Chemical reaction kettle cluster cooperative control method and system and computer equipment

Abstract

The application discloses a chemical reaction kettle cluster cooperative control method and system and a computer device, and relates to the field of high-end chemical flexible manufacturing. The cooperative control method comprises the following steps: obtaining first time sequence data of each reaction kettle based on an edge controller of each reaction kettle in the reaction kettle cluster; synthesizing a first state feature vector of a preset dimension based on the first time sequence data; uploading the first state feature vector to a cloud server by the edge controller, and performing decentralized processing on the first state feature vector based on a mathematical model stored in the cloud server to obtain a global state feature of the reaction kettle cluster; and sending the global state feature to each edge controller by the cloud server, so that the edge controller determines a target action performed by the corresponding reaction kettle according to the global state feature. Through the above method, the communication bandwidth occupied by the transmission data is reduced, the decision-making efficiency of the cloud is improved, and the anti-interference ability and rapid adaptation ability in a dynamic environment are improved.

Description

Technical Field

[0001] This application relates to the field of flexible manufacturing of high-end chemicals, and in particular to a collaborative control method, system, and computer equipment for a cluster of chemical reaction vessels. Background Technology

[0002] In the field of flexible manufacturing of high-end chemicals, the use of miniaturized and modular reactors (such as 10L standard reactors) to form production clusters has become an important development direction.

[0003] Existing systems typically upload raw data (temperature, pressure, etc.) collected at the edge (reactor) directly to the cloud. This "data transporter" model not only consumes a large amount of communication bandwidth, but more importantly, the cloud receives unprocessed, low-level data and cannot directly perceive the real-time process status of the reactor (such as "whether the reaction is in a period of intense exothermic reaction" or "whether the heat transfer efficiency has decreased"). This results in delayed cloud-based decision-making and difficulty in responding to rapidly changing processes.

[0004] The system's reconfiguration logic (such as which reactors are combined to perform a specific process) largely relies on preset rules or fixed models. When faced with unforeseen changes in operating conditions (such as batch differences in raw materials or slight degradation in equipment performance) or when multiple objectives need to be optimized (such as simultaneously achieving the lowest energy consumption, highest yield, and shortest production cycle), the system lacks the ability to learn online and dynamically adjust its reconfiguration strategy, resulting in limited flexibility. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a collaborative control method, system, and computer equipment for a cluster of chemical reactors, which can reduce the communication bandwidth occupied by data transmission, improve the decision-making efficiency in the cloud, and enhance the anti-interference capability and rapid adaptability in dynamic environments.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] In a first aspect, this application provides a method for the coordinated control of a cluster of chemical reactors, which includes the following steps:

[0008] The first time-series data of each reactor is obtained based on the edge controller of each reactor in the reactor cluster. The first time-series data includes the temperature, pressure, stirring speed and / or feed flow rate in the reactor.

[0009] Based on the first time series data, a feature set within the sliding time window is determined, and the feature set is combined into a first state feature vector of a preset dimension. The preset dimension is smaller than the dimension of the first time series data. The feature set includes statistical features, time series features, and frequency domain features.

[0010] The edge controller uploads the first state feature vector to the cloud server and performs decentralized processing on the first state feature vector based on the mathematical model stored on the cloud server to obtain the global state features of the reactor cluster.

[0011] The cloud server sends global state characteristics to each edge controller, enabling the edge controller to determine the target action to be performed by the corresponding reactor based on the global state characteristics.

[0012] In some implementations, a feature set within a sliding time window is determined based on the first time-series data, and the feature set is combined into a first state feature vector of a preset dimension, including:

[0013] The first temporal data is preprocessed based on at least two stacked residual blocks in the temporal neural network to obtain a high-level feature sequence; wherein each residual block contains at least two layers of dilated causal convolution.

[0014] Global average pooling is performed on at least two high-level feature sequences to transform variable-length first-time-series data into a first-state feature vector.

[0015] In some implementations, the first time-series data is preprocessed based on at least two stacked residual blocks in the temporal neural network to obtain a high-level feature sequence, including:

[0016] The first time series data is used as input data into a residual block, and the input data is processed through two layers of dilated causal convolution.

[0017] The output data after two layers of dilated causal convolution is added to the input data to obtain a high-level feature sequence.

[0018] In some implementations, before the cloud server sends global state characteristics to each edge controller, the collaborative control method includes:

[0019] The current production target of the reactor cluster is determined. The first state feature vector and the current production target are used as input data, and the graph neural network model stored on the cloud server compares the similarity between the input data and the historical process map.

[0020] The graph neural network is obtained by aggregating process nodes and process edges in the historical process graph, and / or process nodes and their adjacent process nodes. Process nodes are used to indicate the operation type, temperature range, pressure range and / or duration of the reactor, and process edges are used to indicate the transfer relationships and conditions of the reactor when performing the operation.

[0021] Global state features are determined based on the historical process map with the highest similarity.

[0022] In some implementations, the collaborative control method also includes:

[0023] The second time-series data of the reactor after the target action is executed is obtained based on the edge controller, and the second state feature vector is determined based on the second time-series data;

[0024] The graph neural network is updated based on the second state feature vector and the current production target.

[0025] In some implementations, the first-state feature vector is decentralized based on a mathematical model stored on a cloud server, including:

[0026] An experience replay pool is established based on historical process maps stored on a cloud server, and training data is randomly sampled from the experience replay pool.

[0027] The value decomposition network calculates the global Q-value of the joint action value function based on the training data. It then minimizes the loss function of the global Q-value using a gradient descent algorithm based on temporal difference error, and updates the Q-values ​​of all agents. i Network parameters;

[0028] The intelligent agent is used to instruct the digital twin model corresponding to the reactor stored on the cloud server, and the network parameters are used to instruct the intelligent agent Q. i Weights in the joint action value function.

[0029] In some implementations, the joint action value function Q total (S t A t The following relationship must be satisfied:

[0030] Q total (S t A t )≈∑i=1 N Qi(S t ,a ti ;θi);

[0031] In the formula, S t A represents the first state feature vector received by all agents. t The target action to be performed by all agents, used to update the first state feature vector; N is the number of agents; a ti θi is the target action to be performed by an agent; θi is the i-th agent Q. i The weights of the network parameters in the joint action value function.

[0032] In some implementations, the collaborative control method also includes:

[0033] Determine the network status between the edge controller and the cloud server, and adjust the rule parameters of the edge controller according to the network status. The rule parameters are used to indicate the PID parameters of the PID controller built into the edge controller.

[0034] When the edge controller is communicating with the cloud server, the edge controller adjusts the PID parameters based on the global state characteristics sent by the cloud server.

[0035] When the edge controller is disconnected from the cloud server, the edge controller adjusts the PID parameters based on pre-established priority rules. The priority rules include a first rule and a second rule. When the first time-series data exceeds the critical state, the first rule is triggered first.

[0036] The first rule refers to the edge controller skipping the PID controller to control the reactor and trigger the shutdown mechanism; the second rule refers to the edge controller determining the error and error change rate based on the process variables after the reactor performs the target action, and dynamically adjusting the PID parameters based on the error and error change rate through fuzzy inference.

[0037] Secondly, this application provides a collaborative control system for a chemical reactor cluster, comprising a reactor cluster, several edge controllers, and a cloud server. The reactor cluster consists of at least several reactors. The edge controllers are respectively connected to the corresponding reactors and are used to acquire first time-series data of each reactor. The first time-series data includes temperature, pressure, stirring speed, and / or feed flow rate in the reactor. Based on the first time-series data, a feature set within a sliding time window is determined, and the feature set is combined into a first state feature vector of a preset dimension. The preset dimension is smaller than the dimension of the first time-series data. The feature set includes statistical features, time-series features, and frequency domain features. The cloud server is used to receive the first state feature vector sent by the edge controllers and perform decentralized processing on the first state feature vector based on a mathematical model stored in the cloud to obtain the global state features of the reactor cluster. The edge controllers are also used to determine the target action to be performed by the corresponding reactor based on the global state features and control the execution terminal in the reactor to perform the target action.

[0038] Thirdly, this application provides a computer device including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor executes any of the above-mentioned collaborative control methods for chemical reactor clusters.

[0039] The collaborative control method for a chemical reactor cluster provided in this application process the first time-series data of each reactor acquired by the edge controller. This transforms the large, structurally chaotic, and redundant first time-series data into a simplified, structurally standardized, and clearly defined first state feature vector. This first state feature vector is then uploaded to a cloud server, reducing the communication bandwidth consumed by data transmission and improving the decision-making efficiency of the cloud. By decentralizing the first state feature vector to obtain global state features, the edge controller controls each reactor to perform target actions based on these global state features. This enables the collaborative control system of the chemical reactor cluster to perform online learning and dynamically adjust its reconfiguration strategy under different operating conditions and / or target scenarios, improving its anti-interference capability and rapid adaptability in dynamic environments. Attached Figure Description

[0040] Figure 1 This is a flowchart of the collaborative control method for a cluster of chemical reactors in the embodiments of this application;

[0041] Figure 2 This is a flowchart illustrating the synthesis of the first state feature vector in an embodiment of this application;

[0042] Figure 3 This is a flowchart illustrating the high-level feature sequence obtained in the embodiments of this application;

[0043] Figure 4 This is a schematic diagram of the collaborative control closed loop in the embodiments of this application;

[0044] Figure 5 This is a flowchart illustrating the decentralization process in the embodiments of this application;

[0045] Figure 6 This is a flowchart illustrating the determination of global state features based on similarity in an embodiment of this application;

[0046] Figure 7 This is a schematic diagram of a historical process diagram in an embodiment of this application;

[0047] Figure 8 This is a flowchart illustrating the updating of the graph neural network in an embodiment of this application;

[0048] Figure 9 This is a flowchart illustrating the adjustment of PID parameters based on network status in an embodiment of this application.

[0049] Figure 10 This is a schematic diagram of the collaborative control system of the chemical reactor cluster in the embodiments of this application;

[0050] Figure 11 This is a schematic diagram of a computer device in an embodiment of this application. Detailed Implementation

[0051] To enable those skilled in the art to better understand the present application, the technical solutions in specific embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0052] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. "Comprising" or "including" and similar terms mean that the elements or objects preceding "comprising" or "including" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. "Connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0053] like Figure 1 As shown, this application provides a collaborative control method for a chemical reactor cluster, which is applied to the collaborative control system of the chemical reactor cluster, wherein the chemical reactor cluster is a production cluster composed of miniaturized and modular reactors (such as a 10L standard reactor).

[0054] In some embodiments, the collaborative control method specifically includes the following steps:

[0055] Step S101: Obtain the first time-series data of each reactor based on the edge controller of each reactor in the reactor cluster.

[0056] The first time-series data is used to indicate the temperature, pressure, stirring speed, and / or feed flow rate in the reactor.

[0057] It should be noted that the first time series data of each reactor consists of a data set with a dimension of T*C, where T is the time step and C is the number of channels of the first time series data (such as temperature, pressure, stirring speed and / or feed flow rate).

[0058] Step S102: Determine the feature set within the sliding time window based on the first time series data, and combine the feature set into a first state feature vector of a preset dimension.

[0059] The preset dimension is smaller than the dimension of the first time series data, and the feature set includes statistical features, time series features and frequency domain features.

[0060] A sliding time window refers to data within a continuous time period obtained by sliding a window of fixed length at a specific frequency. The data obtained by a sliding time window is mainly determined by two parameters: the window size and the sliding step size. The window size refers to the time range covered by the window, and the sliding step size refers to the frequency at which the window is generated.

[0061] In some implementations, before determining the feature set within the sliding time window based on the first time-series data, the set of the first time-series data is first passed through a standard causal convolutional layer to update the first time-series data, so that the number of channels of the first time-series data is adjusted from C to D suitable for subsequent calculations, where C is greater than D.

[0062] like Figure 2 As shown, in some implementations, the feature set within the sliding time window is determined based on the first time-series data, and the feature set is combined into a first state feature vector of a preset dimension, including the following steps:

[0063] Step S201: Preprocess the first time series data based on at least two stacked residual blocks in the temporal neural network to obtain a high-level feature sequence.

[0064] Each residual block contains at least two layers of dilated causal convolutions.

[0065] In this implementation, TCN (Temporal Convolutional Network) is used to preprocess the first temporal data. TCN includes causal convolution, dilated convolution, and residual connections.

[0066] The formula for causal convolution is as follows:

[0067]

[0068] in, This is the reading vector of the edge controller at time tk. is the weight matrix of the convolution kernel at position k, which is used to extract local temporal patterns. This is a bias term. The size of the convolution kernel determines the length of the window for viewing historical data at one time. It is a non-linear activation function. The feature map is obtained after convolution at time t.

[0069] The formula for calculating the receptive field after dilated convolution is as follows:

[0070]

[0071] in, The receptive field represents the temporal span of input data that an output neuron can capture. The larger the value, the stronger the model's ability to capture long-range dependencies (such as trends throughout the reaction phase). For the first The expansion factor of the layer, which grows exponentially, in Under certain conditions, expansion can increase the receptive field. The number of residual blocks.

[0072] In some possible implementations, after each layer of dilated causal convolution, the results need to be weighted, ReLU, and Dropout in sequence to improve the efficiency and accuracy of the model.

[0073] Weight normalization, by decomposing the weights of convolutional layers into amplitude and direction components, eliminates the influence of weight scale on gradients, accelerating model training convergence. The ReLU activation function maps the linear convolution output to a non-linear space, enhancing the model's ability to capture complex temporal patterns. Dropout randomly discards some neurons to prevent the model from overfitting to the training data, improving the model's generalization ability.

[0074] Residual linking is the core step in forming the final output of the residual block, and its formula is as follows:

[0075]

[0076] in, This is the final output. This is the activation function. This is the input for the residual block. This is the output after two layers of dilated causal convolution transformation. Residual connections can avoid gradient vanishing, enabling lightweight deep models running on edge devices to be trained effectively and improving the stability of the model learning process.

[0077] like Figure 3 As shown, in some implementations, the first time-series data is preprocessed based on at least two stacked residual blocks in the temporal neural network to obtain a high-level feature sequence, including the following steps:

[0078] Step S301: Input the first time series data as input data into a residual block, and then pass the input data through two layers of dilated causal convolution.

[0079] For example, the first Inflation factor of each residual block Set as For example, 1, 2, 4, 8, and so on, to achieve exponential growth of the receptive field.

[0080] In some implementations, temporal causality is preserved through left padding in causal convolution computation. For example, regarding the width of the convolution kernel... and expansion factor pad the left side of the input sequence. Adding zeros ensures that the two sequences being multiplied have the same dimension, guaranteeing that the output value at time t depends only on the inputs at time t and before time t, conforming to the causal law of time series data. Causality refers to the principle that in time series data, the causal relationship between variables follows a specific temporal order and logical necessity; that is, the causal event occurs before the outcome event in time and directly influences the change in the outcome through some mechanism.

[0081] In other implementations, zeros are padded on both sides during causal convolution computation so that the two sequences being multiplied have the same dimension.

[0082] In some implementations, dilated convolution calculations are performed by inserting [elements] between the elements of the convolution kernel. A zero is used to reduce the receptive field without increasing the number of parameters. Upgraded to .

[0083] It should be noted that causal convolution calculations and dilated convolution calculations can be performed in parallel, which improves computational efficiency.

[0084] Step S302: Add the output data after two layers of dilated causal convolution to the input data to obtain the high-level feature sequence.

[0085] Advanced feature sequences can characterize the temperature, pressure, stirring speed, and / or feed flow rate in each reactor.

[0086] By following the steps above, we can ensure the matching of the dimensions of the two sequences being multiplied during causal convolution calculations and expand the coverage length of the receptive field during dilated convolution calculations. This improves computational efficiency and enables efficient long-range dependency modeling.

[0087] Step S202: Perform global average pooling on at least two high-level feature sequences to convert the variable-length first time series data into a first state feature vector.

[0088] The formula for global average pooling is as follows:

[0089]

[0090] in, It is a vector that contains global information about the entire input sequence. For time series dimension. To extract all data at time t.

[0091] After global average pooling Dimension mapping and feature fine-tuning are performed through one or more fully connected layers to output a first-state feature vector of a preset dimension.

[0092] In this fully connected layer, each neuron is fully connected to all neurons in the previous layer, and feature integration and decision mapping are achieved through linear transformation and nonlinear activation functions.

[0093] Through the above steps, efficient extraction and transformation of time-series data features at the edge are achieved, reducing the communication bandwidth occupied by the data and improving the subsequent data transmission rate.

[0094] Step S103: The edge controller uploads the first state feature vector to the cloud server and performs decentralized processing on the first state feature vector based on the mathematical model stored on the cloud server to obtain the global state features of the reactor cluster.

[0095] In some implementations, the problem is modeled as a Markov game, and the following concepts are abstracted:

[0096] state space : The first state feature vector of each reactor Together constitute This represents the overall operating condition of the entire reactor cluster at time t.

[0097] Action space It is composed of the target actions performed by each reaction vessel, for example... ={Temperature setpoint + 1℃, Temperature setpoint 1℃, stirring speed +10rpm,...} It represents the set of all control commands issued to the entire reactor cluster at time t.

[0098] State transition probability : Represents in Execute Afterwards, the environment was changed by Transfer to The probability of.

[0099] Total reward function The total reward function consists of a local reward *r* and a global reward *R*. The local reward *r* represents the efficiency of an individual reactor in completing its own production task, while the global reward *R* represents the efficiency of the entire reactor cluster in completing its respective production tasks. The formula is as follows:

[0100]

[0101] Through the total reward function It can achieve a balance between the efficient completion of production tasks by individuals and the overall system within a reactor cluster.

[0102] like Figure 4 The diagram shown illustrates the implementation of a collaborative control loop using the Markov game described above. First, the state space is generated. Each reactor selects the target action to be performed and generates the action space. The digital twin calculates the total reward function. and the new state space and the state space Updated to The target actions performed by each reactor are updated to optimize the production target of the reactor cluster.

[0103] In this context, a digital twin is a high-fidelity digital image of a physical entity or process in virtual space. In the implementation of this application, the digital twin is a digital image of a reactor cluster and the produced products.

[0104] like Figure 5 As shown, in some implementations, the first state feature vector is decentralized based on a mathematical model stored on a cloud server, including the following steps:

[0105] Step S501: Establish an experience replay pool based on the historical process map stored on the cloud server, and randomly sample training data from the experience replay pool.

[0106] The data in the experience replay pool includes , , as well as By establishing an experience replay pool, the temporal correlation between data can be broken.

[0107] The randomly sampled training data is mini-batch data, which means that the dataset in the experience replay pool is divided into several small subsets, and the training is performed on several small subsets in batches to further destroy the correlation between data and improve the stability of training.

[0108] Step S502: Based on the value decomposition network, calculate the global Q-value of the joint action value function according to the training data, minimize the loss function of the global Q-value using the gradient descent algorithm through temporal difference error, and update the Q-values ​​of all agents. i Network parameters.

[0109] The intelligent agent is used to instruct the digital twin model corresponding to the reactor stored on the cloud server, and the network parameters are used to instruct the intelligent agent Q. iWeights in the joint action value function. Temporal difference error represents the difference between the predicted value at the current time step and the corrected predicted value at the next time step. The gradient descent algorithm minimizes the loss by calculating the gradient of the loss function with respect to the model parameters and updating the parameters in the opposite direction of the error.

[0110] Specifically, the value function of joint actions The following relationship must be satisfied:

[0111] ;

[0112] In the formula, This is the first state feature vector received by all agents. The target action to be performed by all agents is used to update the first state feature vector. N is the number of agents. The target action to be performed by an intelligent agent. For the i-th intelligent agent The weights of the network parameters in the joint action value function.

[0113] In some implementations, a periodically updated target network is utilized. Calculate a more stable target value. The following relationship must be satisfied:

[0114]

[0115] Where y represents the current time interval. A further estimate of y includes the total reward function at the current time. and the new state space The discount value for the best return. As a discount factor, it measures current and future returns in order to select the individual actions of each reactor that are most beneficial to the overall situation.

[0116] In some implementations, the loss function is calculated using the mean squared error loss, and its formula is as follows:

[0117]

[0118] Here, L is a quantitative indicator of the model prediction error, and the model parameters can be optimized by minimizing L. B is the size of the mini-batch, such as the selected data dimension.

[0119] Through the above steps, the intelligent agent Q is continuously adjusted. i By optimizing the network parameters, the robustness of the model is improved, enabling the reactor cluster to achieve not only local optimization of a single reactor but also global optimization among the reactors.

[0120] Step S104: The cloud server sends the global state characteristics to each edge controller, so that the edge controller can determine the target action to be performed by the corresponding reactor based on the global state characteristics.

[0121] The collaborative control method provided in this application processes the first-time-series data of each reactor acquired by the edge controller. This transforms the large-scale, structurally chaotic, and informationally redundant first-time-series data into a simplified, structurally standardized, and clearly defined first-state feature vector. This first-state feature vector is then uploaded to a cloud server, reducing the communication bandwidth consumed by data transmission and improving the decision-making efficiency in the cloud. By decentralizing the first-state feature vector to obtain global state features, the edge controller controls each reactor to perform target actions based on these global state features. This enables the collaborative control system of the chemical reactor cluster to perform online learning and dynamically adjust its reconfiguration strategy under different operating conditions and / or target conditions, improving its anti-interference capability and rapid adaptability in dynamic environments.

[0122] like Figure 6 As shown, in some implementations, before the cloud server sends the global state characteristics to each edge controller, the collaborative control method also includes the following steps:

[0123] Step S601: Determine the current production target of the reactor cluster, use the first state feature vector and the current production target as input data, and make the graph neural network model stored on the cloud server compare the similarity between the input data and the historical process map.

[0124] The graph neural network is obtained by aggregating process nodes and process edges from the historical process graph, and / or process nodes and their adjacent process nodes. Process nodes are used to indicate the operation type, operation parameters, material information, equipment requirements, etc. of the reactor, while process edges are used to indicate the transfer relationships and conditions of the reactor when performing operations.

[0125] For example, the operation types indicated by the process node include, but are not limited to, heating, stirring reaction, cooling, crystallization, and filtration; the operation parameters include, but are not limited to, temperature range, pressure range, and / or duration; the material information includes, but is not limited to, the types of input and output materials and their approximate proportions; and the equipment requirements include, but are not limited to, the required reactor type and stirring method. The transfer relationships and conditions indicated by the process node include, but are not limited to, the requirements for temperature and pH value during material transfer, the direction and method of material transfer from the previous operation to the next operation, and the time requirements for the transfer process.

[0126] like Figure 7The diagram shown is a schematic of a historical process flow. Each process node represents feeding A, heating to 80°C, cooling to 25°C, product drying, and the final product. Each process edge represents direct transfer, heat preservation for 2 hours, and filtration, respectively.

[0127] In some implementations, cosine similarity comparison is used to compare the similarity between the input data and the historical process map. This is achieved by measuring the degree of directional similarity between the input data and the historical process map, and by calculating the cosine value of the angle between the vectors to quantify the similarity between the input data and the historical process map.

[0128] Step S602: Determine the global state features based on the historical process map with the highest similarity.

[0129] Through the above steps, the complex process relationships are abstracted into a computable graph structure, and the global state characteristics are determined based on the historical process graph with the highest similarity, so as to determine the optimal target action to be performed by the corresponding reactor, thereby further realizing the intelligent control of the reactor.

[0130] like Figure 8 As shown, in some implementations, the cooperative control method further includes the following steps:

[0131] Step S801: Obtain the second time-series data of the reactor after the target action is executed based on the edge controller, and determine the second state feature vector based on the second time-series data.

[0132] Step S802: Update the graph neural network based on the second state feature vector and the current production target.

[0133] Through the above steps, the graph neural network is continuously optimized, thereby optimizing the production process and completing the closed loop of production, feedback, and learning.

[0134] like Figure 9 As shown, the cooperative control method also includes the following steps:

[0135] Step S901: Determine whether the network status between the edge controller and the cloud server is in a communication state. If yes, proceed to step S902; if no, proceed to step S903.

[0136] Adjust the rule parameters of the edge controller according to the network status.

[0137] The rule parameters are used to indicate the PID parameters of the PID controller built into the edge controller. The PID parameters include the proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd. The proportional coefficient Kp determines the strength of the response to the current error. The larger Kp is, the faster the system response. The integral coefficient Ki eliminates static error by accumulating the integral of historical errors. The derivative coefficient Kd can predict the future trend of error changes based on the error change rate, thus suppressing overshoot.

[0138] Step S902: The edge controller adjusts the PID parameters based on the global state characteristics sent by the cloud server.

[0139] Step S903: The edge controller adjusts the PID parameters based on pre-established priority rules.

[0140] The priority rules include a first rule and a second rule. When the first time series data exceeds the critical state, the first rule is triggered first.

[0141] The first rule refers to the edge controller skipping the PID controller and triggering the shutdown mechanism of the reactor. The second rule refers to the edge controller determining the error and error rate of change based on the process variables after the reactor performs the target action, and dynamically adjusting the PID parameters based on the error and error rate of change through fuzzy inference.

[0142] It should be noted that the critical state refers to a pre-defined state that indicates the reactor can no longer continue to operate, which includes, but is not limited to, limitations on temperature, pressure, and pressure change rate.

[0143] For example, the first rule is to trigger an emergency shutdown sequence when the temperature inside the reactor is greater than or equal to a preset temperature, and / or the pressure inside the reactor is greater than or equal to a preset pressure, and / or the rate of pressure change inside the reactor is greater than or equal to a preset rate of change. The emergency shutdown sequence is a predefined and safe set of operations, such as turning off the heating power, opening all cooling valves, activating the pressure relief device, and stopping the feed.

[0144] The second rule is that when the reactor is in a constant temperature stage and the standard deviation of the temperature inside the reactor is less than 0.1℃ in the past two minutes, the edge controller adaptively and dynamically adjusts the PID parameters to achieve precise control of the parameters inside the reactor.

[0145] Through the above steps, when the edge controller is in communication with the cloud server, the edge controller is controlled by the cloud server to adjust the PID parameters. When the edge controller is disconnected from the cloud server, the edge controller adjusts the PID parameters autonomously according to the state and priority rules inside the reactor, avoiding safety hazards when the edge controller is disconnected from the cloud server and improving the accuracy and stability of the PID controller.

[0146] like Figure 10 As shown, this application also provides a collaborative control system 100 for a chemical reactor cluster, which includes a reactor cluster 11, several edge controllers 12, and a cloud server 13.

[0147] The reactor cluster 11 consists of at least a number of reactors 111. Several edge controllers 12 are connected to corresponding reactors 111. The edge controllers 12 acquire first time-series data from each reactor 111, including temperature, pressure, stirring speed, and / or feed flow rate. The edge controllers 12 can also determine a feature set within a sliding time window based on the first time-series data and combine the feature set into a first state feature vector of a preset dimension. The preset dimension is smaller than the dimension of the first time-series data. The feature set includes statistical features, time-series features, and frequency domain features. A cloud server 13 receives the first state feature vector sent by the edge controllers 12 and performs decentralized processing on the first state feature vector based on a mathematical model stored in the cloud to obtain the global state features of the reactor cluster 11.

[0148] The edge controller 12 is also used to determine the target action to be performed by the corresponding reactor 111 based on the global state characteristics, and to control the execution terminal in the reactor 111 to perform the target action.

[0149] The edge controller 12 also includes a PID controller 121, which can adjust its own PID parameters to change parameters (such as temperature, pressure, etc.) within the reactor 111.

[0150] Through the above settings, the large-scale, structurally chaotic, and informationally redundant first-time-series data is transformed into a simplified, structurally standardized, and clearly defined first-state feature vector. This first-state feature vector is then uploaded to the cloud server 13, reducing the communication bandwidth occupied by data transmission and improving the decision-making efficiency in the cloud. The edge controller 12 can control each reactor 111 to perform target actions based on global state characteristics. This enables the collaborative control system 100 of the chemical reactor cluster to perform online learning and dynamic adjustment and reconstruction strategies under different operating conditions and / or targets, improving the anti-interference capability and rapid adaptability of the collaborative control system 100 in dynamic environments.

[0151] like Figure 11 As shown, this application also provides a computer device 200, which includes a memory 21 and a processor 22. The memory 21 stores a computer program. When the computer program is executed by the processor 22, the processor 22 executes the above-mentioned collaborative control method for the chemical reactor cluster.

[0152] Specifically, processor 22 may include a central processing unit, or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0153] In some implementations, memory 21 may include a large-capacity memory for data or instructions.

[0154] For example, memory 21 includes a hard disk drive (HDD), a floppy disk drive, flash memory, optical disk, magneto-optical disk, Universal Serial Bus (USB) drive, or any combination of the above-mentioned memory.

[0155] For example, memory 21 may be located inside or outside the computer device 200.

[0156] In some possible implementations, the computer device 200 also includes a communication interface 23 and a bus 24. The processor 22, memory 21, and communication interface 23 are connected via the bus 24 and communicate with each other.

[0157] The communication interface 23 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0158] Bus 24 includes hardware and / or software that couples components of computer device 200 together. For example, and not to limit, bus 24 may include an accelerated graphics port or other graphics bus, an enhanced industry standard architecture bus, a front-side bus, a low pin count bus, a memory bus, or other suitable bus 24 or any combination of the above buses.

[0159] This application also provides a computer-readable storage medium (not shown) storing a computer program that, when executed by processor 22, implements the steps of the above-described collaborative control method for the chemical reactor cluster.

[0160] Computer-readable storage media include, but are not limited to, electronic, magnetic, optical, infrared, or other physical storage devices or apparatuses that may contain or store information such as executable instructions, data, etc. More specific examples of computer-readable storage media include electrical connections based on one or more wires, RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), SSDs (Solid State Disks), any type of storage disk (such as optical discs), or similar memory, or any suitable combination of the foregoing.

[0161] It should be noted that, in order to avoid repetition and improve the conciseness of the manual, the descriptions of the same or substantially similar technical features, structures or method steps in different implementations may be simplified or omitted.

[0162] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for coordinated control of a cluster of chemical reaction vessels, characterized in that, The collaborative control method includes: The first time-series data of each reactor is obtained based on the edge controller of each reactor in the reactor cluster. The first time-series data includes the temperature, pressure, stirring speed and / or feed flow rate of the reactor. Based on the first time series data, a feature set within a sliding time window is determined, and the feature set is combined into a first state feature vector of a preset dimension; the preset dimension is smaller than the dimension of the first time series data, and the feature set includes statistical features, time series features, and frequency domain features. The edge controller uploads the first state feature vector to the cloud server and performs decentralized processing on the first state feature vector based on the mathematical model stored on the cloud server to obtain the global state features of the reactor cluster. The cloud server sends the global state features to each edge controller, so that the edge controller determines the target action to be performed by the corresponding reactor based on the global state features. Specifically, determining a feature set within a sliding time window based on the first time-series data, and combining the feature set into a first state feature vector of a preset dimension, includes: The first time-series data is preprocessed based on at least two stacked residual blocks in the temporal neural network; the first time-series data is used as the input data of the residual blocks, and the input data is passed through two layers of dilated causal convolution; the output data after the two layers of dilated causal convolution is added to the input data to obtain a high-level feature sequence; global average pooling is performed on at least two of the high-level feature sequences to convert the variable-length first time-series data into a first state feature vector of a preset dimension.

2. The collaborative control method according to claim 1, characterized in that, Each residual block contains at least two layers of dilated causal convolutions.

3. The collaborative control method according to claim 1, characterized in that, Before the cloud server sends the global state features to each of the edge controllers, the collaborative control method includes: The current production target of the reactor cluster is determined, and the first state feature vector and the current production target are used as input data. The graph neural network model stored on the cloud server performs a similarity comparison between the input data and the historical process map. The graph neural network is obtained by aggregating process nodes and process edges in the historical process graph, and / or process nodes and their adjacent process nodes. The process nodes are used to indicate the operation type, temperature range, pressure range and / or duration of the reactor, and the process edges are used to indicate the transfer relationships and conditions of the reactor when performing the operation. The global state features are determined based on the historical process map with the highest similarity.

4. The collaborative control method according to claim 3, characterized in that, The collaborative control method further includes: The edge controller acquires the second time-series data of the reactor after the target action is executed, and determines the second state feature vector based on the second time-series data; The graph neural network is updated based on the second state feature vector and the current production target.

5. The collaborative control method according to claim 1, characterized in that, The process of decentralizing the first state feature vector based on the mathematical model stored on the cloud server includes: An experience replay pool is established based on the historical process maps stored on the cloud server, and training data is randomly sampled from the experience replay pool; The value decomposition network calculates the global Q-value of the joint action value function based on the training data, minimizes the loss function of the global Q-value using a gradient descent algorithm through temporal difference error, and updates all agents. Network parameters; The intelligent agent is used to instruct the digital twin model corresponding to the reactor stored in the cloud server, and the network parameters are used to instruct the intelligent agent. The weights in the value function of the joint action.

6. The cooperative control method according to claim 5, characterized in that, The value function of the joint action The following relationship must be satisfied: ； In the formula, The first state feature vector received by all the said agents; The target action performed by all the agents is used to update the first state feature vector; N is the number of the agents; The target action performed by one of the intelligent agents; For the i-th intelligent agent The weights of the network parameters in the joint action value function.

7. The cooperative control method according to claim 1, characterized in that, The collaborative control method further includes: The network status between the edge controller and the cloud server is determined, and the rule parameters of the edge controller are adjusted according to the network status. The rule parameters are used to indicate the PID parameters of the PID controller built into the edge controller. When the edge controller is in communication with the cloud server, the edge controller adjusts the PID parameters based on the global state characteristics sent by the cloud server. When the edge controller is disconnected from the cloud server, the edge controller adjusts the PID parameters based on pre-established priority rules. The priority rules include a first rule and a second rule. When the first time-series data exceeds a critical state, the first rule is triggered first. Wherein, the first rule refers to the edge controller skipping the PID controller to control the reactor to trigger the shutdown mechanism; the second rule refers to the edge controller determining the error and error change rate based on the process variables after the reactor performs the target action, and dynamically adjusting the PID parameters based on the error and the error change rate through fuzzy inference.

8. A collaborative control system for a cluster of chemical reaction vessels, characterized in that, include: A cluster of reactors, which consists of at least several reactors; A plurality of edge controllers, each connected to a corresponding reactor, are used to acquire first time-series data of each reactor, the first time-series data including temperature, pressure, stirring speed and / or feed flow rate in the reactor; determine a feature set within a sliding time window based on the first time-series data, and combine the feature set into a first state feature vector of a preset dimension, the preset dimension being smaller than the dimension of the first time-series data, the feature set including statistical features, temporal features and frequency domain features; wherein, determining the feature set within a sliding time window based on the first time-series data and combining the feature set into a first state feature vector of a preset dimension includes: preprocessing the first time-series data based on at least two stacked residual blocks in a temporal neural network; using the first time-series data as input data for the residual blocks, and passing the input data through two layers of dilated causal convolution; adding the output data after the two layers of dilated causal convolution to the input data to obtain a high-level feature sequence; and performing global average pooling on at least two of the high-level feature sequences to convert the variable-length first time-series data into a first state feature vector of a preset dimension; A cloud server is used to receive the first state feature vector sent by the edge controller, and to perform decentralized processing on the first state feature vector based on the mathematical model stored in the cloud to obtain the global state features of the reactor cluster. The edge controller is further configured to determine the target action to be performed by the corresponding reactor based on the global state characteristics, and control the execution terminal within the reactor to perform the target action.

9. A computer device, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the collaborative control method for the chemical reactor cluster according to any one of claims 1-7.

Images