A method for early warning and time optimization of bed collapse in fluidized bed granulation process of drug particles

Abstract

The present invention discloses a method for early warning and time optimization of bed collapse in the fluidized bed granulation process of drug particles, comprising the following steps: Step S1: exporting the time series data of the granulation stage of Yangwei granules in the data acquisition and monitoring control (SCADA) system, and performing data preprocessing; Step S2: setting two golden batch selection modes (mode 1 and mode 2) and determining the specific parameter range; Step S3: performing correlation analysis on the data and extracting features; Step S4: constructing a data set and using the linear SVC method to establish an early warning model. The present invention effectively utilizes the industrial production time series data, extracts the data features of the granulation stage of Yangwei granules from it, and can effectively predict the occurrence of Yangwei granules according to the pre-production t F The time series parameter curve within a certain period of time determines whether the corresponding production batch has the risk of bed collapse, and prompts in time to avoid bed collapse, effectively reducing production time and improving production efficiency.

Description

Technical Field

[0001] The present invention relates to the application of industrial time series data analysis in the pharmaceutical field, and in particular to a bed collapse warning and time optimization method for a fluidized bed granulation process of drug particles. Background Art

[0002] Industrial time series data refers to the general term for data that can be collected through technologies such as the Internet of Things and the Internet in a complete industrial chain or the entire life cycle of a product and has a logical sequence relationship. Industrial data is not only characterized by time, but any data that is logically distributed in sequence belongs to the category of industrial time series big data. The sources of industrial time series data mainly include three aspects: enterprise management time series data, equipment Internet of Things time series data, and external time series data. The characteristics of industrial time series data are mainly: large data volume, rapidity, accuracy, closed loop, time series, strong correlation, and high dimensionality. The feature representation of industrial time series data refers to the analysis of time series data in the industry, finding the characteristics of time series data through various means and expressing them for subsequent research on time series data mining. This method depicts the trend characteristics of time series data as a whole, but the determination of the two parameters of the length of each segment and the number of segments increases the difficulty of feature representation.

[0003] Chinese invention patent CN111815053B discloses a prediction method and system for industrial time series data. The method does not rely on an accurate mathematical model and can filter and correct errors in sensor monitoring values.

[0004] Under the demand environment of national policy promotion, strengthening of quality supervision and improvement of production efficiency, many Chinese medicine manufacturers have embarked on the road of digital transformation. At present, it is a common method to integrate real-time data such as temperature, pressure, flow rate and so on in the extraction, concentration and granulation of Chinese medicine production based on the data acquisition and monitoring system (SCADA) and monitor production. As production proceeds, a large amount of industrial time series production data is stored in the system. Utilizing these data and extracting the process laws hidden in the production data through data preprocessing, feature extraction, data modeling and other steps is of great significance to the realization of production fault warning and pharmaceutical process optimization.

[0005] Chinese invention patent CN111444246B discloses a method and system for automatically generating granular linear trends based on time series data. This method uses the original time series data and the time granularity table to finally obtain trend items of each granularity, helping people to clearly decompose the trends of monitoring indicators at different granularities, and then analyze the impact of different events in production.

[0006] Chinese invention patent CN112807370B discloses a method for preparing a stomach-nourishing granule preparation, which is a compound preparation composed of roasted astragalus, codonopsis, tangerine peel, cyperus, white peony root, yam, black plum, and licorice, and has the effects of nourishing the stomach and spleen, regulating qi and having a moderate therapeutic effect.

[0007] The granulation process of Yangwei Granules adopts fluidized bed granulation technology with high granulation strength and fast production speed. However, in actual production, this variety is prone to bed collapse risk during the granulation stage, affecting production efficiency and quality. At present, in order to prevent the risk of bed collapse from becoming a fact, workers need to adjust equipment parameters in time to intervene in the granulation process. However, it is difficult to deeply understand the production rules and causes of failures by relying solely on human experience, and subjective judgments vary from person to person, making it difficult to formulate unified standards. Therefore, it is particularly important to provide a dynamic early warning method for fluidized bed granulation bed collapse based on production big data. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a method for early warning of bed collapse and time optimization in a fluidized bed granulation process of drug particles.

[0009] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0010] A method for early warning and time optimization of bed collapse in a fluidized bed granulation process of drug particles, comprising the following steps:

[0011] Step S1: exporting the time series data of the drug granulation stage in the data acquisition and monitoring control (SCADA) system and performing data preprocessing;

[0012] Step S2: setting two golden batch selection modes (mode 1 and mode 2) and specific parameters;

[0013] Step S3: perform correlation analysis on the data and extract features;

[0014] Step S4: Construct a data set and use the linear SVC method to establish an early warning model.

[0015] Furthermore, in step S1, the time series data of the drug granulation stage in the SCADA system is exported, and the process of data preprocessing is as follows:

[0016] Step S11: exporting the time series data of seven parameters, namely, bed pressure difference, spray gun pressure, peristaltic pump flow, material temperature, air inlet temperature, air volume, and air outlet temperature, during the drug granulation stage from the SCADA system;

[0017] Step S12: Automatically segment the time series data to obtain the time series data of a single production batch, comprehensively compare the time series characteristics of the seven parameters, and determine the start time of liquid injection in the granulation process of each batch of drug particles according to the flow rate Q characteristics of the peristaltic pump;

[0018] Step S13: distinguishing batches with bed collapse risk and batches without bed collapse risk according to the peristaltic pump flow rate Q characteristics, and recording the time when the bed collapse trend first appeared and the number of bed collapse trends in the batches with bed collapse risk;

[0019] Furthermore, the rule for the start time of liquid spraying in the granulation process of each batch of drug particles in step S12 is: a binary threshold is set to convert the peristaltic pump flow Q into a binary variable Qbinary, and the expression is as follows:

[0020]

[0021] Where i is a time point in the spraying stage; t is the first 20 minutes of the spraying process;

[0022] Then calculate the first-order difference of Qbinary, recorded as diff(Qbinary), the first point with a value of 1 in diff(Qbinary) is the candidate starting point, and the first point with a value of -1 is the candidate ending point; further eliminate the noise starting points based on the comparison of the number of alternative starting points and the number of alternative ending points in diff(Qbinary): if the number of starting points is greater than the number of ending points, the last starting point is recorded as the injection starting point; if the number of ending points is greater than the number of starting points, the first ending point is recorded as the injection ending point; the time between the injection starting point and the injection ending point is recorded as the injection duration.

[0023] Furthermore, the rule for distinguishing batches with bed collapse risk and batches without bed collapse risk in step S13 is: except for the start point of liquid injection, the point in the middle where diff(Qbinary)=-1 is recorded as the bed collapse time point during the liquid injection stage.

[0024] Furthermore, in step S2, the batches with no bed collapse risk and short liquid spraying time are called golden batches. The two golden batch selection modes are as follows:

[0025] The selection criteria for the golden batch of model 1 are: 1) no bed collapse trend; 2) the spraying time is within the range of 3 times the standard deviation of the average spraying time, that is:

[0026] D gk,1 ∈|mean(D)±3σ(D)|#(2)

[0027] Where D is the duration of each batch of spraying, D gk,1 It is the golden batch spraying duration selected according to mode 1.

[0028] The golden intervals of the seven time series parameters in step S11 of mode 1 are defined as the range of n times the standard deviation of the mean of each parameter of the golden batch as the parameter golden range, that is:

[0029] P g,1∈|Mean(P)±n*σ(P)|#(3)

[0030] Where P is the timing parameter value, P g,1 is the golden range of parameters defined by mode 1.

[0031] The selection criteria for the golden batch of mode 2 are: 1) no bed collapse trend; 2) the spraying time is less than the average spraying time - the difference between the average spraying time and the shortest spraying time in history (C%), and 3) outliers whose spraying time exceeds 3 times the standard deviation after screening are eliminated.

[0032]

[0033] D gk,2 ∈|Mean(D k,2 )±3σ(D k,2 )|#(5)

[0034] Where D k,2 is the golden batch spraying duration selected according to mode 2, D gk,2 For D k,2 The golden batch spraying time is determined after eliminating outliers.

[0035] The golden intervals of the seven timing parameters in step S11 in Mode 2 are defined as the range covered by all parameters of each golden batch, namely:

[0036] P g,2 ∈[Min(P),Max(P)]#(6)

[0037] Where P g,2 is the golden range of parameters defined by mode 2.

[0038] Furthermore, the specific process of performing the correlation analysis in step S3 is as follows:

[0039] Step S31: Use the Spearman correlation coefficient to perform correlation analysis on the bed pressure difference, spray gun pressure, peristaltic pump flow, material temperature, air inlet temperature, air volume, air outlet temperature and spraying time in the drug granulation production process, and select the parameters with high correlation and controllable with the spraying time or bed collapse risk, which are recorded as characteristic parameters P j .

[0040] Step S32: Determine the characteristic time period t of the spraying stage according to the distribution of the first bed collapse risk time F , which satisfies the first bed collapse risk occurrence time of more than 85% batches at t F back.

[0041] Furthermore, the specific process of feature extraction in step S3 is as follows:

[0042] For the selected golden batch, take the previous time t F Parameter P j Extract features based on the changing trend. Suppose a parameter P j The golden parameter range within a certain time range is The true value of the parameter is P j ={p i ,i∈t F}. Let the time series feature be F j ={f i ,i∈t F}, if the timing point is within the golden range, it is recorded as 0, and if it is outside the range, it is recorded as 1.

[0043]

[0044] In the formula, f i is the feature of each time point, and pi is the parameter value of each time point.

[0045] Furthermore, the process of constructing a data set in step S4 and establishing an early warning model using the linear SVC method is as follows:

[0046] Step S41: The model input data dimension is n k ×P j ×t F , n k Indicates batch, P j is the characteristic parameter in step S31, t F The characteristic time period in step S32 is the indicator of whether there is a risk of bed collapse; the model input data and the corresponding indicators are divided into a training set and a test set according to time to construct a data set;

[0047] Step S42: A linear SVC model is established by using 5-fold cross validation, and the model evaluation indicators are the area under the ROC curve of the training set and the prediction accuracy of the test set; and the parameter n in formula 3 and the parameter C in formula 4 in step S2 are examined to select the best golden batch selection mode and the optimal parameters, and the optimal model is established. Finally, the SVC linear binary classification equation of the optimal model is used to infer whether there is a warning risk of bed collapse.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. This method can be based on the drug granulation process before t F Predict whether there is a risk of bed collapse in time, effectively guide production, reduce production time, and improve production efficiency;

[0050] 2. This method relies on the dynamic early warning of fluidized bed granulation bed collapse based on production big data to avoid the bed collapse risk from becoming a bed collapse fact;

[0051] 3. This method is based on an industrial time series data model, which effectively avoids human intervention in the granulation process by adjusting equipment parameters, and further avoids relying on human experience to make it difficult to deeply understand production rules and causes of failures, making it easier to formulate unified standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of an industrial data analysis method according to an embodiment of the present invention;

[0053] Figure 2 This is a diagram of the parameter correlation analysis result of an embodiment of the present invention;

[0054] Figure 3 5-fold cross validation ROC curves of SVC modeling with different parameters in the embodiment of the present invention; in the figure: (a) mode 2, C = 0.7; (b) mode 1, n = 1; (c) mode 1, n = 2; (d) mode 1, n = 3;

[0055] Figure 4 It is a schematic diagram of the golden interval range of each parameter under the optimal C value of an embodiment of the present invention; in the figure: the dark area is the golden interval, the light area is the abnormal fluctuation interval, and the lines represent the parameter values ​​of a certain batch. DETAILED DESCRIPTION

[0056] In order to make the purpose and technical solution of the present invention clearer and easier to understand and grasp, the technical solution of the present invention will be clearly and completely described below in conjunction with embodiments.

[0057] Example 1

[0058] A method for dynamic early warning of bed collapse and production time optimization in the fluidized bed granulation process of Yangwei granules based on industrial time series data. The main process is shown in Figure 1 , the specific steps are:

[0059] S1: Export the time series data of the granulation stage of Yangwei Granules in the data acquisition and supervisory control (SCADA) system and perform data preprocessing.

[0060] S11: The time series data of seven parameters, including bed pressure difference, spray gun pressure, peristaltic pump flow, material temperature, inlet air temperature, air volume, and outlet air temperature, were exported from the SCADA system during the granulation stage of Yangwei Granules.

[0061] S12: Automatically segment the time series data to obtain the time series data of a single production batch, comprehensively compare the time series characteristics of the seven parameters, and determine the start time of liquid injection during the granulation process of each batch of Yangwei Granules based on the flow Q characteristics of the peristaltic pump.

[0062] Determine the start time of liquid injection during the granulation process of each batch of Yangwei Granules: Set the binary threshold to convert the peristaltic pump flow Q into a binary variable Qbinary, the expression is as follows:

[0063]

[0064] Where i is a time point in the spraying stage; t is the first 20 minutes of the spraying process.

[0065] Then calculate the first-order difference of Qbinary, recorded as diff(Qbinary), the first point with a value of 1 in diff(Qbinary) is the candidate starting point, and the first point with a value of -1 is the candidate ending point; further eliminate the noise starting points based on the comparison of the number of alternative starting points and the number of alternative ending points in diff(Qbinary): if the number of starting points is greater than the number of ending points, the last starting point is recorded as the injection starting point; if the number of ending points is greater than the number of starting points, the first ending point is recorded as the injection ending point; the time between the injection starting point and the injection ending point is recorded as the injection duration.

[0066] Determine the alternative starting point, alternative ending point, spraying starting point, spraying ending point and spraying duration.

[0067] Sp={i|i∈t,diff(Qbinary) i =1}#(2)

[0068] Ep={i|i∈t,diff(Qbinary) i =-1}#(3)

[0069] In the above formula, Sp represents the set of candidate starting points, and Ep represents the set of candidate ending points. If the number of elements in the two sets is the same, they are regarded as the starting point sets respectively. With the end point set If the number of candidate starting points is greater than the number of candidate ending points, the first element of the candidate starting point set is removed, and the remaining elements are regarded as the starting point set, and the candidate ending point set is the ending point set; if the number of candidate starting points is less than the number of candidate ending points, the last element of the alternative ending point set is removed, and the remaining elements are regarded as the ending point set, and the candidate starting point set is regarded as the starting point set. The difference between the ending point set and the starting point set is the injection duration set, denoted as D:

[0070]

[0071] In the formula, e i The set of end points A certain end time point in i is the starting point set The corresponding starting time point in .

[0072] S13: Distinguish batches with bed collapse risk and batches without bed collapse risk according to the peristaltic pump flow rate Q characteristics, and record the time when the bed collapse trend first appears and the number of times the bed collapse trend occurs in the batches with bed collapse risk;

[0073]

[0074] Determine the batches with and without bed collapse risk: Except for the starting point of liquid injection, the point where diff(Qbinary)=-1 is recorded as the bed collapse time point during the liquid injection stage. See the table below for specific values:

[0075] Table 1 Data analysis results of the fluidized bed granulation spraying stage of Yangwei granules (partial)

[0076] Spray start point Spray end point Spraying time (min) Bed collapse time (min) 2021-12-02 09:27:30 2021-12-02 10:49:30 82 0 2021-12-02 09:38:30 2021-12-02 10:58:00 79.5 0 2021-12-02 12:16:30 2021-12-02 13:39:30 83 0 2021-12-02 12:29:00 2021-12-02 13:55:00 86 0 2021-12-03 09:25:00 2021-12-03 11:03:00 98 29.5 2021-12-03 09:13:00 2021-12-03 10:50:30 97.5 19.5 2021-12-03 12:10:30 2021-12-03 14:28:00 137.5 22 2021-12-03 12:34:00 2021-12-03 14:12:00 98 28.5 2021-12-03 15:27:00 2021-12-03 17:05:30 98.5 30.5 2021-12-03 15:00:30 2021-12-03 16:57:30 117 27.5 2021-12-04 09:37:00 2021-12-04 10:58:00 81 0 2021-12-04 09:29:30 2021-12-04 10:57:00 87.5 0 2021-12-04 09:37:30 2021-12-04 11:12:30 95 0 2021-12-04 09:37:30 2021-12-04 10:55:00 77.5 0 2021-12-04 12:27:00 2021-12-04 14:23:00 116 0 2021-12-04 12:10:00 2021-12-04 13:51:30 101.5 0

[0077] Note: In this table, the bed collapse time is expressed by the relative time from the start of the spraying, that is, the bed collapse occurs at a certain time after the start of the spraying.

[0078] S2: Use mode 1 and mode 2 to select the golden batch respectively.

[0079] The selection criteria for the golden batch of model 1 are: 1) no bed collapse trend; 2) the spraying time is within the range of 3 times the standard deviation of the average spraying time, that is:

[0080] D gk,1 ∈|mean(D)±3σ(D)|#(6)

[0081] Where D is the duration of each batch of spraying, D gk,1 It is the golden batch spraying duration selected according to mode 1.

[0082] The golden intervals of the seven time series parameters in step S11 of mode 1 are defined as the range of n times the standard deviation of the mean of each parameter of the golden batch as the parameter golden range, that is:

[0083] P g,1 ∈|Mean(P)±n*σ(P)|#(7)

[0084] Where P is the timing parameter value, P g,1 is the golden range of parameters defined by mode 1.

[0085] The selection criteria for the golden batch of mode 2 are: 1) no bed collapse trend; 2) the spraying time is less than the average spraying time - the difference between the average spraying time and the shortest spraying time in history (C%), and 3) outliers whose spraying time exceeds 3 times the standard deviation after screening are eliminated.

[0086]

[0087] D gk,2∈|Mean(D k,2 )±3σ(D k,2 )|#(9)

[0088] Where D gk,2 It is the golden batch spraying duration selected according to mode 2.

[0089] The golden intervals of the seven timing parameters in step S11 in Mode 2 are defined as the range covered by all parameters of each golden batch, namely:

[0090] P g,2 ∈[Min(P),Max(P)]#(10)

[0091] Where P g,2 is the golden range of parameters defined by mode 2.

[0092] The number of gold batches obtained through the above steps is shown in the following table:

[0093] Table 2: Number of golden batches divided by mode 1

[0094] n 1 2 3 Number of gold batches 3 3 3

[0095] Table 3 Number of golden batches divided by mode 2

[0096] C 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Yellow batch times 2 5 8 12 21 33 44 59 80

[0097] S3: Perform correlation analysis on the data and extract features.

[0098] S31: Spearman correlation analysis was used to analyze the correlation of 9 parameters in the production process of Yangwei Granules, including spray gun pressure, peristaltic pump flow, material temperature, air inlet temperature, air volume, air outlet temperature, spray time, and whether there is a risk of bed collapse during the spray stage. The results are shown in the figure. Figure 2 , the spraying time is strongly negatively correlated with the peristaltic pump flow rate and the inlet temperature; the bed collapse trend is strongly negatively correlated with the inlet temperature, material temperature, and outlet temperature; the spray gun pressure and air volume are both correlated with the spraying time and the bed collapse trend. In addition, the inlet temperature, material temperature, and outlet temperature are strongly correlated. Considering that the outlet temperature is highly controllable, the spray gun pressure P, peristaltic pump flow Q, inlet temperature T1, and air volume V are finally selected as characteristic parameters, denoted as P j .

[0099] S32: Correlation analysis between bed collapse risk rate and spraying time, and determination of characteristic time period t of spraying stage F , which meets the requirement that the first bed collapse risk of 85% batches will occur at t F After that, in this embodiment, t F =20.

[0100] For the selected golden batch, take the previous time t F Parameter P j Extract features based on the changing trend. Suppose a parameter P j The golden parameter range within a certain time range is The true value of the parameter is P j ={p i ,i∈t F}. Let the time series feature be F j ={f i ,i∈t F}, if the timing point is within the golden range, it is recorded as 0, and if it is outside the range, it is recorded as 1.

[0101]

[0102] In the formula, f i is the feature of each time point, and pi is the parameter value of each time point.

[0103] S4: Construct a data set and use the linear SVC method to establish an early warning model.

[0104] S41: The model input data dimension is n k ×P j ×t F , n k Indicates batch, P j is the characteristic parameter in step S31, t F The characteristic time period in step S32, the indicator is whether there is a risk of bed collapse; the model input data and the corresponding indicators are divided into a training set and a test set according to time to construct a data set. In this embodiment, a total of 341 batches of Yangwei Granule granulation time series data were used for analysis, 240 batches of data from December 2021 to March 2022 were used as training sets, and 101 batches of data from April 2022 were used as test sets.

[0105] S42: A linear SVC model is established using 5-fold cross validation, and the model evaluation indicators are the area under the ROC curve of the training set and the prediction accuracy of the test set; and the parameter n in Formula 3 and the parameter C in Formula 4 in step S2 are examined to select the best golden batch selection mode and optimal parameters to establish the optimal model. The golden batches are divided according to Mode 1, and the modeling effects when n values ​​are 1, 2, and 3 are examined, as shown in Table 2; the golden batches are divided according to Mode 2, and the modeling effects of C values ​​of 0.1-0.9 are examined, as shown in Table 3. The modeling results show that the model effect is best when the golden batch is selected according to Mode 2 and C is 0.7. At this time, the area under the ROC curve is 0.95, and the accuracy of the test set is 0.81, indicating that the model has good predictive performance. Its ROC curve is shown in Figure 3 Under the best model effect, the golden intervals of each parameter are calculated according to step S2. Figure 4 shown.

[0106] The SVC linear binary classification equation obtained at this time is as follows:

[0107] Y=0.0544P+0.0872Q-0.0044T1+0.1191V-3.7574#(12)

[0108] Where Y is an indicator of bed collapse risk at the time of definition. If it is greater than 0, it indicates that this batch of production has a greater risk of bed collapse and requires attention. P, Q, T1, and V represent t F The spray gun pressure, peristaltic pump flow, air inlet temperature and air volume within a certain period of time (the first 20 minutes in this embodiment).

[0109] Table 4 Modeling effect of golden batch division in mode 1

[0110] n 1 2 3 Training set AUROC 0.92 0.91 0.91 Test set accuracy 0.71 0.59 0.59

[0111] Table 5 Modeling effect of golden batch division in mode 2

[0112] C 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Training set AUROC 0.91 0.94 0.95 0.95 0.92 0.89 0.89 0.84 0.74 Test set accuracy 0.69 0.70 0.81 0.79 0.73 0.64 0.72 0.57 0.53

[0113] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. A method for bed collapse warning and time optimization in a drug granulation process, characterized in that: The following steps are involved: Step S1: exporting the time series data of the drug granulation stage in the SCADA system and performing data preprocessing; Step S2: setting two golden batch selection modes and specific parameters; Step S3: perform correlation analysis on the data and extract features; Step S4: construct a data set and use the linear SVC method to establish an early warning model; The golden batch described in step S2 is a batch with no bed collapse risk and a short liquid spraying time; The extraction features described in step S3 are: for the selected golden batch, take its previous certain time t F Parameter P j Extract features based on changing trends; The F According to the distribution of the first bed collapse risk time, determine the characteristic time period t of the injection stage F ; The P j The parameter with high correlation with the duration of spraying or the risk of bed collapse and which is controllable is selected and recorded as the characteristic parameter P j ; The linear SVC method establishes the early warning model as follows: the model input data dimension is n k ×P j ×t F , n k Represents batch; 5-fold cross validation is used to establish a linear SVC model.

2. The method according to claim 1, characterized in that The step S1 specifically includes: Step S11: exporting the time series data of seven parameters, namely, bed pressure difference, spray gun pressure, peristaltic pump flow, material temperature, air inlet temperature, air volume, and air outlet temperature, during the drug granulation spraying stage from the SCADA system; Step S12: Automatically segment the time series data and determine the start time of liquid injection during the granulation process of each batch of drug particles according to the flow rate Q characteristics of the peristaltic pump; Step S13: Distinguish batches with bed collapse risk and batches without bed collapse risk according to the peristaltic pump flow rate Q characteristics, and record the time when the bed collapse trend first appears and the number of times the bed collapse trend occurs in the batches with bed collapse risk.

3. The method according to claim 2, characterized in that The rule for the start time of liquid spraying in the granulation process of each batch of drug particles described in step S12 is: set a binary threshold to convert the peristaltic pump flow Q into a binary variable Qbinary, and the expression is as follows: In the formula, i is the injection time point; t is the first 20 minutes of the injection process; Calculate the first-order difference of Qbinary, denoted as diff Qbinary , according to diff Qbinary The values ​​determine the spraying start point, spraying end point, spraying duration and bed collapse time point during the spraying stage.

4. The method according to claim 2, characterized in that: The two gold batch selection modes are as follows: The selection criteria for the golden batch of model 1 are: 1) no bed collapse trend; 2) the spraying time is within the range of 3 times the standard deviation of the average spraying time, that is: D gk，1 ∈|mean(D)±3σ(D)|#(2) Where D is the duration of each batch of spraying, D gk,1 is the golden batch spraying duration selected according to mode 1; The golden interval of the time series data of the seven parameters described in step S11 in mode 1 is defined as the parameter golden range, which is n times the standard deviation range above and below the mean of each parameter in the golden batch, that is: P g，1 ∈|Mean(P)±n*σ(P)|#(3) Where P is the timing parameter value, P g，1 is the golden range of parameters defined by mode 1; The selection criteria for the golden batch of mode 2 are: 1) no bed collapse trend; 2) the spraying time is less than the average spraying time - the difference between the average spraying time and the shortest spraying time in history (C%), and 3) outliers whose spraying time exceeds 3 times the standard deviation after screening are eliminated. D gk，2 ∈|Mean(D k，2 )±3σ(D k，2 )|#(5) Where D k,2 is the golden batch spraying duration selected according to mode 2, D gk,2 For D k，2 The golden batch spraying time is determined after removing outliers on the basis of the above; The golden interval of the time series data of the seven parameters described in step S11 in mode 2 is defined as the range covered by all parameters of each golden batch, namely: P g，2 ∈[Min(P), Max(P)]#(6) Where P g，2 is the golden range of parameters defined by mode 2.

5. The method according to claim 1, characterized in that The specific process of performing correlation analysis on the data in step S3 is as follows: Step S31: using the Spearman correlation coefficient to perform correlation analysis on the bed pressure difference, spray gun pressure, peristaltic pump flow, material temperature, air inlet temperature, air volume, air outlet temperature and spraying time in the drug granulation production process; Step S32: The first bed collapse risk of more than 85% of the batches occurs at t F back.

6. The method according to claim 1, characterized in that The specific process of extracting features in step S3 is as follows: Suppose the golden parameter interval of a certain parameter Pj within a certain time range is The true value of the parameter is P j ={p i , i∈t F }; let the time series feature be F j ={f i , i∈t F If the timing point is within the golden range, it is recorded as 0, and if it is outside the range, it is recorded as 1, that is, In the formula, f i is the feature of each time point, and pi is the parameter value of each time point.

7. The method according to claim 1, characterized in that The process of constructing a data set and establishing an early warning model using the linear SVC method described in step S4 is as follows: Step S41: Divide the model input data and corresponding indicators into a training set and a test set according to time to construct a data set; Step S42: The model evaluation indicators are the area under the ROC curve of the training set and the prediction accuracy of the test set; and the parameter n in Formula 3 and the parameter C in Formula 4 in the step S2 are examined to select the best golden batch selection mode and the optimal parameters, establish the optimal model, and finally use the SVC linear binary classification equation of the optimal model to infer whether there is a warning risk of bed collapse.

8. The method according to any one of claims 1 to 7, characterized in that: The drug granules are stomach-nourishing granules.

Images