improved. Existing methods often fail to detect potential system anomalies in a timely manner, or fail to accurately diagnose the type and severity of faults. The optimization control of the power-on and power-off process is not intelligent and flexible enough, and it is difficult to achieve the best balance between energy efficiency, safety and system performance. Existing control methods generally lack adaptive capabilities and are difficult to effectively respond to changes in system parameters and external environmental disturbances. These technical problems seriously restrict the performance and reliability of isolated circuit power-on and power-off control, and it is urgent to develop more advanced and intelligent control methods to solve them. In some scenarios, such as when placing many computer processing devices, server devices, communication devices and other devices, if these devices are suddenly shut down and powered off when they are working normally, the functions and information processed by the devices will be lost, affecting the service life of the devices. More seriously, the devices will be directly damaged, causing huge losses.
[0004] However, the prior art cannot realize delayed shutdown control. Therefore, further research and innovation are needed to solve the above problems in the prior art. Summary of the invention
[0005] The purpose of the invention is to provide an isolated power supply startup and delayed shutdown control circuit and method to solve the above-mentioned problems existing in the prior art.
[0006] Technical solution, a method for controlling startup and delayed shutdown of an isolated power supply, comprising the following steps:
[0007] S1. Initialize the power management system, obtain the initial threshold of the device configuration instruction set, and based on the initial threshold, collect and pre-process the real-time data of each device to obtain a feature vector;
[0008] S2. Based on the feature vector, the pre-configured long short-term memory network model is called to obtain the forecast sequence and calculate the statistical characteristics of the forecast sequence. The historical data and the forecast sequence are integrated using the adaptive weight method to obtain the comprehensive load index.
[0009] S3, taking the feature vector as input, calling the pre-configured anomaly detection model and fault diagnosis model, and combining the comprehensive load indicators to obtain potential anomalies and fault prediction results;
[0010] S4. Obtain the current system status and fault prediction results, perform adaptive delay calculation, obtain the delay time, and configure it to each device in the system.
[0011] In another embodiment of the present application, an isolated power supply startup and delayed shutdown control circuit is used to execute the isolated power supply startup and delayed shutdown control method, including:
[0012] Photoelectric isolation unit, used to completely isolate input signals from output signals;
[0013] A data acquisition unit, used to collect real-time data of the equipment;
[0014] The intelligent control unit includes a data processing module, an abnormality detection module and a fault diagnosis module.
[0015] Beneficial effects: The present invention provides high-quality feature vectors for the system through multi-dimensional data collection and advanced signal processing; uses the LSTM model to achieve accurate load prediction, and integrates historical data to provide reliable comprehensive load indicators and then combines multiple machine learning algorithms to achieve efficient anomaly detection and fault diagnosis; through dynamic safety margin calculation, multi-objective optimization and adaptive delay control, the intelligent management of the power on and off process is achieved; the present invention improves the accuracy, efficiency and reliability of the power on and off control of the isolated circuit; during specific execution, the power supply main power module can be delayed to power off, so that each computing and processing slave device can power off by itself first, so as to achieve safe power off of all devices without damage; in the present invention, load prediction and fault prediction can be achieved in parallel, which improves processing efficiency and explainability. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a flow chart of the present invention.
[0017] Figure 2 This is a flow chart of step S1 of the present invention.
[0018] Figure 3This is a flow chart of step S2 of the present invention.
[0019] Figure 4 This is a flow chart of step S3 of the present invention.
[0020] Figure 5 This is a flow chart of step S4 of the present invention.
[0021] Figure 6 It is a system structure topology diagram of the present invention.
[0022] Figure 7 1 is a circuit diagram of an isolating switch according to an embodiment of the present invention. DETAILED DESCRIPTION
[0023] like Figure 1 As shown, the present application proposes an isolated power supply startup and delayed shutdown control method, comprising the following steps:
[0024] S1. Initialize the power management system, obtain the initial threshold of the device configuration instruction set, and based on the initial threshold, collect and pre-process the real-time data of each device to obtain a feature vector;
[0025] S2. Based on the feature vector, the pre-configured long short-term memory network model is called to obtain the forecast sequence and calculate the statistical characteristics of the forecast sequence. The historical data and the forecast sequence are integrated using the adaptive weight method to obtain the comprehensive load index.
[0026] S3, taking the feature vector as input, calling the pre-configured anomaly detection model and fault diagnosis model, and combining the comprehensive load indicators to obtain potential anomalies and fault prediction results;
[0027] S4. Obtain the current system status and fault prediction results, perform adaptive delay calculation, obtain the delay time, and configure it to each device in the system.
[0028] like Figure 2 As shown, according to one aspect of the present application, in step S1, based on the initial threshold, real-time data of each device is collected and preprocessed to obtain a feature vector, which is further:
[0029] S11. Based on the initial threshold, collect real-time data of each device, including power data, usage data, performance data, system environment parameters and operation data;
[0030] S12, applying wavelet transform based on the operation data to obtain the denoised operation data; using fast Fourier transform algorithm to extract time domain features and frequency domain features from the denoised operation data;
[0031] S13, based on the power data, usage data and performance data, an adaptive moving average filtering algorithm is used to perform smoothing processing to obtain filtered data;
[0032] S14. Based on the filtered data, time domain features and frequency domain features, a principal component analysis method is used to perform dimensionality reduction to obtain a feature vector after dimensionality reduction.
[0033] In one embodiment of the present application, a preconfigured initial threshold value of a key performance indicator of the system is obtained, wherein the process of configuring the initial threshold value is as follows: receiving system configuration information, generating a device registry R={r1,r2, ..., rn} containing information of n devices, wherein ri represents the basic information and functional characteristics of the ith device; constructing an m×k-dimensional power state matrix M according to the device registry R, wherein m represents the number of devices, and k represents the number of predefined power states; generating a system strategy set Π={π1, π2, ..., πj} based on preset rules and historical data, wherein πi represents the ith power management strategy; using a machine learning algorithm, training an energy consumption model E=f(D, L, C) according to historical energy consumption data, wherein D represents a set of device parameters, L represents load conditions, and C represents environmental conditions; for each device di, training a corresponding power curve model Pi=gi(ui, si), where ui represents the equipment utilization rate, si represents the equipment status, and gi represents the power curve function; collect system operation constraints and establish the system constraint set Ω, including minimum performance requirements and maximum power limits; based on historical performance data and expert knowledge, set the initial thresholds of key performance indicators T={t1, t2, ..., tp}, where ti represents the threshold of the i-th performance indicator.
[0034] Through the distributed sensor network, the power data Pi={pi1, pi2, ...,pim} of each device di is collected in real time, where pij represents the jth power parameter; the system monitoring module is used to obtain the usage status Ui={ui1, ui2, ..., uil} of each device di, where uij represents the jth usage indicator; high-precision performance monitoring equipment is used to collect the performance data Vi={vi1, vi2, ..., vik} of each device di, where vij represents the jth performance indicator; the system environmental parameters E={e1, e2, ..., ej} are obtained through the environmental sensor array, where ei represents the i-th environmental factor; the dedicated data acquisition unit is used to collect the operation data Oi={oi1, oi2, ..., oin} of each device di, including n parameters such as temperature, vibration, and current.
[0035] Wavelet transform is applied to the collected raw data Oi to obtain the denoised data O'i. The fast Fourier transform (FFT) algorithm is used to extract the time domain features FT={ft1, ft2, ..., ftp} and frequency domain features FF={ff1, ff2, ..., ffq} from the denoised data O'i. The adaptive moving average filtering algorithm is used to smooth the power data Pi, usage data Ui and performance data Vi to obtain the filtered data P'i, U'i and V'i. The principal component analysis (PCA) method is used to reduce the dimension of all preprocessed data to obtain the reduced dimension feature vector X={x1, x2, ..., xm}.
[0036] In this embodiment, the initial threshold is used as a benchmark to comprehensively collect real-time data of the equipment, including power data, usage data, performance data, system environment parameters and operation data, and establish a multi-dimensional data foundation. This all-round data collection ensures that subsequent analysis can fully consider all aspects of the system and improve the accuracy and reliability of decision-making. The wavelet transform is applied to the operation data for noise reduction, which effectively removes high-frequency noise and improves the signal-to-noise ratio of the signal. This step lays a clear data foundation for subsequent feature extraction, making the extracted features more accurate and representative. The fast Fourier transform (FFT) algorithm is used to extract time domain features and frequency domain features from the noise-reduced operation data. This dual analysis of time and frequency domains enhances the representativeness and discrimination of the features, enabling the system to simultaneously capture the time variation law and frequency distribution characteristics of the data. For power data, usage data and performance data, an adaptive moving average filtering algorithm is used for smoothing, which effectively eliminates the impact of short-term fluctuations and highlights the long-term trend of the data. This processing method improves the stability and reliability of subsequent predictions. The principal component analysis (PCA) method is used for dimensionality reduction, which not only reduces the redundancy of the data, but also retains the most important information in the data, improving the efficiency of subsequent processing. This multi-level, multi-angle data preprocessing and feature extraction method provides high-quality input data for the power on and off control of isolated circuits, allowing the system to perceive the current state more accurately.
[0037] In another embodiment of the present application, if the system is relatively complex (the complexity is greater than a preset value), segmentation processing is performed, and step S11 also includes a dynamic adaptive feature extraction process, specifically:
[0038] S111, receiving real-time data from each device, segmenting the real-time data to obtain multiple data segments; performing adaptive discrete cosine transform on each data segment to obtain DCT coefficients; calculating the optimal threshold according to the SURE (Stein's Unbiased Risk Estimate) criterion; based on the optimal threshold, selecting key DCT coefficients to form a preliminary feature set;
[0039] S112, input the preliminary feature set into the non-subsampled contourlet transform system to generate a multi-scale decomposition result; for the decomposition result of each scale, apply the adaptive DCT feature extraction method to obtain the feature subset of each scale; integrate the feature subsets of all scales to construct a multi-scale feature matrix, which is the real-time data after decomposition according to the time period.
[0040] In another embodiment of the present application, in step S12, wavelet transform is applied based on the operation data to obtain the noise-reduced operation data, specifically:
[0041] S121, using wavelet transform to decompose the operation data into different frequency components corresponding to different time scales;
[0042] S122, extracting statistical features for each frequency component, including mean, standard deviation, skewness and kurtosis;
[0043] S123. Based on system characteristics and pre-stored expert knowledge, construct and optimize at least two time window sizes, and create a sliding window for each window size;
[0044] S124. Calculate the above statistical features in each sliding window and output a multi-scale feature matrix, in which each row represents a time point and each column represents a combination of different time scales and statistical features.
[0045] In another embodiment of the present application, in step S12, wavelet transform is applied based on the operating data to obtain the denoised operating data, and denoising can also be performed through adaptive multi-fractal wavelet transform, that is: the original operating data sequence is obtained, and a normalization processing method is used to generate a standardized data set; using the standardized data set, multi-fractal spectrum analysis is performed to extract spectral feature parameters; based on the spectral feature parameters, an adaptive basis function family is constructed to calculate the wavelet coefficients; adaptive threshold processing is performed on the wavelet coefficients to reconstruct the signal; the denoising effect is evaluated, and the transformation parameters are optimized when necessary, and the final denoised data is output.
[0046] Specifically:
[0047] S12a, receiving the original running data sequence, performing a normalization operation, and obtaining a standardized data set; applying a sliding window method to the standardized data set, and calculating a local Hölder index; constructing a multifractal spectrum according to the Hölder index distribution; extracting characteristic parameters from the multifractal spectrum, including spectrum width and spectrum asymmetry;
[0048] S12b, based on the multi-fractal spectrum characteristic parameters, define a new basis function family; use dynamic parameter adjustment rules to optimize the basis function family parameters; use the optimized basis function family to perform wavelet transform on the standardized data, and calculate the wavelet coefficients under different scales and translations;
[0049] S12c, calculating the standard deviation of the wavelet coefficients at each scale and determining the adaptive threshold; performing soft threshold processing on the wavelet coefficients to obtain the noise reduction coefficient; using the noise reduction coefficient and the optimized basis function family, performing inverse transformation to reconstruct the signal; performing denormalization processing on the reconstructed signal to obtain preliminary noise reduction data;
[0050] S12d, calculating the signal-to-noise ratio improvement between the original data and the denoised data;
[0051] S12e, determine whether the signal-to-noise ratio improvement reaches a preset threshold; if not, adjust the basis function parameters and repeat steps S12b to S12d; after reaching the threshold or completing a predetermined number of iterations, output the final noise reduction data.
[0052] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:
[0053] S21. Based on the feature vector, construct a long short-term memory network model;
[0054] S22. Based on the long short-term memory network model, the load situation in the future time window is predicted to obtain a prediction sequence;
[0055] S23, calculating the statistical characteristics of the forecast sequence, including mean, standard deviation, peak and valley values;
[0056] S24. Based on statistical characteristics, an adaptive weight method is used to fuse historical data and forecast sequences to obtain a comprehensive load index.
[0057] In one embodiment of the present application, a long short-term memory network (LSTM) model M is constructed based on the feature vector X; the model M is used to predict the load conditions in the future time window W to obtain a prediction sequence F={f1, f2, ..., fw}; the statistical characteristics of the prediction sequence F are calculated, including the mean μ, standard deviation σ, peak value fmax and valley value fmin; and the historical data and the predicted data are fused using an adaptive weight method to obtain a comprehensive load index L.
[0058] In this embodiment, an LSTM model is constructed based on the high-quality feature vector obtained in the previous step. The special structure of the LSTM model can effectively capture the long-term dependencies in the time series data, which is crucial for power system load forecasting. Compared with traditional time series forecasting methods, LSTM can better handle nonlinear and non-stationary time series, and therefore can more accurately predict future load changes. The constructed LSTM model is used to predict the load conditions in the future time window to obtain a forecast sequence. This forecasting method based on deep learning can automatically learn complex patterns and features in the data without manually specifying the specific form of the forecasting model, thereby improving the flexibility and adaptability of the forecast. Again, statistical features are calculated for the forecast sequence, including mean, standard deviation, peak and valley values. These statistical features not only provide an overall estimate of the future load, but also reflect the volatility and extreme value of the load, providing more comprehensive information for subsequent decision-making. The adaptive weight method is used to fuse historical data and forecast sequences to obtain a comprehensive load index. The advantage of this fusion method is that it can dynamically balance the stability of historical data and the forward-looking nature of forecast data, and adaptively adjust the weights according to the current system state and forecast reliability, thereby obtaining a more reliable and comprehensive load assessment. This method of combining history and prediction can effectively reduce the deviation and uncertainty that may be caused by a single data source, and improve the accuracy and robustness of load assessment. Accurate load prediction and assessment are essential for the on / off control of isolated circuits. It can better predict future load changes, provide a reliable basis for subsequent on / off decisions, and help achieve more accurate and efficient power management.
[0059] In another embodiment of the present application, in step S21, the process of constructing the adaptive time window module is specifically as follows:
[0060] S211, construct a set of candidate time window sizes; for each window size, calculate the prediction accuracy using historical data;
[0061] S212, using autocorrelation function and partial autocorrelation function to analyze the periodicity and autocorrelation of the operating data, and using AHP method to select the optimal window size for each state based on historical performance and system dynamic characteristics;
[0062] S213. Based on the optimal window size, construct an optimal prediction time window size set for subsequent calls.
[0063] In another embodiment of the present application, step S2 further includes:
[0064] S25, receiving the feature vector after dimension reduction, constructing a fractional-order differential equation to describe the load change of the system; using the variational Bayes method, estimating the parameters of the fractional-order differential equation to obtain an optimized equation model;
[0065] S26. Input the optimized equation model into the fractional-order Adams numerical solver to calculate the future load value; post-process the calculation results and integrate the comprehensive load index obtained in step S24 to form the final load prediction sequence; the post-processing includes smoothing and outlier removal.
[0066] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:
[0067] S31, inputting the feature vector into a preconfigured Gaussian mixture model, and calculating the log-likelihood value of each data point;
[0068] S32, inputting the feature vector into a preconfigured autoencoder model to calculate a reconstruction error;
[0069] S33, based on the log-likelihood value, the reconstruction error and the comprehensive load index, a weighted sum method is used to calculate an anomaly score; if the anomaly score exceeds a preset threshold, the data point is marked as a potential anomaly point;
[0070] S34, based on the potential abnormal points, using a pre-trained multi-class support vector machine model to perform fault classification and obtain a classification result;
[0071] S35. Obtain historical fault data, and evaluate the severity of the fault based on the historical fault data and the classification results; and calculate a comprehensive risk index based on the severity of the fault;
[0072] S36. Based on the comprehensive risk index, a recurrent neural network model is used to obtain fault prediction results.
[0073] In one embodiment of the present application, a Gaussian mixture model G and an autoencoder model AE are trained using preprocessed data (feature vectors); model G is applied to a real-time data stream to calculate a log-likelihood value LL for each data point; real-time data is input into the autoencoder AE to calculate a reconstruction error RE; a weighted sum method is used to calculate an anomaly score AS by combining the log-likelihood value LL and the reconstruction error RE; if the anomaly score AS exceeds a preset threshold value θ, the data point is marked as a potential anomaly.
[0074] For the detected anomalies, a pre-trained multi-class support vector machine (SVM) model C is used to classify the faults. Based on the classification results and historical fault data, the fault severity S is evaluated. The comprehensive risk index RI is calculated by combining the fault category and severity. The recurrent neural network (RNN) model R is used to predict the fault development trend T based on historical fault data and the current system status. Based on the predicted fault trend T, the potential downtime Δt is estimated.
[0075] In this embodiment, the feature vector is input into a preconfigured Gaussian mixture model (GMM) to calculate the log-likelihood value of each data point. GMM can effectively model complex multimodal data distributions and is particularly suitable for processing multiple normal working conditions that may exist in power systems. By calculating the log-likelihood value, the system can quantify the degree of deviation of each data point from the normal state, providing an important indicator for anomaly detection. The feature vector is input into a preconfigured autoencoder model to calculate the reconstruction error. As an unsupervised learning method, the autoencoder can learn the internal representation of the data and reflect the degree of abnormality of the data through the reconstruction error. This method is particularly suitable for processing high-dimensional data and can capture subtle anomalies that may be overlooked by traditional methods. Then, based on the log-likelihood value, the reconstruction error and the comprehensive load index, the anomaly score is calculated using a weighted summation method. This multi-index fusion method can comprehensively consider abnormal information from different angles and improve the accuracy and robustness of anomaly detection. By setting a preset threshold, the system can flexibly balance the sensitivity and specificity of detection. For the detected potential anomalies, the system uses a pre-trained multi-class support vector machine (SVM) model for fault classification. As a powerful classifier, SVM can effectively handle classification problems in high-dimensional feature space and is particularly suitable for handling multiple types of faults that may exist in power systems. This classification method can not only identify the type of fault, but also provide confidence in the classification, providing more information for subsequent decision-making. In addition, the system also introduces fault severity assessment and comprehensive risk index calculation. By analyzing historical fault data and current classification results, the system can assess the severity of the fault and calculate the comprehensive risk index. This method not only considers the type of fault, but also its potential impact, providing a more comprehensive risk assessment for the switch decision. The system uses a recurrent neural network (RNN) model to predict the fault development trend based on the comprehensive risk index. The RNN model can effectively capture long-term dependencies in time series data and is particularly suitable for predicting the evolution of faults. This prediction capability enables the system to predict possible serious faults in advance, providing valuable time for preventive maintenance and emergency processing. For the switch control of isolated circuits, this embodiment improves the safety and reliability of the system. By timely detecting anomalies, accurately diagnosing faults, assessing risks and predicting trends, the system can take appropriate measures in the early stages of a fault to minimize downtime and potential losses.
[0076] In another embodiment of the present application, the operating data Oi = {oi1,oi2, ..., oim} of each device di is continuously collected, including m parameters such as temperature, vibration, and current. The original data Oi is subjected to noise reduction processing, and the cleaned data O'i is obtained using wavelet transform. Feature extraction is performed to calculate time domain features (mean, variance, peak factor, etc.) and frequency domain features (FFT coefficients, power spectrum density, etc.). Normal working mode modeling is performed, specifically: using Gaussian mixture model (GMM) Gi to fit the data distribution of each device under normal working conditions; calculating the normal working boundary Bi of each device, defined as the area in the GMM where the probability density is greater than the threshold θ1; using the autoencoder AEi to learn the low-dimensional representation of the normal operating data of each device.
[0077] Anomaly detection is performed, specifically: the log likelihood Li = logP(O'i | Gi) of the real-time data O'i in the GMM is calculated; if Li <θ2 (preset threshold), it is marked as a potential anomaly; the data is reconstructed using the autoencoder AEi, and the reconstruction error Ei = ||O'i - AEi(O'i)|| is calculated. 2 ; If Ei >θ3 (another preset threshold), further confirm the abnormal state. Construct a multi-class support vector machine (SVM) classifier Ci to map abnormal data to predefined fault types. For the detected abnormal data O'i, use the classifier Ci to perform preliminary fault classification and obtain the fault type probability distribution Pi; if the highest probability is less than the confidence threshold θ4, mark it as "unknown fault type". For each known fault type fj, establish a regression model Rj to evaluate its severity; input the abnormal data O'i into the corresponding regression model to obtain the fault severity score Sj; combine the fault type probability Pi and the severity score Sj to calculate the comprehensive risk index RI= Σ(Pi,j * Sj).
[0078] Perform time series pattern analysis, specifically: use the long short-term memory network (LSTM) Lj to learn the time series development pattern of each fault type; for the detected anomaly, input its historical data into the LSTM model to predict the future development trend Ti of the fault; based on the predicted trend Ti, estimate the downtime Δt that may be caused by the fault.
[0079] Construct a Bayesian network BN to model the causal relationship between equipment components; input the observed abnormal symptoms into the Bayesian network BN to infer the most likely root cause. Use the Markov Monte Carlo method to perform probabilistic reasoning on the Bayesian network BN to obtain the posterior probability of each potential cause.
[0080] Define the maintenance decision function: MD = f(RI, Δt, cost, availability), where cost is the maintenance cost and availability is the availability of spare parts and personnel; classify the maintenance tasks into "immediate execution", "planned" or "continued monitoring" based on the output of the maintenance decision function MD; for "planned" tasks, use genetic algorithms to optimize maintenance scheduling to minimize overall maintenance costs and downtime. Integrate the results of fault detection, classification, severity assessment, and root cause analysis; use natural language generation (NLG) technology to generate human-readable diagnostic reports; include visual elements in the report, such as fault trend graphs, component relationship diagrams, etc. Collect feedback from maintenance personnel and actual maintenance results; use the collected new data to update various models, including GMM, autoencoder, SVM, regression model, and LSTM; regularly evaluate model performance and trigger the model retraining process if performance degrades; use transfer learning technology to apply the learned knowledge to similar equipment or systems.
[0081] In another embodiment of the present application, it also includes:
[0082] S37. Based on the pre-stored system historical data and the final load forecast sequence, define the probability distribution family and construct the statistical manifold; calculate the Fisher information matrix to obtain the geometric structure information of the manifold. Solve the geodesic equation on the statistical manifold to describe the system state evolution trajectory; map the geodesic solution back to the original state space to obtain the system state prediction result;
[0083] S38, obtaining system state prediction results and real-time monitoring data, using algebraic topology methods to perform anomaly detection, and obtaining anomaly detection results; according to the anomaly detection results and system parameters, using algebraic geometry methods to perform fault diagnosis, and obtaining fault diagnosis results;
[0084] S39, receive system status prediction results and real-time monitoring data, construct a high-order Sheaf complex to represent the topological relationship between data; define open cover and restricted mapping, and establish Sheaf structure. Calculate the continuous homology on Sheaf and analyze the changes of homology groups; identify abnormal states according to the change pattern of homology groups and generate abnormal detection results;
[0085] S310. Based on the abnormal detection results and system parameters, a fault polynomial system is established to construct a fault manifold; the ideal Grobner basis of the fault polynomial is calculated using the Buchberger algorithm. The real-time observation data is substituted into the Grobner basis polynomial to calculate the residual value; the residual pattern is analyzed to identify the specific fault type and form a fault diagnosis result;
[0086] S311 . Compare the fault diagnosis result with the fault prediction result, and output the final fault prediction result.
[0087] like Figure 5 As shown, according to one aspect of the present application, in step S4, the current system state and the fault prediction result are obtained, and an adaptive delay calculation is performed to obtain the delay time, which is further:
[0088] S41, obtaining the current system state, constructing a safety margin function based on the current system state and the fault prediction result; and calculating the safety margin of the current system based on the safety margin function;
[0089] S42, based on the initial threshold of the device configuration instruction set, construct an energy optimization model, and calculate the Pareto frontier solution set; based on the current system state and the preset management goal, construct a judgment matrix of the hierarchical analysis method, calculate the eigenvector of the judgment matrix, and obtain the target weight vector;
[0090] S43, based on the target weight vector, calculating the weighted score of each solution in the Pareto front solution set; selecting the solution with the highest weighted score as the optimal configuration;
[0091] S44. Based on the optimal configuration and the safety margin, the delay time is calculated.
[0092] According to one aspect of the present application, the configuration to each device in step S4 is specifically:
[0093] S45. Generate an execution plan based on the optimal configuration and delay time;
[0094] S46, executing device configuration and shutdown instructions according to the execution plan;
[0095] S47, using the real-time monitoring module to continuously collect response data and status data of each device;
[0096] S48. Update the parameters of the power management system based on the response data and the status data.
[0097] In one embodiment of the present application, based on the current system state S and the fault prediction result T, a safety margin function h(S, T) is constructed; the safety margin m=h(S, T) of the current system is calculated; based on the system objectives and constraints, an objective function set F={f1, f2, ..., fm} is defined, where fi represents the i-th optimization objective; the system constraint set Ω is converted into a mathematical inequality constraint G={g1, g2, ..., gn}; and a multi-objective optimization problem model P is constructed: minimize F(x) subject to G(x)≤0, where x represents the system configuration vector.
[0098] Initialize the parameters of the non-dominated sorting genetic algorithm NSGA-III, including population size N, number of iterations T, crossover rate Pc, and mutation rate Pm; use the NSGA-III algorithm to solve the constructed optimization problem P and obtain the Pareto front solution set PS; apply the simulated annealing algorithm SA to each solution xi in the solution set PS for local search to obtain the optimized solution set PS'; calculate the hypervolume index HV of the solution set PS' to evaluate the quality and diversity of the solution. Based on the current system status and management goals, construct the judgment matrix J of the analytic hierarchy process (AHP); calculate the eigenvector of the judgment matrix J to obtain the target weight vector W={w1, w2, ..., wm}; for each solution xi, calculate its weighted score Si=Σ(wj*fj(xi)), where fj(xi) represents the performance or fitness value of solution xi on target j; select the solution x* with the highest score as the optimal configuration; map x* to a specific device configuration instruction set I={i1,i2, ..., in}. Get the execution time sequence T={t1, t2, ..., tn} corresponding to the optimal configuration x*; calculate the final delay time t delay =max(T)+m.
[0099] According to the instruction set I and the delay time t delay , generate an execution plan E; execute device configuration and shutdown instructions in sequence according to plan E; use the real-time monitoring module M to continuously collect the response data Ri and status data Si of each device di; based on the collected data, update the power state matrix M and the system parameter set P.
[0100] In another embodiment of the present application, in the data collection stage, a device list D = {d1, d2, ..., dn} is obtained, where n is the number of connected devices; for each device di, a real-time status parameter set Si = {si1, si2, ..., sim} is collected, including m parameters such as current load, temperature, energy consumption, etc.; historical data Hi = {hi1, hi2, ..., hik} of each device is read, including the past k shutdown times; system-level parameters P = {p1, p2, ..., pj} are obtained, such as total power consumption, ambient temperature, and other j parameters.
[0101] Normalize the state parameter Si of each device di to obtain the normalized state data S'i = {s'i1, s'i2, ..., s'im}; calculate the historical average shutdown time t of each device avg_i = avg(Hi), where avg( ) represents the average function; the moving weighted average method is used to process historical data to obtain the trend adjustment factor f i =weighted moving_average(Hi), where weighted moving_average represents the moving weighted average function.
[0102] Construct a multi-layer perceptron neural network model M, with the number of input layer nodes being m+1 (m state parameters plus 1 historical average time); use historical data to train model M; for each device di, input the normalized state data S'i and the historical average shutdown time t avg_i To model M, get the estimated shutdown time t est_i ; Apply the trend adjustment factor and calculate the adjusted estimated time t adj_i = t est_i *f i . Construct a device dependency graph G, where nodes are devices and edges represent shutdown order dependencies; use a topological sorting algorithm to sort the device dependency graph G and obtain a shutdown sequence Q = {q1, q2, ..., qn}; calculate the cumulative shutdown time series T = {t1, t2, ..., tn}, where ti = sum(t adj_j for j in Q[:i]).
[0103] Dynamic priority adjustment, specifically: define the priority scoring function priority score (di) = w1*criticality(di) + w2*energy efficiency (di) + w3*health index (di), where w1, w2 and w3 represent weight coefficients, criticality(di) represents the criticality of device di, and energy efficiency (di) represents the energy efficiency of device di, health index (di) represents the health index of device di; the priority score E = {e1, e2,..., en} is calculated for each device; the optimal shutdown subsequence Q' is found by combining the topological sorting Q and the priority score E using the improved Bron-Kerbosch algorithm.
[0104] Define the system safety margin function safety margin (P) = base margin + f(p1, p2, ..., pj), where base margin represents the basic safety margin, and f(p1, p2, ..., pj) represents the additional safety margin calculated based on the current system state and various parameters (such as load, environmental conditions, etc.); calculate the current safety margin of the system m = safety margin (P); Calculate the final delay time t delay= max(T) + m.
[0105] Start the shutdown procedure and start monitoring the actual shutdown progress of each device; define the deviation function delta(di,t) = actual progress (di, t) - expected progress (di, t), where actual progress (di, t) represents the actual shutdown progress of device di at time t, and expected progress (di, t) represents the expected shutdown progress of device di at time t; if for any device di, |delta(di, t)| > threshold, trigger the recalculation process: update the device state Si and system parameter P; repeat the above steps to get the new delay time t delay _ new ; if |t delay _ new -t delay | >adjustment threshold , update t delay = t delay _ new , where adjustment threshold Indicates adjusting the threshold.
[0106] The final calculated delay time t delay Output to the main control system; execute the optimized shutdown sequence Q', and each device is shut down according to the calculated time point; throughout the process, the real-time adjustment mechanism is continuously executed until all devices are safely shut down.
[0107] Record the actual shutdown time t of each device actual _ i ; Calculate the prediction error error i = t actual _ i -t adj_i ; Use these new data to update the neural network model M and the historical data set Hi; adjust the weights w1, w2, w3 of the priority scoring function, using reinforcement learning methods such as Q-learning. Analyze the accumulated data regularly (such as weekly) to identify system performance bottlenecks and abnormal patterns; use genetic algorithms to optimize the structure and hyperparameters of the neural network M; update the device dependency graph G to reflect changes in system configuration; fine-tune the safety margin function safety margin (P), adapt to long-term changes in system behavior.
[0108] In this embodiment, a safety margin function is constructed based on the current system state and fault prediction results, and the safety margin of the current system is calculated. This dynamic safety margin calculation method can reflect the operating risk of the system in real time and provide an important safety reference for subsequent decision-making. An energy optimization model is constructed based on the initial threshold, and an advanced multi-objective optimization algorithm (such as improved NSGA-III) is used to calculate the Pareto front solution set. This method can simultaneously consider multiple optimization objectives (such as minimizing energy consumption, maximizing performance, etc.) and find the best balance between multiple objectives. By constructing a judgment matrix and calculating the target weight vector through the analytic hierarchy process (AHP), the system can dynamically adjust the importance of each objective according to the current state, realizing the flexibility and adaptability of the optimization process. Based on the target weight vector, the system calculates the weighted score of each solution in the Pareto front solution set and selects the solution with the highest score as the optimal configuration. This method not only considers multiple optimization objectives, but also flexibly adjusts the decision preference according to the current system state, realizing the intelligent and dynamic decision-making. Next, based on the optimal configuration and safety margin, the system calculates the delay time. This delay calculation method takes into account both the optimization objectives and the system safety, and can achieve a good balance between efficiency and safety. In the execution phase, the system first generates an execution plan based on the optimal configuration and delay time, and then executes the device configuration and shutdown instructions according to the plan. This execution plan based on the optimization results can ensure that the system performs power on and off operations according to the optimal strategy, maximize energy efficiency and minimize potential risks. During the execution process, the system uses the real-time monitoring module to continuously collect the response data and status data of each device. This real-time monitoring mechanism can capture the abnormal response or state change of the device in time and provide instant feedback to the system. Based on the collected response data and status data, the system updates the parameters of the power management system. This closed-loop feedback mechanism enables the system to continuously learn and adapt, improving the stability and efficiency of long-term operation. For the power on and off control of isolated circuits, this embodiment realizes intelligent management of the entire process from decision-making to execution. Through dynamic safety margin calculation, multi-objective optimization, adaptive delay control and real-time feedback update, the system can achieve efficient and flexible power on and off control while ensuring safety, thereby improving the efficiency and reliability of power management.
[0109] In another embodiment of the present application, the process of calculating the delay time in step S4 may also be:
[0110] S441. Based on the pre-calculated optimized control parameters and system safety requirements, a fuzzy cognitive map is constructed, and concept nodes and weight matrices are defined; the weights of the fuzzy cognitive map are dynamically updated using nonlinear Hebbian learning rules.
[0111] S442, iteratively calculating the fuzzy cognitive map until convergence, obtaining a stable state; converting the stable state into a specific delay time through a mapping function, and generating a final delay control strategy.
[0112] In step S441, the calculation process of the optimized control parameters is specifically as follows: receiving the system state prediction results and fault diagnosis results, constructing a fractional-order PID controller model; defining a comprehensive performance index function, including error terms and control input terms. Using an improved particle swarm optimization algorithm, optimizing controller parameters; applying Matignon's stability theorem, analyzing the stability of the closed-loop system, ensuring that the controller parameters meet the stability conditions, and obtaining optimized control parameters.
[0113] According to one aspect of the present application, the method further includes S5: monitoring and updating the configured system status, specifically:
[0114] S51. Based on the energy optimization model, use the model predictive control framework to obtain the prediction data of the equipment;
[0115] S52, continuously monitoring and recording the actual data of the configured device, calculating the deviation between the predicted data and the actual data, and triggering the adjustment mechanism if the deviation exceeds a preset deviation threshold;
[0116] S53, constructing a Kalman filter based on historical data and actual data; using the Kalman filter to update the system state;
[0117] S54. Recalculate the optimal configuration based on the updated system status.
[0118] In one embodiment of the present application, the deviation Δ=|OP| between the actual observation value O and the predicted value P is calculated; if Δ exceeds the preset threshold ε, the adjustment mechanism is triggered; a Kalman filter KF is constructed to fuse historical data H and real-time observation data O; the Kalman filter KF is used to update the system state estimate S'; based on the updated state S', the optimal control action a* is recalculated.
[0119] In another embodiment of the present application, a device registry R = {r1, r2, ..., rn} is created, where ri contains basic information and functional characteristics of device di; a power state matrix M is established, where rows represent devices and columns represent different power states (such as full power, low power, standby, and off); and a system policy set Π = {π1, π2, ..., πk} is initialized, where each policy πi defines power management rules for a specific scenario.
[0120] For each device di, collect real-time power data Pi = {pi1, pi2, ..., pim}, including m parameters such as voltage, current, power factor, etc. Obtain device usage Ui = {ui1, ui2, ..., uil}, including l indicators such as CPU usage, memory usage, network traffic, etc. Monitor environmental parameters E = {e1, e2, ..., ej}, such as room temperature, humidity, time, etc.
[0121] Carry out load forecasting, specifically: use the long short-term memory network (LSTM) model L, input historical load data H = {h1, h2, ..., ht}; predict the load situation in the future time window W, and obtain the forecast sequence F = {f1, f2, ..., fw}; calculate the statistical characteristics of the forecast load, such as mean μ, standard deviation σ, peak value fmax and valley value fmin. Construct a decision tree model D, input the current system state S = {P1, ..., Pn, U1, ..., Un, E, μ, σ, fmax,fmin}; traverse the decision tree D, and select the most suitable strategy π*∈Π according to the input current system state S; if the decision confidence is lower than the threshold θ, start the reinforcement learning module to optimize the strategy.
[0122] Perform reinforcement learning optimization, specifically: define the state space S, action space A (different power management operations) and reward function R(s, a); use the deep Q network (DQN) Q(s, a) to approximate the action value function; execute the ε-greedy strategy, randomly select actions with probability ε, and select the action with the largest Q(s, a) with probability 1-ε; update the Q network: Q(s, a) is updated to: Q(s, a) +α[R + γmax Q(s', a') - Q(s, a)], where α is the learning rate, γ is the discount factor, s' is the next state reached after executing action a in the current state s, and a' is the action selected in the next state s'.
[0123] Perform power state transition, specifically: calculate the target power state Ti of each device according to the selected strategy π* and the current state S; for each device di, if Ti is different from the current state, generate a state transition command Ci; use the finite state machine (FSM) Fi to simulate the state transition process of each device to ensure the safety and feasibility of the transition.
[0124] Send state transition commands Ci to the corresponding devices in order of device priority; monitor the response and state changes of each device in real time, and update the power state matrix M; if an abnormality is detected (such as no response or inconsistent state of the device), trigger the abnormality handling process. For the detected abnormality ei, query the predefined abnormality handling rule base; execute the corresponding processing flow, which may include retry, rollback or alarm operations; record the abnormal information to the system log for subsequent analysis and improvement.
[0125] Carry out performance evaluation and learning, specifically: calculate the performance indicators of this management cycle, including energy efficiency η, response time τ and stability ρ;<S, π*, η, τ, ρ> Update the decision tree model D as a new training sample; use the new experience data to update the DQN experience playback buffer. Periodically analyze system logs and performance data to identify long-term trends and patterns; use genetic algorithms to optimize the strategy set Π and generate new candidate strategies; evaluate new strategies in a simulation environment and add strategies with excellent performance to Π; regularly retrain the LSTM model L and DQN model to adapt to long-term changes in the system.
[0126] In another embodiment of the present application, an energy consumption model E = f(D, L, C) is constructed, where D is the device parameter, L is the load condition, and C is the environmental condition; for each device di, a power curve model Pi = gi(ui, si) is established, ui is the utilization rate, and si is the device state; a system constraint set Ω is created, including minimum performance requirements, maximum power limits, etc. The energy consumption data Ei={ei1, ei2, ..., eim} and performance data Vi={vi1, vi2, ..., vik} of each device are collected in real time; environmental data C = {c1, c2, ..., ck}, such as temperature, humidity, time, etc. are obtained; moving average filtering is used to remove short-term fluctuations to obtain smoothed data E'i and V'i. Using time series decomposition, historical load data is decomposed into trend, seasonality, and residual components; the trend component is predicted using the ARIMA model, and the seasonal component is predicted using Fourier analysis; combined with the predicted trend and seasonality, the future load forecast L' = {l'1, l'2, ..., l't} is reconstructed.
[0127] Based on historical data, kernel density estimation (KDE) is used to establish the distribution model Di of each performance indicator; the dynamic threshold Ti = μi - k *σi of each indicator is calculated, where k is an adjustable parameter; according to the current load condition L' and environmental conditions C, the fuzzy logic system is used to dynamically adjust the k value.
[0128] Construct a multi-objective optimization problem, specifically: define the objective function set F = {f1, f2, ..., fm}, including energy minimization, performance maximization, temperature balance, etc.; transform the system constraint Ω into a mathematical inequality constraint; construct a multi-objective optimization problem: minimize F(x) subject toΩ, where x is the device configuration vector. Use the improved non-dominated sorting genetic algorithm (NSGA-III) to generate the initial Pareto frontier; apply local search algorithms (such as simulated annealing) to fine-tune the Pareto frontier to improve the quality of the solution; use the hypervolume indicator to evaluate the quality and diversity of the solution set. Use the analytic hierarchy process (AHP) to construct a decision matrix and assign weights to each objective according to the current system state; calculate the comprehensive score of each Pareto solution and select the solution with the highest score as the optimal configuration x*. Map the optimal configuration x* back to the actual device configuration parameters to generate the configuration instruction set I = {i1, i2, ..., in}.
[0129] Use the model predictive control (MPC) framework to predict future states based on the energy consumption model E and the power curve model Pi; solve the finite time domain optimization problem in each control cycle to obtain the optimal control sequence; execute the optimal control action at the current moment and update the system state. Continuously monitor the actual energy consumption Ea and performance index Pa, calculate the deviation ΔE and ΔP from the predicted value; if |ΔE| >εE or |ΔP| >εP (predefined tolerance threshold), trigger the correction mechanism; use the Kalman filter to update the system state estimate, combining the measured data and model prediction; recalculate the optimal control action based on the updated state estimate.
[0130] According to one aspect of the present application, a dynamic strategy selection method is also included, specifically: using preprocessed data to train a decision tree model D; inputting the current system state S into the model D to select the most suitable power management strategy π*; calculating the confidence C of the strategy π*; if C is lower than a preset threshold θ, starting the reinforcement learning module RL to optimize the strategy; using the RL module to generate an optimized strategy π**, and updating the strategy set Π.
[0131] In another embodiment of the present application, an isolated power supply startup and delayed shutdown control circuit is used to execute the isolated power supply startup and delayed shutdown control method, including:
[0132] Photoelectric isolation unit, used to completely isolate input signals from output signals;
[0133] A data acquisition unit, used to collect real-time data of the equipment;
[0134] The intelligent control unit includes a data processing module, an abnormality detection module and a fault diagnosis module.
[0135] According to one aspect of the present application, the photoelectric isolation unit is specifically an isolation switch circuit, which includes a power self-locking switch J1, a capacitor C1, a resistor R2, a resistor R3, a resistor R4, a resistor R5, a resistor R6, a resistor R7, a resistor R8, a resistor R9, a resistor R10, a resistor R11, a resistor R18, a resistor R19, an optical coupler D1A, an optical coupler D1B and an optical coupler D1D, wherein the emitter of the photosensitive transistor of the optical coupler D1A is connected to one end of the resistor R6, and the other end of the resistor R6 is connected to The collector of the phototransistor of the optocoupler D1A is connected to one end of the resistor R3, the other end of the resistor R3 is connected to the operating voltage, the anode of the light-emitting diode of the optocoupler D1A is connected to one end of the resistor R2 and the anode of the light-emitting diode of the optocoupler D1B at the same time, the other end of the resistor R2 is connected to one end of the resistor R19 and pin 2 of the power self-locking switch J1 at the same time, the other end of the resistor R19 is connected to the negative power supply reference ground, the pin 1 of the power self-locking switch J1 is connected to the standby power supply and one end of the capacitor C1 at the same time, the capacitor C The other end of 1 is connected to the negative power supply reference ground, the cathode of the light-emitting diode of the optocoupler D1A is connected to one end of the resistor R8, the other end of the resistor R8 is connected to the negative power supply reference ground, the cathode of the light-emitting diode of the optocoupler D1B is connected to one end of the resistor R9, the other end of the resistor R9 is connected to the negative power supply reference ground, the emitter of the phototransistor of the optocoupler D1B is connected to one end of the resistor R5, the other end of the resistor R5 is connected to the negative power supply reference ground, the collector of the phototransistor of the optocoupler D1B is connected to one end of the resistor R4, and the resistor The other end of R4 is simultaneously connected to one end of resistor R18 and the collector of the phototransistor of optocoupler D1D, the other end of resistor R18 is connected to the standby power supply, the emitter of the phototransistor of optocoupler D1D is connected to the negative power supply reference ground, the positive electrode of the light-emitting diode of optocoupler D1D is connected to one end of resistor R7, the other end of resistor R7 is connected to one end of resistor R10, the other end of resistor R10 is grounded, the negative electrode of the light-emitting diode of optocoupler D1D is connected to one end of resistor R11, and the other end of resistor R11 is grounded.
[0136] Compared with the prior art, the advantages of this embodiment are: the optocoupler switch circuit realizes multi-level signal control of complex circuits, so that the input and output ends are completely electrically isolated, and the anti-interference ability is strong, the service life is long, the response speed is fast, and frequent switching is still stable and reliable. When used with an intelligent control unit, the delayed shutdown time becomes reliable, accurate and adjustable, and the delay time can be set from seconds to minutes. The circuit of the present invention has strong versatility and can be applied to isolated startup control and delayed shutdown requirements in a variety of scenarios. After actual long-term use tests, the circuit of the present invention is very stable and reliable.
[0137] In another embodiment of the present application, when the power self-locking switch J1 is closed, the switch signal Pwrkey is disconnected from the power supply 24V_stby, and Pwrkey is pulled down to 24V_IN- through R19. At this time, the optocouplers D1A and D1B are not conducting; when the power self-locking switch J1 is opened, Pwrkey is connected to the power supply 24V_stby, and the optocouplers D1A and D1B are conducting. After D1B is conducting, PWR_delay is pulled down to 24V_IN-. When PWR_delay is at a high level, that is, when both D1B and D1D are cut off, the DC-DC power supply module does not work, and there is no 24V power output; when PWR_delay becomes a low level, that is, when D1B or D1D is conducting, the power supply module outputs the power supply 24V_IN+. Therefore, after D1B or D1D is turned on, the power supply module outputs the voltage normally, and each slave device also receives 24V_IN+ and starts working normally. After 24V_IN+ is output, 3.3V also has voltage, Figure 7 The minimum system of the intelligent control unit U1 works normally, and the V_detect and pwr_con signals are connected to the pins of the MCU as the input detection signal and the output control signal respectively. pwr_con is pulled down to GND by R10 by default, and D1D does not work at this time; when the device is powered normally, U1 also works normally, and both PWR_BIN and pwr_con output 3.3V high level. At this time, D1D is turned on, and PWR_delay is strongly pulled down to 24V_IN- through the light-receiving transistor inside D1D. PWR_delay is always at a low level, and the power module always outputs normally. In addition, D1A is also turned on, and the voltage of V_detect is about 0 after being divided by R3 (10KΩ) and R6 (330Ω), which is regarded as a low level, and the low level signal is transmitted to U1. Therefore, U1 believes that V_detect is a low level when the power self-locking switch J1 is turned on. When the power self-locking switch J1 is closed, Pwrkey is disconnected from the power supply 24V_stby and pulled down to 24V_IN-. At this time, D1A is cut off and D1B is also cut off. Since 3.3V still exists, D1D is still on, and PWR_delay is always at a low level. After D1A is cut off, V_detect is pulled up to 3.3V. At this time, V_detect is at a high level. When U1 detects that V_detect is at a high level, it is determined that the shutdown signal is recognized, so U1 outputs a low-level signal from PWR_BIN to all slave devices. After receiving the low-level signal from PWR_BIN, the slave device starts to shut down automatically. After U1 counts for 40 seconds (the time can be changed by the program), it outputs a low level from pwr_con. At this time, D1D is cut off, and PWR_delay is pulled up to 24V_stby again. The power module immediately shuts down and does not output 24V_IN+, thereby realizing the shutdown process in which the slave device shuts down first and the main power supply 24V_IN+ delays 40 seconds to power off.
[0138] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.
Claims
1. A method for controlling startup and delayed shutdown of an isolated power supply, characterized in that: The steps include: S1. Initialize the power management system, obtain the initial threshold of the device configuration instruction set, and based on the initial threshold, collect and pre-process the real-time data of each device to obtain a feature vector; S2. Based on the feature vector, the pre-configured long short-term memory network model is called to obtain the forecast sequence and calculate the statistical characteristics of the forecast sequence. The historical data and the forecast sequence are integrated using the adaptive weight method to obtain the comprehensive load index. S3, taking the feature vector as input, calling the pre-configured anomaly detection model and fault diagnosis model, and combining the comprehensive load indicators to obtain potential anomalies and fault prediction results; S4. Obtain the current system status and fault prediction results, perform adaptive delay calculation, obtain the delay time, and configure it to each device in the system; S2 is further: S21. Construct a long short-term memory network model and an adaptive time window module for the feature vector; S22, based on the feature vector after dimension reduction, a long short-term memory network model is used to predict the load situation in the future time window to obtain a prediction sequence; S23, calculating the statistical characteristics of the forecast sequence, including mean, standard deviation, peak and valley values; S24, based on statistical characteristics, an adaptive weight method is used to fuse historical data and forecast sequences to obtain a comprehensive load index; In step S4, the current system state and fault prediction results are obtained, and adaptive delay calculation is performed to obtain the delay time, which is further: S41, obtaining the current system state, constructing a safety margin function based on the current system state and the fault prediction result; and calculating the safety margin of the current system based on the safety margin function; S42, based on the initial threshold of the device configuration instruction set, construct an energy optimization model and calculate the Pareto frontier solution set; Based on the current system status and preset management goals, the judgment matrix of the hierarchical analysis method is constructed, the eigenvector of the judgment matrix is calculated, and the target weight vector is obtained; S43, based on the target weight vector, calculating the weighted score of each solution in the Pareto front solution set; selecting the solution with the highest weighted score as the optimal configuration; S44. Based on the optimal configuration and the safety margin, the delay time is calculated.
2. The isolated power supply startup and delayed shutdown control method according to claim 1, characterized in that: In step S1, based on the initial threshold, the real-time data of each device is collected and preprocessed to obtain a feature vector, which is further: S11. Based on the initial threshold, collect real-time data of each device, including power data, usage data, performance data, system environment parameters and operation data; S12, applying wavelet transform based on the operation data to obtain the denoised operation data; using fast Fourier transform algorithm to extract time domain features and frequency domain features from the denoised operation data; S13, based on the power data, usage data and performance data, an adaptive moving average filtering algorithm is used to perform smoothing processing to obtain filtered data; S14. Based on the filtered data, time domain features and frequency domain features, a principal component analysis method is used to perform dimensionality reduction to obtain a feature vector after dimensionality reduction.
3. The isolated power supply startup and delayed shutdown control method according to claim 2, characterized in that: Step S3 is further as follows: S31, inputting the feature vector into a preconfigured Gaussian mixture model, and calculating the log-likelihood value of each data point; S32, inputting the feature vector into a preconfigured autoencoder model to calculate a reconstruction error; S33, based on the log-likelihood value, the reconstruction error and the comprehensive load index, a weighted sum method is used to calculate an anomaly score; if the anomaly score exceeds a preset threshold, the data point is marked as a potential anomaly point; S34, based on the potential abnormal points, using the pre-trained multi-class support vector machine model to perform fault classification and obtain classification results; S35. Obtain historical fault data, and evaluate the severity of the fault based on the historical fault data and the classification result; Based on the severity of the fault, a comprehensive risk index is calculated; S36. Based on the comprehensive risk index, a recurrent neural network model is used to obtain fault prediction results.
4. The isolated power supply startup and delayed shutdown control method according to claim 3, characterized in that: The configuration to each device in step S4 is specifically as follows: S45. Generate an execution plan based on the optimal configuration and delay time; S46, executing device configuration and shutdown instructions according to the execution plan; S47, using the real-time monitoring module to continuously collect response data and status data of each device; S48. Update the parameters of the power management system based on the response data and the status data.
5. The isolated power supply startup and delayed shutdown control method according to claim 4, characterized in that: The system also includes S5, monitoring and updating the configured system status, specifically: S51. Based on the energy optimization model, use the model predictive control framework to obtain the prediction data of the equipment; S52, continuously monitoring and recording the actual data of the configured device, calculating the deviation between the predicted data and the actual data, and triggering the adjustment mechanism if the deviation exceeds a preset deviation threshold; S53, constructing a Kalman filter based on historical data and actual data; Use Kalman filter to update system status; S54. Recalculate the optimal configuration based on the updated system status.
6. The isolated power supply startup and delayed shutdown control method according to claim 5, characterized in that: In step S12, wavelet transform is applied based on the operation data to obtain the denoised operation data, specifically: S121, using wavelet transform to decompose the operation data into different frequency components corresponding to different time scales; S122, extracting statistical features for each frequency component, including mean, standard deviation, skewness and kurtosis; S123. Based on system characteristics and pre-stored expert knowledge, construct and optimize at least two time window sizes, and create a sliding window for each window size; S124. Calculate the above statistical features in each sliding window and output a multi-scale feature matrix, in which each row represents a time point and each column represents a combination of different time scales and statistical features.
7. The isolated power supply startup and delayed shutdown control method according to claim 6, characterized in that: In step S21, the process of constructing the adaptive time window module is specifically as follows: S211, construct a set of candidate time window sizes; for each window size, calculate the prediction accuracy using historical data; S212, using autocorrelation function and partial autocorrelation function to analyze the periodicity and autocorrelation of the operating data, and using AHP method to select the optimal window size for each state based on historical performance and system dynamic characteristics; S213. Based on the optimal window size, construct an optimal prediction time window size set for subsequent calls.
8. An isolated power supply startup and delayed shutdown control circuit, used to execute the isolated power supply startup and delayed shutdown control method according to any one of claims 1 to 7, characterized in that: include: Photoelectric isolation unit, used to completely isolate input signals from output signals; A data acquisition unit, used to collect real-time data of the equipment; The intelligent control unit includes a data processing module, an abnormality detection module and a fault diagnosis module.
Citation Information
Patent Citations
Intelligent power module management method and system
CN117454093A
Wind turbine generator remote fault diagnosis method and system based on cloud platform
CN117930815A
Fault positioning method and system based on line carrier
CN118444086A
