Method for predicting outlet moisture and controlling water addition ratio in loose conditioning process

Abstract

The present application relates to the technical field of tobacco processing, and discloses an outlet moisture prediction and water addition proportion control method for a loose rehydration process, which comprises the following steps: obtaining real-time production data to calculate water vapor proportion characteristics, extracting an instantaneous flow rate change rate to match a physical lag time, dynamically determining a delay queue length to complete causal alignment of data; constructing a multiple linear regression equation to generate a forward prediction model, and inversely solving the prediction model to obtain total water vapor demand; configuring a preliminary proportion instruction according to a deviation, and when it is determined that the water vapor demand is out of range, combining an evaporation compensation coefficient to convert a compensation amount and add it to the water addition proportion to generate a final instruction; constructing a loss function to extract a gradient data sequence, and feeding back and updating a model intercept. The present application excludes transmission lag interference, overcomes insufficient humidification capacity caused by single-channel hardware limitation, eliminates model static deviation caused by environmental drift, and improves moisture prediction and control precision.

Description

Technical Field

[0001] This invention relates to the field of tobacco processing technology, specifically to a method for predicting the outlet moisture content and controlling the water addition ratio in the loosening and rehydration process. Background Technology

[0002] In the tobacco processing, the loosening and rehumidification process is used to heat and humidify the material to meet the needs of subsequent processing. Existing moisture control schemes use proportional-integral-derivative control logic or static regression models, issuing water and steam addition adjustment commands based on inlet environmental parameters. In actual production, the material needs to go through a certain period of time to be transferred inside the drum, which causes a time misalignment between the outlet moisture content value detected by the sensor and the environmental variable value collected at the inlet. Conventional control systems often directly read the data at the same sampling time for fitting calculations without dynamically aligning the physical lag time caused by the transfer process in combination with the production status. This leads to a misalignment of the causal relationship between the model input and output, reducing the accuracy of the prediction model in calculating transient moisture changes.

[0003] In the execution control stage, the loosening and rehumidification process involves the dual effects of moisture increase and heat transfer. When the inlet material flow rate changes abruptly, the system's set steam demand increases accordingly. Due to the inherent physical opening limit of the steam regulating valve in the bottom pipeline, when the calculated steam demand exceeds the hardware limit, the existing control program can only output the valve's full-load opening command. The existing program cannot transfer the steam task exceeding the limit to the ambient temperature water supply channel with a margin, resulting in a limitation on the overall humidification capacity of the system, causing the actual outlet moisture content to fail to reach the set target.

[0004] In addition to transmission alignment and instruction allocation issues, as the continuous operation time of the equipment increases, long-term changes in ambient temperature and humidity and scaling inside the equipment pipes will cause the constant terms in the original prediction model to deviate from the actual physical state. The existing control system lacks a closed-loop feedback correction mechanism for the underlying parameters of the regression model and cannot synchronously update the model intercept and feature sensitivity weights, resulting in the control system continuously generating static control deviations that are difficult to eliminate during long-term operation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting outlet moisture content and controlling the water addition ratio in a loose rehumidification process. This method solves the problems of insufficient prediction accuracy due to lack of data time causal alignment in conventional control models, lack of total humidification capacity due to the inability to compensate across channels when steam addition channels are limited, and long-term static control deviation caused by the lack of a low-level parameter feedback and update mechanism.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the outlet moisture content and controlling the water addition ratio in a loosening and rehydration process, the method comprising the following steps: Acquire real-time production data including inlet moisture content, return air temperature, material flow rate, water addition, steam addition, and actual outlet moisture content; calculate water addition ratio characteristics and steam addition ratio characteristics; extract instantaneous flow rate change rate, match physical lag time, dynamically determine delay queue length, and complete causal alignment between input feature parameters and actual outlet moisture content; Using the input feature parameters as input independent variables and the actual outlet moisture content as the target dependent variable, a multiple linear regression equation is constructed to generate a forward prediction model and model intercept; the target outlet moisture content is substituted into the forward prediction model to perform inverse calculation to obtain the total water vapor demand. Extract the absolute deviation value, dynamically configure the steam addition distribution coefficient based on the comparison status, and calculate the initial steam addition ratio and the initial water addition ratio; when it is determined that there is a thermodynamic over-limit requirement, add the water compensation term value to the initial water addition ratio to calculate and generate the final water addition ratio instruction, and limit the final steam addition ratio to the steam addition ratio limit boundary value. Construct a target loss function, calculate the standard deviation of the bias time series, perform derivative operation on the backpropagation error function to extract the gradient data sequence of the corresponding interval, multiply by the base learning rate parameter to generate the intercept correction compensation amount, and feed back the intercept correction compensation amount to update the model intercept.

[0007] The underlying principle is as follows: First, by dynamically determining the delay queue length, the temporal causal alignment of multidimensional feature variables is achieved, eliminating the interference of physical transmission delays on the modeling process. Second, a forward and reverse calculation mechanism is constructed to convert multivariate coupled parameters into a total water and steam demand term. Third, a thermodynamic boundary compensation mechanism is established based on the physical limits of the regulating valve. When the steam supply channel is limited, the demand is compensated by the water supply amount using the evaporation compensation coefficient, thus achieving decoupling of control commands. Fourth, a backpropagation error function based on deviation segmentation and gradient calculation are introduced to update the model intercept, eliminating the static control residual error introduced by environmental temperature and humidity drift.

[0008] In a preferred embodiment of the present invention, the steps of calculating the water addition ratio feature and the steam addition ratio feature include: dividing the collected water addition amount and steam addition amount by the material flow rate at the same moment to generate the water addition ratio feature and the steam addition ratio feature respectively; when the material flow rate at the same moment is less than the preset minimum effective flow rate threshold, calling the material flow rate at the previous effective sampling moment to participate in the ratio feature calculation, or marking the current sampling point as an invalid sample and excluding it from the subsequent modeling dataset.

[0009] The principle behind the above settings is to prevent the denominator from approaching zero in division operations due to equipment idling or transient material interruption, thus avoiding calculation abnormalities, ensuring the numerical continuity of the proportional feature sequence in the continuous production process, and avoiding invalid data from interfering with the subsequent model fitting dataset.

[0010] Furthermore, the steps of extracting the instantaneous flow rate change rate, matching the physical lag time, dynamically determining the delay queue length, and completing the causal alignment between the input feature parameters and the actual outlet moisture content include: using a dynamic time warping algorithm to calculate the time step offset when the historical input environmental parameter sequence and the historical actual outlet moisture content sequence reach the maximum morphological similarity, multiplying it by the sampling time interval to calculate the physical lag time, and constructing a priori mapping matrix; performing linear interpolation retrieval in the priori mapping matrix based on the extracted instantaneous flow rate change rate to extract the physical lag time corresponding to the current transient condition, dividing the physical lag time by the sampling time interval and rounding to obtain the queue depth control parameter to dynamically determine the delay queue length; adjusting the residence time of data nodes inside the delay queue and the delay queue length based on the queue depth control parameter, and extracting the historical input features popped from the head of the delay queue and combining them with the actual outlet moisture content collected at the current moment.

[0011] The principle behind the above setting is as follows: the maximum morphological similarity of the nonlinear time series is calculated using the dynamic time warping algorithm, the unsteady hysteresis time in fluid dynamics is extracted, and the causal synchronization of the feature sequence at the input end of the model and the actual water content at the output end is ensured through linear interpolation and depth adjustment of the first-in-first-out queue.

[0012] Furthermore, the steps for fitting and constructing a multiple linear regression equation using the input feature parameters as input independent variables and the actual outlet moisture content as the target dependent variable, and generating a positive prediction model and model intercept, include: using Laida's rule to remove outlier values ​​in the actual outlet moisture content sequence whose absolute deviation is greater than three times the standard deviation of the distribution, and simultaneously removing the input feature parameter samples at the timestamps corresponding to the outlier values; using the max-min normalization algorithm to linearly map the extreme value range of the input feature parameter sequence to a closed interval between 0 and 1, and simultaneously storing the maximum and minimum values ​​corresponding to each input feature parameter; and calling the ordinary least squares algorithm to calculate the sensitivity weights of each feature in the multiple linear regression equation and the model intercept through fitting operations.

[0013] The principle behind the above setup is as follows: outliers are removed using statistical methods, dimensional differences and numerical range gaps between multi-source sensor data are eliminated using linear mapping, and the linear sensitivity of each physical quantity to outlet moisture is extracted using the least squares algorithm to complete the construction of the underlying regression equation.

[0014] In a preferred embodiment of the present invention, the step of substituting the target outlet moisture content into the forward prediction model to perform reverse calculation and obtain the total water vapor demand includes: performing forward normalization transformation on the inlet moisture content input parameter and the return air temperature input parameter collected in real time at the current moment; replacing the predicted outlet moisture content variable on the left side of the forward prediction model equation with the externally set target outlet moisture content value; performing algebraic transposition on the linear equation of the forward prediction model, moving the product term containing the inlet moisture content and the return air temperature to the target outlet moisture content side, and extracting the product term containing the water addition ratio feature and the steam addition ratio feature to construct the total water vapor demand.

[0015] The principle behind the above setting is to transform the prediction model into an equation solution form, separate uncontrollable environmental variables from controllable execution variables through algebraic rearrangement, and calculate the total water and vapor load required by the system based on the target water demand, providing basic data for the decoupled allocation of subsequent control commands.

[0016] Furthermore, the steps for dynamically configuring the steam allocation coefficient based on the comparison status and calculating the initial steam and water allocation ratios include: multiplying the system index tolerance value by a preset ratio coefficient to generate a judgment threshold; configuring a first steam allocation coefficient when the absolute deviation value is greater than the judgment threshold, and configuring a second steam allocation coefficient when the absolute deviation value is less than or equal to the judgment threshold, with the value of the first steam allocation coefficient being forcibly set to be greater than the value of the second steam allocation coefficient; calculating the initial steam and water allocation ratios by combining the total water and steam demand item, the configured steam allocation coefficients, and the feature sensitivity weights, and performing an inverse normalization transformation on the calculation results.

[0017] The principle behind the above settings is as follows: when the outlet moisture content deviates significantly from the set value, the adjustment weight of the steam injection channel is increased to improve the system response speed by utilizing the physical characteristics of rapid steam diffusion and heat exchange; when the deviation decreases, the weight of the steam injection channel is reduced to prevent overshoot fluctuations in the steady-state stage caused by excessive heat input.

[0018] Furthermore, when a thermodynamic overrun requirement is determined, the steps of superimposing the moisture compensation value into the initial water addition ratio to calculate the final water addition ratio instruction and limiting the final steam addition ratio to the steam addition ratio limit boundary value include: retrieving the hardware physical limit opening value associated with the regulating valve of the underlying steam addition pipeline, mapping the hardware physical limit opening value to the steam addition ratio limit boundary value; comparing the initial steam addition ratio with the steam addition ratio limit boundary value to determine the overrun status, and determining that a thermodynamic overrun requirement exists when the initial steam addition ratio is greater than the steam addition ratio limit boundary value. Before executing the cutoff, the initial steam addition ratio is temporarily stored as the original initial steam addition ratio. The difference in steam addition demand beyond the limit is calculated by subtracting the limit boundary value from the original initial steam addition ratio. The difference in steam addition demand beyond the limit is multiplied by the thermodynamic evaporation compensation coefficient and converted into a moisture compensation value. The moisture compensation value is added to the initial water addition ratio to calculate and generate the final water addition ratio instruction. The final steam addition ratio is limited to the limit boundary value of the steam addition ratio.

[0019] The principle behind the above setting is as follows: based on the conversion relationship between thermodynamic enthalpy and latent heat of vaporization, the steam demand exceeding the limit is converted into water compensation amount to prevent insufficient overall humidification caused by the saturation limitation of the steam supply channel.

[0020] In a preferred embodiment of the present invention, the steps of constructing a target loss function, calculating the standard deviation of the deviation time series, performing derivative operation on the backpropagation error function to extract the gradient data series of the corresponding interval, multiplying by the basic learning rate parameter to generate the intercept correction compensation amount, and feeding back the intercept correction compensation amount to update the model intercept include: determining that the system is in the steady-state mass transfer interval when the instantaneous flow rate change rate is less than or equal to the steady-state flow fluctuation threshold, calculating the deviation time series between the actual outlet moisture content and the target outlet moisture content, and extracting the standard deviation of the deviation time series; Based on the comparison results between the absolute deviation value and the judgment threshold, a piecewise backpropagation error function is configured. The standard deviation value of the specified historical time domain and the dynamic judgment threshold are introduced as piecewise scaling factor coefficients to construct the target loss function. The derivative operation is performed on the backpropagation error function to extract the gradient data sequence of the corresponding interval. The gradient data sequence is multiplied by the basic learning rate parameter to generate the intercept correction compensation amount. The intercept correction compensation amount is fed back and superimposed into the intercept constant term of the original model of the multiple linear regression equation to perform parameter overwriting operation.

[0021] The principle behind the above setting is to limit parameter updates to only be performed within the steady-state range to eliminate transient disturbances, and to output local penalties or fine-tune the convergence gradient in different deviation ranges by using piecewise scaling factors to correct the constant term of the regression equation and compensate for the static deviations of the model caused by pipe scaling and changes in ambient temperature.

[0022] Furthermore, the step of introducing the standard deviation value of the specified historical time domain and the dynamic judgment threshold as piecewise scaling factor coefficients to construct the target loss function includes: when the standard deviation value in the specified historical time domain is less than the preset minimum standard deviation threshold, the standard deviation value is replaced with the preset minimum standard deviation threshold to participate in the error gradient calculation.

[0023] The principle behind the above setting is to set a lower limit constraint to prevent the standard deviation from approaching zero due to long-term stable operating conditions, which could lead to abnormal amplification of the gradient result in the division operation and ensure the stability of the operation process.

[0024] Furthermore, while performing the step of feeding back and superimposing the intercept correction compensation to update the model intercept, a long-term parameter update operation is also triggered. The steps include: calculating the coefficient of variation of the series of environmental variables including inlet water content within the long-term period; when the coefficient of variation exceeds the set limit, extracting the corresponding input feature parameter series, using each sampled input feature parameter after positive normalization within the long-term window as a partial derivative multiplier factor to generate the weight error correction compensation for each feature channel; feeding back and superimposing the weight error correction compensation of each feature sensitivity weight to the original feature sensitivity weight corresponding to the multiple linear regression equation to perform a dynamic overwrite operation; when the prediction error exceeds the preset error limit, reverting to the feature sensitivity weight before the update.

[0025] The principle behind the above settings is as follows: the long-term fluctuation of environmental variables triggers the overall update of sensitivity weights, long-term normalized data is used to synchronously update the parameter matrix, and a backoff mechanism based on test error is established to ensure the stability and reliability of the model parameter adjustment process.

[0026] This invention provides a method for predicting the outlet moisture content and controlling the water addition ratio in a loosening and rehydration process. It has the following beneficial effects: 1. This invention extracts the instantaneous flow rate change rate, matches the physical lag time between the input feature parameters and the actual outlet moisture content, and dynamically adjusts the delay queue length to complete the causal alignment of the data. The alignment step eliminates the time misalignment interference between material transmission and sensor sampling in the loose rehydration process, ensures the time synchronization of the dataset fitted by the multiple linear regression equation, and improves the calculation accuracy of the positive prediction model.

[0027] 2. This invention obtains the total water vapor demand by reverse solving the forward prediction model, and when it is determined that the steam addition ratio exceeds the physical limit of the hardware valve, it uses the thermodynamic evaporation compensation coefficient to convert the excess steam addition demand into a moisture compensation term and adds it to the water addition ratio command. The decoupled compensation process avoids the problem of limited total humidification capacity caused by the full load of the bottom steam addition valve, and ensures the outlet moisture compliance rate of the system under the condition of sudden change in material flow.

[0028] 3. This invention utilizes a piecewise backpropagation error function to perform differentiation, combines the basic learning rate to calculate the intercept correction compensation amount, and feeds it back into the model intercept. When environmental variables change, the feature sensitivity weights are updated synchronously. The parameter update action eliminates the static calculation deviation of the model caused by pipe scaling or long-term drift of environmental temperature and humidity, ensuring the predictive stability of the control system in long-term continuous production processes. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the system timing interaction of the present invention; Figure 2This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the data processing and causal alignment process of the present invention; Figure 4 This is a schematic diagram illustrating the algebraic deformation of the target outlet moisture content and the decoupling of water vapor demand in this invention. Figure 5 This is a schematic diagram of the decoupling control and feedback evolution process of the present invention; Figure 6 This is a schematic diagram of the comparative test curves of the control effect of the present invention under sudden inlet moisture interference.

[0030] Among them, 100 is the data processing module; 200 is the predictive modeling module; 300 is the decoupling control module; and 400 is the feedback evolution module. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] See attached document Figure 1 The present invention provides an outlet moisture prediction and water addition ratio control system for a loosening and rehydration process. The system includes: a data processing module 100, a prediction modeling module 200, a decoupling control module 300, and a feedback evolution module 400.

[0033] The data processing module 100 receives real-time production data including inlet moisture content, return air temperature, material flow rate, water addition, steam addition, and actual outlet moisture content. The data processing module 100 divides the water addition and steam addition by the material flow rate to obtain water addition ratio characteristics and steam addition ratio characteristics, and performs invalid sample marking or substitution calculation processing on sampling points that are lower than the preset minimum effective flow rate threshold. The data processing module 100 performs differential operation on the material flow rate to extract the instantaneous flow rate change rate, matches the physical lag time in the prior mapping matrix based on the instantaneous flow rate change rate, and dynamically adjusts the delay queue length in combination with the sampling time interval to achieve causal alignment between the input feature parameters and the actual outlet moisture content.

[0034] The prediction modeling module 200 receives the causally aligned dataset, uses inlet moisture content, return air temperature, water addition ratio characteristics, and steam addition ratio characteristics as input independent variables, and uses the actual outlet moisture content as the target dependent variable to fit and construct a multiple linear regression equation. The prediction modeling module 200 generates a positive prediction model, feature sensitivity weights, and model intercept, and stores the maximum and minimum values ​​in the normalization process as extreme value parameters for subsequent physical dimension recovery.

[0035] The decoupling control module 300 substitutes the externally set target outlet moisture content into the forward prediction model to calculate the total water and steam demand. The decoupling control module 300 calculates the absolute deviation between the actual outlet moisture content and the target outlet moisture content. Based on the comparison between the absolute deviation and the preset threshold, the total water and steam demand is allocated to the initial steam addition ratio and the initial water addition ratio, and physical dimension restoration is performed. The decoupling control module 300 determines the out-of-bounds demand based on the physical limits of the regulating valve. The out-of-bounds demand exceeding the extreme value is converted into a moisture compensation term by combining the evaporation compensation coefficient and superimposed on the initial water addition ratio. The final water addition ratio command and the final steam addition ratio command are determined based on the state where there is no out-of-bounds demand. The output is performed after the upper and lower limit constraints are executed.

[0036] The feedback evolution module 400 establishes a piecewise loss function based on absolute deviation. In the short time domain period, it calculates the standard deviation of the deviation sequence between the actual outlet moisture content and the target outlet moisture content, and updates the model intercept based on the derivative of the piecewise loss function in combination with the instantaneous flow rate change rate. In the long time domain period, it calculates the coefficient of variation including environmental variables such as inlet moisture content. When the coefficient of variation exceeds the set limit, it extracts the corresponding input feature parameter sequence, updates the feature sensitivity weights according to the gradient descent algorithm, and performs error verification on the updated feature sensitivity weights to determine whether to activate or revert to the feature sensitivity weights before the update.

[0037] See attached document Figure 2 This invention provides a method for predicting the outlet moisture content and controlling the water addition ratio in a loosening and rehydration process, the method comprising the following steps: S100 acquires real-time production data including inlet moisture content, return air temperature, material flow rate, water addition, steam addition, and actual outlet moisture content; calculates water addition ratio characteristics and steam addition ratio characteristics; performs invalid sample marking or alternative calculation processing on sampling points where the material flow rate is lower than the preset minimum effective flow rate threshold; performs differential operation on the material flow rate to extract the instantaneous flow rate change rate; matches the physical lag time in the prior mapping matrix; dynamically determines the delay queue length by dividing the physical lag time by the sampling time interval; and completes the causal alignment between the input feature parameters and the actual outlet moisture content.

[0038] S200 receives the causally aligned dataset, uses inlet moisture content, return air temperature, water addition ratio characteristics, and steam addition ratio characteristics as input independent variables, and uses the actual outlet moisture content as the target dependent variable to fit and construct a multiple linear regression equation, generating a positive prediction model, feature sensitivity weights, and model intercept, storing the maximum and minimum values ​​in the normalization process, and maintaining the effective model or refitting when the verification does not meet the preset error conditions; substitutes the target outlet moisture content into the positive prediction model to perform inverse calculation and obtain the total water and steam demand item.

[0039] S300, based on the comparison between the absolute deviation of the actual outlet moisture content and the target outlet moisture content and the preset threshold, allocates the total water and steam demand into a preliminary steam addition ratio and a preliminary water addition ratio and performs physical dimension restoration; judges out-of-bounds demand based on the physical limits of the regulating valve, if there is an out-of-bounds demand, it is converted into a moisture compensation item by combining the evaporation compensation coefficient and added to the preliminary water addition ratio, and determines the final water addition ratio command and the final steam addition ratio command respectively based on the non-out-of-bounds case, and issues the final water addition ratio command and the final steam addition ratio command after applying upper and lower limit constraints.

[0040] S400 establishes a loss function that includes the deviation segment interval. In the short time domain period, it calculates the standard deviation of the time series of deviations between the actual outlet moisture content and the target outlet moisture content. When the standard deviation is less than the preset minimum standard deviation threshold, it replaces the standard deviation with the preset minimum standard deviation threshold for calculation. It calculates and updates the model intercept based on the derivative value of the loss function in combination with the instantaneous flow rate change rate. In the long time domain period, it calculates the coefficient of variation of the series of environmental variables including inlet moisture content. When it exceeds the set limit, it extracts the input feature parameter series and updates all feature sensitivity weights according to the gradient descent algorithm. After error verification, it decides whether to participate in subsequent cycles or revert to the feature sensitivity weights before the update.

[0041] See attached document Figure 3 The specific implementation of step S100 is carried out through the following sub-steps: S101, the data processing module 100 acquires real-time production data of the loosening and rehumidification process through the field communication network. The real-time production data includes the inlet moisture content collected by the equipment set on the inlet side of the heating and humidifying cylinder, the return air temperature collected by the return air duct sensor, and the material flow rate collected by the electronic belt scale. At the same time, the data processing module 100 acquires the water and steam addition amount fed back by the control pipeline and the actual outlet moisture content fed back by the equipment on the outlet side of the cylinder. The data processing module 100 synchronously generates real-time production data according to the set sampling frequency. The data processing module 100 configures the value of the set sampling frequency to be a continuous real number within a closed interval of 1Hz to 5Hz. The data processing module 100 stores the real-time production data in the local working memory. The data processing module 100 divides the collected water addition and steam addition amounts by the material flow rate at the same moment to generate water addition ratio characteristics and steam addition ratio characteristics. When the material flow rate at the same moment is less than the preset minimum effective flow rate threshold, and the value range of the preset minimum effective flow rate threshold is configured to be 0.05 times to 0.1 times the nominal material flow rate.

[0042] The data processing module 100 calls the material flow rate from the previous valid sampling time to participate in the proportional feature calculation, or marks the current sampling point as an invalid sample and excludes it from the subsequent modeling dataset, in order to avoid abnormal proportional feature calculation caused by material flow rate being too low or zero. The data processing module 100 performs the proportional feature conversion calculation according to the following mathematical formula: ; ; In the formula, Input parameters for the water addition ratio feature; This refers to the real-time data on the amount of water added. This refers to the real-time collected material flow rate; Input parameters for the steam refueling ratio characteristic; This is the real-time data on the amount of steam added.

[0043] S102, the data processing module 100 retrieves the material flow time series in the local working memory, the data processing module 100 retrieves the sampling time interval corresponding to the set sampling frequency, the data processing module 100 performs a first-order forward difference operation on the material flow time series, and the data processing module 100 subtracts the material flow at the previous sampling time from the material flow at the current moment and divides it by the sampling time interval to extract the instantaneous flow rate change. The data processing module 100 records the instantaneous flow rate change rate into the state monitoring register as the basic trigger value for determining the dynamic transition state of the physical mass transfer process; for other difference processing details included in the discrete time series differential operation, the data processing module 100 synchronously performs boundary smoothing operation according to the existing numerical analysis algorithm.

[0044] S103, the data processing module 100 constructs a priori mapping matrix based on the historical operating dataset. The priori mapping matrix stores the mapping values ​​between the instantaneous flow rate change rate and the physical lag time in the form of a data dictionary. The data processing module 100 extracts the historical input environmental parameter sequence and the historical actual outlet moisture content sequence. The data processing module 100 uses a dynamic time warping algorithm to calculate the time step offset when the historical input environmental parameter sequence and the historical actual outlet moisture content sequence reach the maximum morphological similarity under different instantaneous flow rate change rate ranges. The data processing module 100 calculates the physical lag time by multiplying the time step offset by the sampling time interval. The data processing module 100 records the paired values ​​of instantaneous flow rate change rate and physical lag time and updates them in the storage space of the prior mapping matrix. For the specific time series distance metric calculation and optimal path backtracking solution operation involved in the dynamic time warping algorithm, the data processing module 100 calls the general time series similarity analysis component for direct processing.

[0045] S104, the data processing module 100 performs linear interpolation retrieval in the prior mapping matrix based on the real-time extracted instantaneous flow rate change rate. The data processing module 100 uses the linear interpolation algorithm to avoid the instantaneous flow rate change rate matching dead zone caused by discrete sampling points. The data processing module 100 extracts the physical lag time corresponding to the current transient condition and converts the physical lag time into a queue depth control parameter. The queue depth control parameter is obtained by dividing the physical lag time by the sampling time interval and then rounding it down to dynamically determine the length of the delay queue. The data processing module 100 establishes a delay queue in memory that follows a first-in-first-out logic. The data processing module 100 controls the delay queue to continuously store the inlet moisture content, return air temperature, water addition ratio, and steam addition ratio in the order of timestamps. The data processing module 100 adjusts the residence time of the data nodes inside the delay queue and the length of the delay queue according to the queue depth control parameter. The data processing module 100 extracts the historical input features popped from the head of the delay queue and combines them with the actual outlet moisture content collected at the current moment. The data processing module 100 outputs an aligned dataset containing the aligned feature combination and the corresponding actual outlet moisture content. The data processing module 100 sends the aligned dataset to subsequent prediction modeling nodes to avoid the feature value misalignment caused by physical pure lag.

[0046] See attached document Figure 4 The specific implementation of step S200 is carried out through the following sub-steps: S201, the prediction modeling module 200 receives the aligned dataset output by the data processing module 100. The aligned dataset contains the input feature parameter sequence and the corresponding actual outlet moisture content sequence in the historical time dimension. The prediction modeling module 200 calculates the mean and standard deviation of the distribution of the input feature parameter sequence and the actual outlet moisture content sequence.

[0047] The prediction modeling module 200 uses the Laida rule to remove outlier values ​​in the actual outlet moisture content sequence whose absolute deviation is greater than three times the standard deviation of the distribution, and simultaneously removes the input feature parameter samples at the timestamps corresponding to the outlier values. The prediction modeling module 200 uses the max-min normalization algorithm to linearly map the extreme value range of the input feature parameter sequence to a closed interval between 0 and 1, and simultaneously stores the maximum and minimum values ​​corresponding to each input feature parameter as the parameter basis for subsequent inverse normalization transformation. The prediction modeling module 200 combines the mapped input feature parameter sequence with the actual outlet moisture content sequence after removing outlier values ​​to synthesize a standard training dataset.

[0048] The prediction modeling module 200 randomly splits the standard training dataset into a training set and a test set at a ratio of 0.7 to 0.3. The prediction modeling module 200 extracts the inlet moisture content, return air temperature, water addition ratio, and steam addition ratio features contained in the standard training dataset to construct the input independent variable matrix. The prediction modeling module 200 extracts the actual outlet moisture content contained in the standard training dataset to construct the target dependent variable matrix. The prediction modeling module 200 declares and initializes the basic operational architecture of the multiple linear regression equation in the memory space and connects the input independent variable matrix and the target dependent variable matrix.

[0049] S202, the prediction modeling module 200 extracts the input independent variable matrix from the training set and constructs the transpose matrix of the input independent variable matrix. The prediction modeling module 200 performs matrix multiplication between the transpose matrix and the original input independent variable matrix to generate the covariance structure matrix. The prediction modeling module 200 performs inverse matrix operation on the covariance structure matrix to generate the inverse feature matrix.

[0050] When the covariance structure matrix does not satisfy the invertibility condition, the prediction modeling module 200 calls the pseudo-inverse matrix algorithm or the ridge regression algorithm with regularization term to generate the inverse feature matrix. The prediction modeling module 200 performs multiplication operations on the inverse feature matrix, the transpose matrix, and the target dependent variable matrix in sequence to generate the initial weight parameter vector distribution. The prediction modeling module 200 calls the ordinary least squares algorithm engine to perform multidimensional space fitting iterative operations on the input independent variable matrix and the target dependent variable matrix in the training set.

[0051] The ordinary least squares algorithm engine uses minimizing the sum of squared errors between the predicted and actual target values ​​as the target loss function for iterative calculation. The prediction modeling module 200 calculates the sensitivity weights of each feature and the model intercept of the multiple linear regression equation through fitting operations. The prediction modeling module 200 inputs the test set into the multiple linear regression equation containing the sensitivity weights of each feature and the model intercept and performs cross-validation operations.

[0052] The prediction modeling module 200 extracts model parameter combinations with a mean squared error of less than 0.02 and a mean absolute error of less than 0.1 and solidifies them to generate a positive prediction model. When the test set validation results do not meet the conditions of a mean squared error of less than 0.02 and a mean absolute error of less than 0.1, the prediction modeling module 200 expands the training sample window and re-executes the fitting operation, or keeps the effective positive prediction model solidified in the previous cycle to continue to participate in the control operation, and marks the current model parameter combination as an inactive state.

[0053] The prediction modeling module 200 constructs and stores a positive prediction model based on the following mathematical formula: ; In the formula, The predicted outlet moisture content is calculated for the model output; The first feature sensitivity weight corresponding to the inlet moisture content; Input parameters for inlet moisture content; The second feature sensitivity weight corresponding to the return air temperature; Input parameters for return air temperature; The sensitivity weight of the third feature corresponding to the water addition ratio feature; Input parameters for the water addition ratio feature; The sensitivity weight of the fourth feature corresponding to the gas refueling ratio feature; Input parameters for the steam refueling ratio characteristic; The intercept constant term of the model generated for the fitting operation.

[0054] S203, the forward prediction model establishes the linear physical correlation between production environment parameters and the amount of water vapor added and the outlet moisture content. The decoupling control module 300 receives the forward prediction model and the sensitivity weights and intercept constants of various features. The decoupling control module 300 reads the target outlet moisture content value set by the system. The decoupling control module 300 replaces the predicted outlet moisture content variable on the left side of the forward prediction model equation with the target outlet moisture content value. The decoupling control module 300 performs an algebraic shifting operation on the linear equation of the forward prediction model. The decoupling control module 300 moves the product term containing the inlet moisture content and return air temperature to the target outlet moisture content side. The decoupling control module 300 extracts the product term containing the water addition ratio feature and the steam addition ratio feature to construct the total water vapor demand term.

[0055] Before performing the reverse calculation, the decoupling control module 300 calls the extreme values ​​of inlet moisture content and return air temperature stored in the prediction modeling module 200, performs a forward normalization transformation on the inlet moisture content input parameters and return air temperature input parameters collected in real time at the current moment, and maps the inlet moisture content input parameters and return air temperature input parameters to a closed interval of 0 to 1.

[0056] The decoupling control module 300 uses known constants to perform reverse calculations based on the following mathematical formula to obtain the total water vapor demand: ; ; In the formula, This is the total water vapor demand value obtained from the reverse calculation; The target outlet moisture content value set externally; The first feature sensitivity weight corresponding to the inlet moisture content; Input parameters for the inlet moisture content after positive normalization transformation; The second feature sensitivity weight corresponding to the return air temperature; Input parameters for the return air temperature after positive normalization transformation; The model intercept constant term generated for the fitting operation; The sensitivity weight of the third feature corresponding to the water addition ratio feature; Input parameters for the water addition ratio characteristics in the normalized space; The sensitivity weight of the fourth feature corresponding to the gas refueling ratio feature; Input parameters for the steam addition ratio characteristics in the normalized space.

[0057] See attached document Figure 5 The specific implementation of step S300 is carried out through the following sub-steps: S301, the decoupling control module 300 receives the actual outlet moisture content collected by the data processing module 100 and the externally set target outlet moisture content, performs the absolute value calculation of the difference to extract the absolute deviation value, and determines the control direction according to the relationship between the actual outlet moisture content and the target outlet moisture content; when the total water and steam demand obtained by reverse calculation is less than or equal to 0, or the target outlet moisture content is lower than the achievable lower limit predicted by the model under the current operating condition, the decoupling control module 300 limits the water addition ratio and steam addition ratio to the minimum control ratio allowed by the equipment, generates a target unachievable state flag, and jumps out of the ratio allocation calculation of the current control cycle, directly outputs the minimum control ratio instruction; When the total water and steam demand is greater than 0, the decoupling control module 300 retrieves the system index tolerance value issued by the host computer and multiplies the system index tolerance value by a preset proportional coefficient to calculate and generate a judgment threshold. Preferably, the preset proportional coefficient is configured to be 0.4. Then, the absolute deviation value is compared with the judgment threshold and the steam distribution coefficient is dynamically configured according to the comparison status.

[0058] When the absolute deviation value is greater than the judgment threshold, a first steam injection allocation coefficient is configured; when the absolute deviation value is less than or equal to the judgment threshold, a second steam injection allocation coefficient is configured. The value range of both the first and second steam injection allocation coefficients is limited to real numbers in a closed interval from 0 to 1. The value of the first steam injection allocation coefficient is forcibly set to be greater than the value of the second steam injection allocation coefficient. This is to increase the proportion of steam injection adjustment when the outlet moisture deviation is large, and to shorten the adjustment time by taking advantage of the fast steam injection response. Conversely, when the outlet moisture deviation is small, the proportion of steam injection adjustment is reduced to reduce steady-state overshoot.

[0059] The decoupling control module 300 calculates the initial steam addition ratio and initial water addition ratio by combining the total water and steam demand item with the configured steam addition distribution coefficient and feature sensitivity weight. When the third feature sensitivity weight or the fourth feature sensitivity weight is less than the preset minimum effective weight threshold, the preset minimum effective weight threshold is configured to be between 0.01 and 0.05. The reverse division calculation based on the corresponding weight is stopped, and the effective control ratio of the previous cycle or the preset safety control ratio is called to participate in the current control cycle calculation. The preset safety control ratio is configured to be between 0.2 and 0.3 times the maximum control ratio allowed by the equipment.

[0060] The decoupling control module 300 calculates the initial steam addition ratio and the initial water addition ratio according to the following mathematical formulas: ; ; In the formula, To calculate the generated normalized initial steam addition ratio; To calculate the generated normalized initial water addition ratio; This is the total water vapor demand value obtained from the reverse calculation; The steam distribution coefficient is dynamically configured based on the absolute deviation value. The sensitivity weight of the fourth feature corresponding to the gas refueling ratio feature; The third feature sensitivity weight is the water addition ratio feature.

[0061] After obtaining the normalized initial steam addition ratio and the normalized initial water addition ratio, the decoupling control module 300 calls the maximum value, minimum value, maximum value and minimum value of the water addition ratio characteristic stored in the prediction modeling module 200, performs an inverse normalization transformation on the solution results, and finally obtains the initial water addition ratio and the initial steam addition ratio in physical dimensions.

[0062] S302, the decoupling control module 300 retrieves the hardware physical limit opening value associated with the regulating valve of the underlying steam injection pipeline. The decoupling control module 300 calibrates the hardware physical limit opening value to 0.95 times the rated full-open stroke of the regulating valve. The decoupling control module 300 maps the hardware physical limit opening value into the steam injection ratio limit boundary value according to the pre-calibrated flow characteristic curve. The decoupling control module 300 compares the initial steam injection ratio with the steam injection ratio limit boundary value to determine the out-of-bounds state.

[0063] When the initial steam addition ratio exceeds the limit boundary value of the steam addition ratio, the decoupling control module 300 determines that there is a thermodynamic over-limit requirement. Before executing the cutoff, the decoupling control module 300 temporarily stores the initial steam addition ratio as the original initial steam addition ratio. The decoupling control module 300 calculates and extracts the difference in over-limit steam addition requirement by subtracting the limit boundary value of the steam addition ratio from the original initial steam addition ratio, and limits the final steam addition ratio to the limit boundary value of the steam addition ratio. For the specific conversion calculation of the flow characteristic curve of the regulating valve, the existing fluid dynamics lookup table algorithm can be used.

[0064] When the initial steam addition ratio is less than or equal to the limit boundary value of the steam addition ratio, the decoupling control module 300 determines that there is no thermodynamic over-limit requirement, determines the initial steam addition ratio as the final steam addition ratio command, determines the initial water addition ratio as the final water addition ratio command, and directly enters the conversion process between the absolute water addition flow command and the absolute steam addition flow command.

[0065] S303, the decoupling control module 300 retrieves the pre-stored thermodynamic evaporation compensation coefficient. The thermodynamic evaporation compensation coefficient comprehensively represents the equivalent conversion relationship between the moisture introduced by steam, the material moisture evaporation loss caused by steam heat release, and water addition compensation. It is used to convert the unmet equivalent steam addition demand into water addition compensation when the steam addition ratio is constrained by the hardware physical limit. The decoupling control module 300 uses the ratio of the saturated steam enthalpy value under the current pipeline working pressure to the specific heat capacity of the input material to determine the specific value of the thermodynamic evaporation compensation coefficient. The decoupling control module 300 multiplies the difference in the excessive steam addition demand by the thermodynamic evaporation compensation coefficient to convert it into the value of the moisture compensation item.

[0066] The decoupling control module 300 adds the moisture compensation value to the initial water addition ratio to calculate the final water addition ratio command. The decoupling control module 300 uses the truncated steam addition ratio limit boundary value as the final steam addition ratio command. The decoupling control module 300 multiplies the final water addition ratio command and the final steam addition ratio command by the real-time collected material flow rate, respectively, and outputs the absolute water addition flow rate command and the absolute steam addition flow rate command associated with the underlying hardware node. Before output, the decoupling control module 300 performs upper and lower limit constraints on the final water addition ratio command and the final steam addition ratio command. When either ratio command is lower than the minimum control ratio allowed by the corresponding equipment, the corresponding ratio command is limited to the corresponding minimum control ratio; when either ratio command is higher than the maximum control ratio allowed by the corresponding equipment, the corresponding ratio command is limited to the corresponding maximum control ratio.

[0067] The decoupling control module 300 performs out-of-bounds compensation calculations based on the following mathematical formula: ; ; In the formula, This is to calculate the final water addition ratio instruction generated; To calculate the initial water addition ratio; To calculate the extracted difference in gas refueling demand beyond the permitted limits; The pre-stored thermodynamic evaporation compensation coefficient; To calculate the final steam ratio instruction generated; The limit boundary value of the gas filling ratio is the result of the hardware physical limit mapping transformation.

[0068] The specific implementation of step S400 is carried out through the following sub-steps: S401, the feedback evolution module 400 receives the actual outlet moisture content sequence output by the data processing module 100 and the target outlet moisture content set externally. The feedback evolution module 400 retrieves the instantaneous flow rate change rate recorded in the status monitoring register. The feedback evolution module 400 extracts the system's preset steady-state flow fluctuation threshold and limits the steady-state flow fluctuation threshold to 0.05 times the nominal capacity.

[0069] The feedback evolution module 400 performs a numerical comparison between the instantaneous flow rate change rate and the steady-state flow fluctuation threshold. When the instantaneous flow rate change rate is less than or equal to the steady-state flow fluctuation threshold, the feedback evolution module 400 determines that the system is in the steady-state mass transfer range. Within the steady-state mass transfer range, the feedback evolution module 400 calculates the time series of the deviation between the actual outlet moisture content and the target outlet moisture content and extracts the standard deviation value of the deviation time series. The feedback evolution module 400 configures the time window length of the specified historical time domain to an integer within a closed interval of 300s to 600s.

[0070] S402, the feedback evolution module 400 calculates the absolute deviation between the actual outlet moisture content and the target outlet moisture content. The feedback evolution module 400 retrieves the system index tolerance value issued by the host computer. The feedback evolution module 400 multiplies the system index tolerance value by 0.4 to generate a dynamic judgment threshold. Based on the comparison result between the absolute deviation value and the dynamic judgment threshold, the feedback evolution module 400 configures a piecewise backpropagation error function. The piecewise backpropagation error function uses the square of the deviation between the actual outlet moisture content and the target outlet moisture content as the base term, and introduces the standard deviation value of the specified historical time domain and the dynamic judgment threshold as piecewise scaling factor coefficients to construct the target loss function. The feedback evolution module 400 performs a derivative operation on the backpropagation error function to extract the gradient data sequence of the corresponding interval, and uses the extracted gradient data sequence as the derivative value of the target loss function.

[0071] The feedback evolution module 400 calculates the error gradient sequence according to the following mathematical formulas when the absolute deviation value is greater than the dynamic judgment threshold and when the absolute deviation value is less than or equal to the dynamic judgment threshold. When the standard deviation value in the specified historical time domain is less than the preset minimum standard deviation threshold, the preset minimum standard deviation threshold is configured to range from 0.01 to 0.05. The feedback evolution module 400 replaces the standard deviation value with the preset minimum standard deviation threshold in the error gradient calculation to avoid abnormal gradient amplification caused by excessively small standard deviation. ; ; In the formula, The local penalty gradient is calculated when the absolute deviation value is greater than the dynamic judgment threshold. The fine-tuned convergence gradient is calculated when the absolute deviation value is less than or equal to the dynamic judgment threshold. This represents the actual outlet moisture content collected at the current moment. The target outlet moisture content value set externally; To calculate the standard deviation value extracted within a specified historical time domain; The dynamic judgment threshold for extracting the allowable tolerance values ​​of system indicators.

[0072] S403, the feedback evolution module 400 configures the base learning rate parameter for parameter updates. The feedback evolution module 400 limits the value range of the base learning rate parameter to a continuous real number within the closed interval from 0.001 to 0.01. The feedback evolution module 400 uses the calculated output local penalty gradient or fine-tuning convergence gradient as the derivative of the target loss function, multiplies it by the base learning rate parameter to generate the intercept correction compensation amount, and retrieves the currently active multiple linear regression equation in the prediction modeling module 200.

[0073] The feedback evolution module 400 feeds back the intercept correction compensation to the original model intercept constant term of the multiple linear regression equation to perform parameter overwriting. The feedback evolution module 400 uses the iteratively updated model intercept constant term to participate in the reverse calculation of the total water vapor demand term in the next control cycle. Through the rolling update of the model intercept constant term, the static control residual error caused by the cumulative drift of ambient temperature and humidity can be avoided.

[0074] The feedback evolution module 400 performs adaptive updates of the model intercept based on the following mathematical formula: ; In the formula, This refers to the intercept constant term of the newly generated model after the update. This is the intercept constant term of the original model stored internally in the multiple linear regression equation; The pre-configured base learning rate parameter; The local penalty gradient or fine-tuning convergence gradient is matched based on the absolute deviation numerical state input.

[0075] S404, the feedback evolution module 400 receives the environmental variable time series output by the data processing module 100. The environmental variable time series includes at least one of the inlet moisture content time series, return air temperature time series, and material flow rate time series. The feedback evolution module 400 calculates the coefficient of variation of each environmental variable time series. When the coefficient of variation is greater than 0.15, the feedback evolution module 400 triggers the full-parameter evolution mechanism.

[0076] The feedback evolution module 400 extracts the input independent variable matrix within the long time-domain window corresponding to the trigger full-parameter evolution mechanism, uses the sampled input feature parameters after positive normalization within the long time-domain window as partial derivative multiplier factors, and performs average or weighted average processing on the weight error correction compensation amount calculated at each sampling time to generate the weight error correction compensation amount corresponding to each feature channel.

[0077] The feedback evolution module 400 feeds back the weight error correction compensation amount of each feature sensitivity weight to the original feature sensitivity weight corresponding to the multiple linear regression equation to perform a dynamic overwrite operation, and uses the updated feature sensitivity weight to perform prediction verification on the aligned dataset in the most recent verification window; when the prediction error does not exceed the preset error limit, the value range of the preset error limit is configured to be 0.05 to 0.1, and the updated feature sensitivity weight is marked as the activation parameter; when the prediction error exceeds the preset error limit, it reverts to the feature sensitivity weight before the update.

[0078] The feedback evolution module 400 performs adaptive updates of the sensitivity weights of the internal features of the multiple linear regression equation according to the following mathematical formula: ; In the formula, The new feature sensitivity weights are generated for the corresponding independent variables after the update; These are the original feature sensitivity weights stored within the multiple linear regression equation; The pre-configured base learning rate parameter; To match the local penalty gradient or fine-tune the convergence gradient based on the absolute deviation numerical state input; The input parameters are acquired in real time for the corresponding feature channels and then processed by positive normalization. Index of a specific feature in the set of independent variables.

[0079] To aid in understanding the technical solution of this invention, the following provides a specific application example and test example of predicting the outlet moisture content and controlling the water addition ratio in a loosening and rehydration process.

[0080] Application Examples: The loose rehydration process equipment of a cigarette factory with a nominal capacity of 5000 kg / h was used as the execution object. The target outlet moisture content of the system was set to 18.50%, the system index tolerance was set to 0.50%, and the sampling frequency was set to 2 Hz.

[0081] The data processing module 100 acquires real-time production data. At a certain moment, the material flow rate is 4812.3 kg / h, the inlet moisture content is 11.24%, the return air temperature is 65.1℃, and the feedback water addition from the previous control cycle is 144.5 kg / h and steam addition is 96.2 kg / h. The data processing module 100 calculates the water addition ratio feature as 0.0300 and the steam addition ratio feature as 0.0199. The data processing module 100 extracts the current instantaneous flow rate change rate, matches it with the prior mapping matrix to obtain a physical lag time of 11.5s, and sets the delay queue length to 23. It extracts the input feature parameters from 23 sampling cycles ago and aligns them with the current actual outlet moisture content for the prediction modeling module 200 to fit.

[0082] The predictive modeling module 200 updates the multiple linear regression equation by combining the aligned dataset. The current model intercept constant term is 1.251. The decoupling control module 300 substitutes the target outlet moisture content of 18.50% into the forward prediction model for reverse calculation. After deducting the contribution terms of inlet moisture content and return air temperature, the total water vapor demand term is obtained. The current measured actual outlet moisture content is 17.62%. The absolute deviation between the actual outlet moisture content and the target outlet moisture content is 0.88%. The system index tolerance is multiplied by 0.4 to calculate the dynamic judgment threshold of 0.20%. Since the absolute deviation of 0.88% is greater than 0.20%, the decoupling control module 300 configures the first steam distribution coefficient of 0.75 and calculates the initial steam distribution ratio of 0.0284.

[0083] The decoupling control module 300 retrieves the steam addition ratio limit boundary value corresponding to the physical limit of the bottom valve, which is 0.0250. The initial steam addition ratio of 0.0284 is greater than the steam addition ratio limit boundary value of 0.0250, indicating an out-of-bounds demand. The difference in out-of-bounds steam addition demand is 0.0034. The decoupling control module 300 retrieves the thermodynamic evaporation compensation coefficient of 1.15 and converts the difference in out-of-bounds steam addition demand into a moisture compensation term of 0.0039, which is added to the initial water addition ratio. Finally, the steam addition ratio command is issued with a limit of 0.0250. The water addition ratio command is output after adding the moisture compensation term. The feedback evolution module 400 uses the absolute deviation of 0.88% to trigger the local penalty gradient calculation of the piecewise loss function, and updates the model intercept constant term in combination with the standard deviation sequence within the time window, which participates in the next cycle calculation.

[0084] Test example: To verify the control effect of the present invention, a comparative test was performed on the same loose rehydration production line. The control group used a single-dimensional proportional integral derivative control algorithm for adjusting the water valve and steam valve, while the experimental group used the outlet moisture prediction and water addition ratio control system of the present invention. The test was set to artificially introduce front-end process fluctuations at the 100th second, causing the inlet moisture content to drop in a step, and the test operation data was recorded.

[0085] The test results are shown in the table below.

[0086] Time-series sampling test data record table under sudden inlet moisture interference: According to the table above and appendix Figure 6 According to the data, within the steady-state range from 0s to 80s, the actual outlet moisture content of both the control group and the experimental group fluctuated around the target outlet moisture content of 18.50%, confirming that the system was in normal production. At 100s, the inlet moisture content suddenly dropped from 11.52% to 9.21% and remained at a low level. Considering the physical mass transfer time within the system, the actual outlet moisture content began to decrease around 115s.

[0087] When faced with sudden disturbances, the actual outlet moisture content of the control group dropped to 17.11% at 125s, resulting in a maximum dynamic deviation of 1.39%. Subsequently, due to the cumulative effect of integration, overcompensation occurred, resulting in a positive overshoot of 18.76% at 180s. After 300s, it tended to be around 18.32%, resulting in a steady-state absolute error of 0.18%. The actual outlet moisture content of the experimental group dropped to 17.92% at 120s, with the maximum dynamic deviation limited to 0.58%. It then recovered to 18.45% at 150s and stabilized between 18.48% and 18.50% after 300s.

[0088] Analysis of the internal output commands of the experimental group revealed that, between 120s and 125s, the absolute steam addition command reached 120.3 kg / h, coinciding with the steam addition ratio limit value set by the physical limit of the regulating valve. The decoupled control module 300 triggered an out-of-bounds demand judgment and multiplied the excess steam addition equivalent water demand by the thermodynamic evaporation compensation coefficient, which was then added to the absolute water addition command. The absolute water addition command climbed to 182.7 kg / h at 125s to perform a water replenishment operation. This invention, by limiting the upper limit of the steam addition command output and performing compensation calculations in the water circuit, prevents the steam addition pipeline valve from entering the fully open dead zone and losing its regulating ability.

[0089] The experimental group did not exhibit significant positive overshoot after encountering a sudden environmental change, and there was no residual error in the steady-state phases of 300s and 400s. The feedback evolution module 400 constructed a piecewise loss function using the absolute deviation between the actual outlet moisture content and the target outlet moisture content, calculated the local penalty gradient based on the short-time domain deviation state, and corrected the model intercept constant term of the multiple linear regression equation. The update operation of the intercept term offset the system offset caused by the low operating of the front-end inlet moisture content. The fixed parameter mechanism of the control group could not adapt to changes in environmental conditions, resulting in a persistent static control error. The actual output results of the experimental group confirmed that the pure time delay causal alignment implemented by the data processing module 100 and the dynamic update logic of the model intercept executed by the feedback evolution module 400 improved the control accuracy of the outlet moisture content under unsteady transition conditions.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the outlet moisture content and controlling the water addition ratio in a loosening and rehydration process, characterized in that, The method includes the following steps: Acquire real-time production data including inlet moisture content, return air temperature, material flow rate, water addition, steam addition, and actual outlet moisture content; Calculate the water addition ratio characteristics and steam addition ratio characteristics; extract the instantaneous flow rate change rate, match the physical lag time, dynamically determine the delay queue length, and complete the causal alignment between the input feature parameters and the actual outlet moisture content; Using the input feature parameters as input independent variables and the actual outlet moisture content as the target dependent variable, a multiple linear regression equation is constructed to generate a positive prediction model and a model intercept; the target outlet moisture content is substituted into the positive prediction model to perform inverse calculation to obtain the total water vapor demand item; Extract the absolute deviation value, dynamically configure the steam addition distribution coefficient based on the comparison status, and calculate the initial steam addition ratio and the initial water addition ratio; when it is determined that there is a thermodynamic over-limit requirement, add the water compensation term value to the initial water addition ratio to calculate and generate the final water addition ratio instruction, and limit the final steam addition ratio to the steam addition ratio limit boundary value. Construct a target loss function, calculate the standard deviation of the bias time series, perform derivative operation on the backpropagation error function to extract the gradient data sequence of the corresponding interval, multiply by the base learning rate parameter to generate the intercept correction compensation amount, and feed back and superimpose the intercept correction compensation amount to update the model intercept.

2. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The steps for calculating the water addition ratio characteristics and steam addition ratio characteristics include: The collected water and steam addition amounts are divided by the material flow rate at the same moment to generate water addition ratio characteristics and steam addition ratio characteristics, respectively. When the material flow rate at any given moment is less than the preset minimum effective flow rate threshold, the material flow rate at the previous effective sampling moment is used to participate in the proportional feature calculation, or the current sampling point is marked as an invalid sample and excluded from the subsequent modeling dataset.

3. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The steps of extracting the instantaneous flow rate change rate, matching the physical lag time, dynamically determining the delay queue length, and completing the causal alignment between the input feature parameters and the actual outlet moisture content include: The time step offset when the historical input environmental parameter sequence and the historical actual outlet moisture content sequence reach the maximum morphological similarity is calculated by using the dynamic time warping algorithm, and the physical lag time is calculated by multiplying it by the sampling time interval. A priori mapping matrix is ​​then constructed. Based on the extracted instantaneous flow rate change, linear interpolation is performed in the prior mapping matrix to extract the physical lag time corresponding to the current transient condition. The queue depth control parameter is obtained by dividing the physical lag time by the sampling time interval and rounding it down, so as to dynamically determine the length of the delay queue. Based on the queue depth control parameters, the dwell time of data nodes inside the delay queue and the length of the delay queue are adjusted, and the historical input features popped from the head of the delay queue are extracted and combined with the actual outlet moisture content collected at the current moment.

4. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The steps for constructing a multiple linear regression equation using the input feature parameters as input independent variables and the actual outlet moisture content as the target dependent variable, and generating a positive prediction model and model intercept, include: The Laida method is used to remove outlier values ​​in the actual outlet moisture content sequence whose absolute deviation is greater than three times the standard deviation of the distribution, and the input feature parameter samples under the timestamps corresponding to the outlier values ​​are removed simultaneously. The maximum-minimum normalization algorithm is used to linearly map the extreme value range of the input feature parameter sequence to a closed interval between 0 and 1, and the maximum and minimum values ​​corresponding to each input feature parameter are stored simultaneously. The ordinary least squares algorithm engine is invoked to calculate the sensitivity weights of each feature and the intercept of the multiple linear regression equation through fitting operations.

5. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The step of substituting the target outlet moisture content into the forward prediction model to perform inverse calculation and obtain the total water vapor demand includes: Perform a positive normalization transformation on the inlet moisture content input parameter and the return air temperature input parameter collected in real time at the current moment; Replace the predicted outlet moisture content variable on the left side of the positive prediction model equation with the externally set target outlet moisture content value; An algebraic shifting operation is performed on the linear equation of the positive prediction model, shifting the product term containing the inlet moisture content and return air temperature to the target outlet moisture content side, and extracting the product term containing the water addition ratio characteristics and steam addition ratio characteristics to construct the total water and steam demand term.

6. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The step of dynamically configuring the steam distribution coefficient based on the comparison status and calculating the initial steam ratio and the initial water ratio includes: The judgment threshold is calculated by multiplying the system index tolerance value by a preset proportional coefficient. When the absolute deviation value is greater than the judgment threshold, a first fuel allocation coefficient is configured; when the absolute deviation value is less than or equal to the judgment threshold, a second fuel allocation coefficient is configured, and the value of the first fuel allocation coefficient is forcibly set to be greater than the value of the second fuel allocation coefficient. The initial steam addition ratio and the initial water addition ratio are calculated by combining the total water and steam demand, the configured steam addition allocation coefficient, and the characteristic sensitivity weight, and the solution results are then subjected to inverse normalization transformation.

7. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The step of adding the moisture compensation term to the initial water addition ratio to calculate the final water addition ratio instruction when a thermodynamic overshoot requirement is determined, and limiting the final steam addition ratio to the steam addition ratio limit boundary value, includes: Retrieve the hardware physical limit opening value associated with the regulating valve of the underlying steam injection pipeline, and map the hardware physical limit opening value into the steam injection ratio limit boundary value; The initial steam addition ratio is compared with the steam addition ratio limit boundary value to determine the boundary state. When the initial steam addition ratio is greater than the steam addition ratio limit boundary value, it is determined that there is a thermodynamic boundary requirement. Before the cutoff is executed, the initial steam refueling ratio is temporarily stored as the original initial steam refueling ratio. The difference in steam refueling demand beyond the limit is calculated and extracted by subtracting the steam refueling ratio limit boundary value from the original initial steam refueling ratio. The difference in steam demand beyond the designated area is multiplied by the thermodynamic evaporation compensation coefficient and converted into a moisture compensation term value. The moisture compensation value is added to the initial water addition ratio to calculate the final water addition ratio instruction, and the final steam addition ratio is limited to the steam addition ratio limit boundary value.

8. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, The steps of constructing the target loss function, calculating the standard deviation of the bias time series, extracting the gradient data sequence of the corresponding interval by performing derivative operation on the backpropagation error function, multiplying by the base learning rate parameter to generate the intercept correction compensation amount, and feeding back the intercept correction compensation amount to update the model intercept include: When the instantaneous flow rate change rate is less than or equal to the steady-state flow rate fluctuation threshold, the system is determined to be in the steady-state mass transfer range. The deviation time series between the actual outlet moisture content and the target outlet moisture content is calculated, and the standard deviation value of the deviation time series is extracted. Based on the comparison results between the absolute deviation value and the judgment threshold, a piecewise backpropagation error function is configured, and the standard deviation value of the specified historical time domain and the dynamic judgment threshold are introduced as piecewise scaling factor coefficients to construct the target loss function. The gradient data sequence of the corresponding interval is extracted by performing a derivative operation on the backpropagation error function, multiplied by the basic learning rate parameter to generate the intercept correction compensation amount, and then fed back into the intercept constant term of the original model of the multiple linear regression equation to perform parameter overwriting operation.

9. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 8, characterized in that, The step of introducing the standard deviation of a specified historical time domain and a dynamic decision threshold as piecewise scaling factor coefficients to construct the target loss function includes: When the standard deviation value within the specified historical time domain is less than the preset minimum standard deviation threshold, the standard deviation value is replaced with the preset minimum standard deviation threshold for error gradient calculation.

10. The method for predicting outlet moisture content and controlling water addition ratio in the loosening and rehydration process according to claim 1, characterized in that, While performing the step of feeding back and superimposing the intercept correction compensation to update the model intercept, a long-term domain parameter update operation is also triggered, the steps of which include: Calculate the coefficient of variation of a series of environmental variables, including inlet moisture content, over a long time period; When the coefficient of variation exceeds the set limit, the corresponding input feature parameter sequence is extracted, and each sampled input feature parameter after positive normalization within the long time window is used as a partial derivative multiplier factor to generate the weight error correction compensation amount corresponding to each feature channel. The error correction compensation for each feature sensitivity weight is fed back and superimposed onto the original feature sensitivity weight corresponding to the multiple linear regression equation to perform a dynamic overwrite operation. When the prediction error exceeds the preset error limit, it reverts to the feature sensitivity weight before the update.

Images