Intelligent information-measuring system of drum drying unit

Full Text

Introduction

Drying is a widespread process in various industries. Drying units belong to the energy-intensive technological devices.

In the drum drying units (DDU), a convective drying process is realized. This type of equipment is used for drying fine bulk and wet materials. The drum drying unit consists of a cylinder installed at an inclination of about 3–6° relative to the horizontal plane, with two belts, with the rotation of the shaft, which slides on support rollers. The distance between the rollers can be adjusted to change the angle of inclination of the shaft. The shaft rotates thanks to a tight fit on the housing, the gears are connected to the motor through a gearbox. In this system, the drying material is actively mixed in contact with the drying agent. Significant heat exchange is carried out, there is a fast drying speed and a high uniformity of the product.

The drum drying unit operates with high productivity at atmospheric pressure and is well suited for technology of processing during the production of feeding alcoholic stillage. The drying in the drum drying unit is characterized by high stability, the quality of the product being dried and meets the technical requirements [1].

An important indicator of the drying process is the quality of the stillage being dried. Monitoring of moisture content in the mode of the real time drying process and making timely decisions on the management of this process has a significant impact not only on the quality of the material being dried, but also on the energy performance of the entire production in whole. The development and implementation of an intelligent information-measuring system (IMS) for the DDU helps to reduce the loss of quality of drying of the distiller’s grain.

Method of indirect measurement of relative humidity of the stillage in the process of its drying in the DDU

The object for which the information-measuring system is realized is a drum shell-and-tube drying plant manufactured by VetterTec GmbH. Generalized block diagram of the DDU is presented in Fig. 1.

 

Fig. 1. Generalized block diagram of the drum type drying unit: 1 – housing; 2 – drum; 3 – steam supply; 4 – dry material outlet; 5 – steam outlet; 6 – blades; 7 – supply of material for drying; 8 – air-inlet windows; 9 – vapor-drain

 

The composition of the DDU includes a fixed housing with a rotating drum located inside. Inside the drum there are two head parts of two pipe slabs and pipes flared in them. The drum with the material is heated by the steam supplied from the factory boiler room. In the tubes inside the drum, the steam moves in the direction opposite to the direction of material movement. At the same time, the quantity, the length and the diameter determine the heating power of the drying unit. At a pressure of 5 – 7 atmospheres in the system, the steam temperature ranges from 160 to 180 °C. The amount of steam supplied is regulated by automatic sealing steam heads.

The material for the drying in the DDU is the distiller’s grain, which is a product left from the raw material after distillation for the producing of the alcohol or the beer. It is a kind of sediment with a strong odor and often cloudy color, as well as impurities.

The stillage can be used to feed the livestock. The mass fraction of distiller’s dried grains is usually (96 ± 1), (93 ± 2), (90 ± 1) %. If the temperature of the material is higher than 120 degrees, then it begins to burn. This adversely affects the quality and the price of the finished product. When the thin stillage is supplied to the inlet of the DDU, it must be diluted with already dried stillage on a mixing screw. This provides that there is no accumulation of intermediate products at the plant.

Residual moisture content of distiller’s dried grains is the main target quality indicator. Traditionally, its humidity content is monitored every 2 hours in the conditions of the laboratory. When determining of the humidity, additional time is required, resulting in a time lag. In order to achieve the required quality of the stillage at the output of the drum drying unit, it is required to know its moisture content in order to make timely decision on controlling the drying process. In order to eliminate this disadvantage it is supposed to use the developed IMS for the DDU.

The IMS is a comprehensive systems that possess information and measurement flows. This complex is accompanied by a multitude of measurement channels, which are of the greatest interest. The power of the measurement channel is determined by the magnitude of the arrangements which are subject to control using the available transducers. The functioning of the created IMS is based on the method of indirect measurement of relative humidity of the stillage in the process of its drying in the DDU on the basis of neural networks. The IMS includes the following primary measuring transducers:

• three primary measuring transducers of the temperature in different zones of the DDU;
• one primary measuring pressure transducer for systems with the supply of heat transfer agent in the form of steam directly into the working area of the DDU;
• primary measuring transmitter of oxygen content;
• distribution primary measuring transducer, which is pointed at the separation of thin and dried stillage;
• pulse-width devices located in the production processes control system (PPCS) and responsible for regulating the operation of the exhaust fan;
• transducer, which is also located in the PPCS and provides the determination of the loading level of the electric motor of the drying unit.

For each primary measuring transducer a separate channel is provided for transmission of information flows directly to the information-measuring system. The created system has eight measurement channels.

To estimate the relative humidity of the material in the mode of the real time in the IMS of the DDU models represented in the form of neural networks are used. The drying of the stillage is a nonlinear complex process, which is difficult to accurately express and control using traditional methods. An artificial neural network is not only capable of processing nonlinearities, realizing a self-organizing tuning and adaptive learning to provide a fault-tolerant noise protection, but also can effectively cope with nonlinear and complex fuzzy processes.

The network inputs are the following values: x₁ – initial moisture content of the stillage at the drying unit inlet, % of the dry matter; x₂ – stillage temperature at the dryer outlet, °С; х₃ – heat transfer agent pressure in atmospheres; х₄ – air temperature in the drying unit, °С; x₅ – exhaust fans power, %; х₆ – oxygen content in the drying unit air, %; x₇ – vapor temperature, °С; х₈ – drying unit electric motor load, %.

Each input parameter usually has different physical values and dimensionalities, and each input sample is equally important. Therefore all input variables are presented in the range [0; 1], and the initial range should cover all possible values of input parameters in the normal mode of operation of the system. The normalization of values is done by the formula:

$$\begin{gathered} x_{iN}=\frac{x_i-{\mu }_i}{{\sigma }_i}; \\ f:T\times U\times X\times S\to Y, \end{gathered}$$ (1)

where $x_{iN}$ is the normalized value of the i-th variable, $x_i$ is value of the i-th variable, ${\mu }_i$ is the mathematical expectation of the i-th variable, ${\sigma }_i$ is the variance of the i-th variable.

The neural network training is organized according to the technological parameters reference metrics in the adaptive range of the drying unit with a fixed accuracy.

The analytical model for evaluating and monitoring the moisture content of the material in a drum drying unit is an artificial neural network, which is presented as a multilayer perceptron, trained using an errors back propagation algorithm. It can be written as follows:

$\phi =100-f_4\left(\sum _{h=1}^{m_3}{f}_3\left(\sum _{k=1}^{m_2}{f}_2\left(\sum _{j=1}^{m_1}{f}_1\left(\sum _{i=1}^n{x}_{iN}{W}_{i,j}^{\left(1\right)}+b_j\right)W_{i,k}^{\left(2\right)}+b_k\right)W_{k,h}^{\left(3\right)}+b_h\right)W_h^{}+b^{\text{out}}\right)$, (2)

where φ is the estimate of the material humidity; $x_{iN}$ is the normalized value of the i-th input variable.2; $f_1,f_2,f_3$ are the activation functions – ReLu-function:

$f_1=\max \left(\mathrm{0,}\sum _{i=1}^n{x}_i{W}_{i,j}^{\left(1\right)}+b_j\right)$; (3)

$f_2=\max \left(\mathrm{0,}\sum _{j=1}^{m_1}\max \left(\mathrm{0,}\sum _{i=1}^n{x}_i{W}_{i,j}^{\left(1\right)}+b_j\right)W_{j,k}^{\left(2\right)}+b_k\right)$; (4)

$f_3=\max \left(\mathrm{0,}\sum _{k=1}^{m_2}\max \left(\mathrm{0,}\sum _{j=1}^{m_1}\max \left(\mathrm{0,}\sum _{i=1}^n{x}_i{W}_{i,j}^{\left(1\right)}+b_j\right)W_{j,k}^{\left(2\right)}+b_k\right)W_{k,h}^{\left(3\right)}+b_h\right)$; (5)

$f_4=\max \left(\mathrm{0,}\sum _{h=1}^{m_3}\max \left(\mathrm{0,}\sum _{k=1}^{m_2}\max \left(\mathrm{0,}\sum _{j=1}^{m_1}\max \left(\mathrm{0,}\sum _{i=1}^n{x}_i{W}_{i,j}^{\left(1\right)}+b_j\right)W_{j,k}^{\left(2\right)}+b_k\right)W_{k,h}^{\left(3\right)}+b_h\right)W_h+b^{\text{out}}\right)$, (6)

n is the number of input variables (in our case n = 8), (i = 1, …, n); m₁ is the number of neurons in the 1^st hidden layer (in our case $m_1=36$) (j = 1, … , $m_1=36$); m₂ is the number of neurons in the 2^nd hidden layer (in our case $m_2=36$) (k = 1, … ,$m_2=36$ ); m₃ is the number of neurons in the 3^rd hidden layer (in our case $m_3=36$) (h = 1, …, $m_3=36$); x_i is the value of the i-th input variable; $W_{i,j}^{\left(1\right)},W_{j,k}^{\left(2\right)},W_{k,h}^{\left(3\right)},W_k^{\text{out}}$ are hidden and output weighting coefficients:

$W_{i,j}^{\left(1\right)}=\left(\begin{array}{cccc}{W}_{\mathrm{1,1}}& W_{\mathrm{1,2}}& \cdots & W_{\mathrm{1,}{m}_1}\\ W_{\mathrm{2,1}}& \cdots & \cdots & W_{\mathrm{2,}{m}_1}\\ ⋮& ⋮& ⋮& ⋮\\ W_{n\mathrm{,1}}& \cdots & \cdots & W_{n,m_1}\end{array}\right)$; (7)

$W_{j,k}^{\left(2\right)}=\left(\begin{array}{cccc}{W}_{\mathrm{1,1}}& W_{\mathrm{1,2}}& \cdots & W_{\mathrm{1,}{m}_2}\\ W_{\mathrm{2,1}}& \cdots & \cdots & W_{\mathrm{2,}{m}_2}\\ ⋮& ⋮& ⋮& ⋮\\ W_{m_1\mathrm{,1}}& \cdots & \cdots & W_{m_1,m_2}\end{array}\right)$; (8)

$W_{k,h}^{\left(3\right)}=\left(\begin{array}{cccc}{W}_{\mathrm{1,1}}& W_{\mathrm{1,2}}& \cdots & W_{\mathrm{1,}{m}_3}\\ W_{\mathrm{2,1}}& \cdots & \cdots & W_{\mathrm{2,}{m}_3}\\ ⋮& ⋮& ⋮& ⋮\\ W_{m_1\mathrm{,1}}& \cdots & \cdots & W_{m_1,m_3}\end{array}\right)$; (9)

$W_h^{\text{out}}=\left(\begin{array}{c}{W}_1\\ W_2\\ \cdots \\ W_{m_3}\end{array}\right)$; (10)

$b_j^{in},b_k,b_h,b^{\text{out}}$ are the displacement vector for the input layer, the displacement vector for the 1-hidden layer, the displacement vector for the 2-hidden layer and the displacement vector for the output layer:

$b_j^{in}=\left(\begin{array}{ccccc}{b}_1& b_2& b_3& \dots & b_{m_1}\end{array}\right)$; (11)

$b_k=\left(\begin{array}{ccccc}{b}_1& b_2& b_3& \dots & b_{m_2}\end{array}\right)$; (12)

$b_h=\left(\begin{array}{ccccc}{b}_1& b_2& b_3& \dots & b_{m_3}\end{array}\right)$. (13)

The used analytical model (2) makes it possible to make an evaluation the humidity of the material in scale of the real time in the drying unit in the process of its drying with a relative error, less than 2 %.

To estimate and monitoring of the humidity of the material during its drying in the drying unit, it is necessary to set up the system once. Using arrays of statistical data obtained from the transducers installed in the drying unit, we train a neural network. The obtained network parameters $W_{i,j}^{\left(1\right)},W_{j,k}^{\left(2\right)},W_{k,h}^{\left(3\right)},W_k^{\text{out}}$, $b_j^{in},b_k,b_h,b^{\text{out}}$ are recorded in the database system, and the resulting analytical model is recorded in the knowledge base. In the database is also stored the information which is coming from the transducers in the process of the system functioning.

The method allows for real-time operational assessment of humidity of the stillage in the process of its drying. The operative assessment of the humidity makes it possible, in the case of the deviation of the value of the humidity from the ideal humidity in the drying unit, to set up the parameters of the drying unit opportunely in order to increase the quality of the released material.

In the developed model of relative humidity estimation the fully connected neural network architecture is used. The proposed architecture has five layers with the corresponding number of neurons in the layers: 8-36-36-36-1. When compared with the experimental measurements, the predicted values followed the law of variation and the size of the experimental values, and the relative error of AARE (Average Absolute Relative Error) does not exceed 2 %. Hence, the neural network model can successfully predict the relative material humidity estimation during the drying process.

At the same time, there are no feedbacks and no connections between neurons in the same layer are utilized. The neurons in each layer are only connected to the neurons in the neighboring layer. The ReLU function was chosen as the activation function in the hidden layers and a linear function was used for the output layer [2].

The performance of the model was evaluated using the correlation coefficient – R_cc, Mean Square Error (MSE) and Mean Relative Error (MRE) of the training sample.

The correlation coefficient is defined as

$R_{\text{cc}}=\frac{{\displaystyle \sum _{i=1}^N\left(\left({\epsilon }_i-\epsilon \right)\left({\rho }_i-\rho \right)\right)}}{\sqrt{{\displaystyle \sum _{i=1}^N{\left({\epsilon }_i-\epsilon \right)}^2\sum _{i=1}^N{\left({\rho }_i-\rho \right)}^2}}}$.

Mean Square Error is calculated as

$\text{MSE}=\sqrt{\frac{1}{N}\sum _{i=1}^N{\left({\epsilon }_i-{\rho }_i\right)}^2}$.

Mean Relative Error

$\text{MRE}=\frac{1}{N}\sum _{i=1}^N\left|\frac{{\epsilon }_i-{\rho }_i}{{\epsilon }_i}\right|\times 100$,

where ${\epsilon }_i$ and ${\rho }_i$ are the experimental and predicted value respectively, and N is the total number of data used in the study. The coefficient $R_{\text{cc}}$ is in [0, 1], and the closer the value is to 1, the better is the model performance, and the closer it is to 0, the worse is the model performance. The smaller is the selective root mean square error, the better is the quality of the prognostication and the model performance [2].

The architecture of the obtained artificial neural network was selected experimentally. From the analysis of the subject area, the number of neurons in the input layer was set to eight in accordance with the parameters that have a significant impact on the drying process in the drum drying unit, and the number of hidden layers and neurons in them was determined based on the accuracy and performance of the model during its training [2].

The optimal number of neurons in each layer was selected using comparative analysis. When the number of hidden layer nodes is equal to 36, the smallest training error is set at the level 0.1896 % and the number of epochs is equal to 50,000. These results show that the neural network model has good generalization ability.

The correlation coefficient $R_{\text{cc}}$, the MSE for the training data set are 99.18 %, 0.1609 % respectively.

The values of the correlation coefficient $R_{\text{cc}}$, the MSE for the set of the testing data are 98.94, 0.1896 % respectively.

The final artificial neural network model was tested on the test data not used in training (Figs. 2, 3).

 

Fig. 2. Correlation between experimental and predicted humidity data during the training

 

Fig. 3. Correlation between experimental and predicted humidity data during the testing

 

The method for indirectly measuring the relative humidity of the stillage during its drying process in the DDU is as follows:

Step 1. Interrogate the transducers of the drying unit and calculate the input values of the neural network.

Step 2. Check the conditions of the ingoing measured values in the adaptive ranges of values used in training the neural network. If the condition is not met, the data is written to the database and a message is sent to the user.

Step 3. Normalize the obtained values (1).

Step 4 Calculate the humidity of the material using analytical model of the neural network (2).

Step 5. Save the calculated value in the database.

Estimation of the material humidity using the example of a random sample of data shown in Table 1.

 

Table 1. Estimation of the material humidity using the example of a random sample of data

x1	x2	x3	x4	x5	x6	x7	x8	Reference metric $\phi$	Estimation of the humidity from the model $\phi$
80.6	117	6.47	114	44	5.5	108.3	75	97.0	96.85234
73.4	115	4.82	109	63	5.4	100.8	75	97.2	97.35859
72.7	115	7.28	117	71	5.8	106.6	77	97.4	96.82810
73.6	116	7.21	111	70	6.2	106.7	76	96.7	96.63859
74.6	116	5.22	107	70	7.4	101.5	77	97.5	97.46047
76.0	115	7.03	131	69	4.6	108.5	75	96.2	96.49960
75.8	106	4.67	125	82	4.9	105.5	70	95.1	95.57700
73.6	106	4.67	123	81	5.4	103.4	71	96.2	96.44693
75.6	98	3.78	115	85	4.4	100.5	75	95.5	95.53036
74.5	114	6.82	119	57	5.1	103.1	80	96.0	95.97664

 

In Table 1, the maximum absolute inaccuracy of humidity monitoring by model (2) is < 0.6 and the relative error is < 0.6 %. To analyze the data, the correlation coefficient was used, which was equal to 0.9894. This result shows that the experimental and predicted values match well. The neural network model with error back-propagation algorithm has good performance and can explain the coincidence of 98.94% of the above experimental values and the values calculated by the model. Based on this, it can be concluded that the obtained artificial neural network can model and predict the change in humidity of the stillage during its drying.

The information-measurement system of the drum drying units

The information-measurement system is realized in the form of a separate program module integrated into the PPCS, which allows determining promptly in real time the humidity of dried stillage at the outlet of the drying unit.

The block diagram of the developed information-measurement system is shown in Fig. 4. The center of this system are knowledge and data bases. In the first case, it is supposed to locate a sufficient amount of information, which is needed to build analytics and decision making.

 

Fig. 4. Structural diagram of the IMS of the relative humidity of the material

 

1. Drying unit of drum type.

1.1 – 1.8. Aggregate of primary measuring transducers of technological parameters of drying process.

2. Means of calculation of input variables.

3. Means of calculation of humidity.

4. Database.

5. Graphical and textual means of the information displaying.

6. Control system of the technological process.

The information-measurement system uses the frame model of the knowledge base [3], which contains the neural network model of the drying process in the DDU. The workflow is defined by the following key points: the humidity calculation module has primary knowledge from pre-collected data and knowledge bases. The latter are formed through neural training and augmented over time. Among the key criteria, the coefficient of normalization of primary information and the acceptable range of characteristics are highlighting. The primary measuring transducers located in the DDU request information through the PPCS: output product temperature, input product humidity, hot steam pressure, evaporation temperature, exhaust fans power, oxygen content in the air, load on the electric motor of the drying unit. The information obtained is also recorded in thedatabase and then sent to the calculation module of humidity estimation. Here, the data are normalized and the calculation and analysis of the material humidity level is performed. The obtained results are recorded in the database. The results are transmitted to the PPCS. The operator of the PPCS can control the drying process in drying units equipped with the IMS on the monitor screen.

The proposed methodology of the IMS construction contains the following stages:

1. Creation of the system requirement taking into account the key control tasks.
2. Technological characteristics determination which have an effect on the level of the stillage relative humidity.
3. The realization of the data collection for the knowledge base about the stillage relative humidity.
4. Division of the information flows into the test and training components.
5. Defining the architecture of the neural traning system.
6. Performing the verification operations regarding the adequacy of the applied model.
7. Development of the algorithmic support and the software.

The information base of the IMS is used to form a system of declarative type, in which the storage and the use of information flows for the work of a special module and of the system as a whole are performed through direct implementation in the PPCS.

The core of the IMS are the knowledge base (KB) and the database (DB).

In the IMS a network model of the KB is used. It contains neural network models of operational estimation of humidity in the process of the material drying in the drying unit. The knowledge here is represented in the following form:

$S_r=\left\{x_n,W_{n,m_1}^{\left(1\right)},W_{m_1,m_2}^{\left(2\right)},W_{m_2,m_3}^{\left(3\right)},W_{m_3}^{\text{out}},b_{m_1}^{\text{in}},b_{m_2},b_{m_3},b^{\text{out}},c_{m_1},c_{m_2},c_{m_3},f_1,f_2,f_3,f_4\right\}$,

where S is a network model; r is a network model number; $x_n$is a multitude of elements of the input layer of the network; $W_{n,m_1}^{\left(1\right)}$ is a connection matrix (synapse weights of the neurons of the first hidden layer), with the dimension $n\times m_1$; $W_{m_1,m_2}^{\left(2\right)}$ is a connection matrix (synaptic weights of the neurons of the second hidden layer), with the dimension $m_1\times m_2$; $W_{m_2,m_3}^{\left(3\right)}$ is a connection matrix (synaptic weights of the neurons of the third hidden layer), with the dimension $m_2\times m_3$; ${\sum }_1$ is a multitude of states of the neurons of the 1st hidden layer, here, the power of the multitude is determined by the calculation number of neurons of the 1st hidden layer equal to $m_1$; ${\sum }_2$ is a multitude of states of neurons of the 2^nd hidden layer, here, the power of the multitude is determined by the calculation number of neurons of the 2^nd hidden layer equal to $m_2$; ${\sum }_3$ is a multitude of states of neurons of the 3^rd hidden layer, here, the power of the multitude is determined by the calculation number of neurons of the 3^rd hidden layer equal to $m_3$; $c_{m_1}$ is a multitude of neurons of the 1st hidden layer, the power of the multitude is equal to $m_1$; $c_{m_2}$ is a multitude of neurons of the 2^nd hidden layer, the power of the multitude is equal to $m_2$; $c_{m_3}$ is a multitude of neurons of the 3rd hidden layer, the power of the multitude is equal to $m_3$; $W_{m_3}^{\text{out}}$ is a vector of links (vector of weight coefficients of the output layer); $b_{m_1}^{in},b_{m_2},b_{m_3},b^{\text{out}}$ are the displacement vector for the 1-hidden layer, the displacement vector for the 2-hidden layer and the displacement vector for the output layer; $f_1,f_2,f_3$ are activation functions ReLu, $f_4$ is a linear function of the activation.

Conclusion

The use of the IMS of the DDU allows to perform estimation of the relative humidity of the distiller’s grain in mode of the real time. The implementation of the IMS of the DDU for estimation and monitoring of material humidity in the process of drying in the drum dryer at Talvis JSC in Novaya Lyada allowed raising the output of quality product. Thus, the assigned technical task has been reached by obtaining a real-time operational assessment of the material humidity during the drying process to control this process in order to ensure the specified quality of the released product.

About the authors

S. V. Artemova

Tambov State Technical University

Email: karasev@mirea.ru

Doctor of Technical Sciences, Head of the Department KB-1 "Information Security"

Russian Federation, Tambov

M. A. Kamenskaya

Tambov State Technical University

Email: karasev@mirea.ru

Candidate of Technical Sciences, Associate Professor of the Department of Electrical Power Engineering

Russian Federation, Tambov

P. I. Karasev

MIREA – Russian Technological University

Author for correspondence.
Email: karasev@mirea.ru

Candidate of Technical Sciences, Associate Professor of the Department of KB-1 "Information Security"

Russian Federation, Moscow

N. S. Ershov

MIREA – Russian Technological University

Email: karasev@mirea.ru

Lecturer, Department of KB-1 "Information Security"

Russian Federation, Moscow

Chien Vu Tri

MIREA – Russian Technological University

Email: karasev@mirea.ru

Postgraduate student of the Department of KB-6 "Devices and Information-Measuring Systems", Institute of Cybersecurity and Digital Technologies

Russian Federation, Moscow

A. A. Domornikova

MIREA – Russian Technological University

Email: karasev@mirea.ru

Student

Russian Federation, Moscow

References

Supplementary files