Keywords

1 Introduction

Agricultural vehicles have an internal combustion engine, which currently can be monitored in real time using sensors. Engine monitoring can provide large amounts of data containing dynamic, combustion, fluid flow processes and major events of current engine conditions, with this data failure detection can be achieved, thus preventing further damage to the engine, as well as reducing engine performance [1].

Different studies have been carried out on the prediction, evaluation and diagnosis of machine failures using artificial intelligence techniques. Artificial Neural Network (ANN) and a humidity sensor were used for modeling sensor failures applying different temperatures [2]. The Wavelet transform and an ANN classifier to predict failures were used to evaluate bearing operating condition monitoring performance [3]. Several authors used sensors and diagnose motor failures, using the Wavelet packet transform, NN and spectrograms for the autonomous prediction of failures [4, 5]. In the same way there are predictive maintenance techniques based on vibrations to know the state of the engine [6]. Research on engine failure prediction mechanisms to reduce fuel consumption and exhaust emissions is vital and necessary [7].

Vibration in agricultural vehicle engines can be detrimental to their proper functioning, and vibration can affect the life of the engine [8]. Different studies focused on the reduction of maintenance costs, for example, for aircraft engines where they have an expenditure of 10.3% of the budget [9]. Thus, one of the objectives of predictive maintenance is to reduce this expense by using on-board sensors to monitor the health status of machine components. Predictive maintenance aims to identify potential malfunctions in advance, allowing a quick intervention before the actual problem arises. With condition-based maintenance, the aim is to schedule maintenance activities only when a functional fault is detected. Proactive maintenance puts primary emphasis on tracing all failures back to their root cause [10, 11].

To obtain measurements with different states, three levels of failure were simulated in the test bench that are in the injectors by adjusting the opening pressure and that directly affect the combustion and therefore the operation of the rest of the system.

2 Methodology

The data is obtained from a compression-ignition engine with three four-stroke cylinders that has an output of 51.2 HP. Four variables of engine operation at a single speed (idle) are studied. Vibration data is obtained using the ICP® 603C01 piezoelectric accelerometer. The signals are captured and processed using a DAQ NI 9250 data acquisition card with software designed in LabVIEW and they are stored in the computer to be processed. The characteristics obtained from each vibration signal are the number of data obtained per sample is 64,000 with a sampling frequency of 8,000 Hz.

The characteristics obtained from the vibration signal are inputs for the autonomous learning system where the different operating conditions of the motor are grouped. The considered conditions are: normal engine operation (BE) and failure in the opening pressure of the injectors with three levels of failure of opening of injectors 25% (ME25), 50% (ME50) and 75% (ME75). The variables considered have influence on the behavior of the engine, but they do not cause the engine to stop working (serious failure). Therefore, if the engine fails, the classification made in autonomous learning will group the defects according to the conditions entered: BE, ME25, ME50, ME75. Each feature for the different operating variables has 200 samples of 64,000 measurement points of vibration signals, where 80% of the samples are used to train learning and 20% to test and validate each case.

2.1 Decision Tree

It is a scheme that is used to decide and solve problems, to achieve this you must consider criteria and make decisions. Decision trees are a classification technique easy to interpret and use, which generate rules in the form of a tree, where the data set is divided into branches until segments of similar behavior are obtained depending on the target variable, and are used in decision making [12].

Decision trees allow to categorize a series of characteristics and parameters. In condition monitoring, and to guarantee the efficiency of this procedure, different operating conditions must be included, and ranges established for each of them [13].

This study will compare possible engine states mentioned above. The training of the decision tree was done using 5 statistical variables of the vibration signal (Median, Mode, Square Mean, Kurtosis and Asymmetry).

2.2 Obtaining Features

To obtain the characteristics is of vital importance in intelligent classification systems, several investigations have used different methods. This work is framed in the analysis of vibration, in the domain of time (the decision tree uses parameters of the time signal) and additionally a frequency analysis was performed. This methodology has been used by several studies [16]. For frequency analysis, the Fourier transform is traditionally used since it can convert a time domain signal into a frequency domain and thus be able to obtain information on frequencies and amplitude for raw signals.

The vibration data obtained in the time domain were transformed into the frequency domain. We recognized the dominant frequencies in each signal. The amplitude and frequency of each peak becomes the dominant characteristics based on frequency.

2.3 Extracted Characteristics

The extraction of the characteristics tries to filter the data with a high level of similarity [17, 18]. The vibration data obtained and characterized according to their state (BE, ME25, ME50 and ME75) are subjected to statistical calculations to determine the values to be used to perform the training, consequently, the important characteristics that contribute to the classifier to select patterns for prediction will be determined. For the different tests the engine revolutions will be considered according to the operating state (BE-850 RPM, ME25-850 RPM, ME50-900, ME75-1400 RPM). For the classification of failures, an intelligent system is created to predict failures in an engine according to the defined states (BE, ME25, ME50 and ME75).

3 Results

3.1 Feature Analysis

The vibration data obtained will be used in the classification of learning and subsequent failure prediction processes. Figure 1 shows the signal recorded for 0.8 s with an accelerometer, for each of the engine operating states: Good condition (BE) (a), ME25 (b), ME50 (c) and ME75 (d), where a considerable difference is observed compared to the change of engine speed (RPM). When causing a serious failure there are changes in the behavior of the engine and increases the RPM.

The characteristics for the conditions of the study (BE, ME25, ME50, ME75) are described below. Table 2 shows the values, in the frequency domain, extracted from the acquired vibration signal. There are 4 variables from which the characteristics were extracted directly from the sensor, and processed with the Fourier Transform, corresponding to the two amplitude values of the signal max. and their amplitude.

Fig. 1.
figure 1

Raw signals

Full size image
Table 1. Characterization values
Full size table

In Fig. 2 a) the area in which the characteristics will be taken is demonstrated to compare and carry out the research, in b) the characterization of the engine in BE status is shown where the main characteristics were selected, which in this study and for all the variables were 2 per sample. Considering (b) we proceed to analyze the other variables. In (c) where the same characteristics can be observed in the ME25 variable, but more pronounced peaks are noted which indicates that a failure may occur in the future, as shown in Table 1, the two characteristics are identical. In (d) presents the ME50 status a difference in the two characteristics is noted where it increases both in frequency and amplitude considering the Fig. 2 (b), having a considerable variation the damage to the engine is identified and in the same way the engine speed varies. The Fig. 2 (e) failure (ME75) a big difference can be noticed with the variable of good condition; this is because it is the strongest fault that can occur, and that the engine can start and work. In this way, characteristics are obtained to achieve adequate training patterns and predict failures of agricultural engines.

Figure 3 presents the scatter plot, which is a method of validation of learning where an efficiency of 95% is obtained and the distribution of the different characteristics obtained based on the variables studied is observed. It is not possible to differentiate or separate the points of the BE and ME25 classes that have similarity with respect to their variance and kurtosis characteristics. However, there is a clear difference of the classes in the characteristics of ME75, where the engine already has considerable failures in its operation, this failure would be the most severe that can be simulated.

Fig. 2.
figure 2

Vibration characteristics

Full size image
Fig. 3.
figure 3

Scatter plot

Full size image

4 Conclusion

An analysis of vibration signal characteristics (statistics of the temporal signal and the frequency domain) was carried out for the monitoring of failures in an agricultural vehicle engine. For this research, a methodology was proposed to diagnose engine failures based on three possible defects in the injectors that directly affect combustion. For the experimental part, a sensor located in the cylinder block as close to the combustion chamber was used that acquired the vibration data. The study is designed to distinguish four engine states: good condition (BE), poor condition varying 25%, 50% and 75% injector opening (ME25, ME50 and ME75).

Regarding the analysis of the characteristics of the signals in the frequency domain, we have worked with FFT analysis, on the frequency band delimited between 40 and 65 Hz, since it has been found that the characteristics are good indicators of state change, as shown in Fig. 3, where differences in amplitude values were observed for the case of BE and ME25 states with respect to ME50 and ME75. Regarding the characteristics of the time domain, they have been extracted from the measured vibration signals, and using trees decision it was concluded that, of all the statistical parameters calculated, the Quadratic Mean, Kurtosis and Asymmetry are the most interesting characteristics to monitor when trying to diagnose engine failures.