COMPARISON AMONG COMPUTATIONAL INTELLIGENCE METHODS FOR ENGINE KNOCK DETECTION. PART 2

Transcription

1 COMPARISON AMONG COMPUTATIONAL INTELLIGENCE METHODS FOR ENGINE KNOCK DETECTION. PART 2 ADRIANA FLORESCU 1, CLAUDIU OROS 1 2, ANAMARIA RADOI Key words: Bayes classifier, Engine knock, Fuzzy Kwan-Cai neural-network, Kohonen self-organizing map (SOM). The novelty this article brings is the use of neural networks and neuro-fuzzy modeling to take the engine knock detection based on pressure and vibration samples taken from an internal combustion engine to the next step from previous achievements, the ultimate goal being higher detection s and also faster response times than the classical non neuro-fuzzy methods. Work started from the theoretical works available in the domain, algorithms were adapted for the case in hand, then the study was led on how they would be affected by the variety of situations that occur in internal-combustion engines with the scope of real-time applications. Following the experiments, results were finally compared showing significantly greater knock detection time improvements than other methods employed so far. 1. INTRODUCTION Engine knock (also known as knocking, self-combustion, detonation, spark knock or pinging) in spark-ignition internal combustion engines occurs when combustion of the mixture of fuel and air in the cylinder starts off correctly because of the ignition by the spark plug, but one or more pockets of the mixture explode outside the normal combustion front. Part 1 of the article presented the importance of engine knock detection, a brief mathematical background of the three tested artificial intelligence methods, experimental results for fuzzy Kwan-Cai neural network and conclusions about this first method [1 5]. Part 2 continues Part 1 with experimental results using Kohonen self-organizing map neural network and Bayes classifier, then a large comparison is made and final conclusions are drawn. We mention that because comparisons and conclusions are drawn from both Part 1 and Part 2, we continued in Part 2 the numbers of figures and Tables from Part 1. 1 Politehnica University of Bucharest, Sp. Independenţei, 313, Sector 6, Bucharest, Romania; s: adriana.florescu@upb.ro, orosclaudiu@gmail.com 2 Ecole Politechnique Federale de Laussane, Route Cantonale1015, Lausanne, Switzerland; rdi_ana@yahoo.com Rev. Roum. Sci. Techn. Électrotechn. et Énerg., 57, 1, p , Bucarest, 2012

2 2 Computational intelligence methods for engine knock detection EXPERIMENTAL RESULTS WITH KOHONEN SELF-ORGANIZING MAP NEURAL NETWORK The Kohonen-Self Organizing Map (SOM) has a sepa learning stage taking place before the detection process begins and being composed of epochs. After the learning stage has ended it does not need to be repeated and the processing of the test batch begins. For this neural network three sizes of neural maps were used nine, one hundred and four hundred neurons, as shown in Tables 4, 5, 6. They were tested on both pressure and vibration samples. Table 4a contains only the pressure sample detection results for the small vector database using the one hundred neuron configuration. By keeping the number of epochs constant at 100 and the learning at 0.2 and by means of a variation of the neighborhood size from 90 down to 10, we obtained the following spike values: a detection of 80% marked bold-italic for the (100; 0.2; 83) group and the maximum value of the detection for the small database 82.85% marked bold for the (100; 0.2; 82) combination. Table 4a Pressure sample detection s for small database Neigh- Detection Learning No. bor- hood Epochs neuron size [%] No. neuron Table 4b Pressure sample detection s for large database Neigh- Learninborhood Epochs size Detection [%] Table 4b contains the pressure sample detection s using the large database. From the start, using the nine neuron map, an important fact appears: the nine neuron map can not cope with the large database due to the small number of neurons that have to remember a large amount of samples, leading to confusion and very low detection s. The variation methods are the same ones as in the complete version of Table 4a but, even by varying each of the parameters and keeping the other two constant, we can not obtain a spike value higher than 29.78% marked italic, value resulting from the combination (100; 0.4; 5). Performing the same variation techniques as in Table 4a, the maximum value for the detection

3 82 Adriana Florescu, Claudiu Oros, Anamaria Radoi 3 in Table 4b results of 90.57% from the (400; 0.2; 400) and (500; 0.2; 400) combinations both marked bold, with lower but not less important spikes of 89.66% for (100; 0.2; 400) and (100; 0.3; 400) marked bold-italic. Table 5a contains the vibration sample detection s for the small database. The same variation methods as those in Tables 4a and 4b were used for the exact same values. The one hundred neuron network encounters its top value of 80% for the (100; 0.2; 95) combination and also a smaller spike of 74.28% for (100; 0.2; 60). The four hundred neuron network tops out at the 82.85% detection for the (300; 0.2; 400) combination of parameters. The same marking methods as in the previous tables were also used here and in the following ones. Table 5a Vibration sample detection s for small database Neigh- Detection Lear- No. bor- Epochs ning neurons hood size [%] Table 5b Vibration sample detection s for large database Neigh- Lea- Bor- Epochs rning hood size No. neurons Detection [%] The large database results for the vibration sample vectors are found in Table 5b. These values have come from the same methods of testing and values used in Tables 4a, 4b and 5a. As in the case of the complete Table 4a (from which only the one hundred neuron section has been presented in this paper), the nine neuron network in the complete Table 5b is not suited for working with such a large database, the network becoming confused. This shows in constant results under 50% which can not be taken into account as valid experimental results. These values can only be used as examples of exceptional cases. The one hundred neuron network section presented in Table 5b obtains a maximum detection of 81.76% for combinations (100; 0.2; 50), another important value over 80% being of % for (100; 0.2; 70). The four hundred neuron network tops out at 89.66% for combinations (100; 0.2; 250) and present other important values of 89.36% for (100; 0.2, 325) and of 89,05% for (100; 0.2; 375).

4 4 Computational intelligence methods for engine knock detection Table 6 represents the average detection times using both pressure and vibration vectors for both small and large databases. With values of s (small database) and s (large database) the pressure samples obtain smaller detection times than the vibration samples with s (small database) and s (large database). This situation is representative for the four hundred neuron network, this also being the slowest solution but with the highest detection s. The nine neuron network, even though it has the best detection times, can not be taken into account as a real application because it is not able to cope with large database. The one hundred neuron network is the best compromise between detection speed and detection s as shown in Table 6. Table 6 Pressure and vibration average detection times for both small and large sample databases Small database Large database Average detection Pressure Vibration Pressure Vibration time [s] samples samples samples samples SOM with 400 neurons SOM with100 neurons SOM with9 neurons As with the previous described algorithms, the SOM results shown in Tables 4 and 5 that an increase in the sample group size (training set case) will lead to an increase in detection s. In this case, the two sepa groups are not sepad by big detection gaps. As in theory, the experimental results in Tables 4, 5 and 6 show that with the increase in neurons there is an increase in detection s but a decrease in detection times because more neurons translate to more detail that can be remembered, so the distinction between knock and non-knock situations can be more precisely done therefore a compromise must be made. Being interested not only in obtaining high detection s but also detection times that would be coherent to the task at hand (samples must be processed in under an engine cycle so the modifications can be brought to the next one), the one hundred neuron map seems to be the best option from the three methods tested. The nine neuron map, even if it produces very high detection times, has a very poor detection in both pressure and vibration groups making it useless for any further applications. The four hundred neuron map presented the highest detection s for this neural network, values that are a little bit smaller than the Fuzzy Kwan-Cai but with detection times very similar to it, the only difference being that the SOM

5 84 Adriana Florescu, Claudiu Oros, Anamaria Radoi 5 needs sepa training. In this case, looking at the detection times in Table 6, the SOM does not seem to make any difference between pressure and vibration signals, the medium detection times showing very small variations. There is a small difference in detection s between pressure and vibration samples; the SOM seems to handle both models very well. 3. EXPERIMENTAL RESULTS WITH BAYES CLASSIFIER The Bayes Classifier, as described by its name, is not a neural network but has been included in this paper as a basic reference point for the evaluation of the two neural networks. It uses a method of calculating the minimum distance from a sample to one of the knock or non-knock class centers-classes that are considered Gaussian by nature. That is why it presents the worst detection times, as shown in Table 8. Table 7a represents the combined pressure and vibration detection s status for the small database. The way the testing has been done for this algorithm is by progressively growing from a small comparison group (the batch of samples chosen to represent the known classes for testing) versus large test group situation, to a large comparison group versus small test group situation. The process starts out with a balance of 11 training vectors and 90 testing ones, which leads to a detection starting from 65.50% for pressure and 55.55% for vibration and grows (for training vectors) versus shrinks (for testing vectors) in a progressive way to 85 training vectors and 16 testing vectors, leading to a detection ending at 43.75% for pressure and 81.25% for vibration. An interesting detail can be observed in Table 7a: the pressure vectors seem to present a constant state even though more and more are added to the learning group every time the detection s stay approximately between 50% and 72.50%, the last value being the highest pressure detection. The change of state occurs at the end of Table 7a where we can observe a decrease in the learning for the combinations of (80 training vectors; 21 testing vectors) with a detection of 42.85% and (85 training vectors; 16 testing vectors) with a detection of 43.75%. This decrease is due to the inclusion in the learning group of vectors that are radically different from their stated class; therefore, the knock or non-knock distinction can not be made. In the case of the vibration sample vectors the progression is of almost uniform growth from 55.55% to 81.25%, the last being also the maximum detection for the small database experiment.

6 6 Computational intelligence methods for engine knock detection Table 7a Pressure and vibration detection s for small database Training vectors Test vectors Pressure detection [%] Vibration detection [%] Table 7b Pressure and vibration detection s for large database Training vectors Test vectors Pressure detection [%] Vibration detection [%] Table 7b follows the same type of progression, only that the large database is used for both pressure and vibration samples. The progression goes from a combination of (371 training vectors; 629 testing vectors) with a detection of 93.64% for pressure and 90.30% for vibration samples to a combination of (671 training vectors; 329 testing vectors) with the maximum detection achieved in this table of 95.44% for pressure samples and 92.40% for vibration samples. Within this progression it can be seen more clearly that the pressure samples are very cohesive in nature and that, given enough samples, the algorithm goes past the problems it has with radically different sample vectors, maintaining a detection over 90% in every case. Table 8 represents the average detection times for both the small and large databases using both pressure and vibration samples. Table 8 Pressure and vibration average detection times for both small and large sample databases Average detection time [s] Pressure Vibration Small sample database Large sample database Being a simple comparative algorithm, we can see in Table 8 that an increase in the database size leads to a slowing down of the process because the comparison must be made with more vectors. In the case of the small database, pressure vectors

7 86 Adriana Florescu, Claudiu Oros, Anamaria Radoi 7 are detected faster (0.0287s) than vibration samples (0.0297s). The large database experiments lead to almost equal average detection times between pressure (0.0948s) and vibration (0.094s) samples, with a tendency to better recognize vibration samples. There is little relevance in the detection s for the small sample group, even though a small variation between pressure and vibration can be seen. The increase in detection s due to a bigger knowledge database can also be seen from Table 7. The greatest importance of the Bayes Classifier in this paper comes from its great sensitivity to change. Given a big enough knowledge database that is also very coherent in the nature of its classes, the detection s go up and can be comparable to the neural networks but at a great cost in speed. 4. COMPARISON AMONG THE THREE TESTED METHODS The first discussion will be based on the database size point of view. As we can see from Fig. 4 and Fig. 5 that summarize results in Tables 1 and 2 from Part 1 and Tables 4, 5 and 7 from Part 2, the size of the learning, training or comparison database is very important in the good functioning of all three tested algorithms. Kwan Cai SOM Bayes 68 82,85 68,57 Kwan Cai SOM Bayes 82,85 81,25 48 Detection s (Pressure samples) [%] (a) Kwan Cai SOM Bayes 95,44 93,4 90,57 Detection s (Vibration samples) [%] (a) Kwan Cai SOM Bayes 93,4 92,4 89,66 Detection s (Pressure samples) [%] (b) Detection s (Vibration samples) [%] Fig. 4 Pressure sample detection s using Fig. 5 Vibration sample detection s using the the small database (a) and the large database (b) small database (a) and the large database (b) for the for the Kwan-Cai, SOM neural networks and the Kwan-Cai, SOM neural networks and the Bayes Bayes Classifier. Classifier. An increase in the database size from one hundred to one thousand sample vectors will lead to a minimum increase of ten percent in the detection s. For (b)

8 8 Computational intelligence methods for engine knock detection the small database, the Fuzzy Kwan-Cai neural network obtains maximum detection s for the pressure samples at 68% that are higher than the ones for vibration samples at 48%, but after using the large data set the maximum pressure and vibration detection s become equal at 93.40%. The difference in detection s for the pressure and vibration samples using the small database shows that the pressure samples are more coherent and therefore easier to classify. The same evolution as shown by the Fuzzy Kwan-Cai is also true for the Kohonen Self- Organizing Map (SOM). Even more so, the increase in learning database size will lead to a theoretical increase in the detection of the Bayes Classifier. The second discussion will be based on the detection point of view. As shown in Fig. 4 and Fig. 5, the Bayes Classifier seems to show the best detection s. Its fault is that it needs large amounts of comparison data in order to create classes that are comprehensive enough. Out of the three algorithms tested in this paper, it is also the less stabile due to the fact that it calculates distances to the center of the comparison classes. If these classes are not well defined and sepad, the detection s fall dramatically. This can be seen in Table 7b. The Fuzzy Kwan-Cai obtains the highest detection s of all three algorithms these being valid detection s that are not influenced by the nature of learned vectors leading to the great stability of this method. The learning method used employs the automatic generation of learning classes as it goes through the sample set. The fuzzy logic creates a more organic representation of the knowledge classes than the boolean one. The Kohonen Self-Organizing Map (SOM) presents the second highest detection s and a more controlled and stabile learning and training environment then the other two algorithms. Because the learning is done prior to the start of the testing process and in repetitive epochs, the neural network has the chance to go through the data set again and again until a complete image is formed. The two neural networks show no considerable preference between pressure and vibration samples and present high stability to drastic variations in training samples which in a non-neural method could cause a decrease in detection s. The nature of these types of signals and their differences are outlined by the Bayes Classifiers sensitivity to unclear classes and the way in which the Fuzzy Kwan-Cai neural network works by showing the internal structure of the classes. The third discussion will be based on the detection time point of view. As present in Fig. 6 and Fig. 7 that summarize results in Table 3 from Part 1 and Tables 6 and 8 from Part 2, it is clear at first glance that the neural networks are far superior to the normal non-neural classification algorithm. The Bayes Classifier obtains the longest detection times due tot the process of comparing each new vector to the knowledge classes. The best, valid, detection times are shown by the

9 88 Adriana Florescu, Claudiu Oros, Anamaria Radoi 9 Kohonen Self-Organizing Map with the one hundred neurons configuration. This configuration, given optimization of the code, can lead to detection times coherent to the engine combustion cycles in which the knock detection needs to take place. The second best detection times that are also very close to one another belong to the Fuzzy Kwan-Cai and SOM with the configuration of four hundred neurons. These two algorithms also show the highest detection s from the methods tested in this paper. In a real-time application there should not be any problem with the SOMs sepa training stage because it would be performed only once inside the factory. It is clear from the information presented in this paper that the best detection s correlated to very good detection times belong to the Kohonen Self- Organizing Map with a configuration of one-hundred-neurons. The SOM with a configuration of four hundred neurons obtains results almost similar to the Fuzzy Kwan-Cai. The Bayes Classifier is very useful for showing the nature of the knock and non-knock classes, how well they are defined and sepad due to its sensitivity to drastic variations in sample vectors. Kwan Cai Small SOM 400 Small SOM 100 Small Bayes Small 0,0287 Kwan Cai Large SOM 400 SOM 100 Large Bayes Large 0,0297 0,0052 0, ,0023 Detection times (Pressure samples) [s] (a) Kwan Cai Large SOM 100 Large SOM 400 Bayes Large 0,0297 0,0022 0, ,0027 0,0022 0, ,0027 Detection times (vibration samples) [s] (a) Kwan Cai Large SOM 100 Large SOM 400 Large Bayes Large 0,094 0,0046 0, ,0028 Detection times (vibration samples) [s] (b) Fig. 6. Pressure sample detection times using the small database (a) and the large database (b) for the Kwan-Cai, SOM neural networks and the Bayes Classifier. Detection times (Vibration samples) [s] (b) Fig. 7. Vibration sample detection times using the small database (a) and the large database (b) for the Kwan-Cai, SOM neural networks and the Bayes Classifier. From a real-world application point of view, in order to further maximize detection s, it is clear that a parallel process composed of a pressure-vibration analysis and detection becomes necessary, based on the experimental results. Due to the developments in digital signal processing (DSP) technology, the parallel process would not lead to an increasing detection times.

10 10 Computational intelligence methods for engine knock detection CONCLUSIONS This final chapter contains general remarks due to the fact that detailed and accu conclusions have been already widely presented in chapter 4. According to experimental results graphicaly compared in Figs. 4 7, the best numerical values are: 95.44% for pressure detection obtained with Bayes classifier and 93.40% for vibration detection obtained with fuzzy Kwan-Cai neural network (both on the large database). Tjhe best pressure detection time of s was obtained with SOM 400-small database and the best vibration detection time of s was obtained with SOM 400-large database. The experiments performed have led to results that prove the superiority of the neural methods in contrast to the normal classification the situation being looked at from a -time point of view as seen in Figs The difference between the neural and non neural methods is represented by an average scale factor of 0.001s in favour of the neural. This superiority should be seen also from a stability to errors point of view as seen in Table 7b where a stray vector can distort the judgement of the non neural Bayes Classifier so that detection s fall. Suggestions for real-world applications were made in the prior chapter leading to further optimizations around the strengths and weaknesses of each algorithm. ACKNOWLEDGEMENTS This work was supported by CNCSIS UEFISCSU, project number PNII IDEI code 1693/2008. Received on 15 March 2011 REFERENCES 1. H. Li and G. A. Karim, Knock in spark ignition hydrogen engines, International Journal of Hydrogen Energy, 29, 8, pp , Anamaria Radoi, V. Lazarescu, Adriana Florescu, Wavelet Analysis to Detect the Knock on Internal Combustion Engines, Rev. Roum. Sci. Techn. Électrotechn. et Énerg., 54, 3, pp , J. Erjavec, Automotive Technology: A System Approach, 5th ed., Delmar Cengage Learning, 2009, pp H.N. Gupta, Fundamentals of Internal Combustion Engines, Prentice-Hall of India Private Limited, New Delhi, 2006, pp Octavian Grigore-Müler, A Neural Controller For On Board Tracking Platform, Rev. Roum. Sci. Techn. Électrotechn. et Énerg., 54, 2, pp , 2009.