Utilizing Neural Networks for Dynamic Performance Improvement of Induction Motor Drive: A Fresh Approach with the Novel IP-Self-Tuning Controller

: This paper introduces a neural network adjustment method for a single gain of an IP speed regulator, to improve the speed control of an induction motor. Thanks to its simplicity and strength the (IP) controller is widely used in the industry for speed control. Yet, in some cases, when the load or mechanical parameters change according to its working conditions, the IP efficiency decreases and the setup quality degrades. In this case, a neural IP-self-tuning seems to overcome these difficulties and ensure a good control performance. The results obtained through the implementation of the proposed control on a dSPACE system and an induction motor clearly demonstrate the effectiveness of this method.


Introduction
During the past few decades, the induction motor has accounted for more than half of industrial processes [1,2]. Thanks to its numerous advantages, it has received special attention in modern industrial installations and has been widely used to enhance energy efficiency and improve process controls [3]. The induction motor offers several benefits compared to other types of rotating electrical machines, including low manufacturing costs and a lower purchase price. Furthermore, it can be directly powered by the three-phase grid, which makes installation easier. Its robustness also reduces maintenance costs since it does not have a mechanical commutator like the DC machine [4]. It is precisely because of these two fundamental qualities (price and robustness) that the combination of the speed controller with the asynchronous machine is currently the most widely used in industrial applications requiring variable speed [5].
It is important to keep in mind that the complexity of controlling the asynchronous machine has led to the development of various control strategies, among which the most popular is field-oriented control in its various versions. The objective of this control is to achieve the performance obtained with the control of a separately excited DC machine [6,7]. Vector control offers the possibility to precisely and rapidly control AC machines. It has numerous advantages over scalar control, such as fast torque response, high precision in speed and torque control, and improved stability through independent component decomposition.
However, vector control also has certain drawbacks, such as reduced robustness to parameter variations, the requirement for a modulator for precise inverter control, and the introduction of control errors due to the use of coordinate transformations dependent on an estimated angle θs [8,9].
The type of controller used in more than 90% of the control systems is the proportional integral derivative one (PID) [10]. These controllers, whose uses are incredibly numerous and varied, date back to the beginning of the twentieth century. The IP controller has both a modulating action to set the speed whenever control is to take place, and an integral one to eliminate the static error [11].
Fuzzy controllers are very suitable to control induction motor thanks to their non-linear feature (a highly non-linear process). By appropriately selecting the membership functions and setting up basic guidelines, non-linearities can be made up in the given system. Membership functions are mathematical models for linguistic terms, such as the triangle function, the trapezoid function or the Gaussian function [12].
In recent years, the sliding mode controller has garnered considerable acclaim owing to its straightforward implementation and resilience against system uncertainties and external disturbances. This controller aims to bring the system's state trajectory towards a designated surface, known as the sliding surface. Through an appropriate switching logic, the controller continuously maneuvers the system around this surface until a balanced state is achieved. This two-phase process involves the approach phase, where the system approaches the sliding surface, and the glide phase, where the system maintains its position and smoothly slides along the surface. The sliding mode controller's effectiveness lies in its ability to handle uncertainties and disturbances, ensuring stable and reliable control even in challenging operating conditions [13], [14], [15], [16].
Adaptive control terminology, whose methods enable a real-time automatic adjustment of the controller settings, is implemented in a control loop. This is done to maintain a desired level of performance when the process parameters are unknown or vary slightly over time [17].
Author in [18], suggested a novel adaptive control approach in order to improve the speed of the induction motor, using an IP self tuning controller set by a fuzzy adjuster, depending on the speed error and its derivative with respect to time. Their aim is to adjust a single gain achieved in real-time by an adaptive IPfuzzy speed controller. Finally, the author conducted a comparative evaluation involving the Fuzzy-PI controller, conventional IP, and the proposed strategy. This study showed that accuracy was improved and the speed was tracked.
In [19], the authors proposed an adaptive control method to improve the speed response of the induction machine in the presence of uncertainties and external disturbances. This method is based on neural networks, with the objective of adjusting the controller gain. The performance of this control strategy was examined through simulation results. The achieved results highlight the control algorithm's resilience to both load disturbances and speed variations.
This paper aims at comparing the performance of the rotor flux-oriented indirect vector control using two types of speed controllers: the classical (IP) and the Self-tuning (IP) based on Artificial Neural Network to improve this control strategy. To showcase the efficacy of the suggested controller, an experimental setup utilizing dSPACE is employed. The experimental system includes a DS1104 controller board from dSPACE, which is a powerful hardware and software platform for the development and validation of real-time control systems.
The remaining of this article is presented as follows. Section 2 introduces the dynamic model of the engine and explains the principles of vector control. It provides details of the engine's operation as well as the fundamental concepts of vector control. In Section 3, the design of the proposed controller is described in detail. The different components and algorithms used in the controller are explained, highlighting the specific techniques employed to improve the system's performance. Experimental results are presented and discussed in Section 4. The experiments are conducted under real conditions to evaluate the effectiveness and robustness of the proposed controller. The system's performance is analyzed and compared to other existing approaches, where applicable. Finally, Section 5 presents the article's final conclusion. The main findings are summarized, highlighting the study's contributions and future prospects for potential improvements to the controller and the system.

Three-Phase Induction Motors
The equations describing the rotor and stator quantities of a three-phase induction motor in the d-q reference frame are as follows [20] Where,"(Vsd, Vsq)" refers to stator voltages, "(Isd, Isq)" corresponds to stator currents, and "(φrd, φrq)" represents ; ; where  is the blondel leakage coefficient.
The electromagnetic torque of the induction machine in the d-q reference frame, along with the mechanical equation, can be expressed as follows: The total moment of inertia of the machine shaft and mechanical load is represented by J. The terms r f and r C correspond respectively to the contributions of viscous friction and the load torque.

Vector control
Vector control uses a Park transformation to decompose current and voltage into two components: torque and flux [21,22]. With this decomposition, vector control is capable of independently controlling the flux component and the torque component, allowing for precise and fast output torque for rapid dynamic response.
Additionally, this control method also enables precise regulation of the machine's speed by utilizing a speed feedback loop. Thus, vector control is considered an advanced control method that allows for precise and fast control of AC machines, such as synchronous and asynchronous machines. Compared to scalar control, vector control offers numerous advantages, including fast torque response, high speed and torque control accuracy, as well as improved stability through the decomposition into independent components [23]. Fig.1 illustrates the fundamental concept of vector control. Where, (φrd, φrq) denote the d-q rotor components fluxes. Ce is the electromagnetic torque. (θs, θ) are the position of the d-q reference frame and the rotor position, respectively.

Standard IP controller
The Proportional Integral (PI) controller differs significantly from the Integral Proportional (IP) controller primarily due to the absence of a zero in the closed-loop transfer function of the IP controller. Also, its output will not exhibit a discontinuity when applying a step-type setpoint [24], [25]. The following figure shows the block diagram of the speed control with the IP controller: The backpropagation of error is a key mechanism in the learning phase of ANNs. It allows for the evaluates of the comparison score of the predicted output, and propagates this error backward through the network to adjust the weights of the synaptic connections according to the formula of equation (8). This enables the gradual improvement of the network's performance and enables it to solve complex problems. The sum of squared errors given by equation (7) [27].
Where, E represents the mean squared error, n is the number of examples, d y represents the actual values, and i y represents the predicted values.
Where, ij  represents the weight of neuron i in the actual layer and neuron j in the next layer,  is the learning rate.
Backpropagation allows for the adjustment of weights in a neural network by calculating the gradient of the error with respect to the weights and using this gradient to update the weights. In equation (9), an additional term is included to optimize the network's performance by minimizing the error between the network's predictions and the desired values [28,29].
The Backpropagation phase is illustrated by  After testing different types of artificial neural network architectures, this study used artificial neural network back propagation. The aim is to train networks to produce reasonable responses to control the induction motor. ANNs are typically composed of three primary layers: the input layer, the hidden layer, and the output layer. The first one with two neurons, i.e., the reference speed and the speed response which can be labelled as *  and  respectively. The output layer has a neuron whose weights are added to each connection line to adjust two input data by a learning process. And the hidden layer is made up of multiple layers which helps the model to do the right learning.
The flowchart of the learning process is described in detail in Fig 5. In most cases, the sigmoid transfer function is frequently used in multi-layer networks built using the back propagation phase. In this study, the sigmoid transfer function is applied to the 1,1  product. Hence, this transfer function takes the input and overwrites the output in the range of 0 to 1. The transfer function for the output layer is a linear function named purlin. Fig.5 demonstrates the flowchart of the process of this backpropagation. To explain, we note that the first one has random weights. The error between the network output and the intended output is limited. Fig.6 shows the block diagram of the improved ANN IP controller The artificial neural network, which collaborates with IP Controller, has two inputs and one output. The inputs are the reference speed and the speed response, and the output is the value of p dK . The proportional total p K can be set as follows: to the analog-to-digital converter unit is established through the CLP1104 connection panel. Furthermore, the motor speed is determined using a tachometric dynamo positioned between the two motors, with a gain of 0.19 V/rad/s.
The drive system enables real-time control of the machine through the utilization of measured data and control signals, including speed, torque, and current. The connection panel is connected to an adaptation card, which serves two primary functions.
Firstly, the adaptation card provides galvanic isolation and modifies the control signals received from the DS1104 card, which operates at TTL level (0/5V). These signals are adjusted to control the SKHI 20opA drivers, requiring a CMOS level of 0/15V.
Secondly, the adaptation card modifies the signals from the LEM current sensors to limit the output voltages to levels acceptable by the analog inputs of the DS1104 card.
The sampling period used in this study is e T = 0.14ms. Table 1 presents the parameters and specifications of the induction motor (IM).

Experimental results for speed control
The experimental results presented above describe a comparative analysis between IP-self-tuning regulation based on neural networks and classical IP regulation. Several practical tests were conducted to evaluate these two approaches. The experimental trials were carried out on a test bench illustrated in Fig. 8. The performance was assessed in terms of speed regulation accuracy, disturbance rejection, and system stability. The obtained results are presented in the figures below.
a. Speed variation tests.  preserving the properties of vector control. Fig. 11 shows that the decoupling between torque and flux is maintained.
b. Load torque test (1000 rpm).  The objective of this comparison test is to evaluate the effectiveness of the two controllers in rejecting disturbances. During this test, the speed was kept constant at 1000 rpm. At t=2 s, a load torque was applied and then removed at t=7 s, as shown in Fig. 12. Throughout this experiment, we observed that disturbances had less impact with the IP-self-tuning compared to the conventional IP.
An important observation is that the q-component of the stator current reacted more quickly with the IP-selftuning than in the case of the conventional IP, as shown in Fig. 13. These results highlight the capabilities of the IP-self-tuning in terms of disturbance compensation and maintaining system stability.
These results emphasize the advantages of IP-self-tuning over the conventional IP in terms of disturbance rejection and system stability restoration. The use of IP-self-tuning can therefore be considered a promising solution for applications requiring a fast and accurate response to disturbances. Fig. 14 shows that during this test, the decoupling between torque and flux is maintained.
c. Load torque test (500 rpm).    The third test aims to evaluate the ability of the two controllers to track a speed profile. As shown in figs.
18(a) and (b), a speed reference of 600 rpm to 0 rpm and -600 rpm to 0 rpm was applied.
The observed results clearly demonstrate the effectiveness of the adaptive IP approach compared to the conventional IP in speed response. In fig. 18 fig. 20 shows that the decoupling between torque and flux is maintained.

Conclusion
This article presents a neural network-based tuning method aimed at improving the speed of an induction Additionally, determining optimal weighting values can also be complex. Despite these limitations, the results obtained with the neural network tuning method show real potential for improving the performance of IP controllers in the specific context of speed control in IRFOC-based induction machines. Further research is needed to develop more systematic approaches that guide the selection of neural network parameters and optimize its operation, thereby facilitating practical application.

Declarations
Ethical Approval: Not applicable.

Competing interests:
The authors declare no conflict of interest. Funding: No institution, agency, or organization provided financial support for this research.

Availability of data and materials:
The datasets used in this study are available upon request. Interested researchers can contact me at [abdellah.elkharki@gmail.com] to obtain access to the data and materials used.