Defect recognition of aluminum alloy cracks based on probabilistic neural network

Bei-tao GUO, Xian ZHANG, Zhen-bo WANG

(Shenyang University of Chemical Technology, Shenyang 110142, China)

Abstract: The artificial neural network method is applied to crack defect recognition of aluminum alloy workpieces to overcome the limitations of traditional manual recognition, thereby improving the accuracy of crack defect recognition. By designing and constructing a water immersion ultrasonic detection system, waveform data of ultrasonic detection defects are obtained, and feature extraction is performed on the collected defect waveform data to extract useful feature information, and the wavelet denoising processing is used to input a probabilistic neural network as a characteristic signal, and the network training is performed to realize intelligent recognition of different crack sizes. Experimental results show that the method can improve the accuracy and detection efficiency of crack defect size identification, and has a good application prospect.

Key words: Ultrasonic testing, Feature extraction, Probabilistic neural network, Crack defect

Aluminum is alloy widely used as a light-weight, high-strength, and corrosion-resistant material. During the casting and extrusion of aluminum alloy parts, the surface of the part usually has defects such as cracks, slag inclusions, and crushing and most of them are crack defects. The current manual detection method can only observe the approximate size of the defect with the naked eye from the time-domain waveform, and cannot accurately determine the actual size of the defect, so it causes great difficulty in analyzing and evaluating the defect. The neural network recognition method can make up for the shortcomings of manual detection and improve the accuracy of detection. With its strong learning ability and organizational ability, neural networks are widely used in artificial intelligence, signal processing, pattern recognition, and other research areas [1-3]. At present, BP and RBP networks are mainly used to classify defects [4-5]. However, due to the difficulty in determining the structure of BP and RBP networks and the slower training speed. Therefore, the probabilistic neural network is applied to crack defect identification of aluminum alloy workpieces, and experimental analysis proves that the probabilistic neural network is feasible and effective to identify crack defects of aluminum alloy workpieces. The results are more reliable.

1 Water immersion ultrasonic detection system

The water immersion ultrasonic flaw detection structure is shown in Fig.1. Its working principle is: placing the aluminum alloy workpiece to be tested on a detecting platform in a water tank. Under the control of the control system, the robot arm holding the probe moves along theXandY-axis to the set coordinate position. Then, theZ-axis movement unit controls the downward movement of the probe. When the probe reaches a certain distance from the workpiece, the probe stops moving downward and starts to detect damage defects, thereby completing the acquisition of the defect waveform. After the collected data is processed, it is input into the neural network for training as feature samples to prepare for pattern recognition.

1.Aluminum alloy frame; 2.Inspection platform; 3.Three-coordinate Y-direction guide; 4.Three-coordinate X-direction guide; 5.Three-coordinate Z-direction guide; 6.Probe holder; 7.Water tank

2 Feature extraction and probabilistic neural network algorithm(PNN)

2.1 Feature extraction

Defects of aluminum alloy crack are influenced by material, acoustic speed, acoustic resistance, scattering, and so on. The echo wave of defects is a very complex signal and representative features should be searched for recognition as input vectors for Probabilistic neural network(PNN). If defects features are extracted, it will be obvious, which will greatly reduce the workload and calculation difficulty for the Probabilistic neural network.

It can be seen from the waveform that the characteristics of the aluminum alloy sample defects are concentrated on a local part which is too dense, it should be classified into a feature group and a vector composed of many features. By calculating the correlation between each of the collected waveform features, we can distinguish whether the features we have selected can be separated and extracted[6]. Therefore, the similarity analysis is performed on the acquired waveform data, calculate the correlation coefficient between the target value and the featurexij,and the formula is as follows:

(1)

2.2 PNN neural network algorithm

Probabilistic neural network is an extended neural network based on radial basis function neural network[7]. The essence of PNN is a parallel algorithm based on Bayesian decision theory[8]. Because it uses the probability density function estimation of Parzen window and the posterior probability output of nonlinear operator mode, it makes the structure simple, fast convergence, and strong classification ability[9]. It is particularly suitable for solving pattern classification problems. Its advantage lies in the use of linear learning algorithms to complete the work done by previous nonlinear learning algorithms while maintaining the high precision of nonlinear algorithms.

The PPN network structure can be divided into input layer, sample layer, summation layer, and output layer[10-11], as shown in Fig.2. The input layer receives the feature signals extracted from the sample and then transmits the feature signals to the nodes of the sample layer through the network. The second layer is the sample layer in which the number of nodes is equal to the sum of the number of training samples. The output value of the sample layer represents a- similar level. The summation layer is the accumulation of the same type of probabilities input from the sample layer to obtain estimates of various probabilities. The output layer outputs the result.

Fig.2 Probabilistic neural network structure

The summation layer is the accumulation of the same type of probabilities input from the sample layer to obtain estimates of various probabilities. Its output is:

(2)

The output value of the sample layer represents its similar level. The output value is then transferred down. The output layer outputs the result, which has only one result, the one with the highest probability, the output is 1, and the other outputs are all 0.

The output of each node is:

(3)

Where:xis the residual in the input layer,αis the type of malfunction,raiis the i-th training vector of a certain malfunction type, andδis the smoothing factor.

PNN absorbs the advantages of radial basis neural network and classical probability density estimation principle. Compared with the traditional feedforward neural network, PNN has more significant advantages in pattern classification[12].

2.3 Neural network simulation and comparison

First, perform random data testing on the traditional linear neural network(BP) and the error curve is shown in Fig.3. The error curve is obtained through the simulation of the PNN after convergence and PNN training error graph is shown in Fig.4

Fig.3 Linear neural network convergence accuracy

Fig.4 Probabilistic neural network convergence accuracy

Compare the convergence speed and accuracy of the traditional linear neural network and the probabilistic neural network, the probabilistic neural network has rapid convergence characteristics, so it is particularly suitable for situations that require real-time classification. The probability nature of the neural network itself also has the same kind of defect classification work in this experiment.

3 Recognition and diagnosis of aluminum alloy crack defects

3.1 Data acquisition

The experiment is based on the CTS-04PC ultrasonic flaw detection system. Water immersion ultrasonic probes are used to detect crack defects of aluminum alloys of various sizes. An ultrasonic water immersion focusing probe is used and its type is 5P20. The piezoelectric chip of the ultrasonic probe is circular, its diameter isD=Φ20 mm, and its frequency is 5MHz. In order to test the accuracy and effectiveness of the PNN network in identifying cracks, five common sizes of artificial cracks of 0.3 mm, 0.4 mm, 0.5 mm, 0.6 mm, and 0.7 mm were prepared. As shown in Fig.5.

Fig.5 0.3～0.7 mm crack defect pattern

The ultrasonic probe inputs the collected analog signal to the Industrial Personal Computer (IPC), and after processing the discrete number table of the analog signal is derived. The collected defect waveform is taken as an example of a crack size of 0.3 mm, as shown in Fig.6.

Fig.6 0.3 mm crack defect detection waveform

3.2 Feature signal extraction

In order to improve the learning speed and recognition rate of PNN, the data needs to be processed to select the feature information that has a greater impact on the waveform. Calculate the similarity between each group of waveforms, and use the correlation coefficient as the basis for discrimination. Only the non-correlated two sets of data can effectively identify and distinguish. Use Equation 1 to calculate the correlation between each set of waveform data collected and each set of waveform data different from itself. The correlation coefficient is shown in Table 1.

Table 1 Correlation coefficient between 0.3～0.7 mm crack defect and 0.3～0.7 mm crack defect

When the correlation coefficient is in the interval [0.5 1] and [-1 -0.5], it indicates that the two defects have a tendency of positive correlation or negative correlation. When the correlation coefficient is within (-0.5 0.5), it can be said that there is no correlation between the two defects. The closer the correlation coefficient is to 0, the smaller the correlation is, and the more feature information can be identified and distinguished.

From the above table, it can be analyzed that the correlation coefficient between any two different defect feature signals that are screened is far less than 0.1, indicating that the feature information that can be identified and distinguished by the PNN network.

After the data collection is completed, the current noise is the main Gaussian white noise in the data, and denoising processing is required. Wavelet denoising processing is performed on each group of collected data. Using the threshold of minimaxi denoising method for denoising 50 000 points in each type of defect waveform were selected as training samples to perform wavelet packet energy feature vector extraction. The d5 wavelet is used to decompose the sample into three layers of wavelet packet, and extract the wavelet packet coefficients of the eight bands in the third layer ([3,0], [3,1],…, [3,7]), and then wavelet denoising. Taking the 0.3mm defect one-period sampling pattern as an example, the time-domain image after wavelet packet decomposition and denoising is shown in Fig.7.

Fig.7 Wavelet packet decomposition of 50,000 samples at 0.3mm defect wave

3.3 Defect recognition based on PNN network

In the test experiment, test waveform data of crack defects of 0.3 mm, 0.4 mm, 0.5 mm, 0.6 mm, 0.7 mm is used as training samples, and neural network learning training is performed on the samples. After training the samples, 100 recognition tests were performed on the test samples of each defect, and the results are shown in Table 2.

Table 2 0.3～0.7 mm crack defect identification results

It can be seen from the experimental results that this method works well. The lowest recognition rate is 0.3 mm crack defects, and the highest recognition rate is 0.7 mm defects, which conforms to the rule that the larger the defect, the easier it is to be identified. The highest recognition rate is 94% instead of 100%, which may be caused by factors such as the surface roughness of the measured workpiece, impurities inside the workpiece and the measurement environment.

4 Conclusion

Probabilistic neural networks have extremely strong non-linear processing capabilities and fault tolerance for sample feature information. The prior knowledge of standard patterns can be used to the greatest extent to identify and classify samples under the Bayes minimum risk criterion, so that the PNN network method has recognition. It has the advantages of high accuracy, fast network training, global optimization, and no local minima. The experimental results show that the application of probabilistic neural network can effectively identify crack defects in aluminum alloy workpieces, and can complete the classification of crack defect sizes, providing an effective solution for the study of automated ultrasonic testing.