A multi-band high-voltage power supply control system and its normalized control and protection method

Abstract

The present invention discloses a multi-band high-voltage power supply control system and a normalized control and protection method thereof, which relates to the field of power supply control technology. The system collects multimodal signals of S, C, and Ku-band high-voltage power supply systems and cooling systems. The system uses the SDAE deep network in combination with the integrated learning AdaBoost framework to construct a fault diagnosis model, and obtains the fault type and confidence level based on signal characteristics. When a fault occurs, the high-voltage output of the corresponding power supply unit is cut off and reported through an isolation interface. A deep reinforcement learning agent is used to make intelligent decisions on protection actions. A digital twin model of the power supply unit is constructed, and an improved gray wolf optimizer is used for multi-task collaborative optimization. The optimized system parameters are obtained and sent to the PLC for execution to ensure the stability of the band output. The present invention achieves high reliability, rapid response, and intelligent operation and maintenance of the multi-band high-voltage power supply system through the deep integration of multi-source signal acquisition, intelligent fault diagnosis, adaptive protection decision-making, and digital twin optimization.

Description

Technical Field

[0001] The present invention relates to the field of power supply control technology, and more specifically, to a multi-band high-voltage power supply control system and a normalized control and protection method thereof. Background Art

[0002] High-voltage power supply systems are key components in powerful electromagnetic pulse (EMP) simulation tests, radar systems, and high-energy physics experiments. They drive microwave components such as klystrons and magnetrons to generate high-power electromagnetic pulses. Traditional multi-band high-voltage power supply systems typically utilize independently designed power modules for different bands (such as S-band, C-band, and Ku-band). This results in a complex system structure, decentralized control logic, and difficulties in achieving efficient coordinated management and fault protection.

[0003] Currently, existing technologies suffer from the following major issues: decentralized control and low management efficiency. High-voltage power supplies of different bands typically utilize independent control systems, resulting in a lack of unified management for parameter adjustment, status monitoring, and protection mechanisms. This increases operational complexity and reduces system reliability. Protection mechanisms also exhibit slow response times. During high-voltage power supply operation, faults such as klystron sparking, titanium pump overcurrent, and cooling failure can cause serious damage to the equipment. Traditional solutions have long protection response times and fail to meet high stability requirements. Electromagnetic compatibility (EMC) issues exist. High-voltage power supply systems operate in complex environments and are susceptible to strong electromagnetic interference. Conventional control modules lack adequate isolation and shielding measures, resulting in unstable signal acquisition and control command transmission. Cooling systems also lack coordination. Multi-band high-voltage power supplies generate significant heat during operation, but existing cooling systems often operate independently and lack intelligent linkage with power supply control. This can lead to localized overheating that impacts overall performance. To address the above issues, there is an urgent need for an integrated, highly reliable multi-band high-voltage power supply control system that adopts a normalized control and protection design to achieve unified management of multiple power modules, rapid fault protection, and intelligent cooling coordination to meet the stringent requirements of modern high-power microwave systems for stability, response speed, and EMC performance. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a multi-band high-voltage power supply control system and a normalized control and protection method thereof. Through the deep integration of multi-source signal acquisition, intelligent fault diagnosis, adaptive protection decision-making and digital twin optimization, the multi-band high-voltage power supply system is realized with high reliability, rapid response and intelligent operation and maintenance.

[0005] The first aspect of the present invention provides a multi-band high-voltage power supply control system, comprising: a multi-source signal acquisition and preprocessing module, an intelligent fault diagnosis module, a protection decision and execution module, and a digital twin and parameter optimization module;

[0006] The multi-source signal acquisition and preprocessing module is responsible for collecting multimodal signals of the S-band high-voltage power supply system, the C-band high-voltage power supply system, the Ku-band high-voltage power supply system and the cooling system, preprocessing the multimodal signals, and extracting signal features;

[0007] The intelligent fault diagnosis module combines the stacked denoising autoencoder network with the ensemble learning AdaBoost framework to build a fault diagnosis model, imports the signal features into the fault diagnosis model, and outputs the fault type and confidence level;

[0008] The protection decision and execution module trains a deep reinforcement learning agent and uses the deep reinforcement learning agent to make intelligent decisions on protection actions based on fault type and fault severity;

[0009] The digital twin and parameter optimization module constructs a digital twin model of the power supply unit. When the multi-band high-voltage power supply system is in a fault state, the digital twin model is used to assist in adjusting the electrical operating point parameters and thermal management parameters. The improved Gray Wolf optimizer is used to perform multi-task collaborative optimization, and the optimized system parameters are obtained and sent to the PLC for execution, and real-time monitoring and feedback are provided.

[0010] In this solution, the multi-source signal acquisition and preprocessing module is specifically:

[0011] A cross-band sensor network is constructed using pre-deployed sensors in the S-band high-voltage power supply system, the C-band high-voltage power supply system, the Ku-band high-voltage power supply system, and the cooling system to synchronously collect multimodal signals, perform data standardization on the multimodal signals, and add timestamps and device IDs;

[0012] Performing digital filtering and time-domain alignment on the multimodal signal to obtain a preprocessed multimodal signal, performing time-frequency domain feature extraction on the preprocessed multimodal signal to construct a feature matrix, obtaining clusters through K-means clustering, and identifying the operating conditions of the multi-band high-voltage power supply system;

[0013] The signal segments corresponding to the working conditions are extracted using a feature extraction layer that combines one-dimensional convolution and dilated convolution for local feature extraction. A channel attention mechanism is introduced to enhance the key frequency band features. The multimodal signals of the preset time window are captured using a BiLSTM network to capture the temporal dependencies. The self-attention layer is used to calculate the cross-band feature correlation and output the global features.

[0014] The local features are aligned with the global features through time axis interpolation, and a weighted splicing fusion strategy is used to perform feature fusion to obtain signal features.

[0015] In this solution, the intelligent fault diagnosis module is specifically:

[0016] A weak classifier is constructed based on a stacked denoising autoencoder network. A layer-by-layer greedy training strategy is used to pre-train the weak classifier. Gaussian noise is added to each hidden layer during the pre-training of the weak classifier for anti-interference training. After the pre-training is completed, a Softmax classifier is added to the top layer and the weak classifier is fine-tuned using labeled data.

[0017] The pre-trained weak classifiers are integrated using the AdaBoost framework. In the first round of iteration, all training samples are given the same weight. During the iteration process, the sample weights are dynamically adjusted according to the classification error of the weak classifiers, and weak classifiers that do not meet the preset classification accuracy are eliminated. When the preset number of iterations is reached, the weighted voting results of each weak classifier are used to form the final prediction. The test samples are used for verification. After verification, a strong classifier is generated to build a fault diagnosis model.

[0018] The fault diagnosis model is used for online monitoring. The acquired signal features are input into all weak classifiers in parallel. Each weak classifier outputs a fault probability prediction value. A comprehensive score is obtained by weighted summation. The threshold is dynamically adjusted according to the current working conditions. When the comprehensive score exceeds the dynamic threshold, a fault alarm is triggered, and the fault type and confidence level are output.

[0019] In this solution, the fault diagnosis model is enhanced in the intelligent fault diagnosis module, specifically:

[0020] After a fault alarm is triggered, the TEO energy characteristics, reflected wave integral characteristics, and phase mutation detection characteristics corresponding to the electrical data are extracted from the S-band high-voltage power supply system, the C-band high-voltage power supply system, and the Ku-band high-voltage power supply system as electrical impact characteristics.

[0021] The cavity temperature rise, vibration spectrum entropy, and flow pulsation corresponding to the physical data are extracted from the C-band high-voltage power supply system, Ku-band high-voltage power supply system, and cooling system as physical anomaly characteristics;

[0022] According to the electrical impact characteristics and physical abnormality characteristics, fault feature fingerprints of different bands are constructed, and the dynamic time warping algorithm is used to match the fault feature fingerprints with the fingerprint templates in the fault mode library. The preliminary screening fault type is obtained through the matching results, and the classification enhancement is performed in the fault diagnosis model based on the preliminary screening fault type.

[0023] In this solution, the protection decision-making and execution module is specifically:

[0024] The multimodal signal is feature-encoded to generate an original state vector, and the original state vector is embedded in the working condition context by combining the working condition label output by K-means clustering and the historical state memory to construct a state space;

[0025] Extract fault type nodes and protection measure nodes based on the fault mode library and historical fault protection measures, build edge connection relationships between nodes according to trigger relationships and exclusion relationships based on constraint conditions, generate a fault-measure knowledge graph, generate a basic action set of protection measures based on the fault-measure knowledge graph, and build an action space;

[0026] A deep reinforcement learning agent is constructed based on the state space and actions using an improved PPO algorithm. A pre-established three-level fault severity evaluation system is used to quantify fault losses. Action costs are evaluated using power loss and equipment wear. A basic reward function is constructed based on these fault losses and action costs.

[0027] Introducing knowledge rewards, constructing a current state vector, performing similarity matching between the current state vector and the fault mode in the fault-measure knowledge graph, constructing a knowledge reward function based on the similarity matching, introducing exploration rewards, calculating the Mahalanobis distance between the current state vector and the historical state vector in the historical state ring buffer, and constructing a knowledge reward function based on the Mahalanobis distance;

[0028] Knowledge graph constraints are introduced into policy updates, and penalty gradients are applied to actions that violate knowledge graph exclusion relationships. Non-compliant actions are dynamically blocked based on the current fault type. Fault records are extracted based on the fault mode library and historical fault protection measures to pre-train a deep reinforcement learning agent, and an ε-greedy strategy is used for online fine-tuning.

[0029] The trained deep reinforcement learning agent basic Q-value network and knowledge Q-value network are used to calculate the Q-value of each action. The temperature coefficient is adjusted according to the working conditions through the dynamic Boltzmann strategy, and the final action is selected as the fault protection measure.

[0030] In this solution, the digital twin and parameter optimization module is specifically:

[0031] Obtaining the electromagnetic field distribution within the S-, C-, and Ku-band klystrons and the three-dimensional temperature field distribution of the multi-band high-voltage power supply, constructing physical constraints based on the electromagnetic field distribution and the three-dimensional temperature field distribution, using historical multimodal signals as input to a physical information neural network, applying physical constraints in the hidden layer of the physical information neural network, and obtaining a state prediction value of the multimodal signal through the output layer;

[0032] Verify the state predictions of the multimodal signals using test data by obtaining a pre-trained digital twin model of the power supply unit and fine-tuning the digital twin model through online incremental training using the latest acquired multimodal signals;

[0033] Real-time multimodal signals are collected and imported into the digital twin model of the power supply unit to load the current state of the multi-band high-voltage power supply. A Gray Wolf optimizer is constructed to perform parameter tuning. An objective function is constructed with the goals of maximizing microwave power, maximizing klystron life, and minimizing cooling energy consumption. This objective function is then associated with the fitness function to calculate the fitness of the Gray Wolf.

[0034] An iterative optimization is performed based on the fitness of the gray wolf. When the number of iterations reaches the maximum number of iterations or the iteration termination condition is met, the position of the α wolf with the highest fitness in the last iteration is output to generate a parameter update suggestion. The update suggestion is simulated and verified using a digital twin, and the simulation results are compared with the expected performance improvement to output an update suggestion that meets the requirements.

[0035] In this solution, the Gray Wolf Optimizer is specifically:

[0036] The fitness of the gray wolf is calculated using a fitness function. The top three optimal solutions are selected as α wolf, β wolf, and δ wolf according to the fitness ranking, and the current global optimal solution is recorded. In the iterative calculation, the positions of α wolf, β wolf, δ wolf and other gray wolves are updated using the steps of surrounding, hunting and attacking prey. The Lévy flight strategy is used for α wolf, and Gaussian perturbations are performed on β wolf and δ wolf to generate a new gray wolf population. An inertia weight that changes with the number of iterations is introduced.

[0037] The fitness of the gray wolves in the new gray wolf population is calculated, and a mutation operator that changes with the number of iterations is introduced to mutate the α wolf with the highest fitness. The fitness of the α wolf before and after the mutation is compared, and the elite retention strategy is used to select the better solution to enter the next generation and guide the search direction.

[0038] When the number of iterations reaches the maximum number of iterations or the iteration termination condition is met, the position of the α wolf with the highest fitness in the last iteration is output to generate parameter update suggestions.

[0039] A second aspect of the present invention provides a normalized control and protection method for a multi-band high-voltage power supply, which is applied to a high-voltage power supply control system and includes the following steps:

[0040] Collecting multimodal signals of an S-band high-voltage power supply system, a C-band high-voltage power supply system, a Ku-band high-voltage power supply system, and a cooling system, preprocessing the multimodal signals, and extracting signal features;

[0041] A fault diagnosis model is constructed by combining a stacked denoising autoencoder network with an ensemble learning AdaBoost framework, the signal features are imported into the fault diagnosis model, and the fault type and confidence level are output;

[0042] When a fault occurs, the high-voltage output of the corresponding power supply unit is cut off and reported to the remote terminal through the isolation interface. A deep reinforcement learning agent is trained and used to make intelligent decisions on protection actions based on the fault type and fault severity.

[0043] A digital twin model of the power supply unit is constructed. When the multi-band high-voltage power supply system is in a fault state, the digital twin model is used to assist in adjusting the electrical operating point parameters and thermal management parameters. An improved Gray Wolf optimizer is used for multi-task collaborative optimization. The optimized system parameters are obtained and sent to the PLC for execution to ensure the stability of the band output.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention achieves high reliability, rapid response, and intelligent operation and maintenance for a multi-band high-voltage power supply system by deeply integrating multi-source signal acquisition, intelligent fault diagnosis, adaptive protection decision-making, and digital twin optimization technologies. The system uses the SDAE-AdaBoost integrated model to perform real-time analysis of multi-dimensional signals, including voltage, current, and vacuum level, from the S / C / Ku-band power supply and cooling system. This allows for precise detection of minor faults (such as abnormal filament current and titanium pump vacuum leaks), improving fault identification accuracy. Incorporating a deep reinforcement learning agent, the system can complete fault classification responses within milliseconds, significantly reducing the risk of klystron damage. A normalized control and protection design constructs a digital twin model using a physical information neural network, enabling virtual rehearsal of parameter tuning and reducing overshoot during the high-voltage ramp. An improved Grey Wolf algorithm collaboratively optimizes multiple objectives, including power output, equipment lifespan, and cooling energy consumption, to improve cooling system energy efficiency while ensuring the preset microwave power. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0047] Figure 1 A block diagram of a multi-band high-voltage power supply control system is shown.

[0048] Figure 2 It shows a flow chart of fault diagnosis performed by the intelligent fault diagnosis module in the embodiment;

[0049] Figure 3 It shows a flow chart of the protection decision and execution module generating the protection decision in the embodiment;

[0050] Figure 4The present invention shows a flow chart of a normalized control and protection method for a multi-band high-voltage power supply. DETAILED DESCRIPTION

[0051] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0053] Figure 1 A block diagram of a multi-band high-voltage power supply control system is shown.

[0054] This embodiment provides a multi-band high-voltage power supply control system, including: a multi-source signal acquisition and preprocessing module 101, an intelligent fault diagnosis module 102, a protection decision and execution module 103, and a digital twin and parameter optimization module 104;

[0055] The multi-source signal acquisition and preprocessing module 101 is responsible for acquiring multimodal signals of the S-band high-voltage power supply system, the C-band high-voltage power supply system, the Ku-band high-voltage power supply system, and the cooling system, preprocessing the multimodal signals, and extracting signal features;

[0056] The intelligent fault diagnosis module 102 combines the stacked denoising autoencoder network with the ensemble learning AdaBoost framework to construct a fault diagnosis model, imports the signal features into the fault diagnosis model, and outputs the fault type and confidence level;

[0057] The protection decision and execution module 103 trains a deep reinforcement learning agent and uses the deep reinforcement learning agent to make intelligent decisions on protection actions based on fault type and fault severity;

[0058] The digital twin and parameter optimization module 104 constructs a digital twin model of the power supply unit. When the multi-band high-voltage power supply system is in a fault state, the digital twin model is used to assist in adjusting the electrical operating point parameters and thermal management parameters. The improved Gray Wolf optimizer is used to perform multi-task collaborative optimization, and the optimized system parameters are obtained and sent to the PLC for execution, and real-time monitoring and feedback are provided.

[0059] It should be noted that the multi-band high-voltage power supply system includes the S-band, C-band, and Ku-band high-voltage power supply systems, as well as the cooling system. The S-band high-voltage power supply system includes two identical power supply systems (including control, titanium pump power supply, filament power supply, and DC high-voltage power supply), a pulse modulator system (pulse modulator and high-voltage pulse-changing components); the C-band high-voltage power supply system includes two identical power supply systems (including control, titanium pump power supply, filament power supply, and DC high-voltage power supply), a pulse modulator system (pulse modulator and high-voltage pulse-changing components); and the Ku-band high-voltage power supply system includes two identical power supply systems (including control, titanium pump power supply, filament power supply, and DC high-voltage power supply), a pulse modulator system (pulse modulator and high-voltage pulse-changing components). The cooling system includes cooling control, a cooling box, a cooling cycle, a heating device, a refrigeration device, a heat exchange device, and supporting sensors and links. The high-voltage power supply system for each band includes two sets of klystron power supply systems. To ensure the effective operation of the high-voltage power supply system, the system normalizes the control of the two power supply systems. Therefore, the system can be divided into a control and protection module, two sets of titanium pump power supplies, two sets of filament power supplies, two sets of DC high-voltage power supplies and two sets of pulse modulators and their pulse oil tank components. All external interfaces adopt isolation measures, including digital, analog and communication interfaces, and are equipped with shielding boxes to ensure their electromagnetic compatibility.

[0060] A cross-band sensor network is constructed using pre-deployed sensors in the S-band, C-band, and Ku-band high-voltage power supply systems and cooling systems. These sensors include high-voltage differential probes, Rogowski coils, capacitance vacuum gauges, electromagnetic flowmeters, and temperature sensor arrays. Multimodal signals, including electrical parameters (cathode voltage, filament current, modulator pulse waveform) and physical parameters (titanium pump vacuum, coolant flow, and tube temperature), are synchronously acquired. These multimodal signals are then normalized and timestamped with device IDs. Digital filtering and time-domain alignment are performed on these multimodal signals to obtain pre-processed multimodal signals. Time-frequency feature extraction is performed on these pre-processed multimodal signals to construct a feature matrix. K-means clustering is then used to obtain clusters and identify the operating conditions of the multi-band high-voltage power supply system. Five typical operating conditions—standby, preheating, steady-state operation, overload, and fault recovery—are automatically recognized. Each operating condition employs a unique feature extraction strategy (for example, transient response is prioritized for overload conditions). The signal segments corresponding to the working conditions are subjected to local feature extraction using a feature extraction layer combining one-dimensional convolution and dilated convolution. A channel attention mechanism is introduced to enhance the key frequency band features. The multimodal signals of the preset time window are captured using a BiLSTM network to capture timing dependencies. The self-attention layer is used to calculate cross-band feature correlations and output global features. The local features (μs level) and the global features (minute level) are aligned using time axis interpolation. The features are fused using a weighted splicing fusion strategy, and the fusion weights are dynamically adjusted according to the working conditions to obtain signal features.

[0061] Figure 2 A flow chart showing fault diagnosis performed by the intelligent fault diagnosis module in the embodiment is shown.

[0062] According to an embodiment of the present invention, the intelligent fault diagnosis module is specifically:

[0063] S202, constructing a weak classifier based on a stacked denoising autoencoder network, pre-training the weak classifier using a layer-by-layer greedy training strategy, and adding Gaussian noise to each hidden layer in the pre-training of the weak classifier for anti-interference training. After the pre-training is completed, adding a Softmax classifier to the top layer and fine-tuning the weak classifier using labeled data;

[0064] S204: Using the AdaBoost framework to integrate the pre-trained weak classifiers, all training samples are assigned the same weight in the first round of iteration. During the iteration process, the sample weights are dynamically adjusted according to the classification errors of the weak classifiers, and weak classifiers that do not meet the preset classification accuracy are eliminated. When the preset number of iterations is reached, the weighted voting results of each weak classifier are used to form the final prediction, which is verified using the test sample. After passing the verification, a strong classifier is generated to build a fault diagnosis model;

[0065] S206, use the fault diagnosis model to perform online monitoring, input the acquired signal features into all weak classifiers in parallel, each weak classifier outputs a fault probability prediction value, and obtains a comprehensive score through weighted summation. The threshold is dynamically adjusted according to the current working conditions. When the comprehensive score exceeds the dynamic threshold, a fault alarm is triggered, and the fault type and confidence level are output.

[0066] It should be noted that the stacked denoising autoencoder network serves as a basic feature extractor, and its three-layer network structure (input layer-hidden layer-output layer) obtains robustness to noise interference by adding Gaussian noise pre-training. The AdaBoost framework integrates multiple denoising autoencoder network weak classifiers into a strong classifier through an iterative mechanism, where the number of input layer nodes of each denoising autoencoder network corresponds to the feature dimension (such as a 50-dimensional feature vector), the number of hidden layer nodes is set to 60% of the input layer, and the ReLU activation function is used. The output layer generates a preliminary judgment result of the fault probability through the Sigmoid function. During the first round of stacked denoising autoencoder network training, 20% intensity Gaussian noise is added to the input data to enhance generalization ability. The classification error rate is calculated after each round of iteration. , when the classification error rate is greater than the preset threshold, the current weak classifier is automatically discarded. For misclassified samples, its weight is increased, so that the next round of stacked denoising autoencoder network pays more attention to the characteristics of difficult samples. After N rounds of iteration, the weight of each weak classifier is According to the formula The stacked denoising autoencoder network with better classification performance receives a higher voting weight. The final decision output of the strong classifier is the weighted sum of the outputs of each weak classifier. A confidence threshold of 0.7 is set to achieve reliable fault determination. The first stacked denoising autoencoder network focuses on capturing microsecond-level transient features, while subsequent stacked denoising autoencoder networks gradually extract features with longer time domains. Using AdaBoost's sample weight adjustment mechanism, the system automatically emphasizes the feature dimensions most sensitive to the current operating conditions. The hidden layer outputs of each stacked denoising autoencoder network are simultaneously input into the attention module to generate a feature importance heatmap for fault tracing. The output layer contains two key pieces of information: fault type classification (sparking, overcurrent, cooling failure, etc.) and a confidence score (0-100%). When an S-band current THD exceeds the limit, the model integrates the outputs of each stacked denoising autoencoder network. If the three main weak classifiers all determine a rectifier bridge fault (with confidence levels of 72%, 85%, and 68%), the final fault probability is 82%, exceeding the threshold and triggering an alert.

[0067] In the intelligent fault diagnosis module, the fault diagnosis model is enhanced. After the fault alarm is triggered, the TEO energy characteristics, reflected wave integral characteristics, and phase mutation detection characteristics corresponding to the electrical data are extracted from the S-band high-voltage power supply system, the C-band high-voltage power supply system, and the Ku-band high-voltage power supply system as electrical impact characteristics. When extracting the TEO energy characteristics, the original signal is subjected to a 5th-order Butterworth bandpass filter to eliminate baseline drift and high-frequency noise. The Teager energy operator is calculated using a sliding window to calculate the instantaneous energy of each sampling point. The peak-to-peak value and standard deviation of the energy in the window are statistically analyzed. Based on the energy distribution of the first n cycles, the energy distribution is calculated. An arc discharge is detected when the energy value exceeds the threshold for three consecutive windows and is accompanied by pulse waveform distortion greater than 50ns. The reflected wave integral feature extraction performs envelope detection on the reflected signal, extracts the pulse waveform envelope, calculates the integrated energy within a single pulse cycle, and divides the integrated value by the incident wave energy to obtain the reflection coefficient. When the reflection coefficient exceeds the preset threshold for five cycles, a waveguide connection abnormality alarm is triggered. Phase jump detection feature extraction uses a Hilbert transform to calculate the instantaneous phase, calculate the phase difference, and calculate the standard deviation of the phase difference over n normal cycles to establish a phase noise floor. Phase jump events are then recorded based on the phase noise floor. The cavity temperature rise, vibration spectrum entropy and flow pulsation corresponding to the physical data are extracted from the C-band high-voltage power supply system, Ku-band high-voltage power supply system and cooling system respectively as physical abnormality characteristics; fault feature fingerprints of different bands are constructed according to the electrical impact characteristics and physical abnormality characteristics, and the dynamic time warping algorithm is used to match the fault feature fingerprints with the fingerprint templates in the fault mode library. The preliminary screening fault type is obtained through the matching result, and the preliminary screening fault type is used to perform classification enhancement in the fault diagnosis model to further improve the fault identification capability of the fault diagnosis model.

[0068] Figure 3 A flow chart showing the generation of protection decisions by the protection decision and execution module in the embodiment is shown.

[0069] According to an embodiment of the present invention, the protection decision and execution module is specifically:

[0070] S302, feature encoding the multimodal signal to generate an original state vector, and embedding the original state vector into the working condition context by combining the working condition label output by K-means clustering and the historical state memory to construct a state space;

[0071] S304: Extracting fault type nodes and protection measure nodes based on the fault mode library and historical fault protection measures, constructing edge connection relationships between the nodes according to the trigger relationship and the exclusion relationship based on the constraint conditions, generating a fault-measure knowledge graph, generating a basic action set of protection measures based on the fault-measure knowledge graph, and constructing an action space;

[0072] S306, using the improved PPO algorithm to construct a deep reinforcement learning agent based on the state space and actions, using a pre-established three-level fault severity evaluation system to quantify the fault loss, and evaluating the action cost through power loss and equipment wear, and constructing a basic reward function based on the fault loss and action cost;

[0073] S308: Introduce knowledge rewards, construct a current state vector, perform similarity matching between the current state vector and the fault mode in the fault-measure knowledge graph, construct a knowledge reward function based on the similarity matching, introduce exploration rewards, calculate the Mahalanobis distance between the current state vector and the historical state vector in the historical state ring buffer, and construct a knowledge reward function based on the Mahalanobis distance.

[0074] S310 introduces knowledge graph constraints into policy updates, applies penalty gradients to actions that violate knowledge graph exclusion relationships, and dynamically blocks non-compliant actions based on the current fault type. Based on the fault mode library and historical fault protection measures, fault records are extracted for pre-training of deep reinforcement learning agents, and an ε-greedy strategy is used for online fine-tuning.

[0075] S312 uses the trained deep reinforcement learning agent basic Q-value network and knowledge Q-value network to calculate the Q-value of each action, adjusts the temperature coefficient according to the working conditions through the dynamic Boltzmann strategy, and selects the final action as the fault protection measure.

[0076] It should be noted that the real-time electrical signals are normalized and segmented into 0.1-second time windows. Within each time window, 12-dimensional time-domain features (including peak, mean, and crest factor) and 8-dimensional frequency-domain features (energy ratio of the main harmonic components) are extracted to form a 20-dimensional electrical feature vector. The physical signal is filtered through a sliding average filter, and trend characteristics (slope, curvature, and fluctuation intensity) are calculated on a 1-minute time scale to generate a 6-dimensional physical state vector. The electrical and physical feature vectors are concatenated to form a 26-dimensional original state vector, which is then reduced to a compact 15-dimensional representation using an autoencoder to achieve multimodal signal feature encoding. Combined with the operating condition labels output by K-means clustering, a 3-dimensional operating condition identifier is appended using one-hot encoding. A historical device state memory is introduced, and an LSTM network is used to extract 8-dimensional time series features of the state changes at the most recent preset time step, resulting in a final state space of 26 dimensions.

[0077] In constructing the fault-action knowledge graph, trigger relationships connect faults and actions, while exclusion relationships identify conflicting actions. The optimal basic action set includes six discrete actions: power reduction, switching to backup power, adjusting cooling flow, triggering pre-charge protection, complete shutdown, and maintaining the current state. A deep reinforcement learning agent based on an improved PPO algorithm establishes a three-level fault severity evaluation system (minor 1, moderate 3, and severe 5) to quantify fault losses, with duration accuracy down to the millisecond level. For example, a klystron spark failure (severity 5) lasting 20ms would result in a contribution loss of 5 × 0.02 = 0.1. In action cost evaluation, power loss is calculated based on the actual percentage reduction, and equipment wear is discounted using a preset lifespan impact coefficient. The knowledge reward compares the current state vector with 20 typical fault modes in the knowledge graph (cosine similarity). The highest match is taken as the baseline value. When the system takes the protection action recommended by the knowledge graph, an additional coefficient of 0.5 is awarded. During the exploration reward activation process, a circular buffer containing 5,000 historical states is maintained. A new state is considered new if the Mahalanobis distance between the current state and the most recent 1,000 records is greater than 3σ. A reward of 0.2 is given for the first discovery of a new state, and this value decays exponentially with repeated occurrences. A rule verification module is added to the policy network output layer to apply a penalty gradient to actions that violate the knowledge graph's exclusionary relationships. For example, when simultaneously outputting power reduction and boost operation, a 10% penalty is added to the gradient backpropagation. Non-compliant actions are dynamically blocked based on the current fault type. For example, if a cooling failure fault is detected, the boost operation option is disabled.

[0078] During the pre-training of a deep reinforcement learning agent, fault records are extracted. Each record contains a state vector, a human-taken action, and a final result. Oversampling is performed to balance the minority class faults. A three-layer fully connected teacher network is used, and the student network (i.e., the policy network) is learned through distillation using a KL divergence loss combined with a knowledge graph compliance loss. During the online fine-tuning phase, the ε-greedy policy schedules an initial exploration probability of ε=0.5, which is linearly decayed by 0.1 after every 500 faults, ultimately maintaining a base exploration rate of 0.1. A soft update strategy is employed, updating the target network by 0.1% of the online network parameters every 1000 steps. Temperature coefficient control in the dynamic Boltzmann policy ensures that the action with the highest Q value is directly selected in the event of a severe fault. The Gumbel-Softmax technique is used to achieve differentiable sampling, ensuring efficient gradient backpropagation while maintaining randomness.

[0079] In a preferred embodiment, state recognition detected excessive TEO energy and a 25% drop in vibration spectrum entropy. Knowledge retrieval matched the electrode microdischarge fault model, generating and executing actions to reduce power by 20% and increase flow by 15%. This reduced the ignition frequency to a safe range within 200ms. Through detailed characterization of the state space and the rule-based constraints of the knowledge graph, deep reinforcement learning enables rapid optimization and decision-making while ensuring safety.

[0080] It should be noted that the digital twin and parameter optimization module obtains the electromagnetic field distribution within the S-, C-, and Ku-band klystrons and the three-dimensional temperature field distribution of the multi-band high-voltage power supply. The electromagnetic field distribution is based on the finite element method to construct a klystron electromagnetic model, solve the three-dimensional Maxwell equations, and accurately simulate the S-, C-, and Ku-band electromagnetic field distribution. Computational fluid dynamics is used to establish a three-dimensional temperature field model, simulating the heat conduction and convection heat transfer in key parts such as the tube body and collector to construct the three-dimensional temperature field distribution. Based on the electromagnetic field distribution and the three-dimensional temperature field distribution, physical constraints are constructed. Historical multimodal signals are used as input to the physical information neural network. Physical constraints are applied to the hidden layer of the physical information neural network, and the state prediction value of the multimodal signal is obtained through the output layer. The state prediction value of the multimodal signal is verified using test data. The verification is verified by obtaining a pre-trained digital twin model of the power supply unit and performing online incremental training using the latest collected multimodal signals to fine-tune the digital twin model.

[0081] When a multi-band high-voltage power supply system is in a faulty state, after temporary protection measures are implemented, the current fault characteristics are injected into the twin and parameter optimization is performed. This ensures that the output power is maximized while preventing further exacerbation of the equipment fault, extending equipment life while reducing cooling energy consumption. Multimodal signals are collected in real time and imported into the power supply unit's digital twin model to load the current state of the multi-band high-voltage power supply. A Gray Wolf optimizer is then constructed to tune the electrical operating point parameters and thermal management parameters. The electrical operating point parameters include the cathode high voltage setpoint, pulse modulator parameters, and titanium pump power supply operating current. Thermal management parameters include the coolant target flow rate, refrigeration unit power setting, and heater start / stop thresholds. An objective function is constructed with the goals of maximizing microwave power, maximizing klystron life, and minimizing cooling energy consumption. The objective function is associated with a fitness function to calculate the fitness of the gray wolf. The fitness of the gray wolf is calculated using the fitness function, and the top three optimal solutions are selected as α wolf, β wolf, and δ wolf according to the fitness ranking. The current global optimal solution is recorded, and the positions of α wolf, β wolf, δ wolf, and other gray wolves are updated in the iterative calculation using the steps of encircling, hunting, and attacking prey. The Lévy flight strategy is adopted for the α wolf, and Gaussian perturbations are performed on the β wolf and δ wolf to generate a new gray wolf population. An inertia weight that changes with the number of iterations is introduced. The fitness of the gray wolves in the new gray wolf population is calculated. A mutation operator that changes with the number of iterations is introduced to mutate the α wolf with the highest fitness. The fitness of the α wolf before and after the mutation is compared, and an elite retention strategy is used to select the optimal solution for the next generation and guide the search direction. When the number of iterations reaches the maximum number of iterations or the iteration termination condition is met, the position of the α wolf with the highest fitness in the last iteration is output to generate parameter update suggestions. These update suggestions are simulated and verified using a digital twin. The simulation results are compared with the expected performance improvement. The update suggestions that meet the requirements are output, executed by the PLC, and the optimization results are monitored. By combining the high-precision simulation of the digital twin with the improved optimization algorithm, the collaborative optimization of multiple objective parameters is achieved.

[0082] Figure 4 A flow chart of a normalized control and protection method for a multi-band high-voltage power supply is shown.

[0083] This embodiment provides a normalized control and protection method for a multi-band high-voltage power supply, which is applied to a high-voltage power supply control system and includes the following steps:

[0084] S402, collecting multimodal signals of the S-band high-voltage power supply system, the C-band high-voltage power supply system, the Ku-band high-voltage power supply system, and the cooling system, preprocessing the multimodal signals, and extracting signal features;

[0085] S404, constructing a fault diagnosis model by combining a stacked denoising autoencoder network with an ensemble learning AdaBoost framework, importing the signal features into the fault diagnosis model, and outputting a fault type and confidence level;

[0086] S406, when a fault occurs, cutting off the high-voltage output of the corresponding power supply unit and reporting to a remote terminal via an isolation interface, training a deep reinforcement learning agent, and using the deep reinforcement learning agent to make intelligent decisions on protection actions based on the fault type and fault severity;

[0087] S408, build a digital twin model of the power supply unit. When the multi-band high-voltage power supply system is in a fault state, the digital twin model is used to assist in adjusting the electrical operating point parameters and thermal management parameters. The improved Gray Wolf optimizer is used to perform multi-task collaborative optimization. The optimized system parameters are obtained and sent to the PLC for execution to ensure the stability of the band output.

[0088] This embodiment provides a computer-readable storage medium, which includes a normalized control and protection method program for a multi-band high-voltage power supply. When the normalized control and protection method program for a multi-band high-voltage power supply is executed by a processor, the steps of a normalized control and protection method for a multi-band high-voltage power supply are implemented.

[0089] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as: multiple blocks or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. In addition, the functional blocks in the various embodiments of the present invention can all be integrated into one processing block, or each block can be a separate block, or two or more blocks can be integrated into one block; the above-mentioned integrated blocks can be implemented in the form of hardware or in the form of hardware plus software functional blocks.

[0090] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0091] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A multi-band high-voltage power supply control system, characterized in that: include: Multi-source signal acquisition and preprocessing module, intelligent fault diagnosis module, protection decision-making and execution module, digital twin and parameter optimization module; The multi-source signal acquisition and preprocessing module is responsible for collecting multimodal signals of the S-band high-voltage power supply system, the C-band high-voltage power supply system, the Ku-band high-voltage power supply system and the cooling system, preprocessing the multimodal signals, and extracting signal features; The intelligent fault diagnosis module combines the stacked denoising autoencoder network with the ensemble learning AdaBoost framework to build a fault diagnosis model, imports the signal features into the fault diagnosis model, and outputs the fault type and confidence level; The protection decision and execution module trains a deep reinforcement learning agent and uses the deep reinforcement learning agent to make intelligent decisions on protection actions based on fault type and fault severity; The digital twin and parameter optimization module constructs a digital twin model of the power supply unit. When the multi-band high-voltage power supply system is in a fault state, the digital twin model assists in adjusting the electrical operating point parameters and thermal management parameters. The improved Gray Wolf optimizer is used for multi-task collaborative optimization. The optimized system parameters are obtained and sent to the PLC for execution, and real-time monitoring and feedback are provided. In the intelligent fault diagnosis module, the fault diagnosis model is enhanced, specifically: After a fault alarm is triggered, the TEO energy characteristics, reflected wave integral characteristics, and phase mutation detection characteristics corresponding to the electrical data are extracted from the S-band high-voltage power supply system, the C-band high-voltage power supply system, and the Ku-band high-voltage power supply system as electrical impact characteristics. The cavity temperature rise, vibration spectrum entropy, and flow pulsation corresponding to the physical data are extracted from the C-band high-voltage power supply system, Ku-band high-voltage power supply system, and cooling system as physical anomaly characteristics; According to the electrical impact characteristics and physical abnormality characteristics, fault feature fingerprints of different bands are constructed, and the dynamic time warping algorithm is used to match the fault feature fingerprints with the fingerprint templates in the fault mode library. The preliminary screening fault type is obtained through the matching results, and the classification enhancement is performed in the fault diagnosis model based on the preliminary screening fault type.

2. The multi-band high-voltage power supply control system according to claim 1, characterized in that: The multi-source signal acquisition and preprocessing module is specifically: A cross-band sensor network is constructed using pre-deployed sensors in the S-band high-voltage power supply system, the C-band high-voltage power supply system, the Ku-band high-voltage power supply system, and the cooling system to synchronously collect multimodal signals, perform data standardization on the multimodal signals, and add timestamps and device IDs; Performing digital filtering and time-domain alignment on the multimodal signal to obtain a preprocessed multimodal signal, performing time-frequency domain feature extraction on the preprocessed multimodal signal to construct a feature matrix, obtaining clusters through K-means clustering, and identifying the operating conditions of the multi-band high-voltage power supply system; The signal segments corresponding to the working conditions are extracted using a feature extraction layer that combines one-dimensional convolution and dilated convolution for local feature extraction. A channel attention mechanism is introduced to enhance the key frequency band features. The multimodal signals of the preset time window are captured using a BiLSTM network to capture the temporal dependencies. The self-attention layer is used to calculate the cross-band feature correlation and output the global features. The local features are aligned with the global features through time axis interpolation, and a weighted splicing fusion strategy is used to perform feature fusion to obtain signal features.

3. The multi-band high-voltage power supply control system according to claim 1, characterized in that: The intelligent fault diagnosis module is specifically: A weak classifier is constructed based on a stacked denoising autoencoder network. A layer-by-layer greedy training strategy is used to pre-train the weak classifier. Gaussian noise is added to each hidden layer during the pre-training of the weak classifier for anti-interference training. After the pre-training is completed, a Softmax classifier is added to the top layer and the weak classifier is fine-tuned using labeled data. The pre-trained weak classifiers are integrated using the AdaBoost framework. In the first round of iteration, all training samples are given the same weight. During the iteration process, the sample weights are dynamically adjusted according to the classification error of the weak classifiers, and weak classifiers that do not meet the preset classification accuracy are eliminated. When the preset number of iterations is reached, the weighted voting results of each weak classifier are used to form the final prediction. The test samples are used for verification. After verification, a strong classifier is generated to build a fault diagnosis model. The fault diagnosis model is used for online monitoring. The acquired signal features are input into all weak classifiers in parallel. Each weak classifier outputs a fault probability prediction value. A comprehensive score is obtained by weighted summation. The threshold is dynamically adjusted according to the current working conditions. When the comprehensive score exceeds the dynamic threshold, a fault alarm is triggered, and the fault type and confidence level are output.

4. The multi-band high-voltage power supply control system according to claim 1, characterized in that: The protection decision and execution module is specifically: The multimodal signal is feature-encoded to generate an original state vector, and the original state vector is embedded in the working condition context by combining the working condition label output by K-means clustering and the historical state memory to construct a state space; Extract fault type nodes and protection measure nodes based on the fault mode library and historical fault protection measures, build edge connection relationships between nodes according to trigger relationships and exclusion relationships based on constraint conditions, generate a fault-measure knowledge graph, generate a basic action set of protection measures based on the fault-measure knowledge graph, and build an action space; A deep reinforcement learning agent is constructed based on the state space and actions using an improved PPO algorithm. A pre-established three-level fault severity evaluation system is used to quantify fault losses. Action costs are evaluated using power loss and equipment wear. A basic reward function is constructed based on these fault losses and action costs. Introducing knowledge rewards, constructing a current state vector, performing similarity matching between the current state vector and the fault mode in the fault-measure knowledge graph, constructing a knowledge reward function based on the similarity matching, introducing exploration rewards, calculating the Mahalanobis distance between the current state vector and the historical state vector in the historical state ring buffer, and constructing a knowledge reward function based on the Mahalanobis distance; Knowledge graph constraints are introduced into policy updates, and penalty gradients are applied to actions that violate knowledge graph exclusion relationships. Non-compliant actions are dynamically blocked based on the current fault type. Fault records are extracted based on the fault mode library and historical fault protection measures to pre-train a deep reinforcement learning agent, and an ε-greedy strategy is used for online fine-tuning. The trained deep reinforcement learning agent basic Q-value network and knowledge Q-value network are used to calculate the Q-value of each action. The temperature coefficient is adjusted according to the working conditions through the dynamic Boltzmann strategy, and the final action is selected as the fault protection measure.

5. The multi-band high-voltage power supply control system according to claim 1, characterized in that: The protection decision-making and execution module and the digital twin and parameter optimization module are specifically: Obtaining the electromagnetic field distribution within the S-, C-, and Ku-band klystrons and the three-dimensional temperature field distribution of the multi-band high-voltage power supply, constructing physical constraints based on the electromagnetic field distribution and the three-dimensional temperature field distribution, using historical multimodal signals as input to a physical information neural network, applying physical constraints in the hidden layer of the physical information neural network, and obtaining a state prediction value of the multimodal signal through the output layer; Verify the state predictions of the multimodal signals using test data by obtaining a pre-trained digital twin model of the power supply unit and fine-tuning the digital twin model through online incremental training using the latest acquired multimodal signals; Real-time multimodal signals are collected and imported into the digital twin model of the power supply unit to load the current state of the multi-band high-voltage power supply. A Gray Wolf optimizer is constructed to perform parameter tuning. An objective function is constructed with the goals of maximizing microwave power, maximizing klystron life, and minimizing cooling energy consumption. This objective function is then associated with the fitness function to calculate the fitness of the Gray Wolf. An iterative optimization is performed based on the fitness of the gray wolf. When the number of iterations reaches the maximum number of iterations or the iteration termination condition is met, the position of the α wolf with the highest fitness in the last iteration is output to generate a parameter update suggestion. The update suggestion is simulated and verified using a digital twin, and the simulation results are compared with the expected performance improvement to output an update suggestion that meets the requirements.

6. The multi-band high-voltage power supply control system according to claim 5, characterized in that: The Gray Wolf Optimizer is specifically: The fitness of the gray wolf is calculated using a fitness function. The top three optimal solutions are selected as α wolf, β wolf, and δ wolf according to the fitness ranking, and the current global optimal solution is recorded. In the iterative calculation, the positions of α wolf, β wolf, δ wolf and other gray wolves are updated using the steps of surrounding, hunting and attacking prey. The Lévy flight strategy is used for α wolf, and Gaussian perturbations are performed on β wolf and δ wolf to generate a new gray wolf population. An inertia weight that changes with the number of iterations is introduced. The fitness of the gray wolves in the new gray wolf population is calculated, and a mutation operator that changes with the number of iterations is introduced to mutate the α wolf with the highest fitness. The fitness of the α wolf before and after the mutation is compared, and the elite retention strategy is used to select the better solution to enter the next generation and guide the search direction. When the number of iterations reaches the maximum number of iterations or the iteration termination condition is met, the position of the α wolf with the highest fitness in the last iteration is output to generate parameter update suggestions.

7. A normalized control and protection method for a multi-band high-voltage power supply, applied to the high-voltage power supply control system according to any one of claims 1 to 6, characterized in that: The following steps are involved: Collecting multimodal signals of an S-band high-voltage power supply system, a C-band high-voltage power supply system, a Ku-band high-voltage power supply system, and a cooling system, preprocessing the multimodal signals, and extracting signal features; A fault diagnosis model is constructed by combining a stacked denoising autoencoder network with an ensemble learning AdaBoost framework, the signal features are imported into the fault diagnosis model, and the fault type and confidence level are output; When a fault occurs, the high-voltage output of the corresponding power supply unit is cut off and reported to the remote terminal through the isolation interface. A deep reinforcement learning agent is trained and used to make intelligent decisions on protection actions based on the fault type and fault severity. A digital twin model of the power supply unit is constructed. When the multi-band high-voltage power supply system is in a fault state, the digital twin model is used to assist in adjusting the electrical operating point parameters and thermal management parameters. An improved Gray Wolf optimizer is used for multi-task collaborative optimization. The optimized system parameters are obtained and sent to the PLC for execution to ensure the stability of the band output.

Images