An industrial AI closed-loop adaptive monitoring control system

Abstract

The application relates to the technical field of monitoring control systems, and provides an industrial AI closed-loop self-adaptive monitoring control system, which comprises a data acquisition module, a data preprocessing module and a feature extraction module; the data acquisition module comprises a multi-source sensor unit and is used for collecting state parameters of an industrial process in real time; the data preprocessing module comprises an abnormal value detection and processing unit and a data normalization unit and is used for cleaning and standardizing the state parameters; the feature extraction module comprises a time domain feature extraction unit, a frequency domain feature extraction unit and a time sequence feature fusion unit based on an attention mechanism and is used for generating a multi-dimensional fusion feature vector from the standardized data. Through multi-source data processing, time-frequency feature fusion and mixed neural network modeling, the industrial process monitoring precision and health evaluation reliability are improved; PID parameter self-adaptive optimization is realized in combination with reinforcement learning, safety protection and online updating are matched, the closed-loop control performance is significantly improved, complex working conditions are adapted, and the stable and efficient operation of the industry is ensured.

Description

Technical Field

[0001] This invention relates to the field of monitoring and control system technology, specifically to an industrial AI closed-loop adaptive monitoring and control system. Background Technology

[0002] Industrial production, as a core pillar of the national economy, is widely applied in key sectors such as chemical engineering, metallurgy, power generation, and machinery manufacturing. The stability, safety, and efficiency of its production processes directly determine product quality, production efficiency, and corporate economic benefits. With the deepening integration of industry and intelligent manufacturing, the complex operating conditions, coupled parameters, and dynamic time-varying characteristics of industrial production sites place higher demands on the real-time performance, accuracy, and adaptability of process monitoring, fault early warning, and closed-loop control. Industrial monitoring and control systems have become core devices for ensuring the stable operation of industrial production.

[0003] Current industrial monitoring and control systems mostly adopt traditional monitoring and control schemes. At the monitoring level, they rely on single sensors to collect data, which is easily affected by on-site noise, outliers cannot be effectively removed, and only simple time-domain features can be extracted, making it difficult to capture the long-term time-series dependence and deep correlation of industrial parameters. This results in low reliability of monitoring data, poor accuracy of state prediction, difficulty in identifying minor equipment anomalies, and easy escalation of faults. At the control level, fixed-parameter PID or simple adaptive PID control is commonly used, with parameters relying on the manual experience of field engineers for tuning. When faced with fluctuations in operating conditions, the adjustment is lagging and the adaptability is poor, which easily leads to problems such as large steady-state deviation, high overshoot, and long settling time. Moreover, there is a lack of fault safety protection and model iterative optimization mechanisms in actual production scenarios, making it unable to adapt to complex dynamic industrial scenarios. Ultimately, this results in high energy consumption, low product qualification rate, and high equipment failure rate in industrial production, making it difficult to meet the high-precision, high-stability, and high-safety production requirements of modern industry. In response to these technical problems, this application proposes an industrial AI closed-loop adaptive monitoring and control system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an industrial AI closed-loop adaptive monitoring and control system, which solves the problems of inaccurate monitoring, poor control adaptation, and difficulty in meeting the needs of complex working conditions in existing industrial monitoring and control systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an industrial AI closed-loop adaptive monitoring and control system, comprising: The data acquisition module includes a multi-source sensor unit for real-time acquisition of state parameters of industrial processes; The data preprocessing module includes an outlier detection and processing unit and a data normalization unit, which are used to clean and standardize the state parameters. The feature extraction module includes a temporal feature extraction unit, a frequency domain feature extraction unit, and an attention-based temporal feature fusion unit, which are used to generate multi-dimensional fused feature vectors from standardized data. The AI ​​prediction and evaluation module has a built-in pre-trained hybrid neural network model, which includes a parallel LSTM layer, a Transformer encoder layer, and a gated fusion unit, used to output the predicted process state value and health score for future time moments based on the fused feature vector. The adaptive control strategy generation module includes a reinforcement learning agent unit and a parameter increment calculation unit, which is used to generate control parameter adjustment amounts based on the process state prediction value, health score and current control parameters; The closed-loop execution and feedback module is used to send the adjusted control parameters to the actuator and receive real-time feedback data after execution, forming a closed-loop circuit. The online model update module is used to incrementally train the hybrid neural network model and the reinforcement learning agent unit using the latest accumulated running data when the preset update trigger conditions are met.

[0006] Preferably, the outlier detection and processing unit uses an outlier detection algorithm based on isolated forest to identify outliers that exceed the normal distribution range and interpolate them using the mean of adjacent normal values.

[0007] Preferably, the attention-based temporal feature fusion unit fuses temporal features and frequency domain features by calculating the interaction attention weights, and the formula for calculating the fused feature vector is as follows: ; in, and These are the time-domain feature matrix and the frequency-domain feature matrix, respectively. , , The weight parameter matrix is ​​a learnable matrix. This is the scaling factor.

[0008] Preferably, in the hybrid neural network model, the LSTM layer is used to capture short-term temporal dependencies, the Transformer encoder layer uses a multi-head self-attention mechanism to capture long-term dependencies, and the hidden states output by the two are dynamically combined through the gating fusion unit.

[0009] Preferably, the calculation formula for the gating fusion unit is: ; ; in, and At time 10:00, respectively, the LSTM layer and the Transformer encoder layer are... The output, For the sigmoid function, Represents element-wise product. and These are the gating parameters.

[0010] Preferably, the health score is generated by quantifying the difference between the predicted process state value at the future time and the preset ideal process envelope, and mapping it to a range of 0 to 1, where 1 represents complete health.

[0011] Preferably, the reinforcement learning agent unit employs a deep deterministic policy gradient algorithm, and its state space consists of the fused feature vector at the current moment, the predicted process state value, and the health score, while its action space consists of the proportional coefficient in the PID control parameters. Integral coefficient and differential coefficients The adjustment amount, the reward function is designed as follows: ; in, The deviation between the process variable and the set value at the next moment. To control the amount of change in output, , These are non-negative weighting coefficients.

[0012] Preferably, the parameter increment calculation unit calculates the action value output by the reinforcement learning agent unit. , , After being limited by the rate of change and the magnitude, the values ​​are incrementally added to the current PID parameters.

[0013] Preferably, the adaptive control strategy generation module further includes a safety protection unit, which is used to forcibly compress the upper limit of the amplitude limit of the parameter increment calculation unit and generate an alarm signal when the received health score is lower than a preset safety threshold.

[0014] Preferably, the preset update trigger condition for the online model update module is: continuous If the health score for a monitoring cycle is lower than a preset threshold, or the moving average of the process status prediction error exceeds a preset upper limit.

[0015] This invention provides an industrial AI closed-loop adaptive monitoring and control system. It has the following beneficial effects: 1. This invention achieves dual capture of short-term fluctuations and long-term trends of industrial parameters through multi-source data cleaning, time-frequency feature fusion and LSTM-Transformer hybrid neural network modeling. It significantly reduces process state prediction error, greatly improves the accuracy of health score, accurately identifies equipment abnormalities, provides early warning of faults, and ensures stable operation of industrial processes.

[0016] 2. This invention combines reinforcement learning to dynamically generate PID parameter adjustment amounts, along with a safety protection mechanism and online model update function, to solve the problems of parameter solidification or adjustment lag in traditional fixed PID and adaptive PID. It effectively reduces steady-state deviation, reduces overshoot, shortens adjustment time, and maintains a high healthy operating rate, adapting to complex and ever-changing industrial conditions. Attached Figure Description

[0017] Figure 1 This is a diagram of the data acquisition module of the present invention; Figure 2 This is a diagram of the data preprocessing module of the present invention; Figure 3 This is a diagram of the feature extraction module of the present invention; Figure 4 This is a diagram of the AI ​​prediction and evaluation module of the present invention; Figure 5 A module diagram for generating the adaptive control strategy of the present invention is provided. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: This invention provides an industrial AI closed-loop adaptive monitoring and control system, comprising: Please see the appendix Figure 1 The data acquisition module includes a multi-source sensor unit for real-time acquisition of state parameters of industrial processes; Specifically, the multi-source sensor unit is configured according to industrial site requirements, including seven types of sensors: temperature, pressure, flow rate, vibration, speed, current, and voltage. The sampling frequency is uniformly set to 10Hz, supporting both 4–20mA analog and Modbus TCP digital interfaces. Real-time acquired status parameters include: medium temperature (0–500℃), pipeline pressure (0–10MPa), and medium flow rate (0–500m³ / s). 3The data includes, but is not limited to, the following parameters: vibration acceleration (0–20g), motor speed (0–3000r / min), operating current (0–500A), and power supply voltage (0–400V). After being filtered by the signal conditioning board, the collected data is uploaded to the edge computing gateway cache, which supports continuous raw data storage for 72 hours.

[0020] ; Please see the appendix Figure 2 The data preprocessing module includes an outlier detection and processing unit and a data normalization unit, which are used to clean and standardize the state parameters. The outlier detection and processing unit adopts an outlier detection algorithm based on isolated forest to identify outliers that exceed the normal distribution range and interpolate them using the mean of adjacent normal values. Specifically, the outlier detection and processing unit uses the isolated forest algorithm with 100 trees, a subsampling size of 256, and an outlier score threshold of 0.6. Scores higher than the threshold are considered outliers. During processing, the mean of the two valid sampling points before and after the outlier is used for interpolation to prevent the spread of single-point outliers. The data normalization unit uses minimum-maximum normalization, with the formula as follows: The upper and lower limits of each parameter are determined based on the historical data of steady-state operation of the equipment. After normalization, the data is mapped to the [0,1] interval to eliminate the difference in dimensions.

[0021] ; Please see the appendix Figure 3 The feature extraction module includes a temporal feature extraction unit, a frequency domain feature extraction unit, and an attention-based temporal feature fusion unit. These units generate multi-dimensional fused feature vectors from the standardized data. The attention-based temporal feature fusion unit fuses temporal and frequency domain features by calculating the interaction attention weights. The formula for calculating the fused feature vector is as follows: ; in, and These are the time-domain feature matrix and the frequency-domain feature matrix, respectively. , , The weight parameter matrix is ​​a learnable matrix. This is the scaling factor; Specifically, the temporal feature extraction unit extracts seven types of temporal features: mean, variance, peak value, kurtosis, skewness, range, and RMS value. Each unit outputs a 7-dimensional vector with a single parameter, resulting in a total of 49-dimensional temporal feature matrices across the seven sensor types. The frequency domain feature extraction unit converts time-series data into the frequency domain using Fast Fourier Transform (FFT) and extracts five types of frequency domain features: dominant frequency, amplitude, harmonic distortion rate, power spectral density, and frequency domain entropy, outputting a 35-dimensional frequency domain feature matrix. scaling factor Set to 64, weight matrix , Random initialization is performed and updated during model training. Interactive attention mechanisms are used to enhance the correlation between time-frequency features, ultimately outputting a 64-dimensional fused feature vector. As the input of subsequent models .

[0022] ; Please see the appendix Figure 4 The AI ​​prediction and evaluation module includes a pre-trained hybrid neural network model. This model comprises parallel LSTM layers, Transformer encoder layers, and a gated fusion unit. It outputs predicted process states and health scores for future timeframes based on fused feature vectors. In the hybrid neural network model, the LSTM layer captures short-term temporal dependencies, while the Transformer encoder layer utilizes a multi-head self-attention mechanism to capture long-term dependencies. The hidden states output by both layers are dynamically combined through the gated fusion unit. The calculation formula for the gated fusion unit is as follows: ; ; in, and At time 10:00, respectively, the LSTM layer and the Transformer encoder layer are... The output, For the sigmoid function, Represents element-wise product. and These are the gating parameters; Specifically, the LSTM layer is stacked in two layers, with 128 hidden units and a dropout rate of 0.2, capturing short-term temporal fluctuations within 10 steps, and outputting... The Transformer encoder has two layers, four multi-head attention heads, 128 hidden dimensions, a positional encoding length of 50, and captures long-term trends within 50 steps. The output... Gating parameters of the gating fusion unit , sigmoid output gate weights Dynamically allocate short-term or long-term feature weights and output the fused hidden state. Prediction and rating The predicted state of the next 5 steps is output through the fully connected layer; the predicted value and the ideal process envelope are quantified by mean square error, normalized and mapped to [0,1] to obtain the health score, where 1 is completely healthy.

[0023] ; Please see the appendix Figure 5 The adaptive control strategy generation module includes a reinforcement learning agent unit and a parameter increment calculation unit. It generates control parameter adjustments based on the predicted process state, health score, and current control parameters. The health score is generated by quantifying the difference between the predicted process state at future times and a preset ideal process envelope, mapping it to a range of 0 to 1, where 1 represents complete health. The reinforcement learning agent unit employs a deep deterministic policy gradient algorithm. Its state space consists of the fused feature vector at the current time, the predicted process state, and the health score. Its action space is the proportional coefficient in the PID control parameters. Integral coefficient and differential coefficients The adjustment amount, the reward function is designed as follows: ; in, The deviation between the process variable and the set value at the next moment. To control the amount of change in output, , The adaptive control strategy generation module also includes a safety protection unit, which is used to forcibly compress the upper limit of the amplitude limit of the parameter increment calculation unit and generate an alarm signal when the received health score is lower than the preset safety threshold. Specifically, the state space of a reinforcement learning agent unit consists of 64 fused features + 7 predicted values ​​+ 1 health dimension, for a total of 72 dimensions. Action space: , , Adjustment amount, range [-0.5, 0.5]; Reward weight: , Prioritizes suppressing bias while also ensuring stability; Training parameters: learning rate The experience playback pool has a capacity of 100,000, and the batch size is 64. The parameter increment calculation unit's action value is limited by the rate of change (single-step change ≤ 0.2) and amplitude. , , Then, the increment is added to the current PID parameter to avoid parameter abrupt changes. The safety protection unit's safety threshold is set to 0.6. When the health status is <0.6, the upper limit of parameter adjustment amplitude is compressed by 50%, and an audible and visual alarm is triggered simultaneously, and fault information is uploaded.

[0024] ; The closed-loop execution and feedback module is used to send the adjusted control parameters to the actuator and receive real-time feedback data after execution, forming a closed-loop circuit. Specifically, the online model update module is used to incrementally train the hybrid neural network model and reinforcement learning agent unit using the latest accumulated running data when the preset update trigger conditions are met. The preset update trigger conditions for the online model update module are: the health score is lower than the preset threshold for N consecutive monitoring periods, or the moving average of the process state prediction error exceeds the preset upper limit.

[0025] ; The online model update module is used to incrementally train the hybrid neural network model and reinforcement learning agent units using the latest accumulated running data when a preset update trigger condition is met. The preset update trigger condition for the online model update module is: continuous... If the health score for a monitoring cycle is lower than the preset threshold, or the moving average of the process status prediction error exceeds the preset upper limit; Specifically, the update trigger condition is N=20 (2000ms), the health threshold is 0.7, and the prediction error upper limit is 0.2. Incremental training is triggered when any condition is met. The incremental training strategy adopts mini-batch incremental learning, taking the latest 5000 valid data points each time, with 5 training rounds and the learning rate reduced to 1 / 10 of the initial value to avoid the model forgetting old knowledge. After the update, the model's prediction accuracy is automatically verified. If the accuracy drops by more than 3%, the model is rolled back to the model before the update to ensure system stability.

[0026] ; Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An industrial AI closed-loop adaptive monitoring and control system, characterized in that, include: The data acquisition module includes a multi-source sensor unit for real-time acquisition of state parameters of industrial processes; The data preprocessing module includes an outlier detection and processing unit and a data normalization unit, which are used to clean and standardize the state parameters. The feature extraction module includes a temporal feature extraction unit, a frequency domain feature extraction unit, and an attention-based temporal feature fusion unit, which are used to generate multi-dimensional fused feature vectors from standardized data. The AI ​​prediction and evaluation module has a built-in pre-trained hybrid neural network model, which includes a parallel LSTM layer, a Transformer encoder layer, and a gated fusion unit, used to output the predicted process state value and health score for future time moments based on the fused feature vector. The adaptive control strategy generation module includes a reinforcement learning agent unit and a parameter increment calculation unit, which is used to generate control parameter adjustment amounts based on the process state prediction value, health score and current control parameters; The closed-loop execution and feedback module is used to send the adjusted control parameters to the actuator and receive real-time feedback data after execution, forming a closed-loop circuit. The online model update module is used to incrementally train the hybrid neural network model and the reinforcement learning agent unit using the latest accumulated running data when the preset update trigger conditions are met.

2. The industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, The outlier detection and processing unit uses an outlier detection algorithm based on isolated forests to identify outliers that exceed the normal distribution range and interpolate them using the mean of adjacent normal values.

3. The industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, The attention-based temporal feature fusion unit fuses temporal and frequency domain features by calculating the interaction attention weights. The formula for calculating the fused feature vector is as follows: ； in, and These are the time-domain feature matrix and the frequency-domain feature matrix, respectively. , , The weight parameter matrix is ​​a learnable matrix. This is the scaling factor.

4. The industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, In the hybrid neural network model, the LSTM layer is used to capture short-term temporal dependencies, and the Transformer encoder layer uses a multi-head self-attention mechanism to capture long-term dependencies. The hidden states output by the two layers are dynamically combined through the gating fusion unit.

5. The industrial AI closed-loop adaptive monitoring and control system according to claim 4, characterized in that, The calculation formula for the gated fusion unit is as follows: ; ； in, and At time 10:00, respectively, the LSTM layer and the Transformer encoder layer are... The output, For the sigmoid function, Represents element-wise product. and These are the gating parameters.

6. The industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, The health score is generated by quantifying the difference between the predicted process state value at the future time and the preset ideal process envelope, and mapping it to the interval between 0 and 1, where 1 represents complete health.

7. The industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, The reinforcement learning agent unit employs a deep deterministic policy gradient algorithm. Its state space consists of the fused feature vector at the current moment, the predicted process state value, and the health score. Its action space is the proportional coefficient in the PID control parameters. Integral coefficient and differential coefficients The adjustment amount, the reward function is designed as follows: ； in, The deviation between the process variable and the set value at the next moment. To control the amount of change in output, , These are non-negative weighting coefficients.

8. The industrial AI closed-loop adaptive monitoring and control system according to claim 7, characterized in that, The parameter increment calculation unit calculates the action value output by the reinforcement learning agent unit. , , After being limited by the rate of change and the magnitude, the values ​​are incrementally added to the current PID parameters.

9. The industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, The adaptive control strategy generation module also includes a safety protection unit, which is used to forcibly compress the upper limit of the amplitude limit of the parameter increment calculation unit and generate an alarm signal when the received health score is lower than the preset safety threshold.

10. An industrial AI closed-loop adaptive monitoring and control system according to claim 1, characterized in that, The preset update trigger condition for the online model update module is: continuous If the health score for a monitoring cycle is lower than a preset threshold, or the moving average of the process status prediction error exceeds a preset upper limit.

Images