Multivariable intelligent monitoring system for pill fluidized bed drying process

Abstract

The invention relates to the technical field of traditional Chinese medicine preparation production, in particular to a multivariable intelligent monitoring system for a pill fluidized bed drying process, which comprises a multi-parameter acquisition module, a multivariable data analysis processing module and an abnormity early warning and intelligent control module. Wherein the multi-parameter acquisition module acquires multi-dimensional heterogeneous monitoring data formed by process and material attributes in real time; the multivariable data analysis processing module identifies a parameter coupling relationship through multivariable statistical analysis based on the data to obtain an analysis result; and the abnormity early warning and intelligent control module establishes an abnormity discrimination model according to the coupling abnormity, early warning is carried out when coupling abnormity is detected, and a control strategy is automatically adjusted according to severity. According to the system, multi-parameter collaborative abnormal millisecond-level response can be realized, the steady-state maintenance rate is improved, and the problems of monitoring blind areas and control lag of a traditional method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traditional Chinese medicine preparation, in particular to a multivariate intelligent monitoring system for the fluidized bed drying process of pills. BACKGROUND

[0002] In the fluidized bed drying production process of pills, the material is dried by efficient heat mass exchange with hot air in a fluidized state. The core process parameters include the coordinated action of multiple variables such as inlet air temperature, air volume, and bed pressure drop, which jointly affect the drying efficiency and product quality.

[0003] However, the traditional monitoring technology has significant limitations in this process:

[0004] First, the single parameter threshold-based judgment mode cannot effectively identify multivariate coupling abnormalities. For example, when the inlet air temperature is within the standard range, but the air volume and bed pressure drop deviate cooperatively, the system often cannot achieve real-time early warning.

[0005] Second, the control strategy based on human experience cannot adapt to the dynamic fluctuations of material properties, resulting in high product quality dispersion.

[0006] Third, the existing monitoring system generally lacks multi-parameter correlation analysis capability, so that the deep value of process data cannot be fully tapped. It is worth noting that Hotelling's T2 statistical method has shown unique advantages in multivariate process monitoring, but its systematic application in the steady-state discrimination and intelligent control of pill drying processes has not been realized. SUMMARY

[0007] Therefore, the technical problem to be solved by the present application is to overcome the problems of traditional monitoring technology in the fluidized bed drying production of pills, such as single parameter threshold judgment, reliance on human experience control, and lack of multi-parameter correlation analysis capability, which makes it difficult to cope with multivariate coupling abnormality identification, adapt to dynamic fluctuations of material properties, and tap deep features of process data.

[0008] To solve the above technical problems, the present application provides a multivariate intelligent monitoring system for the fluidized bed drying process of pills, which takes Hotelling's T2 statistical theory as the core, integrates artificial intelligence algorithms to build the steady-state online monitoring and intelligent control strategy of the fluidized bed drying process of pills, and aims to realize multivariate collaborative abnormality detection and precise regulation of the process. Specifically, the system is composed of the following modules: a multi-parameter acquisition module, a multivariate data analysis processing module, and an abnormality early warning and intelligent control module. Among them,

[0009] The multi-parameter acquisition module is configured to acquire real-time multi-dimensional heterogeneous monitoring data composed of process parameters and material attribute parameters in the fluidized bed drying process.

[0010] The multivariate data analysis processing module is configured to perform correlation analysis based on the multidimensional heterogeneous monitoring data by a multivariate statistical analysis method, identify coupling relationships between parameters, and obtain a multivariate data analysis result.

[0011] The abnormality early warning and intelligent control module is configured to establish an abnormality discrimination model based on the multivariate data analysis result, timely issue an early warning when a multivariate coupling abnormality is detected, and automatically adjust a control strategy according to the severity of the abnormality to realize intelligent control of the pellet fluidized bed drying process.

[0012] In an embodiment of the present application, obtaining the multivariate data analysis result comprises:

[0013] Performing a denoising and standardization preprocessing operation on the multidimensional heterogeneous monitoring data to obtain preprocessed monitoring data;

[0014] Determining feature variables and target variables to be analyzed based on the preprocessed monitoring data, the target variables being steady-state state data of the drying process;

[0015] Calculating mutual information values between each feature variable and its corresponding target variable, denoted as first mutual information values, and screening out potential relevant features according to the first mutual information values to form a candidate feature pool;

[0016] Calculating mutual information values between different features in the candidate feature pool, denoted as second mutual information values;

[0017] Integrating the first mutual information values and the second mutual information values to form a mutual information matrix, the row index values and the column index values of the mutual information matrix being feature variables, and the matrix elements being corresponding mutual information values;

[0018] Screening out a set of key feature variables strongly related to the drying steady state from the candidate feature pool based on the mutual information matrix;

[0019] Calculating a Hotelling's T2 statistic according to the set of key feature variables, which is the multivariate data analysis result.

[0020] In an embodiment of the present application, screening out a set of key feature variables strongly related to the drying steady state from the candidate feature pool based on the mutual information matrix comprises:

[0021] Representing a feature subset in the candidate feature pool by using a binary coding mode, comprising: representing each feature in the candidate feature pool by using a binary bit, wherein “1” indicates that the feature is included in the current subset, and “0” indicates that the feature is not included;

[0022] Based on the mutual information values ​​in the mutual information matrix, a fitness function is designed that maximizes the correlation between the features and the dry steady-state target variable and minimizes the redundancy between features.

[0023] Performing a selection operation includes: calculating the fitness value of each feature subset based on the fitness function, selecting high-quality feature subsets according to the fitness value and a preset probability, and retaining the high-quality feature subsets to form a breeding population;

[0024] Performing crossover operations includes: for the feature subsets in the breeding population obtained by the selection operation, performing single-point crossover according to a preset crossover probability: randomly selecting a crossover position from the binary encoding strings of any two parent feature subsets, and then swapping all the binary bits after the crossover position in the two parent encoding strings to generate a new offspring feature subset;

[0025] Perform mutation operations, including: for a subset of offspring features, randomly flip any binary bit according to a preset mutation probability, and further introduce new feature combinations;

[0026] Repeatedly perform selection, crossover, and mutation operations, iterating until the convergence condition is met that the change in fitness values ​​over a preset number of generations is within a preset range. The subset of features with the highest fitness values ​​is the set of key feature variables X = [X1, X2, ..., X...]. n-1 ,X n ], where n is the dimension.

[0027] In one embodiment of the present invention, the fitness function is as follows:

[0028]

[0029] Where F is the fitness value, MI(X) i MI(X, Y) represents the mutual information value between the i-th feature variable and the dry steady-state target variable Y in the mutual information matrix; i ,Y j ) represents the mutual information value between the i-th and j-th feature variables in the mutual information matrix; m represents the number of features in the current feature subset; α and β are weighting coefficients used to adjust the influence weights of correlation and redundancy.

[0030] In one embodiment of the present invention, calculating the Hotelling's T2 statistic based on the set of key feature variables includes:

[0031] A sliding time window is used to sample historical data of a set of key feature variables within a specified time period before the current moment in real time.

[0032] Based on the historical data, a sample mean vector and a covariance matrix of the set of key characteristic variables are calculated; when a material batch switching or a process parameter change is detected, the current sliding window data is automatically emptied, and the sample mean vector and the covariance matrix are recalculated;

[0033] Based on the sample mean vector and the covariance matrix, a Hotelling's T2 statistic is calculated:

[0034]

[0035] Wherein, n is the sample size in the sliding time window, X is the current sampling data, is the sample mean vector, T is the transpose of the matrix, S -1 is the inverse matrix of the covariance matrix.

[0036] In an embodiment of the present application, based on the multivariate data analysis result, an abnormality discrimination model is established, and timely warning is given when a multivariate coupling abnormality is detected, including:

[0037] The historical normal production data of multiple batches are obtained, based on which a dynamic correction coefficient λ related to production time, environmental temperature and humidity is introduced, and a control threshold value capable of being self-adaptively adjusted according to environmental and process changes is calculated

[0038] Based on the real-time obtained multi-dimensional heterogeneous monitoring data, a Hotelling's T2 statistic value T 2 is calculated, that is, a multivariate data analysis result, and according to the comparison result of the control threshold value and the multivariate data analysis result, abnormality diagnosis is performed from a micro perspective and a macro perspective:

[0039] When , a micro perspective abnormality diagnosis process is performed, including:

[0040] The contribution degree C 2 of the i-th key characteristic variable to the abnormality is calculated based on the multi-dimensional heterogeneous monitoring data: i : Wherein, σi2 represents the variance of the i-th key characteristic variable in the sliding time window, i=1, 2, …, p, and p is the total number of key characteristic variables; X i represents the monitoring value of the i-th key characteristic variable at the current sampling time, represents the sample mean vector of the i-th key characteristic variable in the sliding time window.

[0041] According to the contribution degree, all key characteristic variables are sorted, and the top K characteristic variables in the contribution degree are selected and determined as the abnormality source;

[0042] Performing a macroscopic angle anomaly diagnosis process, including:

[0043] Inputting the latest n minutes of T2 statistical quantity sequence into the trained LSTM model to obtain the T2 prediction value in the future time period, if the T2 prediction value meets any of the following conditions, triggering an early warning:

[0044] Condition one, the prediction value exceeds the control threshold The preset proportion;

[0045] Condition two, the rising rate of the prediction value exceeds M times of the historical normal trend, M>1.

[0046] In an embodiment of the application, the control threshold The calculation method is as follows:

[0047]

[0048] Wherein, p is the number of variables in the key feature variable set selected from the multi-dimensional heterogeneous monitoring data, n is the sample size in the sliding time window, F p,n-p,α is the F distribution quantile with degrees of freedom p and n-p and significance level alpha.

[0049] In an embodiment of the application, according to the severity of the anomaly, the control strategy is automatically adjusted, including:

[0050] Based on the comparison result of Hotelling’s T2 statistical value T 2 And the control threshold Trigger a three-level gradient response according to the severity of the anomaly to realize closed-loop precise regulation of the drying process, the steps are as follows:

[0051] When Determine the early warning state, start the first level response, select the abnormal source based on the contribution of each key feature variable, adjust the weight parameter of the abnormal source, ensure that the control resources are preferentially used for the abnormal source, and use the fuzzy PID method to adjust the device operation parameter corresponding to the abnormal source until

[0052] When Determine the significant abnormal state, start the second level response, generate a global control scheme including temperature curve reconstruction and hot air circulation path switching based on historical control experience and real-time state, execute the global control scheme, and monitor T 2 Change in real time; if T 2 Fall back to Within, automatically switch to the first level response for fine tuning;

[0053] When When the abnormal parameter is determined to be in a serious fault state, a three-stage response is started, a safety shutdown program is immediately executed, and a diagnosis report is outputted for tracing the abnormal parameter and providing a process adjustment suggestion.

[0054] In an embodiment of the present application,

[0055] When Based on historical control experience and real-time state, a global control scheme including temperature curve reconstruction and hot air circulation path switching is generated, including:

[0056] A control model including a current Q network and a target Q network and corresponding state space S, action space A and policy space π(a|s) are constructed, wherein the current Q network is a neural network structure including an LSTM layer, a Dropout layer and a fully connected layer, the structure of the target Q network is the same as that of the current Q network, s i represents the state at the current time i, including the current T 2 value, real-time data of the key feature variable set X, initial moisture content of the material and environmental temperature and humidity; a i represents an executable global control action, including a temperature curve reconstruction strategy and a hot air circulation path switching strategy; the policy space π(a|s) represents a probability distribution of selecting an action a under a state s;

[0057] The parameters of the current Q network are initialized, the Q values Q(s,a) of all possible actions under the current state s are calculated based on the current Q network, and an action a is selected according to an ε-greedy exploration strategy;

[0058] The action a is executed in the environment, and a reward function is constructed based on the T 2 value, so that the immediate reward r is obtained and transferred to the next state s ′ ; the target Q value y of the selected action a is calculated by the target Q network in combination with the immediate reward r and the next state s ′

[0059] Based on the target Q value y and the Q value Q(s,a) predicted by the current Q network, a mean square error loss value is calculated, the parameters of the current Q network are updated by minimizing the loss value through back propagation, and the parameters of the current Q network are copied to the target Q network every preset iteration step;

[0060] The iteration process is repeated until the network converges or the maximum iteration round is reached, and a trained control model is obtained;

[0061] The real-time state features are input into the trained control model, and the action that maximizes the reward value, i.e., the global control scheme, is output.

[0062] ​In one embodiment of the present application, the reward function is as follows:

[0063] R = ω1 R1 + ω2 R2 - ω3 R3 + ω4 R4

[0064] Wherein, R1 represents T 2 Statistical quantity to control threshold Close speed, R1 = T of the current sampling time t 2 Statistical quantity, T of the last sampling time of T of the last sampling time of 2 Statistical quantity; R2 represents T 2 Stability in the statistical quantity falling process, T of the last N sampling times 2 Statistical quantity, σ (·) represents standard deviation; R3 represents the adjustment amplitude reflecting temperature and air volume parameters, R3 = ρ1 T ΔT represents the current air inlet temperature adjustment amount, ΔT max The maximum step length of air inlet temperature adjustment in the primary response, ΔQ is the current air volume adjustment amount, ΔQ max The maximum step length of air volume adjustment in the primary response, ρ1, ρ2 are weight coefficients, ρ1 + ρ2 = 1; R4 is when T 2 Statistical quantity falls to the control threshold The reward value given when R4 = ω1, ω2, ω3, ω4 are weight coefficients, satisfying ω1 + ω2 + ω3 + ω4 = 2.

[0065] The above technical solutions of the present application have the following advantages compared with the prior art:

[0066] The present application realizes accurate monitoring of multiple parameters in the fluidized bed drying process of pills by constructing a multi-dimensional heterogeneous data perception network; the dynamic adaptive model constructed based on Hotelling's T2, combined with feature engineering optimization and dynamic covariance matrix updating, can effectively screen core variables and adapt to production dynamics; the steady-state intelligent discrimination system dominated by Hotelling T2 can quickly locate the abnormal source and give early warning of potential risks through dynamic control limit setting and multi-scale anomaly diagnosis; the hierarchical intelligent control strategy can take different response measures according to the abnormal degree, forming a closed-loop control; the overall millisecond-level response to multiple parameter collaborative anomalies is realized, the steady-state maintenance rate is improved, and the monitoring blind area and control lag problem of traditional methods are effectively solved. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to make the content of the present application more easily understood, the present application will be further described in detail below according to specific embodiments of the present application and in conjunction with the drawings, wherein,

[0068] Figure 1 is a structural schematic diagram of a multivariate intelligent monitoring system for a pellet fluidized bed drying process provided in an embodiment of the present application;

[0069] Figure 2 is a flowchart of obtaining multivariate data results provided in an embodiment of the present application;

[0070] Description of the Drawings: 100, multi-parameter acquisition module; 200, multivariate data analysis processing module; 300, abnormal early warning and intelligent control module. DETAILED DESCRIPTION

[0071] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it. The embodiments are not limiting to the present application.

[0072] As shown in Figure 1 , the present application provides a multivariate intelligent monitoring system for a pellet fluidized bed drying process, comprising: a multi-parameter acquisition module 100, a multivariate data analysis processing module 200 and an abnormal early warning and intelligent control module 300; wherein,

[0073] The multi-parameter acquisition module 100 is configured to acquire multivariate heterogeneous monitoring data composed of process parameters and material attribute parameters in real time during the fluidized bed drying process; wherein the multivariate heterogeneous monitoring data comprises:

[0074] a. Continuous data: continuous numerical sequence changing with time, such as drying temperature, humidity, air speed, etc.

[0075] b. Discrete / classification data: numerical representation of discrete labels such as equipment operation mode, initial state classification of material, etc.

[0076] c. Other heterogeneous data: standardized pressure, flow and other specific physical quantity monitoring data.

[0077] The multivariate data analysis processing module 200 is configured to perform correlation analysis based on the multivariate heterogeneous monitoring data through multivariate statistical analysis method, identify the coupling relationship between parameters, and obtain multivariate data analysis results.

[0078] The abnormal early warning and intelligent control module 300 is configured to establish an abnormality discrimination model based on the multivariate data analysis results, timely warn when multivariate coupling abnormality is detected, and automatically adjust the control strategy according to the severity of the abnormality, to realize intelligent control of the pellet fluidized bed drying process.

[0079] Further, as shown in Figure 2 The multivariate data analysis processing module 200 performs correlation analysis based on the multidimensional heterogeneous monitoring data through a multivariate statistical analysis method to identify the coupling relationship between parameters and obtain multivariate data analysis results, including:

[0080] The multidimensional heterogeneous monitoring data (such as temperature, humidity, bed pressure drop, equipment operation mode, etc.) is preprocessed by denoising and standardization to obtain preprocessed monitoring data, including:

[0081] The missing values are filled in by interpolation method (such as linear interpolation); the abnormal values are identified by 3σ rule or box plot, and are corrected (such as replaced by adjacent normal values) or removed after combining with process logic judgment;

[0082] For continuous data (such as drying temperature, humidity, wind speed, etc. changing with time), Z-score standardization method is used to convert to uniform scale with mean 0 and variance 1; for discrete / categorical data (such as equipment operation mode "automatic / manual", material initial state "high humidity / medium humidity / low humidity"), label encoding or one-hot encoding is used to convert to numerical representation (such as "automatic = 1, manual = 0"); for other heterogeneous data (such as preprocessed pressure, flow, etc. physical quantity), the standardization rule consistent with continuous data is used for conversion to ensure that all data are in uniform numerical format;

[0083] Based on the preprocessed monitoring data, determine the characteristic variables and target variables to be analyzed, and the target variables are the steady state data of the drying process; use labels for the characteristic variables and target variables, such as "1" for steady state and "0" for non-steady state;

[0084] Calculate the mutual information value between each characteristic variable and its corresponding target variable, denoted as the first mutual information value, which is used to reflect the importance of the characteristic variable for steady state discrimination; screen out potential relevant features according to the first mutual information value to form a candidate feature pool;

[0085] Calculate the mutual information value between different features in the candidate feature pool, denoted as the second mutual information value, which is used to quantify the redundancy between features. The higher the value, the more overlapping information the features have, and the stronger the redundancy is;

[0086] Integrate the first mutual information value and the second mutual information value to form a mutual information matrix, the row index value and the column index value of the mutual information matrix are both characteristic variables, and the matrix element is the corresponding mutual information value. This matrix can intuitively present the correlation between features and targets and the redundancy between features;

[0087] Based on the mutual information matrix, screen out a set of key characteristic variables strongly related to drying steady state from the candidate feature pool;

[0088] According to the set of key feature variables, a Hotelling's T2 statistic is calculated, which is a multivariate data analysis result.

[0089] Further, in the embodiment, based on the mutual information matrix, a set of key feature variables strongly related to the drying steady state is screened from the candidate feature pool by a genetic algorithm, including:

[0090] The feature subsets in the candidate feature pool are represented by a binary coding method, including: each feature in the candidate feature pool is represented by a binary bit, where "1" represents that the feature is included in the current subset, and "0" represents that the feature is not included; for example, if the candidate feature pool contains 5 feature variables, the code "10110" represents that the 1st, 3rd, and 4th features are selected.

[0091] Based on the mutual information values in the mutual information matrix, a fitness function is designed to maximize the correlation between the features and the drying steady state target variable and minimize the redundancy between the features:

[0092]

[0093] where F is the fitness value, MI(X i ,Y) is the mutual information value of the ith feature variable and the drying steady state target variable Y in the mutual information matrix; MI(X i ,Y j ) is the mutual information value between the ith feature variable and the jth feature variable in the mutual information matrix; m is the number of features in the current feature subset; and a and β are weight coefficients for adjusting the influence weight of correlation and redundancy according to process requirements.

[0094] The selection operation is performed, including: calculating the fitness values of the feature subsets based on the fitness function, selecting high-quality feature subsets according to the fitness values by a roulette method according to a predetermined probability, the higher the fitness value, the higher the probability of being selected, and retaining the high-quality feature subsets to form a breeding population;

[0095] The crossover operation is performed, including: for the feature subsets in the breeding population obtained by the selection operation, performing single-point crossover according to a predetermined crossover probability: randomly selecting a crossover position from the binary code strings of any two parent feature subsets, for example, randomly selecting the 3rd bit as the crossover point when the code string length is 10, and then exchanging all binary bits after the crossover position in the two parent code strings to generate new child feature subsets;

[0096] For example, the encoding strings of two parent feature subsets are "101100" and "010011" respectively, and if the 3rd bit is randomly selected as the crossover position, the binary bits after the crossover are exchanged, and the resulting offspring encoding strings are "101011" and "010100". Through this operation, the offspring feature subsets inherit part of the feature selection information of the parents and generate new feature combinations, which helps to explore better feature subsets in the iteration process, i.e., a core variable set that is strongly related to the drying steady state and has low redundancy, and provides accurate feature input for subsequent steady state monitoring and control of the drying process;

[0097] The mutation operation is performed, including: for the offspring feature subset, randomly flipping any one binary bit according to a preset mutation probability, to further introduce new feature combinations;

[0098] The selection, crossover and mutation operations are repeatedly performed, and the iteration is continued until a convergence condition that the change value of the fitness value meets a continuous preset number of generations is within a preset range, and the feature subset with the highest fitness value is obtained as the key feature variable set X = [X1, X2, …, Xn] that is strongly related to the drying steady state. n-1 n n is the dimension.

[0099] Specifically, in the embodiment, according to the key feature variable set, a Hotelling's T2 statistic is calculated, including:

[0100] A sliding time window with a step size of 300s is used to sample the historical data of the key feature variable set X in a specified time period (the previous 300s) before the current time in real time;

[0101] Based on the historical data, the sample mean vector (the mean value of each variable in the window) and the covariance matrix (reflecting the linear correlation between variables) of the key feature variable set are calculated; when a material batch switching or a process parameter (such as the inlet air temperature set value) change is detected, the current sliding window data is automatically emptied, and the sample mean vector and the covariance matrix are recalculated to ensure that the model adapts to the new working condition;

[0102] To ensure that the calculation time of the single-sample T2 statistic is <10ms, meeting the real-time requirement (millisecond-level response) of online monitoring of the pellet fluidized bed drying process, the operation process of the sample mean vector and the covariance matrix is accelerated through GPU (graphics processing unit) parallel computing technology, and the Hotelling's T2 statistic is obtained:

[0103]

[0104] where n is the sample size in the sliding time window, X is the current sampling data, is the sample mean vector, T is the transpose of the matrix, and S is the covariance matrix.​-1 is the inverse matrix of the covariance matrix.

[0105] Further, in the present embodiment, based on the multivariate data analysis result, an abnormality discrimination model is established, and timely warning is given when a multivariate coupling abnormality is detected, including:

[0106] A plurality of batches (more than 200 batches) of historical normal production data (including a complete set of key characteristic variables X and corresponding steady-state identifiers) are obtained, ensuring that the data covers normal conditions under different seasons, material batches, and environmental conditions;

[0107] Based on the data, a dynamic correction coefficient λ related to production time, environmental temperature and humidity is introduced, and Bootstrap sampling and F distribution are used to calculate a control threshold that can be adaptively adjusted according to environmental and process changes :

[0108]

[0109] where p is the number of variables in the set of key characteristic variables selected from the multi-dimensional heterogeneous monitoring data, n is the sample size in the sliding time window, F p,n-p,α is the F distribution quantile with degrees of freedom p and n-p and significance level α, and α is usually 0.05, i.e. 95% confidence level;

[0110] Based on the real-time obtained multi-dimensional heterogeneous monitoring data, the Hotelling’s T2 statistical value T 2 is calculated, i.e. the multivariate data analysis result, and a comparison result between the control threshold and the multivariate data analysis result is obtained to perform abnormality diagnosis from a micro perspective and a macro perspective:

[0111] When , a micro perspective abnormality diagnosis process is performed, including:

[0112] The contribution degree C 2 of the i-th key characteristic variable selected from the multi-dimensional heterogeneous monitoring data to the abnormality of T i is calculated: S ii represents the variance of the i-th key characteristic variable in the sliding time window, i = 1, 2, …, p, and p is the total number of key characteristic variables; X i represents the monitoring value of the i-th key characteristic variable at the current sampling time, represents the sample mean vector of the i-th key characteristic variable in the sliding time window;

[0113] According to the contribution degree, all key characteristic variables are sorted, and the top K characteristic variables in the contribution degree are selected and determined as the abnormality source.

[0114] Performing a macroscopic angle anomaly diagnosis process, including:

[0115] Taking the latest n minutes of T 2 The statistical sequence input into the trained LSTM model obtains the T 2 The predicted value, if the T 2 The predicted value meets any of the following conditions, triggering an early warning:

[0116] Condition one, the predicted value exceeds the control threshold The preset proportion (90%);

[0117] Condition two, the rising rate of the predicted value exceeds M=1.5 times the historical normal trend.

[0118] Further, in this embodiment, according to the severity of the anomaly, the control strategy is automatically adjusted, including:

[0119] Based on the comparison result of the Hotelling's T2 statistical value T 2 And the control threshold Trigger a three-level gradient response according to the severity of the anomaly to achieve closed-loop precise regulation of the drying process, as follows:

[0120] When Determine the early warning state, start the first level response, use the fuzzy PID method to adjust the operation parameters of the equipment corresponding to the abnormal source, the specific steps are as follows:

[0121] Based on the contribution degree C i Of each key feature variable, set the contribution degree threshold C thresh = 0.1·T 2 , screen out variables with C i >C thres As the main abnormal source, and calculate the dynamic change rate R If R i >2σ i , standard deviation Marked as a rapidly changing abnormal source;

[0122] For the screened abnormal source, its control weight ω i According to the contribution degree dynamically allocated: Where m is the number of abnormal sources, ensuring

[0123] Set the total control resource upper limit R total Including the total power adjustment amount, the total air volume change rate, and the abnormal source allocation resource R i = ω i ·Rtotal , high contribution variables are given priority;

[0124] Subsequently, a fuzzy PID controller is constructed with error and its rate of change Δe(t) as inputs and the device operating parameter adjustment amount Δu(t) as output, and PID parameters (K p , K p0 + ΔK p , K i = K i0 + ΔK i , K d = K d0 + ΔK d ) are dynamically adjusted through a pre-set fuzzy rule base (e.g., if e is large and the air volume contribution is high, then quickly reduce the air volume).

[0125] Finally, under the condition that each adjustment amount is ≤ the device safety threshold, the PID parameters are used to drive the actuator to perform gradual adjustment, and the T 2 change trend is measured after adjustment. If the rate of decline is insufficient, the adjustment amplitude is increased, and if it continues to rise, the response level is upgraded until T 2 ≤ , at which point the steady-state monitoring mode is entered.

[0126] When T , it is determined to be a significant abnormal state, and a secondary response is started. Based on historical control experience and real-time state, a global control scheme is generated that includes temperature curve reconstruction and hot air circulation path switching. The global control scheme is executed by the actuator, and T 2 change is monitored in real time. If T 2 falls back to within , it is automatically switched to a primary response for fine tuning.

[0127] When T , it is determined to be a serious fault state, and a tertiary response is started. The safety shutdown program is immediately executed, and a diagnostic report is output that traces the abnormal parameters and provides process adjustment recommendations. The specific implementation steps are as follows: immediately send an emergency command to the device control system: stop the air inlet heating module and cut off the heat source input; turn off the air supply fan and the induced draft fan to terminate the hot air circulation; start the emergency cooling device (such as introducing normal temperature inert gas) to prevent material quality deterioration (such as degradation of active ingredients) due to high temperature continuous drying;

[0128] Based on the results of microscale abnormality analysis combined with multiscale abnormality diagnosis, 1-2 core variables with the highest contribution (such as a sudden increase in bed pressure drop, which may correspond to material caking, and a sudden decrease in air volume, which may correspond to fan failure) are recorded, along with their specific values and change curves at the time of abnormality.

[0129] A diagnostic report is automatically generated that includes the following: abnormality occurrence time, T2 value peak and over-amplitude; main abnormal source variable and specific data deviating from steady state; suggested process adjustment measures (such as cleaning the fluidized bed screen to solve the bed pressure drop anomaly or calibrating the air volume sensor); synchronously pushing the diagnosis report to the production management system for reference by the operation and maintenance personnel.

[0130] Specifically, in the present embodiment, when , based on historical control experience and real-time state, a global control scheme including temperature curve reconstruction and hot air circulation path switching is generated, including:

[0131] Based on the DQN algorithm, a control model including a current Q network and a target Q network is constructed, as well as a corresponding state space S, action space A and policy space π(a|s), wherein the current Q network is a neural network structure including an LSTM layer, a Dropout layer and a fully connected layer, the structure of the target Q network is the same as that of the current Q network, s i ∈S represents the state at the current time i, including the current Hotelling T 2 value, real-time data of the key feature variable set X, initial moisture content of the material, environmental temperature and humidity; a i ∈A represents the executable global control action, including the temperature curve reconstruction strategy and the hot air circulation path switching strategy; the policy space π(a|s) represents the probability distribution of selecting action a under state s;

[0132] The parameters of the current Q network are initialized, the Q values Q(s,a) of all possible actions under the current state s are calculated based on the current Q network, and the action a is selected according to the ε-greedy exploration strategy;

[0133] The action a is executed in the environment, and the reward function is constructed based on the T 2 value, through which the immediate reward r is obtained and transferred to the next state s ′ ; combined with the immediate reward r and the next state s ′ , the target Q value y of selecting action a is calculated using the target Q network;

[0134] Based on the target Q value y and the Q value Q(s,a) predicted by the current Q network, the mean square error loss value is calculated, the parameters of the current Q network are updated by minimizing the loss value through back propagation, and the parameters of the current Q network are copied to the target Q network every preset iteration step;

[0135] The above iteration process is repeated until the network converges or the maximum iteration round is reached, and the trained control model is obtained;

[0136] The real-time state features are input into the trained regulation model, and an action maximizing the reward value is output, i.e. a global regulation scheme, i.e. a temperature curve reconstruction strategy: for example, adjusting 35℃×15min to 45℃×20min, and then to 50℃ constant temperature; a hot air circulation path switching strategy: for example, combining the ring-shaped air inlet path with an instant switching time sequence.

[0137] Further, the reward function encourages T 2 The statistical quantity quickly returns to the steady state, ensuring the stability of the adjustment process, avoiding system oscillation caused by excessive parameter adjustment, and highly adapting to the demand for balance and stability of the global regulation scheme, and its expression is as follows:

[0138] R = ω1·R1 + ω2·R2 - ω3·R3 + ω4·R4

[0139] R1 represents the T 2 statistical quantity, the speed of approach, T 2 statistical quantity at the current sampling time t, is the T statistical quantity at the last sampling time of T 2 statistical quantity, i.e. R1 is positive when T drops sharply to 0; when T rises, R1 = 0, and there is no reward;

[0140] R2 represents the stability of the T 2 statistical quantity during the drop, represents the T 2 statistical quantity at the last N sampling times, and σ(·) represents the standard deviation, and the smaller the fluctuation amplitude (σ), the closer R2 is to 1; if the fluctuation amplitude exceeds R2 = 0;

[0141] R3 represents the adjustment amplitude of the temperature and air volume parameters, ΔT represents the current air inlet temperature adjustment amount, ΔT max is the maximum step length of the air inlet temperature adjustment in the primary response, and ΔQ is the current air volume adjustment amount, ΔQ max is the maximum step length of the air volume adjustment in the primary response, and ρ1, ρ2 are weight coefficients, ρ1 + ρ2 = 1; the larger the adjustment amplitude, the larger the value of R3, and the stronger the punishment on the total reward value.

[0142] R4 is a reward value when T2 statistic falls back to control threshold reward value given at the inner time, ω1, ω2, ω3, ω4 are weight coefficients, ω1+ω2+ω3+ω4=2 is satisfied, preferably, ω1=0.6, ω2=0.5, ω3=0.5, ω4=0.4. The value range of the reward value R is [-1, 2], positive and negative feedback is clear, which can effectively guide the reinforcement learning model to learn the optimal strategy of "fast T 2 , stable volatility, and less adjustment".

[0143] In summary, based on the above technical solutions, the present application realizes the dual perception of "real-time monitoring+trend prediction" through multi-scale diagnosis, and realizes the gradient response of "fine tuning-global regulation-emergency treatment" through hierarchical control, which together guarantees the stability of the drying process: the positioning of abnormal sources of microscopic diagnosis improves the accuracy of control, the early warning of macro trends improves the initiative of control, and the dynamic adjustment of hierarchical strategies ensures the efficiency of control, ultimately achieving the goal of a stability maintenance rate of ≥99%.

[0144] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0145] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0146] These computer program instructions can also be stored in a computer-readable memory that can cause the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocksFigure 1 the function specified in the one or more blocks.

[0147] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 the flow or flows and / or blocks Figure 1 the function specified in the one or more blocks.

[0148] Obviously, the above embodiments are only examples for clearly illustrating the present application, and are not intended to limit the implementation modes. Based on the above description, other different forms of changes or variations can also be made by those skilled in the art. Here, all the implementation modes are not required or can not be exhausted. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A multivariate intelligent monitoring system for a pellet fluidized bed drying process, characterized in that, The method comprises the following steps: a multi-parameter acquisition module configured to acquire multi-dimensional heterogeneous monitoring data composed of process parameters and material attribute parameters in a fluidized bed drying process in real time; a multivariate data analysis processing module configured to perform correlation analysis on the multi-dimensional heterogeneous monitoring data by a multivariate statistical analysis method, identify the coupling relationship between parameters, and obtain multivariate data analysis results; and an abnormality early warning and intelligent control module configured to establish an abnormality discrimination model based on the multivariate data analysis results, timely warn when multivariate coupling abnormalities are detected, and automatically adjust the control strategy according to the severity of the abnormality to realize intelligent control of the fluidized bed drying process of the pellets.

2. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 1, wherein, The multivariate data analysis results are obtained by the following steps: performing denoising and standardization preprocessing operations on the multi-dimensional heterogeneous monitoring data to obtain preprocessed monitoring data; determining feature variables and target variables to be analyzed based on the preprocessed monitoring data, wherein the target variables are steady-state data of the drying process; calculating mutual information values between each feature variable and its corresponding target variable, denoted as first mutual information values, and screening out potential related features according to the first mutual information values to form a candidate feature pool; calculating mutual information values between different features in the candidate feature pool, denoted as second mutual information values; integrating the first mutual information values and the second mutual information values to form a mutual information matrix, wherein the row index values and the column index values of the mutual information matrix are feature variables, and the matrix elements are corresponding mutual information values; screening out a set of key feature variables strongly related to the drying steady state from the candidate feature pool based on the mutual information matrix; calculating Hotelling's T2 statistics according to the set of key feature variables, which are the multivariate data analysis results.

3. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 2, characterized in that, The set of key feature variables strongly related to the drying steady state is screened out from the candidate feature pool based on the mutual information matrix, which comprises the following steps: using a binary encoding method to represent feature subsets in the candidate feature pool, including: using binary bits to represent each feature in the feature pool, wherein "1" indicates that the feature is included in the current subset, and "0" indicates that the feature is not included; designing a fitness function containing the maximum correlation between features and the drying steady-state target variables and the minimum redundancy between features based on the mutual information values in the mutual information matrix; performing a selection operation, including: calculating the fitness values of each feature subset based on the fitness function, screening out high-quality feature subsets according to the fitness values according to a preset probability, and retaining the high-quality feature subsets to form a breeding population; performing a crossover operation, including: for the feature subsets in the breeding population obtained by the selection operation, performing a single-point crossover according to a preset crossover probability: randomly selecting a crossover position from the binary encoding strings of any two parent feature subsets, and then exchanging all binary bits after the crossover position in the two parent encoding strings to generate new offspring feature subsets; performing a mutation operation, including: randomly flipping any binary bit according to a preset mutation probability for the offspring feature subsets to further introduce new feature combinations; The selection, crossover and mutation operations are repeatedly performed, and the iteration is continued until a convergence condition that a variation value of the fitness value meets a preset number of generations is satisfied, and a feature subset with the highest fitness value is obtained, that is, a key feature variable set X=[X1, X2, …, Xn] highly correlated with the drying steady state, n is a dimension. n-1 , n ​ 4. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 3, wherein, The fitness function is as follows: where F is the fitness value, MI(X i ,Y) is the mutual information value of the ith feature variable and the dry steady-state target variable Y in the mutual information matrix; MI(X i ,Y j ) is the mutual information value between the ith feature variable and the jth feature variable in the mutual information matrix; m is the number of features in the current feature subset; and a and β are weight coefficients for adjusting the influence weight of the correlation and redundancy.

5. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 2, wherein, According to the set of key characteristic variables, a Hotelling's T2 statistic is calculated, including: Using a sliding time window, real-time sampling of the historical data of the set of key characteristic variables within a specified time period before the current time is performed; Based on the historical data, a sample mean vector and a covariance matrix of the set of key characteristic variables are calculated; when a material batch switching or process parameter change is detected, the current sliding window data is automatically emptied, and the sample mean vector and the covariance matrix are recalculated; Based on the sample mean vector and the covariance matrix, a Hotelling's T2 statistic is calculated: where n is the sample size in the sliding time window, X is the current sampling data, is the sample mean vector, T is the transpose of the matrix, S -1 is the inverse of the covariance matrix.

6. The multivariate intelligent monitoring system for pellet fluidized bed drying process of claim 1, wherein, Based on the results of multivariate data analysis, an abnormality discrimination model is established, and timely warning is given when a multivariate coupling abnormality is detected, including: A plurality of batches of historical normal production data are acquired, based on the data, a dynamic correction coefficient λ related to production time, environmental temperature and humidity is introduced, and a control threshold value capable of being self-adaptively adjusted according to environmental and process changes is calculated Based on the real-time acquired multi-dimensional heterogeneous monitoring data, a Hotelling's T2 statistic value T is calculated 2 , that is, a multivariate data analysis result, according to the comparison result of the control threshold value and the multivariate data analysis result, abnormal diagnosis is performed from a micro perspective and a macro perspective: When a micro-angle abnormality diagnosis procedure is performed, including: The contribution degree C of the anomaly is calculated by the following formula: 2 The contribution degree C of the anomaly is calculated by the following formula: i : S ii Xi represents the variance of the i-th key characteristic variable in the sliding time window, i = 1, 2, …, p, and p is the total number of key characteristic variables; X i Xi represents the monitoring value of the i-th key characteristic variable at the current sampling time, Xi represents the sample mean vector of the i-th key characteristic variable in the sliding time window; According to the contribution degree, all key characteristic variables are sorted, and the top K characteristic variables with the highest contribution degree are selected as the abnormal sources; A macroscopic angle abnormality diagnosis process is performed, including: The latest n minutes of T2 statistic sequence is input into the trained LSTM model to obtain the T2 prediction value in the future time period, and if the T2 prediction value meets any of the following conditions, an alarm is triggered: Condition one, the predicted value exceeds the control threshold a preset proportion; Condition two, the rising rate of the prediction value exceeds M times of the historical normal trend, M>1.

7. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 6, characterized in that, The control threshold The calculation method is as follows: where p is the number of variables in the set of key feature variables filtered from the multi-dimensional heterogeneous monitoring data, n is the sample size in the sliding time window, F p,n-p,α is the F-distribution quantile with degrees of freedom p and n-p, and significance level a.

8. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 6, wherein, According to the severity of the abnormality, the control strategy is automatically adjusted, including: Based on the value of Hotelling's T2 statistic T 2 In comparison with the control threshold value A three-level gradient response is triggered according to the severity of the anomaly, realizing closed-loop precise regulation of the drying process, and the steps are as follows: When , it is determined as an early warning state, a first level response is started, an abnormal source is screened out based on the contribution degree of each key characteristic variable, the weight parameter of the abnormal source is adjusted, it is ensured that the regulation and control resources are preferentially used for the abnormal source, the fuzzy PID method is used to adjust the device operation parameter corresponding to the abnormal source, and the process is repeated until When , it is determined that the significant abnormal state, the secondary response is started, the global control scheme including temperature curve reconstruction and hot air circulation path switching is generated based on historical control experience and real-time state, the global control scheme is executed, and T 2 is monitored in real time; if T 2 falls back to , the primary response is automatically switched to fine-tune. When a serious fault condition is determined, a three-stage response is initiated, an immediate safety shutdown procedure is executed, and a diagnostic report is outputted that traces the abnormal parameters and suggests process adjustments.

9. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 8, wherein, When a global control scheme containing temperature curve reconstruction and hot air circulation path switching is generated based on historical control experience and real-time state, including: A regulatory model including a current Q network and a target Q network and corresponding state space S, action space A and policy space π(a|s) are constructed, wherein the current Q network is a neural network structure including LSTM layer, Dropout layer and full connection layer, the structure of the target Q network is the same as that of the current Q network, s i ∈S represents the state at the current time i, including the current T 2 value, real-time data of the key feature variable set X, initial moisture content of the material, environmental temperature and humidity; a i ∈A represents the executable global regulation action, including temperature curve reconstruction strategy and hot air circulation path switching strategy; the policy space π(a|s) represents the probability distribution of selecting action a under state s; Initializing the parameters of the current Q network, calculating the Q value Q(s,a) of all possible actions under the current state s based on the current Q network, and selecting the action a according to the ε-greedy exploration strategy; performing an action a in an environment, based on T 2 a reward function by which an immediate reward r is obtained and a transition to a next state s is made ′ ; combining the immediate reward r and the next state s ′ using the target Q-network to calculate a target Q-value y for the selected action a; Based on the target Q value y and the Q value Q(s,a) predicted by the current Q network, the mean square error loss value is calculated, and the parameters of the current Q network are updated by backpropagation to minimize the loss value. The parameters of the current Q network are copied to the target Q network every preset iteration step; Repeat the above iteration process until the network converges or the maximum iteration round is reached to obtain the trained regulation and control model; The real-time state features are input into the trained regulation and control model, and the action that maximizes the reward value, i.e. the global regulation and control scheme, is output.

10. The multivariate intelligent monitoring system for pellet fluidized bed drying process according to claim 9, wherein, The reward function is as follows: R=ω1·R1+ω2·R2-ω3·R3+ω4·R4 wherein R1 represents T 2 statistical quantity to control threshold speed of approaching, T of the current sampling time t 2 statistical quantity, is T of the last sampling time of 2 statistical quantity; R2 represents T 2 stability in the statistical quantity falling process, T of the last N sampling times 2 statistical quantity, σ(·) represents standard deviation; R3 represents adjustment amplitude reflecting temperature and air volume parameters, ΔT represents current air inlet temperature adjustment amount, ΔT max is the maximum step length of air inlet temperature adjustment in the primary response, ΔQ is current air volume adjustment amount, ΔQ max is the maximum step length of air volume adjustment in the primary response, ρ1, ρ2 are weight coefficients, ρ1+ρ2=1; R4 is the reward value given when T 2 statistical quantity falls back to the control threshold inside, ω1, ω2, ω3, ω4 are weight coefficients, satisfying ω1+ω2+ω3+ω4=2.

Images