Weakly Supervised Transfer Learning for Multi-label Appliance Classification

Tanoni, Giulia; Principi, Emanuele; Mandolini, Luigi; Squartini, Stefano

doi:10.1007/978-3-031-24801-6_26

Cited by 4 publications

(16 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Differently, in [7], the authors adopt a weakly supervised multilabel approach to reduce the labeling effort to train a CRNN, using both weakly labeled data (labels provided for a group of consecutive samples, e.g., for a 4 hours period) and strongly labeled data (i.e., labeled sample-by-sample). Successively, a transfer learning approach based on weak labels has been proposed in [13].…”

Section: A Multilabel Appliance Classificationmentioning

confidence: 99%

“…We investigate a real-world scenario where the network model is initially trained on a large quantity of publicly available measurements, annotated with strong and weak labels. Then, only weakly annotated data are available, labeled by end-users in a target environment, to fine-tune the network [13] and to distil the less complex model. Note that in the proposed method, end users are asked only for weak information about their appliance usage, with a significant reduction of labeling effort compared to strong labels and improved performance compared to unlabeled data.…”

Section: Contributionsmentioning

confidence: 99%

“…Due to the availability of a large quantity of low-frequency electrical load measurements from smart meters, deep learning (DL) approaches have recently become popular, representing the current state of the art in NILM both for regression and classification tasks [5], [6], [7], [8], [9], [10], [11], [12], [13].…”

Section: Introductionmentioning

confidence: 99%

“…Transfer learning typically requires collecting new data directly from the target environment to fine-tune pretrained models, requiring the engagement of end users for data annotation, which is made simpler by weak supervision. Previous studies have demonstrated the effectiveness of weak labels in improving performance in both disaggregation [24] and multilabel appliance classification tasks [7], [13], and this study proposes using weak labels to jointly distil knowledge and reduce network complexity during transfer learning. The method uses a Convolutional Recurrent Neural Network (CRNN), which has been successfully used in a centralized NILM scenario [7].…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Knowledge Distillation for Scalable Nonintrusive Load Monitoring

Tanoni,

Stankovic,

Stankovic

et al. 2024

IEEE Trans. Ind. Inf.

Self Cite

View full text Add to dashboard Cite

Smart meters allow the grid to interface with individual buildings and extract detailed consumption information using nonintrusive load monitoring (NILM) algorithms applied to the acquired data. Deep neural networks, which represent the state of the art for NILM, are affected by scalability issues since they require high computational and memory resources, and by reduced performance when training and target domains mismatched. This article proposes a knowledge distillation approach for NILM, in particular for multilabel appliance classification, to reduce model complexity and improve generalization on unseen data domains. The approach uses weak supervision to reduce labeling effort, which is useful in practical scenarios. Experiments, conducted on U.K.-DALE and RE-FIT datasets, demonstrated that a low-complexity network can be obtained for deployment on edge devices while maintaining high performance on unseen data domains. The proposed approach outperformed benchmark methods in unseen target domains achieving a F 1 -score 0.14 higher than a benchmark model 78 times more complex.

show abstract

Section: A Multilabel Appliance Classificationmentioning

confidence: 99%

Section: Contributionsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Knowledge Distillation for Scalable Nonintrusive Load Monitoring

Tanoni,

Stankovic,

Stankovic

et al. 2024

IEEE Trans. Ind. Inf.

Self Cite

View full text Add to dashboard Cite

show abstract

“…In terms of the aforementioned manual annotations, under this approach, users would only need to indicate whether an appliance was used or not within a certain time window. Also, for the transfer learning procedure, in [41] weak supervision was demonstrated to be effective compared to a supervised strategy, especially in the practical scenario of acquiring labels from the user feedback. Considering the multi-label appliance classification task, a weak label is provided for an entire temporal segment of the aggregate signal indicating whether an appliance is ON or OFF within that segment.…”

Section: Introductionmentioning

confidence: 99%

A weakly supervised active learning framework for non-intrusive load monitoring

Tanoni,

Sobot,

Principi

et al. 2024

Integrated Computer-Aided Engineering

View full text Add to dashboard Cite

Energy efficiency is at a critical point now with rising energy prices and decarbonisation of the residential sector to meet the global NetZero agenda. Non-Intrusive Load Monitoring is a software-based technique to monitor individual appliances inside a building from a single aggregate meter reading and recent approaches are based on supervised deep learning. Such approaches are affected by practical constraints related to labelled data collection, particularly when a pre-trained model is deployed in an unknown target environment and needs to be adapted to the new data domain. In this case, transfer learning is usually adopted and the end-user is directly involved in the labelling process. Unlike previous literature, we propose a combined weakly supervised and active learning approach to reduce the quantity of data to be labelled and the end user effort in providing the labels. We demonstrate the efficacy of our method comparing it to a transfer learning approach based on weak supervision. Our method reduces the quantity of weakly annotated data required by up to 82.6–98.5% in four target domains while improving the appliance classification performance.

show abstract