A method and system for controlling waste gas recovery in resin production

Abstract

This application belongs to the field of control and relates to a control method and system for recovering waste gas from resin production. It addresses the technical problems in existing technologies, such as the inability of operating condition division to match the true state of a multivariable system and the need for PID parameters to balance multiple control objectives. The control method includes the following steps: S1, dividing neurons into stable operating condition neurons, transitional operating condition neurons, and disturbance operating condition neurons; S2, obtaining a Pareto optimal solution set composed of PID parameter groups, including the ITAE optimal solution, the control output rate of change optimal solution, and the equilibrium solution; S3, determining the best matching unit for the latent eigenvectors in a self-organizing map network; S4, when the best matching unit is a stable operating condition neuron, selecting the ITAE optimal solution when the absolute value of the current control error is greater than a high threshold; selecting the control output rate of change optimal solution when the absolute value of the current control error is less than a low threshold; and selecting the equilibrium solution when it is between the two. This achieves accurate identification and division of the operating state of the resin production waste gas recovery process.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to a method and system for controlling the recovery of waste gas from resin production. Background Technology

[0002] Waste gas recovery systems in resin production processes can recover waste gas, which not only reduces environmental pollution but also enables waste utilization. However, the recovery efficiency varies considerably under different conditions due to factors such as fluctuations in raw material composition, changes in production load, and variations in ambient temperature.

[0003] In the control of waste gas recovery from resin production, the most common method is PID control because it is relatively simple. However, when the operating conditions deviate significantly, PID control struggles to adapt, leading to deteriorated control performance, slow response, large overshoot, and even system oscillations. This makes it impossible to simultaneously meet the dual requirements of waste gas recovery rate and system operational stability.

[0004] As an improvement on PID control, gain-scheduled PID control pre-sets several typical operating points and offline tunes an independent set of PID parameters for each point. During operation, it determines the current operating condition based on the values ​​of one or more key process variables and switches to the corresponding parameter set. However, the division of operating conditions often relies on a single or a few directly measurable variables, which is overly simplistic and fails to accurately reflect the true operating state of the system affected by multiple variables. Furthermore, it is difficult to achieve a balance between multiple conflicting objectives, such as tracking performance and control output stability, in the PID parameters for each operating point. Summary of the Invention

[0005] This invention provides a method and system for controlling the recovery of waste gas from resin production, in order to solve the technical problems in the prior art where the division of operating conditions cannot match the real state of a multivariable system and the PID parameters need to be balanced across multiple control objectives.

[0006] In a first aspect, the present invention provides a method for controlling and recovering waste gas from resin production, comprising the following steps:

[0007] S1. Acquire historical time-series data of exhaust gas concentration, temperature, and pipeline pressure, divide them into several data windows, and use a pre-trained variational autoencoder to transform each data window into a latent space of a preset dimension to obtain a single latent feature vector. Input the historical latent feature vector set formed by summing all latent feature vectors into a self-organizing map network for offline training to obtain a two-dimensional neuron topology grid. Based on the variance of the process variables associated with the latent feature vectors, the neurons are divided into stable operating condition neurons, transitional operating condition neurons, and perturbation operating condition neurons.

[0008] S2, obtain the local linearized process model corresponding to each stable working condition neuron, and solve it using a non-dominated sorting genetic algorithm. With minimizing ITAE and minimizing the rate of change of control output as the dual optimization objectives, obtain the Pareto optimal solution set composed of PID parameter groups, which includes the optimal solution of ITAE, the optimal solution of control output rate of change, and the equilibrium solution.

[0009] S3 acquires the current exhaust gas concentration, temperature and pipeline pressure in real time during each control cycle, calculates the potential feature vector of the current operating condition using a variational autoencoder, and determines the best matching unit of the potential feature vector in the self-organizing map network. The best matching unit is the neuron in S1 that has completed the operating condition classification.

[0010] S4. When the best matching unit is the steady-state neuron, the optimal solution is selected when the absolute value of the current control error is greater than the high threshold; the optimal solution is selected when the absolute value of the current control error is less than the low threshold; and the balanced solution is selected when it is between the two.

[0011] Furthermore, in S1, the process of dividing neurons into stable operating condition neurons, transitional operating condition neurons, and perturbed operating condition neurons includes the following steps:

[0012] Calculate the variance of exhaust gas concentration, temperature and pipeline pressure corresponding to the latent feature vector associated with each neuron to obtain a comprehensive variance index representing the fluctuation of the operating conditions corresponding to each neuron.

[0013] Sort all neurons in ascending order of their combined variance index;

[0014] Neurons with a comprehensive variance index below a preset low quantile are designated as stable operating neurons, neurons with a comprehensive variance index above a preset high quantile are designated as perturbed operating neurons, and the remaining neurons are designated as transitional operating neurons.

[0015] Furthermore, the variational autoencoder consists of an encoder and a decoder, both of which employ a multilayer perceptron structure;

[0016] During offline training, historical time series data is input into the VAE, and the network weights are adjusted by minimizing the reconstruction error and KL divergence.

[0017] Further, in S2, in an optional embodiment, the process of obtaining the Pareto optimal solution set composed of the PID parameter set includes the following steps:

[0018] Initialize a PID parameter set [K] with multiple sets of parameters. p K i K d The population constituted by;

[0019] Offspring populations are generated by performing crossover and mutation operations on individuals in the population.

[0020] Based on non-dominated sorting and crowding calculation, superior individuals are selected from the parent and offspring populations to form a new generation population.

[0021] Repeat the iteration until the termination condition is met, and use the resulting non-dominated solution set as the Pareto optimal solution set.

[0022] Furthermore, in S4, when the best matching unit is a transitional working condition neuron, the neighboring stable working condition neurons of the best matching unit on the two-dimensional neuron topology grid are determined, and the balance solutions corresponding to each neighboring stable working condition neuron are weighted and fused according to the Gaussian kernel function value between the potential feature vector of the current working condition and the weight vector of each neighboring stable working condition neuron to obtain the current PID parameters.

[0023] Furthermore, in an optional embodiment, when the best matching unit is a transitional working condition neuron, the process of weightedly fusing the equilibrium solutions corresponding to each neighboring stable working condition neuron to obtain the current PID parameters includes the following steps:

[0024] On a two-dimensional neuron topology grid, multiple stable working condition neurons that are adjacent to the best matching unit are selected based on topological distance;

[0025] The similarity between the latent feature vector of the current working condition and the weight vector of each neighboring stable working condition neuron is calculated using the Gaussian kernel function, and this similarity is used as the weight.

[0026] The calculated weights are normalized, and the PID parameters corresponding to the equilibrium solutions of neurons in each nearby stable condition are weighted and summed to obtain the current PID parameters.

[0027] Furthermore, if the current best matching unit is a transitional neuron, search for its eight neighboring neurons on the two-dimensional neuron topology grid, and select all neighboring neurons marked as stable neurons, which are denoted as the neighboring stable neurons.

[0028] Furthermore, in S4, when the optimal matching unit is the disturbance condition neuron, the proportional and derivative components of the PID controller are maintained, the integral component is reset according to the recovery function, and the feedforward compensation filter associated with the disturbance condition neuron is enabled.

[0029] Furthermore, the process of maintaining the proportional and derivative components of the PID controller, resetting the integral component according to the recovery function, and enabling the feedforward compensation filter associated with the disturbance condition neuron includes the following steps:

[0030] Keep the proportional P and derivative D parameters of the PID controller unchanged, and reset the cumulative state value of the integral I element to the preset value;

[0031] Enable the pass function as The feedforward compensation filter, where the gain Here are the gain parameters of the feedforward compensation filter. , It is a time constant. It is a variable in the complex frequency domain. The output of the feedforward compensation filter is superimposed on the output of the PID controller, and together they act on the control actuator.

[0032] Secondly, the present invention provides a resin production waste gas recovery control system, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned resin production waste gas recovery control method is implemented.

[0033] Beneficial effects:

[0034] 1. This invention introduces a variational autoencoder to extract features from multidimensional process data and combines it with a self-organizing map network for operating condition clustering, thereby achieving accurate identification and classification of the operating state of the resin production waste gas recovery process; for the different stable operating conditions identified, a multi-objective genetic algorithm is used to solve for the Pareto optimal PID parameter set; during the control implementation phase, the optimal control strategy is selected based on the error magnitude within the stable operating conditions, which can balance multiple control objectives.

[0035] 2. In the transition zone of operating conditions, the transition of control action is achieved by weighted fusion of adjacent parameters, avoiding sudden output changes caused by parameter switching.

[0036] 3. When encountering large disturbances, the integral reset and feedforward compensation mechanism can suppress integral saturation and enhance the system's anti-disturbance capability. Attached Figure Description

[0037] Figure 1 A flowchart for a method of controlling waste gas recovery in resin production;

[0038] Figure 2 A time series plot of pipeline pressure;

[0039] Figure 3 This is a time series graph of exhaust gas concentration;

[0040] Figure 4 A time series plot of temperature;

[0041] Figure 5 A schematic diagram of the Pareto optimal solution set for PID parameters;

[0042] Figure 6 A schematic diagram illustrating the SOM (Self-Organizing Map) network's operating conditions.

[0043] Figure 7 This is a structural block diagram of a resin production waste gas recovery and control system. Detailed Implementation

[0044] Specific embodiments of the resin production waste gas recovery and control method provided by the present invention:

[0045] like Figure 1 As shown, a method for controlling and recovering waste gas from resin production includes the following steps:

[0046] S1. Obtain historical time series data of exhaust gas concentration, temperature, and pipeline pressure, divide them into several data windows, and use a pre-trained variational autoencoder to transform each data window into a latent space of a preset dimension to obtain a single latent feature vector. Input the set of historical latent feature vectors formed by summing all latent feature vectors into a self-organizing map network for offline training to obtain a two-dimensional neuron topology grid. According to the variance of the process variable associated with the latent feature vector, the neurons are divided into stable operating condition neurons, transitional operating condition neurons, and perturbation operating condition neurons.

[0047] In one optional embodiment, the resin production waste gas mainly includes benzene compounds, aldehydes, esters, HCl, ammonia, and sulfides, all of which can pollute the environment. Different sensors in the resin production waste gas recovery system (such as waste gas analyzers, thermocouples, and pressure transmitters) continuously collect waste gas concentration, temperature, and pipeline pressure data for several months or days to obtain historical data, such as... Figure 2 , Figure 3 and Figure 4 As shown, historical data is divided into data windows containing continuous sampling points in chronological order. Each data window forms a multi-dimensional input vector. For example, if a data window has 10 sampling points, then the input vector for each data window is 30-dimensional. These multi-dimensional input vectors constitute the historical time series data.

[0048] A variational autoencoder (VAE) consists of an encoder and a decoder, both of which employ a multilayer perceptron structure, with a latent space dimension preferably of 2. During offline training, historical time-series data is input into the VAE, and the network weights are adjusted by minimizing reconstruction error and KL divergence. After training, all historical time-series data are sequentially passed through the trained encoder, mapping each 30-dimensional data window to a 2-dimensional latent feature vector. The latent feature vectors of all data windows together constitute the historical latent feature vector set.

[0049] Preferably, a 10×10 two-dimensional neuron topology grid is established, where the dimension of the weight vector of each neuron in the two-dimensional neuron topology grid is the same as the dimension of the latent feature vector (i.e., 2-dimensional). The two-dimensional neuron topology grid is a two-dimensional planar grid obtained by training a neural network using a self-organizing map network (SOM). During training, the neurons in the SOM spontaneously organize themselves to form a two-dimensional planar grid. Similar operating conditions in the original high-dimensional space are mapped to neighboring neurons on the two-dimensional neuron topology grid. Each neuron represents a specific production condition. By analyzing the variance of the data mapped to each neuron, these neurons can be further labeled as stable, transitional, or perturbed.

[0050] The SOM (Search Engine Module) is trained using a set of historical latent feature vectors. After training, for each neuron in the 2D planar grid, all historical latent feature vectors that are the best matching units are identified, and the corresponding raw process data (i.e., exhaust gas concentration, temperature, and pipeline pressure) are obtained. The variance of each variable in the raw process data is calculated. If the variance of the raw process data corresponding to a neuron is less than a low variance threshold, it is marked as a stable operating condition neuron; if the variance is greater than a high variance threshold, it is marked as a disturbed operating condition neuron; if the variance is between the two, it is marked as a transitional operating condition neuron, such as... Figure 5 , Figure 6 As shown.

[0051] In an optional embodiment, the process of dividing neurons into stable-condition neurons, transitional-condition neurons, and perturbation-condition neurons includes the following steps:

[0052] Calculate the variance of exhaust gas concentration, temperature and pipeline pressure corresponding to the latent feature vector associated with each neuron to obtain a comprehensive variance index representing the fluctuation of the operating conditions corresponding to each neuron.

[0053] Sort all neurons in ascending order of their combined variance index;

[0054] Neurons with a comprehensive variance index below a preset low quantile are designated as stable operating neurons, neurons with a comprehensive variance index above a preset high quantile are designated as perturbed operating neurons, and the remaining neurons are designated as transitional operating neurons.

[0055] For example, suppose a self-organizing map network contains 100 neurons. For one neuron, all data points mapped to that neuron are collected. These data points include exhaust gas concentration, temperature, and pipeline pressure. The variances of these variables over time are calculated separately; for example, the variance of exhaust gas concentration is 0.05, the variance of temperature is 1.2, and the variance of pipeline pressure is 0.8. These individual variances are then combined into a comprehensive variance index using methods such as weighted summation. The higher the comprehensive variance index, the greater the fluctuation in the operating conditions corresponding to the neuron. The weights of each variance can be determined using business experience. For instance, in chemical production, exceeding the exhaust gas concentration standard directly triggers an environmental warning, and its impact on operating condition stability is far greater than that of pipeline pressure. Therefore, the variance of exhaust gas concentration can be given a larger weight.

[0056] After calculating the combined variance index of all 100 neurons, the index values ​​are sorted from smallest to largest. A low quantile and a high quantile are pre-defined, for example, the low quantile is set to the 20th percentile and the high quantile is set to the 80th percentile. After sorting, the 20 neurons with the lowest combined variance index represent the state with the least fluctuation in the process variable, which are the steady-state neurons. The 20 neurons with the highest combined variance index represent the state with the most volatile fluctuation in the process variable, which are the perturbation neurons. The remaining 60 neurons in the middle represent the transitional neurons from steady-state to perturbation or from perturbation to steady-state.

[0057] S2. Obtain the local linearized process model corresponding to each stable operating condition neuron, and solve it using a non-dominated sorting genetic algorithm. With minimizing ITAE and minimizing the rate of change of control output as the dual optimization objectives, obtain the Pareto optimal solution set composed of PID parameter groups, which includes the optimal solution for ITAE, the optimal solution for the rate of change of control output, and the equilibrium solution.

[0058] In an optional embodiment, for each neuron labeled as a steady-state neuron, a first-order plus pure-time FOPDT model is identified using a system identification algorithm (such as the subspace method) based on the associated subset of original process data (i.e., the original process variables: exhaust gas concentration, temperature, and pipeline pressure). The FOPDT model includes three parameters: process gain K, time constant T, and delay time τ. The FOPDT model is a type of locally linearized process model, which uses a linear mathematical model to approximate the dynamic characteristics of complex nonlinear systems near a specific operating point. A corresponding FOPDT model is established for each steady-state neuron divided in the two-dimensional neuron topology grid. A bi-objective optimization problem is constructed for the FOPDT model, with the PID parameter K as the optimization variable. p K i K dThe two objective functions are the integral time multiplied by absolute error (ITAE) performance index under the closed-loop step response and the control output rate of change index, respectively. A non-dominated sorting genetic algorithm (NSGA-II) is used for optimization, and after multiple generations of evolution, a set of Pareto optimal solutions is obtained. From this set of Pareto optimal solutions, three representative solutions are selected: the solution with the smallest ITAE value is selected as the ITAE optimal solution, the solution with the smallest control output rate of change is selected as the control output rate of change optimal solution, and the solution located near the Pareto front inflection point is selected as the equilibrium solution.

[0059] In an optional embodiment, the process of obtaining the Pareto optimal solution set consisting of the PID parameter set includes the following steps:

[0060] Initialize a PID parameter set [K] with multiple sets of parameters. p K i K d The population constituted by;

[0061] Offspring populations are generated by performing crossover and mutation operations on individuals in the population.

[0062] Based on non-dominated sorting and crowding calculation, superior individuals are selected from the parent and offspring populations to form a new generation population.

[0063] Repeat the iteration until the termination condition is met, and use the resulting non-dominated solution set as the Pareto optimal solution set.

[0064] For example, to obtain PID parameters that balance control performance and control cost, an initial population is randomly generated, containing 50 different PID parameters, each representing an independent individual. Here, K in the PID parameters is set... p =2.1、K i =0.5, K d =0.08. These individuals undergo crossover and mutation operations: for example, averaging or weighting the parameters of two individuals to generate new individuals, or adding small random perturbations to the parameters of an individual, thus creating 50 offspring individuals. The parent and offspring individuals, a total of 100 individuals, are merged and evaluated based on two conflicting objectives: minimizing the ITAE and minimizing the rate of change of the control output. Individuals are assigned to different levels through non-dominated sorting, and the crowding degree within each level is calculated to ensure solution diversity. The 50 best individuals are selected from these 100 individuals to enter the next generation. After 200 iterations, the algorithm converges, yielding a set of non-dominated solutions that cannot be substituted for each other, which is the Pareto optimal solution set, containing various PID parameter combinations ranging from those emphasizing fast response to those emphasizing stationary control.

[0065] S3 acquires the current exhaust gas concentration, temperature and pipeline pressure in real time during the control cycle, calculates the potential feature vector of the current operating condition using a variational autoencoder, and determines the best matching unit of the potential feature vector in the self-organizing map network. The best matching unit is the neuron in S1 that has completed the operating condition classification.

[0066] In an optional embodiment, during online system operation, exhaust gas concentration, temperature, and pipeline pressure are collected for the current control cycle and the past nine control cycles, forming a 30-dimensional real-time data window vector. This real-time data window vector is input into a pre-trained offline VAE encoder, which outputs a 2-dimensional latent feature vector for the current operating condition. The Euclidean distance between the latent feature vector and the weight vectors of all 100 neurons in the self-organizing map network is calculated. The neuron with the smallest Euclidean distance is the best matching unit for the current operating condition; this best matching unit is the neuron that has already completed the operating condition classification.

[0067] S4, when the best matching unit is a stable operating condition neuron, if the absolute value of the current control error is greater than the high threshold, the optimal solution of the ITAE is selected; if the absolute value of the current control error is less than the low threshold, the optimal solution of the control output change rate is selected; if the absolute value of the current control error is between the low threshold and the high threshold, the balanced solution is selected.

[0068] In an optional embodiment, after determining that the current best matching unit is the steady-state neuron, the difference between the current setpoint and the measured value is further calculated to obtain the control error e(t). The absolute value of the control error |e(t)| is compared with the preset high-order threshold and low-order threshold. Assuming the high-order threshold is 5 and the low-order threshold is 1, if |e(t)| is greater than 5, it indicates a large deviation, and the controller uses the PID parameter set corresponding to the optimal solution of ITAE. If |e(t)| is less than 1, it indicates that the process is close to stabilization, and to avoid overtuning and output jitter, the controller switches to the PID parameter set corresponding to the optimal solution of the control output change rate. If the value of |e(t)| is between 1 and 5, the PID parameter set corresponding to the balanced solution that takes into account both response speed and stability is selected.

[0069] In an optional embodiment, in S4, when the best matching unit is a transitional working condition neuron, the neighboring stable working condition neurons of the best matching unit on the two-dimensional neuron topology grid are determined, and the balance solutions corresponding to each neighboring stable working condition neuron are weighted and fused according to the Gaussian kernel function value between the potential feature vector of the current working condition and the weight vector of each neighboring stable working condition neuron to obtain the current PID parameters.

[0070] Specifically, if the current best matching unit is a transitional neuron, we search for its eight neighboring neurons on the two-dimensional neuron topology grid, and then select all neighboring neurons marked as stable conditions, assuming N such neurons are found. For each stable condition neighboring neuron, we calculate the Euclidean distance between the current latent feature vector Z and the weight vector of the neighboring neuron, and substitute the Euclidean distance into the Gaussian kernel function to obtain a weight value; the smaller the distance, the larger the weight value. We normalize the obtained N weight values ​​so that their sum is 1. We extract the PID parameter set of the equilibrium solution corresponding to each of the N stable condition neighboring neurons, where the PID parameter set is obtained by weighted averaging of the N PID parameters, for example, K... p The value is equal to the K value of the neighboring neurons in each steady-state condition. p The values ​​are multiplied by their corresponding normalized weights and then summed.

[0071] In an optional embodiment, when the best matching unit is a transient condition neuron, the process of weightedly fusing the equilibrium solutions corresponding to each neighboring stable condition neuron to obtain the current PID parameters includes the following steps:

[0072] On a two-dimensional neuron topology grid, multiple stable working condition neurons that are adjacent to the best matching unit are selected based on topological distance;

[0073] The similarity between the latent feature vector of the current working condition and the weight vector of each neighboring stable working condition neuron is calculated using the Gaussian kernel function, and this similarity is used as the weight.

[0074] The calculated weights are normalized, and the PID parameters corresponding to the equilibrium solutions of neurons in each nearby stable condition are weighted and summed to obtain the current PID parameters.

[0075] Specifically, when the operating condition switches from an unstable region to another stable region, and the state is mapped to the transitional operating condition neuron, the controller parameters need to be smoothly adjusted. For example, the current best matching unit is located at coordinates (5, 6) in the two-dimensional neuron topology grid, which is a transitional neuron. Centered on this transitional neuron, three neighboring stable operating condition neurons are found within a radius of 2, located at coordinates (4, 6), (6, 5), and (7, 7), respectively. Each of the three stable operating condition neurons has a set of pre-calculated PID parameters corresponding to the equilibrium solution. The similarity between the current latent feature vector and the weight vectors of the three stable operating condition neurons is calculated using a Gaussian kernel function. Assuming the calculated similarities are 0.8, 0.5, and 0.3, they are normalized by dividing each value by the sum of 1.6, resulting in new weights of 0.5, 0.3125, and 0.1875. The three normalized weights are multiplied by the PID parameters corresponding to the equilibrium solutions of their respective stable operating condition neurons, and then summed to obtain a set of smoothly interpolated PID parameters. For example, K after smooth interpolation p The value of K is the neuron in the first stable condition. p Multiply the value by 0.5, and add the K value of the second stable neuron. p Multiply the value by 0.3125, and add the K of the third stabilizing neuron. p Multiply the value by 0.1875.

[0076] In an optional embodiment, in S4, when the best matching unit is the disturbance condition neuron, the proportional and derivative links of the PID controller are maintained, the integral link is reset according to the recovery function, and the feedforward compensation filter associated with the disturbance condition neuron is enabled.

[0077] Specifically, if the current optimal matching unit is the disturbance condition neuron, the proportional parameter K in the PID controller... p and differential parameter K d The value from the previous moment remains unchanged. The integral term is reset, and its new value is calculated by a recovery function. For example, the new integral value equals an estimated steady-state output value associated with the disturbance condition minus the current proportional term output, to eliminate integral saturation. Simultaneously, the feedforward compensator set for the disturbance is activated. The feedforward compensator calculates the compensation amount based on one or more measurable disturbance source signals and adds the compensation amount to the total output of the PID controller, thereby intervening before the disturbance affects the system. Here, the disturbance source signal can be a sudden change in the upstream feed flow rate.

[0078] In an optional embodiment, the process of maintaining the proportional and derivative elements of the PID controller, resetting the integral element according to the recovery function, and enabling the feedforward compensation filter associated with the disturbance condition neuron includes the following steps:

[0079] Keep the proportional P and derivative D parameters of the PID controller unchanged, and reset the cumulative state value of the integral I element to the preset value;

[0080] Enable the pass function as The feedforward compensation filter, where the gain Here are the gain parameters of the feedforward compensation filter. , It is a time constant. It is a variable in the complex frequency domain. The output of the feedforward compensation filter is superimposed on the output of the PID controller, and together they act on the control actuator.

[0081] Specifically, when a large, predictable disturbance (such as a sudden change in composition) causes the operating condition to be identified as a disturbance, the proportional parameter P, responsible for the immediate response, and the derivative parameter D, responsible for the predictive response, in the PID controller will remain unchanged to maintain basic control. The integral element I, responsible for eliminating static error, may experience overshoot or saturation due to accumulated error over time, i.e., integral saturation, when a disturbance occurs. To avoid this problem, the accumulated value of the integral element is reset to a safe initial value, such as zero. A feedforward compensator specifically designed to handle such disturbances is activated. The feedforward compensation filter is preferably a dynamic filter, whose parameters (such as gain) are... and time constant , The feedforward controller identifies the specific disturbance based on historical data. It measures the disturbance source and calculates a compensating control signal, where the measured disturbance source can be a raw material composition analysis value or temperature, etc. The compensating control signal is added to the output signal of the PID controller after integral reset to form a control command, which is then sent to the actuator.

[0082] In one embodiment, during the offline analysis phase, typical disturbance events and their impact on the exhaust gas system have been identified through historical data analysis. Gain The time constant determines the strength of the compensation effect. , These factors collectively determine the response speed and form of the compensation effect (e.g., whether it's a rapid impact followed by a decline, or a gradual increase), thus generating a control action. This control action is equal in magnitude and opposite in direction to the impact of the disturbance on the system, canceling it out when the disturbance occurs. A typical disturbance condition is the periodic discharge of material from the upstream polymerization reactor, which leads to VOCs in the exhaust gas pipeline. s The concentration experienced a short-lived, pulsed spike. Upon detecting this condition, the online monitoring system activated the compensator. Given that the spike disturbance would require an increase of 15Hz in the recovery fan frequency for effective suppression, and that the disturbance's impact lasted approximately 60 seconds, therefore... Set to 15 to ensure sufficient compensation; at the same time, to take action before the concentration fully increases, it can be set to... For 10 seconds The time is set to 5 seconds, allowing the compensation action to respond faster than the disturbance itself, thus preventing excessive exhaust emissions.

[0083] An embodiment of the resin production waste gas recovery control system provided by the present invention:

[0084] like Figure 7 As shown, the resin production waste gas recovery control system includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned resin production waste gas recovery control method is implemented.

[0085] The resin production waste gas recovery control system also includes other components well known to those skilled in the art, such as communication interfaces. Their setup and functions are known in the art and will not be described in detail here.

Claims

1. A resin production waste gas recovery control method characterized by, The method comprises the following steps: S1, obtaining historical time series data of exhaust gas concentration, temperature and pipeline pressure, dividing the data into several data windows, converting each data window to a preset dimension latent space by using a pre-trained variational autoencoder to obtain a single latent feature vector, and inputting a set of historical latent feature vectors formed by collecting all the latent feature vectors into a self-organizing mapping network for offline training to obtain a two-dimensional neuron topology grid; according to the variance of the process variables associated with the latent feature vectors, the neurons are divided into stable working condition neurons, transition working condition neurons and disturbance working condition neurons; S2, obtaining a local linear process model corresponding to each stable working condition neuron, and solving by using a non-dominated sorting genetic algorithm to minimize ITAE and minimize the control output rate as a double optimization target to obtain a Pareto optimal solution set composed of PID parameter groups, including ITAE optimal solution, control output rate optimal solution and balance solution; S3, in each control period, real-time acquisition of the current exhaust gas concentration, temperature and pipeline pressure, calculation of the latent feature vector of the current working condition by using the variational autoencoder, and determination of the best matching unit of the latent feature vector in the self-organizing mapping network, the best matching unit being the neuron in S1 which has completed working condition classification; S4, when the best matching unit is a stable working condition neuron, selecting the ITAE optimal solution when the absolute value of the current control error is greater than the high threshold value, selecting the control output rate optimal solution when the absolute value of the current control error is less than the low threshold value, and selecting the balance solution when the absolute value of the current control error is between the two.

2. The resin production exhaust gas recovery control method according to claim 1, characterized by, In S1, the process of dividing the neurons into stable working condition neurons, transition working condition neurons and disturbance working condition neurons comprises the following steps: Calculate the variance of the exhaust gas concentration, temperature and pipeline pressure corresponding to the latent feature vector associated with each neuron to obtain a comprehensive variance index representing the working condition volatility of each neuron; Arrange the comprehensive variance indexes of all neurons in ascending order; The neurons with a comprehensive variance index lower than a preset low quantile point are regarded as stable working condition neurons, the neurons with a comprehensive variance index higher than a preset high quantile point are regarded as disturbance working condition neurons, and the remaining neurons are regarded as transition working condition neurons.

3. The resin production exhaust gas recovery control method according to claim 1, characterized by, The variational autoencoder is composed of an encoder and a decoder, and both the encoder and the decoder adopt a multilayer perceptron structure; During offline training, the historical time series data is input into the VAE, and the network weights are adjusted by minimizing the reconstruction error and the KL divergence.

4. The resin production exhaust gas recovery control method according to claim 1, characterized by, In S2, the process of obtaining a Pareto optimal solution set composed of PID parameter groups comprises the following steps: Initialize a population of multiple sets of PID parameters [K p , K i , K d ]; Generate a child population by performing crossover and mutation operations on the parent population; Select excellent individuals from the parent population and the child population to form a new generation population based on non-dominated sorting and crowding degree calculation; Repeat the iteration until the termination condition is met, and the obtained non-dominated solution set is regarded as the Pareto optimal solution set.

5. The resin production exhaust gas recovery control method according to claim 1, characterized by, In S4, when the best matching unit is a transition working condition neuron, neighboring stable working condition neurons of the best matching unit on the two-dimensional neuron topological grid are determined, and the equilibrium solutions corresponding to the neighboring stable working condition neurons are weighted and fused according to the Gaussian kernel function values between the potential feature vector of the current working condition and the weight vectors of the neighboring stable working condition neurons, to obtain the current PID parameters.

6. The resin production exhaust gas recovery control method according to claim 5, characterized by, When the best matching unit is a transition working condition neuron, the process of weighted fusion of the equilibrium solutions corresponding to the neighboring stable working condition neurons to obtain the current PID parameters includes the following steps: On the two-dimensional neuron topological grid, a plurality of stable working condition neurons neighboring the best matching unit are selected according to the topological distance; The similarity between the potential feature vector of the current working condition and the weight vectors of the neighboring stable working condition neurons is calculated through a Gaussian kernel function, and the similarity is taken as the weight; The calculated weight is normalized, and the PID parameters corresponding to the equilibrium solutions of the neighboring stable working condition neurons are weighted and summed, to obtain the current PID parameters.

7. The resin production exhaust gas recovery control method according to claim 5, characterized by, If the current best matching unit is a transition working condition neuron, eight neighbor neurons around the best matching unit are found on the two-dimensional neuron topological grid, and all the neighbor neurons marked as stable working condition are selected from the eight neighbor neurons, and are recorded as the neighboring stable working condition neurons.

8. The resin production exhaust gas recovery control method according to claim 1, characterized by, In S4, when the best matching unit is a disturbance working condition neuron, the proportional and derivative links of the PID controller are maintained, the integral link is reset according to the recovery function, and the feedforward compensation filter associated with the disturbance working condition neuron is enabled.

9. The resin production exhaust gas recovery control method according to claim 8, characterized by, The process of maintaining the proportional and derivative links of the PID controller, resetting the integral link according to the recovery function, and enabling the feedforward compensation filter associated with the disturbance working condition neuron includes the following steps: The proportional P and derivative D link parameters of the PID controller are maintained, and the accumulated state value of the integral I link is reset to a preset value. The transfer function of the feedforward compensation filter is enabled wherein the gain is a gain parameter of the feedforward compensation filter, , is a time constant, is a complex frequency domain variable, the output of the feedforward compensation filter is superimposed with the output of the PID controller, and the two jointly act on the control actuator.

10. A resin production waste gas recovery control system, characterized in that, The device comprises a memory and a processor, and the memory stores computer program instructions, which, when executed by the processor, implement the resin production waste gas recovery control method of any one of claims 1-9.

Images