Noise evaluation of MF285 and U650 tractors by using Adaptive Neuro-Fuzzy Inference Systems (ANFIS) method

2.1. Data acquisition

This study was conducted in Razi university of Kermanshah, Iran. Medium textured soil is dominant in the region. Personal noise exposures of the agricultural tractor operators were measured during tractor operations with MF285 and U650 tractors. Table 1 shows some information about the tractors, and Table 2 about other agricultural machinery used in this study.

In this study noise of MF285 and U650 tractors has been investigated. These tractors are the most commonly used tractors in Iran and they usually are used without cabin. This is the reason of choosing this kind of tractors for this study.

The sources of noise of any vehicle can be separated to 3 categories. First, the aerodynamical noise which comes from the effect of wind on machine. This type of noise is not as important as the other types of noise in the tractors, because the field speed of tractors usually is not so high, which affects on the noise of tractor. Second, the frictional noise which comes from the movement of mechanical part and frictional force between them. The last type of sources of noise of vehicle is noise which comes from combustion in the engine.

Before testing, all the necessary technical visits were carried out on the tractors. Test site was prepared and maintained according to ISO [19, 20] sound measurement standard. The test area consisted of a flat open space free from obstacles and the effect of signboards, buildings and hillsides for at least 15 m from measurement zone. The suggested wind speed and other climate limitations were kept in mind during measurements. The background noise was at least 30 dB(A) lower than that for the tractor. Fig. 2 shows the dimensions of the area in which the tractor noise measurement was carried out.

Fig. 1Dimensions of the test site

Here R stands for the distance from the obstacles to the measurement zone; $L$ and $W$ are the length and width of measurement zone, respectively. The minimum values of $R$, $L$ and $W$ were 40, 20 and 8 m, respectively.

It is important that the measuring equipment be properly selected to monitor and measure sound properties. When we have a basic situation in which we need to assess the severity of environmental noise, we may need to measure only the overall sound pressure level or the A-weighted level, using a simple sound level meter [21].

In this study, a SLM (TES-52 SOUND LEVEL METER) was mounted 1.2 m above the ground surface and 5 m away from the center of the tractor path way in a horizontal position and pointed in the direction of travel. To measure the noise near the operator’s ear, the microphone of dosimeter (Model TES-1354/1355) was attached to the operator clothes.

The SLM and dosimeter calibration was performed by using TES-1356 Sound Level Calibrator before data gathering. Most sound level meters have three “weighting” networks, called the A-, B-, and C-scales. Originally, the A-scale was designed to correspond to the response of the human ear for a sound pressure level of 40 dB at all frequencies. The B-scale was designed to correspond to the response of the human ear for a sound pressure level of 70 dB at all frequencies. The C-scale was approximately flat (constant) for frequencies between 63 Hz and 4000 Hz. The A-scale is widely used as a single measure of possible hearing damage, annoyance caused by noise, and compliance with various noise regulations. The sound levels indicated by the A-scale network are denoted by LA, and the units are designated dB(A) [21].

It was needed to choose the range of variables before trying to perform the tests, i.e., forward speed, gear ratio and the microphone position. The ranges of variables were selected to sustain most of operational condition of MF285 and U650 tractors regard to agriculture operation. After determining the range of variables, table test was developed based on completely randomized split plot factorial test (shown in Table 3). All tests were done with 5 replications.

The field speed of tractor in different operations had been selected according to Table 4 [22].

The data obtained in this study were analyzed using SPSS 19.0 software. Variance analyses for sound pressure level (dB) at A-weighted equivalent sound pressure level (dB(A)) were done for each working condition. Duncan test was used for multiple comparisons of mean values of sound pressure level.

2.2. Development of ANFIS model

ANFIS is an adaptive network that permits the application of neural network topology and fuzzy logic. It not only includes the characteristics of both methods but also eliminates some disadvantages of their lonely used case. ANFIS uses the learning ability of ANN to define the input–output relationship and construct the fuzzy rules by determining the inputs structure. The system results were obtained by thinking and reasoning capability of the fuzzy logic.

The ANFIS network (Fig. 2) works as follows: let x and y be the two typical input values fed at the two input nodes, which then transforms those values to the membership functions (say bell-shaped) and give the output as follows [note in general, $w=$ output from a node; $\mu =$ membership function; and $Mi$ and $Ni=$ fuzzy sets associated with nodes $x$, $y$ in Eq. (1)]:

where $a_1$, $b_1$ and $c_1=$ changeable basis parameters. Similar computations are carried out for the input of $y$ to obtain ${\mu }_{Ni}\left(y\right)$. The membership functions are then multiplied in the second layer, e.g.:

Such products or firing strengths are then averaged:

Nodes of the fourth layer use the above ratio as a weighting factor. Furthermore, using fuzzy if-then rules produces the following output (an example of an if-then rule is: if $x$ is $M_1$ and $y$ is $N_1$, then $f_1=p_{1}x+q_{1}y+r_1$):

where $p$, $q$, and $r=$ changeable consequent parameters. The final network output $f$ is produced by the node of the fifth layer as a summation of all incoming signals, which is exemplified in the Eq. (4). The description of the learning algorithm is given by Jang and Sun [23].

The following scenarios are considered in building the ANFIS model (Fig. 3) with the inputs and output shown in the network. A computer program (MATLAB code) was developed to perform the analysis, and can be obtained from the corresponding author.

Fig. 2ANFIS network architecture

Fig. 3The ANFIS architecture for sound pressure

The performance of ANFIS model in training and testing sets is validated in terms of the common statistical measures; $R^2$ (coefficient of determinationwhich presents the degree of association between predicted and true values), $RMSE$ (which is preferred in many iterative prediction and opti-mization schemes), $SSE$, and $MAE$ (+ or −, which is a parameter commonly understood in engineering applications and which considers algebraic difference between predicted and true values):

where $Y_{estimate}$ is the estimated data, $Y_{observed}$ is the observed data. These statistical measures provide information on the strength of linear relationship between the observed and the estimated values. If the model is “perfect”, $SSE$ is zero and $R^2$ is 1. If the model is a total failure, $R^2$ is zero.