INTRODUCTION:
Copeland (1964) “Man consumes no material except water in such tremendous quantities”. It is no doubt that with the development of human civilization, concrete will continue to be a dominant construction material in the future. Versatility to make concrete from materials easily available and moulding into desired shape has made concrete the second largest material to be consumed. It is a composite material composing of coarse granular material embedded in hard material matrix (binder) which fills the spaces and binds the aggregates.
The ultimate goal in designing of concrete mixes is to obtain desired strength, durability, performance with economy. Among these requirements compressive strength plays a predominant role in designing concrete mixes. Artificial Neural Networks also termed as “Neural Nets”, “Artificial Neural Nets”, or ANN in short is a computation tool modeled on the basis of human nervous system or brain. The human nervous system consists of massively large parallel interconnection of neurons. These neurons work in smaller interval of time. These Neural Nets are ideally suited for a wide range of tasks where there are no available algorithm to complete the task. ANN is preliminarily trained to solve problems adapting a teaching method and data applicable. Based on training received ANN has the capability to generalize and thereby recognize similarities among different inputs provided. Neural nets are nonlinear massive parallel computational models. ANN consists of simple processors linked by weighted connections; these processing nodes are called as “Neurons”. The neurons are multiple-input, multiple-output systems (MIMO). They receive signals from the inputs provided which generates a resultant signal and thereby produce similar signals to all outputs. These neurons in an ANN are arranged as layers where the layer that interacts with the environment in receiving inputs is input layer and the final layers to present processed data is output layer. Several other layers in between these two are termed as hidden layers. The connections between neurons are called as synapses.
Overview of artificial neural networks:
The power of neural nets lies in their ability to generalize by learning and parallel processing. Uses of ANN have the following capabilities:
· Non-Linearity
· Adaptivity
· Input-Output Mapping
· Fault Tolerance
· VLSI implement ability
· Uniformity of Analysis and Design
· Contextual Information
· Neurobiological Analogy.
In ANN each neuron communicates its output to other node using activation functions. These networks may be a single or multi-layered one. Single layered networks have an input layer that projects onto output layer of neurons unlike multi-layered which have hidden layers. Basic methodology consists of 3 phases namely Network Training, Testing and Validation. In the process of training, algorithms are adapted to modify the weights of connections. This process is referred to as learning which is done by assigning weights and biasing the computed from a set of data available as per conditions applicable. ANN learns from examples and generalize for data apart from the training data. Then the test data is used to check generalization. Learning situation is categorized into Supervised, Unsupervised and Reinforced Learning. Supervised learning/Associative is where the network is trained by providing suitable input matching output patterns. In case of unsupervised learning, output is trained to respond to clusters of input patterns and in case of Reinforced, learning is intermediate since learning machine does some action on the environment and gets the response from the environment(Rogers (1994)). There are several Neural Network Architectures classified based on Biological Networks:
Full Connected:
Every node is connected to every other node.
Single-Layer feed forward:
one input layer that projects onto an output layer but not vice-versa.
Multilayer-feed forward:
network with hidden layers in between input and output.
Recurrent:
consists of feedback loops.
Modular:
tasks are done using smaller modules and then combined Back Propagation Network is used in layered feed-forward ANNs. Here neurons are organized in layers and signals are sent forward and errors are propagated backward. It uses supervised training. Number of hidden layer is fixed as per the application, complexity and the count of input- output samples. Input layer has neurons equal to the number of input samples and so is the output. Number of hidden layer is determined by trial and error based on experimental data. Some important Terminologies include:
Feed Forward Computation:
input vector representing the pattern to be recognized is incident onto input layer and subsequently distributed to the hidden layers and transferred to output layer in the form of weighted connection.
Error Back-Propagation:
Networks deal with supervised values, hence first aspect required to be recognized in the training is the need for measuring of classes of network to produce target values. The measure of this value is termed as network error. In Back-propagation learning algorithm the error measure is found using mean square root given by
Ep=
Tpj-Opj)2
where
Tpj= Target Value of jthOutput unit for pattern p
Opj= Actual Output for jth unit for pattern p
Problem Definition:
Design of concrete mixes is done by certain empirical relationships between design parameters and experiments. A normal concrete mix of desired strength is achieved by several trials. Hence ANN is used for this problem where no solution algorithm is available. The feature of ANN in obtaining relationship between output and input is utilized to establish relationship between the sample data. This will considerably reduce the number of trails. For developing this Neural nets, sufficient set mixing proportions with corresponding strengths, water content, fineness modulus are required in training the network.
MATERIALS AND METHODS:
In the present work implementation and training is developed using MATLAB. Neural network toolbox in MATLAB version 7.7.0 is utilized for the study. A multilayer perceptron approach is used here in developing this model. Back Propagation training algorithms is used. There are several activation functions available which include Threshold, Ramp, Sigmoid and Gaussian. The activation function used here is the log-sigmoidal function. Sigmoid is a function whose graph is S shaped. It is an increasing function and balances between linear and non-linear.
Fig. 1 Logistic Curve
Curve shown in Fig. 1 is defined by the formula
S(t)=
The procedure that follows in determining the output is as follows:
Step 1: Sum of weighed inputs is found
Step 2: Transform weighed inputs using above formulae
Step 3: To sum hidden node outputs
Step 4: Transform weighed sum
Learning rate affects the speed of arriving at solutions. In this back propagation the learning rate is analogous to step-size parameter from gradient-descent algorithm. Momentum parameter is used to prevent the system from convergence to a local minimum or saddle point. High momentum parameter is also used to increase the speed of convergence in the system. The values of learning rate L and momentum M for the study is 0.2 and 0.1. Training time T=400 for each network. Normalization is done using WEKA used in carrying out ANN analysis. In pre-process minimum, maximum, mean and standard deviation for each of the features is computed and used in sigmoidal function transformation. Hence this helps in maintaining values within standard deviation of mean. Normalized range is in between 0 and 1. Cross-Validation technique with folds grossing 20 is implemented which is a standard tool in analytics and is important for fine-tune of data mining models. Success of the model to predict the 7, 14 and 28 days compressive strength depends on training data and magnitude. Predicted strengths are compared with the actual strengths. Training of the network was carried out using a set of inputs and corresponding output data. Input and the target accurate values are obtained from various available sources. Table 1 contains the input and target output data followed and Table 2 contains the range of input data.