In recent years, artificial intelligence (AI) has played a critical role in driving innovation across various
industries. Advances in vision and voice technologies continue to foster the development of large intelligent
models,
resulting in new use cases and improved user experiences.
There is also a growing demand for AI that can run on edge
devices with microcontrollers and microprocessors, which offer benefits like lower latency, reduced energy
consumption
and enhanced data privacy. Among those applications, time series data, a sequence of data points recorded at
consistent,
evenly spaced intervals over time, is often used to develop three types of primary tasks: anomaly detection,
classification and regression.
Applications that Require Time Series Data
Anomaly detection, as the name suggests, is designed to identify behaviors that are outside the expected, and rely
on
time series data to identify deviations from normal behavior, which can trigger alerts or emergency stops to
minimize
damage.
Classification trains models to recognize and categorize input data by learning patterns within. This includes
assigning
labels to data points during training, enabling the models to make informed decisions. Once developed, those models
can
effectively identify patterns observed in data and classify new inputs.
Regression tasks aim to predict continuous values from the data, such as battery life prediction based on historical
battery discharge data or predictive maintenance for motors. Temperature and vibration sensor data can be used to
predict the likelihood of potential failure over time. While there are many other use cases that benefit from
machine learning and AI development, time series is one of the most complex and dynamic types of data.
To advance AI on the edge, we are introducing the eIQ® Time Series Studio (eIQ TSS), a new tool in our eIQ AI and
machine learning development software family. eIQ TSS features an automated machine-learning workflow that
streamlines
the development and deployment of time series-based machine learning models across microcontroller (MCU) class
devices
such as the MCX portfolio of
MCUs and i.MX RT portfolio
of crossover MCUs.
Time Series Studio supports a wide range of sensor input signals, including voltage, current, temperature, vibration,
pressure, sound, time of flight, among others, as well as combinations of these for multimodal sensor fusion. The
automatic machine learning capability enables developers to extract meaningful insights from raw time-sequential
data
and quickly build AI models tailored to meet accuracy, RAM and storage criteria for microcontrollers. The tool
offers a
comprehensive development environment, including data curation, visualization and analysis, as well as model
autogeneration, optimization, emulation and deployment.
Step-by-Step Workflow for eIQ® Time Series Studio
Sample Application
To jumpstart development, sample applications and datasets are provided for three primary tasks: anomaly detection,
classification and regression. Details for each application and step-by-step instructions are provided in the tool
to
help developers get started with the development workflow.
Homepage of the eIQ Time Series Studio user interface.
Data Input
Data curation is important to ensure that the data is clean, organized and aligned. For example, when data is
collected
from multiple sensors in an outdoor environment (where data can be noisy due to environmental factors) and with
varying
sampling rates, aligning and synchronizing these data in a way that preserves temporal relationships is crucial for
accurate model performance.
Developers can define the number of channels and classes when importing their own custom time series data, and Time
Series Studio also provides
different data viewing options – raw, temporal, statistical and spectral.
The dataset input page within the eIQ TSS user interface.
Training and Optimization
Model training and optimization are easier when automated machine learning replaces the traditional manual iterative
development process for parameter tuning, model and algorithm searching. Generate models with one click and sort by
accuracy or by flash/RAM size. This reduces the model training and optimization time from several weeks to hours.
The training page within the eIQ TSS user interface.
Emulation
After training is complete, models can be tested and verified within a virtual edge environment using different,
unseen
test datasets. This replicates the target device environment, enabling developers to validate model performance and
accuracy before deploying it to the actual hardware.
The emulation page within the eIQ TSS user interface.
Deployment
Once the selected model is compiled, a custom library can be generated for the application. Using the library is
straightforward as it consists of just two API calls – one to initialize the model and another to run the inference.
It
can generate a library compatible with MCUXpresso and Code Warrior IDE.
The deployment page within the eIQ TSS user interface.
Data Intelligence
Users often import time series datasets based on their prior knowledge. However, without comprehensive data analysis,
this may hinder the effectiveness of the data used for training. For example, the sampling frequency might exceed
the
application’s requirement, or the amount of training data in each class may be unbalanced for a classification task.
To address these challenges, data intelligence is a utility that can be used to evaluate dataset balance and assess
the
importance of individual data channels. This utility not only detects data imbalance but also identifies redundant
channels that can potentially be removed for resource optimization. Additionally, the utility recommends optimal
sampling frequencies and window size, enabling users to refine their datasets for improved quality and more accurate
analytical results.
The data intelligence page within the eIQ TSS user interface.
In this example, we can identify:
- Two out of the twelve channels may not be required and can be removed to save resources
- The raw continuous data may have over sample rate, and it’s recommended to reduce it to 1/16
Based on the intelligent analysis, the user can modify the dataset used for future training and expect better
results.
Time Series Studio offers a seamless end-to-end solution, and it is designed to reduce the barrier to entry and time
required for developers, partners and customers to develop AI solutions with their data. Combining this new tool
with
NXP’s full range of MCU and applications processor products and NPUs that further accelerate AI workloads, we look
forward to enabling organizations of all sizes to leverage the power of AI to innovate and solve complex problems.
eIQ Time Series Studio is included as part of the eIQ toolkit starting with version 1.13.1.
