Edge MLOps: Dynamic fine- tuning of predictive maintenance models

By Gabriel Antonio Valverde Castilla , Machine Learning Lead of DATAIS

The industrial world is evolving rapidly, driven by advanced solutions that go beyond mere prediction. The need to directly intervene in systems and automate maintenance decisions has led to the adoption of a more dynamic and adaptive approach, especially in the field of predictive maintenance (PM). This article delves into the practical and technical essence of dynamic model fitting for PM, merging the power of causal models, reinforcement learning (RL), and the latest trends in distributed modeling.

Technological advances herald the dawn of a new era in industrial systems management, where anticipation and proactive action stand as fundamental pillars to optimize efficiency and avoid costly disruptions.

In this context, Barbara and DATAIS are at the forefront of this revolution, showing how the transition from predictive maintenance to prescriptive maintenance can radically transform the management of critical systems.

In the framework of the recent event: "Adaptive AI in Industrial Systems" ACCIONA presented its latest successful case of Automatic Fine-Tuning of distributed models on the Edge together with Barbara . I highlighted the inherent challenge of going beyond simple failure prediction, towards the ability to make automatic and preventive decisions regarding systems. The main objective lies not only in monitoring and detecting potential problems but also in intervening before they occur, driven by causal models that combine probabilistic graphical predictive models and expert information based on differential equations. This article aims to review and elaborate on these fundamental points that emerged at the event.

Challenges and Solutions in Predictive Maintenance

Imagine being able to predict when a membrane in a water plant is so degraded that it needs to be replaced before it fails. That's precisely what Predictive Maintenance aims for anticipating potential faults in industrial systems. However, the ambition goes further: Prescriptive Maintenance aims not only to predict but also to take preventive action to stop those errors from happening. This approach isn't solely about detecting issues but actively enhancing system performance.

DATAIS 's approach centers around a specific case study: predictive maintenance applied to the steel furnace environment. This environment, characterized by the buildup of isolating residues, poses critical challenges for system efficiency and safety.

For instance, in a steel manufacturing furnace, "Fouling," the accumulation of residues isolating the furnace walls, can lead to costly failures. This phenomenon can be prevented by making decisions based on understanding the relationships between variables like temperature, used reagents, and various indicators of raw material quality.

The key to addressing these challenges lies in combining real-time data, sensors, and the valuable input of expert knowledge. Implementing causal models, which can alleviate data scarcity and ensure decision-making guided by relational structures, becomes a fundamental pillar.

Causal Models and Graphs in Action

These models, often represented in graph structures, enable more informed decisions and mitigate the risk associated with dynamic decision-making based on predictions in complex industrial systems.

Causality-based reinforcement learning emerges as a valuable tool in this context. This approach allows building an artificial intelligence capable of making safer decisions, considering the inherent uncertainty in industrial system dynamics and providing the ability to adjust how far it deviates from current expert decisions.

Practical Implementation with MLOps and Reinforcement Learning

Implementing these techniques in distributed industrial environments poses an additional challenge of synchronization. In this context, Barbara 's Edge MLOps Platform emerges as a crucial solution by enabling real-time data collection, adaptable model training, and efficient deployment across multiple plants.

The synergy between MLOps and Reinforcement Learning is crucial to automate and continually enhance model performance. The ability to make decisions based on measurements of uncertainty and risk in outcomes represents a balanced and essential approach in dynamic and complex environments such as steel manufacturing.

Regarding the causal model, its utility lies not solely in relying on data but in understanding the interrelationships among measured indicators. By constructing a solution with expert knowledge, from differential equations to insights from on-site personnel, a reinforcement learning system capable of making lower-risk decisions can be developed.

Exploring and evaluating potential decisions in a dynamic machine learning system aims to identify optimal ones (e.g., maximizing product quality). However, this exploration can carry high business risks. Measuring uncertainty in decision-making allows discarding options that, despite seeming optimal, exhibit significant uncertainty levels regarding possible outcomes.

In the specific context of a steel manufacturing system, the issue of "Fouling" in furnaces is addressed. This phenomenon, involving the accumulation of soot during the reactive process, acts as an insulator and can lead to temperatures surpassing the furnace walls' melting point, causing costly damages and production halts.

Caution against this risk affects both the efficiency and the quality of the process, restricting the selection of certain chemicals that accelerate soot formation. This is where expert knowledge comes into play: the differential equations and relationships between indicators translate into graph structures forming causal models. These models understand how various variables, such as temperatures and chemical characteristics of different reagents, mutually affect each other and influence the reactive process of the furnace.

This approach unfolds in three proposals:

1. Based on expert knowledge
2. Integrating data with expert knowledge
3. Exclusively focused on data.

Each has its advantages and limitations, from performing well in a specific plant to adapting to multiple locations with different conditions, more efficiently integrated into the Barbara platform.

Comparing the application of reinforcement learning based on causal methodology shows how these decisions influence system performance. The proposed models, in most cases, do not surpass the melting temperature, whereas traditional models, despite mitigating errors, exhibit recurring failures directly impacting profits, thus highlighting the robustness and transferability of the expert knowledge-based solution.

Results: Effectiveness and Traceability

The data presented earlier demonstrate a clear advantage in the effectiveness and transferability of your prescriptive approach compared to conventional methodologies. Minimizing errors and adaptability to different environments are key points for success.?

The ultimate goal is the proactive automation of industrial processes, eliminating the fear of catastrophic failures and enabling optimal and safe performance in critical industrial environments.

To delve deeper into the practical implementation of these solutions, let's consider each deployment option in detail:

• Local Deployment: Implementing specific models for each plant offers advantages in adaptability and local monitoring capability. However, scalability might pose a challenge, and transferring to other locations may require significant adjustments.

• Data Centralization: Data centralization within a unified platform provides a comprehensive view across multiple plants, easing model implementation and applying expert knowledge at scale. However, it may pose challenges in terms of adapting to specific conditions at each plant.

• Centralized Dynamic Adjustment: This more complex strategy, which combines data centralization with specific dynamic adjustments for each plant, provides a more versatile and adaptable solution. The ability to make dynamic adjustments and real-time data-driven updates offers a significant advantage in continuously optimizing systems.

Exploring the practical application of these strategies in specific industrial contexts can further illustrate the effectiveness and versatility of each approach. Let's examine, for example, how these strategies are applied in sectors such as the steel industry, power production, or advanced manufacturing.

Conclusions and Future Perspectives

Automation and proactive intervention in industrial systems are the future.

The goal is to alleviate the fear of catastrophic failures, enabling more efficient and safer operations management.

The combination of causal models, reinforcement learning, and Edge MLOps Platforms opens an exciting landscape for Predictive and Prescriptive Maintenance.

The challenge now lies in the practical integration of these techniques into a complex and ever-evolving industrial environment. As we move towards real-time automation and adaptability, collaboration between domain experts and AI professionals will be crucial to successfully implement these transformations.

The dynamic adjustment of models for predictive maintenance is an ever-evolving journey, but its promise of safety, efficiency, and optimal performance in industrial systems deserves continuous exploration and investment in research and development.

This article provides a deep and practical insight into how the dynamic tuning of models is transforming the landscape of Predictive Maintenance in the industry, exploring advanced techniques and their practical applications in distributed industrial environments.

Source: Valverde, G., Quesada, D., Larra?aga, P., & Bielza, C. (2023). Causal reinforcement learning based on Bayesian networks applied to industrial settings. Engineering Applications of Artificial Intelligence, 125, 106657. https://doi.org/10.1016/j.engappai.2023.106657??