YARN TENSION PATTERN RETRIEVAL SYSTEM BASED ON GAUSSIAN MAXIMUM LIKELIHOOD. Received July 2010; revised December 2010

Transcription

1 International Journal of Innovative Computing, Information and Control ICIC International c 2011 ISSN Volume 7, Number 11, November 2011 pp YARN TENSION PATTERN RETRIEVAL SYSTEM BASED ON GAUSSIAN MAXIMUM LIKELIHOOD Chuan-Pin Lu 1 and Jiun-Jian Liaw 2, 1 Department of Health Industry Information Technology Meiho University No. 23, Pingguang Rd., Neipu, Pingtung 91202, Taiwan chuan.pin.lu@gmail.com 2 Department of Information and Communication Engineering Chaoyang University of Technology No. 168, Jifong E. Rd., Wufong District, Taichung 41349, Taiwan Corresponding author: jjliaw@cyut.edu.tw Received July 2010; revised December 2010 Abstract. The unusual yarn tension inspection plays an important role in yarn quality measurement. The patterns of unusual tension, which are captured and recorded by an on-line yarn tension monitor system, can ensure precise recognition of the unusual type of tension in order to solve the problem immediately. However, it is not easy for operators of twister machines to recognize the patterns of unusual tension without related training. The traditional on-line yarn tension monitor systems only detect unusual variation in tension, but cannot identify the patterns of unusual tension for operators, especially in the improved quality yarns. To assist the operators in the pattern recognition of unusual tension, in this paper, we propose an unusual yarn tension retrieval system based on Gaussian maximum likelihood classification algorithm. The proposed system uses four features to describe the tension patterns, and includes the following processes: pattern generation, feature calculation, similarity degree measurement, new class detection and pattern retrieval. Experimental results show that the proposed system can serve as an efficient and fast tool to identify unusual tension patterns. Keywords: Unusual tension, Yarn, Retrieval system, Gaussian maximum likelihood 1. Introduction. Many automatic methods and systems [1-3] have been developed to improve the quality of products and the operation efficiency, such as on-line yarn tension monitor systems (like Barmag s Unitens [4], Yu Hwa s QAI [5] and FAG s product), which serve as an important tool for automatic package grading, real time data view, historical data report, unusual yarn tension detection and recording. Yarn quality grading is based on the number of occurrences of unusual tension. Thus, the records of unusual tension can enable operators to track the unusual events, find the yarn s defects and improve it. These traditional monitor systems only carry out the patterns detection and record for unusual variation in tension, but cannot help operators to identify patterns, especially in yarn quality improvement. In factory, unusual tension pattern identification is now performed by operators. It is difficult for an operator to perfectly identify many kinds of patterns without the assistance of equipment. The curve is crucial in the unusual tension pattern identification. However, the patterns curves are not the same while the yarn s production equipment, manufacturing conditions or manufacturing processes are different. Sometimes, patterns of a new class can also be found. As mentioned above, even welltrained and experienced operators cannot remember all class patterns, especially when they are tired. To assist operators in the identification of unusual tension patterns, we 6261

2 6262 C.-P. LU AND J.-J. LIAW have developed an unusual tension pattern retrieval system based on Gaussian maximum likelihood classification algorithm [6] and described in this paper. The system provides not only pattern recognition, but also new class pattern detection for users. Less execution time is needed for pattern retrieval in the proposed system, which is necessary and useful for the on-line yarn tension monitor systems. Retrieval systems are used in various fields, such as images [7-9], sounds [10], cars [11], patents [12], knowledge [13], laws, standards and philology. The retrieval systems use a query keyword as feature to search similar patterns or the correlative data from database. These keywords can be a simple text with a specific significance or a visual image with multi-information (such as colors, shapes and textures). However, little attention has been given to the retrieval system for unusual yarn tension. Therefore, we propose an unusual yarn tension retrieval system in this paper. In the proposed system, the tension data are measured by tension sensors and plotted as a curve. The curve is regarded as the pattern of unusual tension [4,5] after normalization and noise reduction. We use the similarity measurement of Gaussian maximum likelihood [6] for patterns matching in the retrieval method. Each unusual tension signal data has its noise reduced by the wavelet filter [14,15] and normalized by rescaling. Then, we calculate the features (such as variance, skewness, gradient cumulative value and entropy) from the query pattern, and measure the similarity degrees of all classes by Gaussian similarity likelihood method. At last, similarity comparison is done and the five similar samples are displayed in retrieval results. In the sections below, our method is introduced, and then four experiments that were carried out to evaluate the performance of the proposed system are described. At last, the conclusion is given. Figure 1. The framework of the proposed retrieval system 2. Methods. The framework of the proposed retrieval system is based on the clientserver architecture, which is appears in Figure 1. The retrieval system (server) collects various unusual tension patterns from on-line tension monitor systems (clients), and provides retrieval services. If the new class patterns are detected, the message will be given to the operator by news flash. The retrieval method contains four processes: Firstly, the query pattern has its noise reduced and normalized; secondly, the four features are calculated; thirdly, the similarity degrees are measured; lastly, new class pattern detection

3 YARN TENSION PATTERN RETRIEVAL SYSTEM 6263 and pattern retrieval are performed. The details of the processes are described in the following sections Pattern generation. The yarn tension signals are always mixed with the noises that increase the difficulty of pattern recognition. Thus, noise reduction is necessary before the features calculation. The common denoise filters, such as Butterworth filter, smoothing filter and wavelet filter, are used for noise reduction in preprocessing. These filters have different individual characteristics. Concluding from our experiments, the wavelet filter performs noise reduction the best. We apply the wavelet filter to reduce noise in this paper. Wavelet transform can efficiently extract both time and frequency information from a time-varying signal. The noise can be reduced by removing smaller wavelet coefficients. We use the Mallat s herringbone algorithm (Mallat, 1989 [14]) with Daubechies-4 wavelet (scaling function) to transform the tension signal data into wavelet coefficients without downsampling. The Daubechies-4 scaling function coefficients are defined as ( , , , ) [15]. An original tension signal s(t) has 512 data points and 8-bit signal resolution, and the data sequence is denoted as t. After the above noise reduction, the signal data s(t) is normalized on a 0 to 100 scale. The pattern is normalized signal data, and we denote it as s(t). The s(t) and s(t) are shown in Figure 2. (a) (b) Figure 2. Pattern generation: (a) original signal s(t); (b) pattern s(t) 2.2. Features calculation. Four features (variance (σ 2 ), skewness (η), gradient cumulate value (gc) and entropy (h)) are used to recognize the unusual tension patterns. The notation x = (σ 2, η, gc, h) T is denoted as feature vector of the pattern s(t). The features are defined as following equations: Variance σ 2 : / M M σ 2 = ( s(t) µ) /M, 2 µ = s(t) M (1) where M is data number of pattern (M = 512), µ is mean value of amplitude. This feature is used to measure the range of variation in tension. Skewness η: /( ) η = M s(t)(t ˆµ) 3 ˆσ 3 M s(t), / ˆµ = M M / M M (2) t s(t) s(t), ˆσ = s(t)(t ˆµ) 2 s(t) Skewness is the measure of the degree of symmetry of a curve.

4 6264 C.-P. LU AND J.-J. LIAW Gradient cumulate value gc: gc = M 1 ( s(t + 1) s(t)) (3) Gradient cumulate value characterizes the degree of variation in tension. Entropy h: If s(t) {a 1,..., a ν,..., a β }, the notation P (a ν ) is the probability of s(t) = a ν, and v = (P (a 1 ),..., P (a ν ),..., P (a β )) T, the entropy h is used to measure the occurrence uncertainty of tension peak, which is defined as follows: β h = H(v) = P (a ν ) log P (a ν ) (4) ν=1 The feature vector x has to be normalized before the similarity degree measurement. The normalized x is re-denoted as x Similarity degree measurement. After obtaining the feature vectors x, the similarity measurement of all classes is performed. In this paper, the probability functions of feature vectors are assumed as having Gaussian distribution. The query pattern is classified according to Bayesian estimation with the Mahalanobis distance, which is a minimum distance classifier [6]. The distance (d) is the decision function, and is regarded as the similarity degree. The similarity degree of the query pattern in class w j is denoted as d j. It is defined as d j( x ) = [ ( x µ j ) T C 1 j ( x µ j ) ] 1/2, j = 1,..., n, (5) where x is the feature vector of the query pattern; n is the number of pattern classes; µ j and C j are the mean vector and the co-variance matrix of the training samples in class w j, respectively. The µ j and C j are defined as µ j = 1 m x m j,i, C j = 1 m i=1 m i=1 x j,i x T j,i µ j µ T j, j = 1,..., n, (6) where x j,i is the feature vector of ith training sample; m is the number of training samples. According to the central limit theorem, the µ j and C j will tend towards finite values when the number of training samples is sufficiently large. Each query pattern has n similarity degrees {d 1, d 2,..., d n} with n classes New class detection and pattern retrieval. Sometimes, the new class patterns (also called the unknown class patterns), which have never been defined in the database, are found when the yarn s production equipment, manufacturing process or manufacturing conditions are different. For such cases, the system needs a method for the new class patterns detection. Therefore, we set the thresholds (T j ) of similarity degree. Each T j represents the least similarity of patterns in each class w j. If all d j are larger than all T j (d j > T j, j = 1,..., n). The query pattern is identified as the new class pattern, and it must be stored on a temporary storage and waited for recognition by the operators. The T j is defined as T j = max{d j,i ( x j,i)}, i = 1,..., m, j = 1,..., n, (7) where d j,i is the similarity degree of ith training sample x j,i in class w j. In pattern recognition, the system finds the minimum similarity degree d z(= min{d j}, z j) from n similarity degrees {d 1, d 2,..., d n}. If d z T z, and the query pattern is

5 YARN TENSION PATTERN RETRIEVAL SYSTEM 6265 assigned to the class w z, otherwise it is regarded as the new class pattern. At last, the five similar patterns in database with smaller values of similarity distance d j,i are increasingly ranked and displayed for users. d j,i is defined by the following equation. d j,i = d j d j,i, i = 1,..., m, j = 1,..., n (8) 3. Experimental Results. The common ten class patterns of unusual tension, such as Knot Transfer, Foreign Matter, Higher Tension, Splice Transfer, Unusual Descent, Unusual Rise, Tail Transfer, Yarn in Out, Broken Yarn and Cut Yarn, are collected from QAI on-line yarn tension monitor systems (Yu Hwa Co., ltd, 2004) (n = 10, see Figure 3), and used to evaluate the performance (the identification accuracy and the execution time of retrieval) of the proposed system. The feature vectors x of the patterns are shown in Table 1. Each class of patterns has forty samples (m = 40), which are used to obtain the µ j, C j and T j. The µ j and T j are shown in Table 2. In additions, we used a personal computer with the Intel Pentium 2.4 GHz CPU to conduct all the experiments. Table 1. The feature vectors x of ten classes of the patterns (see Figure 3) unusual tension pattern Knot Transfer Foreign Matter Higher Tension Splice Transfer Unusual Descent Unusual Rise Tail Transfer Yarn In Out Broken Yarn Cut Yarn x = (σ 2, η, gc, h) T (26.006, 1.182, 5.479, 0.067) T (18.385, 1.123, 2.113, 0.003) T (19.840, 1.209, 0.011, 0.343) T (35.305, 1.107, , 0.093) T (27.281, 1.462, , 0.017) T (28.954, 1.090, , 0.019) T (14.960, 1.223, 1.905, 0.167) T (23.033, 1.308, , 0.064) T (22.381, 1.304, , 0.282) T (26.414, 1.267, , 0.174) T Table 2. The mean of feature vectors µ j, and the thresholds T j j unusual tension pattern µ j T j 1 Knot Transfer (0.571, 0.333, 0.514, 0.139) T Foreign Matter (0.251, 0.210, 0.515, 0.014) T Higher Tension (0.404, 0.477, 0.632, 0.731) T Splice Transfer (0.685, 0.254, 0.728, 0.193) T Unusual Descent (0.534, 0.881, 0.094, 0.069) T Unusual Rise (0.629, 0.461, 0.589, 0.053) T Tail Transfer (0.097, 0.435, 0.525, 0.382) T Yarn In Out (0.514, 0.557, 0.541, 0.096) T Broken Yarn (0.355, 0.642, 0.276, 0.518) T Cut Yarn (0.600, 0.539, 0.240, 0.375) T Four experiments were carried to evaluate our method as Figures 4-7 show. Figure 4 shows the experiment of the query pattern Knot Transfer (Figure 4(a)), which is a training sample. In the experiment, the retrieval results are ranked by similarity distance d, and shown in Figures 4(b)-4(f). The feature vector ( x ) and the similarity degree (d j) of the query pattern are shown in Table 3. In Table 3, d 1(= 0.353) has the minimum

6 6266 C.-P. LU AND J.-J. LIAW (a) Knot Transfer w 1 (b) Foreign Matter w 2 (c) Higher Tension w 3 (d) Splice Transfer w 4 (e) Unusual Descent w 5 (f) Unusual Rise w 6 (g) Tail Transfer w 7 (h) Yarn In Out w 8 (i) Broken Yarn w 9 (j) Cut Yarn w 10 Figure 3. The common ten classes of the patterns of unusual tension

7 YARN TENSION PATTERN RETRIEVAL SYSTEM 6267 Table 3. Experiment 1 for Knot Transfer (class w 1 ), d 1 = min{d j} is (< T 1 ), the execution time of retrieval is 171 milliseconds: feature vector x and similarity degree d j of the query pattern and the top five similar samples in the database d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 d 9 d E Figure 4 x similarity degree similarity distance (a) (0.624, 0.324, 0.516, 0.113) T (b) (0.498, 0.364, 0.489, 0.162) T d 1 = d 1,3 = d 1,3 = (c) (0.543, 0.333, 0.478, 0.139) T d 1,1 = d 1,1 = (d) (0.572, 0.305, 0.540, 0.175) T d 1,13 = d 1,13 = (e) (0.542, 0.325, 0.489, 0.183) T d 1,9 = d 1,9 = (f) (0.615, 0.288, 0.552, 0.120) T d 1,29 = d 1,29 = (a) Query pattern (b) Rank 1 (c) Rank 2 (d) Rank 3 (e) Rank 4 (f) Rank 5 Figure 4. The retrieval results of the experiment 1 for Knot Transfer : (a) query pattern; (b)-(f) the top five similar samples value and is less than the T 1 (= ); the execution time of retrieval is 171 milliseconds. In Figure 4, the classes of five similar patterns (rank 1-5) are the same. Therefore, the pattern is assigned to class w 1. Judging from the experimental results, the retrieval result is correct.

8 6268 C.-P. LU AND J.-J. LIAW Table 4. Experiment 2 for Broken Yarn (class w 9 ), d 9 = min{d j} is (< T 9 ), the execution time of retrieval is 166 milliseconds: feature vector x and similarity degree d of the query pattern and the other patterns in the database d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 d 9 d E Figure 5 x similarity degree similarity distance (a) (0.296, 0.628, 0.306, 0.365) T (b) (0.517, 0.573, 0.257, 0.306) T d 9 = d 9,13 = d 9,13 = (c) (0.294, 0.735, 0.253, 0.491) T d 9,21 = d 9,21 = (d) (0.263, 0.617, 0.298, 0.449) T d 9,6 = d 9,6 = (e) (0.293, 0.672, 0.265, 0.583) T d 9,5 = d 9,5 = (f) (0.344, 0.630, 0.299, 0.655) T d 9,9 = d 9,9 = (a) Query pattern (b) Rank 1 (c) Rank 2 (d) Rank 3 (e) Rank 4 (f) Rank 5 Figure 5. Retrieval results of the experiment 2 for Broken Yarn : (a) query pattern; (b)-(f) the top five similar samples The second experiment is shown in Figure 5. The query pattern is Broken (Figure 5(a)), which is a test sample. After similarity measurement and comparison, the d 9 is (< T 9 ), which is the minimum value. This pattern is assigned to class w 9 after performing the new class detection and the pattern retrieval (see Figures 5(b)-5(f) and

9 YARN TENSION PATTERN RETRIEVAL SYSTEM 6269 (a) Query pattern (b) Rank 1 (c) Rank 2 (d) Rank 3 (e) Rank 4 (f) Rank 5 Figure 6. Retrieval results of the experiment 3 for an artificial pattern: (a) query pattern; (b)-(f) the top five similar samples Table 5. Experiment 3 for an artificial pattern, d 8 = min{d j} is (> T 8 ), the execution time of retrieval is 183 milliseconds: feature vector x and similarity degree d of the query pattern and the other patterns in the database d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 d 9 d E Figure 6 x similarity degree similarity distance (a) (0.582, 0.207, 0.987, 0.038) T (b) (0.752, 0.670, 0.139, 0.096) T d 8 = d 8,25 = d 8,25 = (c) (0.585, 0.587, 0.830, 0.055) T d 8,36 = d 8,36 = (d) (0.393, 0.615, 0.732, 0.157) T d 8,16 = d 8,16 = (e) (0.594, 0.746, 0.295, 0.103) T d 8,17 = d 8,17 = (f) (0.397, 0.471, 0.491, 0.133) T d 8,24 = d 8,24 = Table 4). As the result of the retrieval, the query pattern can be determined correctly. The execution time of retrieval is 166 milliseconds. In the third experiment, an artificial pattern, which does not belong to any class, is used to test the method of new class detection. The retrieval results, similarity degrees

10 6270 C.-P. LU AND J.-J. LIAW (a) Query pattern (b) Rank 1 (c) Rank 2 (d) Rank 3 (e) Rank 4 (f) Rank 5 Figure 7. Retrieval results of the experiment 4 for the pattern with mass variance of material: (a) query pattern; (b)-(f) the top five similar samples Table 6. Experiment 4 for Tail Transfer (class w 7 ), d 7 = min{d j} is (> T 7 ), the execution time of retrieval is 178 milliseconds: feature vector x and similarity degree d of the query pattern and the top five similar samples in the database d 1 d 2 d 3 d 4 d 5 d 6 d 7 d 8 d 9 d E Figure 7 x similarity degree similarity distance (a) (0.235, 0.544, 0.544, 0.069) T (b) (0.577, 0.180, 0.716, 0.039) T d 7 = , d 4 = d 4,37 = d 4,37 = (c) (0.072, 0.369, 0.621, 0.133) T d 7,11 = d 7,11 = (d) (0.218, 0.478, 0.522, 0.199) T d 7,16 = d 7,16 = (e) (0.258, 0.418, 0.498, 0.261) T d 4,2 = d 4,2 = (f) (0.404, 0.378, 0.670, 0.140) T d 4,19 = d 4,19 = and feature vectors are shown in Figure 6 and Table 5, respectively. The similarity degree calculation shows that all d j are large than all T j (j = 1,..., 10) (see Table 5). Although, the d 8 has the least value among others, it is still larger than T 8. This pattern is therefore assigned to the new class pattern correctly.

11 YARN TENSION PATTERN RETRIEVAL SYSTEM 6271 In the last experiment, a test pattern Knot Transfer with mass variation of material, which is a special case, is carried to evaluate our method. The mass variance is not an unusual type of tension; it mixes with tension pattern when the quality of yarn s material is poor. As retrieval results (see Table 6 and Figure 7), all d j are large than all T j, the five similar patterns include two classes ( Tail Transfer and Splice Transfer ). Therefore, this pattern is regarded as a new class pattern and stored in a temporary storage for recognition by the operators. Although, this pattern is failed to identify as a new class pattern and excluded with regard as a training sample, but it is conducive to preserve the identification accuracy for the classifier. In reality, this pattern is always regarded as the one with a blemish and as the nullity sample by operators. From the results of experiments above, it was found that the identification accuracy is determined by the quality of training samples. If the training sample is mixed with the mass variance, it will make the T j large, and lead to the reduction of the identification accuracy. Therefore, we need to cautiously select training samples in the proposed system. As seen from the experimental results, all experiments can successfully obtain the correct retrieval results with less computation time. 4. Conclusions. In this paper, a retrieval system of unusual tension patterns based on the Gaussian maximum likelihood is developed. This system is useful for operators to track the unusual events, find new class patterns, and improve the yarn quality. We design four features for recognition of the common unusual tension patterns and the thresholds of similarity degree for new class pattern detection. Experimental results show that the proposed system is an efficient and fast tool for operators in unusual tension pattern retrieval. Acknowledgment. This work was supported in part by the National Science Council, Taiwan, under grant NSC E REFERENCES [1] T.-C. Lin, M.-J. Kuo and C.-H. Hsu, Robust adaptive tracking control of multivariable nonlinear systems based on interval type-2 fuzzy approach, International Journal of Innovative Computing, Information and Control, vol.6, no.3(a), pp , [2] A. Chouder and S. Silvestre, Automatic supervision and fault detection of PV systems based on power losses analysis, Energy Conversion and Management, vol.51, no.10, pp , [3] Z. Q. Lang and Z. K. Peng, A novel approach for nonlinearity detection in vibrating systems, Journal of Sound and Vibration, vol.314, no.3-5, pp , [4] Barmag Web Site, [5] C.-P. Lu and L.-K. Tsou, QAI On-Line Yarn Quality Monitor System User Manual, Yu Hwa Co. ltd, Taiwan, [6] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Academic Press, London, [7] J. Zhang and Y. Lei, Content based image retrieval using unclean positive examples, IEEE Trans. on Image Processing, vol.18, no.10, pp , [8] E. Aptoula and S. Lefevre, Morphological description of color images for content-based image retrieval, IEEE Trans. on Image Processing, vol.18, no.11, pp , [9] H. Liu, C. Zhang, X. Ji and Y. Zhang, An algorithm for co-training in medical image retrieval, International Journal of Innovative Computing, Information and Control, vol.5, no.11(b), pp , [10] K. Bozena, Computing with words concept applied to musical information retrieval, Electronic Notes in Theoretical Computer Science, vol.82, no.4, pp.1-12, [11] T. Samatsu, K. Tachikawa and Y. Shi, Usability improvement for a car retrieval system employing the important degrees of fuzzy grades, International Journal of Innovative Computing, Information and Control, vol.5, no.12(b), pp , [12] V. Hassler, Open source libraries for information retrieval, IEEE Software, vol.22, no.5, pp.78-82, 2005.

12 6272 C.-P. LU AND J.-J. LIAW [13] S. Mukherjea, B. Bamba and P. Kankar, Information retrieval and knowledge discovery utilizing a biomedical patent semantic web, IEEE Trans. on Knowledge and Data Engineering, vol.17, no.8, pp , [14] S. G. Mallat, A theory for multiresolution signal decomposition: The wavelet representation, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol.11, pp , [15] C. S. Burrus, R. A. Gopinath and H. Guo, Introduction to Wavelet and Wavelet Transforms, Prentice-Hall, 1998.