**4.3. Classification results**

In this chapter, we focused on 10 conditions of AUV, which are listed in **Table 1**.

There are two types of these AUV conditions. One of them is failure situation and the other is functional condition and both include five motion status of the AUV. We carried out 20

**Figure 10.** The AUV's fault simulator diagram.

**Figure 11.** The wave maker of the NCKU's ship model towing tank.

Design and Estimation of an AUV Portable Intelligent Rescue System Based on Attitude Recognition Algorithm http://dx.doi.org/10.5772/intechopen.79980 63


**Table 1.** Activities performed in this experiment.

experiments for capturing data and verifying classifier. The data from 17 experiments were adopted in the training program of the recognition scheme; the data obtained from the other experiments were used for testing the recognition performance. Note that, since the sampling frequency is 27.5 Hz, the total number of the short-term and long-term samplings for each activity of each experiment is 550 and 2200, respectively, which means 20 seconds per short-term window and 80 seconds per long-term window. The feature extraction of this chapter was based on 50% overlapping windows using 550 samples of window sizes to avoid information loss at the boundary of a single window. The dimension of a feature vector was 45 (an accelerometer 3 axes 9 features + a gyroscope 3 axes 6 features). **Figure 12** illustrates the first 2200 data of accelerations and Euler orientations collected from the first experiment. The selected features of sensor's data enabled effective recognition of the conditions and were suggested for BPN training procedure. A computation program adopted the input features and activated the feature classifier learning procedure with the BP algorithm, and outputted the results to short-term classifier. Then, an AUV condition was distinguished by a long-term classifier, of which the input is from the short-term classifier to raise the accuracy of failure recognition. The number of neurons in each hidden layer is 4, 6, and 7 for the feature classifier, short-term classifier, and long-term classifier, respectively, and the number of epochs is 700 for each neural training. The BPN classifier was trained on the training data set and tested on the test set which are from the experiment values. The classifier was created by neural network toolbox of MATLABTM for practical implementation and for validating the proposed model.

After building up our prediction algorithm, we apply our chosen prediction algorithm on our new test set which is from the real signal of AUV, in order to have an idea about the algorithm's performance on unseen data. The confusion matrix measured in the real AUV test is shown in **Table 2**, which recorded the results from 20 times experiments on each condition of AUV. We have implemented in two different ways under MATLABTM environment. In the first, we conducted in our proposed classifier system with feature extraction and the results indicate that the AUV failure detection on the average 97% of the time, and a successful functional condition accuracy of 93% is achieved. Second, we chose a classifier in ANN learning algorithm without feature extraction for comparing with the classifier that we proposed. The performance indicated

**Figure 12.** The accelerations and Euler orientation of the first experiment.



**Table 2.** Confusion matrix for all the testing experiments.

that the ANN classifier without feature extraction performs poorer than our proposed classifier. From the confusion matrix, we can know that the malfunction and functional conditions are not easy to be confused. However, the motions within functional condition may be misclassified between each other, because these activities contain similar amplitude peaks and waveforms at the AUV.
