Valve Fault Diagnosis in Internal Combustion Engines Using Acoustic Emission and Artificial Neural Network

This paper presents the potential of acoustic emission (AE) technique to detect valve damage in internal combustion engines. The cylinder head of a spark-ignited engine was used as the experimental setup. The effect of three types of valve damage (clearance, semicrack, and notch) on valve leakage was investigated. The experimental results showed that AE is an effective method to detect damage and the type of damage in valves in both of the time and frequency domains. An artificial neural network was trained based on time domain analysis using AE parametric features (AErms, count, absolute AE energy, maximum signal amplitude, and average signal level). The network consisted of five, six, and five nodes in the input, hidden, and output layers, respectively. The results of the trained system showed that the AE technique could be used to identify the type of damage and its location.


Introduction
The valves in an internal combustion engine play a significant role in engine performance.Moreover they are the most important components in the valvetrain and face high temperatures and gas pressure impulses.In the failure analysis of a valvetrain, valve failures represent the most common problems.The major causes of valve failure are distortion of the valve seat, deposits on the valve, small tappet clearance, burnt valve, erosion of valves, heat fatigue, pitting, breaks, and wear [1,2].In recent years, implementing condition monitoring techniques for internal combustion (IC) engines has led to reliable nonintrusive methods for engine diagnosis [3].
It is possible for turbulent flow, such as air flow through valves, to induce vibrations in their structures.The frequency of oscillation depends on the stiffness of the structure.The amplitude depends on mean flow velocity and on how close the forcing frequency is to the natural frequency of the structure.Lighthill's formulation of the aeroacoustic problem showed that the mixing region of a jet could be equated to a volume of quadrupoles with strength proportional to the stress tensor in the moving fluid.Such a mathematical formulation allows radiated jet acoustic power levels to be estimated and proportional to V 8 [4]: where   is the sound power (W),  is the diameter of the leakage hole, V is the average turbulence jet velocity (m/s), and  is the sound velocity in the fluid (m/s).Equation ( 1) clearly shows that flow velocity has the most effect on generating sound.The mass of air flow in/out to the cylinder is related to the air density, velocity, and area through which the gas flows: where ṁ is the mass flow (kg/s),  represents the area through which the gas flows (m 2 ),  relates the flow density (kg/m 3 ), and V is the flow velocity (m/s).In this study, the valve diameters were  ev = 30.52mm for the exhaust valve and  iv = 36.24mm for the intake valve.Vibration analysis is a well-established technique for fault diagnosis in rotating machinery, particularly for IC engines.Application of this technique to an IC engine is strongly affected by engine faults, auxiliary devices, and mechanical events in the engine [5].The high signal-to-noise ratio characteristic of the AE signal in comparison to vibration analysis has led to widespread interest in the last decade for AE application in the field of fault diagnosis of engines [6].
There are comprehensive investigation methods for fault diagnostic of VL using the AE method.All of them simulate the VL fault using valve clearance change [16][17][18][19][20][21]; however, significant causes exist besides clearance that eventuate in VL such as valve head crack and valve head notch.Using AE signals in VL to identify the type of fault is a new area of research that was addressed in this study.Three fault types (valve clearance, valve head crack, and valve head notch) were artificially simulated as follows.
(i) The valve clearance fault was simulated by a very small lift in valve.
(ii) Valve head crack fault was simulated by a semicrack on the exhaust/intake valve head.
(iii) The removal of a small amount of valve head material simulates valve head notch.
The test rig was a cylinder head for a spark-ignited engine.The aim of this investigation at this stage was to use AE signals in VL to identify fault types in an engine that is not running.

Test Rig and Measurement
System.The test setup consisted of the aluminium cylinder head of a small four-cylinder spark-ignited engine.Figure 1(a) shows the position of four wideband AE sensors on the cylinder head and a schematic of the test rig.Pressurized air entered the combustion chamber through the spark plug hole.Air flow was measured using an air flow meter with an accuracy of 1 L/min.

AE Data Acquisition
System.AE signals were acquired using four wideband sensors.The operating frequency was filtered from 100 kHz to 1 MHz.The AE sensors were coupled to the cylinder head using thin layers of vacuum grease.The signals were then amplified using PAC 2/4/6 preamplifiers before being recorded.Then the signals were digitized using a PAC PCI-2 (18 bit A/D, 40 MHz, 4 channels) AE data acquisition system.Raw data was acquired at a sampling rate of 2 MHz.AE sensors were mounted on the cylinder head near the spark plug hole on all four cylinders, as shown in Figure 1.

Simulated Valve Fault
Types.Three types of valve fault were artificially simulated.The test procedures were similar for all cases: inserting pressurized air through the spark plug hole and leakage measurement at the target valve port caused by its faulty valve.These faults are detailed below.
(i) Semicracked Valve.This fault was simulated by making a semicrack on the valve head using the wire cut method.The exhaust valve of cylinder number 1 and intake valve of cylinder number 2 were simulated using this method.The area of the removed surface of the valve head was 0.27 × 3.72 mm, as shown in Figure 2. (ii) Notched Valve.This fault was simulated by removing a slight amount of material from the seating face width of the exhaust valve, as shown in Figure 3. (iii) Valve Clearance.This fault was simulated by a 0.1 mm lift to the exhaust and intake valves of cylinder number 4. Figure 4 shows a healthy valve in cylinder 4 that has no leakage at zero lift.External lift simulated extra clearance.

Test
Procedure.The valves of cylinder 4 were healthy and they had no leakage at zero lift; therefore the exhaust and intake valves were separately lifted 0.1 mm using a lifter bolt.Lift was measured by a dial gauge with an accuracy of 0.01 mm as shown in Figure 1(b).At each step, one valve (exhaust or intake) was opened.Inlet pressurized air was controlled using a pressure regulator from 1 to 6 bar.At each pressure, all parameters were recorded (4 AE sensors, leakage rate, lift, and inlet air pressure).There was no lift for cylinders 1 to 3 because one valve of each cylinder was faulty and had leakage at zero lift.). Figure 10 shows the AE signals recorded for healthy valves with no leakage at zero lift and healthy valves from the same cylinders with no leakage at zero lift.Figures 6-9 show the AE signal acquired for 0.02 sec.

Results and Discussion
Equation (1) shows that the acoustic power level was proportional to V 8 , which means that acoustic power level was strongly dependent on leakage velocity; therefore, an increase in the leakage rate caused a very clear change in AE signal amplitude.In all cases, it was observed that AE signals were typically continuous and the amplitude of the AE signal increased as the inlet pressure increased.Figures 6-8 show the results for semicracked and notched valves in cylinders 1-3.It can be seen that the AE signals were very similar in wave form and amplitude; thus, it was not easy to distinguish between signals produced by different faults.
Figure 9 shows the results for the lifted valves in cylinder 4 and indicates that the signals have only a slightly different wave.It can be seen that the AE signals recorded from the exhaust and intake lifted valves are almost the same.
The signals in Figure 10 can be regarded as basic AE amplitudes for VL and are related to the existence of compressed air in the cylinder chamber, the test setup, environment, and background noise.

Frequency Domain Signal Analysis.
Frequency analysis of the AE signal from the VL shows the energy distribution of the signal in the frequency domain.Figures 11,12,13, and 14 show the power spectral density (PSD) of the signals shown in Figures 6-9.Frequency analysis of the AE signals demonstrates that gas leakage generates broadband AE signal.Moreover comparing Figure 15 for the spectrum of healthy valves with those in Figures 11-14 shows that there are some picks in the spectrum of faulty valves, while these do not exist in healthy ones.
In Figures 11-13, the frequency analysis of AE signals for semicracked and notched valves exhibited similar behavior, as would be expected from their waveforms in Figures 6-8. Figure 14 shows the results of frequency analysis of the AE signal from the exhaust and intake VL of cylinder 4. It can be seen that both exhaust and intake were very similar in their PSD analyses.Figure 15 presents the PSD analysis of the AE signal recorded for healthy valves with no leakage.
Although the frequency analysis detected the healthy and simulated faulty valves in the frequency domain, this method faced difficulties in showing a clear difference between the types of defective valves.For example, it was difficult to distinguish between the semicracked and notched valves.Separating the different types of faults requires an expert and systematic method.An artificial neural network (ANN) based on AE parameters extracted from raw AE signals was utilized for this stage.The parameters were the AE basic parameters: AE rms , count, absolute AE energy, maximum signal amplitude, and average signal level.

Artificial Neural Network for Determining Valve Fault Type
Artificial neural networks (ANNs) can identify correlated patterns between the input data set and corresponding target values.ANNs have good ability for prediction, pattern recognition, data compression, and decision-making [22].
Figure 16 shows a model of an artificial neuron.ANN was adopted to use five parameters (AE rms , count, absolute AE energy, maximum signal amplitude, average signal level) to identify valve fault type.The input layer consisted of five nodes corresponding to the five input features.The hidden layer had six nodes and the output layer had six nodes, of which five nodes represent the valve fault type and one node is reserved for the healthy condition.The output of each node is defined by the following sigmoid function: where net is the training data,  is the vector of weights,  is vector of inputs, and () is the transfer function.
The backpropagation method was used to learn the weights between each pair of nodes.The total data set for  1 shows the output condition.
Figure 17 is a sample that shows the efficiency of maximum amplitude, AE rms , and count in classifying the six classes of valve condition.As seen, some valve faults can be categorized visually, but other classes have very similar properties.A more efficient expert system is necessary to classify them with high performance.Table 2 presents the confusion matrix of the ANN method.The confusion matrix shows that ANN model was able to classify AE signals correctly.The most significant issue for a trained expert system is that the classifier may

Conclusion
This study constructed an expert system based on acoustic emission data to distinguish between valve fault types.Three valve faults (cracked valve, notched valve, and valve clearance) were simulated and used in the cylinder head of an internal combustion gasoline engine.It was shown that using time and frequency domain analysis only detects the difference between faulty and healthy valves.Healthy valves have no leakage and their AE signals have low amplitude.In order to distinguish between the fault types in valves, an ANN was used based on AE features.It was trained using five AE parameters (AE rms , count, absolute AE energy, maximum signal amplitude, and average signal level) for fault separation.These parameters had very good correlation with valve leakage and ANN was well trained using random segmented data.The results from this expert system showed good performance (over 92%) in distinguishing between faulty and healthy valves and fault types.

Figure 1 :
Figure 1: Test rig overview.(a) AE sensors positions and (b) schematic of measurement system.

3. 2 .
Time Domain Signal Analysis.Analysis in the time domain reveals the overall signal amplitude, periodic features, and AE signal type.Figures 6, 7, 8, and 9 show the raw AE waveform caused by VL at each cylinder.The data in Figures 6-9 are for the AE sensor nearest to the studied cylinder (sensors 1-4 for cylinder 1-4, resp.

Figure 4 :Figure 5 :
Figure 4: Healthy valves of cylinder number 4. Clearance fault was simulated by lift on valves.

Figure 6 :Figure 7 :
Figure 6: Raw AE signal due to VL through semicracked exhaust valve in cylinder 1.

Figure 8 :Figure 9 :
Figure 8: Raw AE signal due to VL through notched exhaust valve in cylinder 3.

Figure 10 :
Figure 10: Raw AE signal relate to healthy valves with no leakage in cylinder 4.

4 Figure 11 :
Figure 11: Power spectral density of raw AE signal due to VL through semicracked exhaust valve in cylinder 1 versus time at 1-6 bar by sensor number 1.
u t p r e s s u r e ( b ) Fre qu enc y (M Hz ) Power/frequency (dB/Hz)

Figure 12 :
Figure 12: Power spectral density of raw AE signal due to VL through semicracked intake valve in cylinder 2.
u t p r e s s u r e ( b ) Fre qu enc y (M Hz ) Power/frequency (dB/Hz)

Figure 13 :
Figure 13: Power spectral density of raw AE signal due to VL through notched exhaust valve in cylinder 3.

Figure 14 :
Figure 14: Power spectral density of raw AE signal due to VL through 0.1 mm lifted valve in cylinder 4. (a) Exhaust valve and (b) intake valve.

4 Figure 15 :
Figure 15: Power spectral density of raw AE signal for healthy valves with no leakage in cylinder 4.

Table 2 :
Confusion matrix of testing data for ANN.
Figure 17: Three features of AE signal in 4 bar inlet pressure.