Development of an artificial neural network model to predict expert judgement of leather handle from instrumentally measured parameters

Leather is a widely used material whose handling character is still assessed manually by experienced people in the leather industry. The aim of this study was to provide a new approach to such characterisation by developing Artificial Neural Network models to investigate the relationship between the subjective assessment of leather handle and its measureable physical characteristics. Two collections of commercial leather samples provided by TFL and PITTARDS were studied in this project. While the handle of the TFL collection covered a varied range, the PITTARDS collection was all relatively soft leather and with less difference within the collection. Descriptive Sensory Analysis was used to identify and quantify the subjective assessment of leather handle. A panel constituted of leather experts was organised and trained to: 1) define attributes describing leather handle; 2) assess specific leather handle by responding to questionnaires seeking information about the above attributes. According to the analysis of the raw data and the assessment observation, the attributes that should be used for training the artificial network models were "stiff", "empty", "smooth", "firm", "high density" and "elastic". Various physical measurements relating to leather handle were carried out as follows: standard leather thickness, apparent density, thickness with 1 gram load and 2 gram load, resistance to compression, resistance to stretching, surface friction, modified vertical loop deformation, drooping angle and BLC softness. The parameters from each measurement were all scaled on range 0 to 1 before being fed into network models. Artificial neural networks were developed through learning from the TFL examples and then tested on the PITTARDS collection. In the training stage, parameters from physical measurements and attribute gradings provided by descriptive sensory analysis were fed into the networks as input and desired output respectively. In the testing stage, physical measurement parameters were input to the trained network and the output of the network, which was the prediction of the leather handle, was compared with the gradings given by the panel. The testing results showed that the neural network models developed were able to judge the handle of a newly presented leather as well as an expert. Statistical methods were explored in the development of artificial neural network models. Principal Component Analysis was used to classify the attributes of leather handle and demonstrated that the predominant and most representative attributes out of the six attributes were "stiff", "empty" and "smooth". A network model called physical2panel, predicting the above three attributes from three physical parameters was built up by adopting a novel pruning method termed "Double-Threshold" which was used to decide the irrelevance of an input to a model. This pruning method was based on Bayesian methodology and implemented by comparing the overall connection weight of each input to each output with the limitation of two thresholds. The pruning results revealed that among the sixteen physical parameters, only three of them, - the reading from BLC softness guage, the compression secant modulus and the leather thickness measured under 1 gram load were important to the model. Another network model, termed panel2panel, that predicts the other three attributes "firm", "high density" and "elastic" from the prediction of the model physical2panel was developed and also proved to work as well as a leather expert panel. The conception of a 3D handle space was explored and shown to be a powerful means of demonstrating the findings.