Unlabeled Sensor Anomaly Detection for Real-Time Device Failure

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional machine learning approaches fail to detect device failures in real-time due to the absence of annotated data and supporting information in industrial settings, making it challenging to identify abnormal behavior from unlabeled sensor observations.

Innovation Solution

A processor-implemented method using a stacked auto-encoder with optimized hyperparameters, applying a windowing technique to create a windowed dataset, followed by dimensionality reduction and density-based clustering to predict the operating status of devices as success, failure, or transition states.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional machine learning approaches are used for device failure detection, then annotated data and domain expertise are required, but such data and expertise are absent in industrial settings

Engineering Contradiction:
Improvedevice failure detection accuracyVSAvoidabsence of annotated data
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

Instead of requiring annotated data to train models (traditional approach), the patent inverts the approach by using unlabeled sensor data and having the model learn normal patterns automatically. The auto-encoder learns to reconstruct normal sensor data, and deviations from reconstruction indicate anomalies, eliminating the need for labeled failure data.

Inventive Principle:
Principle #13The other way round (Inversion)

Solution Approach 2:

The system performs self-service by automatically learning device behavior patterns from unlabeled sensor data without requiring domain expertise or manual annotation. The stacked auto-encoder and clustering algorithms autonomously identify normal operating patterns and detect deviations, making the system self-sufficient in industrial settings.

Inventive Principle:
Principle #25Self-service

2Measurement precision

If complex deep learning models are applied to unlabeled sensor data, then detection accuracy improves, but computational power requirements increase

Engineering Contradiction:
Improveabnormal behavior detection accuracyVSAvoidcomputational power consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent segments the complex detection task into multiple manageable components: (1) stacked auto-encoder for feature extraction from sensor data, (2) dimensionality reduction for compact representation, and (3) clustering for pattern recognition. This segmentation allows each component to be computationally efficient while maintaining high overall detection accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms high-dimensional sensor data into lower-dimensional latent representations through the auto-encoder and dimensionality reduction steps. This dimensionality change reduces computational complexity and memory requirements while preserving the essential features needed for accurate anomaly detection.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP3961338B1Detection of abnormal behaviour of devices from associated unlabeled sensor observations
Publication Date: 2024.01.03 TATA CONSULTANCY SERVICES LTD
  • EP3961338B1 patent drawingFigure 1
  • EP3961338B1 patent drawingFigure 2
  • EP3961338B1 patent drawingFigure 3

AI summary

Conventionally, detecting time when a device is going to fail in real time has been a real challenge given the associated constraints and requirements. Due to absence in any supporting information or annotated data, traditional approaches have failed to detection abnormality in devices. Present disclosure provide systems and methods for detecting abnormal behaviour of a device from associated unlabeled sensor observations wherein KPIs are computed based on unlabeled sensor observations of at least two sensor parameters and windowing technique is applied on modified dataset to obtain windowed dataset based on which hyper-parameters of deep learning-based auto-encoder are optimized to obtain set of embeddings. Dimensionality reduction technique is applied on the embeddings to obtain embeddings with reduced dimension. Density based clustering technique with hyper-parameters is applied on embeddings with reduced dimension and clusters) for unlabeled sensor observations are obtained. Cardinality is assigned to clusters) to predict abnormal behaviour of the device.