Artificial Bee and Ant Colony-assisted Performance Improvements in Artificial Neural Network-based Rotor Fault Detection

motor failures used in production be detected and easily. In this article, sound signals were recorded during the operation of the asynchronous motor, is operational and with a rotor bar crack; and filtering, normalization, and Fast Fourier Transform were performed. The detection of rotor broken bar error was examined using the feed-forward backpropagation Artificial Neural Network (ANN) method. With intuitive algorithms such as the artificial bee colony and artificial ant colony, improvements to the ANN results were investigated. experimental intuitive algorithms the estimation performance of the neural network.


I. INTRODUCTION
Asynchronous motors, which have important features such as cost effectiveness, easy speed control, low maintenance cost, and low maintenance requirement, are used in most production lines. Stator and rotor failures and, generally, mechanical failures can be seen in these motors. These failures must be detected during operation or in advance. In this way, by providing the proper operating conditions, the production system is prevented from being disrupted.
An Artificial Neural Network (ANN) is a structure based on a human learning mechanism. ANNs have been used in the diagnosis of motor failures based on learning ability. In many studies, detection methods related to asynchronous motor bearing and stator failures were examined [1], [2]. Glowacz and Glowacz [3] detected stator failures of a single-phase asynchronous motor that employs acoustic signals. In the article, the sound signals of the one-phase asynchronous motor were examined through ANNs and the short circuit fault detection in the stator windings were analyzed. The comparison was made by using three different ANN methods. Delgado-Arredondo, Morinigo-Manuscript received 12 January, 2022; accepted 11 February, 2022.
Sotelo, Osornio-Rios, Avina-Cervantes, Rostro-Gonzalez, and Romero-Troncoso [4] examined a method known as motor current analysis, which identifies errors in asynchronous motors through sound and vibration signals. In this article, the error detection methodology was examined by acoustic signals in squirrel-cage induction motors. The study was carried out to detect rotor bar fractures, mechanical imbalances, and bearing defects in the motors. Khater, El-Sebah, Osama, and Sakkoury [5] created a special model for the detection of rotor fractures of the asynchronous motor driven by an alternating current (AC) driver. They examined the determination of the number of broken rotor bars of the asynchronous motor. In a modelbased study [6], alterations in three-phase currents were observed to diagnose rotor fractures and bearing failures of induction motors. Asad, Vaimann, Belachen, Kallaste, Rassölkin, and Iqbal [7] used the applied voltage variations on the motor to determine the broken rotor bar of the motor, whose torque was controlled by an inverter. The study revealed the results that the method was effective. Deus, Sobrinho, Belo, Brito, de Souza Ramos, and Lima-Filho [8] reported a less costly method than conventional methods by analyzing the stator current values of the asynchronous motor, detecting broken rotor bar failures at different levels. Garcia-Calva, Morinigo-Sotelo, Garcia-Perez, Camarena-Martinez, and Romero-Troncoso [9] obtained results in the detection of broken rotor rod failure of inverter-driven motors by analyzing current harmonicity. Dias, da Silva, and Alves [10] detected broken rotor bars in AC motors, which brought up a new approach depending on current analysis. Soleimani, Cruz, and Haghjoo [11] reported a method for the detection of broken rotor bars in AC motors based on the examination of the air gap magnetic field between the stator and the rotor. Elez, Car, Tvoric, and Vaseghi [12] added two coils to the motor stator to measure the field change in the air gap. Magnetic field changes obtained by these coils were shown to detect rotor bar fractures in the motor. Gritli, Rossi, Casadei, Filippetti, and Capolino [13] examined the active and reactive powers of a three-phase winding rotor induction motor, and detected electrical faults of the rotor with rotor spectral current and voltage values. In the study in [14], before applying the monitoring methods based on the current analysis, the presence of various harmonics in healthy machines was thoroughly analyzed, the harmonics of the rotor fractures were compared, and the failure was detected. Keskes and Braham [15] achieved highly accurate results in the detection of broken rotor bar failures, which are very close to the main frequency in the spectrum and difficult to realize, with both wavelength package conversion and a graphic support vector combination. The authors in [16] performed the detection of rotor bar semi-fracture by applying multiple signal analysis to the stator current signal with a Kalman filter. St-Onge, Cameron, Saleh, and Scheme [17] detected rotor rod fracture using the attribute extraction method with pattern-based diagnostics in three-phase asynchronous motors. Trachi, Elbouchikhi, Choqueuse, and Benbouzid [18] detected broken bearings and rotor bars using a high-resolution spectral analysis technique based on stator current measurements. Xu et al. [19] achieved positive results using the multi-feature extraction method in winding short circuit detection in motors. Ramu, Irudayaraj, Subramani, and Subramaniam [20] proposed a new approach to detect broken rotor bar failures using Hilbert Transform and ANN when the asynchronous three-phase motor is controlled by a motor driver. Wang, Liu, Guo, and Wang [21] conducted a study of engine error analysis using the noise-based incomplete wavelet packet analysis -ANN model in gasoline engines.
When all these studies are examined, it is seen that the usage of ANN has increased in recent years in the detection of motor failures. And in this study, it is mainly proposed how the diagnostic performance of the ANN can be improved by employing intuitive algorithms. If the literature is examined, it is seen that no optimization process was performed during the submission of input data to the neural network [3], [20]- [23]. If metaheuristic optimization algorithms are viewed, it is seen that Artificial Bee Colony (ABC) and Artificial Ant Colony (AAC) algorithms have a simple and flexible structure [24]. They also provide ease of use for discrete problems. Therefore, in this study, the data set to be submitted to ANN was reordered using ABC and AAC algorithms, and the estimation performance of ANN was improved.
In this study, audio signs of a three-phase induction machine with healthy and broken rotor bars were recorded. Noises were minimized by employing the windowing method. Fast Fourier Transformation (FFT) was applied to the filtered signs. The resulting samples were applied to the feed-forward backpropagation ANN, and the rotor crack detection performance of the developed ANN was examined. After that, the data order was arranged using the artificial ant and bee colony algorithms. Data were ordered by calculating the shortest distance in artificial ant optimization and determining the best source of artificial bee colony optimization. New weights were obtained by training the ANN with the input data set that had been reordered with the ABC algorithm or the AAC algorithm.
The explanations on experimental results, intuitiveassisted ANN details, performance, and evaluation, and ultimately the results, are presented in the following sections of the article, respectively.

II. ARTIFICIAL NEURAL NETWORKS
An ANN consists of a series of interconnected processing elements called "neurons". A general view of the multilayered ANN structure, which is selected for employing in this study, is given in Fig. 1  The general representation of an artificial neuron is presented in Fig. 2. In this topology, the outputs of the neurons in the previous layer are taken as input parameters to the neurons in the next layer. And in the neuron, the input parameter is multiplied by its own weight value and incorporated into the operations. The neuron collects these weighted signals and its threshold value. In a classical neuron structure, the neuron runs its activation function with the total result. Finally, it sends the output of the transfer function to the neurons of the next layer. The relationship between the input signals and the output signal of a neuron is defined as where y j is the output value of the jth neuron, x i is the ith input value, w ij is the weight value of the connection between the jth neuron and the ith input, θ j is the threshold value of the jth neuron, n is the number of inputs, and f is the activation function used. Transfer functions such as the logarithmic sigmoids below can be preferred, since solutions to nonlinear problems are sought using ANNs. Furthermore, neuron function was selected as sigmoid as the output value would be in the range of 0 to 1 in this study The input signals are normalized to be compatible with the operating range of the selected transfer function. Denormalization is performed on the attained output signals; however, denormalization is not needed in this study.
To obtain the estimation ability, the ANN must undergo the training phase. In the training phase, the weight and threshold values are updated at each iteration, depending on the estimation error. In this study, the ANN parameters (weights and thresholds) are updated by performing the backpropagation algorithm which is given in detail in [25]. According to the termination criterion, the optimization process is stopped. The iteration number or the estimation performance can be chosen as the stopping criterion. Here, the iteration number is applied as a stopping criterion for the training.

III. THE ARTIFICIAL BEE COLONY ALGORITHM
The ABC algorithm was developed by modeling the behavior of bees searching for food in the natural world [24], [26]. It is accepted that the number of bees involved in this algorithm will be equal to the number of food sources available. In the algorithm, the bee colony is considered to consist of three different classes of bees: worker bees, onlooker bees, and scout bees. It is assumed that half of the colony are worker bees and the other half are onlooker bees. The scout bee, on the other hand, represents status and can be described as a temporary phase in which the worker bee, which is running out of food supply, randomly searches for a source of food. The worker bee, which runs out of nectar, turns into a scout bee. The algorithm uses the following equation to produce initial food sources min max min where i: 1... SN and j: 1... D. Here, SN is the number of food sources and D is the number of parameters to optimize. The lower limit of the j th parameter is min j x and the upper limit of the jth parameter is max j x . Each source has a worker bee. Therefore, the number of food sources and the number of worker bees are equal. The simulation to determine the new resource in the existing resource neighborhood is defined by the expression given in where j is a randomly selected value in the range of ,.
The conformity value of the new resource found is presented as follows where f i is the cost value of the v i resource that is the solution.
After completing their research in one cycle, all worker bees return and transfer the nectar information of the nectar sources they find to the onlooker bees in the form of dances. This is done using the conformity value. For this, the roulette wheel method is used. According to the roulette wheel method, the more the same option is in the areas on the rotating wheel, the more likely the arrow will point with that option.
All conformity values are collected and formulated with a larger place on the wheel for those with good conformity values.
The ratio of a resource's conformity value to the sum of the conformity value of all resources gives that resource's likelihood of being selected relative to other resources where fitness i is the quality of resource i, SN is the number of worker bees.
After the probability values are calculated in the algorithm, a random number is produced in the range of [0, 1] for each resource in the selection process according to the roulette wheel using these values. And if the value of p i is greater than the number produced, onlooker bees also produce a new solution, such as worker bees, in this resource region by using (4).
The new solution is evaluated and its quality is calculated. The new and old solutions are compared and the optimal one is selected. This process continues until all onlooker bees spread to food areas.
At the end of a cycle, when all worker and onlooker bees complete the search process, the worker bee in the finished resource now becomes a scout bee and randomly initiates the new resource search process, as given in (3).
The optimization process continues until the termination criterion is reached. In this study, the number of iterations is chosen as the termination criterion.

IV. THE ARTIFICIAL ANT COLONY ALGORITHM
Animals that live in colonies and solve their problems through cooperation are described as social animals. Ants are also animals that fall into this group [27]. The AAC algorithm is an algorithm based on mathematical models of real ant behavior. However, the difference from artificial ants is that they have a certain memory, are not completely blind, and live in a discrete-time environment. This algorithm yields good results in the solution of the traveler seller problem, which is widely used in the literature [28]. This problem consists of creating a closed tour with a minimum distance, provided that the seller visits each of the n number of cities once. Let d ij be the distance between cities i and j 12 22 If m indicates the number of ants in the colony, and b i (t) gives the number of ants at moment t, then (9) can be presented 1 ( ).
If τ ij (t) indicates the amount of pheromone substance stored in the line (i, j) between i and j, the pheromone substance at moment t + 1 is given as below where ρ is the evaporation rate. The amount of pheromone delivered in per unit time is calculated as 1 ( , 1) ( , 1).
For an ant, the heuristic information controlling the move from ith city to jth city is shown as follows Finally, the moving probability for the kth ant from ith city to jth city is presented in (13) Above and are two positive real parameters that specify the relative influence of the pheromone trail and the heuristic information. For the kth ant when it is at the ith city, k i N is the set of cities in the taboo list that has not been visited yet. A taboo list is prepared for each ant to visit n number of cities. When the tour is completed, this taboo list is reset too.
For this study, in the optimization study with the artificial ant colony algorithm, the distances between the spectral distribution points of each sign were taken into account and the shortest distance was calculated. The feed-forward backpropagation algorithm for the shortest distance values was reapplied to find the best solutions and new error values were given.

V. EXPERIMENTAL STUDY FOR ROTOR FAILURE DETECTION
The recorded motor sound signals for the detection of rotor bar fractures are processed step by step in accordance with the block diagram given in Fig. 3. Sound measurements made around the motor at various speeds and loads are taken in a laboratory environment and stored as data. Both for the healthy motor and for the motor where the rotor failure is created, these measurements are repeated and the audio data are transferred to the Matlab software.
The windowing method was used to filter sound signals. After some trials, the number of windows taken as 80. Figures 4(a) and 5(a) show raw audio data. A new signal was obtained by taking the arithmetic average of the 80 subparts of recorded audio data. Figures 4(b) and 5(b) show sound signals after the windowing process. The FFT transform of this resulting signal was performed, and the spectrums given in Fig. 4(c) and Fig. 5(c) were obtained. Spectral changes are obtained by using the FFT. Here, the number of spectral lines is 84, the bandwidth (f max ) is 200 Hz and the frequency resolution (Δf) is 2.37 Hz. The sampling frequency (f s ) was taken as 44 kHz and the block size (N) was 4096 in the time domain. In this study, to detect the faulty motor, the specific spectrum area, which includes the signals with an amplitude greater than 0.05, is considered. Figure 6 shows the spectrums selected for the healthy motor and faulty motor. This operation was applied to all signals. The harmonics presented in Fig. 6 were normalized with regard to main omponent and switched to comparison processes. A brief info about the experimental setup can be helpful to talk the laboratory studies. A 3-phase 0.25 kW 1390 rpm, 380 V AC, 0.8 A squirrel cage asynchronous motor was used. An image of the motor examined is given in Fig. 7. AC Driver: An ABB brand ACS150 model driver of 0.37 kW power was used as an AC driver to perform acoustic measurement at various speeds of the motor in the test circuit. A K&H brand, EM3320-1. A model magnetic braking unit was used to get acoustic measurements of the motor under load.
Microphone: A condenser-type microphone was used for acoustic measurements. The sensitivity range of the microphone is between 100 Hz and 10 kHz.
To create the bar fracture, a 2 mm-diameter hole was drilled from the point where the aluminum short circuit bars of the rotor were located after the rotor was removed from the motor. The resulting space was closed with silicone to prevent it from creating undesirable sounds. Figure 8 shows the hole that was drilled into the rotor. The motor was controlled by employing the AC driver. The motor was connected to the braking unit with the appropriate coupling. The microphone, located close to the body of the motor, was connected to the computer and audio files were recorded in *.m4a format.
While the speed of the motor was controlled by the driver, the change of load was adjusted in Nm via the braking controller. The motor was loaded at 0 % (no load), 25 %, 50 %, 75 %, 100 %, and 125 % of the nominal torque value. In the trials, it was observed that when the torque values increased in low-speed revolutions, the motor did not revolve and was locked. Therefore, a total of 54 data were recorded. In Table I, the data that cannot be recorded are shown with x. rses-i, i: 1, 2, ..., 30 are audio files recorded . The windowing method was applied in the Matlab software to filter the sound signal. FFT was performed for filtered audio data. For example, the spectrums obtained after the evaluation of the audio file of the motor with a solid rotor and the one with a single broken rotor bar with the Matlab are given in Figs. 9 and 10.

VI. DEVELOPMENT OF INTUITIVELY SUPPORTED ANN
Here, 44 data sets were used for training and 10 for testing. 22 of the 44 available data sets belong to the healthy motor, and the other 22 belong to the faulty motor.The first optimization study was performed using randomly ordered data sets. The second and third were made with data sets, which were ordered by ant colony and bee colony algorithms. The shortest distance value for the ant colony and the best source data for the bee colony algorithm were employed. The ABC algorithm usesm(3) to produce initial food sources [24]. The Mean Squared Error (MSE) is used as the performance index. Parameters for the ABC optimization algorithm: the number of food sources SN = 44 (number of data sets used for training steps), the number of parameters to optimize D = 45 (number of inputs from ANN), iteration number = 10, lower limit min where m is the number of data sets, E m is the estimation error where B m is the expected output (healthy motor: 1, faulty motor: 0), Q m is the ANN output.
In the AAC algorithm, this problem consists of creating a closed tour with a minimum distance, provided that the seller visits each of the n numbers of cities once [27], [28]. Distance d ij is calculated by employing (8). The distances between the spectral distribution points of each sign were taken into account and the shortest distance was calculated. The feedforward backpropagation algorithm for the shortest distance values was reapplied to find the best solutions, and new error values were given. The MSE is used as an error function. Parameters for the AAC optimization algorithm: ant number = 50, iteration number = 10, pheromone parameter α = 1, parameter for the heuristic information β = 1, evaporation rate δ = 0,5. These are the optimization parameters determined with trial-error, which provide the best estimation results. The MSE changes through the optimization process for those two cases are presented in Figs. 11 and 12. Fig. 11. The performance indices of ANN training for the case of ABC assistance. In the figures, "backpropagation" shows the MSE of ANN with randomly ordered input data sets. "ABC" is the MSE variation of ABC-assisted ANN according to the iteration number, and "AAC" is the MSE variation of AACassisted ANN according to the iteration number.

VII. PERFORMANCE EVALUATION
In the feed forward back propagation algorithm model, in which 54 data points are used, 44 data points are used for training and 10 data points for testing. There are 45 inputs for ANN. The inputs have been selected by taking amplitudes greater than 0.05 as seen in Fig. 6. In the feedforward backpropagation application with a single hidden layer with 10 neurons improved by an individual Matlab code, the test error rate was obtained as MSE of 0.0117872. At this step, a randomly ordered data set was used.
ANN training was done with input data set sorted by ABC algorithm. The same operations were performed for the data set sorted with the AAC algorithm. Then, test data were applied to both ANNs, and MSE values were calculated for each one. Now we have the MSE values of three different ANNs. These values are presented in Table  II. In the table, ANN-E is the MSE of ANN obtained using the randomly sorted data set, ABC-E is the MSE of ANN obtained using the data set sorted using the ABC algorithm, and AAC-E is the MSE of ANN obtained using the data set sorted using the AAC algorithm.
It is shown that ABC-supported ANN presents the best failure diagnosis performance with a MSE of 0.0107476.

VIII. CONCLUSIONS
As in rotor fractures, rotor cracks cause an imbalance in bar currents and ultimately vibrations due to the moment induced in the rotor. This difference can be evaluated by various comparison methods, and it is possible to detect fractures in the rotor bars of the motor by acoustic measurements.
With this study, it was experimentally demonstrated that an ABC-assisted ANN can successfully diagnose rotor fractures. The test results indicate that ABC can improve the performance of the neural network faulty motor detection. If Table II is examined, it is seen that ABC-assisted ANN presents a 9 % improvement in rotor crack detection compared to ANN.
Finally, future work will consider phase current signals and different metaheuristic optimization algorithms in the detection of motor failure.

CONFLICTS OF INTEREST
The authors declare that they have no conflicts of interest.