Time-series classification identifies specific categories, operational states, or predefined labels within time-series data, and is widely applied across domains including healthcare, wireless monitoring, smart security, and automated testing.

The waveform classifier solution implements a lightweight, 13K-parameter time-series classification model (CLS_13k_NPU), designed for efficient deployment on embedded neural processing units (NPUs). The design employs a 256-point Real Fast Fourier Transform (RFFT) feature extraction pipeline combined with 8-frame concatenation per AI inference task, accepting a 128-element 1D time-series feature vector as input and producing a 3-element output vector where each element represents the confidence score of the corresponding class.

The exported model file outputs an INT8 (8-bit quantized) tensor from the final fully connected layer. Since the raw output consists of per-class confidence scores in quantized form, a post-processing step is required to derive the final classification result.

Waveform Classifier Model Architecture

The mode network is structured as a sequential feature transformation pipeline, where each stage builds directly upon the output of the previous to progressively refine and abstract the input representation toward a final classification decision. Figure 5-2 shows the workflow.

Figure 5-2 Time-series Classification Model (CLS_13k) Flow Chart

The pipeline begins with a Batch Normalization Layer, which standardizes the raw input sequence in real time, ensurinsg consistent feature scaling regardless of input signal amplitude variations before any learned transformation is applied. With the input stabilized, the network then progressively deepens the feature representation through six Consecutive Convolutional Blocks. Each block follows a consistent internal structure - a Conv2D Layer with bias, followed by Batch Normalization and ReLU activation, forming a repeating pattern that captures increasingly abstract temporal patterns across successive stages. The Batch Normalization within each block maintains training stability as depth increases, while the ReLU activation introduces the non-linearity necessary for the model to learn complex signal characteristics.

After the six convolutional stages, the resulting feature map passes through an Adaptive Average Pooling Layer, which collapses the spatial resolution to a fixed size of 1x1 regardless of the preceding temporal length. This design is particularly significant: by decoupling the model from any fixed temporal dimension, this inherently grants the model compatibility with varying time-series input lengths, providing flexibility for future deployment scenarios without requiring architectural changes.

The fixed-size pooled feature map is then flattened into a 2D tensor and passed through a final Fully Connected (FC) Layer, which maps the compressed, high-dimensional feature representation directly to the target class output scores.

Model Output and Post-Processing

The exported model file outputs an INT8 (8-bit quantized) tensor from the final fully connected layer. Since the raw output consists of per-class confidence scores in quantized form, a post-processing step is required to derive the final classification result. Typically, this is accomplished by applying an argmax operation over the output vector, which identifies the class index carrying the highest confidence score as the predicted classification outcome.