Using the LLM to troubleshoot the Building Automation

What is sort of neat is ChatGPT has somewhat of a newer feature which people can upload pictures to have analyzed.

Why not see if ChatGPT can troubleshoot the building automation system (BAS) with snip-it's of the GUI for an HVAC control systems? For the untrained building operator this can be a very difficult problem in validating HVAC is operating correctly. The picture in the snip below a screenshot I found online where I made up my own sensor readings and valve positions by modifying the image in Microsoft Paint to mimic a broken heating coil on the air handling unit (AHU) where if this was in real life it would cause significant temperature complaints the building.

My prompt to ChatGPT (newest 4o model) to give the LLM some context: can you tell me what is wrong with this air handling unit and please layout some steps in troubleshooting it?

And ChatGPT’s response which I can vouch it's not bad, it recognized in the BAS screenshot the heating and cooling coils but cannot distinguish which temperature sensors are the AHU return, mixed, and leaving air readings...but it does know the temperature sensor setpoint and that there is a “Temperature Discrepancy” which can be half the battle IMHO is seeing something wrong. What is also neat is ChatGPT can tell the difference in the heating and cooling valves and understands that the readings are a percent valve command to the actuator motor on the coils. This is ChatGPT response back to me:

Based on the image of the air handling unit (AHU) with its current readings and component states, here are some observations and potential issues:

1.    Temperature Discrepancy: The temperature readings show significant variation:

·        Left side: 74.6°F

·        Right side: 78.6°F and 88.8°F

·        Supply Setpoint: 55.0°F

The supply setpoint temperature (55.0°F) is much lower than the actual temperatures, indicating the system is not cooling effectively.

2.    Cooling Valve Position: The cooling valve is 67.8% open, which suggests the system is trying to cool the air but is not reaching the setpoint.

3.    Heating Valve Position: The heating valve is 0% open, indicating no heating is being applied, which is correct since cooling is needed.

4.    Damper Position: The damper position is 32%, which might be insufficient for proper airflow. The damper position should be checked to ensure it is providing the right amount of outside air and return air for effective cooling.

Troubleshooting Steps

1. Check the Cooling Coil:
2. Inspect the Cooling Valve:
3. Assess Airflow:
4. Evaluate Temperature Sensors:
5. System Controls:
6. Check for Air Leaks:
7. Review Historical Data:

The experienced building operator or commissioning (Cx) agent would quickly recognize the elevated leaving air temperature and the closed heating valve, indicating a probable leak in the heating valve but I personally find this pretty cool that it can even do just that.

So, just for fun, let's try it again with the same AHU graphic snip in a new session to the LLM without any context or prompt to see what the language model generates without any background information...and the response back to me.

It looks like you've uploaded an image related to an HVAC system, specifically an Air Handling Unit (AHU) simulation. The image displays various temperature readings, damper positions, and valve positions.

- Supply Setpoint: 55.0°F

- Temperatures:

- 74.6°F (likely the return air temperature)

- 78.6°F (possibly the mixed air temperature)

- 88.8°F (likely the discharge air temperature)

- Damper Position: 32%

- Heating Valve: 0%

- Cooling Valve: 67.8%

The image seems to depict a scenario where the cooling valve is active, and the heating valve is closed. The supply air temperature setpoint is quite low, indicating that the system is trying to cool down the air significantly. If the goal is to troubleshoot or analyze the system based on this image, it appears that the cooling is engaged quite heavily (67.8%), while heating is not in use (0%). The damper position indicates a partial opening, likely allowing a mix of return and fresh air. Could you please provide more context or specify what you need assistance with regarding this image?

I think this is pretty cool.

In facilities management, maintenance staff are often too busy or find HVAC systems too complicated to troubleshoot effectively. For example, the building operator on shift might have a background as a master carpenter or plumber, and without HVAC specialists available, assistance from AI would be extremely helpful.

Let's try another one.

I found a random BAS graphic of a VAV box, which in real life would serve a single zone like an office or something for HVAC temperature regulation. I modified the graphic in Microsoft Paint again to artificially show zero CFM airflow. To the untrained eye, it may appear that the zone is maintaining its temperature setpoint. However, in an occupied building, there should always be airflow for ventilation, especially in large HVAC systems. A seasoned Commissioning (Cx) professional or an advanced building operator should quickly spot this anomaly in any BAS graphic. This would suggest a flow sensing issue, a potential mechanical problem with the air damper, or even scenarios where the sensing elements do not function well under low flow conditions.

And without any context just plopping that into the chat prompt, ChatGPT responded.

The image you uploaded shows a graphical interface of a Variable Air Volume (VAV) module within an HVAC system. It includes real-time data such as flow rate, setpoint, damper position, zone temperature, and setpoints for heating and cooling. Additionally, it shows occupancy status, weather information, and options for navigating between floor zones. Is there anything specific you would like to know or analyze regarding this VAV module interface?

Where then I responded to push it further with a "tell me what you see."

The image appears to be a graphical interface for monitoring and controlling a Variable Air Volume (VAV) box in an HVAC system. Here are the details visible in the interface:

1. Weather Information (Top Left Corner):

- Temperature: 30.6°F

- Relative Humidity: 46.7%

- Enthalpy: 12.8 BTU/lb

2. Navigation Links (Top Center):

- "1st Floor Zones"

- "2nd Floor Zones"

3. Location Information (Top Center):

- Floor: 1st Floor

- Room: Room 8 / RTU-14 / VMA-40

4. Main Display:

- Shows a VAV module with a label indicating the airflow direction.

- The module includes wiring and a control unit labeled "METASYS."

5. Current Readings (Near the Main Display):

- RTU-14 Temperature: 59.2°F

- VAV Module Flow: 0 CFM (Cubic Feet per Minute)

- Setpoint: 100 CFM

- Damper Position: 19%

6. Notes Section (Bottom Left):

- An empty box for adding notes.

7. Occupancy and Temperature Control (Bottom Center):

- Occupancy status: "OCCUPIED"

- Heating Setpoint: 68°F

- Cooling Setpoint: 71°F

- Zone Temperature: 71.2°F

- Setpoint Adjustment: ±2.0°F

- Effective Cooling Setpoint: 73.0°F

- Effective Heating Setpoint: 70.0°F

- Occupied until: 10:00 PM

8. Graphical Indicators (Bottom Center):

- A bar graph showing the heating and cooling range and the current temperature within the zone.

9. Temperature Control Device (Right Side):

- Displays current zone temperature, setpoint adjustments, and occupancy status.

The interface appears to be part of a building management system, providing real-time monitoring and control of HVAC components to maintain desired environmental conditions within a specific zone.

And pushing it further, "do you see anything incorrect or broken?" And the results are not that bad once it is prompted enough, it does see the air flow rate reading zero which is awesome!

Here are a few observations that might indicate potential issues or areas for further investigation:

1. Flow Rate:

- The VAV Module Flow is showing 0 CFM, while the setpoint is 100 CFM. This discrepancy could indicate a problem with the airflow, such as a blockage or a malfunctioning fan.

2. Temperature Discrepancies:

- The RTU-14 temperature is 59.2°F, which is significantly lower than the zone temperature of 71.2°F. Depending on the system design and the location of sensors, this might be normal, but it could also indicate an issue with temperature regulation or sensor accuracy.

3. Damper Position:

- The damper position is at 19%, which is relatively low. This might be appropriate if the system is trying to reduce airflow to meet the current conditions, but in combination with the 0 CFM flow, it could suggest a potential issue.

4. Occupancy Status:

- The system is showing "Occupied until 10:00 PM," but without additional context, it's hard to determine if this is correct. If the building is expected to be unoccupied, this setting might need adjustment.

5. Setpoints and Adjustments:

- The effective cooling and heating setpoints (73.0°F and 70.0°F) are close to the current zone temperature of 71.2°F, which is within the normal range, but it’s important to verify if these settings are appropriate for occupant comfort and energy efficiency.

Overall, the key concern is the airflow rate being zero despite the setpoint being 100 CFM. This is the most critical issue that likely requires immediate attention. The other points may need further investigation to ensure optimal performance of the HVAC system.

In conclusion

I think this is awesome and can greatly help building operators. I hope the facilities management world embraces this technology. Typically, building operators are pressed for time with work order completions and efficiency demands, so any gaps that can be filled to operate buildings more efficiently and effectively are welcome. It's important to note that sometimes the building operator is also the custodian, especially in school districts, where the person may not have an advanced background in mechanical or electrical engineering and HVAC sequencing. In these cases, AI could be a tremendous help in operating buildings better. With its extensive knowledge of indoor air quality (IAQ), an AI like this can support operators in ensuring the air people breathe inside buildings is within acceptable thresholds. As anyone in the HVAC controls or commissioning (Cx) industry knows, we've all encountered horror stories of buildings with terrible IAQ due to a lack of understanding by the building operator. This is another great use case for AI, ensuring healthier indoor environments.

I also find it fascinating that AI can often provide troubleshooting steps almost without prompting. For instance, during a particularly harsh Wisconsin winter a few years ago, I recall a situation where the school district superintendent had to step in because the maintenance staff couldn't resolve the issue. The superintendent, eager to manage the situation, urgently needed troubleshooting steps to prevent school from being canceled for the day. I actually posted about that experience here.

How does AI even understand troubleshooting HVAC?

I cannot help but wonder this question as there is not a lot of troubleshooting manuals that I can find online for HVAC controls. Actually, there is zero that I can find online and asking around some, and Tristan Hair , Drew DePriest, MCR.w, WELL AP , and others some having referred me to the Honeywell Gray Manual which is a classic in industry. It is a very in-depth manual and not that bad but it's all theory no real troubleshooting steps. For example, it will state the theory of a PID loop calculation and heating load calcs which is good for the control's technician or the mechanical engineering student but what about for the common building operator? For me I learned to troubleshoot HVAC controls sitting on constructions sites for 10 years setting up BAS' and working some with Adam Mangrich, PE, CCP in the field on existing building commission projects where you can apply theory to what you are looking at but what about the LLM? I know they learn by data which is a bit different than the human learning through experience just commissioning HVAC in the field and responding to service calls.

I couldn't help myself but feed in the troubleshooting steps back into ChatGPT and ask it "How do you know how to troubleshoot?"

The response you see regarding HVAC troubleshooting and verification processes is generated based on a combination of general knowledge about HVAC systems and best practices for building management, alongside an understanding of user needs in emergency and maintenance scenarios. As an AI language model, I don't learn from specific documents, sources, or databases but am instead trained on a mixture of licensed data, data created by human trainers, and publicly available data. This training involves large-scale datasets that help develop a broad understanding of various topics, including technical subjects like HVAC systems.

The guide provided is a generic procedure that reflects common practices in troubleshooting and verifying HVAC systems, especially within contexts that involve Building Automation Systems (BAS). It's designed to be useful for individuals like yourself—a school district superintendent without specialized knowledge of HVAC systems—by detailing a step-by-step approach to ensure that heating and cooling systems operate efficiently and that issues are promptly addressed. This guidance helps facilitate effective communication and coordination with technical staff to maintain a comfortable and safe environment for school operations.

My ability to provide troubleshooting steps or guidelines across various domains, like HVAC systems, is based on patterns and general knowledge extracted during my training. Here’s a bit more about how that works:

1. Training on Diverse Data: I am trained on a vast and diverse dataset that includes text from books, articles, websites, and other sources. This data covers a wide range of subjects, including technical topics like engineering, technology, and practical how-to guides.

2. Pattern Recognition: During training, I develop the ability to recognize patterns in data, including common troubleshooting steps, problem-solving strategies, and technical procedures. This capability enables me to generate responses based on typical practices and generalized knowledge.

3. Generalization: From the training data, I learn to generalize and apply the information to new but similar situations. For example, common steps in troubleshooting technical equipment—like verifying operation, checking for error codes, and testing system responses—are applicable across many types of machinery and systems.

4. Simulating Expertise: Although I'm not an expert or a human, I can simulate expertise by integrating and rephrasing the information embedded in my training data. This allows me to provide useful and contextually appropriate responses.

My responses aim to be helpful by suggesting actions that are typically effective based on general knowledge. However, for specialized or highly technical issues, consulting with a qualified professional who can assess specific situations in detail is always recommended.

Closing remarks

One thing that comes to mind is that, when I have the time, I love listening to the Lex Fridman podcast. Quite often, there are deep dives into technical topics, including AI and LLMs, which I find fascinating and valuable. I can't recall which interview it was, but on the podcast—possibly the one with Elon Musk or Sam Altman—they described using LLMs as an exploration process because the results can be unexpected and sometimes surprising. For me, it's mind-boggling how these models work. Despite the complexity, there's not much code involved (according to the podcast); it's just a very very large model or lots of numbers. I hope people find this interesting and I am optimistic that the industry will eventually adopt such technologies, potentially integrating them into the BAS of the future and/or IoT to help buildings operate more efficiently and effectively.