An automatic control method for exhaust of an integrated experiment table

Abstract

This invention provides an automatic exhaust control method for an integrated experimental bench, belonging to the technical field of automatic exhaust control for integrated experimental benches. This invention collects wind speed sensor signals from each operating surface and extracts effective face wind speed signals through sliding median window filtering and wavelet decomposition filtering. It employs a model predictive control framework to uniformly calculate the valve opening commands of each variable air volume valve. Based on the Mahalanobis distance between the indoor transient pressure difference and the standard pressure difference vector, it adaptively adjusts the advance compensation timing of the exhaust-makeup air coordinated predictive feedforward controller. It utilizes an artificial intelligence-based dynamic exhaust control optimization model to dynamically allocate computational resources to each control thread. A variational expectation propagation sparse decoupling algorithm is used to update the pipeline network sensitivity response matrix online to achieve model parameter self-correction. Furthermore, it updates the compensation gain parameters in real time based on the pressure difference dynamic adjustment function value. This solves the technical problem of beat frequency oscillations in the face wind speed control system caused by pipeline static pressure coupling during the coordinated adjustment of multiple exhaust hoods.

Description

Technical Field

[0001] This invention belongs to the field of automatic exhaust control technology for integrated experimental benches, and more specifically, relates to an automatic exhaust control method for integrated experimental benches. Background Technology

[0002] Integrated laboratory bench exhaust control systems are core equipment for safe ventilation in laboratories, used to maintain the air velocity in the workbench space above a safe threshold, protecting laboratory personnel from harmful gases. In existing technologies, each exhaust hood is equipped with an independent variable air volume valve and face velocity sensor, achieving single-loop closed-loop regulation through a proportional-integral-derivative controller or a simple feedforward controller. Exhaust-makeup air coordination relies on a fixed delay strategy, and pipeline parameters are obtained through offline calibration and then used consistently.

[0003] In existing technologies, when multiple exhaust hood openings change their openings simultaneously, the variable air volume valves share the same exhaust main pipe. The action of a single valve causes fluctuations in the static pressure of the main pipe, which in turn affects the actual airflow of the remaining exhaust hood openings. Because the adjustment cycles of each independent control loop differ, their adjustment actions are superimposed in time, forming low-frequency periodic fluctuations with a frequency much lower than the adjustment frequency of a single loop—a phenomenon known as beat frequency oscillation. Existing single-loop control architectures cannot perceive the global coupling state of the pipeline network. Each loop outputs adjustment commands independently based on its own deviation, failing to predict the impact of other loop actions on its own loop, nor to coordinate the adjustment sequence of valves at the global level. This results in beat frequency oscillations that are difficult to converge, leading to long-term deviations in face velocity from the set value. In other words, existing technologies suffer from the technical problem of beat frequency oscillations in the face velocity control system caused by pipeline network static pressure coupling during the coordinated adjustment of multiple exhaust hood openings. Summary of the Invention

[0004] In view of this, the present invention provides an automatic exhaust control method for an integrated experimental platform, which can solve the technical problem in the prior art where static pressure coupling of the pipeline network causes beat frequency oscillations in the face wind speed control system when multiple exhaust hoods are linked for adjustment.

[0005] This invention is implemented as follows: This invention provides an automatic exhaust control method for an integrated experimental bench, comprising the following steps:

[0006] The face wind speed sensor signals of each exhaust hood are collected. The face wind speed sensor signals are then subjected to sliding median window filtering and wavelet decomposition filtering in sequence. The low-frequency components whose duration exceeds the effective duration threshold are extracted as the effective face wind speed signals.

[0007] Using the effective surface wind speed signal as input, the valve opening command of each variable air volume valve is uniformly calculated through the pipeline static pressure-air volume coupled state space model and the model predictive control framework, and the valve opening command is sent to each variable air volume valve for execution.

[0008] The system collects the indoor transient pressure difference, calculates the Mahalanobis distance between the indoor transient pressure difference and the standard vector of the pressure difference, adjusts the advance compensation sequence of the exhaust-makeup air coordination prediction feedforward controller according to the interval to which the Mahalanobis distance belongs, and drives the makeup air valve to act in advance according to the advance compensation sequence.

[0009] The effective face velocity signal of each exhaust hood opening, the indoor transient pressure difference, and the valve position feedback signal of each variable air volume valve are input into the exhaust dynamic control optimization model. The exhaust dynamic control optimization model outputs a resource allocation weight vector. Based on the resource allocation weight vector, the memory allocation ratio and control cycle duration of the corresponding control thread of each variable air volume valve are adjusted.

[0010] The variational expectation propagation sparse decoupling algorithm is used to decouple the surface wind speed sensor signal and the valve position feedback signal online, update the pipeline network sensitivity response matrix, and substitute the decoupled air volume estimates of each exhaust hood inlet back into the pipeline network static pressure-air volume coupled state space model to achieve self-correction of model parameters.

[0011] The system calculates the dynamic adjustment function value of the differential pressure in real time, selects the corresponding compensation gain parameter based on the range to which the dynamic adjustment function value of the differential pressure belongs, updates the compensation gain parameter to the exhaust-makeup air coordinated prediction feedforward controller, and returns to the step of collecting the space wind speed sensor signal of each operating platform after completing one control cycle.

[0012] The sliding median window filter has a window length of 1 to 5 seconds. The wavelet decomposition filter decomposes the surface wind speed sensor signal into a low-frequency sub-band and a high-frequency sub-band, retaining only the components in the low-frequency sub-band whose duration exceeds the effective duration threshold.

[0013] The effective duration threshold is determined iteratively through no less than 30 personnel operation disturbance experiments, with the goal of achieving the optimal weighted comprehensive balance between false trigger rate and false detection rate, and the value range is 3 to 5 seconds.

[0014] The pipeline static pressure-air volume coupled state space model uses the valve opening command of each variable air volume valve as the control input, and the estimated values ​​of static pressure of each pipeline node and air volume of each exhaust hood outlet as the state variables. A multi-input multi-output linear state space equation is established, and the coefficient matrix is ​​identified by the step response method. The model is iterated for no less than 5 rounds until the residual converges.

[0015] The model predictive control framework uses the current static pressure of the pipeline node and the estimated air volume of each exhaust hood as initial conditions in each control cycle, predicts forward 3 to 10 control steps, and uses the minimum sum of squared deviations of the space wind speed of all operating platforms as the objective function to uniformly optimize the valve opening commands of each variable air volume valve.

[0016] The pressure difference standard vector is obtained by continuously collecting indoor transient pressure difference time-series data for no less than 7 days, and statistically processing the data to obtain a vector composed of the mean and standard deviation of the pressure difference for each time period; when the Mahalanobis distance is less than 1, the lead compensation time series is not adjusted; when the Mahalanobis distance is within the interval... When the advance compensation timing is shortened by 5 to 15 seconds, when the Mahalanobis distance is greater than or equal to 3, the advance compensation timing is advanced to the moment the exhaust volume change command is issued, and the Smith predictor is superimposed to compensate for the pure lag.

[0017] The Smith predictor is based on the transfer function of the make-up air system. The transfer function of the make-up air system is identified by the step response experiment of the make-up air valve. The experiment is repeated no less than 10 times and the average value is taken. The stroke time of the make-up air valve is 60-120s and the equivalent time constant of the temperature regulation heat exchange is 30-90s.

[0018] The exhaust dynamic control optimization model includes an input layer, an encoding layer, a backbone network, an unscented Kalman filter, and an output layer. The backbone network consists of three hidden layers of leaky integral firing neurons, each containing 256 leaky integral firing neurons. The synaptic weights are updated online using pulse temporal dependence plasticity rules.

[0019] Among them, the exhaust dynamic control optimization model sets up an emergency jump channel. When the cumulative pulse firing rate of any leaky integral firing neuron exceeds the emergency trigger threshold within 10ms, the bypass direct connection layer is activated to directly output the emergency regulation code and skip the regular classification layer. The emergency trigger threshold ranges from 0.6 to 0.9.

[0020] The output layer takes the Bayesian optimal estimation result and the emergency adjustment code as input, and outputs the resource allocation weight vector in parallel with the traditional spectrum analysis diagnosis result. The final resource allocation weight vector is obtained by fusing decision through Dempster-Shafer evidence theory.

[0021] The variational expectation propagation sparse decoupling algorithm introduces a Laplace sparse prior to the observation matrix composed of the wind speed sensor signal and the valve position feedback signal, and introduces a matrix normal prior to the pipeline network sensitivity response matrix. The variational inference framework is used to transform the posterior approximation problem into minimizing the KL divergence. The expectation propagation algorithm iteratively updates the message node by node on the factor graph.

[0022] The variational expectation propagation sparse decoupling algorithm updates the posterior in batches using a sliding window in online mode, only locally refreshing the factor graph nodes affected by the new data, with a global recalculation frequency of once per minute.

[0023] Wherein, the differential pressure dynamic adjustment function value The result is obtained by weighted summation of three factors: the normalized value of the indoor transient pressure difference, the normalized value of the current opening of the make-up air valve, and the normalized value of the current advance compensation timing of the exhaust-make-up air coordination prediction feedforward controller. The sum of the weight coefficients is 1, and the result is determined by at least 20 rounds of on-site debugging experiments.

[0024] Among them, when When the compensation gain parameter is increased to 1.2 to 1.5 times its current value; when When the compensation gain parameter remains constant; when When the compensation gain parameter is reduced to 0.7 to 0.9 times its current value; when The compensation gain parameter is reset to the initial calibration value and the pipeline static pressure-air volume coupled state space model is triggered to re-identify the transfer function of the make-up air system.

[0025] The training dataset of the exhaust dynamic control optimization model has a total collection time of no less than 720 hours, covering scenarios of disturbance of a single exhaust hood, simultaneous disturbance of multiple exhaust hoods, and delayed response of the make-up air valve. The training set, validation set, and test set are divided in an 8:1:1 ratio. Training is stopped when the control effect score of the validation set no longer improves after 10 consecutive rounds.

[0026] This invention establishes a state-space model of pipeline static pressure-airflow coupling, incorporates the adjustment actions of each variable airflow valve into a unified model predictive control framework, and proactively predicts global pressure and airflow changes in each control cycle. With the objective function of minimizing the sum of squared wind speed deviations across the entire operating platform space, it uniformly optimizes the valve opening commands of each variable airflow valve, thus solving the technical problem of beat frequency oscillation caused by pipeline static pressure coupling when multiple exhaust hoods are linked for adjustment.

[0027] This invention utilizes a coupled state-space model of pipeline static pressure and airflow to accurately describe the impact of each valve's actions on the static pressure of the global pipeline nodes. This allows the model's predictive control framework to incorporate the pipeline coupling effect as a constraint in the optimization when uniformly solving for the opening of each valve, rather than superimposing the results after independent calculations of each loop. The adjustment actions of each variable airflow valve are globally coordinated in terms of timing and amplitude, eliminating the phase superposition conditions caused by differences in the timing of single-loop adjustments, thereby cutting off the formation mechanism of beat frequency oscillations.

[0028] In summary, the present invention solves the technical problem mentioned in the background art of beat frequency oscillation in the face wind speed control system caused by static pressure coupling of the pipeline network during the linkage adjustment of multiple exhaust hoods. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 The graph shows the convergence process of the spatial wind speed deviation of the six operating tables over time.

[0031] Figure 3 The dynamic adjustment function value of differential pressure within a typical control cycle Curve showing changes over time.

[0032] Figure 4 The graph shows the static pressure dynamic response curves of the six pipeline nodes after the disturbance.

[0033] Figure 5 This is a schematic diagram of the first type of test bench in Example 2.

[0034] Figure 6 This is a second schematic diagram of the test bench in Example 2. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0036] like Figure 1 The diagram shown is a flowchart of an automatic exhaust control method for an integrated experimental bench provided by the present invention. This method includes the following steps:

[0037] S01. Collect the face wind speed sensor signals from each exhaust hood opening, and sequentially perform sliding median window filtering and wavelet decomposition filtering on the face wind speed sensor signals to extract the low-frequency components whose duration exceeds the effective duration threshold as the effective face wind speed signals.

[0038] S02. Using the effective surface wind speed signal as input, the valve opening command of each variable air volume valve is uniformly calculated through the pipeline static pressure-air volume coupled state space model and the model predictive control framework, and the valve opening command is sent to each variable air volume valve for execution.

[0039] S03. Collect the indoor transient pressure difference, calculate the Mahalanobis distance between the indoor transient pressure difference and the standard vector of the pressure difference, adjust the advance compensation sequence of the exhaust-makeup air coordination prediction feedforward controller according to the interval to which the Mahalanobis distance belongs, and drive the makeup air valve to act in advance according to the advance compensation sequence.

[0040] S04. Input the effective face velocity signal of each exhaust hood opening, the indoor transient pressure difference, and the valve position feedback signal of each variable air volume valve into the exhaust dynamic control optimization model. The exhaust dynamic control optimization model outputs a resource allocation weight vector. Based on the resource allocation weight vector, adjust the memory allocation ratio and control cycle duration of the control thread corresponding to each variable air volume valve.

[0041] S05. The variational expectation propagation sparse decoupling algorithm is used to decouple the surface wind speed sensor signal and the valve position feedback signal online, update the pipeline network sensitivity response matrix, and substitute the decoupled air volume estimates of each exhaust hood inlet back into the pipeline static pressure-air volume coupled state space model of step S02 to realize the self-correction of model parameters.

[0042] S06. Calculate the differential pressure dynamic adjustment function value in real time, select the corresponding compensation gain parameter according to the interval to which the differential pressure dynamic adjustment function value belongs, update the compensation gain parameter to the exhaust-makeup air coordinated prediction feedforward controller, complete one control cycle, and return to step S01.

[0043] The sliding median window filter has a window length of 1–5 s to suppress isolated pulse interference caused by personnel moving in and out of the room or by surrounding airflow disturbances. The wavelet decomposition filter decomposes the face wind speed sensor signal into low-frequency and high-frequency sub-bands, retaining only the components in the low-frequency sub-band whose duration exceeds the effective duration threshold as the effective face wind speed signal. The effective duration threshold is obtained by conducting no less than 30 personnel operation disturbance experiments during the laboratory field debugging phase, collecting the original face wind speed sensor signal, and using the weighted comprehensive optimization of false trigger rate and false detection rate as the objective, iteratively analyzing and determining the effective duration threshold, with a value range of 3–5 s.

[0044] The pipeline static pressure-airflow coupled state-space model uses the valve opening command of each variable air volume valve as the control input and the estimated static pressure of each pipeline node and the airflow of each exhaust hood outlet as the state variables to establish a multi-input multi-output linear state-space equation. The coefficient matrix of the state-space equation is obtained by identifying the actual pipeline topology and airflow test data in the laboratory. The identification method is the step response method, which applies a step excitation to each variable air volume valve in sequence, records the response of all pipeline nodes, and iterates for no less than 5 rounds until the residual converges.

[0045] The model predictive control framework uses the current static pressure of the pipeline node and the estimated air volume of each exhaust hood as initial conditions in each control cycle. It predicts the global pressure and air volume changes 3 to 10 control steps forward. The objective function is to minimize the sum of squared deviations of the air velocity in the space of all operating platforms, and to uniformly optimize the valve opening commands of each variable air volume valve. The specific values ​​of the prediction step and sampling cycle are determined by the pipeline time constant. After measuring the adjustment time of each pipeline node through field step response experiments, the values ​​are determined based on the principle of covering the longest adjustment time. The model predictive control framework unifies and coordinates the adjustment actions of all variable air volume valves through global look-ahead optimization, eliminates the pipeline static pressure coupling interference caused by the shared exhaust duct of each exhaust hood, and fundamentally suppresses the beat frequency oscillation generated when multiple exhaust hoods are linked for adjustment.

[0046] The pressure difference standard vector is obtained as follows: Under normal laboratory conditions, continuous indoor transient pressure difference time-series data are collected for no less than 7 days. After statistical processing, a vector composed of the mean and standard deviation of the pressure difference for each time period is obtained as the pressure difference standard vector. The Mahalanobis distance is measured by the normalized deviation between the current indoor transient pressure difference and the pressure difference standard vector, eliminating the influence of inconsistencies in dimensions and dimensional correlations on distance calculation. When the Mahalanobis distance is less than 1, the indoor transient pressure difference is considered normal, and the advance compensation time series is not adjusted. When the Mahalanobis distance falls within an interval... When the time is 5-15 seconds, the advance compensation timing will be shortened; when the Mahalanobis distance is greater than or equal to 3, the advance compensation timing will be advanced to the moment the exhaust volume change command is issued, and the Smith predictor will be superimposed to compensate for the pure time delay; the Smith predictor is based on the transfer function of the make-up air system, which is identified by the step response experiment of the make-up air valve, and the experiment is repeated no less than 10 times and the average value is taken; the stroke time of the make-up air valve is 60-120 seconds, and the equivalent time constant of the temperature regulation heat exchange is 30-90 seconds. All of the above parameters are determined by the on-site step response experiment.

[0047] The specific structure of the exhaust dynamic control optimization model is as follows: The exhaust dynamic control optimization model uses a multi-dimensional time-series vector composed of the normalized values ​​of the effective surface wind speed signals of each exhaust hood opening, the normalized values ​​of the indoor transient pressure difference, and the normalized values ​​of the valve position feedback signals of each variable air volume valve as input; the input layer converts the multi-dimensional time-series vector into a sparse pulse sequence through Delta modulation; the encoding layer maps the sparse pulse sequence to a high-dimensional pulse feature space; the backbone network consists of three layers of hidden layers of leaky integral firing neurons, each layer containing 256 leaky integral firing neurons, and the synaptic weights are updated online using the pulse time-dependent plasticity rule; the pulse firing rate vector output by the hidden layer of the third layer of leaky integral firing neurons is input into an unscented Kalman filter, which uses the pulse firing rate vector as an observation and the predicted value of the pipeline static pressure-air volume coupled state space model as the state transition equation to achieve Bayesian optimal estimation of the control state; the network is equipped with an emergency jump channel, which is triggered when any leaky integral firing neuron... When the cumulative pulse firing rate of a variable air volume valve exceeds the emergency trigger threshold within 10ms, the bypass direct connection layer is activated to directly output the emergency adjustment code, skipping the conventional classification layer. The output layer takes the Bayesian optimal estimation result and the emergency adjustment code as input and outputs the resource allocation weight vector of each variable air volume valve. The resource allocation weight vector is parallel to the traditional spectrum analysis diagnosis result and is fused with the Dempster-Shafer evidence theory to obtain the final resource allocation weight vector. Each element in the resource allocation weight vector is linearly positively correlated with the memory allocation ratio of the corresponding variable air volume valve control thread and linearly negatively correlated with the control cycle duration. The emergency trigger threshold is obtained by applying different degrees of abrupt disturbances to each exhaust hood in the laboratory, recording the correspondence between the cumulative pulse firing rate of the leaked integral firing neuron and the control response effect, aiming at the weighted comprehensive optimality of the false trigger rate and the leaked trigger rate, and iterating for no less than 20 experiments to determine the value range of 0.6 to 0.9 (normalized).

[0048] The steps for establishing the training dataset for the exhaust dynamic control optimization model specifically include: applying step excitation with different valve opening combinations to each exhaust hood port on the actual laboratory pipeline network, simultaneously recording multi-dimensional time-series data such as face velocity sensor signals, static pressure at each pipeline node, indoor transient pressure difference, and valve position feedback signals of each variable air volume valve, collecting operating conditions covering single exhaust hood port disturbance, simultaneous disturbance of multiple exhaust hood ports, and delayed response scenarios of the make-up air valve, with a total collection time of no less than 720 hours; cleaning the collected data to remove sensor fault segments and communication packet loss segments; using manually labeled control effect scores as supervision labels, the control effect scores being composed of a weighted average of face velocity deviation, indoor transient pressure difference deviation, and adjustment time; and dividing the cleaned data into training set, validation set, and test set in an 8:1:1 ratio.

[0049] The specific steps for training the exhaust dynamic control optimization model include: initializing the backbone network synaptic weights to random small values ​​and initializing the covariance matrix of the unscented Kalman filter to an identity matrix; inputting the training set data in batches according to time sequence, updating the backbone network synaptic weights online iteratively through the impulse time-dependent plasticity rule, and updating the unscented Kalman filter parameters once per round through maximum likelihood estimation; using the control effect score on the validation set as the early stopping criterion, stopping training when the control effect score on the validation set no longer improves after 10 consecutive rounds; evaluating on the test set after training, and if the control effect score on the test set is lower than 95% of the control effect score on the validation set, then expanding the training data and retraining.

[0050] The technical advantages of the proposed exhaust dynamic control optimization model are as follows: The time-coded pulse neural network expresses surface wind speed sensor signal events in a sparse pulse manner, generating calculations only when the signal changes, resulting in extremely low power consumption during static periods, making it suitable for long-term operation on embedded platforms of laboratory controllers; the pulse timing-dependent plasticity rule adjusts synaptic weights in real time based on the temporal differences in pulse firing by neurons before and after leakage integration, enabling the exhaust dynamic control optimization model to continuously adapt to network characteristic drift during operation without offline retraining; the unscented Kalman filter integrates the nonlinear feature mapping of the backbone network with the predicted values ​​of the network static pressure-airflow coupled state-space model, maintaining the uncertainty estimate of the control state through a Bayesian framework, thus ensuring statistical optimality in the calculation of the resource allocation weight vector; the emergency jump channel ensures real-time response under high-frequency disturbance scenarios, and its integration with Dempster-Shafer evidence theory jointly guarantees the robustness of the output resource allocation weight vector.

[0051] The variational expectation propagation sparse decoupling algorithm establishes a linear hybrid model for the multi-sensor system. It uses the observation matrix composed of surface wind speed sensor signals and valve position feedback signals as input, introduces a Laplace sparse prior for the estimated airflow vectors at each exhaust hood opening, and a matrix normal prior for the uncertainty of the pipeline sensitivity response matrix. A variational inference framework is used to transform the posterior approximation problem into minimizing the KL divergence. The expectation propagation algorithm iteratively updates messages node by node on the factor graph, with local computations at each node executed in parallel. In online mode, the posterior is updated in batches using a sliding window, only locally refreshing factor graph nodes affected by new data, with a global recalculation frequency of once per minute. The decoupled airflow estimates of each exhaust hood opening are used as the output of step S05 and substituted back into the static pressure-airflow coupled state space model of the pipeline network. The technical effect of the variational expectation propagation sparse decoupling algorithm is as follows: through the Laplace sparse prior constraint, the airflow estimates of each exhaust hood opening still converge to the sparse solution when the number of sensors is less than the theoretically complete number, avoiding the underdetermined problem that leads to non-unique solutions. The factor graph local refresh mechanism of expectation propagation makes the computational amount of online updates sublinearly related to the pipeline network scale, ensuring that the embedded controller completes the decoupling calculation within the real-time control cycle, while maintaining the Bayesian confidence interval of the airflow estimates of each exhaust hood opening to quantify the decoupling uncertainty.

[0052] The differential pressure dynamic adjustment function takes the normalized value of the indoor transient differential pressure in the current control cycle, the normalized value of the current opening degree of the make-up air valve, and the normalized value of the current advance compensation timing of the exhaust-make-up air coordination predictive feedforward controller as inputs. The weighted sum of these three values ​​yields the differential pressure dynamic adjustment function value, denoted as . The formula is as follows:

[0053] ;

[0054] in The unit is the indoor transient pressure difference, expressed in Pa. This is the rated differential pressure, in Pa. The current opening degree of the air supply valve, in degrees. The rated opening of the make-up air valve is expressed in degrees (°). For the current advance compensation time series, the unit is s. The time series advance compensation rating is expressed in seconds. , , These are weighting coefficients, the sum of which is 1. Their specific values ​​are determined through on-site commissioning experiments by iterating at least 20 times to minimize the indoor transient pressure difference deviation and the weighted index of the adjustment time. When this occurs, the compensation gain parameter will be increased to 1.2 to 1.5 times its current value; when... When, keep the compensation gain parameter unchanged; when When this occurs, the compensation gain parameter will be reduced to 0.7 to 0.9 times its current value; when When the compensation gain parameter is reset to the initial calibration value, the pipeline static pressure-air volume coupled state space model is triggered to re-identify the transfer function of the make-up air system; the compensation gain parameter is the multiplicative coefficient used to adjust the advance compensation amplitude in the exhaust-make-up air coordinated prediction feedforward controller.

[0055] Among them, beat frequency oscillation is a low-frequency oscillation phenomenon formed by the superposition of the regulation cycles of multiple regulation loops. It is manifested as the face wind speed deviation fluctuating periodically at a frequency much lower than the regulation frequency of a single loop.

[0056] Among them, the leak-integral firing neuron is a spiking neuron model that simulates the accumulation and leakage behavior of membrane potential in biological neurons. The membrane potential accumulates with the input pulse and decays exponentially with the time constant. When the membrane potential exceeds the threshold, a pulse is fired and the neuron is reset.

[0057] Among them, the pulse timing-dependent plasticity rule is an online learning rule that adjusts synaptic weights based on the time difference between the firing pulses of the preceding and following omission integral firing neurons. When the preceding omission integral firing neuron fires before the following omission integral firing neuron, the weight is enhanced, and vice versa. No global error signal or backpropagation is required.

[0058] Delta modulation is an encoding method that converts continuous analog signals into sparse pulse sequences. It generates positive or negative pulses only when the signal amplitude changes beyond the Delta modulation threshold, and does not generate pulses during static segments. The Delta modulation threshold is determined by the sensor range and the real-time computing capability of the embedded platform. It is determined through field calibration experiments and ranges from 1% to 5% of the sensor's full scale.

[0059] Among them, the Dempster-Shafer evidence theory is a multi-source information fusion framework that represents diagnostic results from different sources as trust functions and fuses them into a unified decision probability through the Dempster merging rule, thus handling uncertainty and conflicting evidence.

[0060] Among them, the Smith predictor is a feedforward compensation structure designed for pure time delay processes. By connecting the predicted output of the transfer function of the air supply system in series in the control loop, the pure time delay is eliminated from the closed-loop characteristic equation, so that the design of the exhaust-air supply coordinated predictive feedforward controller only needs to be carried out for the time-free part.

[0061] Here, the KL divergence is the Kullback-Leibler divergence, used to measure the difference between two probability distributions; the variational inference framework achieves posterior approximation by minimizing the KL divergence between the approximate posterior distribution and the true posterior distribution.

[0062] Mahalanobis distance is a statistical distance metric that takes into account the variance of each dimension and the correlation between dimensions. It calculates the normalized deviation between the current indoor transient pressure difference vector and the pressure difference standard vector, eliminating the influence of dimensional inconsistency and dimensional correlation on distance calculation.

[0063] Among them, the pipeline network sensitivity response matrix is ​​a matrix describing the influence of the valve opening change of each variable air volume valve on the air volume of each pipeline network node. The matrix elements are updated online by the variational expectation propagation sparse decoupling algorithm.

[0064] Among them, the unscented Kalman filter is a Bayesian filter for handling nonlinear state estimation problems. It avoids the linearization error of the extended Kalman filter by using Sigma point sampling to approximate the probability distribution of the nonlinear function and maintains the mean and covariance matrix of the control state.

[0065] Optionally, the present invention also provides a computer-based method for forming an integrated experimental bench exhaust automatic control system, wherein the computer is provided with a readable storage medium storing program instructions, and the program instructions can execute the above-described method when the computer is run.

[0066] The specific implementation of step S01 is as follows: Analog voltage signals output by face velocity sensors installed at each exhaust hood opening are collected and discretized with a fixed sampling period before entering the signal preprocessing flow. First, a sliding median window filter is applied to the original discrete signal. The window length is set to the number of sampling points corresponding to 1–5 seconds. The sliding median window filter replaces the current time value with the median of all sampled values ​​within the window. Its statistical characteristics ensure that isolated abnormal pulses cannot affect the output, making it suitable for suppressing short-term impact interference caused by personnel's arms entering or exiting the exhaust hood opening or surrounding airflow disturbances. The signal after median filtering enters wavelet decomposition filtering. A wavelet basis of appropriate order is selected to perform multi-resolution decomposition of the signal, separating the low-frequency subband and the high-frequency subband. The fast-fluctuation components corresponding to the high-frequency subband are directly discarded. The duration of each segment of the low-frequency subband is detected. Components whose duration exceeds the effective duration threshold are retained. This threshold is determined iteratively through no less than 30 personnel operation disturbance experiments, with the goal of achieving the optimal weighted comprehensive result of false trigger rate and false detection rate. The reference value range is 3–5 seconds. The retained low-frequency continuous components are used as the effective face velocity signal and output to subsequent steps.

[0067] The specific implementation of step S02 is as follows: The pipeline static pressure-airflow coupled state-space model uses the valve opening commands of each variable airflow valve as the control input vector, and the vector composed of the static pressure of each pipeline node and the estimated airflow of each exhaust hood outlet as the state variable, to establish a multi-input multi-output linear state-space equation. The coefficient matrix is ​​obtained by the on-site step response identification method. Step excitation is applied to each variable airflow valve in sequence, and the static pressure response curves of all pipeline nodes are recorded synchronously. The coefficient matrix elements are extracted by least squares fitting, and the iteration is repeated for no less than 5 rounds until the residual converges. At the beginning of each control cycle, the model predictive control framework uses the current static pressure of the pipeline node and the estimated airflow of each exhaust hood outlet as the initial conditions, and uses the state-space equation to recursively extrapolate 3 to 10 control steps to predict the dynamic evolution trajectory of global pressure and airflow. With the minimum sum of squared deviations of the airflow velocity in all operating platform spaces as the objective function, a quadratic programming problem is constructed and solved to obtain the optimal valve opening command of each variable airflow valve at the current moment. The valve opening commands are synchronously sent to each variable air volume valve actuator via the communication network to achieve global coordinated adjustment and eliminate the static pressure coupling interference of the pipeline network caused by the shared exhaust duct of each exhaust hood.

[0068] The specific implementation of step S03 is as follows: Under normal laboratory operating conditions, continuously collect indoor transient pressure difference time-series data for no less than 7 days. Statistical processing is performed to obtain the mean and standard deviation of the pressure difference for each time period, forming a pressure difference standard vector. After real-time acquisition of indoor transient pressure difference, calculate the Mahalanobis distance between the current pressure difference vector and the pressure difference standard vector. The Mahalanobis distance is normalized through the covariance matrix of the pressure difference standard vector, eliminating the influence of inconsistent dimensions and dimensional correlation. Based on the interval to which the Mahalanobis distance belongs, adjust the advance compensation timing of the exhaust-makeup air coordinated prediction feedforward controller in three levels: when the Mahalanobis distance is less than 1, the pressure difference is considered normal, and no timing adjustment is made; when the Mahalanobis distance is less than 1, the pressure difference is considered normal, and no timing adjustment is made; when the Mahalanobis distance is less than 1, the pressure difference is considered normal, and no timing adjustment is made; when the Mahalanobis distance is less than 1, the pressure difference is considered normal, and no timing adjustment is made; when the Mahalanobis distance is less than 1, the pressure difference is considered normal, and no timing adjustment is made. When the advance compensation timing is shortened by 5-15s, and when the Mahalanobis distance is greater than or equal to 3, the advance compensation timing is advanced to the moment the exhaust volume change command is issued, and the Smith predictor is superimposed to compensate for the pure time delay of the make-up air system. The Smith predictor is based on the make-up air system transfer function, which is identified by the make-up air valve step response experiment. The experiment is repeated no less than 10 times and the average value is taken. The reference range for the make-up air valve stroke time is 60-120s, and the reference range for the equivalent time constant of temperature regulation heat exchange is 30-90s.

[0069] The specific implementation of step S04 is as follows: The exhaust dynamic control optimization model receives the normalized values ​​of the effective face wind speed signals of each exhaust hood opening, the normalized values ​​of the indoor transient pressure difference, and the normalized values ​​of the valve position feedback signals of each variable air volume valve. These are combined into a multi-dimensional time-series vector and then converted into a sparse pulse sequence via Delta modulation. The Delta modulation threshold is set to 1%–5% of the sensor's full-scale range, and positive or negative pulses are generated only when the signal amplitude changes beyond the threshold. The encoding layer maps the sparse pulse sequence to a high-dimensional pulse feature space. The backbone network consists of three hidden layers with 256 leaky integral firing neurons per layer. The synaptic weights are updated online using pulse time-dependent plasticity rules, eliminating the need for a global error signal. The pulse firing rate vector output from the third hidden layer is input to an unscented Kalman filter. The unscented Kalman filter uses the predicted values ​​from the pipe network static pressure-air volume coupled state-space model as the state transition equation to achieve Bayesian optimal state estimation. When the cumulative firing rate of any missing integral firing neuron exceeds the emergency trigger threshold within 10ms, the bypass direct connection layer is activated to output an emergency regulation code. The reference range for the emergency trigger threshold is 0.6–0.9. The output layer fuses the Bayesian optimal estimation result with the emergency regulation code using Dempster-Shafer evidence theory and spectral analysis diagnostic results, outputting the final resource allocation weight vector. Each element is linearly positively correlated with the memory allocation ratio of the corresponding strain airflow valve control thread and linearly negatively correlated with the control cycle duration.

[0070] The specific implementation of step S05 is as follows: The variational expectation propagation sparse decoupling algorithm uses the surface wind speed sensor signal and the valve position feedback signal to form an observation matrix. It introduces a Laplace sparse prior for the estimated airflow at each exhaust hood opening and a matrix normal prior for the pipeline sensitivity response matrix. A variational inference framework is used to transform the posterior approximation problem into a KL divergence minimization problem. The expectation propagation algorithm iteratively updates messages node by node on the factor graph, and the local computations at each node can be executed in parallel. During online operation, a sliding window is used to batch update the posterior, only locally refreshing factor graph nodes affected by new data. The global recalculation frequency is once per minute, ensuring that the embedded controller completes the decoupling calculation within the real-time control cycle. After decoupling, the estimated airflow at each exhaust hood opening and its Bayesian confidence interval are output. The airflow estimates are substituted back into the pipeline static pressure-airflow coupled state space model to achieve online self-correction of the coefficient matrix, while simultaneously updating the pipeline sensitivity response matrix.

[0071] The specific implementation of step S06 is as follows: The differential pressure dynamic adjustment function is obtained by weighted summing of three factors: the normalized value of the indoor transient differential pressure in the current control cycle, the normalized value of the current opening degree of the make-up air valve, and the normalized value of the current advance compensation timing of the exhaust-make-up air coordination prediction feedforward controller. Weighting coefficient , , The sum is 1, determined through at least 20 iterations of on-site debugging experiments. Based on... The interval is divided into four levels for updating compensation gain parameters: The value will increase to 1.2 to 1.5 times the current value. Time remains unchanged It will decrease to 0.7 to 0.9 times the current value. The system is reset to its initial calibration value and the pipeline static pressure-airflow coupled state-space model is triggered to re-identify the transfer function of the make-up air system. After the update is complete, the system returns to step S01 to enter the next control cycle.

[0072] It should be noted that the key technologies of this invention include: First, the combination of the pipeline static pressure-airflow coupling state-space model and the model predictive control framework transforms the adjustment action of each variable airflow valve from decentralized single-loop calculation to global joint optimization, turning the pipeline coupling effect from interference into a known constraint, thereby eliminating the temporal superposition conditions for beat frequency oscillation. Second, the variational expectation propagation sparse decoupling algorithm ensures airflow estimation convergence when the number of sensors is insufficient through Laplace sparse priors, and achieves sublinear computational complexity through a factor graph local refresh mechanism, making model parameter self-calibration feasible within the real-time control cycle. Third, the exhaust-makeup air coordination predictive feedforward controller uses Mahalanobis distance as the criterion for adaptive switching compensation strategy, combined with a Smith predictor to eliminate pure time delay in the makeup air system, preventing external pressure difference disturbances from becoming a new source of beat frequency oscillation excitation. The synergistic effect of these three technologies enables pipeline parameter identification, global predictive optimization, and makeup air coordination to form a closed adaptive control loop, maintaining stable convergence of face velocity even under complex operating conditions such as pipeline characteristic drift and simultaneous disturbances at multiple exhaust hoods.

[0073] It should be noted that under conditions where multiple exhaust hood openings simultaneously undergo significant changes in opening, the make-up air system exhibits a significant pure lag due to its long travel time and large equivalent heat exchange time constant. The transient negative pressure in the room deepens rapidly after a sudden increase in exhaust volume, while the actual increase in make-up air volume requires a 60-120s valve travel time and a 30-90s equivalent heat exchange delay to reach the room. This results in a severe lag between make-up air compensation and exhaust volume changes. The persistent negative pressure in the room then negatively impacts the static pressure at the exhaust duct network nodes, further disturbing the actual face velocity at each exhaust hood opening. The above technical problem arises because the pure lag characteristic of the make-up air system prevents closed-loop control from compensating promptly after the occurrence of negative pressure in the room. Traditional fixed-delay feedforward strategies cannot adaptively adjust the triggering timing under conditions with varying exhaust volume changes in magnitude and rate. Furthermore, after the negative pressure reacts to the exhaust duct network, each variable air volume valve responds independently to new deviations, further exacerbating network coupling disturbances and creating a vicious cycle where make-up air lag and exhaust oscillations amplify each other. The common solutions to the aforementioned technical problems are to increase the power of the motor driving the make-up air valve to shorten the travel time, set a fixed time advance to trigger the make-up air, or introduce cascade control on the make-up air side. However, increasing the drive power is limited by the system's mechanical structure, the fixed advance cannot cover operating conditions with large differences in the rate of change of exhaust air volume, and cascade control still relies on the actual flow feedback of the make-up air, failing to eliminate the impact of pure time delay on closed-loop stability. Furthermore, none of the above methods consider the identification uncertainty of the make-up air system transfer function, resulting in a significant degradation in control performance due to parameter drift after long-term operation. This invention effectively solves this technical problem by identifying the make-up air system transfer function online and establishing a Smith predictor, eliminating the pure time delay of the make-up air system from the closed-loop characteristic equation. This allows the exhaust-make-up air coordinated predictive feedforward controller to be designed only for the time-free dynamic part of the make-up air, significantly improving compensation accuracy. Meanwhile, the Mahalanobis distance criterion assesses the degree of deviation of the indoor transient pressure difference in real time, and dynamically adjusts the timing of the advance compensation based on the degree of deviation. When there is a significant surge in exhaust air volume, it immediately triggers makeup air and superimposes Smith predictor compensation. When the pressure difference is normal, it avoids unnecessary premature action, balancing the timeliness of compensation and control stability. The dynamic adjustment function of the pressure difference further uses the opening degree of the makeup air valve and the timing of advance compensation as comprehensive indicators to correct the compensation gain parameters in real time. This ensures that the exhaust-makeup air coordinated feedforward controller maintains the optimal compensation amplitude even after long-term drift of the pipeline characteristics, fundamentally cutting off the propagation path of secondary disturbances in the exhaust pipeline caused by makeup air lag.

[0074] Specifically, the principle of this invention is:

[0075] The fundamental reason why this invention can solve the above-mentioned technical problems is that it replaces the original decentralized single-loop control architecture with a centralized predictive control architecture based on the pipeline static pressure-air volume coupling state space model, so that the pipeline coupling relationship is transformed from a source of interference into predictable and usable known information.

[0076] The physical essence of beat frequency oscillation is that multiple regulating loops in a shared pipe network mutually excite each other. The error signal of each loop contains coupling components introduced by the actions of other loops. The single-loop controller misinterprets this component as a setpoint deviation of its own loop and responds accordingly, generating a new round of interference, and so on in a cycle. This invention first establishes the pipe network static pressure-airflow coupling state space equation through a step response identification method, accurately quantifying the influence coefficient matrix of the valve opening changes on the static pressure of all pipe network nodes, and parameterizing the pipe network coupling relationship. On this basis, the model predictive control framework uses the current pipe network node static pressure and the estimated airflow of each exhaust hood as initial conditions, predicts the global state evolution of multiple control steps forward, and uses the minimum sum of squared deviations of the spatial wind speed of all operating platforms as the objective function to simultaneously solve for the optimal opening sequence of all variable air volume valves in the future control time domain. This solution process naturally incorporates the mutual influence of valve actions into the constraints, making the output valve commands coordinated in a global sense, eliminating the condition of phase superposition caused by differences in regulation timing, thereby eliminating beat frequency oscillation.

[0077] Furthermore, the variational expectation propagation sparse decoupling algorithm continuously updates the network sensitivity response matrix online, ensuring that the state-space model parameters self-correct in real time as they drift with network characteristics, maintaining the model predictive control framework's accurate description of the actual coupling relationship. The exhaust dynamic control optimization model, through the fusion of a spiking neural network and an unscented Kalman filter, dynamically allocates computational resources to each control thread, enabling higher-priority exhaust hoods to obtain shorter control cycles, further shortening the global coordination response time. The exhaust-makeup air coordination predictive feedforward controller judges the degree of indoor pressure difference deviation based on Mahalanobis distance, adaptively adjusts the advance compensation timing, and combines a Smith predictor to compensate for the pure time delay of the makeup air system, preventing indoor negative pressure disturbances caused by insufficient makeup air from being superimposed on the exhaust network, and avoiding external pressure difference disturbances becoming a new excitation source for beat frequency oscillations. All the above components support each other, jointly ensuring global stable convergence during multi-exhaust hood linkage adjustment.

[0078] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0079] The specific implementation method of step S01 is as follows.

[0080] Collect raw signals from the face velocity sensors at each exhaust hood opening. ,unit First, perform a sliding median window filter on it, with a window length of... The value range is 1 to 5 seconds, and the filtered signal is obtained. ,unit This is used to suppress isolated pulse interference. Subsequently, [the following was done / impled / etc.]. Wavelet decomposition is performed to divide the signal into low-frequency subbands. With high-frequency subband All units are Retain low-frequency subbands where the duration exceeds the effective duration threshold The component of the effective surface wind speed signal ,unit . The value range is 3-5s, and it is determined by no fewer than 30 on-site personnel disturbance experiments, with the false trigger rate being [value missing]. With false negative rate The weighted optimal summation is used to determine the objective iteratively, and the objective function is expressed as follows:

[0081] ;

[0082] In the formula, The target value is optimized for the threshold; it is dimensionless. The false trigger rate weight is a dimensionless positive real number. The weight for the false negative rate is a dimensionless positive real number. To be at the threshold The false trigger rate obtained from the statistics is dimensionless. To be at the threshold The false negative rate obtained from the statistics is dimensionless.

[0083] The specific implementation method of step S02 is as follows.

[0084] Effective surface wind speed signal As input, establish a coupled state-space model of the pipeline network static pressure and air volume. Assume the pipeline network has a total of... Each node has both the number of exhaust hood openings and the number of variable air volume valves. The state variables are 1D nodal hydrostatic vector The unit is Pa. Vector estimation of air volume at the exhaust hood opening ,unit , merged into augmented state vector The control input is Valve opening command vector The unit is °, and the state equation is expressed as follows:

[0085] ;

[0086] ;

[0087] In the formula, for 3D augmented state vector The system matrix is ​​dimensionless. The input matrix has the following dimensions: That is, the unit of each state quantity divided by ° Let be the observation matrix, with dimensions of Let be the process noise vector, with dimensions equal to... Consistent To observe the noise vector, the dimensions are... Consistent For the observed output vector, in units For discrete time step indexes, dimensionless This represents the space of real vectors or real matrices of the corresponding dimension. Coefficient matrix. , , The step response method is used to identify the variable air volume valve. Step excitation is applied to each variable air volume valve in sequence, and the response of all pipeline nodes is recorded. The fitting is iterated for no less than 5 rounds until the residual converges.

[0088] The model predictive control framework, in each control cycle, uses the current state... Using initial conditions, predict forward. One control step size The value range is 3 to 10, and the objective function is expressed as follows:

[0089] ;

[0090] In the formula, For the first The exhaust hood opening is at the first The predicted face wind speed at step, in units For the first The face velocity setting value for each exhaust hood opening, in units To control the incremental penalty coefficient, a dimensionless coefficient is used, with an empirical value of 0.01–0.1. For the first The valve opening increment vector, in degrees. This represents the maximum valve opening, in degrees (°). Euclidean norm The objective function value is dimensionless; the optimal valve opening command is obtained by solving the above quadratic programming problem. The unit is °, and the instructions are issued for implementation.

[0091] The specific implementation method of step S03 is as follows.

[0092] Acquisition of indoor transient pressure difference vector Unit: Pa This represents the number of dimensions for differential pressure data acquisition. Standard vector of differential pressure. The data was obtained from statistical analysis of indoor transient pressure difference time-series data collected continuously for no less than 7 days at the site. The unit is Pa, and the covariance matrix is ​​[missing information]. Estimate from the same batch of data using the sample covariance formula:

[0093] ;

[0094] In the formula, The total number of samples collected is dimensionless. For the first One sample was collected, in Pa. unit Mahalanobis distance The calculation is expressed as follows:

[0095] ;

[0096] In the formula, The inverse of the covariance matrix, in units. Unit: Pa; Square root of product This is a dimensionless statistical distance. When, the advance compensation timing is not adjusted; when At that time, advance compensation timing Shorten by 5-15 seconds; when At that time, the advance compensation timing is triggered simultaneously with the issuance of the exhaust volume change command, and the Smith predictor compensation for pure time lag is superimposed. The Smith predictor is based on the make-up air system transfer function. Establish, as described below:

[0097] ;

[0098] In the formula, The static gain of the make-up air system is dimensionless. This is the equivalent time constant, in seconds, with a value ranging from 30 to 90 seconds. This is the pure time delay, in seconds, corresponding to a stroke time of 60–120 seconds for the make-up air valve. For the Laplace operator, unit It is a dimensionless transfer function. The step response of the make-up air valve was identified through experiments, with the results repeated at least 10 times and the average value taken. A Smith predictor was connected in series in the control loop. The predicted output eliminates pure time delay from the closed-loop characteristic equation, allowing the exhaust-makeup air coordinated predictive feedforward controller design to be performed only for the time-free portion.

[0099] The specific implementation method of step S04 is as follows.

[0100] Normalize the effective face velocity of each exhaust hood opening. Normalized value of indoor transient pressure difference and the normalized values ​​of the valve position feedback signals of each variable air volume valve Composition of multidimensional time-series input vectors , dimensionless. Among them This is the current indoor transient pressure differential scalar, in Pa. This is the rated differential pressure, in Pa. For the first Each variable air volume valve position feedback signal, in degrees. The rated opening of the make-up air valve is in degrees. The input layer uses delta modulation to... Convert to sparse pulse sequence The superscript 0 indicates the input layer, and the delta modulation rule is described as follows:

[0101] ;

[0102] In the formula, For the input layer 3D pulse output for No. Dimensionless components For the first Dimensionless reference level, updated after each trigger. The delta modulation threshold is dimensionless and ranges from 1% to 5% of the sensor's full scale. The coding layer will... Mapped to a high-dimensional pulse feature space, the backbone network consists of three layers of hidden layers with 256 leaky integral firing neurons each. Layer The membrane potential of a leaky integral firing neuron The dynamic equations are expressed as follows:

[0103] ;

[0104] In the formula, For the first Layer The membrane potential of a single neuron, dimensionless normalized value. The membrane time constant is expressed in milliseconds (ms), with an empirical value of 10–20 ms. For the first Layer The input neuron to the first Synaptic weights of individual neurons, dimensionless For the first Layer The pulse output of each neuron Dimensionless unit The right side of the equals sign unit , The sum is a dimensionless pulse weighted sum, and its contribution is expressed through synaptic weights. Converted to Dimensions consistent with the left side; when Exceeding the threshold At that time, neurons fire impulses And Reset to 0, The empirical value is 0.5 (normalized). Synaptic weights are updated online using impulse timing-dependent plasticity rules, with the update amount... The statement is as follows:

[0105] ;

[0106] In the formula, The time of post-neuron firing is measured in milliseconds (ms). The firing time of the preneuron is measured in milliseconds (ms). This is the weighting enhancement coefficient, dimensionless, with an empirical value of 0.01. This is the weighting weakening coefficient, dimensionless, with an empirical value of 0.012. To enhance the time window constant, in milliseconds, the empirical value is 20ms. To reduce the time window constant (unit: milliseconds), an empirical value is 20 milliseconds. and It is an indicator function, dimensionless. This represents the dimensionless weight update quantity; all terms on the right-hand side of the equation are dimensionless. The output pulse firing rate vector of the 3rd hidden layer. ,unit Input unscented Kalman filter, with For the observables, the predicted values ​​from the pipeline static pressure-airflow coupled state-space model are used as the state transition equations to achieve the Bayesian optimal estimate of the control state, thus obtaining the estimated state. With estimation of covariance matrix When any neuron accumulates within a 10ms window Cumulative pulse firing rate within Exceeding the emergency trigger threshold When this occurs, the bypass direct-connect layer outputs an emergency adjustment code. , The calculation is expressed as follows:

[0107] ;

[0108] In the formula, The cumulative window length is set to 10ms. For the 3rd layer One neuron in Pulse output at any moment Dimensionless unit , The value range is 0.6 to 0.9 (normalized, unit). The output layer is determined through at least 20 field-based mutation perturbation experiments, using a weighted optimal iteration based on false triggering rate and missed triggering rate. and Assign weight vectors to the input and output resources. The vector is dimensionless and is fused with the spectral analysis diagnostic results using the Dempster-Schafer evidence theory to obtain the final resource allocation weight vector. Dimensionless, based on Adjust the memory allocation ratio and control cycle duration of each variable air volume valve control thread.

[0109] The specific implementation method of step S05 is as follows.

[0110] The variational expectation propagation sparse decoupling algorithm is used to establish a linear hybrid model for a multi-sensor system. Let the observation matrix be... It consists of a face wind speed sensor signal and a valve position feedback signal. The total number of sensors, The sliding window batch length is expressed in units of sampling points. The hybrid model is described as follows:

[0111] ;

[0112] In the formula, The network sensitivity response matrix has the following dimensions: That is, the sensor signal unit divided by This is a matrix of estimated airflow values ​​for each exhaust hood opening, in units of... Let be the residual matrix, with dimensions and Consistent. Yes. Each column Introducing Laplace sparse prior, for of Norm, unit , For sparse regularization coefficients, in units of , so that the prior exponential parameter Dimensionless, empirical value is 0.1 to 1.0; for Introducing a matrix normal prior, the probability density is expressed as follows:

[0113] ;

[0114] In the formula, for Prior mean matrix Let be the row covariance matrix, positive definite. Let be the column covariance matrix, positive definite. The trace operation is performed on the matrix; the product of matrices within the exponent is dimensionless. The variational inference framework minimizes the approximate posterior. The Kolb-Leibler divergence between the true posterior and the actual posterior achieves a posterior approximation. The Kolb-Leibler divergence is expressed as follows:

[0115] ;

[0116] In the formula, The Korbeck-Leibler divergence is dimensionless. For approximate posterior probability density This represents the true posterior probability density. The expectation propagation algorithm iteratively updates messages node by node on the factor graph, and after decoupling, obtains the estimated airflow at each exhaust hood opening. ,unit Substituting back into the pipeline static pressure-airflow coupled state-space model from step S02, self-calibration of model parameters is achieved, and the pipeline sensitivity response matrix is ​​determined. The global recalculation is performed once every minute.

[0117] The specific implementation method of step S06 is as follows.

[0118] Differential pressure dynamic adjustment function value The calculation formula is expressed as follows:

[0119] ;

[0120] In the formula, The unit is the indoor transient pressure difference, expressed in Pa. This is the rated differential pressure, in Pa. The current opening degree of the air supply valve, in degrees. The rated opening of the make-up air valve is expressed in degrees (°). For the current advance compensation time series, the unit is s. The time series advance compensation rating is expressed in seconds. , , Let be the weighting coefficient, satisfying Dimensionless, determined by minimizing the weighted index of indoor transient pressure difference deviation and adjustment time through at least 20 rounds of on-site commissioning iterations. The value is a dimensionless adjustment function; all terms on the right-hand side of the equation are normalized to dimensionless by dividing by the nominal value. When At that time, compensation gain parameter Upgraded to ~ ;when hour, Remain unchanged; when hour, Reduce to ~ ;when hour, Reset to initial calibration value It then triggers the pipeline static pressure-air volume coupled state space model to re-identify the transfer function of the make-up air system, and returns to step S01 after completing one control cycle.

[0121] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To verify the effectiveness of the invention, technicians built a test environment and deployed the automatic exhaust control method described in this invention on an integrated experimental platform in a chemical laboratory to conduct a full-process verification of an exhaust duct network system containing 6 exhaust hoods. The laboratory floor height is 3.5m, the diameter of the main exhaust pipe is 400mm, the rated face velocity setting of each exhaust hood is 0.5m / s, the rated stroke time of the make-up air valve is 90s, and the equivalent time constant of temperature regulation heat exchange is 60s.

[0122] The integrated experimental platform's exhaust hood consists of an exhaust hood body, a built-in micro-differential pressure / wind speed sensor, connecting branch pipes, and a variable air volume valve (VAV) installed at the junction of the branch pipe and the main exhaust pipe. The sensor captures the flow velocity at the hood opening in real time and feeds the signal back to the controller. The controller then performs global calculations using a state-space model and uniformly issues valve opening commands to the VAV located at the branch pipe. This ensures the efficiency of pollutant collection at each workstation while eliminating static pressure fluctuations in the pipeline network caused by sharing the main pipe.

[0123] During the system initialization phase, technicians sequentially applied step excitations to each variable air volume (VAV) valve, recording the static pressure response curves of all six pipeline nodes. After five rounds of iterative fitting, the coefficient matrix residuals of the pipeline static pressure-air volume coupled state-space model converged to below 0.003. Simultaneously, a step response experiment was conducted on the make-up air valve, repeated 12 times and the average value was taken to identify the transfer function parameters of the make-up air system, upon which the Smith predictor was established. The effective duration threshold was determined through 34 personnel operation disturbance experiments, ultimately set at 4 seconds. The pressure difference standard vector was obtained from eight consecutive days of indoor transient pressure difference time-series data through statistical processing. The training dataset for the exhaust dynamic control optimization model covered scenarios including disturbances at a single exhaust hood opening, simultaneous disturbances at three exhaust hood openings, and delayed response scenarios at the make-up air valve, with a total acquisition time of 750 hours. After cleaning, the dataset was divided in an 8:1:1 ratio.

[0124] During the verification phase, technicians simulated actual experimental operations, simultaneously applying opening variation disturbances to six exhaust hood openings. Before the disturbance, the air velocity in the space of each operating platform was stable near the set value. After the disturbance was triggered, the model predictive control framework, within the current control cycle, used the minimum sum of squared face velocity deviations of the six exhaust hood openings as the objective function, simultaneously solving for the optimal valve opening command for all variable air volume valves and issuing it for execution. When the indoor transient pressure difference reached 3.6 Mahalanobis distance, the exhaust-makeup air coordinated predictive feedforward controller immediately advanced the advance compensation timing to the moment the exhaust volume change command was issued, and activated the Smith predictor to compensate for pure time lag. The exhaust dynamic control optimization model detected that the cumulative impulse firing rate of the leak integral firing neurons corresponding to three exhaust hood openings exceeded the emergency trigger threshold of 0.72, activating the emergency jump channel. The bypass direct connection layer directly output the emergency adjustment code, increasing the memory allocation ratio of the corresponding control thread and shortening the control cycle. The convergence process of the air velocity deviation in the space of each operating platform over time is as follows: Figure 2 As shown.

[0125] The values ​​of each key control parameter during one complete control cycle are shown in Table 1:

[0126] Table 1. Record of key parameters for a typical control cycle

[0127]

[0128] During the online operation of the variational expectation propagation sparse decoupling algorithm, the pipeline sensitivity response matrix is ​​globally recalculated every minute, and the local factor graph nodes are updated instantly upon the arrival of new data. After decoupling, the mean width of the Bayesian confidence interval for the estimated airflow at each exhaust hood outlet remains at 0.018. Below, after substituting the estimated airflow values ​​back into the coupled state-space model of the pipeline static pressure-airflow, the coefficient matrix residuals converge further. The dynamic adjustment function value of the differential pressure... The calculation results are as follows Figure 3 As shown, the weighting coefficients , , The values ​​were set to 0.5, 0.3, and 0.2 respectively, and determined through 22 rounds of on-site debugging experiments.

[0129] The dynamic response curve of static pressure at pipeline nodes after disturbance is shown in the figure. Figure 4 As shown, the static pressure fluctuations at the six nodes gradually converged under coordinated control, without any sustained low-frequency oscillations. In contrast to the traditional single-loop proportional-integral-derivative (PID) control scheme, which exhibits significant beat-frequency oscillations in the static pressure at each pipeline node under the same disturbance, with an oscillation period of approximately 80 seconds, the method described in this invention uses global prediction optimization to ensure that the valve adjustment actions are coordinated in phase, eliminating the timing superposition conditions required for beat-frequency oscillations. Under coordinated optimization control, the pipeline static pressure monotonically converges without any low-frequency oscillation components. The entire test bench in Example 2 is as follows... Figure 5or Figure 6 As shown.

[0130] From a technical principle perspective, the advancements of this invention compared to traditional methods are reflected in the following aspects: Traditional single-loop controllers, when dealing with pipeline coupling, can only passively respond to the error signal of their own loop, failing to incorporate the actions of other loops into their own loop's decision-making. This results in the responses of each loop superimposed on the time axis, exciting beat frequency oscillations. This invention, based on a state-space model, parameterizes the global coupling relationship and incorporates it into the optimization constraints of model predictive control, unifying the adjustment actions of all valves at the solution level and eliminating the source of superimposed excitation from a fundamental mechanism. Traditional fixed-delay feedforward strategies cannot adapt to changes in the pure time delay of the supplementary air system. This invention, however, eliminates the pure time delay from the closed-loop characteristic equation through a Smith predictor, allowing the compensation controller design to focus only on the time-delay-free portion, improving compensation accuracy and stability. Traditional offline calibration parameters become invalid after long-term operation due to pipeline characteristic drift. This invention, however, uses a variational expectation propagation sparse decoupling algorithm to continuously update the pipeline sensitivity response matrix during online operation, maintaining the model's accurate description of the actual pipeline network. This ensures that the look-ahead calculation of predictive control is always based on effective parameters, maintaining long-term consistency in control performance.

[0131] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0132] Table 2. Variable Explanation Table (Part 1)

[0133]

[0134] Table 3. Variable Explanation Table (Part Two)

[0135]

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic exhaust control method for an integrated experimental platform, characterized in that, Includes the following steps: The face wind speed sensor signals of each exhaust hood are collected. The face wind speed sensor signals are then subjected to sliding median window filtering and wavelet decomposition filtering in sequence. The low-frequency components whose duration exceeds the effective duration threshold are extracted as the effective face wind speed signals. Using the effective surface wind speed signal as input, the valve opening command of each variable air volume valve is uniformly calculated through the pipeline static pressure-air volume coupled state space model and the model predictive control framework, and the valve opening command is sent to each variable air volume valve for execution. The system collects the indoor transient pressure difference, calculates the Mahalanobis distance between the indoor transient pressure difference and the standard vector of the pressure difference, adjusts the advance compensation sequence of the exhaust-makeup air coordination prediction feedforward controller according to the interval to which the Mahalanobis distance belongs, and drives the makeup air valve to act in advance according to the advance compensation sequence. The effective face velocity signal of each exhaust hood opening, the indoor transient pressure difference, and the valve position feedback signal of each variable air volume valve are input into the exhaust dynamic control optimization model. The exhaust dynamic control optimization model outputs a resource allocation weight vector. Based on the resource allocation weight vector, the memory allocation ratio and control cycle duration of the corresponding control thread of each variable air volume valve are adjusted. The variational expectation propagation sparse decoupling algorithm is used to decouple the surface wind speed sensor signal and the valve position feedback signal online, update the pipeline network sensitivity response matrix, and substitute the decoupled air volume estimates of each exhaust hood inlet back into the pipeline network static pressure-air volume coupled state space model to achieve self-correction of model parameters. The system calculates the dynamic adjustment function value of the differential pressure in real time, selects the corresponding compensation gain parameter based on the range to which the dynamic adjustment function value of the differential pressure belongs, updates the compensation gain parameter to the exhaust-makeup air coordinated prediction feedforward controller, and returns to the step of collecting the space wind speed sensor signal of each operating platform after completing one control cycle.

2. The automatic exhaust control method according to claim 1, characterized in that, The sliding median window filter has a window length of 1 to 5 seconds. The wavelet decomposition filter decomposes the surface wind speed sensor signal into a low-frequency sub-band and a high-frequency sub-band, retaining only the components in the low-frequency sub-band whose duration exceeds the effective duration threshold.

3. The automatic exhaust control method according to claim 2, characterized in that, The effective duration threshold is determined iteratively through no less than 30 human operation disturbance experiments, with the goal of achieving the optimal weighted comprehensive balance between false trigger rate and false detection rate, and the value range is 3 to 5 seconds.

4. The automatic exhaust control method according to claim 3, characterized in that, The pipeline static pressure-airflow coupled state-space model uses the valve opening command of each variable airflow valve as the control input and the estimated static pressure of each pipeline node and the airflow of each exhaust hood outlet as the state variables. A multi-input multi-output linear state-space equation is established, and the coefficient matrix is ​​identified by the step response method. The model is iterated for no less than 5 rounds until the residual converges.

5. The automatic exhaust control method according to claim 4, characterized in that, The model predictive control framework uses the current static pressure of the pipeline node and the estimated air volume of each exhaust hood as initial conditions in each control cycle, predicts forward 3 to 10 control steps, and uses the minimum sum of squared deviations of the spatial wind speeds of all operating platforms as the objective function to uniformly optimize the valve opening commands of each variable air volume valve.

6. The automatic exhaust control method according to claim 5, characterized in that, The pressure difference standard vector is obtained by continuously collecting indoor transient pressure difference time-series data for no less than 7 days, and statistically processing it to obtain a vector composed of the mean and standard deviation of the pressure difference for each time period; when the Mahalanobis distance is less than 1, the lead compensation time series is not adjusted; when the Mahalanobis distance is within the interval... When the advance compensation timing is shortened by 5 to 15 seconds, when the Mahalanobis distance is greater than or equal to 3, the advance compensation timing is advanced to the moment the exhaust volume change command is issued, and the Smith predictor is superimposed to compensate for the pure lag.

7. The automatic exhaust control method according to claim 6, characterized in that, The Smith predictor is based on the transfer function of the make-up air system, which is identified by the step response experiment of the make-up air valve. The experiment is repeated no less than 10 times and the average value is taken. The stroke time of the make-up air valve is 60-120s, and the equivalent time constant of the temperature regulation heat exchange is 30-90s.

8. The automatic exhaust control method according to claim 7, characterized in that, The exhaust dynamic control optimization model includes an input layer, an encoding layer, a backbone network, an unscented Kalman filter, and an output layer. The backbone network consists of three hidden layers of leaky integral firing neurons, each containing 256 leaky integral firing neurons. The synaptic weights are updated online using pulse temporal dependence plasticity rules.

9. The automatic exhaust control method according to claim 8, characterized in that, The exhaust dynamic control optimization model sets up an emergency jump channel. When the cumulative pulse firing rate of any leaky integral firing neuron exceeds the emergency trigger threshold within 10ms, the bypass direct connection layer is activated to directly output the emergency regulation code, skipping the regular classification layer. The emergency trigger threshold ranges from 0.6 to 0.

9.

10. The automatic exhaust control method according to claim 9, characterized in that, The output layer takes the Bayesian optimal estimation result and the emergency adjustment code as input, and outputs the resource allocation weight vector in parallel with the traditional spectrum analysis diagnosis result. The final resource allocation weight vector is obtained by fusing decision through Dempster-Shafer evidence theory.

Images