Insights from our experts: How CIM’s use of AI transforms building operations

Following our recent post on the advanced AI capabilities within CIM’s PEAK Platform, we’re diving deeper into the specific AI-driven technologies that make PEAK a game-changer in building operations. This follow-up features insights from our Senior Data Scientist, David Waterworth, and Head of Engineering, Antonious Mickael, addressing key questions about how our AI delivers unparalleled value.

By blending CIM’s expertise in software, UX, and engineering with cutting-edge AI applications, PEAK streamlines building onboarding, enhances operational team productivity, and delivers measurable value quickly. Leveraging multiple AI models tailored for distinct use cases, PEAK ensures operational efficiency, energy optimization, and reliability. Let’s break down how it all comes together.

AI models tailored for building operations

Does CIM offer a single AI model or multiple models in your platform?

David Waterworth: We use multiple models, each targeting specific use cases. For example, large language models (LLMs) accelerate onboarding and validate data mappings, while expert systems handle FDD. This dual approach ensures both speed and accuracy across different aspects of building management.

What decisions are automated using these AI models?

David Waterworth: Our models operate as "recommenders." LLMs assist engineers by suggesting data stream categories, equipment associations, and identifying inconsistencies. Engineers decide whether to accept these recommendations, ensuring human oversight. Meanwhile, expert systems deploy rule-based FDD, monitoring equipment status, schedules, environmental conditions, and energy consumption to provide actionable insights. Automation helps identify and apply the optimal set of rules for each site, streamlining the process.

How do you train these AI models, and what data sources are used?

David Waterworth: Our LLMs are trained using a large dataset of BACNet and Haystack metadata from building management systems (BMS). Each building deployment contributes new data, creating a continuous learning loop. This human-in-the-loop approach refines the models over time, leveraging the expertise of engineers during the portfolio onboarding process.

What data is essential for PEAK to function effectively?

David Waterworth: We require equipment-level BMS time-series data at 15-minute granularity. This ensures our models have the granularity and frequency needed to deliver accurate fault detection and actionable insights.

Describe whether you offer a pre-trained model and what additional training is necessary using a site-specific dataset and human-based expertise.

David Waterworth: Our models are not site-specific. However, the deployment of expert-system rules is customized to each site. Our large language models (LLMs) are pre-trained and specifically adapted to the unique language and structure of BMS text descriptions. They are further fine-tuned for specific tasks through an active learning process, ensuring accuracy and relevance.

How do you ensure the AI applications operates as intended once deployed?

David Waterworth: Our engineering team monitors decisions made by the FDD model and tunes rules to balance false positives and negatives. Continuous oversight ensures the AI maintains high accuracy and delivers reliable recommendations.

Streamlined onboarding with AI

A key differentiator of the PEAK Platform is its proprietary onboarding process, which integrates AI to reduce setup time, improve data accuracy, and provide insights faster than traditional methods.

Steps in the process:

1. Equipment register creation: Establishing an equipment inventory within PEAK.
2. Parent-child relationship mapping: Configuring associations between systems like air handling units (AHUs), variable air volume (VAV) units, and central plant equipment.
3. Metadata prediction: Using machine learning to label data points based on our extensive dataset, expediting commissioning.
4. Anomaly detection: Identifying and correcting inconsistencies using text-to-vector algorithms and semantic similarity.
5. Automated rules deployment: Leveraging scripts to apply fault detection rules quickly and accurately, cutting deployment time by 90%.

This approach allows insights to typically be available within 30 days of deployment, delivering immediate value.

Real-world impact of AI

Building on the foundational insights shared in our earlier blog, here’s how our AI capabilities address client challenges:‍

• ‍Accessibility for non-technical team members - CIM’s application of AI and ML transforms complex fault detection and diagnosis into a more accessible and intuitive process. This makes sophisticated technical information accessible to non-technical team members, effectively upskilling them and reducing the reliance on hiring of highly technical staff.
• ‍Faster ROI - Using AI greatly accelerates the commissioning process of sites. The rapid commissioning ensures that the value from the platform is realised quickly, enabling clients to benefit from PEAK’s intelligent monitoring capabilities without the typical delays associated with traditional commissioning processes. By reducing the time to operational efficiency, these technologies directly contribute to faster ROI and immediate improvements in building performance.‍
• Improved accuracy and coverage - The integration of AI into CIM's processes significantly enhances accuracy and expands the coverage of automated FDD. This means that more faults across a broader range of operational plant and equipment are identified with precision. By capturing a wider spectrum of potential issues, these technologies ensure a more thorough and reliable monitoring of operations, improving efficiency and reducing equipment downtime.