Aerogel production flow control method and system based on online monitoring

Abstract

The invention provides an aerogel production flow control method and system based on online monitoring, and relates to the technical field of intelligent manufacturing and industrial automation, and the method comprises the steps: inputting state parameters into a flow control AI model; the AI model executes the following steps: based on the current and historical state parameter sequences, determining a first flow control instruction required for enabling the reaction process to return to a preset optimization path; predicting a state parameter trajectory based on the first flow control instruction; judging whether the state parameter track can be effectively stabilized in a tolerance range of a preset optimization path or not; if the judgment result is yes, the conveying flow of the materials is automatically adjusted; if the judgment result is no, triggering to generate a man-machine cooperation task; and the user is assisted in executing the man-machine cooperation task, a second flow control instruction is generated, and the conveying flow of the corresponding materials is adjusted. Key state parameters in aerogel production are collected in real time through an online sensor system, and dynamic prediction and accurate regulation and control of the reaction process are achieved in combination with a flow control AI model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial automation, and particularly relates to an aerogel production flow control method and system based on online monitoring. BACKGROUND

[0002] As a high-performance material, aerogels have wide applications in aerospace, energy storage and environmental protection due to their low density, high porosity and excellent thermal insulation performance. However, the sol-gel reaction process in aerogel production is complex, involving multivariate coupling and nonlinear dynamics, and traditional flow control methods are difficult to achieve real-time and accurate regulation. Existing technologies mainly rely on manual experience or control systems based on fixed rules, lacking dynamic prediction and optimization capabilities for state parameters (such as temperature, pH, viscosity, etc.) in the reaction process. This leads to low production efficiency, unstable product quality, and lack of effective automatic adjustment mechanisms or human-machine collaborative intervention means when the reaction deviates from the optimal path. In addition, traditional methods cannot fully utilize the correlation between historical data and real-time data when processing time series data, making it difficult to respond to rapid changes in process parameters.

[0003] To solve the above problems, there is an urgent need for a flow control method based on online monitoring and artificial intelligence, which dynamically predicts and regulates material flow by real-time acquisition of reaction state parameters, combined with advanced deep learning models, while introducing a human-machine collaborative mechanism to improve the stability and intelligence level of the production process. SUMMARY

[0004] One of the purposes of the present application is to provide an aerogel production flow control method and system based on online monitoring to solve the problems identified in the background.

[0005] In a first aspect, the present application provides an aerogel production flow control method based on online monitoring, comprising: S1: Real-time acquisition of state parameters in the aerogel sol-gel reaction process through an online sensor system; S2: Inputting the state parameters into a pre-trained flow control AI model; S3: The flow control AI model executes the following steps: Based on the current and historical state parameter sequence, determine the first flow control instruction required to return the reaction process to the pre-set optimal path; Based on the first flow control instruction, predict the state parameter trajectory in the future period of time; Determine whether the state parameter trajectory can be effectively stabilized within the tolerance range of the pre-set optimal path; S4: If the result of step S3 is yes, automatically adjust the delivery flow of at least one of the precursors, catalysts or solvents according to the first flow control instruction; S5: If the result of step S3 is no, triggering a human-machine collaborative task to be generated; S6: The auxiliary user executes the human-machine collaborative task to generate a second flow control instruction, and adjusts the conveying flow of the corresponding material according to the second flow control instruction.

[0006] Optionally, the state parameters include at least three parameter combinations of a reaction kettle temperature, a system pH value, a fluid viscosity, a precursor concentration and a gel time.

[0007] Optionally, the pre-training step of the flow control AI model includes: collecting time-series state parameters and corresponding flow control instructions recorded in a historical production process to constitute a sample set, and performing normalization and sliding window segmentation processing on the sample set; based on the processed sample set, constructing a multi-step prediction model with a long short-term memory network as the core, taking a state parameter sequence of a previous period as the input of the multi-step prediction model, and taking a deviation of a state parameter trajectory of a subsequent period and a preset optimization path as a training target; in the training process of the multi-step prediction model, using a time-series convolution network to extract local features of the input sequence, and focusing on key process stages through an attention mechanism; adopting mean square error and trajectory smoothness as a joint loss function, and introducing a trajectory stability evaluation index based on a tolerance range as a model verification standard; adopting a gradient descent algorithm to optimize network parameters of the multi-step prediction model until the multi-step prediction model has the ability to determine a first flow control instruction based on a current and historical state parameter sequence, predict a future state parameter trajectory, and judge whether the state parameter trajectory can be effectively stabilized within a tolerance range of a preset optimization path, and obtain a flow control AI model.

[0008] Optionally, the triggering of the human-machine collaborative task includes: when the flow control AI model judges that the state parameter trajectory cannot be stabilized within the tolerance range of the preset optimization path, generating a collaborative task triggering signal containing a deviation parameter and an out-of-control time window; sending the triggering signal to a human-machine interaction system to activate a collaborative task generation engine therein; based on the deviation parameter and the out-of-control time window, constructing a human-machine collaborative task containing a material type to be adjusted, a target flow range and an adjustment emergency level through the collaborative task generation engine.

[0009] Optionally, the auxiliary user executing the human-machine collaborative task includes: Based on the defined material type to be adjusted, target flow range, and adjustment urgency level in the human-machine collaborative task, a visual decision-making model is constructed, which integrates a real-time state visualization window of the reaction process, a historical optimization path comparison view, and a multi-material flow collaborative regulation interface. Whenever a user inputs a preliminary flow control sub-instruction based on the visual decision-making model, the following operations are performed in parallel: Real-time tracking of the user's hesitation state for the preliminary flow control sub-instruction; Synchronously evaluating the joint risk state if the preliminary flow control sub-instruction is actually executed in real time; wherein the evaluation method of the joint risk state is: extracting the currently input preliminary flow control sub-instruction and the historical preliminary sub-instruction stored in the to-be-implemented instruction set, constructing an instruction sequence; inputting the instruction sequence into the risk prediction network trained by reaction kinetics data to calculate its comprehensive influence index on the aerogel sol-gel reaction process; the comprehensive influence index includes: reaction path deviation expected value, gel time standard deviation increase ratio, product porosity deviation degree; based on the comprehensive influence index, a joint risk value is generated by weighted fusion; When the joint risk value exceeds the preset risk threshold, according to the remaining time of the defined out-of-control time window from the current time point in the human-machine collaborative task, the decision progress reflected by the user's input instructions, and the size of the joint risk value, a decision reversible time window is calculated by a dynamic window planning algorithm; wherein the decision reversible time window is the maximum allowed period for the user to revoke the input preliminary sub-instruction and the system to have enough time to recover to the pre-decision state; Generating a hesitation state trigger constraint condition for the decision reversible time window; When entering the decision reversible time window, real-time detection is performed on whether the user's current hesitation state meets the trigger constraint condition; If it meets, based on the remaining length of the decision reversible time window, the user's operation trajectory and cognitive focus sequence in the visual decision-making model within the previously set memory time window, the cognitive information fragments received by the user and their decision logic gaps are inferred by a cognitive content reconstruction engine; Fusing the remaining time and cognitive gap information, an adaptive auxiliary cognitive strategy is output by a reinforcement learning strategy selection model; Executing the auxiliary cognitive strategy to enhance the user's awareness of the joint risk state; If the user does not perform a revocation operation on the input preliminary flow control sub-instruction after executing the auxiliary cognitive strategy, the instruction is added to the to-be-implemented instruction set; When the user confirms that the decision is complete, a second flow control instruction is generated based on all preliminary flow control sub-instructions in the to-be-implemented instruction set, and is sent to the execution mechanism to adjust the corresponding material delivery flow.

[0010] Optionally, the visual decision model assists users in forming multi-dimensional decision-making cognition by dynamically rendering deviation areas of the reaction parameter evolution curve from the preset optimization path and displaying coupling influence coefficients of different material flow adjustments.

[0011] Optionally, the hesitation state is determined based on fusion of a first behavior indicator before user input and a second behavior indicator after the input; the first behavior indicator includes: frequency and duration of viewing a relevant parameter panel in the visual decision model, historical instruction modification times, and deviation degree of the current input instruction from the system recommended instruction; the second behavior indicator includes: number of re-views of instructions related parameters after inputting the instruction, trigger delay of instruction modification operation, and mouse focus dwell jitter characteristics in the instruction input area; a long short-term memory network-based hesitation evaluation model is used to perform time series modeling on the first behavior indicator and the second behavior indicator, and output a current hesitation state quantitative value.

[0012] Optionally, the trigger constraint condition is used to determine whether to assist users in risk cognition in the current window, and is defined as: the user's hesitation state quantitative value is continuously higher than the hesitation threshold value in the latest continuous period, and the hesitation fluctuation variance is lower than the stability threshold value.

[0013] Optionally, the auxiliary cognition strategy includes: highlighting the risk parameter change trend, pushing a historical similar decision case comparison view, and automatically simulating execution of the instruction sequence and visualizing the predicted results.

[0014] In a second aspect, the online monitoring-based aerogel production flow control system provided by the embodiments of the present application includes: An online monitoring module is configured to collect state parameters in an aerogel sol-gel reaction process in real time through an online sensor system. A parameter input module is configured to input the state parameters to a pre-trained flow control AI model. An artificial intelligence module is configured to enable the flow control AI model to perform the following steps: Determine a first flow control instruction required to make the reaction process return to a preset optimization path based on current and historical state parameter sequences. Predict a state parameter trajectory in a future period of time based on the first flow control instruction. Determine whether the state parameter trajectory can be effectively stabilized within a tolerance range of the preset optimization path. A first flow control module is configured to automatically adjust a delivery flow of at least one of a precursor, a catalyst, or a solvent according to the first flow control instruction if the determination result is yes. A man-machine collaboration module is configured to trigger generation of a man-machine collaboration task if the determination result is no. The second flow control module is used for assisting the user in performing a man-machine collaborative task, generating a second flow control instruction, and adjusting the conveying flow of the corresponding material according to the second flow control instruction.

[0015] The present application has the following beneficial effects: The technical solution of the present application realizes dynamic prediction and precise regulation of the reaction process by collecting key state parameters in aerogel production in real time through an online sensor system and combining a pre-trained flow control AI model, significantly improving production efficiency and product quality stability compared with traditional methods. The model accurately captures the time sequence characteristics and key stage information of the process parameters through a long short-term memory network, a time sequence convolution network and an attention mechanism, generates optimized flow control instructions, and ensures that the reaction trajectory is stably within the tolerance range of the preset optimization path. In addition, when the predicted trajectory deviates from the optimization path, the system generates an adjustment task containing deviation parameters and emergency levels through a man-machine collaborative task generation mechanism, assists the user in efficient intervention, and reduces the complexity and failure rate of manual operation. The scheme not only improves the automation level of aerogel production, but also guarantees the smoothness and reliability of the control instruction through the joint loss function and trajectory stability evaluation, providing intelligent and precise technical support for industrial production.

[0016] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structure particularly pointed out in the written description and the accompanying drawings.

[0017] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are intended to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, for explaining the present application, and do not constitute a limitation of the present application. In the drawings: Figure 1 A flow chart of the aerogel production flow control method based on online monitoring in the embodiment of the present application; Figure 2 A schematic diagram of the aerogel production flow control system based on online monitoring in the embodiment of the present application. DETAILED DESCRIPTION

[0019] The preferred embodiments of the present application will be described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application.

[0020] The research and development idea of the present application originates from the demand for precise control of complex reaction processes in aerogel production, aiming to solve the shortcomings of traditional methods in dynamic regulation and stability through the combination of artificial intelligence and online monitoring technology. In the research and development process, first, the online sensor system is used to collect real-time data of key state parameters such as reaction kettle temperature, pH value, and viscosity, forming high-dimensional time series data; then, a multi-step prediction model is designed with a long short-term memory network as the core, using a time series convolution network to extract local features and combining an attention mechanism to focus on key process stages to predict the reaction trajectory and generate optimized control instructions; to ensure the robustness of the model, a joint loss function of mean square error and trajectory smoothness is used, and a stability evaluation index based on tolerance range is introduced; finally, for abnormal situations, a human-machine collaborative task generation mechanism is developed, which quickly generates adjustment tasks through deviation analysis and emergency evaluation to assist users in achieving precise intervention. This systematic design from data collection, model training to human-machine collaboration ensures the intelligence and efficiency of flow control, providing an innovative solution for the industrial application of aerogel production.

[0021] Figure 1 The flow chart of the aerogel production flow control method based on online monitoring provided by the embodiments of the present application is shown in FIG. 1, which comprises the following steps: Figure 1 S1: Collecting state parameters in real time during the sol-gel reaction process of aerogel sol by an online sensor system. The state parameters include at least three parameter combinations of reaction kettle temperature, system pH value, fluid viscosity, precursor concentration, and gel time.

[0022] ​The online sensor system monitors at least three parameter combinations of reaction kettle temperature, system pH, fluid viscosity, precursor concentration, and gel time in real time by integrating multiple high-precision sensors to fully characterize the dynamic changes of the reaction process. The reaction kettle temperature is measured by a thermocouple or an infrared thermometer, and the sensor is installed on the inner wall of the reaction kettle or the surface of the reaction liquid to record the temperature value (unit: °C, accuracy ±0.1°C). The system pH is measured by a pH electrode, which is immersed in the reaction liquid to output the pH value in real time (range 0-14, accuracy ±0.01). The fluid viscosity is measured by an online viscometer, which uses a vibration or rotary sensor to calculate the viscosity value based on fluid resistance (unit: mPa·s, accuracy ±1%). The precursor concentration is measured by a UV-visible spectrophotometer or a refractometer, which determines the concentration value based on the standard curve of absorbance or refractive index and concentration (unit: mol / L, accuracy ±0.01 mol / L). The gel time is monitored by an optical turbidimeter or a dynamic light scattering instrument, which records the time length from sol to gel transition (unit: seconds, accuracy ±1 seconds) by detecting the sudden change of system turbidity or particle size growth. The sensor system collects data at a fixed sampling frequency (such as once every second), converts the analog signal to digital signal through an analog-to-digital converter, and stores it in a local controller or cloud database for subsequent analysis. The collected parameter combination should contain at least three parameters to ensure multi-dimensional characterization of the reaction state, such as temperature, pH, and viscosity, reflecting the thermodynamic, chemical equilibrium, and fluid mechanics characteristics.

[0023] A specific implementation example is provided as follows: In a sol-gel reaction system of an aerogel production factory, an online sensor system is deployed, including a thermocouple (model K, installed on the inner wall of the reactor, temperature measurement range 0-200℃), a pH electrode (model E-201-C, immersed in the reaction solution, calibrated to pH 4-10), and a vibrating viscometer (model SV-10, installed on the discharge pipeline of the reactor, measurement range 1-1000 mPa·s). Taking the production of aerogel using tetraethyl orthosilicate (TEOS) as precursor as an example, the sensor system collects data at a frequency of 1 Hz: the thermocouple records the temperature of the reactor as 50.2℃ (±0.1℃), the pH electrode records the pH value of the system as 6.85 (±0.01), and the viscometer records the fluid viscosity as 120 mPa·s (±1.2 mPa·s). The precursor concentration is measured by a UV-visible spectrophotometer (model UV-1800), based on the absorbance of TEOS at 240 nm wavelength, combined with a previously established concentration-absorbance standard curve, and determined as 0.5 mol / L (±0.01 mol / L). The gel time is monitored by an optical turbidimeter (model HACH2100Q), and the moment when the turbidity of the system changes from 0.1 NTU to 10 NTU is recorded as the gel time of 1800 seconds (±1 second). The collected data is transmitted to the PLC controller through the RS485 interface and stored in the local SQLite database in real time for subsequent call by the flow control AI model. The system calibrates the sensors once a minute to ensure data consistency.

[0024] S2: inputting the state parameters into a pre-trained flow control AI model. The pre-training step of the flow control AI model includes: S21, collecting the time series state parameters and corresponding flow control instructions recorded in the historical production process to form a sample set, and performing normalization and sliding window segmentation processing on the sample set.

[0025] The time series state parameters and corresponding flow control instructions in the historical production process are collected to construct a sample set for training the flow control AI model. The time series state parameters include the temperature of the reaction kettle, the pH value of the system, the fluid viscosity, the precursor concentration, and the gel time, which are obtained in the same way as S1, recorded by the online sensor system at a fixed frequency (such as 1 Hz), and stored as time series data (format: timestamp-parameter value pair). The flow control instruction refers to the specific instruction for controlling the flow rate of the precursor, catalyst or solvent (unit: mL / min, accuracy ±0.1 mL / min), which is obtained through the flow controller log of the production equipment and recorded as a timestamp-flow value pair. The sample set is composed of multiple sets of time series data, each containing a sequence of state parameters and a corresponding sequence of flow control instructions for a continuous time period (such as 1 hour). The normalization process uses the min-max normalization method to map each parameter value to the [0, 1] interval, with the calculation formula being: (x-xmin) / (xmax-xmin), where xmin and xmax are the historical minimum and maximum values of the parameter, respectively, which are obtained by traversing the historical data. The sliding window segmentation process slides with a fixed time window (such as 60 seconds) and a step size of 1 second, generating multiple time series segments of fixed length, each containing state parameters and corresponding flow instructions, to capture the dynamic characteristics of the reaction process. The processed sample set is stored as a structured data set for subsequent model training.

[0026] For example, in an aerogel production plant, historical data is extracted from the sol-gel reaction records of the past 6 months, containing 100,000 sets of time series data, each recording the temperature (40-60℃), pH value (6-8), viscosity (100-150 mPa·s), precursor concentration (0.4-0.6 mol / L), and gel time (1500-2000 seconds) collected at a frequency of 1 Hz within 1 hour, as well as the corresponding precursor (TEOS) flow instruction (10-50 mL / min). The data is extracted from the production database through SQL query to generate a sample set, with each data set stored in CSV format (columns: timestamp, temperature, pH, viscosity, concentration, gel time, flow). The normalization process uses the NumPy library in Python to traverse the sample set and calculate xmin and xmax for each parameter (e.g. temperature xmin=40℃, xmax=60℃), normalizing the parameter values to [0, 1] (e.g. temperature 50℃ normalized to 0.5). The sliding window segmentation uses the Pandas library with a 60-second window and a 1-second step size to generate approximately 99,000 sequence segments, each containing data for 60 time points (5 state parameters + 1 flow instruction). The processed sample set is stored in HDF5 format, occupying about 500MB, for subsequent multi-step prediction model training, ensuring efficient data reading.

[0027] S22, based on the processed sample set, a multi-step prediction model with a long short-term memory network as the core is constructed, the state parameter sequence of the previous period is taken as the input of the multi-step prediction model, and the deviation of the state parameter trajectory of the subsequent period from the preset optimization path is taken as the training target.

[0028] Based on the processed sample set, a multi-step prediction model with a long short-term memory network (LSTM) as the core is constructed, which is used to predict the future state parameter trajectory and generate flow control instructions. LSTM is a recurrent neural network suitable for processing time series, which captures long-term dependencies through forget gate, input gate and output gate mechanism, and adapts to the dynamic changes of aerogel reaction process. The multi-step prediction model takes the normalized state parameter sequence (length T, for example, 60 seconds) as input, and outputs the state parameter trajectory (including temperature, pH value, viscosity, etc.) of future N steps (for example, 300 seconds) and the corresponding flow control instruction sequence. The model structure includes input layer, 2-3 layers of LSTM unit (128-256 neurons per layer), fully connected layer and output layer. The input layer receives a T×M matrix (M is the number of state parameters, such as 5), the LSTM unit controls the retention of historical information through the forget gate (based on sigmoid function, output [0,1]), updates the unit state through the input gate (combined with tanh activation function), and generates hidden state through the output gate. The fully connected layer maps the LSTM output to the state parameter trajectory and flow instruction. The training target is to minimize the deviation of the predicted trajectory from the preset optimization path (defined by process experts, for example, temperature is stable at 50±1℃), and the deviation is calculated by mean square error. The model is built on GPU using TensorFlow or PyTorch framework, and the parameters are optimized by back propagation.

[0029] For example, in the aerogel production scenario, a multi-step prediction model based on LSTM is built using the PyTorch framework. The model input is a 60-second normalized sequence of state parameters (5 dimensions: temperature, pH, viscosity, concentration, and gelation time), and the output is a 5-dimensional state parameter trajectory and a 1-dimensional flow command sequence for the next 300 seconds. The model contains three LSTM layers (256 neurons per layer, tanh activation function), followed by a fully connected layer (128 neurons, ReLU activation), and the output layer is 6-dimensional (5-dimensional state parameters + 1-dimensional flow command). Input data is loaded from the HDF5 sample set of S21, with preset optimization paths such as temperature 50±1℃, pH 7±0.1, and viscosity 120±5 mPa·s. The model is built on an NVIDIA RTX 3060 GPU, using the nn.LSTM module of PyTorch to implement LSTM units. The forget gate and input gate control the information flow through the sigmoid function, and the output gate generates the prediction sequence. The fully connected layer uses the nn.Linear module to map the output. During training, the sample set was divided into an 80% training set and a 20% validation set, with a batch size of 64 and an initial learning rate of 0.001. The model was optimized through backpropagation, iterating for 100 epochs to ensure that the predicted trajectory deviation was less than 5%. The trained model was saved as a .pth file for subsequent real-time prediction.

[0030] S23. During the training of the multi-step prediction model, a temporal convolutional network is used to extract local features of the input sequence, and an attention mechanism is used to weight and focus on key process stages.

[0031] During the training of the multi-step prediction model, a Temporal Convolutional Network (TCN) and an attention mechanism are integrated to enhance the model's ability to perceive local features and key process stages of the input sequence. The TCN extracts local patterns (such as abrupt temperature changes or rapid viscosity increases) from the time series using one-dimensional convolutional kernels (size 3-5, stride 1), and captures features at different time scales through multi-layer residual connections (4-8 convolutional kernels per layer), outputting a feature vector (dimension matching the LSTM input). The attention mechanism, based on a self-attention model, calculates the weights at each time point of the input sequence (normalized using a softmax function), focusing on key process stages (such as the gelation initiation point). The parameters of the TCN include the kernel size (selected experimentally, e.g., 3), the number of layers (4-6), and the dilation factor (a power of 2), determined through grid search. The parameters of the attention mechanism include the number of attention heads (4-8) and the query-key dimension (64-128), optimized through validation set performance. The outputs of the TCN and the attention mechanism are concatenated with the LSTM input as input to the multi-step prediction model. TCN and attention mechanisms are implemented using PyTorch or TensorFlow, and are jointly optimized with LSTM during training to reduce local feature loss and misjudgment at critical stages.

[0032] For example, in model training, TCN adopts torch.nn.Convl d module of PyTorch, sets 4 layers of convolution, 8 convolution kernels for each layer (size 3, expansion factor 1, 2, 4, 8 respectively), input is 60x5 dimensional state parameter sequence, output is 60x128 dimensional feature vector. Attention mechanism uses torch.nn.MultiheadAttention module, sets 4 attention heads, query-key dimension is 64, input is the feature vector output by TCN, calculates the weight of each time point (for example, the weight of the gelation stage is 0.8). The output of TCN and attention mechanism is spliced with the original sequence through torch.cat, and input into the LSTM model in S22. Parameter optimization is determined through grid search: try convolution kernel size 3, 4, 5, layer number 4, 5, 6, attention head number 4, 6, 8, select the combination with the smallest bias on the validation set (convolution kernel size 3, layer number 4, head number 4). Training is carried out on NVIDIA RTX3060 GPU, batch size is 64, learning rate is 0.001, and iteration is 100 epochs. TCN extracts the mutation feature of temperature from 49℃ to 51℃, and attention mechanism concentrates the weight on the stage near 1800 seconds of gelation time (weight 0.75-0.85), which significantly improves the prediction accuracy of the model on the key process stage, and reduces the bias of the validation set by 10%.

[0033] S24, adopts mean square error and trajectory smoothness as a joint loss function, and introduces a trajectory stability evaluation index based on a tolerance range as a model validation standard.

[0034] The joint loss function combines mean square error (MSE) and trajectory smoothness to optimize the prediction accuracy and stability of the multi-step prediction model. MSE calculates the deviation of the predicted state parameter trajectory from the preset optimized path, defined as the average of the squared difference between the predicted value and the target value, with a unit of normalized dimensionless value, calculated through the real trajectory in the sample set. Trajectory smoothness is measured by the sum of the absolute values of the second-order difference of the predicted trajectory, reflecting the degree of fluctuation of the trajectory, with a unit of dimensionless value, calculated by numerical differentiation (for example, the difference between adjacent time points of the predicted value). The joint loss function is a weighted sum: the weight of MSE is 0.7, and the weight of smoothness is 0.3, determined through validation set experiments. The trajectory stability evaluation index is based on the tolerance range (for example, temperature ±1℃, pH value ±0.1), calculates the proportion of the predicted trajectory within the tolerance range (unit: percentage), determined by comparing the predicted value with the tolerance boundary, and considered stable if the proportion is higher than 90%. During training, the joint loss function optimizes the model parameters through backpropagation, and the stability index evaluates the model performance on the validation set to ensure that the predicted trajectory is both accurate and stable.

[0035] For example, in model training, the joint loss function is implemented using PyTorch, the MSE is calculated by torch.nn.MSELoss, and the input is the predicted trajectory (5-dimensional state parameters, 300 seconds) and the preset optimization path (e.g., temperature 50°C, pH 7). The trajectory smoothness is calculated by torch.diff to calculate the second-order difference of the predicted trajectory, and the sum of the absolute values is taken as the smoothness loss. The weights are set to MSE 0.7, smoothness 0.3, which are determined by validation set experiments (try 0.6-0.8 range, 0.7 best effect). The stability index is calculated by a Python script, which checks the proportion of points in the predicted trajectory that meet the tolerance range of temperature 49-51°C and pH 6.9-7.1, for example, a 300-second predicted trajectory has 270 seconds that meet the tolerance, and the stability is 90%. During training, the joint loss function is optimized on an NVIDIA RTX3060 GPU, with a batch size of 64, a learning rate of 0.001, and 100 epochs of iteration. On the validation set, the MSE is reduced to 0.01, the smoothness loss is reduced to 0.005, and the stability index reaches 92%, indicating that the model's predicted trajectory is accurate and has small fluctuations, meeting production requirements.

[0036] S25, the network parameters of the multi-step prediction model are optimized using the gradient descent algorithm until the multi-step prediction model has the ability to determine the first flow control instruction based on the current and historical state parameter sequence, predict the future state parameter trajectory, and judge whether the state parameter trajectory can effectively stabilize within the tolerance range of the preset optimization path, obtaining a flow control AI model.

[0037] The gradient descent algorithm iteratively updates the network parameters of the multi-step prediction model (LSTM weights, TCN convolution kernel weights, attention mechanism weights, etc.), minimizes the joint loss function, and realizes flow control instruction generation and trajectory prediction based on the state parameter sequence. The algorithm uses the Adam optimizer, combined with adaptive learning rate and momentum method, to accelerate convergence. Parameters include learning rate (initial value 0.001, adjusted by learning rate decay), batch size (32-128, determined by memory capacity), number of iterations (50-200, determined by convergence of validation set). The training process is as follows: input the sample set of S21, calculate the joint loss function (MSE + smoothness), calculate the parameter gradient by back propagation, and update the weights. The model capabilities include: generating the first flow control instruction (unit: mL / min) based on the current and historical state parameter sequence, predicting the future 300-second state parameter trajectory, and judging whether the trajectory is within the tolerance range (e.g., temperature ±1°C). The Adam optimizer is implemented by PyTorch's torch.optim.Adam, and the training is performed on the GPU. A stability index of more than 90% on the validation set is considered to be trained.

[0038] For example, gradient descent is implemented using PyTorch's torch.optim.Adam with an initial learning rate of 0.001, a batch size of 64, and 100 epochs. The training data is the HDF5 sample set of S21, with input as a 60-second x 5-dimensional state parameter sequence and output as a 300-second x 5-dimensional trajectory and 1-dimensional flow command. The joint loss function (MSE weight 0.7, smoothness weight 0.3) is calculated through S24 definition, backpropagation is implemented through torch.autograd, and the weights of LSTM, TCN, and attention mechanism are updated. Training is performed on an NVIDIA RTX 3060 GPU with a memory usage of approximately 2 GB. The learning rate is decayed by 0.1 every 20 epochs (minimum 0.0001), and convergence is monitored through the validation set. After training, the model generates flow commands (e.g., TEOS flow 30 mL / min) for input sequences (e.g., temperature 50°C, pH 7) and predicts 300-second trajectories (temperature 49.8-50.2°C) with a stability of 93%. The model is saved as a.pth file and deployed to the production system to generate flow commands and predict trajectories in real-time, meeting the tolerance requirements.

[0039] S3: The flow control AI model performs the following steps: Based on the current and historical state parameter sequence, determine the first flow control command required to return the reaction process to the preset optimization path.

[0040] Based on the first flow control command, predict the state parameter trajectory for a future period of time.

[0041] Determine whether the state parameter trajectory can effectively stabilize within the tolerance range of the preset optimization path.

[0042] The flow control AI model is based on a multi-step prediction model trained on S2, which processes current and historical state parameter sequences in real-time to perform three tasks: generating first flow control instructions, predicting future state parameter trajectories, and judging trajectory stability. The first flow control instructions are determined through model output, with units of mL / min, based on LSTM-predicted flow values, reflecting the delivery amount of precursors, catalysts, or solvents. The future state parameter trajectory is a 5-dimensional parameter sequence (temperature, pH, viscosity, concentration, gel time) for the next N seconds (e.g., 300 seconds), predicted by LSTM and TCN combined, with an attention mechanism focusing on key stages. The trajectory stability is calculated by comparing the predicted trajectory with the preset optimization path tolerance range (e.g., temperature ±1°C), with the stability index being the percentage of points meeting the tolerance (over 90% for stability). The model runs on an embedded controller or industrial PC, with input being real-time data collected by S1 (60 seconds x 5 dimensions), output being instructions and trajectories, and storage in local cache for S4 and S5. The model is optimized to inference mode by torch.jit.trace of PyTorch, reducing computational delay (<100ms).

[0043] For example, in an aerogel production system, the flow control AI model is deployed on an industrial PC (configuration: Inteli7, 16GB RAM), loads the.pth file saved by S25, and uses torch.jit.trace of PyTorch to convert to inference mode. The input is a 60-second state parameter sequence collected by S1 (e.g., temperature 50.1°C, pH 6.95, viscosity 121 mPa·s), the model generates first flow control instructions (e.g., TEOS flow 32.5 mL / min, ±0.1 mL / min), and predicts the trajectory for the next 300 seconds (temperature 49.9-50.3°C, pH 6.9-7.0). The stability judgment is calculated by a Python script, with 280 seconds in the trajectory meeting the tolerance (temperature ±1°C, pH ±0.1), and the stability being 93.3%. The inference time is 80ms, meeting the real-time requirements. The output is stored in a SQLite database, with fields including timestamp, flow instructions, predicted trajectory, and stability index. The model processes the input once a minute, ensuring real-time monitoring and control of the reaction process, with the predicted trajectory deviating from the actual production data by less than 5%.

[0044] S4: If the result of step S3 is yes, automatically adjust the delivery flow of at least one of the precursors, catalysts, or solvents according to the first flow control instructions.

[0045] When S3 determines that the state parameter trajectory is within the tolerance range (stability ≥ 90%), the delivery flow rate of the precursor, catalyst or solvent is automatically adjusted according to the first flow control instruction to ensure that the reaction process returns to the preset optimization path. The flow control instruction is a numerical value (unit: mL / min, accuracy ± 0.1 mL / min) output by S3, which is converted into a valve opening signal (0-100%) through the PID algorithm of the industrial controller to control the metering pump or solenoid valve. The adjustment objects include precursors (such as TEOS), catalysts (such as ammonia water) and solvents (such as ethanol), and the actual flow rate (accuracy ± 0.1 mL / min) is monitored through respective flowmeters and compared with the instruction value to form a closed-loop control. The PID parameters (proportional, integral and derivative coefficients) are determined through production equipment debugging, for example, the proportional coefficient Kp = 0.5 and the integral time Ti = 10 seconds. The adjustment process is executed by the PLC controller, the instruction is transmitted to the flow control equipment through the Modbus protocol, and the actual flow rate is fed back to the database in real time, and a deviation of less than 2% is considered to be a successful adjustment.

[0046] For example, in an aerogel production system, the first flow control instruction generated by S3 is TEOS flow rate 32.5 mL / min, with a stability of 93%. The instruction is sent to the Siemens S7-1200 PLC through the Modbus protocol, and the built-in PID algorithm (Kp = 0.5, Ti = 10 seconds, Td = 1 second) in the PLC converts the instruction into an opening signal (60%) for the metering pump (model Grundfos DDA). The pump is connected to the TEOS tank, and the actual flow rate is monitored by an electromagnetic flowmeter (model Endress+Hauser Proline) to be 32.4 mL / min (± 0.1 mL / min). The adjustment process feeds back flow rate data to the PLC every second, and a deviation of 0.1 mL / min (< 2%) is considered to be successful. The system simultaneously adjusts the catalyst (ammonia water) flow rate to 5 mL / min and the solvent (ethanol) flow rate to 20 mL / min, both of which are achieved through similar closed-loop control. The adjustment data is recorded in the SQLite database, including timestamp, instruction value, actual flow rate and deviation, to ensure that the reaction temperature is stable at 50 ± 1°C and the pH value is at 7 ± 0.1.

[0047] S5: If the result of step S3 is no, a human-machine collaborative task is triggered. The triggering of the human-machine collaborative task includes: S51, when the flow control AI model determines that the state parameter trajectory cannot be stabilized within the tolerance range of the preset optimization path, a collaborative task trigger signal containing the deviation parameter and the out-of-control time window is generated.

[0048] When S3 determines that the state parameter trajectory cannot be stabilized within the tolerance range (stability < 90%), the flow control AI model generates a collaborative task trigger signal, including a deviation parameter and an out-of-control time window. The deviation parameter is the difference between the predicted trajectory and the preset optimized path (for example, temperature deviation 2°C, pH value deviation 0.2), which is calculated by comparing the trajectory prediction result of S3 with the tolerance boundary. The out-of-control time window is the time period (unit: seconds) during which the trajectory exceeds the tolerance, which is determined by detecting the start and end points of the continuous non-tolerance in the trajectory (for example, 100-150 seconds of 300 seconds of trajectory exceed the tolerance). The trigger signal is structured data, including the deviation parameter (5-dimensional vector), the out-of-control time window (start and end time stamp), and is transmitted to the human-computer interaction system through the MQTT protocol. The signal generation is realized through a Python script running on an industrial PC, and the signal generation delay is less than 50ms, ensuring fast response.

[0049] For example, in production, S3 predicts a 300-second trajectory showing that the temperature rises to 52.5°C (exceeding 50±1°C) at 100-150 seconds, with a stability of 85% (<90%), triggering a collaborative task. The Python script calculates the deviation parameter (temperature deviation 2.5°C, pH value deviation 0.15, etc.), and determines the out-of-control time window as 100-150 seconds. The trigger signal generation includes the deviation vector ([2.5, 0.15, 0, 0, 0]) and the time window (100s-150s), which is sent to the human-computer interaction system (topic: / collaborative_task) through the Paho-MQTT library in MQTT protocol. The signal generation takes 40ms, and is stored in the SQLite database (fields: timestamp, deviation parameter, time window). After the signal is sent, the system logs record the triggering event, ensuring accurate reception and processing of subsequent human-computer collaborative tasks.

[0050] S52, sends the trigger signal to the human-computer interaction system to activate the collaborative task generation engine therein.

[0051] The trigger signal is sent to the human-computer interaction system through the MQTT protocol to activate the collaborative task generation engine therein. The human-computer interaction system is a software platform running on an industrial PC or a cloud server, including a display interface and a task management module, which receives MQTT messages (topic subscription, format is structured data). The collaborative task generation engine is a rule-based decision module that analyzes the deviation parameter and the out-of-control time window in the trigger signal to generate a task description. The engine is implemented through the PythonFlask framework and deployed on a local server to analyze the deviation parameter (e.g., temperature deviation 2°C) and the time window (e.g., 50 seconds) of the signal and generate a task according to the preset rules (e.g., adjust the TEOS flow if the deviation > 1°C). The MQTT protocol ensures that the signal transmission delay is less than 100ms, and the engine activation time is less than 50ms, ensuring real-time performance.

[0052] For example, the trigger signal is sent to the human-machine interaction system through the MQTT protocol (server: localhost:1883, topic: / collaborative_task), running on an industrial PC (Ubuntu 20.04, Flask 2.0). The signal contains a temperature deviation of 2.5°C and a loss of control time window of 100-150 seconds. The Flask server subscribes to the topic and parses the signal data. The collaborative task generation engine is a Python module in the Flask application, which loads preset rules (e.g., temperature deviation > 1°C, recommend reducing TEOS flow). The engine takes 30ms to activate and generates a task description after parsing the signal (stored as a SQLite record, fields: deviation 2.5°C, time window 100-150 seconds, recommended material TEOS). The task is displayed through a web interface, including the deviation value and the time window, prompting the operator to adjust the flow. The system logs the signal reception and engine activation time to ensure that the task generation is not missed.

[0053] S53, through the collaborative task generation engine, based on the deviation parameter and the loss of control time window, constructs a human-machine collaborative task containing the type of material to be adjusted, the target flow range and the adjustment emergency level.

[0054] The collaborative task generation engine constructs a human-machine collaborative task based on the deviation parameter and the loss of control time window, including the type of material to be adjusted (TEOS, ammonia or ethanol), the target flow range (unit: mL / min, accuracy ± 0.1 mL / min) and the adjustment emergency level (high, medium, low). The type of material is determined by analyzing the deviation parameter, for example, if the temperature deviation is greater than 1°C, the TEOS flow is adjusted first. The target flow range is calculated based on historical data statistics and process rules, for example, the TEOS flow is adjusted within 20-40 mL / min based on the flow distribution of successful cases in the sample set. The emergency level is determined according to the loss of control time window and the deviation size, for example, the time window 2°C is high. The task is generated in JSON format and displayed on the web interface of the human-machine interaction system for the operator to refer. The engine is implemented through Flask and SQLAlchemy, the rules are loaded from the database, and the task generation delay is less than 100ms.

[0055] For example, the collaborative task generation engine runs on a Flask server, parsing the signal (temperature deviation 2.5 °C, time window 100-150 seconds) of S51. The engine queries the rules in the SQLite database (temperature deviation > 1 °C, adjust TEOS flow; deviation > 2 °C, high urgency), determines that the material is TEOS, and the target flow range is 25-35 mL / min (based on historical data statistics). The urgency is high due to the deviation of 2.5 °C and the time window of 50 seconds. The task generation JSON data (fields: material = TEOS, flow range = 25-35 mL / min, urgency = high) is displayed through the Flask web interface, which uses the Bootstrap framework and includes material type, flow range, and urgency labels (red for high). The task generation takes 80 ms, is stored in the SQLite database, and is available for operators to view and execute, with system logs recording the task generation time and content.

[0056] S6: Assist the user to perform the human-machine collaborative task, generate a second flow control instruction, and adjust the delivery flow of the corresponding material according to the second flow control instruction.

[0057] The user assists in performing the human-machine collaborative task generated in S53, receives operator input through the human-machine interaction system, generates a second flow control instruction (unit: mL / min, precision ± 0.1 mL / min), and adjusts the delivery flow of the corresponding material. The operator views the task (material type, target flow range, urgency) on the web interface and inputs the specific flow value (e.g., TEOS flow 30 mL / min). The system verifies whether the input value is within the target range (e.g., 25-35 mL / min). The second flow control instruction is sent to the PLC through the Modbus protocol to control the metering pump or electromagnetic valve, and the adjustment method is the same as S4, using the PID algorithm (Kp = 0.5, Ti = 10 seconds) to convert to valve opening. The adjustment process provides real-time feedback of the actual flow (monitored by a flow meter, precision ± 0.1 mL / min), and a deviation of less than 2% is considered successful. The system is implemented through Flask and Modbus, and the operation logs are stored in the database to ensure traceability.

[0058] For example, in the web interface of the human-machine interaction system (running on the Flask server), the operator views the S53 task (TEOS, 25-35 mL / min, high urgency), and inputs the flow value of 30 mL / min. The system verifies that the input is within the range, sends it to the Siemens S7-1200 PLC through the Modbus protocol, and converts it to a metering pump opening signal (58%). The pump adjusts the TEOS flow to 30.1 mL / min (flowmeter monitoring, deviation 0.1 mL / min), meeting the <2% requirement. The adjustment data is recorded to the SQLite database (fields: timestamp, instruction value 30 mL / min, actual flow 30.1 mL / min). After the operator confirms the adjustment, the system displays "adjustment successful" through the web interface, and logs the operation time and flow value. The reaction temperature drops to 50.2°C within 5 minutes, the pH value returns to 7.0, the trajectory stability reaches 95%, and the task validity is verified.

[0059] The technical solution realizes dynamic prediction and precise regulation of the reaction process by collecting key state parameters in aerogel production in real time through an online sensor system and combining a pre-trained flow control AI model, significantly improving production efficiency and product quality stability compared to traditional methods. The model accurately captures the time sequence characteristics and key stage information of the process parameters through long short-term memory networks, time sequence convolution networks, and attention mechanisms, generates optimized flow control instructions, and ensures that the reaction trajectory is stable within the tolerance range of the preset optimization path. In addition, when the predicted trajectory deviates from the optimization path, the system generates adjustment tasks containing deviation parameters and urgency levels through a human-machine collaborative task generation mechanism, assisting users in efficient intervention and reducing the complexity and error rate of manual operation. This scheme not only improves the automation level of aerogel production, but also guarantees the smoothness and reliability of the control instructions through the joint loss function and trajectory stability evaluation, providing intelligent and precise technical support for industrial production.

[0060] In complex industrial processes such as aerogel sol-gel reactions, human-machine collaborative tasks involve multi-material flow regulation, reaction path optimization, and real-time risk management, facing challenges such as high decision-making complexity, heavy user cognitive burden, and potential loss of control risks caused by hesitation. Traditional human-computer interaction systems usually rely on static interfaces or single parameter monitoring, which are difficult to dynamically present the coupling relationship of multi-dimensional reaction parameters and the comparison of historical optimization paths, leading to users' difficulty in quickly forming accurate regulation cognition in high-dimensional decision-making scenarios. In addition, users' hesitation behavior (such as repeatedly checking parameters, modifying instructions, or input delay) is often ignored, and the decision uncertainty reflected by it cannot be effectively utilized to optimize the auxiliary strategy. At the same time, existing technologies lack real-time evaluation mechanisms for instruction execution risks, especially for the comprehensive impact of multi-instruction sequences on the reaction process (such as reaction path deviation, gel time fluctuation, and product quality deviation), and cannot provide dynamic decision reversibility analysis and risk intervention within the key time window.

[0061] To solve the above problems, there is an urgent need for a comprehensive human-machine collaborative decision support system that can integrate real-time state visualization, user hesitation state analysis, risk prediction, and dynamic auxiliary strategy, to improve the decision-making efficiency and accuracy of users in complex industrial scenarios, reduce the risk of loss of control, and ensure the stability of the reaction process and the quality of the product.

[0062] Therefore, in some embodiments, the auxiliary user performs a human-machine collaborative task, including: S61, based on the defined material types to be adjusted, target flow range, and adjustment urgency level in the human-machine collaborative task, a visual decision model integrating real-time state visualization window, historical optimization path comparison view, and multi-material flow collaborative regulation interface is constructed. The visual decision model dynamically renders the deviation area of the reaction parameter evolution curve and the preset optimization path, and displays the coupling influence coefficient of different material flow adjustments, to assist users in forming multi-dimensional decision-making cognition.

[0063] The visualized decision model is an integrated user interface aiming to assist users in multi-dimensional decision making by dynamically presenting real-time state, historical data comparison and material flow adjustment suggestions during the aerogel sol-gel reaction process. The type of material to be adjusted (e.g. tetraethyl orthosilicate TEOS, ammonia or ethanol) is determined by the human-machine collaborative task generated by S53 and stored as a string identifier (e.g. "TEOS"). The target flow range is defined in the form of a numerical interval (unit: mL / min, precision ±0.1 mL / min, for example 25-35 mL / min), which is calculated by historical production data statistics and process rules, and the flow distribution of successful cases in the production database is obtained by querying. The adjustment urgency level (high, medium, low) is based on the size of the deviation parameter and the length of the time window of loss of control, which is calculated by preset thresholds (e.g. deviation > 2℃ for high). The real-time state visualization window dynamically renders the evolution curve of the reaction parameters (e.g. temperature, pH, viscosity) through the Web interface, and the curve data is obtained from the online sensor system of S1 (sampling frequency 1 Hz, precision ±0.1℃, ±0.01pH, etc.), using the Plotly library to draw a line chart, with the horizontal axis as time (unit: seconds) and the vertical axis as parameter value, and the tolerance area of the preset optimization path (e.g. temperature 50±1℃) is superimposed (deviation is represented by shading). The historical optimization path comparison view extracts the state parameter trajectory of past successful production cases (e.g. cases with temperature stable at 50±0.5℃) by querying the historical database, calculates the average trajectory and standard deviation using the Pandas library, and plots the comparison curve. The multi-material flow collaborative adjustment interface displays the coupling influence coefficient of different material (e.g. TEOS, ammonia) flow adjustment, which is calculated by the reaction kinetics model based on the partial derivative of the reaction parameter with respect to the material flow change (e.g. ∂temperature / ∂TEOS flow, unit: ℃ / (mL / min)), which is estimated from historical data by finite difference method. The model is deployed on the Flask server, and the data is updated in real time through the MQTT protocol, with a rendering delay of <100ms, ensuring that users can perceive the reaction state and adjustment effect in real time.

[0064] A specific implementation example is provided as follows: In the aerogel production factory, the visual decision model is deployed on an industrial PC running Ubuntu 20.04 (configuration: Inteli7, 16GB RAM) using the Flask framework and the Plotly library. The S53 task specifies that the material to be adjusted is TEOS, the target flow range is 25-35 mL / min, and the urgency is high. The real-time status window draws the line chart of temperature (50.2°C ± 0.1°C), pH value (6.85 ± 0.01), and viscosity (120 ± 1.2 mPa•s) through Plotly, the data is read from the SQLite database of S1 at a frequency of 1 Hz, the horizontal axis is the last 300 seconds, and the tolerance area is displayed in green shadow (such as temperature 49-51°C). The historical optimization path view queries the database for the past 6 months, extracts 10 cases where the temperature is stable at 50 ± 0.5°C, Pandas calculates the average trajectory (for example, temperature 50.1°C, standard deviation 0.3°C), and draws it as a blue contrast curve. The multi-material flow interface displays the coupling influence coefficient of TEOS and ammonia, the partial derivative is calculated based on historical data (1000 groups, TEOS flow 20-40 mL / min, temperature change 0.5-1°C) through Python script, and the temperature rises by 0.02°C / (mL / min) for every 1 mL / min increase in TEOS flow. The interface is beautified through the Bootstrap framework, displaying "TEOS flow suggestion: 30 mL / min, temperature impact +0.6°C". The data is obtained in real time from the sensor system through the Paho-MQTT library (topic: / process_data), and the rendering takes 80ms. The operator observes that the temperature is 2°C higher through the interface, combines the historical comparison and coupling coefficient, and judges that the TEOS flow needs to be reduced. The interface logs the user's viewing duration (15 seconds) and click frequency (3 times) and stores them in the SQLite database to ensure that the decision is traceable.

[0065] S62, whenever a user inputs a preliminary flow control sub-instruction based on the visual decision model, the following operations are performed in parallel: S621, real-time tracking of the user's hesitation state for the preliminary flow control sub-instruction. The hesitation state is judged based on the fusion of the first behavior indicator before user input and the second behavior indicator after input; the first behavior indicator includes: the frequency and duration of viewing related parameter panels in the visual decision model, the number of historical instruction modifications, the deviation degree of the current input instruction from the system recommended instruction; the second behavior indicator includes: the number of re-views of the instruction related parameters after the instruction is input, the trigger delay of the instruction modification operation, the mouse focus dwell jitter characteristics in the instruction input area; through the hesitation evaluation model based on long short-term memory network, the first behavior indicator and the second behavior indicator are time series modeled, and the current hesitation state quantitative value is output.

[0066] Real-time tracking of user hesitation state is achieved by fusing user interaction behaviors in the visualized decision model to evaluate their decision confidence on the prepared flow control sub-instruction. The prepared flow control sub-instruction is the flow value (unit: mL / min, precision ±0.1 mL / min, e.g., TEOS flow 30 mL / min) input by the user in the web interface. The first behavior indicators include: viewing parameter panel frequency (times / minute, counted by JavaScript click event), viewing duration (seconds, recorded by setinterval to record mouse hover time), historical instruction modification times (queried from the database for the last 1 hour modification log, unit: times), and deviation of current input instruction from system recommended instruction (calculated as the absolute difference between input value and S61 recommended value such as 30 mL / min, unit: mL / min). The second behavior indicators include: parameter re- viewing times after instruction input (recorded by click event to record the number of times to view the temperature panel), modification triggering delay (time from input to click the modification button, unit: seconds, counted by setTimeout), and mouse focus dwell jitter characteristics (calculated by mousemove event to calculate the trajectory variance of the mouse in the input box, unit: pixel2). The hesitation evaluation model is based on a long short-term memory network (LSTM), with 6-dimensional behavior sequences (3-dimensional first indicators + 3-dimensional second indicators) of 60 seconds as input, each dimension normalized to [0, 1] (min-max normalization). The model structure is 2-layer LSTM (128 neurons per layer, tanh activation), followed by a fully connected layer (64 neurons, ReLU activation), outputting a hesitation state quantitative value (0-1, 1 indicating extreme hesitation). The model is trained by historical behavior data (5000 groups, containing operator interaction logs), implemented using PyTorch, Adam optimizer (learning rate 0.001), trained for 100 epochs, and the validation set accuracy is >90%. The hesitation value is calculated every second, stored in the SQLite database, and used for subsequent steps.

[0067] For example, in an aerogel production system, an operator inputs a preliminary sub-instruction (TEOS flow 30 mL / min) at S61's web interface. The hesitation evaluation model is deployed on a Flask server (Inteli7, 16 GB RAM) using PyTorch to load a pre-trained LSTM model. The system records the first behavioral indicators through JavaScript: the operator views the temperature panel (50.2°C) 5 times per minute for a total of 20 seconds; database queries show that the instruction was modified 3 times in the past 1 hour; the input of 30 mL / min deviates from the recommended value of 32 mL / min by 2 mL / min. The second behavioral indicators: view the pH value panel 2 times after input, modification delay 5 seconds, mouse track variance 10 pixels² (calculated through the mousemove event). After normalizing the behavioral data (e.g., viewing frequency 5 times normalized to 0.5, range 0-10 times), a 60-second × 6-dimensional sequence is formed, which is input into the LSTM model (2 layers, 128 neurons per layer). The model outputs a hesitation value of 0.75 (indicating high hesitation), and the time consumption is 50 ms. The results are stored in a SQLite database (fields: timestamp, hesitation value, behavioral indicators). The operator frequently views the panel due to the high temperature (52°C), and the hesitation value reflects his uncertainty about the flow adjustment. The log records the interaction details (viewing duration 20 seconds, modification times 3), providing data support for S65.

[0068] S622, synchronously evaluate the joint risk state of the preliminary flow control sub-instruction if actually executed in real time; wherein the evaluation method of the joint risk state is: extracting the currently input preliminary flow control sub-instruction and the historical preliminary sub-instruction stored in the set of instructions to be implemented, constructing an instruction sequence; inputting the instruction sequence into the risk prediction network trained by reaction kinetics data to calculate its comprehensive influence index on the aerogel sol-gel reaction process; the comprehensive influence index includes: reaction path deviation expected value, gel time standard deviation increase ratio, product porosity deviation degree; based on the comprehensive influence index, a joint risk value is generated by weighted fusion.

[0069] The joint risk state assessment analyzes the impact of the preliminary flow control sub-instructions on the aerogel reaction process through a risk prediction network. The preliminary sub-instruction (e.g., TEOS flow 30 mL / min) and historical sub-instructions in the to-be-implemented instruction set (SQLite database storage, the latest 10, such as 28-32 mL / min) form an instruction sequence (length 11, unit: mL / min). The risk prediction network is based on a convolutional neural network (CNN), and the input is the instruction sequence and the current state parameters of S1 (temperature 50.2°C, pH value 6.85, etc.), and the output is a comprehensive impact index. The network structure includes 3 layers of one-dimensional convolution (16 convolution kernels per layer, size 3, ReLU activation), followed by a fully connected layer (128 neurons, ReLU activation), and outputs a 3-dimensional index: reaction path deviation expected value (mean square error of predicted trajectory and optimized path, unit: dimensionless, calculated by S3 trajectory comparison), gel time standard deviation increase ratio (percentage of predicted gel time fluctuation relative to historical mean value, calculated by historical data statistics), and product porosity deviation degree (deviation of predicted porosity from target value 5 nm, unit: nm, estimated by reaction kinetics model). The network is trained by 5000 sets of historical production data (including instructions, state parameters, and porosity results) using PyTorch, Adam optimizer (learning rate 0.001), 100 epochs, and verification set error <5%. The comprehensive impact index is generated by weighted fusion to generate a joint risk value (0-1, weights: path deviation 0.4, gel time 0.3, porosity 0.3, optimized by verification set), and the calculation time is <50 ms, and the storage is to the database.

[0070] For example, in a production system, an operator inputs a preliminary sub-instruction (TEOS flow 30 mL / min), and the to-be-implemented instruction set contains 10 historical instructions (average 31 mL / min). The risk prediction network (PyTorch implementation, deployed on an Intel i7 PC) loads the trained model, inputs an 11-dimensional instruction sequence and the current state (temperature 52°C, pH value 6.9). The network 3-layer convolution (16 kernels per layer, size 3) extracts sequence features, and the fully connected layer outputs the indexes: path deviation 0.03 (trajectory deviation 3%), gel time standard deviation increase 10% (historical mean value 1800 seconds), and porosity deviation 0.5 nm (target 5 nm). Weighted fusion (weights 0.4, 0.3, 0.3) calculates the joint risk value 0.65 (> threshold 0.5), and the time consumption is 40 ms. The results are stored in the SQLite database (fields: timestamp, risk value, index value). The production case shows that the 30 mL / min flow causes the temperature to rise to 52.5°C, the porosity is too large, and the risk value reflects potential quality problems, triggering S63. The log records the input instructions and risk assessment details to ensure traceability.

[0071] S63. When the joint risk value exceeds the preset risk threshold, the decision reversible time window is calculated by a dynamic window planning algorithm based on the remaining time from the current time point to the uncontrollable time window defined by the human-machine collaborative task, the decision progress reflected by the user's input instructions, and the magnitude of the joint risk value; wherein, the decision reversible time window is the maximum allowable period during which the user can revoke the input preparatory sub-instructions and the system has enough time to recover to the state before the decision.

[0072] When the combined risk value (e.g., 0.65) exceeds a preset threshold (e.g., 0.5, determined through historical data statistics), the dynamic window planning algorithm calculates the decision reversible time window to ensure the user has sufficient time to rescind high-risk sub-instructions. The remaining time is the difference (in seconds, obtained via the system clock) between the current time and the end of the S53 out-of-control time window (e.g., 150 seconds). Decision progress is quantified by the number of input sub-instructions (database queries, unit: times) and the deviation from the recommended value (unit: mL / min). For example, 5 instructions with an average deviation of 2 mL / min indicates high progress. The dynamic window planning algorithm is based on dynamic programming (DP), with the state being (remaining time, number of instructions, risk value) and the objective being to maximize the reversible time window (unit: seconds). The algorithm assumes that the system recovery time is proportional to the risk value and the instruction deviation (e.g., for every 0.1 increase in risk value, the recovery time increases by 10 seconds), and recursively calculates the optimal window: W(t) = min(Trecovery, Tremaining), where Trecovery is the recovery time (estimated through historical experiments), and Tremaining is the remaining time. The algorithm is implemented using Python's NumPy, iterating through 100 time steps (1 second each) with a time consumption of <20ms. The windowed results are stored in a database for use by S64.

[0073] For example, with a combined risk value of 0.65 (>0.5), an out-of-control time window of 100-150 seconds, a current time of 120 seconds, and 30 seconds remaining, a database query shows that 5 sub-instructions have been entered (TEOS flow rate 28-32 mL / min, deviating from the recommended value by 2 mL / min). A dynamic window planning algorithm (implemented in NumPy, running on an Intel i7 PC) takes the state (30 seconds, 5 sub-instructions, 0.65) as input and estimates the recovery time (risk value of 0.65 corresponds to 20 seconds, deviation of 2 mL / min adds 5 seconds, totaling 25 seconds). The algorithm recursively calculates a reversible time window of 25 seconds (min(25,30)). The calculation takes 15 ms, and the window is stored in an SQLite database (fields: timestamp, window duration). In production, the operator can revoke the 30 mL / min instruction within 25 seconds, and the system can be restored to a state of 50℃ and pH 7. The algorithm input (30 seconds remaining, risk value 0.65) and output (25-second window) are logged to ensure traceability of decision support.

[0074] S64, generate a hesitation state trigger constraint for the decision reversible time window. The trigger constraint is used to determine whether to assist the user in risk awareness within the current window, which is defined as: the user's hesitation state quantification value is continuously higher than the hesitation threshold in the latest continuous period, and the hesitation fluctuation variance is lower than the stability threshold.

[0075] The hesitation state trigger constraint is based on the hesitation state quantification value (0-1) of S621 to determine whether to assist the user in the decision reversible time window. The continuous period is the latest 60 seconds (defined by the system clock), and the hesitation threshold is 0.6 (reflected by the high hesitation tendency through historical operator behavior statistics). The hesitation fluctuation variance is the variance of the hesitation value within 60 seconds (unit: dimensionless, calculated by NumPy.var), and the stability threshold is 0.05 (historical data shows that the variance is 0.6 and the variance is <0.05, triggering risk awareness assistance. The condition is calculated in real time by a Python script, with the input being the hesitation value sequence of S621 (1Hz sampling), updated every second, with a time consumption of <10ms. The result is stored in the SQLite database (fields: timestamp, hesitation value, variance, trigger state) for S65. The constraint ensures that assistance is provided to users with high hesitation and stability, avoiding false triggers.

[0076] For example, the decision reversible time window is 25 seconds, and S621 provides a 60-second hesitation value sequence (average 0.75, range 0.7-0.8). The Python script (running on the Flask server) calculates the latest 60-second hesitation value mean 0.75 (>0.6) and variance 0.02 (<0.05), satisfying the trigger constraint. The calculation time is 8ms, and the result is stored in the SQLite database (fields: timestamp, mean 0.75, variance 0.02, trigger = yes). In production, the operator hesitates due to high temperature (52°C) and continuously checks the parameter panel (hesitation value 0.75), with stable behavior (variance 0.02), triggering S65 assistance. Log the constraint calculation details (input sequence, mean, variance) to ensure that the trigger is verifiable.

[0077] S65, when entering the decision reversible time window, real-time detect whether the user's current hesitation state meets the trigger constraint.

[0078] After entering the decision reversible time window (e.g. 25 seconds), the system real-time detects whether the hesitation value of S621 meets the trigger constraint of S64 (hesitation value >0.6, variance 0.6 and variance <0.05, trigger state is "yes", otherwise "no"). The detection result is stored in the database (fields: timestamp, trigger state), with a time consumption of <10ms. The detection process is run through the Flask server to ensure real-time performance, and abnormal situations (such as data missing) are handled by default values (hesitation value 0) to avoid interruption.

[0079] For example, in a 25-second window, the system reads the hesitation value sequence (60 seconds, mean 0.75, variance 0.02) of S621. The Python script calculates the mean 0.75 (> 0.6) and variance 0.02 (< 0.05) per second, triggering the state as “Yes” in 9 ms. The result is stored in the SQLite database. In production, the operator keeps high hesitation due to frequent checking of the temperature (52 °C), triggering S66 assistance. The log records the input (hesitation value sequence) and output (trigger = Yes), ensuring real-time and accuracy.

[0080] S66, if yes, based on the remaining length of the decision reversible time window, the user's operation trajectory and cognitive focus sequence in the visualization decision model within the previous setting memory time window, the cognitive content reconstruction engine infers the user's received cognitive information fragments and decision logic gaps.

[0081] When S65 triggers assistance, the cognitive content reconstruction engine analyzes the operation trajectory and cognitive focus sequence in the remaining length of the decision reversible time window (e.g., 20 seconds, system clock calculation) and the memory time window (the last 300 seconds, database query), infers the user's cognitive information and logic gaps. The operation trajectory includes click events (panel viewing times, JavaScript records), input records (flow values, database queries). The cognitive focus sequence is the order of the parameter panels viewed by the user (e.g., temperature -> pH value, click event record). The engine is based on a rule model, with rules such as: frequently viewed parameters (e.g., temperature 5 times) are the main cognitive focus, and frequent modification instructions (> 3 times) indicate logic gaps (uncertain effect on flow). The engine is implemented by Python's Pandas, which parses the trajectory data and outputs cognitive information fragments (e.g., “focus on temperature 52 °C”) and gaps (e.g., “do not understand the effect of TEOS flow on temperature”). The time cost is < 50 ms. The results are stored in the database for S67.

[0082] For example, the window remaining 20 seconds, the memory time window 300 seconds, the operator clicks the temperature panel 5 times, the pH value 3 times, and inputs the TEOS flow 2 times (30, 31 mL / min). The cognitive content reconstruction engine (Pandas implementation) parses the trajectory and infers the cognitive information: “focus on temperature 52 °C and pH value 6.9”, and the gap: “do not understand the effect of TEOS flow on temperature adjustment”. The time cost is 40 ms, and the results are stored in the SQLite database (fields: timestamp, cognitive information, gap). In production, the operator frequently checks due to high temperature, and the engine identifies his focus and gap, and the log records the details to support S67.

[0083] S67, fuse the remaining time and the cognitive gap information, use the reinforcement learning strategy to select a suitable auxiliary cognitive strategy, execute the auxiliary cognitive strategy, and enhance the user's cognition of the joint risk state. The auxiliary cognitive strategy includes: highlighting the risk parameter trend, pushing the historical similar decision case comparison view, and automatically simulating the execution instruction sequence and visualizing the predicted results.

[0084] The reinforcement learning strategy selection model fuses the remaining time (for example, 20 seconds) and the cognitive gap (for example, "not understanding the TEOS flow impact") of S66 to select the optimal auxiliary cognitive strategy. The model is based on Q-learning, the state is (remaining time, gap type, risk value 0.65), and the action includes highlighting (Plotly highlights the temperature curve), pushing historical cases (database queries similar trajectories), and simulating instruction sequences (S3 model predicts trajectories). The reward function is the user's cancellation rate (cancel high-risk instructions +1, keep -1, historical data statistics) and response time (2℃ cases), and simulation is performed by S3 model to predict 30mL / min trajectories. The execution time is <100ms, and the results are stored in the database.

[0085] For example, the remaining time is 20 seconds, the gap is "TEOS flow impact", and the risk value is 0.65. The Q-learning model (NumPy implementation) selects the "highlight" strategy, Plotly highlights the temperature curve (52℃, red), and the time consumption is 50ms. The database queries 3 historical cases (temperature deviation 2.5℃, reduce TEOS to 28mL / min to restore normal), which are displayed through the Web interface. The S3 model simulates a 30mL / min trajectory (temperature rises to 53℃), and the results are presented in a line chart. The execution time is 90ms, and the results are stored in the SQLite database. The operator views the highlighted curve and cases, understands the flow risk, logs the strategy selection and execution details.

[0086] S68, if the user does not perform a cancellation operation on the input preliminary flow control sub-instruction after executing the auxiliary cognitive strategy, the instruction is added to the to-be-implemented instruction set.

[0087] After executing the S67 strategy, the system monitors whether the user cancels the sub-instruction (30mL / min, click event detection of the cancel button in JavaScript). If there is no cancellation within 10 seconds (setTimeout timing), the instruction is added to the to-be-implemented instruction set (SQLite database, fields: timestamp, flow value, material) by a Python script. The addition time is <20ms, and the instruction set can store up to 20 instructions (FIFO strategy). If it is cancelled, the instruction is not added, and the system records the operation log.

[0088] For example, after S67 highlights the temperature curve, the operator does not click the undo button within 10 seconds (30 mL / min). The Python script adds the instruction to the SQLite database (fields: timestamp, TEOs, 30 mL / min), which takes 15 ms. The instruction set is updated, containing 11 instructions. The operator's behavior (no undo) and the addition of time are logged, ensuring traceability. S69, when the user confirms that the decision is complete, a second flow control instruction is generated based on all the prepared flow control sub-instructions in the pending instruction set and sent to the actuator to adjust the corresponding material delivery flow.

[0089] The user clicks the "Confirm" button through the Web interface (JavaScript records), and the system extracts all sub-instructions from the pending instruction set (SQLite database) (e.g., 11, TEOs flow 28-32 mL / min). A second flow control instruction is generated through weighted averaging (weights based on input time, recent instructions weight 0.2, early 0.05) (unit: mL / min, precision ± 0.1 mL / min). The instruction is sent to the PLC through the Modbus protocol to control the metering pump (PID algorithm, Kp = 0.5), adjust the flow, which takes < 50 ms. The adjustment result is fed back to the database (actual flow, deviation).

[0090] For example, the operator clicks "Confirm", and the instruction set contains 11 TEOs sub-instructions (average 30.5 mL / min). The Python script calculates the weighted average (recent instructions 30 mL / min weight 0.2) to generate a second instruction 30.4 mL / min, which is sent to the Siemens S7-1200 PLC through Modbus to adjust the metering pump to 30.3 mL / min (deviation 0.1 mL / min). It takes 40 ms to store in the SQLite database. The temperature returns to 50.2°C, and the instruction generation and adjustment details are logged.

[0091] The technical solution significantly improves the decision-making efficiency and reliability in human-machine collaborative tasks through the construction of a visual decision-making model, real-time hesitation state evaluation, joint risk prediction, and dynamic auxiliary cognitive strategy. Its integrated visual interface intuitively presents the coupling effects of reaction parameter evolution and multi-material flow, reducing the cognitive burden of users; through the precise quantification of user hesitation state by the long short-term memory network, combined with the risk prediction network to evaluate the comprehensive impact of the instruction sequence, ensuring that high-risk instructions are timely intervened within the decision reversible time window; the dynamic auxiliary strategy enhances the user's awareness of potential risks by highlighting risk trends, pushing historical cases, and simulating prediction results, reducing the probability of loss of control. The overall solution effectively improves the stability of the aerogel sol-gel reaction process and the quality of the product, while shortening the decision-making time and improving the efficiency of human-machine collaboration in complex industrial scenarios.

[0092] In particular, step S63 calculates the decision reversible time window through a dynamic window planning algorithm, providing the user with a clear time range to reverse high-risk instructions, significantly reducing the risk of loss of control due to operational errors. The algorithm considers the remaining time, decision progress, and joint risk value, accurately estimating the time required for the system to recover to a safe state through dynamic programming. This allows the user to quickly determine whether to reverse the instruction within the critical time node, avoiding reaction path deviation or product quality deviation.

[0093] Step S64 triggers the constraint condition by defining the hesitation state, accurately screening high-hesitation users who need assistance, avoiding ineffective intervention and improving the targeting of assistance strategies. Based on the mean and variance of the hesitation value sequence in the last 60 seconds, the real-time and stability of the trigger condition are ensured. For example, when the operator frequently checks the parameters due to high temperature, the system accurately identifies his high-hesitation and stable state, triggering subsequent assistance. This mechanism effectively balances the assistance frequency and user autonomy, reducing unnecessary interference while ensuring timely intervention in high-risk scenarios, improving the efficiency of decision support.

[0094] Step S65 detects whether the user's hesitation state meets the trigger constraint in real time within the decision reversible time window, ensuring that the assistance strategy is accurately triggered at critical moments. The system updates the hesitation value every second to quickly respond to changes in user behavior, such as frequent parameter checking due to high temperature. This real-time detection mechanism makes up for the shortcomings of traditional systems in responding to dynamic user behavior, ensuring the timeliness and accuracy of assistance. The storage and logging of detection results also provide traceability for subsequent analysis, enabling the system to effectively support user decision-making in high-dynamic industrial scenarios and reduce the risk of loss of control.

[0095] Step S66 analyzes the operation trajectory and cognitive focus sequence through the cognitive content reconstruction engine, accurately inferring the cognitive information the user has received and the decision logic gap, providing key evidence for personalized assistance. The engine analyzes user behavior based on a rule model, such as frequent clicking on the temperature panel indicating a focus point and multiple instruction modifications reflecting a logic gap. This analysis addresses the shortcomings of traditional systems in lacking deep understanding of user cognitive state, making the assistance strategy more tailored to user needs. For example, after identifying the operator's confusion about the flow adjustment effect, the system can push relevant information, thereby improving the user's understanding of the complex reaction process and decision-making confidence.

[0096] Figure 2 A schematic diagram of an aerogel production flow control system based on online monitoring is provided for the embodiments of the present application, as shown in Figure 2 The system comprises: An online monitoring module 1 for real-time acquisition of state parameters during the sol-gel reaction process of the aerogel sol-gel through an online sensor system; a parameter input module 2 configured to input state parameters to a pre-trained flow control AI model; an AI module 3 configured to cause the flow control AI model to perform the following steps: determine a first flow control instruction required to cause the reaction process to return to a preset optimization path based on a current and historical sequence of state parameters; predict a state parameter trajectory in a future period of time based on the first flow control instruction; determine whether the state parameter trajectory can be effectively stabilized within a tolerance range of the preset optimization path; a first flow control module 4 configured to automatically adjust a delivery flow of at least one of a precursor, a catalyst or a solvent according to the first flow control instruction if the determination result is yes; a man-machine collaboration module 5 configured to trigger generation of a man-machine collaboration task if the determination result is no; a second flow control module 6 configured to assist a user to perform the man-machine collaboration task, generate a second flow control instruction, and adjust a delivery flow of a corresponding material according to the second flow control instruction.

[0097] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for controlling the flow rate of aerogel production based on online monitoring, characterized in that, include: S1: Real-time acquisition of state parameters during the aerogel sol-gel reaction process via an online sensor system; S2: Input the state parameters into the pre-trained flow control AI model; S3: The supply flow control AI model performs the following steps: Based on the current and historical state parameter sequences, the first flow control command required to bring the reaction process back to the preset optimization path is determined. Based on the first flow control command, predict the trajectory of state parameters over a future period of time; Determine whether the trajectory of the state parameters can be effectively stabilized within the tolerance range of the preset optimized path; S4: If the judgment result in step S3 is yes, then the conveying flow rate of at least one material among the precursor, catalyst or solvent is automatically adjusted according to the first flow control command. S5: If the judgment result in step S3 is negative, then trigger the generation of human-machine collaborative task; S6: Assists users in performing human-machine collaborative tasks, generates a second flow control command, and adjusts the conveying flow of the corresponding material according to the second flow control command.

2. The aerogel production flow control method based on online monitoring as described in claim 1, characterized in that, The state parameters include at least three of the following: reactor temperature, system pH value, fluid viscosity, precursor concentration, and gelation time.

3. The aerogel production flow control method based on online monitoring as described in claim 1, characterized in that, The pre-training steps of the flow control AI model include: Collect time-series status parameters and corresponding flow control commands recorded during historical production processes to form a sample set, and perform normalization and sliding window segmentation on the sample set. Based on the processed sample set, a multi-step prediction model with a long short-term memory network as its core is constructed. The state parameter sequence of the previous time period is used as the input of the multi-step prediction model, and the deviation between the state parameter trajectory of the subsequent time period and the preset optimization path is used as the training target. During the training of the multi-step prediction model, a temporal convolutional network is used to extract local features of the input sequence, and an attention mechanism is used to weight and focus on key process stages. Mean squared error and trajectory smoothness are used as the joint loss function, and a trajectory stability evaluation index based on tolerance range is introduced as the model validation standard. The gradient descent algorithm is used to optimize the network parameters of the multi-step prediction model until the multi-step prediction model has the ability to determine the first flow control command based on the current and historical state parameter sequences, predict the future state parameter trajectory, and determine whether the state parameter trajectory can be effectively stabilized within the tolerance range of the preset optimization path, thus obtaining the flow control AI model.

4. The aerogel production flow control method based on online monitoring as described in claim 1, characterized in that, The triggering of the human-machine collaborative task includes: When the flow control AI model determines that the trajectory of the state parameters cannot be stabilized within the tolerance range of the preset optimization path, it generates a collaborative task trigger signal containing deviation parameters and the runaway time window. The trigger signal is sent to the human-computer interaction system to activate the collaborative task generation engine within it; Through the collaborative task generation engine, based on deviation parameters and runaway time windows, a human-machine collaborative task is constructed, which includes the type of material to be adjusted, the target flow range, and the adjustment urgency level.

5. The aerogel production flow control method based on online monitoring as described in claim 1, characterized in that, The assistance provided to users in performing human-machine collaborative tasks includes: Based on the material type to be adjusted, target flow range, and adjustment urgency level defined in the human-machine collaborative task, a visual decision-making model is constructed that integrates a real-time status visualization window of the reaction process, a historical optimized path comparison view, and a multi-material flow collaborative control interface. Whenever a user inputs a preliminary flow control sub-instruction based on the aforementioned visual decision model, the following operations are executed in parallel: Real-time tracking of the user's hesitation state in response to the prepared flow control sub-command; The joint risk status of the prepared flow control sub-instruction if it is actually executed is evaluated synchronously in real time. The evaluation method of the joint risk status is as follows: extract the currently input prepared flow control sub-instruction and the historical prepared sub-instructions stored in the instruction set to be implemented to construct an instruction sequence; input the instruction sequence into a risk prediction network trained with reaction kinetic data to calculate its comprehensive impact index on the aerogel sol-gel reaction process; the comprehensive impact index includes: expected value of reaction path deviation, increase ratio of standard deviation of gel time, and deviation of product porosity; and generate a joint risk value based on the weighted fusion of the comprehensive impact index. When the combined risk value exceeds the preset risk threshold, the decision reversible time window is calculated using a dynamic window planning algorithm based on the remaining time from the current time point to the uncontrollable time window defined by the human-machine collaborative task, the decision progress reflected by the user's input instructions, and the magnitude of the combined risk value. The decision reversible time window is the maximum allowable period during which the user can revoke the input preparatory sub-instructions and the system has sufficient time to recover to the state before the decision. Generate hesitant state triggering constraints for the aforementioned decision reversible time window; When entering the decision reversible time window, it is detected in real time whether the user's current hesitation state meets the triggering constraint condition; If satisfied, based on the remaining duration of the decision reversible time window, the user's operation trajectory and cognitive focus sequence in the visual decision model within the previously set memory time window, the cognitive content reconstruction engine infers the cognitive information fragments that the user has received and the gaps in their decision logic. By integrating information on remaining time and cognitive gap, a reinforcement learning strategy is used to select a suitable auxiliary cognitive strategy to output from the model. The aforementioned cognitive assistance strategy enhances the user's awareness of the joint risk status. If the user does not cancel the input of the pre-entered flow control sub-command after the auxiliary cognitive strategy is executed, the command will be added to the set of commands to be implemented. When the user confirms the completion of the decision, a second flow control instruction is generated based on all the prepared flow control sub-instructions in the instruction set to be implemented, and sent to the actuator to adjust the corresponding material conveying flow.

6. The aerogel production flow control method based on online monitoring as described in claim 5, characterized in that, The visualization decision-making model dynamically renders the deviation area between the reaction parameter evolution curve and the preset optimization path, and displays the coupling influence coefficient of different material flow rate adjustments to help users form a multi-dimensional decision-making understanding.

7. The aerogel production flow control method based on online monitoring as described in claim 5, characterized in that, The hesitation state is determined by fusing a first behavioral indicator before user input and a second behavioral indicator after input. The first behavioral indicator includes: the frequency and duration of viewing relevant parameter panels in the visual decision model, the number of historical command modifications, and the deviation between the current input command and the system-recommended command. The second behavioral indicator includes: the number of times the relevant parameters of the command are viewed again after the command is input, the triggering delay of the command modification operation, and the jitter characteristics of the mouse focus on the command input area. Through a hesitation evaluation model based on a long short-term memory network, the first and second behavioral indicators are time-series modeled to output the quantitative value of the current hesitation state.

8. The aerogel production flow control method based on online monitoring as described in claim 7, characterized in that, The triggering constraint is used to determine whether to provide risk perception assistance to the user within the current window. It is defined as follows: the user's hesitancy state quantification value is continuously higher than the hesitancy threshold in the most recent continuous period, and the hesitancy fluctuation variance is lower than the stability threshold.

9. The aerogel production flow control method based on online monitoring as described in claim 5, characterized in that, The auxiliary cognitive strategies include: highlighting the changing trends of risk parameters, pushing a comparison view of similar historical decision cases, automatically simulating the execution of instruction sequences and visualizing the prediction results.

10. An aerogel production flow control system based on online monitoring, characterized in that, include: The online monitoring module is used to collect state parameters in real time during the aerogel sol-gel reaction process through an online sensor system; The parameter input module is used to input state parameters into the pre-trained flow control AI model; The artificial intelligence module is used to enable the flow control AI model to perform the following steps: Based on the current and historical state parameter sequences, the first flow control command required to bring the reaction process back to the preset optimization path is determined. Based on the first flow control command, predict the trajectory of state parameters over a future period of time; Determine whether the trajectory of the state parameters can be effectively stabilized within the tolerance range of the preset optimized path; The first flow control module is used to automatically adjust the flow rate of at least one of the precursor, catalyst or solvent according to the first flow control command if the judgment result is yes. The human-machine collaboration module is used to trigger the generation of human-machine collaboration tasks if the judgment result is negative. The second flow control module is used to assist users in performing human-machine collaborative tasks, generate second flow control commands, and adjust the conveying flow of corresponding materials according to the second flow control commands.

Images