Bi-LSTM Alarm Sequence Prediction for Long-Horizon Operator Guidance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current alarm prediction systems are limited in predicting long duration alarm sequences and rely on LSTM architectures that fail to effectively learn dependencies between input and output sequences, resulting in inaccurate and short-term predictions, which are insufficient for proactive maintenance and operational efficiency.

Innovation Solution

The implementation of a Bidirectional Long short-term memory (BiLSTM) neural network with an encoder-decoder technique that uses true output sequences as input during training, allowing for the prediction of longer alarm sequences and the generation of a mapping table to recommend operator actions based on a Matching Sequence Score (MSS) for optimal preventive measures.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Duration of action of moving object

If LSTM architecture is used for alarm prediction, then the system can process sequential alarm data, but the prediction duration is limited to minutes and accuracy deteriorates for longer sequences

Engineering Contradiction:
Improveprediction durationVSAvoidprediction accuracy
Core Design Contradiction:
Duration of action of moving objectVSMeasurement precision

Solution Approach 1:

The alarm sequence prediction problem is segmented into multiple time steps, where the model predicts one alarm at a time in sequence. The BiLSTM processes the input alarm sequence in segments, generating predictions step-by-step for extended time horizons while maintaining accuracy at each prediction step rather than degrading over the entire sequence.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from unidirectional LSTM to bidirectional LSTM, adding a temporal dimension by processing sequences in both forward and backward directions. This dimensional change allows the model to capture dependencies from both past and future contexts, enabling longer prediction durations without sacrificing accuracy.

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

2Device complexity

If LSTM uses only past information for training, then the model architecture remains simple, but the ability to learn dependencies between input and output sequences is insufficient

Engineering Contradiction:
Improvemodel architecture complexityVSAvoidsequence dependency learning
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent applies the inversion principle by using bidirectional LSTM that processes sequences in both forward and backward directions, rather than only forward as in traditional LSTM. This allows the model to learn dependencies by considering future context during training, significantly improving sequence dependency learning while maintaining manageable architectural complexity.

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

3Loss of time

If the forecasting window is small (minutes), then the LSTM model can provide predictions, but operators have insufficient time to plan and execute preventive measures

Engineering Contradiction:
Improveresponse timeVSAvoidpreventive action effectiveness
Core Design Contradiction:
Loss of timeVSEase of manufacture

Solution Approach 1:

The BiLSTM model performs preliminary predictions of alarm sequences extending several hours into the future, allowing operators to take preventive actions before faults actually occur. By predicting the sequence of alarms in advance, the system enables operators to plan and execute mitigation strategies proactively rather than reactively.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4481520A1Long duration alarm sequence predictions using bi-LSTM and operator sequence recommendations thereof
Publication Date: 2024.12.25 TATA CONSULTANCY SERVICES LTD
  • EP4481520A1 patent drawingFigure 1A
  • EP4481520A1 patent drawingFigure 1B
  • EP4481520A1 patent drawingFigure 2A

AI summary

Techniques used by the state of the art alarm prediction systems rely mostly on LSTM, which have technical limitation in predicting a sequence with long durations. Embodiments of the present disclosure provide a method and system for long duration alarm sequence predictions from past alarm sequence using a Bi-LSTM and operator sequence recommendations thereof. The BiLSTM with an encoder decoder technique uses true output sequence as input to decoder at each time step during training. This allows the BiLSTM to learn dependencies among input and output sequence effectively. The operator sequence recommended is identified based on closeness of the predicted future output alarm sequence with one among the unique alarm sequences in a mapping table. The alarm sequence closeness is computed using a Matching Sequence Score (MSS) disclosed by the method, since known sequence evaluation metrics such as Blue Score has limitations to be directly applied in alarm sequence evaluation.