Recently, deep learning models such as convolutional neural network (CNN), convolutional autoencoder (CAE), CNN-based support vector machine (SVM), YOLO, fully convolutional network (FCN), fully convolutional data description (FCDD) and so on have been being applied to defect detections and anomaly detections of various kinds of industrial products, materials and systems. In those models, downsampled images including target features to be fitted to the resolution of each input layer are basically used for training and testing data. In this paper, intelligent anomaly diagnosis system for numerical control (NC) machine tools is considered, in which a simple microphone is positioned as the only sensor we use. Generally, mechanical sound and vibration generated from a machine tool itself or machining sound and vibration generated from a router bit, i.e., end mill cutter attached to the spindle head can be recorded and used for training NN models. Dataset used in this paper consists of operating sounds recorded from five different types of machines using a smartphone microphone. Since cost reduction is generally required in building systems, no special vibration or acceleration sensors are used. For experimental evaluation, nine kinds of mechanical sounds are collected from the five machine tools, and then training datasets consisting of sound blocks are prepared. Each sound block (SB) is time series data extracted from WAV (Waveform Audio File Format) files (.wav). For example, if a WAV file is recorded with a sampling rate 44100 [Hz] and an extracted time for forming a SB is set to 0.005 [s], then the data length of the sound block approximately becomes 220. The extracted SBs from a WAV file are employed for training three types of NN models for classification. As for the NN models for comparison, conventional shallow NN, RNN and 1D CNN are designed and trained using the nine kinds of mechanical sounds. Classification results of test SBs by the three models are shown. Then, an autoencoder is designed and considered for identifier by training it using only SBs of a machine tool. One of the technical needs in dealing with time series data such as SB data by NN is how to clearly visualize and understand anomalous regions in concurrent with identification. In this paper, finally, we propose the SB data-based FCDD model to meet this need. The superiority of the FCDD model is verified in terms of anomaly detection accuracy and concurrent visualization of understanding.