Agricultural Machine Anomaly Detection for Unforeseen Obstacles

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing agricultural machine monitoring systems struggle to detect unforeseen obstacles in unstructured environments due to the need for extensive training on predefined object classes, leading to potential collisions and reduced operator trust in operator assistance systems and increased workload.

Innovation Solution

A control system that utilizes an autoencoder architecture to encode and decode image data, generating an anomaly map by comparing reconstructed images with input images, allowing for the detection of any anomalies in the environment without requiring extensive training on multiple object classes, and can control operational components such as steering and propulsion systems to avoid detected obstacles.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If object detection algorithms are trained on predefined object classes, then detection accuracy for common objects is improved, but the system fails to detect unforeseen objects in unstructured environments

Engineering Contradiction:
Improvedetection accuracyVSAvoidability to detect unforeseen objects
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

Instead of training the system to detect specific objects (positive approach), the patent inverts the approach by training the system to recognize normal environments and detecting anything that deviates from this normal state as an anomaly. This allows detection of unforeseen objects without requiring predefined object classes.

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

Solution Approach 2:

The system uses the environment data itself to define what is normal, without requiring external training data for each object class. The system serves itself by learning from the operational environment and automatically adapting to detect anomalies.

Inventive Principle:
Principle #25Self-service

2Adaptability or versatility

If extensive training data for multiple object classes is collected, then detection coverage is improved, but computational power and training time requirements increase significantly

Engineering Contradiction:
Improvedetection coverageVSAvoidcomputational power consumption
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the essential feature of detecting anomalies rather than detecting every possible object class. By taking out the need for extensive object-specific training, the system achieves good detection coverage with significantly reduced computational requirements.

Inventive Principle:
Principle #2Taking out (Extraction)

3Device complexity

If predefined object classes are used for detection, then system complexity is reduced, but the workload for operators increases and operator trust decreases

Engineering Contradiction:
Improvesystem complexityVSAvoidoperator workload
Core Design Contradiction:
Device complexityVSEase of operation

Solution Approach 1:

The anomaly detection system serves multiple detection purposes universally without requiring separate detection mechanisms for each object type. This multi-functional approach reduces operator workload as a single system handles diverse detection needs.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4239597A1Monitoring agricultural operations
Publication Date: 2023.09.06 AGCO INT GMBH
  • EP4239597A1 patent drawingFigure 1~2
  • EP4239597A1 patent drawingFigure 3
  • EP4239597A1 patent drawingFigure 4

AI summary

Methods and systems are provided for monitoring operation of an agricultural machine. Image data indicative of an input image of a working environment of the agricultural machine is receive and encoded utilising an encoder network to map the image data to a lower-dimensional feature space. The encoded data is then decoded form a reconstructed image of the working environment which is compared with the input image to identify anomalies within the working environment. One or more operable components associated with the machine may be controlled based on the identification of one or more anomalies.