Titulo Estágio
Expainability Improvement in Fault Detection for Industrial Machinery Using Supervised Learning Sound Identification Models
Local do Estágio
Laboratório de Informática Industrial e Sistemas do DEI/CISUC (LIIS@DEI-FCTUC)
Enquadramento
Diagnostic routines in household appliances suffer from low explainability because they merely identify the effects of the failure mode without the capability to pinpoint the root cause, including the specific failed part and occurrence of the failure.
The issue of explainability is even worse regarding mechanical failures, since the most common diagnostics have low efficiency due to the lack of a specific sensor to measure the system dynamics and usually rely on indirect measurements.
This lack of explainability severely hampers the appliance repair process, since technicians struggle to understanding the underlying causes of the malfunction, leading to a lack of trust in the diagnostic results. Consequently, technicians may resort to arbitrary repair procedures, increasing the likelihood of human errors.
Nevertheless, the opposite scenario also presents challenges. Enhancing explainability by presenting technicians with all sensor readings would overly complicate the diagnostic process. Technicians may find it overwhelming to analyze numerous readings, thereby also increasing the likelihood of errors.
Efficient explainability is paramount for automated and accurate diagnosis and repair of household appliances, avoiding financial and consumer experience losses.
This project proposes the use of audio recordings of noises emitted by industrial machines as a method of identifying the root cause of failures, identifying the specific part and occurrence leading to the malfunction. Also, the audio recording is captured using a common external device (such as a smartphone) for more practical implementation.
Different supervised learning models will be compared to achieve the best explainability along with a satisfactory accuracy. Annotation of audio recordings will be performed manually, mimicking the investigative process employed in diagnosing faults in industrial machinery.
Taking into consideration real-world applications, where acquiring noise samples is resource-intensive and time-consuming, the training method will preferably utilize small audio recording datasets.
A subset of the MIMII dataset [1] [2], containing diverse industrial machine noises, will be used in this project. Despite not directly utilizing audio samples from household appliances, the selected industrial machine noises will closely resemble those of household appliances. The project will ensure the extrapolation of findings for application in household appliances.
[1] https://zenodo.org/records/3384388
[2] Purohit, Harsh, et al. "MIMII Dataset: Sound dataset for malfunctioning industrial machine investigation and inspection." arXiv preprint arXiv:1909.09347 (2019).
Objetivo
The objective of this project is to improve explainability in fault diagnosis using noise identification in industrial machines without compromising accuracy and maintaining low complexity in practical application.
This study aims to assess explainability, accuracy, and application complexity among various supervised noise identification models, while comparing them with the baseline.
Additionally, the proposal seeks to establish a parallel between noise identification in industrial machines and its potential application in diagnosing faults in household appliances.
Plano de Trabalhos - Semestre 1
1st Semester
1. Study state of the art models for fault detection in industrial applications.
2. Study state of the art methods for audio identification concerning fault detection (industry application).
3. Define implementation requirements and evaluation procedures for the proposed models.
4. Define the dataset used (as a subset of the MIMII dataset).
5. Develop model prototypes.
6. Produce a report with initial results and assess the full feasibility implementation.
Plano de Trabalhos - Semestre 2
2nd Semester
7. Develop final models and fine-tune them.
8. Conduct analyzes and tests of experimental models.
9. Evaluate accuracy and comparative explainability analysis for each model.
10. Write a scientific article.
11. Write the final documentation and dissertation.
Condições
The work will be developed within the scope of the PRR CRAI project, and a workplace will be made available at LIIS-CISUC.
Observações
The internship may benefit from the award of a research grant for graduates for a period of 6 months, possibly renewable, supported by the ongoing project.
Orientador
Alberto Cardoso
alberto@dei.uc.pt 📩