SLMs for Root Cause Analysis

Flanders Make

Flanders Make is the strategic research centre for the manufacturing industry. Our mission is to strengthen the long-term international competitiveness of the Flemish manufacturing industry. That’s why we work together with SMEs and large companies on pre-competitive, industry-driven technological research, resulting in concrete product and production innovation in the vehicle industry, the manufacturing industry, and production environments.

Problem Context

Root Cause Analysis (RCA) is usually performed manually by technicians and engineers. It requires deep knowledge of the equipment and a thorough understanding of its functioning principles.

Goal of the thesis

The main challenge concerning RCA is the retirement of experienced workers and the long period required by new hires to get accustomed to the different equipment. The goal of this master thesis is to support the new staff in reaching high productivity as soon as possible, while they still settle into their new positions.

Technical approach

One solution to the aforementioned problem are language models. Large Language Models (LLMs) are widely used across different domains to generate answers to queries, sometimes by leveraging internal documents (through RAG) or after having been fine-tuned on a use case. Small Language Models (SLMs) are smaller versions of LLMs, which can run locally on edge hardware (such as PCs or laptops). While they may not be as capable as LLMs, they are suitable for specific tasks and do not require expensive hardware to run on, nor a subscription to a cloud service provider.

To overcome the limitations of SLMs, an agentic workflow based on collaboration will be devised. One SLM will act as the Subject Matter Expert (SME) of a certain equipment, and it will collaborate with others acting as Electrical and Mechanical experts (SE and ME). They will have to collaborate between them to answer prompts such as “What is the cause of the high vibrations in the machine” which are input by the operators. The answer of the group of agents will be a list of the X most probable causes of the problem.

To test the proposed workflow, the Modular multi-actuated system (MMAS) will be used as a test set-up. It mimics the functioning of a weaving loom. Please follow the link to find out more. There is already a dataset acquired from the platform, which include faults. A fault detector has already been developed, and its results can be used to devise prompts for the agentic LLMs, for test purposes.

Learning targets

Knowledge and skills gained and/or developed at the end of this thesis:

• SLMs: RAG, agentic workflows, collaboration
• ML Ops: LangChain, vector database
• Software development: Python scientific and ML stacks
• Project management: working in an Agile environment

Profile Student

• Bachelor degree in machine learning, signal processing, applied mathematics, computer science or similar;
• Programming: Python scientific and ML stacks (NumPy, SciPy, pandas, scikit-learn, Pytorch), optimization;
• Basic knowledge of LLMs, RAG, multi-agent systems;
• Passion for research, curiosity and eagerness to learn;
• Result oriented, responsible and proactive;
• Creative and open minded:
• A good communicator, able to communicate in English;
• Eager to learn and a team player.

Practical Data

• • Thesis: This assignment is a topic for a master thesis for a Belgian university.
  • Follow up moments take place at the offices of Flanders Make located in Kortrijk or Lommel , Belgium. Depending of the choice of the student.

File attachments