Pelletizing process fault monitoring and diagnosing method based on variational auto-encoder and contribution diagram

Through the variational autoencoder and contribution graph method, the problem that traditional methods are difficult to deal with nonlinear and non-Gaussian data is solved, and high-precision monitoring and fault diagnosis of drug particles production process is realized, and monitoring accuracy and fault diagnosis capabilities are improved. Adapt to different production environments, reduce costs, and ensure the quality of drugs.

CN119066569APending Publication Date: 2024-12-03NORTHEASTERN UNIV AT QINHUANGDAO
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411085834.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Traditional multivariate statistical analysis methods are difficult to process complex nonlinear and non-Gaussian data in industrial production, resulting in limited monitoring performance and inability to effectively capture complex relationships and non-linear features in the data, affecting the monitoring and failure of drug particles production process. Diagnostic effect.

Method used

采用基于变分自编码器和贡献图法的方法,通过变分自编码器模型对历史生产数据进行预处理和特征提取,将高维数据映射到低维潜在空间,自适应提取数据中的复杂非线性特征,并通过监测统计量和控制限进行故障诊断。

Benefits of technology

It realizes high-precision monitoring and fault diagnosis of drug granules production process, can detect and locate potential faults earlier, improves monitoring accuracy and fault diagnosis capabilities, adapts to different production environments and process conditions, reduces production costs, and ensures drug quality and Patient safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_1
    Figure SMS_1
  • Figure SMS_3
    Figure SMS_3
  • Figure SMS_5
    Figure SMS_5
Patent Text Reader

Abstract

The invention provides a drug continuous granulation process monitoring and fault diagnosis method based on a variational auto-encoder and a contribution diagram, and aims to solve the defects in processing complex nonlinear and non-Gaussian data in the prior art. According to the method, dimensionality reduction and feature extraction of high-dimensional data are realized by constructing and training a variational auto-encoder model, potential vectors are generated through a re-parameterization technique, and then input data are reconstructed. According to the method, monitoring statistics based on KL divergence are designed and used for monitoring key quality attributes and key process parameters in the production process in real time. In addition, the influence of each variable on monitoring statistics is evaluated by using a contribution graph method, and accurate diagnosis and positioning of faults are realized. Compared with a traditional multivariate statistical process monitoring method, the method has higher adaptability and higher monitoring precision, potential faults can be found in the early stage, quality problem accumulation is avoided, production efficiency is improved, production cost is reduced, and medicine quality and patient safety are guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of raw material mixing and continuous granulation, and relates to the problem of key quality attribute monitoring and fault diagnosis of drug particles, and specifically is a method for monitoring and fault diagnosis of continuous production process of drug particles based on variational autoencoder and contribution graph method. Background Art

[0002] In the past decade, with the promotion of Pharmaceutical 4.0 (i.e. the application of Industry 4.0 in the pharmaceutical field), the continuous pharmaceutical model has gradually become dominant in the pharmaceutical industry due to its advantages such as improving drug production efficiency, reducing drug manufacturing costs, improving drug product quality and enhancing production flexibility.

[0003] The quality of drugs is directly related to the health and safety of patients, and strict monitoring must be carried out to ensure that each batch of drugs meets the standards. Due to the complexity and efficiency of the continuous pharmaceutical process, each production link needs to be precisely controlled. Real-time monitoring of key quality attributes and key process parameters can detect abnormal conditions in a timely manner, quickly diagnose faults, avoid the accumulation of quality problems and production interruptions, thereby improving production efficiency and reducing costs. This is of great significance for ensuring drug quality and patient safety.

[0004] With the rapid development of industrial process automation and informatization, a large amount of data is collected and stored in industrial production, and data-driven process monitoring technology has developed rapidly. In the field of continuous pharmaceutical process monitoring, one of the most popular monitoring technologies is Multivariate Statistical Process Monitoring (MSPM), such as Principal Component Analysis (PCA) and Partial Least Squares (PLS).

[0005] However, modern industrial production processes are becoming increasingly complex, and there is a high degree of correlation and nonlinearity between process variables. Traditional MSPM methods are difficult to handle this situation, which in turn affects their monitoring performance. In order to solve the monitoring problem of nonlinear processes, researchers have proposed different types of monitoring methods. Among them, the most commonly used methods are kernel methods, such as kernel principal component analysis (KPCA), kernel partial least squares (KPLS), etc. However, due to the limitations of kernel technology, these methods are only suitable for processing small and medium-sized samples. In order to further improve the monitoring performance, some scholars have applied deep learning methods to the field of process monitoring, such as autoencoder (AE), denoising autoencoder (VAE), restricted Boltzmann machine (RBM), etc.

[0006] However, none of the above networks can handle the non-Gaussian problem of data well. The non-linearity and non-Gaussianity of data directly affect the extraction of data features. By processing the non-linearity and non-Gaussianity of data, the essential features of the data can be extracted. Summary of the invention

[0007] In order to solve the problems existing in the prior art, the present invention provides a method for monitoring and fault diagnosis of continuous production process of drug particles based on variational autoencoder and contribution graph method. The invention is not limited by prior knowledge, can capture the complex relationship and nonlinear characteristics between variables in the data, adaptively extract the feature information hidden in the original data, map high-dimensional data to low-dimensional latent space, realize data dimensionality reduction and feature extraction, and has strong adaptability.

[0008] The technical solution of the present invention is:

[0009] The method for monitoring and diagnosing the continuous production process of drug particles based on variational autoencoders and contribution graph method is characterized by comprising the following steps:

[0010] Step 1: Collection and preprocessing of historical data:

[0011] The established rotary agitator and twin-screw extruder wet granulation model was used to simulate the actual production process, and the properties and operating variables during normal operation, including the agitator inlet raw material mass flow rate, excipient mass flow rate, agitator speed, agitator outlet mass flow rate, adhesive mass flow rate, twin-screw extruder screw speed, average residence time of particles in the equipment, particle size, and water content, were collected.

[0012] The collected historical production data of the granulation process were standardized using the following formula as a training set.

[0013]

[0014] Among them, x jk is the jth data of the kth variable in the data set, is the mean of the kth variable in the data set, σ k is the standard deviation of the kth variable in the data set.

[0015] Step 2: Build a variational autoencoder model:

[0016] The variational autoencoder model consists of two parts: a probability encoder and a probability decoder.

[0017] The encoder module consists of three fully connected layers, with an input dimension of 9. After being processed by the ReLU activation function, the output dimensions are 128 and 64 respectively, and finally the mean and logarithmic variance of the latent space are output; the latent vector z is generated by the reparameterization technique. The decoder module also consists of three fully connected layers, with an input dimension of 3. After being processed by the ReLU activation function, the output dimensions are 64 and 128 respectively, and the final output dimension is 9. The Sigmoid activation function is used to decode the latent space vector into a high-dimensional input data form and reconstruct the original input data.

[0018] Step 3: Train the variational autoencoder model established in step 2:

[0019] During the training process, the optimization steps of the model include forward propagation, loss calculation, and parameter update. During forward propagation, the input data is passed through the encoder to generate the mean and log variance, the latent vector is obtained through the reparameterization technique, and then the decoder is used to generate the reconstructed data. The loss function consists of reconstruction loss and KL divergence loss. The former measures the difference between the reconstructed data and the original input data, and the latter measures the difference between the latent space distribution and the standard normal distribution. Finally, back propagation is performed based on the total loss to update the parameters of the encoder and decoder.

[0020] Step 4: Design monitoring statistics and calculate control limits:

[0021] The distribution of data generated in the latent space during normal operation is regarded as the benchmark distribution. The KL divergence between the distribution of monitoring data in the latent space and the benchmark distribution is regarded as a monitoring statistic, which is used to measure the degree to which the monitoring data deviates from normal working conditions and is recorded as KLD.

[0022] In the process of calculating the control limits, we first input K groups of normal working data into the variational autoencoder model, calculate the KLD between it and the benchmark distribution, and record it. Then, we use the kernel density estimation method to estimate the probability density function of this group of KLDs, and calculate the control limits at a confidence level of 0.98.

[0023] Step 5: Conduct production process monitoring.

[0024] Step 6: Calculate the contribution of each variable to the monitoring statistics to achieve fault diagnosis.

[0025] In order to evaluate the impact of each variable on the monitoring statistic KLD, the present invention adopts the following method for processing.

[0026] First, there is a historical normal data set X:

[0027]

[0028] In the formula, Represents the nth data corresponding to the i-th variable in the normal data set.

[0029] Secondly, the detected anomaly data set X′:

[0030]

[0031] In the formula, is the nth data corresponding to the pth variable in the abnormal data set.

[0032] In order to calculate the impact of each variable on the monitoring statistic KLD, the data is processed as follows:

[0033]

[0034] Where, X D ( i ) means that the i-th variable in the abnormal data is used to replace the i-th variable in the normal historical data, and p means that the production process is sampled p times.

[0035] After replacing the I group X D ( i ) are sent to the variational autoencoder monitoring model to calculate the KL divergence between it and the reference distribution, denoted as KLD( i ).

[0036] Therefore, the contribution R of each variable to KLD can be calculated by the following formula:

[0037]

[0038] Through the above method, the contribution of each variable to KLD can be analyzed, thereby locating the fault location more accurately.

[0039] The beneficial effects of the present invention are as follows: the present invention realizes high-precision monitoring and fault diagnosis of the drug particle production process by introducing a variational autoencoder combined with a contribution graph method. The method adaptively extracts complex nonlinear features hidden in the original data, and maps high-dimensional data to a low-dimensional latent space, thereby improving monitoring accuracy and fault diagnosis capabilities. Compared with traditional methods, the present invention can detect and locate potential faults earlier, avoid the accumulation of quality problems, has strong adaptability, adapts to different production environments and process conditions, improves production efficiency, and reduces production costs. At the same time, the present invention performs well in processing the non-Gaussianity of data, can extract the essential characteristics of the data, and ensure drug quality and patient safety. It is an advanced and effective method for monitoring and fault diagnosis of drug particle production processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is the schematic diagram of the variational autoencoder model;

[0041] Figure 2 is a flow chart of the method of the present invention;

[0042] Figure 3 It is the calculation process diagram of control limit;

[0043] Figure 4 This is the normal data distribution and each fault monitoring result diagram;

[0044] Figure 5 A comparison diagram of the fault diagnosis results of the method of the present invention and the traditional PCA method;

[0045] Figure 6 It is a comparison chart of monitoring performance indicators between the method of the present invention and the traditional PCA method; DETAILED DESCRIPTION

[0046] The specific implementation of the present invention is described in detail below in conjunction with the technical scheme and the accompanying drawings.

[0047] This implementation case selects 9 key variables in the pellet production process for monitoring and fault diagnosis. Figure 2 The flow chart of the drug particle quality monitoring method for this process includes the following steps:

[0048] Step 1: Collection and preprocessing of historical data:

[0049] The established rotary agitator and twin-screw extruder wet granulation model was used to simulate the actual production process, and the properties and operating variables during normal operation, including the agitator inlet raw material mass flow rate, excipient mass flow rate, agitator speed, agitator outlet mass flow rate, adhesive mass flow rate, twin-screw extruder screw speed, average residence time of particles in the equipment, particle size, and water content, were collected.

[0050] The collected historical production data of the granulation process were standardized using the following formula as a training set.

[0051]

[0052] Among them, x jk is the jth data of the kth variable in the data set, is the mean of the kth variable in the data set, σ k is the standard deviation of the kth variable in the data set.

[0053] Step 2: Build a variational autoencoder model:

[0054] Attached Figure 1 The data flow in the variational autoencoder is shown. The encoder module consists of three fully connected layers with an input dimension of 9. After being processed by the ReLU activation function, the output dimensions are 128 and 64 respectively. Finally, the mean and logarithmic variance of the latent space are output; the latent vector z is generated by the reparameterization technique. The decoder module also consists of three fully connected layers with an input dimension of 3. After being processed by the ReLU activation function, the output dimensions are 64 and 128 respectively, and the final output dimension is 9. The Sigmoid activation function is used to decode the latent space vector into a high-dimensional input data form and reconstruct the original input data.

[0055] Step 3: Train the variational autoencoder model established in step 2:

[0056] During the training process, the optimization steps of the model include forward propagation, loss calculation, and parameter update. During forward propagation, the input data is passed through the encoder to generate the mean and log variance, the latent vector is obtained through the reparameterization technique, and then the decoder is used to generate the reconstructed data. The loss function consists of reconstruction loss and KL divergence loss. The former measures the difference between the reconstructed data and the original input data, and the latter measures the difference between the latent space distribution and the standard normal distribution. Finally, back propagation is performed based on the total loss to update the parameters of the encoder and decoder.

[0057] Step 4: Design monitoring statistics and calculate control limits:

[0058] The distribution of data generated in the latent space during normal operation is regarded as the benchmark distribution. The KL divergence between the distribution of monitoring data in the latent space and the benchmark distribution is regarded as a monitoring statistic, which is used to measure the degree to which the monitoring data deviates from normal working conditions and is recorded as KLD.

[0059] Attached Figure 3 The kernel density estimation result is the specific result of the control limit. In the process of calculating the control limit, firstly, the K group of normal working data is input into the variational autoencoder model, and the KLD between it and the benchmark distribution is calculated and recorded. Secondly, the kernel density estimation method is used to estimate the probability density function of this group of KLD, and the control limit is calculated at a confidence level of 0.98.

[0060] Step 5: Conduct production process monitoring. Figure 4 The normal data distribution and the monitoring results of the three types of faults designed in Table 1 are shown. Taking the working conditions shown in Table 1 as an example, the monitoring values ​​of each variable fluctuate around the reference value, and the fluctuation range does not exceed 5%. The other three figures show the performance of the process monitoring method based on variational autoencoder when facing types 1 to 3 faults. It can be observed that when these three types of faults occur, the value of the monitoring statistic KLD increases rapidly and is significantly higher than the control limit. It can be concluded that the monitoring method based on variational autoencoder can timely monitor the occurrence of faults such as oscillation, step, and slow drift, which confirms the effectiveness of the monitoring method.

[0061] Table 1 Example fault design

[0062] Step 6: Calculate the contribution of each variable to the monitoring statistics to achieve fault diagnosis.

[0063] In order to evaluate the impact of each variable on the monitoring statistic KLD, the present invention adopts the following method for processing.

[0064] First, there is a historical normal data set X:

[0065]

[0066] In the formula, Represents the nth data corresponding to the i-th variable in the normal data set.

[0067] Secondly, the detected anomaly data set X′:

[0068]

[0069] In the formula, is the nth data corresponding to the pth variable in the abnormal data set.

[0070] In order to calculate the impact of each variable on the monitoring statistic KLD, the data is processed as follows:

[0071]

[0072] Where, X D(i) It means that the i-th variable in the normal historical data is replaced by the i-th variable in the abnormal data, and p means that the production process is sampled p times.

[0073] The 9 groups of X after replacement D(i) They are respectively sent to the variational autoencoder model to calculate the KL divergence between them and the reference distribution, denoted as KLD (i) .

[0074] Therefore, the contribution R of each variable to KLD can be calculated by the following formula:

[0075]

[0076] Through the above method, the contribution of each variable to KLD can be analyzed, thereby locating the fault location more accurately.

[0077] Attached Figure 5 The comparison results of fault diagnosis between the proposed method and the traditional PCA method are shown. Figure 1 ), 3), and 5) represent the fault variable identification results based on the variational autoencoder fault graph method. Figure 2 ), 4), and 6) represent the fault identification results based on PCA.

[0078] It can be seen that the T2 statistic in PCA has a good effect on identifying oscillation type faults, but cannot identify step and drift type faults. The SPE statistic has poor recognition effects on the three types of faults. The non-fault variables are highly contaminated by the fault variables and cannot accurately identify the location of the fault. In contrast, the monitoring method based on variational autoencoders can better identify the locations of oscillation, step, and drift faults. Based on the experimental results, it can be preliminarily concluded that the monitoring method based on variational autoencoders is superior to the traditional monitoring method based on PCA.

[0079] In order to compare the effects of the two monitoring methods more deeply, it is necessary to analyze the monitoring results. In process monitoring, the false alarm rate (FAR) and the abnormal detection rate (ADR) are the main indicators for evaluating the effectiveness of monitoring methods, collectively referred to as monitoring performance indicators.

[0080] The calculation method of ADR and FAR is as follows:

[0081]

[0082]

[0083] Where N Ta is the number of true anomaly samples detected, N Fa is the number of false anomaly samples detected, N a is the number of abnormal samples in the data set, N n is the number of normal samples in the dataset.

[0084] Still using the above collected data set, the false alarm rate of the two monitoring methods in normal operation and the fault detection rate of the two monitoring methods when faults 1-3 occur are calculated respectively. The results are shown in the attached figure. Figure 6 shown.

[0085] From the attached Figure 6 It can be seen that when processing normal data, the FAR of both monitoring methods remains at a very low level and will not affect the normal production. However, when processing different types of fault data, the monitoring method based on the variational autoencoder model performs well. For three types of faults, such as oscillation, step, and slow drift, the ADR remains above 85%, showing a high level of performance. In contrast, when facing oscillation-type faults, the ADR of the PCA-based method is less than 60%, and it cannot effectively complete the monitoring task.

[0086] Therefore, it can be concluded that the variational autoencoder has greater advantages than PCA in both process monitoring and fault diagnosis, and can effectively complete the corresponding monitoring and diagnosis tasks.

Claims

1. A method for monitoring and fault diagnosis of continuous production process of drug particles based on variational autoencoder and contribution graph method, which is characterized by comprising the following steps: Step 1: Collection and preprocessing of historical data: The established rotary agitator and twin-screw extruder wet granulation model was used to simulate the actual production process. The key raw material properties and operating variables during normal operation were collected, and the collected historical production data of the granulation process were standardized as a training set. Step 2: Build a variational autoencoder model: Step 3: Train the variational autoencoder model established in step 2: During the training process, the optimization steps of the model include forward propagation, loss calculation, and parameter update. During forward propagation, the input data is passed through the encoder to generate the mean and log variance, the latent vector is obtained through the reparameterization technique, and then the decoder is used to generate the reconstructed data. The loss function consists of reconstruction loss and KL divergence loss. The former measures the difference between the reconstructed data and the original input data, and the latter measures the difference between the latent space distribution and the standard normal distribution. Finally, back propagation is performed based on the total loss to update the parameters of the encoder and decoder. Step 4: Design monitoring statistics and calculate control limits: The distribution generated by the data during normal operation in the latent space is regarded as the benchmark distribution. The KL divergence between the distribution of the monitoring data in the latent space and the benchmark distribution is regarded as a monitoring statistic, which is used to measure the degree to which the monitoring data deviates from normal working conditions, denoted as KLD. In the process of calculating the control limit, firstly, the K groups of data during normal operation are input into the variational autoencoder model, and the KLD between them and the benchmark distribution is calculated and recorded. Secondly, the kernel density estimation method is used to estimate the probability density function of this group of KLD, and the control limit is calculated at a confidence level of 0.

98. Step 5: Conduct production process monitoring. Step 6: Calculate the contribution of each variable to the monitoring statistics to achieve fault diagnosis. In order to calculate the impact of each variable on the monitoring statistic KLD, the data is processed as follows: In the formula, X D(i) It means that the ith variable in the abnormal data is used to replace the ith variable in the normal historical data. p means that the production process is sampled p times. Indicates the nth data corresponding to the i-th variable in the normal data set, is the nth data corresponding to the pth variable in the abnormal data set. After replacing the I group X D(i) They are respectively sent to the variational autoencoder monitoring model to calculate the KL divergence between them and the reference distribution, denoted as KLD (i) . Therefore, the contribution R of each variable to KLD can be calculated by the following formula: The above method can be used to analyze the influence of each variable on KLD, thereby locating the fault location more accurately.

2. The drug particle quality detection method based on variational autoencoder and contribution graph method according to claim 1 is characterized in that: The encoder module consists of three fully connected layers, with an input dimension of 9. After being processed by the ReLU activation function, the output dimensions are 128 and 64 respectively, and finally the mean and logarithmic variance of the latent space are output; the latent vector z is generated by the reparameterization technique. The decoder module also consists of three fully connected layers, with an input dimension of 3. After being processed by the ReLU activation function, the output dimensions are 64 and 128 respectively, and the final output dimension is 9. The Sigmoid activation function is used to decode the latent space vector into a high-dimensional input data form and reconstruct the original input data.

3. The method according to claim 1, characterized in that The key attributes and operating variables described in step 1 include the mass flow rate of raw materials at the agitator inlet, the mass flow rate of excipients, the agitator speed, the mass flow rate of the agitator outlet, the mass flow rate of the binder, the screw speed of the twin-screw extruder, the average residence time of the particles in the equipment, the particle size and the water content of the particles.

4. The method according to claim 1, characterized in that In step 6, the i-th variable in the abnormal data is replaced by the i-th variable in the normal historical data, and the KL divergence between the replaced data and the benchmark distribution is calculated to evaluate the impact of each variable on the monitoring statistic KLD.