A DCS artificial intelligence module and its optimization method

Abstract

A DCS artificial intelligence module and an optimization method thereof are disclosed. The DCS artificial intelligence module adopts an incremental learning model. The model can output target variable parameters as an applied machine learning model and can perform machine learning training. During the training process, feature parameters and target variable parameters corresponding to a timestamp and a specified time thereafter are input, so that the incremental learning model has the ability to predict on the time axis; a weight coefficient is set for the target variable and an optimized variance algorithm is used to evaluate the model and calculate the Gaussian distribution deviation degree of the target variable parameter error, so as to provide an accurate basis for model fine-tuning; a DCS distributed control system is combined with artificial intelligence for collaborative control to improve the control optimization capability of the distributed control system; the optimized training / application integrated incremental learning model is integrated into the DCS to realize abnormality recognition and automatic startup solutions.

Description

Technical Field

[0001] The present invention relates to the general field of control or regulation systems, and in particular to a DCS artificial intelligence module and an optimization method thereof. Background Art

[0002] DCS (Distributed Control System) is a computer control system that centrally manages and decentralized controls the production process, also known as a distributed control system. Several commonly used programming languages ​​in DCS distributed control systems include Ladder Diagram (LD), Function Block Diagram (FBD), Sequential Function Chart (SFC), Structured Text (ST), Instruction List (IL), etc., which are mainly for distributed process control programming, and it is very difficult to realize artificial intelligence control.

[0003] DCS distributed control systems are widely used in modern automation control systems such as petroleum, chemical industry, electric power, and metallurgy. With the increasing complexity of industrial systems, traditional system control methods can no longer fully take into account the requirements of efficiency, quality, and energy saving. The combination of DCS distributed control systems and AI artificial intelligence to achieve collaborative control has great value and practical significance. First, through big data and machine learning algorithms, industrial systems can be optimized and controlled to reduce energy consumption and improve product quality. Secondly, intelligent manufacturing can be realized to improve production efficiency and flexibility. Finally, fault diagnosis and analysis, fault warning, and equipment reliability and maintenance efficiency can be improved, maintenance time and manpower requirements can be saved, and operation and maintenance costs can be reduced.

[0004] The combination of DCS distributed control system and AI artificial intelligence to achieve collaborative control has its own particularity. First of all, DCS distributed control system includes numerous control processes and corresponding different process processing techniques, as well as many distributed control subsystems. The control subsystem may also include lower-level subsystems, which in turn include various types of equipment containing embedded controllers and PLC programmable controllers. For example, in the power industry, the DCS distributed control system of coal-fired power plants includes three main boilers, steam turbines, and generator control systems, as well as auxiliary systems such as denitrification, dust removal, desulfurization, and water treatment and environmental protection. The dust removal system also includes lower-level systems such as high-voltage power supply, low-voltage vibration heating, and pneumatic ash conveying. The AI ​​artificial intelligence system needs to achieve optimal control of the DCS distributed control system and establish a model. There is a lack of model training data sets and test data sets. Since the DCS distributed control system belongs to a private dedicated network, there are many internal subsystems and equipment, and the operating conditions and process processing techniques are related to specific engineering projects. Therefore, it is necessary to obtain the operating data of specific engineering projects and establish a proprietary model training data set and test data set. This is a difficult problem. Secondly, it is also a problem how AI can be seamlessly embedded into the DCS distributed control system to achieve optimized control without changing the existing architecture of the DCS distributed control system. Thirdly, it is also a problem for AI to achieve optimized control across multiple layers of networks and various field buses in the DCS distributed control system, and to change specific parameters in the system, subsystem, and equipment. At the same time, the specific equipment of the DCS distributed control system has very high real-time requirements, requiring AI to have a short response time. The existing technologies in this field are:

[0005] China's invention patent CN105824234A industrial furnace intelligent RTU discloses an industrial furnace intelligent RTU, which is based on STM32F103ZET6 and integrates multiple universal I / O interfaces. It can be used with various gas and oil industrial furnaces (including heating furnaces, hot air furnaces, incinerators, reactors, boilers, etc.), and can realize the purging, ignition, flame monitoring, fuel pressure, flow high and low limits, air low limit control of the combustion process, as well as the precise adjustment of flue gas parameter content and furnace pressure, and can independently constitute a combustion automatic control system. It supports multiple field bus protocols, has multiple communication interfaces such as RS485, CAN, and RS232, and is easy to connect to a computer to form a distributed control system. The internal solidification software adopts an embedded multi-tasking processing mechanism, simulates artificial intelligence, and modular multiple function program blocks can be directly called; it has an intelligent network unit controller with strong applicability, high control accuracy, and good reliability.

[0006] Chinese invention patent CN110442100A, a thermal control intelligent DCS diagnosis and early warning method and system, discloses a thermal control intelligent DCS diagnosis and early warning method and system. The method includes: obtaining data related to the operation of the power plant; according to the data, determining the power plant operation devices associated with the data; setting a value less than the data to obtain reset data; according to the reset data, issuing an early warning to the power plant operation devices to obtain a processing plan; according to the processing plan, testing the correctness of the early warning. In the present invention, after real-time data collection, the equipment status can be automatically monitored, the equipment status can be quickly determined, and equipment with serious defects or major hidden dangers can be quickly screened out from a large amount of test data, which greatly reduces the workload of operating personnel and users, improves work efficiency, and enhances the effectiveness of diagnostic results.

[0007] Although these existing technologies have expanded the functionality of DCS distributed control systems, they lack a high degree of structural integration with artificial intelligence technologies to achieve highly intelligent upgrades and improvements to DCS distributed control systems. Summary of the invention

[0008] In view of the problems described in the background technology, the present invention proposes a technical solution of a DCS artificial intelligence module and its optimization method, which realizes the combination of DCS distributed control system and AI artificial intelligence, performs collaborative control, and improves the control optimization capability of the distributed control system. The optimized training / application integrated incremental learning model is integrated into the DCS to realize abnormal identification and automatic startup solutions.

[0009] Under the adjustment of control signals, the DCS distributed control system triggers the transmission of various parameters to the artificial intelligence module through the communication interface. The artificial intelligence module runs and outputs two data results: A. Model parameters: This is the most direct training output, which refers to the parameter values ​​such as weights and biases that are continuously optimized during the training process; B. Output target variable parameters or decisions based on the input data model, category labels in classification problems, and numerical predictions in regression problems. The application output results are returned to the DSC distributed control system as control parameter instructions. The specific technical solutions are as follows:

[0010] A DCS artificial intelligence module is based on a high-performance industrial computer hardware and software platform composed of CPU and GPU. The core optimization control of this module includes two major parts: a virtual controller and an artificial intelligence module. The virtual controller consists of a communication protocol driver module, a configuration parameter input module, a dynamic parameter input module, and an optimization control output module. The artificial intelligence module adopts an incremental learning model, which can be used as an application machine learning model to output predicted target variable parameters, and can also perform machine learning training.

[0011] During the machine learning training process, the incremental learning model inputs the feature parameters and target variable parameters corresponding to the specified timestamp T, presets the duration N, and simultaneously inputs the target variable parameters corresponding to the timestamp T+N value in the time series, and participates in the probability fitting training of the artificial intelligence model. When the incremental learning model is used as an applied machine learning model, the predicted output content includes the target variable parameters at the next N moments, so that the incremental learning model has the ability to predict on the timeline.

[0012] As a preferred technical solution of the present invention, when evaluating the accuracy of the incremental learning model, the target variable weight coefficient output by the preset machine learning is used to optimize the evaluation algorithm. The impact of the target variable error on the evaluation result is related to the weight of the target parameter. The exponential function in the variance calculation formula is replaced by the product of the base of the exponential function and the absolute value of the base of the exponential function. The positive and negative values ​​of the variance exponential function are retained to perform calculations for the accuracy evaluation of the training model. The calculated result value is used as an evaluation indicator of the Gaussian distribution offset of the target variable parameter error, and at the same time as an evaluation indicator for measuring the degree of discreteness of the target variable parameter error.

[0013] As a preferred technical solution of the present invention, the incremental learning model accuracy evaluation: the input feature data and target variable parameters are subjected to the probability fitting operation of the artificial intelligence model, and the incremental learning model outputs a set of predicted target variable parameters y ( , , .... ), where the elements correspond one to one to the weight coefficient k ( , , .... ), standard target variable parameter b ( , , .... ), the error evaluation parameter η of the target variable offset discrete is calculated by retaining the discrete positive and negative attributes:

[0014]

[0015] Where: η is the error evaluation parameter of the target variable offset discrete, k is the weight coefficient of the target variable, is the target variable parameter output by the incremental learning model, is the weight coefficient, are standard target variable parameters.

[0016] During the machine learning training process, the incremental learning model inputs feature data and target variable parameters including exception codes and solution codes, and performs probability fitting training. When the incremental learning model is output as an application model, the output content includes exception codes and solution codes.

[0017] When the incremental learning model is output as an application model, the output content includes an exception code and a solution code. Each solution code corresponds to a solution. Each solution includes a communication person in charge and his or her mobile phone number. When a solution is started, the exception information is automatically sent to the corresponding mobile phone number through a text message to notify the corresponding communication person in charge.

[0018] The artificial intelligence module is connected to the communication network or field communication bus of the DCS distributed control system through a communication interface (such as serial port RS232, RS485, CAN, Ethernet), so as to realize the optimization control of the DCS distributed control system or its subsystem by the artificial intelligence module.

[0019] The incremental learning model in the artificial intelligence module outputs the target variable parameters as an application model, and can compare the field target variable parameters obtained through the communication network, field communication bus connection and DCS control, and issue an alarm when the comparison error exceeds a specified threshold.

[0020] The DCS artificial intelligence module supports the expansion of multiple different types of DCS proprietary artificial intelligence modules in a DCS distributed control system, and supports synchronous parallelism.

[0021] The DCS artificial intelligence module collects data regularly when the DCS distributed control system is in operation, optimizes and trains the model to suit the latest actual environment, solves the matching of equipment aging, temperature changes, humidity environmental parameters and the model, and forms a self-evolution and self-training technical effect.

[0022] Compared with the prior art, the beneficial technical effects of the present invention are:

[0023] 1. A DCS artificial intelligence module and its optimization method. The evaluation of the accuracy of the training model sets the weight coefficient through the target variable output by machine learning, retains the attributes of the positive and negative values ​​of the variance to calculate the accuracy evaluation of the training model. The calculated value is used as an evaluation indicator of the central benchmark offset of the target variable parameter, which can accurately measure the offset of the peak center of the normal distribution of the output value error, thereby providing a reliable reference value for the fine-tuning of the machine learning model, greatly improving the efficiency of the fine-tuning of the training model; the calculated value is also used as an evaluation indicator to measure the discrete degree of the target variable parameter. The evaluation of the accuracy of the training model enlarges the interval of discrete evaluation and increases the accuracy of the basic scale of the correction evaluation.

[0024] 2. A DCS artificial intelligence module and its optimization method. The artificial intelligence module adopts an incremental learning model, which can be used as an application machine learning model output to predict target variable parameters, and can also perform machine learning training. The machine learning training and application output are integrated. When the distributed control system is operating, data is collected regularly and the model is optimized and trained to suit the latest actual environment, so as to solve the problems of equipment aging, temperature changes, humidity environmental parameters and model matching, and form a self-evolution and self-training technical effect.

[0025] 3. A DCS artificial intelligence module and its optimization method, through the virtual controller inside the DCS artificial intelligence module, access to the communication network or field communication bus of the existing DCS distributed control system, to achieve seamless embedding into the DCS distributed control system to achieve optimized control. The access cost is low, the solution is simple, reliable, and the implementation is fast.

[0026] 4. A DCS artificial intelligence module and an optimization method thereof, which automatically establishes proprietary training data sets and test data sets for corresponding models of systems, subsystems, and equipment in the DCS artificial intelligence module through a dynamic parameter input module, a training data set module, and a test data set module in a virtual controller inside the DCS artificial intelligence module.

[0027] 5. A DCS artificial intelligence module and its optimization method. Through the optimization control output module in the virtual controller inside the DCS artificial intelligence module, the existing DCS distributed control system is optimized and controlled, and the parameters of the equipment of the DCS distributed control system and subsystems are changed and adjusted across multiple layers of networks and multiple field buses.

[0028] 6. A DCS artificial intelligence module and its optimization method can easily expand multiple different types of proprietary DCS artificial intelligence modules in a DCS distributed control system, and support synchronous parallelism, so that the incremental learning model of the DCS distributed control system and artificial intelligence is highly integrated, greatly improving the intelligence and automation of the DSC distributed control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is an overall logical architecture diagram of a DCS artificial intelligence module and its optimization method.

[0030] Figure 2 It is a data flow structure diagram of the incremental learning model application in a DCS artificial intelligence module and its optimization method.

[0031] Figure 3 It is a comparison chart of the actual dust concentration value and the model output value during the incremental learning model evaluation in the DCS artificial intelligence module of the dust removal system.

[0032] Figure 4This is a simplified diagram of the logical architecture of the DCS artificial intelligence module of the dust removal system. DETAILED DESCRIPTION

[0033] A DCS artificial intelligence module and its optimization method are based on a high-performance industrial computer hardware and software platform composed of CPU and GPU. The core optimization control of the DCS artificial intelligence module includes two parts: a virtual controller and an artificial intelligence module. The artificial intelligence algorithm module is connected to the communication network or field communication bus of the DCS distributed control system through a communication interface (such as serial port RS232, RS485, CAN, Ethernet), so as to realize the optimization control of the DCS distributed control system or its subsystems by the artificial intelligence module.

[0034] Under the adjustment of the control signal, the DCS distributed control system triggers the transmission of various parameters to the artificial intelligence module through the communication interface. The artificial intelligence module runs and outputs two data results: A. Model parameters: This is the most direct training output, which refers to the parameter values ​​such as weights and biases that are continuously optimized during the training process. B. Application output results: Based on the input data, the model outputs the target variable parameters or decisions, the category labels in the classification problem, and the numerical predictions in the regression problem, and returns them to the DSC distributed control system as the control parameter instructions.

[0035] The DCS artificial intelligence module and its optimization method support the expansion of multiple different types of DCS proprietary artificial intelligence modules in a DCS distributed control system, and support synchronous parallelism.

[0036] The core optimization control virtual controller part of the DCS artificial intelligence module includes a communication protocol driver module, a configuration parameter input module, a dynamic parameter input module, and an optimization control output module.

[0037] The communication protocol driver module is implemented according to the communication protocol rules of the communication network or field communication bus of the DCS distributed control system to realize the interconnection between the DCS artificial intelligence module communication interface and the DCS distributed control system. At the same time, the DCS artificial intelligence module is virtualized into an identifiable embedded controller in the DCS distributed control system on the computer platform on which it runs, so as to realize the DCS artificial intelligence module for fast and convenient control of the DCS distributed control system. The configuration can be selected according to the protocol of the existing DCS distributed control system, such as supporting Modbus, Modbus TCP / IP, Profibus, Ethernet / IP, FoundationFieldbus, CDT, DL / T645, IEC60870-5 series (such as 101, 102, 103, 104).

[0038] The configuration parameter input module is used to receive parameters input by users from the HMI human-machine interface in the DCS distributed control system to enable users to change and configure the parameters of the artificial intelligence module.

[0039] The dynamic parameter input module is used to receive real-time operation data from DCS distributed control systems, subsystems, equipment, etc. The real-time operation data of systems, subsystems, equipment, etc.: one is used to establish the training data set and test data set of the DCS artificial intelligence module, and the other is used as the parameter input of the real-time optimization control module in the DCS artificial intelligence module. After the model analysis and prediction of the real-time optimization control module, the optimized control parameters are generated. Figure 1 As shown, the DCS distributed control system will obtain the S1, Sx and other subsystem related parameters through the communication network or the field communication bus, the A1, Ax, B1, Bx, C1, Cx, D1, Dx, E1, Ex, F1, Fx and other embedded controller related parameters in the subsystem, and the PLC1, PLCx and other programmable controller related parameters through the communication network or the field communication bus. The parameters are refreshed into the dynamic parameter input module of the virtual controller at the same time point of the fixed time interval cycle.

[0040] The optimization control output module receives the control parameters from the artificial intelligence module and is used to optimize the control of the DCS distributed control system and change the control parameters of the control system, subsystems, equipment, etc. Figure 1 As shown, the optimized control output module in the virtual controller of the DCS artificial intelligence module updates the optimized control parameters into the DCS distributed control system through the communication network or the field communication bus. The DCS distributed control system then writes the specific control parameters into the corresponding subsystems such as S1 and Sx, and the embedded controllers such as A1, Ax, B1, Bx, C1, Cx, D1, Dx, E1, Ex, F1, Fx, and programmable controllers such as PLC1 and PLCx in the subsystem.

[0041] The artificial intelligence module part of the DCS artificial intelligence module includes a training data set module, a test data set module, an optimization model training module, an intelligent model screening module, and a real-time optimization control module.

[0042] Training data set module: The real-time operation data of the DCS distributed control system, subsystems, equipment, etc. received from the dynamic parameter input module is pre-processed to generate a training data set for the deep learning model training of the DCS artificial intelligence module. Since the real-time operation data of the DCS distributed control system is time-varying data, the training data set should synchronously record the data of the system, subsystem, and equipment at the same time point, and record the real-time operation data at fixed intervals to continuously generate sufficient time series data samples distributed in the system, subsystem, and equipment corresponding to various working conditions. The training data set is used in key links such as model learning, feature extraction, weight optimization, overfitting and generalization control, model selection and tuning, and performance benchmark comparison of the DCS artificial intelligence module.

[0043] Test data set module: The real-time operation data of the DCS distributed control system, subsystems, equipment, etc. received from the dynamic parameter input module, excluding the data set used for training, is preprocessed to generate a test data set for the deep learning model test of the DCS artificial intelligence module. Since the real-time operation data of the DCS distributed control system is time-varying data, the test data set should synchronously record the data of the system, subsystem, and equipment at the same time point, and record the real-time operation data at a fixed interval time period to generate sufficient time series data samples distributed in the system, subsystem, and equipment corresponding to various working conditions. The test data set is used for model performance evaluation, optimal model selection, hyperparameter tuning, model robustness verification, comparison of different algorithms, and final inspection before deployment of the DCS artificial intelligence module. The model performance evaluation usually uses mean square error (MSE), mean absolute error (MAE), etc.

[0044] Data preprocessing: In engineering practice, the data we obtain may contain missing values, duplicate values, etc., and data preprocessing is required before use. The common process of data preprocessing is: removing unique attributes, processing missing values, attribute encoding, data standardization and regularization, feature selection, and principal component analysis. Since the real-time operation data of the DCS distributed control system is time-varying data, standardization and normalization methods can help the DCS distributed control system better capture the characteristics of the time series.

[0045] Standardized processing method:

[0046]

[0047] Among them, x is the original data, μ is the mean of the data, and δ is the standard deviation of the data.

[0048] Normalization method:

[0049]

[0050] Among them, x is the original data, min is the minimum value of the data, and max is the maximum value of the data.

[0051] Optimize model training module: including DCS artificial intelligence model establishment and model training.

[0052] DCS artificial intelligence model establishment: First, determine that the objects of the DCS artificial intelligence module are the global objects, subsystem objects, and subsystem equipment of the DCS distributed control system. Secondly, according to the characteristics of the specific objects of the DCS distributed control system, determine the parameters required by the DCS artificial intelligence module, obtain data through the virtual controller, and establish training and test data sets through the training data set module and the test data set module. Thirdly, according to the selected parameters of the specific objects of the DCS distributed control system, divide the selected parameters into two groups, one group as the input samples of the deep learning neural network, and the other group as the input samples of the deep learning neural network and the corresponding output labels of the deep learning neural network to construct a DCS artificial intelligence model based on the deep learning neural network. According to different grouping methods of the selected parameters, different multiple DCS artificial intelligence incremental learning models based on deep learning neural networks can be constructed. Different grouping methods for the same DCS distributed control parameters in the same training and test sets produce different input parameter samples, different output labels, and constitute different models. Refer to the data flow structure diagram for the incremental learning model application. Figure 2 .

[0053] Secondly, select the appropriate network structure and choose the appropriate deep learning model according to the characteristics of the object, such as convolutional neural network (CNN), recurrent neural network (RNN) and its variants (such as LSTM, GRU). Finally, according to the characteristics of the specific object of the DCS distributed control system, determine the number of layers of the appropriate network, the number of neurons in each layer, the activation function, etc.

[0054] As the preferred technical solution of the present invention, an incremental learning model is adopted, which can be used as an application machine learning model to output predicted target variable parameters, and can also perform machine learning training; the DCS artificial intelligence module, when the DCS distributed control system is in operation, regularly collects data and optimizes the model for training to suit the latest actual environment, solves equipment aging, temperature changes, humidity environmental parameters and the matching of the model, and forms a self-evolution and self-training technical effect. Example 1

[0055] Use Python and a library that supports online learning, such as SGDClassifier in sklearn or tensorflow, to train the incremental learning model process. The code is as follows:

[0056] from sklearn.linear_model import SGDClassifier

[0057] import pandas as pd

[0058] # Initialize the model

[0059] model = SGDClassifier()

[0060] # Set a function to get new data

[0061] def get_new_data():

[0062] # This should be the logic for getting new data

[0063] pass

[0064] # Call this function periodically while the application is running

[0065] while True:

[0066] # Get new data

[0067] new_data = get_new_data()

[0068] if new_data is not None:

[0069] # Preprocess new data

[0070] X_new, y_new = preprocess(new_data)

[0071] # Update the model with new data

[0072] model.partial_fit(X_new, y_new, classes=np.unique(y_new))

[0073] The incremental learning model uses known input data (features) and corresponding output labels (target variables) to train the model so that the model can learn the mapping relationship between input and output from these data. Once the model is trained, it can be used to make predictions on new, unknown data.

[0074] Furthermore, during the machine learning training process, the incremental learning model inputs the feature parameters and target variable parameters corresponding to the specified timestamp T, presets the duration N, and simultaneously inputs the target variable parameters corresponding to the timestamp T+N value in the time series to participate in the probability fitting training of the artificial intelligence model. When the incremental learning model is used as an applied machine learning model, the predicted output content includes the target variable parameters at the next N moments, so that the incremental learning model has the ability to predict on the timeline.

[0075] Intelligent model screening module: Use different data test sets of specific objects of DCS distributed control system under various working conditions to evaluate the performance of the model under different working conditions of DCS objects, including accuracy, recall rate, F1 score and other indicators. Discover the overfitting and underfitting phenomena of the model to ensure that the model has good generalization ability. Evaluate and screen the model with the best performance and the most suitable for the specific object requirements of DCS distributed control system from many candidate models.

[0076] As a preferred technical solution of the present invention, when evaluating the accuracy of the incremental learning model, the target variable weight coefficient output by the preset machine learning is used to optimize the evaluation algorithm. The impact of the target variable error on the evaluation result is related to the weight of the target parameter. The exponential function in the variance calculation formula is replaced by the product of the base of the exponential function and the absolute value of the base of the exponential function. The positive and negative values ​​of the variance exponential function are retained to perform calculations for the accuracy evaluation of the training model. The calculated result value is used as an evaluation indicator of the Gaussian distribution offset of the target variable parameter error, and at the same time as an evaluation indicator for measuring the degree of discreteness of the target variable parameter error.

[0077] As a preferred technical solution of the present invention, the incremental learning model accuracy evaluation: the input feature data and target variable parameters are subjected to the probability fitting operation of the artificial intelligence model, and the incremental learning model outputs a set of predicted target variable parameters y ( , , .... ), where the elements correspond one to one to the weight coefficient k ( , , .... ), standard target variable parameter b ( , , .... ), the error evaluation parameter η of the target variable offset discrete is calculated by retaining the discrete positive and negative attributes:

[0078]

[0079] Where: η is the error evaluation parameter of the target variable offset discrete, k is the weight coefficient of the target variable, is the target variable parameter output by the incremental learning model, is the weight coefficient, is the standard target variable parameter. When η is less than or equal to the specified threshold, the incremental learning model is open to use as an application output model. When η is greater than the specified threshold, the incremental learning model continues to be used as a machine learning training model.

[0080] Real-time optimization control module: select the incremental learning model of the DCS artificial intelligence module with the best performance in evaluation and screening, and switch it into the real-time operation state, continuously obtain the real-time operation data from the DCS distributed control system, subsystems, equipment, etc. in the dynamic parameter input module, and the DCS artificial intelligence model outputs the optimized control parameters according to the real-time operation data. Through the optimized control output module, the DCS distributed control system is optimized and the control parameters of the control system, subsystems, equipment, etc. are changed.

[0081] As a preferred technical solution of the present invention, during the machine learning training process, the incremental learning model inputs feature data and target variable parameters including exception codes and solution codes, and performs probability fitting training. When the incremental learning model is output as an application model, the output content includes exception codes and solution codes.

[0082] As a preferred technical solution of the present invention, when the incremental learning model is output as an application model, the output content includes an exception code and a solution code. Each solution code corresponds to a solution. Each solution includes a communication person in charge and his or her mobile phone number. When a solution is started, the exception information is automatically sent to the corresponding mobile phone number via text message to notify the corresponding communication person in charge.

[0083] The specific process of incremental learning model training in a DCS artificial intelligence module and its optimization method is as follows:

[0084] S1. Initialize the DCS artificial intelligence model parameters. Before training begins, randomly generate the parameters that need to be initialized.

[0085] S2. Forward propagation: Use the data in the training data set of the specific object of the DCS distributed control system under various working conditions to forward propagate through the model and calculate the output result.

[0086] S3.DCS artificial intelligence model calculates the loss and uses the loss function to calculate the difference between the predicted result and the true value to obtain the loss value.

[0087] S4. Back propagation, the DCS artificial intelligence model calculates the gradient of each parameter in the model through the back propagation algorithm based on the loss value.

[0088] S5. Update parameters. The DCS artificial intelligence model uses an optimization algorithm to update the parameters of the model according to the gradient.

[0089] S6. Iterative training: the DCS artificial intelligence model repeats the above steps until the preset training rounds are reached or the loss value drops below a certain threshold. Example 2

[0090] A DCS artificial intelligence module for a dust removal system of a subsystem in a DCS distributed control system of a coal-fired power plant in the power industry.

[0091] The DCS artificial intelligence module logic architecture diagram of the dust removal system is as follows: Figure 4 As shown in the figure, the DCS distributed control system of the dust removal system of a coal-fired power plant includes the control systems of the three main engines, boilers, steam turbines, and generators, as well as auxiliary systems such as denitrification, dust removal, desulfurization, and water treatment and environmental protection. The dust removal system also includes lower-level systems such as high-voltage power supply, low-voltage vibration heating, and pneumatic ash conveying. The specific implementation steps of the DCS artificial intelligence module of the dust removal system are as follows:

[0092] Step 1: According to the characteristics of the dust removal system object, determine the parameters required by the DCS artificial intelligence module, including: boiler coal feed rate, total air volume, flue gas temperature, oxygen content in flue gas, power generation, electric field voltage and current of A1~Ax, B1~Bx electric field high-voltage power supply controller in the ESP1~ESPx electrostatic precipitator of the dust removal system, and the dust outlet concentration.

[0093] Step 2: Obtain the parameters required by the DCS artificial intelligence module from the DCS distributed control system through the dynamic parameter input module of the virtual controller: boiler coal feed rate, total air volume, flue gas temperature, oxygen content in flue gas, power generation, dust removal system ESP1~ESPx, A1~Ax / B1~Bx in the electrostatic precipitator, electric field high voltage, electric field voltage and current of the power controller, and dust concentration at the precipitator outlet, and input the abnormal code and solution code for probability fitting training.

[0094] Step 3: The training data set module of the DCS artificial intelligence module obtains the parameters of the DCS distributed control system in real time and at specified future times through the dynamic parameter input module of the virtual controller: the coal feed rate of the boiler, the total air volume, the flue gas temperature, the oxygen content in the flue gas, the power generation, the electric field voltage and current of the A1~Ax and B1~Bx electric field high-voltage power supply controllers in the ESP1~ESPx electrostatic precipitators of the dust removal system, and the dust collector outlet concentration. The data is preprocessed and saved in time series and fixed time intervals to obtain training data set samples of various working conditions of the dust removal system.

[0095] Step 4. The test data set module of the DCS artificial intelligence module obtains the parameters of the DCS distributed control system in real time and at specified future times through the dynamic parameter input module of the virtual controller: the coal feed rate of the boiler, the total air volume, the flue gas temperature, the oxygen content in the flue gas, the power generation, the electric field voltage and current of the A1~Ax and B1~Bx electric field high-voltage power supply controllers in the ESP1~ESPx electrostatic precipitators of the dust removal system, and the emission concentration value at the outlet of the dust collector. The data is preprocessed and saved in time series and fixed time intervals to obtain test data set samples of various working conditions of the dust removal system.

[0096] Step 5: DCS artificial intelligence model is established. The goal is to ensure that the dust collector outlet concentration emission meets the standard under different operating conditions formed by different coal feed rates, total air volume, flue gas temperature, oxygen content in the flue gas, and power generation in the coal-fired power plant boiler, and to achieve the minimum energy consumption output of the A1~Ax and B1~Bx electric field high-voltage power supplies in the ESP1~ESPx electrostatic precipitators of the dust removal system, so as to achieve energy-saving optimization control.

[0097] Step 6, energy consumption prediction model: The selected parameters are divided into two groups according to the energy consumption characteristics. The coal supply, total air volume, flue gas temperature, oxygen content in flue gas, power generation, and electric field voltage of A1~Ax and B1~Bx electric field high-voltage power supply controllers in the dust removal system ESP1~ESPx electrostatic precipitator in the training data set are divided into one group as the input samples of the incremental learning neural network. The electric field current of A1~Ax and B1~Bx electric field high-voltage power supply controllers in the dust removal system ESP1~ESPx electrostatic precipitator in the training data set is used as the corresponding output label of the deep learning neural network, and the appropriate deep learning neural network structure is selected to build an incremental learning model for energy consumption prediction.

[0098] Step 7, outlet emission prediction model: The selected parameters are divided into two groups according to the outlet emission characteristics, and the coal supply, total air volume, flue gas temperature, oxygen content in flue gas, power generation, and electric field voltage of A1~Ax, B1~Bx electric field high-voltage power supply controller in the dust removal system ESP1~ESPx electrostatic precipitator in the training data set are used as input samples of the deep learning neural network. The outlet emission concentration value of the dust collector in the training data set is used as the corresponding output label of the deep learning neural network, and a suitable deep learning neural network structure is selected to construct an incremental learning model for outlet emission prediction.

[0099] Step 8: DCS artificial intelligence model training. The energy consumption prediction model and the export emission prediction model are trained based on the DCS training data set. The model parameters and model network are optimized. The training is iterated until the preset training rounds are reached or the loss value is reduced to below a certain threshold. That is, the incremental learning model is qualified for accuracy evaluation and then used. The input feature data and target variable parameters are subjected to the probability fitting operation of the artificial intelligence model. The incremental learning model outputs a set of predicted target variable parameters y( , , .... ), where the elements correspond one to one to the weight coefficient k ( , , .... ), standard target variable parameter b ( , , .... ), the error evaluation parameter η of the target variable offset discrete is calculated by retaining the discrete positive and negative attributes:

[0100]

[0101] Where: η is the error evaluation parameter of the target variable offset discrete, k is the weight coefficient of the target variable, is the target variable parameter output by the incremental learning model, is the weight coefficient, is the standard target variable parameter. When η is less than or equal to the specified threshold, the incremental learning model is open to use as an application output model. When η is greater than the specified threshold, the incremental learning model continues to be used as a machine learning training model.

[0102] Step 9, DCS artificial intelligence model screening: The energy consumption prediction model and the outlet emission prediction model are tested on the trained models based on the DCS test data set to examine the performance of the model on the test set, including evaluation indicators such as accuracy, precision, and recall, to identify overfitting and underfitting phenomena of the model to ensure that the model has good generalization ability. If it does not meet the requirements, retrain the energy consumption prediction and outlet emission prediction models. Until the evaluation test results and the final performance of the model meet the requirements of the dust removal system for the model, refer to the comparison chart of the actual dust concentration value and the model output value during the incremental learning model evaluation in the DCS artificial intelligence module of the dust removal system. Figure 3 .

[0103] Step 10. Use of DCS artificial intelligence model: After the energy consumption prediction and export emission prediction models are trained and tested, the incremental learning models of energy consumption prediction and export emission are put into operation. The dynamic parameter input module of the virtual controller is used to obtain the relevant real-time and future specified time target variables of the dust removal system of the DCS distributed control system subsystem. The parameters enter the energy consumption prediction and export emission prediction models. The model gives the optimized control parameters according to the operating parameters at real-time and future specified times. Through the optimized control output module of the virtual controller, the incremental learning model has the ability to predict on the time axis; realize the optimized control of the DCS distributed control system, change the parameters of the A1~Ax, B1~Bx electric field high-voltage power supply controllers in the ESP1~ESPx electrostatic precipitators of the dust removal system, adjust the power output of the high-voltage power supply, and achieve energy saving, efficiency improvement and emission reduction.

[0104] Special statement: The "embodiment" and the like described in this specification refer to the specific features, elements or characteristics described in conjunction with the embodiment included in the embodiment generally described in this application. The appearance of the same expression in multiple places in the specification does not limit the specific reference to the same embodiment. In other words, when describing a specific feature, element or characteristic in conjunction with any embodiment, it is claimed that such feature, element or characteristic is implemented in conjunction with other embodiments and is included in the scope of the claims for protection of the present invention; The embodiment describes the present invention with reference to multiple explanatory embodiments of the logical architecture and ideas of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art can design many other modifications and implementation methods within the framework of the technical solution of the present invention, and can make various non-essential modifications and improvements to the key points of the technical solution. For those skilled in the art, other uses will also be obvious, and non-substantial changes or replacements that can be easily thought of for implementation will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A DCS artificial intelligence module, the core optimization control of which includes two parts: a virtual controller and an artificial intelligence module, wherein the virtual controller is composed of a communication protocol driver module, a configuration parameter input module, a dynamic parameter input module, and an optimization control output module; characterized in that: The artificial intelligence module uses an incremental learning model as the output target variable parameter of the applied machine learning model and performs machine learning training at the same time; During the machine learning training process, the incremental learning model inputs the feature parameters and target variable parameters corresponding to the specified timestamp T, presets the duration N, and simultaneously inputs the target variable parameters corresponding to the timestamp T+N value in the time series, and participates in the probability fitting training of the artificial intelligence model. When the incremental learning model is used as an applied machine learning model, the predicted output content includes the target variable parameters at the next N moments, so that the incremental learning model has the ability to predict on the time axis; When evaluating the accuracy of the incremental learning model, the target variable weight coefficient output by the preset machine learning is used to optimize the evaluation algorithm. The impact of the target variable error on the evaluation result is related to the weight of the target parameter. The variance algorithm is optimized, and the exponential function in the variance calculation formula is replaced by the product of the base of the exponential function and the absolute value of the base of the exponential function. The positive and negative values ​​of the variance exponential function are retained to improve the calculation of the training model accuracy evaluation. The input feature data and target variable parameters are subjected to the probability fitting operation of the artificial intelligence model, and the incremental learning model outputs a set of target variable parameters. , , .... , one-to-one corresponding weight coefficient , , .... , standard target variable parameters , , .... , use the optimized variance algorithm to calculate the error evaluation parameter η of the target variable offset discrete: Where: η is the error evaluation parameter of the target variable offset discrete, k is the weight coefficient of the target variable, is the target variable parameter output by the incremental learning model, is the weight coefficient, is the standard target variable parameter; The calculated result value is used as an evaluation indicator of the Gaussian distribution deviation of the target variable parameter error. It is also used as an evaluation indicator to measure the discrete degree of the target variable parameter error, providing an accurate basis for fine-tuning the incremental learning model.

2. A DCS artificial intelligence module according to claim 1, characterized in that: During the machine learning training process, the incremental learning model inputs the exception code and the solution code in addition to the feature data and the target variable parameters to perform probability fitting training. When the incremental learning model is output as an application model, the output content includes the exception code and the solution code.

3. A DCS artificial intelligence module according to claim 2, characterized in that: When the incremental learning model is output as an application model, the output content includes an exception code and a solution code. Each solution code corresponds to a solution. Each solution includes a communication person in charge and his / her mobile phone number. When a solution is started, the exception information is automatically sent to the corresponding mobile phone number through a text message to notify the corresponding communication person in charge.

4. A DCS artificial intelligence module according to claim 1, characterized in that: The artificial intelligence module is connected to the communication network or field communication bus of the DCS distributed control system through a communication interface to realize the optimization control of the DCS distributed control system or its subsystems by the artificial intelligence module.

5. A DCS artificial intelligence module according to claim 1, characterized in that: The incremental learning model in the artificial intelligence module outputs target variable parameters as an application model, which can be compared with the on-site target variable parameters obtained through the communication network, on-site communication bus connection and DCS control, and an alarm is issued when the comparison error exceeds a specified threshold.

6. A DCS artificial intelligence module according to claim 1, characterized in that: It supports the expansion of multiple different types of proprietary DCS artificial intelligence modules in a DCS distributed control system, and supports synchronous parallelism.

7. The DCS artificial intelligence module according to claim 1, characterized in that: When the DCS distributed control system is operating, the incremental learning model regularly collects data and optimizes the model to adapt to the latest actual environment, solves the matching of equipment aging, temperature changes, humidity environmental parameters and the model, and forms a self-evolution and self-training technical effect.

Images