A Dataset Fusion Algorithm for Generalised Anomaly Detection in Homogeneous Periodic Time Series Datasets

arXiv:2305.08197 · cs.LG, cs.AI, eess.SP · Submitted 2023-05-14 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "A Dataset Fusion Algorithm for Generalised Anomaly Detection in Homogeneous Periodic Time Series Datasets".

Jane: The gist The proposed approach, tested on a case study of 3-phase current data from 2 different homogeneous Induction Motor (IM) fault datasets using an unsupervised LSTMCaps NN,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Moving into how they actually do this, the paper introduces something called "Dataset Fusion," which is their novel algorithm for composing these periodic time series datasets <ref:2305.08197#pg1>. They are merging multiple homogeneous datasets into one unified dataset for anomaly detection models <ref:2305.08197#pg1>.

Jane: So, how does this fusion work in practice? The summary explains that it starts by down-sampling each signal to the sampling frequency of the lowest frequency dataset, which they do using a Fourier method involving a low-pass windowed-sinc filter <ref:2305.08197#pg1>.

Lu: After that re-sampling, they normalize each dataset using Z-score normalization to handle those varying motor currents so everything is on a comparable scale <ref:2305.08197#pg1>.

Meng: Then comes period batching, where they use a zero-crossing algorithm to find one period and then concatenate 'n' periods based on a parameter you set <ref:2305.08197#pg1>.

Tom: And here’s another key part: for each dataset, the batch order is shuffled randomly <ref:2305.08197#pg1>. They do this specifically to mitigate distribution shift and stop noise from one area of the signal dominating everything else <ref:2305.08197#pg1>.

Jane: They then construct a new signal by alternating batches from each dataset—taking one batch from A, then one from B, and so on—which creates that final unified training set <ref:2305.08197#pg1>. It sounds like they are carefully engineering the composition to ensure the model sees a balanced view of all fault types.

Lalam: So, the summary is about taking these separate signals, making them compatible through down-sampling and normalization, then intelligently mixing their periods in a shuffled alternating way to create one robust training set <ref:2305.08197#pg1>. It’s a clever way to represent population distributions <ref:2305.08197#pg1>.

The paper's summary: Tom: The paper really highlights two big improvements in their approach. First is the ability to fuse time series data, meaning one model can learn from diverse sensor inputs like different motor fault types simultaneously <ref:2305.08197#pg1>.

Jane: That’s powerful because it moves past models that are specialized for just one thing; this fusion allows a single model to be more versatile across all those different conditions <ref:2305.08197#pg2>.

Meng: I need to press on that efficiency point—that's what makes it practical for real-world deployment, isn't it? That huge reduction in power usage when you only use a fraction of the data is something we have to talk about when we're trying to ship things one.

The paper's improvements: Tom: So, wrapping up, the main thing is that this Dataset Fusion Algorithm for Generalised Anomaly Detection in Homogeneous Periodic Time Series Datasets shows a way to make AI models much more robust when they have to deal with multiple different sources of data <ref:2305.08197#pg1>.

Jane: It achieves this by creating a single unified dataset through careful down-sampling, normalization, and period batching, which allows the unsupervised LSTMCaps Neural Network to achieve an F1 score of zero point eight seven nine <ref:2305.08197#pg1>.

Conclusion: Tom: So we looked at "A Dataset Fusion Algorithm for Generalised Anomaly Detection in Homogeneous Periodic Time Series Datasets" and essentially we saw how they take multiple different motor fault datasets—like interturn shorts and broken rotor bars—and they fuse them into one dataset so an AI model can learn from both simultaneously.

Jane: Right, so the core idea is this novel Dataset Fusion algorithm that merges periodic signals into a single unified dataset to make anomaly detection more robust across all those different motor conditions.

Lu: The methodology involves down-sampling each signal to the lowest sampling frequency using a Fourier method with a low-pass windowed-sinc filter, and then Z-score normalization to handle those varying motor currents.

Meng: And then they do this period batching thing, using a zero-crossing algorithm to identify one period and concatenating 'n' periods based on some user parameter you set.

Tom: They also shuffle the batch order randomly for each dataset, which is smart because it helps mitigate distribution shift and stops noise in one area from dominating the whole signal.

Jane: They then construct a new signal by alternating batches from each dataset—one batch from A, then one from B, and so on—which creates that final unified training set for the unsupervised LSTMCaps Neural Network.

Lu: The results show they achieve an average F1 score of zero point eight seven nine across both datasets, which is better than models specialized in just one of those individual datasets <ref:2305.08197#pg1>.

Tom: And the efficiency part is really compelling because they prove that using only six point two five percent of the training data results in only a four point zero four percent decrease in performance, which means you get this high accuracy while cutting computational power by ninety-three point seven percent.

Jane: That's pretty impressive for deployment, isn't it? It shows that the fusion method isn't just about getting a marginally better score; it’s about making the system much lighter and more practical for real-world use cases.

Lu: Experiment three suggests that training on a higher number of samples from Dataset B, as opposed to Dataset A, is actually what leads to more generalized performance across the whole thing <ref:2305.08197#pg2>.

Meng: From my side, I'm interested in how scalable this fusion process is when you move beyond just two datasets; if we can maintain this efficiency while fusing ten different sensor types, that’s where the real production value comes in.

Tom: It really pushes the idea that we don't always need massive amounts of perfectly labeled data to get good generalization if you structure your input representation intelligently.

Jane: So, "A Dataset Fusion Algorithm for Generalised Anomaly Detection in Homogeneous Periodic Time Series Datasets" gives us a concrete way to build more versatile AI models without needing a massive training budget.

Lu: It’s a strong time-series approach that addresses the challenge of achieving generalized performance across multiple homogeneous data sources <ref:2305.08197#pg2>.

Meng: It's solid for initial prototyping, but we'll need to see how it handles drift in real-time streaming scenarios before we can say it’s fully production ready.

Tom: Yeah, so the next big thing is seeing if this fusion concept works when the signals aren't perfectly periodic anymore—we should probably check out some of those other papers on time-series generalization next.

cs.LG, cs.AI, eess.SP

Submitted: 2023-05-14

Updated: 2023-05-14

Comments: This work has been submitted to the IEEE for possible publication

DOI: 10.1109/ACCESS.2023.3326725

Code: https://github.com/http:/arxiv.org

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: The gist The proposed approach, tested on a case study of 3-phase current data from 2 different homogeneous Induction Motor (IM) fault datasets using an unsupervised LSTMCaps NN, significantly

Key concepts

Dataset Fusion: This is a novel algorithm designed to combine several homogeneous time series signals into a single dataset. The goal is to merge data from different motor fault types so that one neural network can learn general patterns instead of needing separate models for each specific fault type.
Period Batching and Construction: The process involves identifying a single cycle in each signal, then grouping multiple cycles together into batches. These batches are then randomly shuffled and interleaved to create new training samples, which helps prevent noise from one specific data source from dominating the learning process.
Green AI Efficiency: The method significantly reduces computational needs while maintaining accuracy. By using only a small fraction of the total training data (6.25%), the model still performs well, resulting in a massive 93.7% reduction in computational power compared to conventional methods.

Terminology

F1 Score — 1× in this paper · explained in 4 episodes
Neural Networks — 1× in this paper

Summary

The gist The proposed approach, tested on a case study of 3-phase current data from 2 different homogeneous Induction Motor (IM) fault datasets using an unsupervised LSTMCaps NN, significantly outperforms conventional training approaches with an Average F1 score of 0.879 and effectively generalises across all datasets

Dataset Fusion Algorithm

The paper introduces Dataset Fusion, a novel dataset composition algorithm for fusing periodic signals from multiple homogeneous datasets into a single dataset while retaining unique features for generalised anomaly detection. This approach is designed to merge multiple homogeneous, periodic time series datasets into a single unified dataset for training anomaly detection NN models. The fusion process is engineered to accurately represent population distributions and increase robustness against potential data distribution shifts.

Dataset Fusion Process

The Dataset Fusion algorithm involves several key steps to create the fused dataset. The process begins by down-sampling each signal to the sampling frequency of the dataset with the lowest sampling frequency, expressed in Equation 1. This re-sampling is implemented using a Fourier method, which involves applying a low-pass windowed-sinc filter and then inverse-transforming back into the time domain. Each dataset is then normalised using Z-score normalisation to overcome varying motor currents.

Period Batching and Construction

To batch the periods together, a zero-crossing algorithm is employed to first identify a single period, and then concatenate n periods based on the user-defined parameter. For each dataset, the batch order is then shuffled randomly in order to mitigate the effect of distribution shift and prevent noise in one area of the signal from being prevalent in other areas of the signal. A new signal is then constructed using the shuffled batches from each dataset by appending a batch from each dataset in an alternating fashion.

Experimental Validation and Green AI

The algorithm was tested on a case study of 3-phase motor current data, where Dataset A contained fault data for an interturn short circuit fault and Dataset B contained current signals for a broken rotor bar fault. Results show that using only 6.25% of the training data, translating to a 93.7% reduction in computational power, results in a mere 4.04% decrease in performance, demonstrating the advantages of the proposed approach in terms of both performance and computational efficiency. This demonstrates the advantages of the proposed approach in terms of both performance and computational efficiency.

Performance Comparison

Experiment 1 demonstrated that the DF method has the best performance overall, with a best Average F1 score of 0.879, and an average of 0.821 from 10 runs. The results imply a considerable advantage in fusing the datasets, as the performance on individual datasets significantly exceeds that of models specialised in each respective datasets. The DF approach remains superior even when reducing the number of data samples, with only a 4.04% reduction in performance despite using only 6.25% of the training data.

Conclusion and Future Work

The conclusions drawn from Experiment 3 suggest that there is a greater benefit to training on a higher number of samples from Dataset B, as opposed to Dataset A, in order to achieve a more generalised performance. The robust performance of the DF approach, combined with its ability to accommodate various experimental settings, positions it as a valuable tool for practical use cases. Future works might investigate the applicability of the algorithm for training classifier models or examine whether using the frequency domain representation of the fused dataset could enhance performance. The paper presents a time-series dataset composition approach called the DF algorithm, designed to address challenges in achieving generalised anomaly detection performance across multiple homogeneous data sources.

Acknowledgments

The authors would like to express their sincere gratitude to Voltvision for their substantial support of this research. The sponsorship has significantly contributed to the progress and achievements of this work. The authors received the BEng (Hons) degree in electronic and computer Engineering from Brunel University London, Uxbridge in 2021 and is pursuing the Ph.D. degree in Electronic and Electrical Engineering with Brunel University London, Uxbridge. Tatiana Kalganova received the B.Sc. (Hons.) and Ph.D. degrees. She is currently a Professor in intelligent systems and the Electronic and Computer Engineering Postgraduate Research Director with Brunel University London, Uxbridge, U.K. The paper's index terms are Generalisation, Dataset Fusion, Data Reduction, Anomaly Detection, Neural Network Training, Green AI, Time Series, Environmental AI. The paper was sponsored in part by Voltvision. The authors are Ayman Elhalwagy and Tatiana Kalganova. The paper was sponsored in part by Voltvision, the distribution of the data sample used for training is not fully representative of the true distribution population, then the input-output function that is represented by this sample will inevitably vary from the function of the population. This work was sponsored in part by Voltvision. The authors would like to express their sincere gratitude to Voltvision for their substantial support of this research. The paper is 2305.08197.

--- Page 1 ---

Experimental Validation and Green AI

Improvements for AI systems

Bold Fusion of Time Series Data: The Dataset Fusion algorithm can merge multiple homogeneous, periodic time series datasets into a single unified dataset for training anomaly detection NN models, allowing a single model to learn from diverse sensor inputs (e.g., different motor fault types) simultaneously.
Green AI Efficiency: The approach demonstrates computational savings by showing that using only 6.25% of the training data, translating to a 93.7% reduction in computational power, results in a mere 4.04% decrease in performance. This enables deployment on resource-constrained environments while maintaining high accuracy for anomaly detection.
Robust Anomaly Detection: The algorithm's effectiveness under non-ideal conditions is highlighted, suggesting the fused model can maintain performance across all datasets and effectively handle data distribution shifts by using a novel composition strategy rather than relying solely on denoising the training sample.

Abstract

The generalisation of Neural Networks (NN) to multiple datasets is often overlooked in literature due to NNs typically being optimised for specific data sources. This becomes especially challenging in time-series-based multi-dataset models due to difficulties in fusing sequential data from different sensors and collection specifications. In a commercial environment, however, generalisation can effectively utilise available data and computational power, which is essential in the context of Green AI, the sustainable development of AI models. This paper introduces "Dataset Fusion," a novel dataset composition algorithm for fusing periodic signals from multiple homogeneous datasets into a single dataset while retaining unique features for generalised anomaly detection. The proposed approach, tested on a case study of 3-phase current data from 2 different homogeneous Induction Motor (IM) fault datasets using an unsupervised LSTMCaps NN, significantly outperforms conventional training approaches with an Average F1 score of 0.879 and effectively generalises across all datasets. The proposed approach was also tested with varying percentages of the training data, in line with the principles of Green AI. Results show that using only 6.25% of the training data, translating to a 93.7% reduction in computational power, results in a mere 4.04% decrease in performance, demonstrating the advantages of the proposed approach in terms of both performance and computational efficiency. Moreover, the algorithm's effectiveness under non-ideal conditions highlights its potential for practical use in real-world applications.