Algorithms
In our business, we offer the power of machine learning algorithms in conjunction with sensor technologies to provide innovative solutions across various industries. These algorithms serve as the backbone of our systems, enabling us to extract valuable insights from sensor data and drive actionable decisions. Leveraging state-of-the-art machine learning technology allows us to unlock the full potential of sensor networks, leading to enhanced efficiency, accuracy, and predictive capabilities in our offerings to customers.
Among the myriad of machine learning algorithms available, we carefully select the most appropriate and tailor them to our customers' business needs. Examples of algorithms we frequently use include:
- The family of Decision Trees, Random Forest, and Gradient Boosting Machines are relevant for both spot measurements and time-series data as they can be used to classify different types of radio signals or detect patterns in signal data over time. The decision tree creates tree-like structures for decision-making, which is interpretable for sensor-based rules. The random forest is an ensemble of decision trees which is robust for complex sensor data. The isolation forest detects anomalies (e.g., intrusion detection, outlier identification) and is effective for identifying unusual sensor readings. The gradient boosting machine boosts weak models (e.g., predicting water quality parameters) and handles noisy or sparse sensor data.
- Support Vector Machines (SVM) are relevant for both spot measurements and time-series data for classification tasks and anomaly detection in radio signal data. It separates data into classes (e.g., identifying different gas emissions) and is useful for multi-class classification.
- Artificial Neural Networks (ANN) / Deep Learning including Convolutional Neural Networks (CNN) are relevant for both spot measurements and time-series data for modelling complex relationships in radio signal data and predicting signal characteristics and using the CNN for image-like pattern recognition and classification tasks. The ANN handles complex patterns (e.g., biometric recognition, image processing) and large-scale sensor data with deep architectures. The CNN uses image and signal processing (e.g., image-based smoke detection) which is excellent for visual and spectral data and data that is pre-processed with, for example, FFT or wavelet transforms.
- Variational Autoencoders (VAEs) are relevant for both spot measurements and time-series data for anomaly detection and feature learning in radio signal and general sensor data.
- Transformers are highly important for time-series data from radio signals for capturing long-range dependencies and identifying specific traits or patterns in signal sequences.
- RNN & Long Short-Term Memory (LSTM) Networks, these are both highly relevant for time-series data from radio signals for modelling temporal dependencies and predicting future signal characteristics. The RNN is useful with temporal dependencies (e.g., time series prediction, motion sensing), and captures sequential sensor readings. The LSTM predicts future values (e.g., electric current levels) and also handles time-dependent sensor data.
- Kalman Filters are specifically useful with sensor fusion (e.g., GPS localization, tracking), and is adapt at combining noisy sensor measurements.
- Gaussian Mixture Models (GMM) work by clustering (e.g., light source detection from spectral data) which identifies sensor data clusters.
- Wavelet Transform uses signal analysis (e.g., RF signal interrogation) and extracts features from sensor signals used either directly or in conjunction with other algorithms.
