Using Artificial Intelligence for Thermographic Image Analysis: Applications to the Arc Welding Process

This study has focus on making a model with high reliability to classify thermography videos in arc welding, for achieving high accuracy through series of steps that we did carefully. Indeed, the main goal is recognizing three different thermography classes by classifying thermography videos of arc welding which will be done by labeling the videos and then it needs to measure the accuracy that shows how much classification is correct. We started by thoroughly preparing and preprocessing the thermography video data. Custom scripts were developed to extract thermal frames and convert them into a format optimized for neural networks. The convolutional layers of our model (InceptionV3) are trained to identify a wide variety of features. The idea in this work is that the lower layers of these networks can extract useful features that are enough to request to other tasks, even if they weren’t specifically trained on our data. In our project, the convolutional layers of the ImageNet layers are frozen, and it means that their weights are not updated during the training. This mechanism is useful and practical when the dataset is small, and this is exactly true for our dataset which is small. In this preprocessing we resized the frames, also we normalized the pixel values to make the dataset more diverse and powerful. These steps were crucial to ensure the model could accurately handle different thermal patterns and conditions. In this model we combined Convolutional Neural Networks (CNNs) with a type of RNN networks which is called GRU and is obviously better than LSTM because it has been improved and its structure and architecture are simpler than LSTM. This architecture is created to control the sequence data. The CNN layers pulled out some spatial features from the thermal frames, while the GRU layers obtained the time-based relationships between frames and then making the model especially efficient at processing video data. In GRU we have two important gates which are update gate and reset gate. In update gate it specifies that how much of the previous information should be passed to the future. In reset gate, it makes decision that how much of previous information should be forget. GRU has a hidden state that acts as the network’s memory. In the first layer of our network, we have a GRU layer that frame_features_input dimensions should be mentioned. Also, it says that what value should be for sequence. Let’s go through the constructure of danse layer which is known as a fully connected layer that connects every neuron from the previous layer to every neuron in the dense layer. The model was trained using the sparse_categorical_crossentropy loss function and the Adam optimizer, both chosen because they’re good at efficiently training deep learning models. Training involved multiple epochs, during which the model’s parameters were fine-tuned to reduce the loss function. To prevent overfitting and ensure the model could generalize well to new data, techniques like early stopping and regularization were used. We did some important adjustments which are setting the learning rate, batch size, and the model’s structure according to the number of layers and neurons that we had. During the process, we kept the model’s efficiency which we used a validation dataset that can focus on accuracy to make sure which was working efficiently. After finishing the training phase, the model reached a final accuracy of about 97%, showing how well the chosen architecture and training process worked.