Neuro-controller implementation for the embedded control system for mini-greenhouse

Development of the structure and algorithm of the functioning of an intelligent data collection and processing tool

In general, the structure of the basic intelligent data collection and processing tool includes three main components: a subsystem for collecting data about the environment; a subsystem for processing input technological data; and a subsystem of influence on the studied environment.

In mathematical form, the corresponding structure can be written using the following tuple: (1)${\displaystyle N{e}_{contr}=〈M_{sensors},M_{hard-softwore},M_{actuators},M_{\mathrm{int}s}〉}$ where M_sensors is a set of sensors and detectors; M_{hard−softwore} is a set of hardware and software tools; M_actuators is a set of actuators that make it possible to influence the studied environment and M_ints is an incidence matrix that allows establishing relationships between sensors, software and hardware, and actuators.

Let us consider an intelligent data processing tool using the example of implementing an intelligent mini-greenhouse management system (Ma, Li & Yang, 2018; Suryawanshi et al., 2018). The intelligent greenhouse provides maintenance of the microclimate and lighting regime for growing plants according to the specified conditions. We implement the corresponding system on the basis of the developed neuro-controller. The structure of the developed control system consists of two main components (Fig. 1): a microcontroller that runs the main loop of the control program and a neuro-controller software. The control system analyzes input data from various sensors and generates control signals for actuators.

There are four main sensors to gather information about mini-greenhouse:

• air temperature sensor, that monitors the temperature of the environment;

• air humidity sensor to gather information on air humidity;

• sensor to track soil moisture;

• light sensor, that checks outdoor lighting level.

Additionally, a real time clock is used for tracking the time of day.

The block diagram of the algorithm of the neuro-controller for controlling the intelligent mini-greenhouse is shown in Fig. 2.

The above sensors make it possible to determine changes in the mini-greenhouse environment (Ma, Li & Yang, 2018; Suryawanshi et al., 2018). The actuators were used to influence the medium of the greenhouse. In particular, the developed intelligent greenhouse control system (Batyuk, Voityshyn & Verhun, 2018) uses the following actuators(executive modules): the subsystem for watering the soil; the ventilation subsystem to reduce temperature and humidity inside the air control and the cleaning system; the subsystem of heating air; the lighting subsystem that turns on in case of insufficient external light.

For the case of the structure shown in Fig. 1, the set of sensors includes elements, namely: ${\displaystyle M_{sensors}=\left(S_1,S_2,S_3,S_4\right),}$ where S₁ is the air temperature sensor; S₂ is the air humidity sensor; S₃ is the soil moisture sensor; S₄ is the illumination sensor of the medium of the mini-greenhouse.

A set of hardware and software tools includes only two components: a microcontroller and software that emulates the operation of an artificial neural network. It should be noted that the appropriate set may include several microcontrollers, which are determined by technical and economic feasibility. The set of actuators also includes four elements (modules): ${\displaystyle M_{actuators}=\left(A_1,A_2,A_3,A_4\right),}$ where A₁ is a module that implements the function of watering a controlled environment; A₂ is a module that provides ventilation of the environment; A₃ and A₄ heating module and lighting module, respectively.

The incidence matrix for the studied structure of the system in Fig. 1 makes it possible to display the connections between structural elements. It has the following form: ${\displaystyle M_{int\mathrm{s}}=\left|\begin{array}{c}\hfill 1111000000\hfill \\ \hfill 0000000000\hfill \\ \hfill 0000000000\hfill \\ \hfill 0000000000\hfill \\ \hfill 1111111111\hfill \\ \hfill 0000100000\hfill \\ \hfill 0000010000\hfill \\ \hfill 0000001000\hfill \\ \hfill 0000000100\hfill \\ \hfill 0000000010\hfill \\ \hfill 0000000001\hfill \end{array}\right|.}$

The algorithm of the neuro-controller operation includes several basic steps (Fig. 3). The developed algorithm provides an initial step, which is designed to establish initial data (system initialization, port initialization, etc.). The following steps are performed sequentially in the cycle: polling and receiving technological data from sensors; the step related to the processing of the received data from sensors by a neural network and the step of forming control signals for the subsystem of influence on the studied environment.

Representation of the structural model of the system in graph form (Eq. (2)) makes it possible to analyze the functioning of the system using existing free software systems. (2)${\displaystyle G=\left(P,I\right),}$ where P is a set of nodes (components) and I is a set of arcs.

To analyze the operation of the intelligent mini-greenhouse control system, it is advisable to use a structural model in the form of a graph (Fig. 4).

Development, training and implementation of the artificial neural network model

Data preparation for neural network training and testing

To train the neural network, a set of data is generated that describe the received values from the sensors and the expected values for the actuators. For training, it is necessary to generate a sufficient sample that will represent the various states where the system can be in.

To prepare a set of data and its normalization a JAVA program was developed. Program randomly selects the values of the sensors and analytically calculates the expected values on the actuators. The generated data sets must be normalized so that the values are in the range [0..1]. For this purpose, at the 2nd stage, the program runs the developed normalization module and outputs the final results. The example of the main class code for data normalization is shown in Fig. 5.

An example of the generated training sample for training a neural network is shown in Fig. 6.

Peculiarities of implementation and training of an artificial neural network

The NeurophStudio environment was used for designing, training and checking neural network correctness functioning. This is free software for various types of neural network design. It allows monitoring of the learning process, modifying the structure of the neural network, determining a set of training values, visualizing the learning results, etc.

A multilayer perceptron was chosen as a neural network and designed with the help of NeurophStudio. It consists of four input neurons and one balancing neuron, six internal neurons and one balancing neuron. The output layer contains four neurons.

The peculiarity of this artificial neural network is that all neurons of the first layer have a connection with all neurons of the second layer (except for the balancing one for the second layer). The situation is similar with the connections between the second and third layers, namely: seven neurons of the second layer have connections with all four neurons of the third layer (Fig. 7).

Within the environment, the structure and connections between neurons can be explored. For correct training of the neural network, it is recommended to divide the set into two parts. The first part is used to train the neural network, while the second part is a control set that can be used to test the neural network on new data. In this case, neural network training takes place on 80% of the data set.

The training of neural network of multilayer perceptron is carried out by the backpropagation method. An example of the dependence of the neural network learning error on the number of iterations is shown in Figs. 8 and 9.

The neural network is tested on 20% of the data that was not involved in the training process, and the results of work on this sample provide an opportunity to check how the neural network functions on independent data.

Checking the accuracy of the functioning of the neural network for the intelligent mini-greenhouse control shows the mean squared error value of 0.032. The trained neural network is used for neuro-controller implementation.

Hardware and software implementation of the neuro-controller

The development of a neuro-controller includes two main parts: hardware and software.

Features of the hardware implementation of the neuro-controller

We developed a structural diagram of the hardware part of the neuro-controller, which includes the following components (Fig. 10):

• STM32-F103C8T6 microcontroller, which is based on the ARM 32 Cortex-M3 MCU with an operating frequency 72 MHz, 64 KB of Flash memory and 20 KB of data memory;

• real time clock DS1307, which is necessary to determine the time of day;

• a number of sensors were used to monitor technological data on microclimate and lighting:

• temperature sensor DS18B20, which makes it possible to monitor the temperature in the range from −10 °C to 85 °C with an error of 0.5 °C;

• temperature and humidity sensor DHT11, which allows to monitor air humidity in the range from 20% to 80% with an error of 5%. Also, the sensor allows to monitor the temperature in the range from 0 °C to 60 °C with an error of 2%. The DS18B20 sensor gives a value with a smaller error, so it is recommended to use it in contrast to the DHT11;

• soil moisture sensor that measures soil conductivity and provides data in the range of 0–5 V depending on the set threshold value;

• photoresistor, on the basis of which the light sensor is organized and an auxiliary balancing resistor is used. The change in the voltage ratio between the balancing resistor and the photoresistor is determined by external lighting.

• relays are used as executive devices to control the power supply of the actuators.

The online EasyEDA development environment was used to design the hardware part of embedded control system for mini-greenhouse. This environment makes it possible to synthesize the electrical schematic diagram of the circuit board (Fig. 11).

The schematic diagram shows each component’s pins and their connections.

The software part of the developed embedded control system for mini-greenhouse

The software for an embedded control system for a mini-greenhouse consists of a set of modules. Each of them implements some operations, including the realization of the neuro-controller and its control. The interconnection of these modules is shown in Fig. 12.

For system initialization, initialization of ports, sensors, actuators, real time clock and loading of initial data a system initialization module is responsible.

A sensor control module is used to periodically read data from sensors, and access the real time clock. It sets up the initial time and periodically reads the current time of day. The sensor control module is accessed by a periphery control module. This module also transmits control signals to the actuator control module by means of relays that change the state of the actuators.

The developed neuro-controller consists of several modules, described below. The neural network loading module is used for the neural network initialization, loading the neural network configuration, neuron function types, and weighting coefficients matrices. The neural network input/output module is responsible for loading input data into the neural network and reading the output data for the actuator’s control. The Data processing module is used to coordinate the work of this modules. It also controls the Neural network emulation module. For emulation, the subroutine takes information about the type of neuron function, weighting coefficients, and connections between neurons and performs calculations of each neuron’s value. As a result, the emulated neural network forms values that are used to control the mini greenhouse.

The main feature of the neuro-controller is its cyclical operation. Since there is no need to frequently poll and update the system state, the interrupt operator is used, which stops the main cycle of program execution and limits the frequency of execution of the main cycle.

During the execution of the main cycle, a subroutine is called to simulate the operation of the neural network. Based on its response, commands are formed to control the actuators.

The algorithm of emulating of the neural network operation is presented in Fig. 13.

The subroutine is responsible for calculating the values of each of the neurons. As input data of the neural network the measured values from sensors and the real time clock are used. At the output, the value of the state to which each of the actuators should be transferred (on/off) is calculated.