Detection of early browning in pears using vibro-acoustic signals

Core browning in Korla pear (Pyrus bretschneideri Rehd.) occurs generally during storage at room temperature. The browning disorder can significantly reduce the shelf stability, and thereby to cause considerable economic losses. Moreover, the browning part of pears can be taken in the juicing process, leading to the juice toxins over the safety limit for drinking. Therefore, a reliable and rapid method has been urgently demanding to nondestructively detect internal disorder for high-quality fruits. In this study, an acoustic system using the piezoelectric beam transducers was developed for nondestructively detecting disorder of pears with different internal browning. The obtained response signals were analyzed to extract 11 statistical features in time domain, and seven statistical features in frequency domain. Accordingly, three modes of feature vectors were formed in the time domain, frequency domain, and time-frequency domain. A Compensation Distance Evaluation Technology (CDET) was also used to evaluate the sensitivities of each parameter in feature vectors. Normally, the larger values of sensitivity evaluation factor can imply the higher sensitivities to the browning classes of pears. Based on sensitivity evaluation factors values in the healthy and browning of pears, the descending order of 11 time-domain features were the mean (T1), shape factor (T11), kurtosis (T6), square root amplitude value (T5), clearance indicator (T8), peak (T3), impulse factor (T9), root mean square (T2), short-time energy (T4), kurtosis factor (T7), and crest factor (T10). The sensitivities of seven frequency-domain features were also ranked in order, the variance (F2), mean square (F6), root mean square (F7), standard deviation (F3), mean (F1), kurtosis (F4), and gravity (F5). Combining two types of features, the descending order of all the features was as follows: T11, F2, T6, T5, F7, F6, T1, F3, F4, F1, F5, T8, T3, T9, T2, T4, T10, and T7. In the slight browning and moderate browning of pears, the sensitivities of time-domain features can be ranked in the descending order of T7, T11, T8, T3, T9, T10, T6, T2, T5, T4, and T1. The obtained order for the frequency-domain features was F4, F3, F7, F6, F1, F5, and F2. In the combined time-frequency features, the order was as follows: T11, T8, F6, T3, F7, F3, T7, F4, F1, F2, T9, F5, T10, T6, T2, T5, T4, and T1. Subsequently, a K-nearest neighbor (KNN) algorithm was utilized to train the classifier using the first n sensitive features as the inputted data. Therefore, a browning discrimination model was constructed for the moderate disorder, whereas, a slight browning discrimination model for the mild disorder. Both models performed the best, when combining the features from the time-domain and frequency-domain. In the browning discrimination model, a high overall accuracy of 91.84 % was obtained with the specific feature vectors, including three time-domain features (T11, T6 and T5), and one frequency-domain feature (F2). In slight browning discrimination model, the slight browning of pears can be further identified with an accuracy of 81.82 %. In the case of slight browning, the specific feature vectors were adopted, including two time-domain features (T11 and T8), and one frequency-domain feature (F6). In the confusion matrix analysis, the high values of F1 indicated that two discrimination models can be used to achieve the high robustness and performance, and further to be generalized for the identification of fruits. These findings can provide a sound theoretical basis and strategy for the industrial real-time in-line detection, and automatic grading in the internal browning disorder of pears.