Anomaly detection for refrigeration systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional anomaly detection techniques for refrigeration systems are inaccurate and fail to account for the nuances of refrigeration systems, leading to unnecessary maintenance and unplanned expenses due to late failure detection or over-maintenance.
Innovation Solution
A self-supervised machine learning model, such as an LSTM autoencoder, processes telemetry data including temperature and setpoint values to generate periodic anomaly metrics, accurately predicting equipment failure by learning the behavior of refrigeration systems through defrost periods and other operational modes, reducing false positives, and providing predictive maintenance schedules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional anomaly detection techniques are used for refrigeration systems, then the system can detect failures, but the detection accuracy is low leading to false positives and unnecessary maintenance
Solution Approach 1:
The patent transforms the anomaly detection approach by changing from conventional rule-based parameters to machine learning models that process multiple telemetry parameters (temperature, humidity, door openings, compressor runtime) simultaneously. This parameter transformation enables the system to distinguish between normal variations (like defrost cycles) and actual anomalies, reducing false positives while improving detection accuracy.
Solution Approach 2:
The patent replaces conventional mechanical anomaly detection methods with machine learning-based detection systems. Instead of using simple threshold-based rules, the system employs trained models that analyze patterns in telemetry data, substituting the mechanical detection approach with an intelligent system that adapts to refrigeration system behavior and reduces unnecessary maintenance trips.
2Reliability
If conventional condition-based maintenance schedules are used, then maintenance can be planned, but the schedules produce many false positives and do not account for refrigeration system nuances
Solution Approach 1:
The patent implements self-service through automated machine learning models that continuously analyze telemetry data and generate maintenance predictions without human intervention. The system automatically learns refrigeration system nuances (such as defrost cycles and operational patterns) and adjusts maintenance schedules accordingly, eliminating the need for complex manual scheduling while improving reliability through data-driven insights.
Solution Approach 2:
The patent incorporates feedback mechanisms where the machine learning model continuously receives telemetry data from refrigeration systems, analyzes patterns, and refines its predictions over time. This feedback loop allows the system to learn from actual system behavior and maintenance outcomes, improving scheduling reliability while the automated nature of the feedback process keeps system complexity manageable.
3Productivity
If reactive maintenance is performed, then maintenance is only done when system fails, but this results in loss of productivity and unplanned expenses
Solution Approach 1:
The patent applies preliminary action by using machine learning models to predict equipment failures before they occur. The system analyzes telemetry data patterns that precede failures and generates early warnings, enabling maintenance to be performed proactively during planned downtime rather than reactively during operational failures, thus maintaining productivity while reducing unplanned downtime.
Data Source
AI summary
Methods and systems are described for anomaly detection in refrigeration systems. A process for providing anomaly detection for refrigeration systems includes receiving telemetry data of one or more refrigeration systems, including measured temperature values and setpoint temperature values; processing the telemetry data to determine machine learning input data based at least in part on at least a portion of the measured temperature values and at least a portion of the setpoint temperature values; and using one or more hardware processors to apply the machine learning input data to a trained anomaly detection machine learning model to determine periodic anomaly metrics. The process provides an automatically determined indication based at least in part on at least a portion of the periodic anomaly metrics.