关键词: 针织机械, 压电选针器, 随机共振, 共位检测, BP神经网络, 失效分析

Abstract:

Objective Piezoelectric needle selector is the driver of jacquard needles. Its performance is related to the quality of jacquard knitting production during the process of needle hooking and looping. Aging of piezoelectric crystal or peeling off intermediate bonding layer would cause gradual failure of the actuator. Once the needle selector fails to act normally during knitting process, it would lead to defects such as off-pattern, holes in cloth surface, and even machine failures e.g., pin impact and pin breakage. The existing piezoelectric needle selector uses open loop control, and the results of needle selector would not be perceived. Hence, the control system would not be able to judge whether the action of the knife head of the needle selector is accurate, and the abnormal operation of the piezoelectric needle selector often causes mechanical failure or abnormal jacquard knitting.

Method Given the difficulty in identifying fault characteristics of the piezoelectric needle selector during the jacquard process, a gradient failure detection scheme based on the stochastic resonance-BP (SR-BP) algorithm for the piezoelectric needle selector was proposed. The study investigated the motion state of the driving component: the twin-crystal piezoelectric cantilever beam during the jacquard process of the piezoelectric needle selector, as well as the electric signal generated by its dual-directional piezoelectric effect. For the internal electric signals of the piezoelectric ceramic driver in a gradient failure state, SR fault diagnosis was conducted, and the SR-BP model was established. By extracting the time-domain and frequency-domain parameters of the vibration sequence and SR parameters to generate training samples, the algorithm was made to match the feature parameters to obtain the optimal SR parameters. The vibration signal of the piezoelectric needle selector was mixed with the noise signal into the nonlinear system formed by the SR parameters, and the time-domain and frequency-domain changes of the system output was observed to achieve the purpose of rapid fault detection.

Results An experimental platform and testing system were considered, and the SR-BP analysis on 1 000 sets of data was conducted. The results show that when the SR-BP samples were fewer than 200, the accuracy was below 75%. However, as the number of samples increased, the network accuracy of SR-BP continuously got risen, exceeding 95% when samples exceeded 1 000. The network accuracy of SR-BP primarily depended on the sample size, and the larger the sample size the higher the prediction accuracy. Simultaneous tests on the influence of different feature parameters on testing accuracy and signal-to-noise ratio (SNR) were conducted. With only three induced parameters, nine time-domain feature parameters and SNR, the network accuracy reached 95.4%, with an SNR of 6.91 dB. In the analysis using three induced parameters, three frequency-domain parameters, and SNR, accuracy reached 94.6%, with an SNR of 5.54 dB, showing a decline in accuracy and SNR compared to time-domain parameters. Hence, time-domain features could enhance network accuracy. This is because time-domain parameters can effectively describe the local features of the original vibration signal, better representing the original signal. Frequency-domain parameters reflected the statistical measure of the original signal over an extended time scale, making it a challenge to differentiate the signal's local features. After considering both time and frequency-domain characteristics, the SR response fault features were more pronounced, indicating optimal fault feature extraction with an accuracy of 97.5% and an SNR of 7.36 dB. The combination of time-domain and frequency-domain features provided the best fault diagnosis results. In conclusion, as the number of feature parameters increased, the predictive accuracies of SR parameters and the SNR were both improved. The signal and SR response characteristics in the samples acted as constraints in network training, and as these constraints increased, training results were gradually improved. Optimal SR could only be achieved by thoroughly extracting signal features and inputting them into BP.

Conclusion After in-depth research and experimentation, it is confirmed that the piezoelectric picker's gradual failure detection scheme based on the SR-BP algorithm possesses an extremely high level of accuracy under large sample conditions, reaching an accuracy rate of 97.5%. Compared to conventional diagnostic methods, this approach is more efficient and rapid, significantly enhancing the reliability of fabric quality. This research suggests that combining time-domain and frequency-domain characteristics can achieve the best fault diagnosis results. In the textile field, applying this method can notably reduce quality problems caused by picker faults, bringing tangible benefits to weaving enterprises. It is recommended that future researchers explore more feature parameters and more complex neural network models to further improve diagnostic accuracy. With technological advancements and increasing data volumes, it is predicted that the accuracy of this detection method will further improve, offering more innovative opportunities for the textile industry.

Key words: knitting machinery, piezoelectric needle selector, stochastic resonance, colocation detection, BP neural network, failure analysis