Investigations On Pattern Recognition Of Damage Mechanisms In Self-reinforced Polyethylene Composites

Ultra high molecular weight polyethylene (UHMWPE) fiber is a kind of high performance organic fiber with highly oriental extended-chain. It has been widely used in many fields for high specific strengthen, high specific modulus, excellent toughness, well abrasion resistance and excellent anti-impact performance, fine chemical resistance, insulation and biocompatibility. With rapid development of thermoplastic composite materials in recent years, their structures were used in more and more engineering fields. Based on "one polymer composites", the interface of PE self-reinforced composites and general performances of fiber were improved due to chemical compatibility. Meanwhile, it has more competitive advantage than other composites such as low price, rich resource and convenient recycle. Research on damage mechanisms of PE self-reinforced composites are important to the assurance of safety of the composites in service and obtain the optimal structure design. Results of damage mechanisms on glass-fiber, carbon-fiber reinforced thermoset composites revealed acoustic emission (AE) is an effective tools which can provide rich and real information during damage progress. Up to now, researches mostly focus on thermoset composites and rarely focus on thermoplastic composites. PE self-reinforced composites is a typical thermoplastic composites, this study will investigate its AE characteristic during damage progress based on AE technology. Then, correlation between damage mechanisms and AE signals will be established and damage mechanisms will be classified and identified. This study can provide a convenient and effective method for damage mechanisms identified on thermoplastic composites based on AE technology during real applications.Damage mechanisms of fiber reinforced composites are very complicated including fiber breakage, matrix crack, interface debonding and delamination. In order to reveal AE feature of time domain and frequency domain generated from different damage mechanisms of PE self-reinforced composites and establish correlation between damage mechanisms and AE signals to classify and identify AE signals, and finally establish AE signal pattern recognition system based on artificial neural networks, this study mainly include following research work. Firstly, according to damage characteristic of thermoplastic composites, using model sample with simple structure to generate AE signals of desired damage mode, analyze AE features of time domain and frequency domain of typical damage mechanisms based on mechanical performance change and Fast Fourier transformation (FFT). Secondly, according to damage process of model sample with simple structure, the cluster analysis on AE signals were investigated including selection of similarity measure and cluster variables and validation of cluster results. Cluster analysis method of AE signals on typical damage modes was established. Results of several common methods on discriminant analysis were compared. Finally, artificial neural networks for AE signal classification and identification on PE self-reinforced composites were established and performance of neural networks due to some optimal algorithms was also compared. The experiment revealed damage process of model sample including different damage mechanisms and AE response can reflect features of different damage stage. Matrix sample has small number of signals and poor AE activities, the damage mechanisms including plastic deformation and fracture.90°laminate generate interface damage signals with different degree. Due to complicated structure and more damage,0°laminate and [+45°/-45°] laminates have more AE signals and high AE activities. The former mainly generate AE signals from fiber breakage, the later mainly generate AE signals from in-layer shear and interlayer damage. AE signals from damage mechanisms with small degree destruction such as matrix plastic deformation and interface initial damage are low amplitude and short duration. However, AE signals from big degree destruction such as matrix fracture, interface debonding, fiber breakage and delamination are high amplitude and long duration. According to damage process, AE signals from early damage of all kinds of specimen are all low amplitude with short duration, while high amplitude with long duration AE signals mainly generated at the moment of material fracture. The results of FFT analysis on AE signals revealed different frequency features between non-damage signals and damage signals as well as among different damage mechanisms. The difference from various damage mechanisms is obvious both on time domain and frequency domain. However, the overlap of distribution results in identification difficulty when only using parameters of time domain or frequency domain.AE signals from damage process of composites are hybrid signals with many different damage modes. The purpose of cluster analysis on AE signals is to establish correlation between typical damage mechanisms and AE signals. For obtaining reliable cluster results, simple model specimens such as pure matrix,90°unidirectional laminate, fiber bundle and [+45°/-45°] laminates were used to generate expected damage mechanisms. Firstly, the results of variable cluster analysis on 8 AE parameters revealed similarity within groups is better than similarity between groups and relation of similarity is same when all parameters were divided into 3 groups. Therefore, amplitude, peak frequency and duration selected from each group can be used as pattern features. Then, based on k-means algorithms, sample cluster analysis on all kinds of model specimens were performed to establish AE training set with 8 typical kinds of damage mechanisms. Cluster results were validated by SEM of each model specimen. According to cluster results, discriminant analysis on AE signals was also performed. The result of hypothesis test on mean equal among different model showed each pattern features were effective for discriminant analysis. Furthermore, discriminant effects of Euclidean distance, Mahalanobis distance and k nearest neighbor discriminance were compared with percentage of right identification. The results revealed Mahalanobis distance discriminance has the best effect. Percentage of right identification over 90% showed different damage mechanisms can be divided with pattern features including amplitude, peak frequency and duration. The main error of identification resulted from different discriminances were all come from identification mistake between matrix fracture and interface debonding as well as matrix plastic deformation-1 and interface damage. Distribution of AE signal in pattern space from typical damages showed better right identification resulted from better separability and identification mistake will take place when overlap each other. Classification and identification of AE signal can be realized by cluster and discriminant analysis. However, these methods are complicated in process, poor efficiency and inconvenient to real application.In order to improving classification and identification, convenience of data analysis and more fit for demand of real application, this study investigated pattern recognition of AE signal based on artificial neural networks. These work included establishing self-organizing competitive (SOC) and error back-propagation (BP) network to realize classification and identification respectively. According to the classification results from SOC network and cluster analysis to the same data set, the classification consistency of two methods is over 98%. Therefore, SOC network is a feasible and effective tool for cluster analysis of AE signals. Based on the same data set and the same architecture of BP network, the right identification percentage with standard BP training algorithm and other improved training algorithms were compared. The results revealed standard BP algorithms is unfit for real application due to long training time, low identification and easy effected by initial value of network. Among heuristic improved method, both BP algorithm with momentum and adaptive learning rate and resilient BP algorithm can improve identification effect and accelerate training process. Among standard numerical optimal improved methods, Levenberg-Marquardt algorithm has the best training effect. It is the best algorithms to training BP networks to identify different damage mechanisms with AE signals.