Real-time monitoring system and method for a plasma generator

Abstract

The application relates to the technical field of ion generators, and discloses a real-time monitoring system and method for a plasma generator, which collects and pre-processes historical operation data, environmental conditions and configuration parameters of the plasma generator, and constructs an operation index prediction model. By real-time acquisition of current data and input into the model, future operation indexes are predicted, compared with real-time operation indexes, and deviation points are marked. The state monitoring process is modeled as a Markov decision process, a state space, an action space, a reward function and an output strategy are defined, and actions are optimized and adjusted through a reinforcement learning algorithm. Real-time monitoring is carried out based on the trained state monitoring decision model, and optimal adjustment actions are executed automatically or after confirmation by an operator. The application realizes intelligent state monitoring and predictive maintenance of the plasma generator, effectively improves the equipment performance and system stability, and has significant industrial application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ion generator, in particular to a real-time monitoring system and method of plasma generator. BACKGROUND

[0002] In industrial production, scientific research experiments and many high-tech fields, plasma generators as an important equipment, its performance and stability have a crucial influence on the running efficiency and result accuracy of the whole system. However, the plasma generator is affected by many factors during operation, such as changes in environmental conditions, aging of the equipment itself and adjustment of configuration parameters, and its running index often shows a complex and variable trend.

[0003] With the advent of the Industry 4.0 era and the development of intelligent technology, real-time monitoring and predictive maintenance of plasma generators have become an urgent demand in the industry. Although the existing monitoring technology has realized real-time monitoring of the running state of the equipment to some extent, it only stays at the level of data collection and simple alarm, lacks in-depth data analysis and prediction ability, and cannot accurately predict the future running state of the equipment, nor can it actively adjust the equipment parameters or take maintenance measures in advance.

[0004] Due to the complex working principle of plasma generator, its performance is affected by many factors, how to comprehensively consider the environmental conditions, equipment configuration and operation history, build a model that can accurately predict the change trend of the running index of the equipment, and on this basis realize intelligent state monitoring and decision-making, is the problem to be solved in the current technology. SUMMARY

[0005] The purpose of the present application is to provide a real-time monitoring system and method of plasma generator to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides the following technical solution: a real-time monitoring method of plasma generator, the method comprising:

[0007] Collecting running index time series data of the plasma generator in historical operation, as well as environmental conditions and configuration parameters of the plasma generator at runtime; and pre-processing the collected data to form a feature data set;

[0008] Building a running index prediction model for the plasma generator, which is based on the feature data set and combines environmental conditions and plasma generator configuration parameters to predict the running index change trend of the plasma generator in the future period of time;

[0009] Real-time acquisition of current configuration parameter data and environmental condition data of the plasma generator, and input of the acquired data into the running index prediction model to receive the predicted running index output by the model;

[0010] Acquisition of real-time running index data of the plasma generator, drawing of the acquired real-time running index data and the predicted running index data output by the model into curves varying with time, time stamp alignment of the two curves, and comparison of the two curves in the same coordinate system; further setting of a performance deviation threshold, and judging whether each corresponding point on the two curves has a deviation exceeding the threshold, and marking the point as a deviation point if there is one;

[0011] Construction of a state monitoring decision model, modeling of the state monitoring process of the plasma generator as a Markov decision process, and specific definitions including: ① state space: defined as a combination of real-time running index data, configuration parameters and environmental condition data of the plasma generator; ② action space: defined as a set of executable monitoring or adjustment operations; ③ reward function: designed according to the stable operation and performance optimization goals of the plasma generator, including a reward mechanism for state stability, performance maintenance and energy efficiency improvement, and combining the monitored performance deviation data to give an additional reward for operations that reduce deviation and an additional penalty for operations that increase deviation; ④ output strategy: defined as a method for selecting the optimal action under the current state, and continuously optimizing the generated adjustment action through reinforcement learning algorithm;

[0012] Real-time monitoring based on the trained state monitoring decision model, real-time collection of configuration parameters and environmental condition data of the plasma generator as model input, selection of the optimal adjustment action according to the current state and strategy, and automatic execution of the decision result or transmission of the decision result to the operator for confirmation and then execution through the monitoring system.

[0013] Preferably, the running index includes plasma density, temperature, voltage, current, power, gas flow and cooling system state; the environmental condition includes environmental temperature, humidity, air pressure and electromagnetic interference condition; and the configuration parameter includes the type, size, working frequency and electrode material of the plasma generator.

[0014] Preferably, the set of executable monitoring or adjustment operations includes: adjusting input power or working frequency, changing gas type or flow, starting or stopping the cooling system, adjusting electrode spacing or replacing electrodes, cleaning or replacing parts, adjusting control system parameters, adjusting data acquisition frequency or monitoring accuracy.

[0015] Preferably, the reward function R consists of multiple sub-reward items, and the total reward value is obtained by weighted summation, specifically using the formula: R = w1*R_stab + w2*R_perf + w3*R_eff + w4*R_maint + w5*R_dev; where R_stab represents the reward item for state stability, calculated based on the fluctuation of the plasma generator's operating state; R_perf represents the reward item for performance maintenance, calculated based on the maintenance of the plasma generator's key performance indicators; and R_eff represents the reward item for energy efficiency improvement. The reward is calculated based on the degree of improvement in energy consumption efficiency; R_maint represents the reward for preventive maintenance, calculated based on the execution and effectiveness of preventive maintenance operations; R_dev represents the reward or penalty for reducing performance deviation, based on the monitored performance deviation data, giving positive rewards for operations that reduce deviation and penalties for operations that increase deviation, and its value is set as a function that is inversely or directly proportional to the change in deviation; w1, w2, w3, w4 and w5 are the weight coefficients of each sub-reward item, and w1+w2+w3+w4+w5=1.

[0016] Preferably, the performance indicator prediction model is constructed using the Gated Recurrent Unit (GRU) algorithm, and its network structure specifically includes:

[0017] Input layer: Used to receive feature datasets as input, which include time-series performance index data, environmental parameter data, system configuration data, and time-related features;

[0018] Hidden layer: Contains at least one GRU hidden layer. Each GRU hidden layer contains several GRU units. Each GRU unit has an update gate and a reset gate to control the flow of information and state updates.

[0019] Output layer: Receives the output of the hidden layer and maps the output of the GRU network to the predicted performance metric value through a fully connected layer or a regression layer.

[0020] Preferably, the steps for training and running the indicator prediction model include:

[0021] S1: Initialize the weights and biases of the GRU network using small random numbers;

[0022] S2: Select a loss function to measure the difference between the model's predicted values ​​and the actual values;

[0023] S3: Using the backpropagation algorithm and gradient descent optimization algorithm, the gradient is calculated based on the loss function, and the weights and biases of the GRU network are updated. The network is then trained iteratively until the loss function converges.

[0024] S4: Save the trained GRU network model as a predictive model for operational indicators in the real-time monitoring method of the plasma generator.

[0025] Preferably, the state monitoring and decision model is trained using the Deep Q-Network (DQN) algorithm.

[0026] Preferably, the steps for training the state monitoring decision model include:

[0027] A1. Initialize the Q-network: Construct a deep neural network to evaluate the value of state-action pairs;

[0028] A2. Initialize the experience playback buffer D: Set the buffer size N to store the state transition experience generated during the plasma generator monitoring process, including state s, action a, reward r, next state s' and whether the termination flag is done;

[0029] A3. For each training round:

[0030] A301: Initialize the state s at the start of the round;

[0031] A302: Before the round ends, the following steps are executed in a loop: i) The Q network selects action a based on the current state s, and uses an ε-greedy strategy to balance exploration and exploitation; ii) Execute action a, observe the reward r and the next state s', and whether to terminate (done); iii) Store the experience (s,a,r,s',done) in the experience replay buffer D; iv) Randomly sample a batch of experience from the experience replay buffer D to update the network parameters; v) Use the sampled experience to update the parameters of the Q network to minimize the value function prediction error; vi) Update the current state s to the next state s'.

[0032] A303: When the round ends, reset state s to the new start state of the round;

[0033] A4: During training, periodically evaluate the performance of the state monitoring decision model, and adjust the weight coefficients of the reward function, the learning rate of the model, and the magnitude of exploration noise based on the evaluation results;

[0034] A5: Repeat steps A3 and A4 until a satisfactory performance level is achieved.

[0035] Preferably, in step A4, the mean squared error (MSE) is used as the performance evaluation metric for the current round of training. The specific implementation steps include:

[0036] A401: Based on the sampled empirical tuples (s,a,r,s'), calculate the target Q value for each empirical tuple using the target Q network;

[0037] A402: Calculate the predicted Q value of the main Q network for each state-action pair (s, a) in the experience tuple;

[0038] A403: Calculate the mean square error (MSE) between the predicted Q value and the target Q value as the loss function value of the current round of iteration, the formula for calculating MSE is:

[0039]

[0040] where N is the number of sampled experience tuples, Q target is the target Q value, and Q predicted is the predicted Q value of the main Q network.

[0041] A404: According to the calculated MSE value, update the parameters of the main Q network using the gradient descent algorithm to minimize the MSE value, thereby optimizing the performance of the main Q network;

[0042] A405: Repeat steps A401 to A404 until the end condition of the current training round is reached, and then enter the next training round.

[0043] Preferably, a real-time monitoring system for a plasma generator, the system comprising:

[0044] a data collection module for collecting time series data of operating indicators of the plasma generator in historical operation, environmental conditions at the time of operation and configuration parameters of the plasma generator, and pre-processing the collected data to form a feature data set;

[0045] a prediction model construction module for constructing an operating indicator prediction model based on the feature data set, in combination with environmental conditions and plasma generator configuration parameters, to predict the operating indicator trend of the plasma generator in the future period of time;

[0046] a prediction data acquisition module for acquiring current configuration parameter data and environmental condition data of the plasma generator in real time, and inputting the acquired data into the operating indicator prediction module to receive and output the model predicted operating indicator;

[0047] a real-time data acquisition module for acquiring real-time operating indicator data of the plasma generator;

[0048] a deviation detection module for plotting the real-time operating indicator data and the predicted operating indicator data into curves changing with time, aligning the time stamps and placing them in the same coordinate system for comparison, further setting a performance deviation threshold, determining whether each corresponding point on the two curves has a deviation exceeding the threshold, and marking the deviation points;

[0049] A state monitoring decision module is used to build a state monitoring decision model, model the state monitoring process of the plasma generator as a Markov decision process, include defining a state space, an action space, a reward function and an output policy, and continuously optimize the generated adjustment actions through a reinforcement learning algorithm;

[0050] A decision execution module is used to perform real-time monitoring based on the trained state monitoring decision model, collect configuration parameters and environmental condition data of the plasma generator in real time as model input, select the optimal adjustment action according to the current state and policy, and automatically execute the decision result or transmit it to the operator for confirmation before execution through the monitoring system.

[0051] Compared with the prior art, the beneficial effects of the present application are:

[0052] The present application collects and analyzes the running data of the plasma generator in real time, combines environmental conditions and configuration parameters, and builds a running index prediction model. This model can accurately predict the running index trend of the device in the future, so that the operator can understand the device state in advance and adjust the parameters or take maintenance measures in time, thereby significantly improving the running efficiency and stability of the plasma generator. Through real-time monitoring and predictive maintenance, the present application can timely discover and solve potential problems, avoid serious damage to the device due to failure, and effectively prolong the service life of the plasma generator. At the same time, based on the monitoring of performance deviation and the optimization adjustment of Markov decision process, the device can always be in the best working state, reducing wear and aging.

[0053] The present application applies reinforcement learning algorithm to the state monitoring decision model of the plasma generator, realizes intelligent state monitoring and decision making by continuously optimizing the generated adjustment actions. This intelligent monitoring method not only improves the accuracy and efficiency of monitoring, but also reduces the dependence on manual intervention and improves the intelligent level of the whole system. The reward function design in the state monitoring decision model fully considers the stable operation, performance maintenance and energy efficiency improvement goals of the plasma generator. By continuously optimizing the adjustment actions, the present application can ensure stable operation of the device while optimizing performance and improving energy efficiency, reducing energy consumption and improving production efficiency.

[0054] The real-time monitoring system and method of the present application have good maintainability and scalability. The system can automatically record the device running history to provide data support for subsequent maintenance and analysis. The system structure is clear and the module division is reasonable, which is convenient for subsequent upgrading and expansion to adapt to the monitoring needs of plasma generators of different types and specifications. BRIEF DESCRIPTION OF DRAWINGS

[0055] Fig. 1 A step diagram of the real-time monitoring method of the plasma generator described in the present application;

[0056] Fig. 2 Training flowchart for running index prediction model;

[0057] Fig. 3 Construction flowchart for state monitoring decision model. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0059] Please refer to Figs. 1-3 The present application provides a real-time monitoring method for a plasma generator, which comprises the following steps:

[0060] Collect running index time series data from the historical operation of the plasma generator, including but not limited to voltage, current, temperature and other key indicators. At the same time, record the environmental conditions (such as ambient temperature, humidity, etc.) and configuration parameters (such as working frequency, power setting, etc.) of the plasma generator during operation. The collected data is preprocessed through cleaning, denoising and normalization, etc. to form a feature data set, providing a basis for subsequent modeling.

[0061] Based on the feature data set, a running index prediction model is constructed using machine learning algorithms. The model considers the environmental conditions and plasma generator configuration parameters, and through training learns the complex relationship between them, so as to predict the running index trend of the plasma generator in the future.

[0062] Real-time acquisition of current configuration parameter data and environmental condition data of the plasma generator as model input, receiving the predicted running index output by the model. At the same time, acquire real-time running index data of the plasma generator for comparison with the prediction results.

[0063] Draw the real-time running index data and the predicted running index data into curves changing with time respectively, and align the time stamps. Compare the two curves in the same coordinate system, set a performance deviation threshold, automatically judge whether there is a deviation exceeding the threshold for each corresponding point, if there is, mark it as a deviation point as the basis for subsequent decision.

[0064] State monitoring decision model construction: model the state monitoring process of the plasma generator as a Markov decision process. Specific definitions include:

[0065] ①State space: composed of real-time running index data, configuration parameters and environmental condition data of the plasma generator.

[0066] ②Action space: contains a set of executable monitoring or adjustment operations, such as adjusting the working frequency, power, etc.

[0067] ③Reward function: designed according to the stable operation and performance optimization goals of the plasma generator, containing reward mechanisms for state stability, performance maintenance and energy efficiency improvement. At the same time, combined with the monitored performance deviation data, additional rewards are given for operations that reduce deviation, and additional penalties are given for operations that increase deviation.

[0068] ④Output strategy: uses reinforcement learning algorithm to continuously optimize the generated adjustment actions and selects the optimal action under the current state.

[0069] Based on the trained state monitoring decision model, real-time collection of configuration parameters and environmental condition data of the plasma generator is performed as model input. According to the current state and strategy, the optimal adjustment action is selected. The decision result can be automatically executed or transmitted to the operator for confirmation and execution through the monitoring system, ensuring the stable operation and performance optimization of the plasma generator.

[0070] The present application will be further described in conjunction with Examples 1 to 4:

[0071] Example 1:

[0072] In the selection of running indicators, the present application comprehensively considers the key performance parameters of the plasma generator, including but not limited to plasma density, temperature, voltage, current, power, gas flow and cooling system state. These indicators can comprehensively reflect the working state and performance level of the plasma generator. Plasma density and temperature are important parameters for describing the state of plasma, directly affecting the generation and maintenance of plasma. Voltage, current and power reflect the working efficiency and energy consumption of the plasma generator. Gas flow determines the generation rate and stability of plasma. The cooling system state is related to whether the plasma generator can run stably for a long time and avoid overheating damage. By monitoring these indicators in real time, the working abnormalities of the plasma generator can be found in time, providing basis for subsequent prediction and decision-making.

[0073] Environmental conditions also have a significant impact on the performance of the plasma generator, including environmental temperature, humidity, air pressure, and electromagnetic interference. Environmental temperature can affect the stability and efficiency of the plasma. Excessive humidity can cause electrode corrosion or increase gas humidity, affecting the quality of the plasma. Changes in air pressure can affect the degree of ionization of the gas and the density of the plasma. Electromagnetic interference may interfere with the control system of the plasma generator, affecting its normal operation. By including these environmental conditions in the monitoring range, the running state of the plasma generator can be more accurately predicted, and corresponding measures can be taken to respond to environmental changes.

[0074] The configuration parameters of the plasma generator determine its working characteristics and performance range. The configuration parameters include the type, size, working frequency and electrode material of the plasma generator. The type determines the generation method and application field of the plasma. The size affects the volume and shape of the plasma, which in turn affects its application effect. The working frequency determines the generation rate and energy distribution of the plasma. The electrode material is related to the durability of the electrode and the quality of the plasma. By monitoring and adjusting these configuration parameters in real time, the performance of the plasma generator can be optimized to meet different application requirements.

[0075] In order to realize real-time monitoring and predictive maintenance of the plasma generator, the present invention defines a series of executable monitoring or adjustment operation sets, including but not limited to:

[0076] Adjust the input power or working frequency: By changing the input power or working frequency, the generation rate and energy distribution of the plasma can be adjusted to optimize the performance of the plasma generator.

[0077] Change the gas type or flow rate: According to the application requirements, different gas types and flow rates can be selected to obtain the required plasma characteristics.

[0078] Start or stop the cooling system: According to the temperature condition of the plasma generator, the cooling system is started or stopped in time to ensure its long-term stable operation.

[0079] Adjust the electrode spacing or replace the electrode: The electrode spacing and electrode material selection will affect the quality and stability of the plasma. By adjusting the electrode spacing or replacing the electrode, the performance of the plasma generator can be optimized.

[0080] Clean or replace parts: Regular cleaning and replacement of parts of the plasma generator can maintain its good working condition and prolong its service life.

[0081] Adjust the control system parameters: According to the actual requirements, adjust the parameters of the control system to achieve more accurate control of the plasma generator.

[0082] Adjusting data collection frequency or monitoring precision: According to the application scenario and monitoring requirements, the frequency of data collection and the precision of monitoring can be adjusted to meet different monitoring requirements.

[0083] Embodiment 2:

[0084] This embodiment is used to describe the composition of the reward function in the real-time monitoring method of the plasma generator and the training process of the state monitoring decision model, which specifically includes:

[0085] Define the reward function: The reward function R is the core part of the state monitoring decision model, which is composed of multiple sub-reward items, and the total reward value is obtained by weighted summation. The specific formula is:

[0086] R = w1*R_stab + w2*R_perf + w3*R_eff + w4*R_maint + w5*R_dev Where: R_stab represents the reward item of state stability, which is calculated according to the fluctuation degree of the running state of the plasma generator. If the running state is stable and the fluctuation is small, a positive reward is given; otherwise, if the fluctuation is large, a negative reward or zero reward is given. R_perf represents the reward item of performance maintenance, which is calculated according to the maintenance of the key performance indicators of the plasma generator. If the performance indicators remain within the set range, a positive reward is given; if the performance indicators deviate from the set range, a negative reward is given. R_eff represents the reward item of energy efficiency improvement, which is calculated according to the improvement degree of energy consumption efficiency. If the energy efficiency is improved, a positive reward is given; if the energy efficiency is reduced, a negative reward is given. R_maint represents the reward item of preventive maintenance, which is calculated according to the execution and effect of preventive maintenance operations. If an effective preventive maintenance operation is performed, such as replacing aged components in advance, a positive reward is given. R_dev represents the reward item or penalty item of performance deviation reduction, which is calculated according to the monitored performance deviation data. A positive reward is given for operations that reduce deviation, and the reward value is proportional to the amount of deviation reduction; a penalty is given for operations that increase deviation, and the penalty value is proportional to the amount of deviation increase.

[0087] The weight coefficients w1, w2, w3, w4 and w5 are used to balance the importance of each sub-reward item, and satisfy w1+w2+w3+w4+w5=1. These weight coefficients can be adjusted according to the actual application scenario and monitoring requirements.

[0088] In order to train the state monitoring decision model, the present application adopts the deep Q network (DQN) algorithm. The training process of the state monitoring decision model includes the following steps:

[0089] A1. Initialize Q network: Build a deep neural network to evaluate the value of state-action pairs. The input of the network is the current state of the plasma generator (including running indicators, environmental conditions and configuration parameters), and the output is the value estimation of each possible action.

[0090] A2. Initialize experience replay buffer D: Set the buffer size N to store state transition experiences generated during the plasma generator monitoring process. Each experience includes the current state s, the selected action a, the obtained reward r, the next state s', and the termination flag done.

[0091] A3. For each training episode:

[0092] A301: Initialize the starting state s of the episode.

[0093] A302: Loop the following steps until the episode terminates:

[0094] i) Q-network selects an action a from the current state s using an ε-greedy policy balancing exploration and exploitation. That is, with probability ε, a random action is chosen, and with probability 1-ε, the action with the highest value is chosen.

[0095] ii) Execute the action a and observe the obtained reward r, the next state s', and the termination flag done.

[0096] iii) Store the experience (s, a, r, s', done) in the experience replay buffer D.

[0097] iv) Randomly sample a batch of experiences from the experience replay buffer D for updating the network parameters. This breaks the temporal correlation between experiences, improving training efficiency.

[0098] v) Update the parameters of the Q-network using the sampled experiences to minimize the value function prediction error. Typically, mean squared error or cross-entropy loss functions are used for training.

[0099] vi) Update the current state s to the next state s'.

[0100] A303: When the episode terminates (e.g., reaching the maximum number of steps or the plasma generator fails), reset the state s to the starting state of the new episode.

[0101] A4. Regular evaluation and adjustment: During the training process, regularly evaluate the performance of the state monitoring decision model. Based on the evaluation results, adjust the weight coefficients of the reward function, the learning rate of the model, and the size of the exploration noise to optimize the model performance.

[0102] A5. Repeat training: Repeat steps A3 and A4 until a satisfactory performance level is reached. This usually means that the model can stably select the optimal action, keep the plasma generator in an efficient and stable operating state, and reduce performance deviations and failures.

[0103] The DQN algorithm combines the advantages of deep learning and reinforcement learning, and can handle high-dimensional state space and continuous action space problems.

[0104] Embodiment 3:

[0105] This embodiment is used to describe in detail the specific implementation steps of using mean square error (MSE) as a performance evaluation index in step A4:

[0106] Step A4: Performance evaluation and parameter update

[0107] During the training of the state monitoring decision model, in order to evaluate the performance of the current round of training and optimize the model parameters, the mean square error (MSE) is used as an evaluation index. The specific implementation steps are as follows:

[0108] A401: Calculate the target Q value: according to the sampled experience tuples (s, a, r, s'), use the target Q network to calculate the target Q value of each experience tuple.

[0109] A402: Calculate the predicted Q value: calculate the predicted Q value of the state-action pair (s, a) in each experience tuple by the main Q network.

[0110] A403: Calculate the mean square error (MSE): calculate the mean square error (MSE) between the predicted Q value and the target Q value as the loss function value of the current round of iteration. The formula for calculating MSE is:

[0111]

[0112] where N is the number of sampled experience tuples, Q target is the target Q value, and Q predicted is the predicted Q value of the main Q network. The MSE value reflects the degree of deviation between the predicted Q value of the main Q network and the target Q value, and is an important indicator for evaluating model performance.

[0113] A404: Parameter update: according to the calculated MSE value, use the gradient descent algorithm to update the parameters of the main Q network to minimize the MSE value. The gradient descent algorithm calculates the gradient of the loss function with respect to the network parameters and updates the parameters in the opposite direction of the gradient, thereby gradually reducing the prediction error and optimizing the model performance.

[0114] During the parameter update process, the learning rate α can be set to control the step size of parameter update. The choice of learning rate has an important impact on the training effect and convergence speed of the model. Too large a learning rate may cause the model to oscillate during training, while too small a learning rate may cause the model to converge too slowly or even fail to converge.

[0115] A405: Repeat training: Repeat steps A401 to A404 until the end condition of the current training round is reached. The end condition can be reaching a preset number of training rounds, the MSE value being below a certain threshold, or the model performance no longer improving significantly, etc. When the end condition is met, proceed to the next training round to continue optimizing the model performance.

[0116] Embodiment 4:

[0117] For the construction and training of the operation index prediction model, this embodiment is used to describe the specific implementation steps and structural details of constructing the prediction model using the gated recurrent unit (GRU) algorithm, which specifically includes:

[0118] ① The operation index prediction model is constructed using the GRU algorithm, and its network structure includes an input layer, a hidden layer, and an output layer:

[0119] Input layer: used to receive feature data sets as input. The feature data set contains time series performance indicator data (such as temperature, pressure, current, etc.), environmental parameter data (such as environmental temperature, humidity, etc.), system configuration data (such as device model, setting parameters, etc.), and time-related features (such as timestamp, date, etc.). After preprocessing, these data are input into the GRU network in sequence.

[0120] Hidden layer: contains at least one GRU hidden layer. Each GRU hidden layer is composed of several GRU units, each of which has an update gate and a reset gate inside. The update gate controls the influence of the previous state on the current state, and the reset gate controls the reset degree of the previous state to the current input information. These gating mechanisms enable the GRU network to capture long-term dependencies in time series data and effectively handle noise and irregularities in sequences.

[0121] Output layer: receives the output of the hidden layer and maps the output of the GRU network to the predicted performance indicator value through a fully connected layer or a regression layer. The output layer can be designed as single output or multiple outputs according to specific requirements to predict the future value of one or more performance indicators.

[0122] ② The steps for training the operation index prediction model include the following stages:

[0123] S1: Initialize network parameters

[0124] Initialize the weights and bias terms of the GRU network, usually using small random numbers to ensure sufficient exploration of the model at the beginning of training.

[0125] S2: Select loss function

[0126] Select an appropriate loss function to measure the difference between the model's predicted values and the true values. For regression problems, commonly used loss functions include Mean Squared Error (MSE), Root Mean Squared Error (RMSE), or Mean Absolute Error (MAE), etc. Choose the appropriate loss function according to the actual needs and data characteristics.

[0127] S3: Iterative training of the network

[0128] Use the backpropagation algorithm and gradient descent optimization algorithm to train the model. The specific steps are as follows:

[0129] Forward propagation: input the feature dataset into the GRU network and calculate the network's predicted output.

[0130] Calculate the loss: calculate the difference between the predicted value and the true value according to the loss function, and get the loss value.

[0131] Backpropagation: calculate the gradient according to the loss value, which is the partial derivative of the loss function with respect to the network parameters.

[0132] Parameter update: use the Adam optimization algorithm to update the weights and bias terms of the GRU network to reduce the loss value.

[0133] Repeat the above steps to iteratively train the network until the loss function converges or reaches the set number of training rounds.

[0134] S4: Save the model

[0135] Save the trained GRU network model as the performance indicator prediction model in the real-time monitoring method of the plasma generator. This model can be deployed in the monitoring system to receive feature data in real time and predict the future value of the performance indicator, providing a basis for monitoring decisions.

[0136] The present application also includes a real-time monitoring system for a plasma generator, comprising:

[0137] A data collection module for collecting time series data of performance indicators in historical operation of the plasma generator, environmental conditions during operation, and configuration parameters of the plasma generator, and preprocessing the collected data to form a feature dataset;

[0138] A prediction model construction module for constructing a performance indicator prediction model based on the feature dataset, combining environmental conditions and plasma generator configuration parameters, to predict the performance indicator trend of the plasma generator in the future period;

[0139] A prediction data acquisition module for acquiring current configuration parameter data and environmental condition data of the plasma generator in real time, and inputting the acquired data into the performance indicator prediction module to receive and output the model predicted performance indicator;

[0140] a real-time data acquisition module configured to acquire real-time operation index data of the plasma generator;

[0141] a deviation detection module configured to plot the real-time operation index data and the predicted operation index data into curves varying with time, align the time stamps, and place them in the same coordinate system for comparison, further set a performance deviation threshold, determine whether each corresponding point on the two curves has a deviation exceeding the threshold, and mark the deviation point;

[0142] a state monitoring decision module configured to build a state monitoring decision model, model the state monitoring process of the plasma generator as a Markov decision process, including defining state space, action space, reward function and output policy, and continuously optimize the generated adjustment action through reinforcement learning algorithm;

[0143] a decision execution module configured to perform real-time monitoring based on the trained state monitoring decision model, collect configuration parameters and environmental condition data of the plasma generator as model input, select the optimal adjustment action according to the current state and policy, and automatically execute the decision result or transmit it to the operator for confirmation before execution through the monitoring system.

[0144] The implementation of the system is described in the above embodiments, and will not be repeated here.

[0145] It should be noted that, in this document, the terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0146] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method of real-time monitoring of a plasma generator, characterized in that, The method comprises: Collecting running index time series data of the plasma generator in historical operation, and environmental conditions and configuration parameters of the plasma generator at the time of operation; and preprocessing the collected data to form a feature data set; Building a running index prediction model for the plasma generator, which is based on the feature data set, combined with the environmental conditions and the configuration parameters of the plasma generator, to predict the running index change trend of the plasma generator in a future period of time; Real-time acquisition of current configuration parameter data and environmental condition data of the plasma generator, and input of the acquired data into the running index prediction model to receive the predicted running index output by the model; Acquisition of real-time running index data of the plasma generator; drawing of the acquired real-time running index data and the predicted running index data output by the model into curves changing with time, time stamp alignment of the two curves, and comparison of the two curves in the same coordinate system; further setting of a performance deviation threshold, judgment of whether there is a deviation exceeding the threshold for each corresponding point on the two curves, and marking of the deviation point if there is one; Building a state monitoring decision model, modeling the state monitoring process of the plasma generator as a Markov decision process, and specific definitions including: ① state space: defined as a combination of real-time running index data, configuration parameters and environmental condition data of the plasma generator; ② action space: defined as a set of executable monitoring or adjustment operations; ③ reward function: designed according to the stable operation and performance optimization goals of the plasma generator, including a reward mechanism for state stability, performance maintenance and energy efficiency improvement, and giving an additional reward for an operation that reduces the deviation and an additional penalty for an operation that increases the deviation; ④ output strategy: defined as a method for selecting the optimal action under the current state, and continuously optimizing the generated adjustment action through reinforcement learning algorithm; Real-time monitoring based on the trained state monitoring decision model, real-time collection of configuration parameters and environmental condition data of the plasma generator as model input, selection of the optimal adjustment action according to the current state and strategy, and automatic execution of the decision result or transmission of the decision result to the operator for confirmation and then execution through the monitoring system.

2. The method of real-time monitoring of a plasma generator according to claim 1, wherein, The running index includes plasma density, temperature, voltage, current, power, gas flow and cooling system state; the environmental conditions include environmental temperature, humidity, air pressure and electromagnetic interference conditions; and the configuration parameters include the type, size, working frequency and electrode material of the plasma generator.

3. The method of claim 1, wherein the step of monitoring the plasma generator in real time comprises the step of: The set of executable monitoring or adjustment operations includes: adjustment of input power or working frequency, change of gas type or flow, start or stop of the cooling system, adjustment of electrode spacing or replacement of electrodes, cleaning or replacement of components, adjustment of control system parameters, adjustment of data acquisition frequency or monitoring accuracy.

4. The method of claim 1, wherein, The reward function R is composed of multiple sub-reward items, and the total reward value is obtained by weighted summation, and the specific formula is: R=w1*R_stab+w2*R_perf+w3*R_eff+w4*R_maint+w5*R_dev; wherein R_stab represents the reward item of state stability, which is calculated according to the fluctuation degree of the running state of the plasma generator; R_perf represents the reward item of performance maintenance, which is calculated according to the maintenance situation of the key performance indicators of the plasma generator; R_eff represents the reward item of energy efficiency improvement, which is calculated according to the improvement degree of energy consumption efficiency; R_maint represents the reward item of preventive maintenance, which is calculated according to the execution and effect of the preventive maintenance operation; R_dev represents the reward item or penalty item of performance deviation reduction, which is calculated according to the monitored performance deviation data, and positive reward is given to the operation of reducing deviation, and penalty is given to the operation of increasing deviation, and its value is set as a function inversely proportional or proportional to the deviation change; w1, w2, w3, w4 and w5 are weight coefficients of each sub-reward item, and w1+w2+w3+w4+w5=1.

5. The method of real-time monitoring of a plasma generator according to claim 4, wherein, The running index prediction model is constructed by using a gated recurrent unit (GRU) algorithm, and the network structure specifically includes: An input layer: used for receiving a feature data set as input, the feature data set including time series performance index data, environmental parameter data, system configuration data and time-related features; A hidden layer: containing at least one GRU hidden layer, each GRU hidden layer containing a plurality of GRU units, each GRU unit being provided with an update gate and a reset gate inside for controlling the flow of information and state updating; An output layer: receiving the output of the hidden layer and mapping the output of the GRU network to the predicted performance index value through a fully connected layer or a regression layer.

6. The method of real-time monitoring of a plasma generator according to claim 5, wherein, The steps of training the running index prediction model include: S1: initializing the weights and bias items of the GRU network, and initializing with small random numbers; S2: selecting a loss function for measuring the difference between the predicted value and the true value of the model; S3: using a back propagation algorithm and a gradient descent optimization algorithm to calculate the gradient according to the loss function, and updating the weights and bias items of the GRU network, and iteratively training the network until the loss function converges; S4: saving the trained GRU network model as the running index prediction model in the real-time monitoring method of the plasma generator.

7. The method of real-time monitoring of a plasma generator according to claim 4, wherein, A deep Q network (DQN) algorithm is used to train the state monitoring decision model.

8. The method of real-time monitoring of a plasma generator according to claim 7, wherein, The steps of training the state monitoring decision model include: A1, initializing the Q network: constructing a deep neural network for evaluating the value of state-action pairs; A2, initializing the experience replay buffer D: setting the buffer size N to store the state transition experience generated in the monitoring process of the plasma generator, including state s, action a, reward r, next state s' and termination flag done; A3, for each training round: A301: initialize the state s at the beginning of the round; A302: Before the round is terminated, the following steps are executed in a loop: i) the Q network selects an action a according to the current state s, balancing exploration and exploitation using an ε-greedy policy; ii) the action a is executed, and the resulting reward r and next state s' and termination flag done are observed; iii) the experience (s, a, r, s', done) is stored in the experience replay buffer D; iv) a batch of experiences is randomly sampled from the experience replay buffer D for updating the network parameters; v) using the sampled experiences, the parameters of the Q network are updated to minimize the value function prediction error; vi) the current state s is updated to the next state s'; A303: When the round is terminated, the state s is reset to the starting state of the new round; A4: During the training process, the performance of the state monitoring decision model is evaluated periodically, and the weight coefficients of the reward function, the learning rate of the model, and the size of the exploration noise are adjusted according to the evaluation results; A5: Steps A3 and A4 are repeated until a satisfactory performance level is reached.

9. The method of real-time monitoring of a plasma generator according to claim 8, wherein, In step A4, the mean square error MSE is used as the performance evaluation index of the current round of training, and the specific implementation steps include: A401: According to the sampled experience tuple (s, a, r, s'), the target Q network is used to calculate the target Q value of each experience tuple; A402: The prediction Q value of the state-action pair (s, a) in each experience tuple is calculated by the main Q network; A403: The mean square error MSE between the prediction Q value and the target Q value is calculated as the loss function value of the current iteration, and the calculation formula of MSE is: Wherein, N is the number of sampling experience tuples, Q target is the target Q value, Q predicted is the predicted Q value of the main Q network; A404: According to the calculated MSE value, the gradient descent algorithm is used to update the parameters of the main Q network to minimize the MSE value, thereby optimizing the performance of the main Q network; A405: Steps A401 to A404 are repeated until the end condition of the current training round is reached, and then the next training round is entered.

10. A real-time monitoring system for a plasma generator, characterized by, The system comprises: A data collection module for collecting running index time series data, environmental conditions and configuration parameters of the plasma generator during historical operation, and pre-processing the collected data to form a feature data set; A prediction model construction module for constructing a running index prediction model based on the feature data set, combining environmental conditions and plasma generator configuration parameters, to predict the running index trend of the plasma generator in the future period of time; A prediction data acquisition module for acquiring real-time configuration parameter data and environmental condition data of the plasma generator, and inputting the acquired data into the running index prediction module to receive and output the model predicted running index; A real-time data acquisition module for acquiring real-time running index data of the plasma generator; A deviation detection module for plotting the real-time running index data and the predicted running index data into curves changing over time, aligning the time stamps and placing them in the same coordinate system for comparison, further setting a performance deviation threshold, determining whether each corresponding point on the two curves has a deviation exceeding the threshold, and marking the deviation points; The state monitoring decision module is configured to build a state monitoring decision model, model the state monitoring process of the plasma generator as a Markov decision process, define a state space, an action space, a reward function and an output policy, and continuously optimize the generated adjustment action through a reinforcement learning algorithm. The decision execution module is configured to perform real-time monitoring based on the trained state monitoring decision model, collect configuration parameters and environmental condition data of the plasma generator as model input in real time, select an optimal adjustment action according to a current state and a policy, and automatically execute the decision result or transmit the decision result to an operator for confirmation before execution through a monitoring system.

Images