Scale-Free Soft Sensor for Monitoring of Water Content in Fluid Bed Granulation Process

Abstract

In-line monitoring of granule water content during fluid bed granulation is important to control drug product qualities. In this study, a practical scale-free soft sensor to predict water content was proposed to cope with the manufacturing scale changes in drug product development. The proposed method exploits two key ideas to construct a scale-free soft sensor. First, to accommodate the changes in the manufacturing scale, the process parameters (PPs) that are critical to water content at different manufacturing scales were selected as input variables. Second, to construct an accurate statistical model, locally weighted partial least squares regression (LW-PLSR), which can cope with collinearity and nonlinearity, was utilized. The soft sensor was developed using both laboratory (approx. 4 kg) data and pilot (approx. 25 kg) scale data, and the prediction accuracy in the commercial (approx. 100 kg) scale was evaluated based on the assumption that the process was scaled-up from the pilot scale to the commercial scale. The developed soft sensor exhibited a high prediction accuracy, which was equivalent to the commonly used near-infrared (NIR) spectra-based method. The proposed method requires only standard instruments; therefore, it is expected to be a cost-effective alternative to the NIR spectra-based method.

Introduction

Fluid bed granulation has been widely applied to the manufacturing of pharmaceutical solid oral dosage forms such as tablets to improve drug product quality and manufacturability. In a top-spray granulation process, a binder solution is sprayed onto the particles fluidized by the heated inlet air from the bottom of the granulator. In-line monitoring of the granule water content profile during granulation is important because it correlates the granule properties and can affect the tablet properties such as hardness and disintegration time finally.¹⁾

Mass balance and heat transfer models have been traditionally used for predicting the water content of the granules.^2–4) These models typically use not only the process parameters (PPs) monitored by the standard instruments, such as inlet air humidity, spray rate, and exhaust air temperature, but also several unknown parameters, which need to be tuned by using experimental data, such as the heat transfer coefficient,²⁾ particle wetting surface ratio,³⁾ and the ability of powder particles to retain moisture.⁴⁾ Mass balance and heat transfer models often have a challenge in determining the values of unknown parameters, which can deteriorate the prediction accuracy. The values of unknown parameters can differ depending on various factors such as the manufacturing condition, the scale of a fluid bed granulator, and the granule formulation. Hence, to construct accurate and robust models, it is necessary to calibrate the models by using the experimental data acquired with various manufacturing conditions and manufacturing scales, and especially to tune the unknown parameters accurately, which is practically challenging.

Soft sensors,⁵⁾ which are statistical models constructed by using the standard PPs, have been utilized to predict the water content of granules. Rambali et al.⁶⁾ demonstrated an optimization of the fluid bed granulation process based on the multiple regression model. In addition, other regression models such as partial least squares regression (PLSR)⁷⁾ that can cope with collinearity among PPs have been used to increase the prediction accuracy and the robustness of the statistical model.^8,9) However, the soft sensors, as well as the mass balance and heat transfer models, are valid only within the process conditions considered during model development. For example, when the manufacturing scale is changed, the prediction accuracy of the model might decrease because the relationship between the majority of PPs and the water content is generally dependent on the manufacturing scale. Similarly, the existing soft sensors used to predict the particle size of granules were valid only for the same granulator for which experimental data was obtained.^9,10) Although the manufacturing scale of the granulation process can be adjusted according to the demand during clinical development and commercial manufacturing, a soft sensor that enables scale-free monitoring of the material attributes of granules during granulation such as water content has not been developed yet.

This study aims to develop a novel robust soft sensor that can counter changes in the manufacturing scale. To develop a scale-free soft sensor for monitoring water content, the effects of the input variables and a regression method considering the accuracy and robustness of the prediction model were evaluated. There are two key ideas for constructing a scale-free soft sensor. First, to accommodate the changes in the manufacturing scale, the input variables were selected based on variable importance in the PLSR model constructed using the experimental data obtained at different manufacturing scales. Second, to construct an accurate statistical model, locally weighted partial least squares regression (LW-PLSR)¹¹⁾ was utilized. The LW-PLSR is a type of just-in-time modeling method that can cope with collinearity and nonlinearity. Recently, LW-PLSR has been applied to various industrial processes such as pharmaceutical,^11–13) petrochemical,^14,15) and semiconductor production.^15,16) The LW-PLSR was adopted since the relationship between basic PPs and water content could be nonlinear. In this study, soft sensors were developed using both laboratory (approx. 4 kg) data and pilot (approx. 25 kg) scale data; the prediction accuracy in the commercial (approx. 100 kg) scale was evaluated based on the assumption that the process was scaled-up from the pilot scale to the commercial scale. A direct, near-infrared (NIR) spectra-based prediction model^17,18) was also constructed as a reference. While this approach requires a high initial investment for equipment such as the NIR spectrometer, probe, spectroscopy software, and their incorporation into the fluid bed granulation process, the NIR spectra-based monitoring methods are generally robust enough to counter changes in the manufacturing scale. Both the soft sensor and the NIR spectra-based models were constructed using a dataset obtained from the same experiments, and their prediction accuracies were compared.

Experimental

Materials

A formulation composed of the drug substance (Daiichi Sankyo Co., Ltd., Japan) and several excipients were granulated using the following three different fluid bed granulators: NFLO-5 (Freund Corp., Japan) for the approx. 4-kg scale, GPCG-30 (Powrex Corp., Japan) for the approx. 25-kg scale, and WSG-120 (Powrex Corp.) for the approx. 100-kg scale. In this study, the manufacturing scales were defined as follows: approx. 4 kg is the laboratory scale, approx. 25 kg is the pilot scale, and approx. 100 kg is the commercial scale. A schematic diagram of the fluid bed granulator is provided in Fig. 1.

Fig. 1. Schematic Diagram of a Fluid Bed Granulator

Granules were sampled during the granulation process at the time points provided in Table 1. In this study, as provided in Table 1, the calibration and validation datasets consist of only spraying process data because the number of PPs in the spraying process is different from that in the drying process; spray rate (g/min) and spray air volume (NL/min) are PPs specific to the spraying process. The spraying process, where liquid supply and evaporation occur simultaneously, is complex and difficult to control water content accurately compared to the drying process. Besides, the scale-up is generally performed based on the water content profile during the spraying process. It is desirable that PPs are determined to keep the granule water content during the spraying process constant at different manufacturing scales.¹⁹⁾ Thus, in-line monitoring of water content in the spraying process is especially important, and the present research focused on the spraying process.

Table 1. Calibration and Validation Datasets

Dataset	Equipment	Batch No.	Time	PP1	PP2	PP3	PP4	PP5	PP6	PP7	PP8	WC
Calibration	NFLO-5	1	0	2.6	76.3	10.0	50	149	41.3	39.7	20	2.2
Calibration	NFLO-5	1	5	2.5	82.8	10.0	52	150	39.2	38.3	40	2.9
Calibration	NFLO-5	1	10	2.5	84.7	10.1	55	150	36.7	36.6	48	3.5
Calibration	NFLO-5	1	15	2.0	84.7	10.0	53	150	35.1	35.5	50	4.3
Calibration	NFLO-5	1	20	2.5	84.1	9.3	69	149	35.8	34.3	55	3.5
Calibration	NFLO-5	1	25	2.5	84.8	9.5	53	149	36.2	34.1	56	4.0
Calibration	NFLO-5	1	30	2.5	85.1	9.7	49	148	36.7	34.1	58	3.8
Calibration	NFLO-5	2	0	2.6	67.4	10.0	70	147	42.0	42.6	16	1.9
Calibration	NFLO-5	2	15	2.0	70.6	10.1	81	148	29.3	31.6	65	9.5
Calibration	NFLO-5	2	20	1.5	70.1	10.2	81	149	28.2	30.3	68	12.5
Calibration	NFLO-5	2	25	1.5	70.1	10.1	71	148	27.7	29.4	69	15.6
Calibration	NFLO-5	3	5	2.5	69.3	9.2	66	147	35.5	41.6	34	3.7
Calibration	NFLO-5	3	10	2.5	69.9	9.5	65	147	30.6	36.8	46	4.7
Calibration	NFLO-5	3	15	2.5	70.0	9.8	67	147	29.2	34.1	53	7.9
Calibration	NFLO-5	3	20	2.5	69.6	8.8	69	148	28.5	30.4	70	7.7
Calibration	NFLO-5	3	25	2.6	69.9	9.4	67	149	28.4	29.6	73	9.2
Calibration	NFLO-5	3	29	2.5	70.0	9.6	66	148	28.4	29.1	75	10.2
Calibration	GPCG-30	4	10	8.0	87.0	5.1	180	178	41.2	41.8	34	3.1
Calibration	GPCG-30	4	20	7.9	87.9	5.2	180	175	38.0	38.9	44	3.8
Calibration	GPCG-30	4	30	8.0	88.0	5.3	179	179	36.5	37.3	51	4.1
Calibration	GPCG-30	4	40	7.9	87.9	5.4	180	181	35.5	36.1	57	4.5
Calibration	GPCG-30	4	50	7.9	87.6	5.5	180	182	34.9	35.2	61	4.4
Calibration	GPCG-30	4	59	7.9	87.8	5.5	180	181	34.6	34.6	65	4.8
Calibration	GPCG-30	5	10	7.9	87.5	6.5	151	148	44.8	42.9	33	2.8
Calibration	GPCG-30	5	20	8.0	88.7	6.6	150	150	43.0	41.3	38	3.0
Calibration	GPCG-30	5	30	7.9	87.9	6.7	149	148	42.1	40.4	41	3.2
Calibration	GPCG-30	5	40	8.0	88.7	6.3	149	151	41.6	40.0	43	3.3
Calibration	GPCG-30	5	50	7.9	88.0	6.7	149	150	41.0	39.6	44	3.5
Calibration	GPCG-30	5	60	7.9	88.0	6.5	150	148	40.7	39.3	46	3.4
Calibration	GPCG-30	5	72	7.9	87.9	6.7	149	151	40.3	38.8	47	3.5
Validation	WSG-120	1	10	30.0	89.9	2.3	699	715	41.3	40.1	51	3.6
Validation	WSG-120	1	20	30.1	90.0	2.2	699	692	38.7	38.1	54	4.0
Validation	WSG-120	1	30	29.9	90.1	2.4	698	698	37.3	36.9	59	4.3
Validation	WSG-120	1	40	30.0	90.0	2.3	699	698	36.5	36.1	64	4.7
Validation	WSG-120	1	50	30.0	90.1	2.2	698	700	35.9	35.5	67	4.9
Validation	WSG-120	1	62	30.0	90.0	2.3	697	699	35.5	35.1	69	5.0
Validation	WSG-120	2	10	27.4	85.0	4.0	848	650	35.7	37.5	45	4.3
Validation	WSG-120	2	20	28.1	85.0	3.8	849	643	32.2	33.9	59	6.2
Validation	WSG-120	2	30	28.1	85.0	3.8	849	649	30.9	32.2	68	7.6
Validation	WSG-120	2	40	28.1	85.0	3.9	848	646	30.3	31.2	74	9.3
Validation	WSG-120	2	51	28.0	85.0	3.9	848	647	30.1	30.8	78	10.8
Validation	WSG-120	3	10	31.9	94.9	3.8	548	730	49.8	47.0	26	2.6
Validation	WSG-120	3	20	32.1	95.1	3.8	550	712	50.3	48.4	23	2.6
Validation	WSG-120	3	30	31.5	95.1	4.1	548	746	50.8	49.2	22	2.8
Validation	WSG-120	3	40	31.9	95.1	3.8	549	743	51.0	49.5	21	2.6
Validation	WSG-120	3	50	32.0	95.0	3.9	549	742	50.9	49.6	21	2.8
Validation	WSG-120	3	60	31.2	95.0	4.1	549	743	51.0	49.7	21	2.7
Validation	WSG-120	3	70	31.8	95.0	3.9	549	741	51.0	49.7	21	2.6
Validation	WSG-120	3	79	31.6	95.1	3.8	548	741	51.1	49.7	20	2.5

Time: spraying time (min), PP1: inlet air volume (m³/min), PP2: inlet air temperature (°C), PP3: inlet air humidity (g-water/kg-air), PP4: spray rate (g/min), PP5: spray air volume (NL/min), PP6: product temperature (°C), PP7: exhaust air temperature (°C), PP8: exhaust air humidity (%RH), and WC: water content of granules (%).

In addition, in all process conditions investigated in this study, it was confirmed that granule growth was proceeding during the spraying process according to the particle size of granules after drying. The particle size of granule after drying obtained through the experiment of WSG-120 Batch No.3, which is the driest process condition, was 95.8 µm and larger than that of the mixture with drug substance and several excipients, i.e., 36.6 µm.

Measurement

The measured values of PPs, NIR spectra of granules, and water content of granules, which were measured at the time points provided in Table 1, were utilized for calibration or validation.

Process Parameters (PPs)

The following eight PPs were measured during the granulation process: inlet air volume (m³/min), inlet air temperature (°C), inlet air humidity (g-water/kg-air), spray rate (g/min), spray air volume (NL/min), product temperature (°C), exhaust air temperature (°C), and exhaust air humidity (%RH). Among the eight PPs, inlet air volume (m³/min), inlet air temperature (°C), spray rate (g/min), and spray air volume (NL/min) are operable PPs. All the PPs were acquired using standard instruments such as a thermometer, hygrometer, flowmeter, and electric balance.

NIR Spectra

The NIR spectra of the granules were measured during the granulation process using a Fourier-transform NIR spectrometer MPA (Bruker Optik GmbH, Germany) or Matrix-F (Bruker Optik GmbH) through a fiber-optic probe attached to the fluid bed granulators. Because the MPA and Matrix-F use an equivalent optical device, the NIR spectra obtained are comparable.

Reference

The water content of the granules sampled during the granulation process was measured by the following loss on drying (LOD) device: HR73 (Mettler-Toledo K.K., Japan) or its equivalent HR83 (Mettler-Toledo K.K.).

Calibration Dataset

In this study, a scenario in which the granulation process was scaled-up from a pilot scale to a commercial scale was investigated. As shown in Table 1, the calibration dataset for both the soft sensors and NIR spectra-based models consists of 30 samples, which were acquired at the laboratory and pilot scale: NFLO-5 (3 lots) and GPCG-30 (2 lots).

Preprocessing

The PPs were normalized, that is, their means were equal to 0 and the variances were equal to 1. For the NIR spectra, the wavenumber region ranging from 8007 to 4258 cm⁻¹ was used to exclude any noisy regions, and the prediction models were constructed using the following five preprocessing methods: first derivative, second derivative, standard normal variate (SNV),²⁰⁾ first derivative and SNV, and second derivative and SNV. The method presenting the highest prediction accuracy was then selected. Each preprocessing method was performed using the OPUS software (Bruker Optik GmbH).

Input Variable Selection

Process Parameters (PPs)

To predict the water content at a commercial scale without using the commercial scale data in the calibration dataset, critical PPs regardless of the manufacturing scales were selected as follows. First, a PLSR model was constructed using all the PPs provided in the calibration dataset. Then, variable importance in the projection (VIP) scores, which indicate contributions of the input variables to the prediction value, were calculated. The VIP score for the j-th variable is defined as follows:

  

(1)

where M is the number of input variables, R is the number of latent variables (LVs), q_r is the r-th regression coefficient vector, t_r is the r-th score vector, and w_jr is the j-th element of the r-th weighting vector w_r. Here, R was determined to minimize the prediction error sums of squares of the leave-one-out cross validation (LOOCV). In general, the input variables whose VIP scores are greater than or equal to one are considered critical to the prediction value.²¹⁾ In this study, PPs with VIP scores greater than or equal to one were selected as the input variables.

NIR Spectra

The spectral fluctuation dividing (SFD)-nearest correlation spectral clustering (NCSC)-PLSR²²⁾ was adopted for wavenumber selection using the Python software (Python Software Foundation). SFD-NCSC-PLSR was reported to achieve a higher prediction accuracy than commonly used variable selection methods such as VIP, PLSR-beta, and interval PLSR.

Regression Methods

PLSR and LW-PLSR were used to construct the prediction model using the selected input variables. To evaluate the effect of the input variable selection on the prediction performance, a PLSR model using all the PPs was constructed and evaluated. The main difference between PLSR and LW-PLSR is the weighting rule for the calibration samples. In PLSR, a prediction model is constructed using fixed weighting values for the calibration samples. In contrast, in LW-PLSR, which is a type of just-in-time modeling method, the weighting values are updated according to the distances between a query and the calibration samples whenever a prediction is required for a query, and a local PLSR model is constructed. That is, LW-PLSR includes PLSR as a special case.¹¹⁾

The PLSR and LW-PLSR models were constructed using MATLAB^® software (MathWorks, Inc., U.S.A.). LOOCV was performed to determine the tuning parameters of PLSR and LW-PLSR: the number of LVs for PLSR and the combination of the number of LVs and the localization parameter for LW-PLSR. The root mean square error of cross validation (RMSECV) and the correlation coefficient (R_CV) were calculated to evaluate the prediction accuracy as follows:

  

(2)

where N is the number of samples, and ŷ_n and y_n are the prediction value and reference value of the water content for the n-th sample, respectively;

  

(3)

where ŷ̅ a̅nd y̅ are the mean values of ŷ̅_n and y_n, respectively.

Validation

As shown in Table 1, the validation dataset consists of 19 samples acquired at the commercial scale: WSG-120 (3 lots), which is not included in the calibration dataset. The validation dataset was used to evaluate the prediction accuracy of the constructed models based on the root mean square error of prediction (RMSEP) and R_P.

  

(4)

  

(5)

To assess the validity of the LW-PLSR model as the soft sensor, two distance criteria were used: Hotelling’s T² and square prediction error (SPE), also known as Q residual.²³⁾

  

(6)

  

(7)

where t_r, query is the r-th LV score of the query, t̅_{r, calibration} is the mean value of t_{r, calibration}, which is the r-th score vector of the calibration dataset samples, s²_{tr, calibration} is the variance of t_{r, calibration}, x_m, query and x̂_m, query are the query’s experimental value and reconstructed value of the m-th input variable.

Muteki et al.²⁴⁾ demonstrated that two distance criteria, T² and Q, could test the validity of the PLSR model for the query. In this study, a 99% confidence limit was adopted as the threshold of T² and Q to test whether the LW-PLSR model was valid for the query.

Results

Soft Sensors

Input Variable Selection

Figure 2 shows the VIP scores of all the PPs, which were calculated using the PLSR model after normalization. The results show that the VIP scores of four PPs, i.e., inlet air temperature (°C), product temperature (°C), exhaust air temperature (°C), and exhaust air humidity (%RH), were greater than one. Therefore, these four PPs were selected as the input variables.

Fig. 2. Importance of PPs Based on VIP Scores

All PPs were normalized before PLSR modeling.

RMSEPs of the PLSR models are provided in Figs. 3A and 3B. RMSEP was significantly reduced from 27.2 to 3.5% by using the four selected PPs as input variables. In addition, RMSECVs are provided in Figs. 3C and 3D. The PLSR model with the four selected PPs showed a notably smaller difference between RMSEP and RMSECV than that with all PPs; their differences were 1.7 and 25.6%, respectively. These results confirm that the proposed method constructed a robust soft sensor to counter the manufacturing scale change.

Fig. 3. Prediction Accuracy of PLSR Models Using the Four Selected PPs (A, C) and All PPs (B, D)

(A, B) RMSEPs and R_P values, (C, D) RMSECVs and R_CV values.

Regression Method

LW-PLSR and PLSR were compared in terms of prediction accuracy. The four selected PPs were used as input variables in both models; RMSEPs were calculated using the validation dataset shown in Table 1. As shown in Figs. 3A and 4A, the RMSEP of 0.5% for the LW-PLSR model was significantly smaller than that of 3.5% for the PLSR model. In addition, the difference between the RMSEP and RMSECV of −0.2% for the LW-PLSR model was smaller than that of 1.7% for the PLSR model as shown in Figs. 3 and 4. These results indicate that the LW-PLSR model is more accurate and robust than the PLSR model. The LW-PLSR model with the four selected PPs was confirmed to show the highest prediction accuracy among the three soft sensors.

Fig. 4. Prediction Accuracy of LW-PLSR Model Using the Four Selected PPs

(A) RMSEP and R_P value, (B) RMSECV and R_CV value.

The number of LVs in the LW-PLSR model is the same as that of the input variables. This indicates that dimension reduction was not performed, and Q was constant at 0. In addition, Hotelling’s T² values of all validation dataset samples were lower than the 99% confidence limit.

NIR Spectra-Based Models

Preprocessing

Table 2 shows the RMSECVs and R_CV values of the five PLSR models constructed by applying various preprocessing methods. As a result, SNV, which showed the highest prediction accuracy, was selected as the preprocessing method for constructing the prediction model using the NIR spectra.

Table 2. Comparison of Five Preprocessing Methods in RMSECVs and R_CV Values

Preprocessing method	RMSECV (%)	RCV
First derivative	0.6	0.98
Second derivative	0.7	0.97
SNV	0.5	0.99
First derivative + SNV	0.7	0.98
Second derivative + SNV	0.8	0.97

Wavenumber Selection

Figure 5 shows the wavenumber region selected by SFD-NCSC-PLSR, i.e., ranging from 7212 to 6935 cm⁻¹, which includes the first overtone wavenumber region for H₂O.

Fig. 5. NIR Spectra, Preprocessed Using SNV, of Granules during Fluid Bed Granulation at the Wavenumber Region Ranging from 8007 to 4258 cm⁻¹

Regression Method

Two models based on LW-PLSR and PLSR were constructed after applying SNV to the calibration set of the NIR spectra with a wavenumber region ranging from 7212 to 6935 cm⁻¹. The R_P values for both models were evaluated using the validation dataset shown in Table 1. The R_P value of 0.98 for the LW-PLSR model was higher than that of 0.93 for the PLSR model, as observed from the results shown in Figs. 6A and 6B. Moreover, Fig. 6 shows the 0.5% difference between the RMSEP and RMSECV for the LW-PLSR model, which was equivalent to that of 0.6% for the PLSR model. Thus, the LW-PLSR model utilizing the NIR spectra was confirmed to be the better estimator of the water content compared to the PLSR model.

Fig. 6. Prediction Accuracy of LW-PLSR (A, C) and PLSR (B, D) Models Using Absorbance at the Wavelength Selected by SFD-NCSC-PLSR

(A, B) RMSEPs and R_P values, (C, D) RMSECVs and R_CV values.

Comparison of the Prediction Accuracy

The best soft sensor being the LW-PLSR model with the four PPs, and the best NIR spectra-based model being the LW-PLSR model with a wavenumber region from 7212 to 6935 cm⁻¹ were compared. As shown in Figs. 4A and 6A, the 0.5% RMSEP for the soft sensor was slightly smaller than that of 1.0% for the NIR spectra-based model. In addition, Figs. 4A, 4B, 6A, and 6C show the −0.2% difference between the RMSEP and RMSECV for the soft sensor, which was equivalent to that of 0.5% for the NIR spectra-based model. These results indicate that the soft sensor was as accurate as the NIR spectra-based model and robust enough to counter manufacturing scale changes.

Discussion

In this study, soft sensors were developed using both laboratory (approx. 4 kg) data and pilot (approx. 25 kg) scale data; the prediction accuracy on a commercial scale (approx. 100 kg) was evaluated on the assumption that the process was scaled-up from the pilot scale to the commercial scale. The developed soft sensor achieved a high prediction accuracy equivalent to that of the NIR spectra-based model and was robust enough to counter the manufacturing scale change, only when an appropriate selection of the input variables and the regression method selection were carried out.

As shown in Fig. 2, the inlet air temperature (°C), product temperature (°C), exhaust air temperature (°C), and exhaust air humidity (%RH) were selected as the input variables of the soft sensor because the VIP scores were larger than 1.0. The soft sensor utilizing the four PPs showed higher prediction accuracy compared to the models using different PPs selected by the various VIP criteria, as shown in Fig. 7. The difference between the RMSEP and RMSECV became remarkably smaller when the VIP criterion was 1.0 or greater. Furthermore, the RMSEP and RMSECV of the model, which used the four PPs selected according to the VIP criterion of 1.0, were smaller than those of the model constructed using the PPs selected at the VIP criterion of 1.2. Based on the results, the four selected PPs were considered critical for constructing a robust soft sensor to counter manufacturing scale changes. As shown in Fig. 8, the four selected PPs were in similar ranges regardless of the manufacturing scale, while the parameter ranges of the other four PPs, i.e., inlet air volume (m³/min), inlet air humidity (g-water/kg-air), spray rate (g/min), and spray air volume (NL/min), were significantly different depending on the manufacturing scale. Reflecting the ranges of the four PPs used in the soft sensor, the values of Hotelling’s T² and Q of all validation dataset samples were within the 99% confidence limit, which contributed to the validity of the LW-PLSR models for the query. The four selected PPs include one operable PP, i.e., inlet air temperature (°C). Thus, constructed soft sensors are expected to be useful for not only in-line monitoring but also water content control by adjusting inlet air temperature.

Fig. 7. Difference between RMSEPs and RMSECVs of LW-PLSR Models Depending on VIP Score Criteria

Fig. 8. Ranges of Input Variables at Three Fluid Bed Granulators: NFLO-5, GPCG-30, and WSG-120

Furthermore, LW-PLSR was effective in improving the prediction accuracy because the relationship between the input and output variables was nonlinear. As shown in Fig. 9, the four selected input variables represent the nonlinearity of the output variable of the water content. In LW-PLSR, a local PLSR model is constructed per a query using a weighted calibration dataset based on the distances between the query and the calibration samples. Because of its applicability to the nonlinearity in the dataset, the LW-PLSR model achieved a better prediction performance and robustness compared to the PLSR model.

Fig. 9. Relationship between Input Variables and an Output Variable (i.e., Water Content) in the Calibration Dataset

Conclusion

In this study, we demonstrated that the proposed soft sensor could predict water content accurately beyond the manufacturing scale range of the calibration dataset. There are two key ideas for constructing a scale-free soft sensor. First, to accommodate the changes in the manufacturing scale, PPs that are critical to water content at different manufacturing scales were used as the input variables. Second, to construct an accurate statistical model, LW-PLSR was utilized to cope with collinearity and nonlinearity. The experiments demonstrated that the soft sensor based on the proposed method exhibited a high prediction accuracy, which was equivalent to the commonly used NIR spectra-based method. Unlike the NIR-based method, the proposed method requires only standard instruments. Therefore, the proposed method is expected to be a cost-effective alternative to the existing NIR spectra-based monitoring method and has the ability to enhance the implementation of in-line monitoring of the fluid bed granulation processes.

Conflict of Interest

The authors declare no conflict of interest.

References

• 1) Hartung A., Knoell M., Schmidt U., Langguth P., Drug Dev. Ind. Pharm., 37, 274–280 (2011).
• 2) Watano S., Fukushima T., Miyanami K., Chem. Pharm. Bull., 44, 572–576 (1996).
• 3) Hu X., Cunningham J., Winstead D., J. Pharm. Sci., 97, 1564–1577 (2008).
• 4) Pla D. L., Kamyar R., Hashemian N., Mehdizadeh H., Moshgbar M., Powder Technol., 326, 69–77 (2018).
• 5) Kadlec P., Gabrys B., Strandt S., Comput. Chem. Eng., 33, 795–814 (2009).
• 6) Rambali B., Baert L., Massart D. L., Int. J. Pharm., 220, 149–160 (2001).
• 7) Wold S., Sjöström M., Eriksson L., Chemom. Intell. Lab. Syst., 58, 109–130 (2001).
• 8) Lipsanen T., Närvänen T., Räikkönen H., Antikainen O., Yliruusi J., AAPS PharmSciTech, 9, 1070–1077 (2008).
• 9) Närvänen T., Antikainen O., Yliruusi J., AAPS PharmSciTech, 10, 1268–1275 (2009).
• 10) Yoshizaki R., Kano M., Tanabe S., Miyano T., IFAC-PapersOnLine, 48, 303–308 (2015).
• 11) Kim S., Kano M., Nakagawa H., Hasebe S., Int. J. Pharm., 421, 269–274 (2011).
• 12) Nakagawa H., Tajima T., Kano M., Kim S., Hasebe S., Suzuki T., Nakagami H., Anal. Chem., 84, 3820–3826 (2012).
• 13) Nakagawa H., Kano M., Hasebe S., Miyano T., Watanabe T., Wakiyama N., Int. J. Pharm., 471, 264–275 (2014).
• 14) Kim S., Okajima R., Kano M., Hasebe S., Chemom. Intell. Lab. Syst., 124, 43–49 (2013).
• 15) Kano M., Fujiwara K., J. Chem. Eng. Japan, 46, 1–17 (2013).
• 16) Hirai T., Kano M., IEEE Trans. Semicond. Manuf., 28, 137–144 (2015).
• 17) Frake P., Greenhalgh D., Grierson S. M., Hempenstall J. M., Rudd D. R., Int. J. Pharm., 151, 75–80 (1997).
• 18) Rantanen J., Antikainen O., Mannermaa J. P., Yliruusi J., Pharm. Dev. Technol., 5, 209–217 (2000).
• 19) Rambali B., Baert L., Massart D. L., Int. J. Pharm., 252, 197–206 (2003).
• 20) Barnes R. J., Dhanoa M. S., Lister S. J., Appl. Spectrosc., 43, 772–777 (1989).
• 21) Chong I. G., Jun C. H., Chemom. Intell. Lab. Syst., 78, 103–112 (2005).
• 22) Miyano T., Fujiwara K., Kano M., Tanabe H., Nakagawa H., Watanabe T., Minami H., Chemom. Intell. Lab. Syst., 148, 85–94 (2015).
• 23) Jackson J. E., Mudholkar G. S., Technometrics, 21, 341–349 (1979).
• 24) Muteki K., Yamamoto K., Reid G. L., Krishnan M., J. Pharm. Innov., 6, 142–156 (2011).