Intelligent control system and method based on time sequence filling network

Abstract

The application relates to artificial intelligence control technology and discloses an intelligent control system and method based on a time sequence filling network, which comprises the following steps: collecting and preprocessing a data set; constructing and pre-training a time sequence filling network; constructing a time sequence prediction model based on an encoder-decoder architecture or a pure decoder architecture for time sequence data of a control variable, time sequence data of an intermediate variable and time sequence data of a controlled variable after preprocessing; and pre-training the constructed time sequence prediction model through a mask modeling strategy; and performing online reasoning and control on the model after pre-training to obtain a control variable. The method designed by the application does not require an accurate mechanism model, has high online calculation efficiency, and has unified architecture and is easy to expand.

Description

Technical Field

[0001] This invention relates to artificial intelligence control technology, and more particularly to an intelligent control system and method based on time-filled networks. Background Technology

[0002] Traditional control system design, whether it's classic PID control or modern control theory methods such as model predictive control (MPC) and adaptive control, relies on a precise mathematical model of the controlled object. However, in real-world industrial processes, energy systems, and advanced manufacturing, many controlled objects exhibit complex characteristics such as strong nonlinearity, large time delay, multivariate coupling, and variable operating conditions, making it difficult to establish accurate and concise mechanistic models.

[0003] As exemplified by the existing technology CN202411023061.X, model-based control methods are complex to design, lack robustness, or suffer severe performance degradation when the model mismatches.

[0004] In recent years, data-driven control methods have gained widespread attention. Among them, machine learning techniques, represented by deep neural networks, have provided new approaches to handling complex system control problems due to their powerful nonlinear fitting and temporal feature extraction capabilities. Existing techniques employ recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or temporal convolutional networks (TCNs) to build a forward prediction model from control commands to system output, and then use this model as a basis to combine optimization algorithms to solve for the control variables. However, these methods typically still follow a two-stage "modeling-optimization" framework, and their performance is limited by the accuracy of the forward prediction model. Furthermore, the online optimization process may face computational delays or getting trapped in local optima.

[0005] Another approach is end-to-end reinforcement learning (RL) control, which learns control policies directly through interaction with the environment. However, RL methods typically require massive amounts of interaction data and suffer from drawbacks such as low training efficiency, security risks during the exploration process, and poor policy interpretability, limiting their application in industrial scenarios with high safety requirements.

[0006] Therefore, designing a control method that can make full use of historical operating data, avoid complex online optimization, and possess strong nonlinear processing and temporal reasoning capabilities has become an urgent problem to be solved. Summary of the Invention

[0007] This invention addresses the problems of complex design, insufficient robustness, and severe performance degradation in model-based control methods in the prior art when the model mismatch occurs, and provides an intelligent control system and method based on time-series filling networks.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution.

[0009] Intelligent control methods based on time-filled networks include: Dataset acquisition and preprocessing: real-time or offline acquisition of time series data of control variables, intermediate variables, and controlled variables; and cleaning, alignment, and standardization of the acquired time series data of control variables, intermediate variables, and controlled variables. The construction and pre-training of the time-series-filled Transformer network involves building a time-series prediction model based on a Transformer encoder-decoder architecture or a pure decoder architecture using preprocessed time-series data of control variables, intermediate variables, and controlled variables. The constructed time-series prediction model is then pre-trained using a masking modeling strategy. Online inference and control involves performing online inference and control on the pre-trained model to obtain control variables.

[0010] Preferably, online inference and control are performed on the pre-trained model to obtain the control variables, including: Construct an input sequence, which includes the actual observed values ​​of all variables at the current time, the actual observed values ​​of all variables within the first preset time period, and the processed values ​​of all variables within the second preset time period. Sequence filling involves inputting the constructed input sequence into a pre-trained temporal filling Transformer network. The temporal filling Transformer network uses the known parts of the input sequence as conditions to automatically infer and fill in the values ​​of the mask marker positions to be filled, thereby obtaining the control command sequence within the second preset time period. The system executes in a rolling manner, extracting the first control command from the control command sequence within the second preset time period and issuing it to the actuator; in the next control cycle, it updates the historical context and returns to the constructed input sequence to achieve closed-loop control.

[0011] Preferably, constructing the input sequence includes: Obtain the actual observed values ​​of all variables at the current moment and the actual observed values ​​of all variables within the first preset time period in the past as historical context; All variables within the processed second preset time period are configured such that the control variables within the second preset time period have their values ​​set to the mask marker to be filled; the controlled variables within the second preset time period have their values ​​set to the desired target value or target trajectory; and the intermediate variables within the second preset time period have their values ​​set to the continuation value.

[0012] Preferably, the continuation value of the intermediate variable is the last observation in the historical window, or a short-term forecast provided by the embedded short-term forecast model.

[0013] Preferably, the construction and pre-training of the temporally filled Transformer network includes: Construct a temporally padded Transformer network, which can be built using a Transformer-based encoder-decoder architecture or a pure decoder architecture; The pre-training of the temporally filled Transformer network employs a mask modeling strategy for self-supervised pre-training.

[0014] Preferably, constructing a time-filled Transformer network includes: The input to the time-stuffed Transformer network is a multivariate time series segment [x1, x2, ..., xt] with a fixed time step, where xt is the feature vector of the t-th time series segment. The feature vector of the time series segment is a feature vector concatenated from the time-t data of the control variable, the time-t data of the intermediate variable, and the time-t data of the controlled variable.

[0015] As a preferred approach, self-supervised pre-training of the Transformer network using a masking modeling strategy includes: For a multivariate time series segment [x1, x2, ..., xt] with a fixed time step, at least one continuous time series segment and variable channel are selected, and different variables and different time periods are randomly masked; for unselected time series segments and variable channels, the original values ​​are kept unchanged. Using the unmasked context information in the input sequence as a condition, the masked time series is input into the network to train the Transformer network to predict the original variable values ​​of the masked positions; The mean squared error between the predicted and true values ​​at the masked locations, or the smoothed L1 loss, is used as the loss function to optimize the network parameters.

[0016] Preferably, the time series data of the control variables are historical input data that are directly set or adjusted, including the valve opening, motor speed and heating power of the control system; The time series data of intermediate variables are physical quantities that reflect the internal state of the system, including the internal pressure of the control system's pipes, the liquid level of the intermediate container, and the temperature gradient of the reaction process. The time series data of the controlled variables are the key performance indicators of the final output of the control system, including product concentration, outlet temperature and platform positioning accuracy.

[0017] To address the aforementioned technical problems, this application also provides an intelligent control system based on a time-filling network. The system, implemented using an intelligent control method based on a time-filling network, includes: The dataset acquisition and preprocessing module acquires time series data of control variables, intermediate variables, and controlled variables in real time or offline, and cleans, aligns, and standardizes the acquired time series data of control variables, intermediate variables, and controlled variables. The module for constructing and pre-training the time-series-filled Transformer network constructs a time-series prediction model based on a Transformer encoder-decoder architecture or a pure decoder architecture for the preprocessed time-series data of the control variables, intermediate variables, and controlled variables. The constructed time-series prediction model is then pre-trained using a mask modeling strategy. The online inference and control module performs online inference and control on the pre-trained model to obtain control variables.

[0018] This invention, by adopting the above technical solutions, has significant technical effects: This invention does not require a precise mechanistic model; it learns system dynamics entirely from historical data, avoiding the complex and difficult process of mechanistic modeling. It is particularly suitable for complex black-box or gray-box systems.

[0019] This invention transforms control into filling, naturally incorporating multiple objectives and constraints. Through input masking and filling patterns during the inference phase—masking CVs, extending IVs, and setting PVs objectives—the pursuit of control objectives and the natural evolution of the system's internal states are taken as known conditions, transforming the solution of control variables into a conditional sequence generation problem. When filling CVs, the network spontaneously considers how to satisfy the PVs objective while keeping IVs within a reasonable range, which is essentially equivalent to handling a multi-objective optimization problem with state constraints.

[0020] The Transformer network of this invention, with its self-attention mechanism, can efficiently capture complex nonlinear dynamic interactions and long-term temporal dependencies among multiple variables, and has a strong ability to model nonlinear and temporal relationships for handling large-lag, strongly coupled systems.

[0021] This invention features high online computation efficiency. During the inference phase, only one forward propagation of the neural network is required to obtain the control sequence for a future period of time, making it suitable for scenarios with high real-time requirements.

[0022] This invention qualitatively understands the basis for the network's control decisions by analyzing the distribution of attention weights on different parts of the network when filling CVs, including historical context, current IVs state, and future PVs target, and is interpretable.

[0023] This invention can easily incorporate more process information or constraints by adding new variables to the input sequence. The same pre-trained model can easily cope with different control tasks by changing the target setting of PVs during inference. Its architecture is unified and easy to expand. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention.

[0025] Figure 2 This is a schematic diagram of the randomly generated mask matrix of the present invention.

[0026] Figure 3 This is an illustration of the effect of fading the mask segment of the multivariate original time series according to the present invention.

[0027] Figure 4 This is a schematic diagram of the complete time series after network filling according to the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0029] Example 1

[0030] Intelligent control methods based on time-filled networks include: Dataset acquisition and preprocessing: real-time or offline acquisition of time series data of control variables, intermediate variables, and controlled variables; and cleaning, alignment, and standardization of the acquired time series data of control variables, intermediate variables, and controlled variables. The construction and pre-training of the time-series-filled Transformer network involves building a time-series prediction model based on a Transformer encoder-decoder architecture or a pure decoder architecture using preprocessed time-series data of control variables, intermediate variables, and controlled variables. The constructed time-series prediction model is then pre-trained using a masking modeling strategy. Online inference and control involves performing online inference and control on the pre-trained model to obtain control variables.

[0031] After the pre-trained model is completed, online inference and control are performed to obtain the control variables, including: Construct an input sequence, which includes the actual observed values ​​of all variables at the current time, the actual observed values ​​of all variables within the first preset time period, and the processed values ​​of all variables within the second preset time period. Sequence filling involves inputting the constructed input sequence into a pre-trained temporal filling Transformer network. The temporal filling Transformer network uses the known parts of the input sequence as conditions to automatically infer and fill in the values ​​of the mask marker positions to be filled, thereby obtaining the control command sequence within the second preset time period. The system executes in a rolling manner, extracting the first control command from the control command sequence within the second preset time period and issuing it to the actuator; in the next control cycle, it updates the historical context and returns to the constructed input sequence to achieve closed-loop control.

[0032] Constructing the input sequence includes: Obtain the actual observed values ​​of all variables at the current moment and the actual observed values ​​of all variables within the first preset time period in the past as historical context; All variables within the processed second preset time period are configured such that the control variables within the second preset time period have their values ​​set to the mask marker to be filled; the controlled variables within the second preset time period have their values ​​set to the desired target value or target trajectory; and the intermediate variables within the second preset time period have their values ​​set to the continuation value.

[0033] The continuation value of the intermediate variable is the last observation in the history window, or the short-term forecast provided by the embedded short-term forecast model.

[0034] The construction and pre-training of the temporally filled Transformer network includes: Construct a temporally padded Transformer network, which can be built using a Transformer-based encoder-decoder architecture or a pure decoder architecture; The pre-training of the temporally filled Transformer network employs a mask modeling strategy for self-supervised pre-training.

[0035] Constructing a temporally filled Transformer network includes: The input to the time-stuffed Transformer network is a multivariate time series segment [x1, x2, ..., xt] with a fixed time step, where xt is the feature vector of the t-th time series segment. The feature vector of the time series segment is a feature vector concatenated from the time-t data of the control variable, the time-t data of the intermediate variable, and the time-t data of the controlled variable.

[0036] Self-supervised pre-training of Transformer networks using a masking modeling strategy includes: For a multivariate time series segment [x1, x2, ..., xt] with a fixed time step, at least one continuous time series segment and variable channel are selected, and different variables and different time periods are randomly masked; for unselected time series segments and variable channels, the original values ​​are kept unchanged. Using the unmasked context information in the input sequence as a condition, the masked time series is input into the network to train the Transformer network to predict the original variable values ​​of the masked positions; The mean squared error between the predicted and true values ​​at the masked locations, or the smoothed L1 loss, is used as the loss function to optimize the network parameters.

[0037] The time series data of the control variables are historical input data that were directly set or regulated, including the valve opening, motor speed and heating power of the control system; The time series data of intermediate variables are physical quantities that reflect the internal state of the system, including the internal pressure of the control system's pipes, the liquid level of the intermediate container, and the temperature gradient of the reaction process. The time series data of the controlled variables are the key performance indicators of the final output of the control system, including product concentration, outlet temperature and platform positioning accuracy.

[0038] Example 2

[0039] Based on Example 1, this example is an intelligent control system based on a timing-filling network, which is implemented through an intelligent control method based on a timing-filling network.

[0040] Data acquisition and preprocessing module: used to acquire time series data of control variables, intermediate variables, and controlled variables in real time or offline, and to perform cleaning, alignment and standardization processing.

[0041] Temporal Filling Network Module: Pre-trains the model and stores the pre-trained temporal filling Transformer network model and its parameters.

[0042] The online inference module is responsible for generating an input sequence based on the current observation data, the set target value / trajectory, and predefined input construction rules within each control cycle; calling the time-series filling network model to perform forward computation to obtain the future control command sequence; and outputting the first control command.

[0043] The control execution and interface module converts the control commands generated by the online inference engine into signals acceptable to the actuator and sends them to the actual controlled object.

[0044] Example 3

[0045] Based on Example 1, this example focuses on historical data acquisition and preprocessing, collecting multivariate time series datasets generated by the target control system during long-term operation. The dataset includes at least three types of variables: control variables (CVs), which are input quantities that can be directly set or adjusted by the operator or the underlying controller, such as valve opening, motor speed, heating power, and fan guide vane opening.

[0046] Intermediate variables (IVs) are observable physical quantities that reflect the internal state of a system but are not usually directly used as control targets, such as the internal pressure of a pipeline, the liquid level in an intermediate container, and the temperature gradient of a reaction process. Controlled variables (PVs) are the final output of the system or key performance indicators that need to be stabilized near the set value, such as product concentration, outlet temperature, and platform positioning accuracy. The collected raw data undergoes cleaning and preprocessing, specifically including: Outlier handling employs statistical methods, such as the 3σ criterion or model-based methods, to identify and remove obviously erroneous or unreliable observation data points.

[0047] For missing values, linear interpolation or the mean of nearest neighbors can be used to fill in a small number of random missing values. For large, continuous missing segments, data from that time period can be removed or a more complex interpolation model can be used.

[0048] For time alignment, since different sensors may have different sampling frequencies, all variable sequences need to be uniformly resampled to the same fixed time interval (such as 1 second), and the timestamps need to be strictly aligned.

[0049] Standardization / Normalization: Standardize each variable sequence separately (e.g., Z-score standardization) or normalize it to the [0,1] interval to eliminate the influence of dimensions and accelerate model training convergence.

[0050] Construction and pre-training of temporally filled Transformer networks A time-series prediction model based on a Transformer encoder-decoder architecture or a pure decoder architecture is constructed, which is referred to as a time-series filling network in this invention.

[0051] The network input is a multivariate time series segment of fixed length (e.g., T time steps), represented as X = x1, x2, ..., xt, where each xt is a concatenated feature vector containing all CVs, IVs, and PVs. For example, if there are 5 CVs, 4 IVs, and only 1 PV, then at time t, these variable values ​​are put into a 10-dimensional feature vector to obtain the feature vector xt.

[0052] The network's pre-training task employs a masking modeling strategy: In each training sample, one or more consecutive segments are randomly selected from the sequence, and the values ​​of all variables (including CVs, IVs, and PVs) within these segments are masked (i.e., replaced with zero values). Assume the matrix X formed by the variables in the sequence is... nxm Where n represents the number of variables and m represents the length of the time series, variables 1 and 3 are randomly selected, and the continuous time segments are [t1, ..., t5] and [t10, ..., t25]. Then the mask sets the original values ​​of variables 1 and 3 in the matrix X to 0 at the corresponding time segments, while keeping the values ​​at other positions unchanged.

[0053] The training objective of the network is to accurately predict the original values ​​of all variables at the masked location based on the unmasked context information. The loss function typically uses the mean squared error (MSE) between the predicted and true values ​​at the masked location or a smoothed L1 loss.

[0054] By performing the aforementioned self-supervised pre-training on a large-scale preprocessed dataset, the network can learn and master the complex dynamic coupling relationships, temporal dependencies, and causal relationships among multiple variables in a system without any additional labels. The Transformer's self-attention mechanism enables it to capture long-distance dependencies, which is crucial for understanding industrial processes with large lag characteristics.

[0055] Based on conditional imputation for online inference and control, after pre-training, the model possesses the ability to infer the complete sequence from partial sequence context. In the online control phase, this ability is used for reverse inference to generate control commands. The specific steps are as follows: Construct the current context and future goal: Obtain the actual observed values ​​of all variables at the current time t and within a past period (such as t-T0 to t), as the known "historical context".

[0056] For the future control time domain from t to t+Tc: control variables (CVs): all are set to the [MASK] marker to be filled. This is the object that the network needs to solve.

[0057] Intermediate variables (IVs): Instead of masking, they are filled with the last observation in the history window (the value at time t), or with short-term predictions provided by a simple predictive model (such as Kalman filtering). This is equivalent to providing the network with prior information about the current state of the system and its natural evolution trend, which is key to guiding the generation of reasonable control actions.

[0058] The controlled variables (PVs) are not masked, but are directly set to the desired target value (for setpoint control) or target trajectory (for tracking control). This explicitly communicates the control task to the network.

[0059] Sequence padding and decoupling control variables: The constructed complete sequence, which is partially known (historical context, IVs continuation value, PVs target value) and partially unknown (future CVs are masked), is input into the pre-trained temporal padding network.

[0060] Based on its learned knowledge of system dynamics, the network automatically infers and fills in the values ​​of masked future CVs positions, using known parts as strong conditions.

[0061] In the network output, the CVs sequence for the future time period is the desired future control instruction sequence.

[0062] Rolling Execution and Closed-Loop Feedback: A rolling optimization strategy common in Model Predictive Control (MPC) is employed. At time t, a time-series filling network is invoked to fill in the CVs values ​​for a future time period. After filling, only the CVs value of the first time step (time t+1) in the generated control command sequence is taken and sent to the actual actuator. After the action is executed, new actual observation data (including PVs and IVs) are collected, completing the S1-S2 update of the filled model. At time t+1, the input sequence is reconstructed (historical window sliding update) to perform the next round of conditional filling inference, achieving closed-loop feedback control and model optimization.

Claims

1. An intelligent control method based on time-series filling networks, characterized in that, The methods include: Dataset acquisition and preprocessing: real-time or offline acquisition of time series data of control variables, intermediate variables, and controlled variables; and cleaning, alignment, and standardization of the acquired time series data of control variables, intermediate variables, and controlled variables. The construction and pre-training of the time-series-filled Transformer network involves building a time-series prediction model based on a Transformer encoder-decoder architecture or a pure decoder architecture using preprocessed time-series data of control variables, intermediate variables, and controlled variables. The constructed time-series prediction model is then pre-trained using a masking modeling strategy. Online inference and control involves performing online inference and control on the pre-trained model to obtain control variables.

2. The intelligent control method based on time-series filling networks according to claim 1, characterized in that, After the pre-trained model is completed, online inference and control are performed to obtain the control variables, including: Construct an input sequence, which includes the actual observed values ​​of all variables at the current time, the actual observed values ​​of all variables within the first preset time period, and the processed values ​​of all variables within the second preset time period. Sequence filling involves inputting the constructed input sequence into a pre-trained temporal filling Transformer network. The temporal filling Transformer network uses the known parts of the input sequence as conditions to automatically infer and fill in the values ​​of the mask marker positions to be filled, thereby obtaining the control command sequence within the second preset time period. The system executes in a rolling manner, extracting the first control command from the control command sequence within the second preset time period and issuing it to the actuator; in the next control cycle, it updates the historical context and returns to the constructed input sequence to achieve closed-loop control.

3. The intelligent control method based on time-series filling networks according to claim 2, characterized in that, Constructing the input sequence includes: Obtain the actual observed values ​​of all variables at the current moment and the actual observed values ​​of all variables within the first preset time period in the past as historical context; All variables within the processed second preset time period are configured such that the control variables within the second preset time period have their values ​​set to the mask marker to be filled; the controlled variables within the second preset time period have their values ​​set to the desired target value or target trajectory; and the intermediate variables within the second preset time period have their values ​​set to the continuation value.

4. The intelligent control method based on time-filling networks according to claim 3, characterized in that, The continuation value of the intermediate variable is the last observation in the history window, or the short-term forecast provided by the embedded short-term forecast model.

5. The intelligent control method based on time-series filling networks according to claim 1, characterized in that, The construction and pre-training of the temporally filled Transformer network includes: Construct a temporally padded Transformer network, which can be built using a Transformer-based encoder-decoder architecture or a pure decoder architecture; The pre-training of the temporally filled Transformer network employs a mask modeling strategy for self-supervised pre-training.

6. The intelligent control method based on time-filling networks according to claim 5, characterized in that, Constructing a temporally filled Transformer network includes: The input to the time-stuffed Transformer network is a multivariate time series segment [x1, x2, ..., xt] with a fixed time step, where xt is the feature vector of the t-th time series segment. The feature vector of the t-th time series segment is a feature vector composed of the time-t data of the control variable, the time-t data of the intermediate variable, and the time-t data of the controlled variable.

7. The intelligent control method based on time-filling networks according to claim 5, characterized in that, Self-supervised pre-training of Transformer networks using a masking modeling strategy includes: For a multivariate time series segment [x1, x2, ..., xt] with a fixed time step, at least one consecutive time series segment and variable channel are selected and randomly masked; the original values ​​of the unselected time series segments and variable channels are left unchanged. Using the unmasked context information in the input sequence as a condition, the masked time series is input into the network to train the Transformer network to predict the original variable values ​​of the masked positions; The mean squared error between the predicted and true values ​​at the masked locations, or the smoothed L1 loss, is used as the loss function to optimize the network parameters.

8. The intelligent control method based on time-series filling networks according to claim 1, characterized in that, The time series data of the control variables are historical input data that were directly set or regulated, including the valve opening, motor speed and heating power of the control system; The time series data of intermediate variables are physical quantities that reflect the internal state of the system, including the internal pressure of the control system's pipes, the liquid level of the intermediate container, and the temperature gradient of the reaction process. The time series data of the controlled variables are the key performance indicators of the final output of the control system, including product concentration, outlet temperature and platform positioning accuracy.

9. An intelligent control system based on a time-series filling network, characterized in that, The system implemented by the intelligent control method based on time-filling networks as described in any one of claims 1-8 includes: The dataset acquisition and preprocessing module acquires time series data of control variables, intermediate variables, and controlled variables in real time or offline, and cleans, aligns, and standardizes the acquired time series data of control variables, intermediate variables, and controlled variables. The module for constructing and pre-training the time-series-filled Transformer network constructs a time-series prediction model based on a Transformer encoder-decoder architecture or a pure decoder architecture for the preprocessed time-series data of the control variables, intermediate variables, and controlled variables. The constructed time-series prediction model is then pre-trained using a mask modeling strategy. The online inference and control module performs online inference and control on the pre-trained model to obtain control variables.

Images