Beware of AI-driven maintenance plans filling your CMMS with non-value work

A recent LinkedIn post suggested using artificial intelligence (AI) to develop preventive maintenance and other checklists as a starting point. Being an early adopter of generative AI technologies such as ChatGPT, I know AI can be powerful in the proper context.

But here is the rift. Generative AI suffers from vagueness and a lack of repeatability. Getting a desired result often requires multiple iterations and applying technical knowledge to improve the AI output into the desired checklist or procedure. 

And therein is the issue. People seek a quick, simple solution without understanding the basic concepts, like defining tasks to address the likely failure modes based on the asset’s operating context. They tend to take the AI-generated output as the end game. Remember that AI output is derived from millions of internet samples used to teach it. Some tasks might apply; some will not; and some, while valid, may be too frequent based on the failure mode.

As a test, I entered the parameters defined in the LinkedIn post of asset type (more specific than the post suggested), the number of tasks, desired frequency, and type of checklist into ChatGPT. The prompt was, “You are a maintenance manager who needs to develop a preventive maintenance checklist. The checklist is for a Goulds Self-Priming Centrifugal Pump. The approximate number of tasks should be fifteen, and the checklist frequency should be every month. Develop the checklist.”

Below is a sampling of the output from ChatGPT v4.0:

Monthly centrifugal pump PM checklist 

1. Inspection of pump casing for leaks or damage - Check the pump casing for any signs of leaks, cracks, or wear. Look for any unusual signs of distress or damage that could impact the pump's performance.

2. Check for seal and gasket integrity - Inspect the seals and gaskets for any signs of wear or leaks. Replace if necessary to ensure no fluid leaks or air ingress could affect the pump's priming capabilities.

3. Lubrication of bearings - Ensure the pump's bearings are adequately lubricated with the appropriate lubricant. Check the manufacturer's recommendations for the specific type and amount of lubricant.

4. Inspect the impeller for wear or damage - Check the impeller for any wear, damage, or clogging signs. Clean if necessary and replace any parts that are worn out or damaged.

5. Verify shaft alignment - Check the alignment of the pump shaft to ensure it is straight and properly aligned with the motor to prevent undue stress or vibration.

6. Check motor electrical connections and wiring - Inspect the motor's electrical connections and wiring for any signs of wear, corrosion, or damage. Ensure all connections are secure.

7. Check for proper rotation and motor performance - Verify that the motor and pump are operating in the correct rotation direction and that the motor performance is stable and within normal parameters.

8. Inspect pump mounting and foundation - Check the pump's mounting and foundation for any signs of loosening or wear. Tighten any loose bolts and make repairs as necessary.

9. Verify suction and discharge hoses/pipes for integrity - Inspect suction and discharge hoses or pipes for any signs of wear, leaks, or damage. Replace or repair as necessary to ensure proper flow.

There are several issues with this AI-generated checklist. First and most important, assuming the tasks are the right ones to begin with, what is missing is the precision or feasibility – i.e., “regrease bearings every 2,000 hours or at a three-month interval”. This lack of precision presents a challenge.

Also, repeats of the prompt generated variations in the output checklist. For example, some outputs suggested that vibration analysis be used on the pump on a monthly frequency. Using RCM2 terminology, while technically feasible, vibration may not be cost-effective unless the pump is critical to the process. I encourage you to run your own prompts to see the results. Likely, the output will vary considerably from each run of the same prompt.

Finally, I picked fifteen tasks arbitrarily (items 10-15 were removed due to column space constraints). If I had selected thirty, then the AI engine would have provided them, which is another issue. We must determine the likely failure modes and then define precision tasks that address them based on the pump’s operating context. Otherwise, plants will continue to suffer where 40-60% of PMs are not value-added, as they don’t address the likely failure modes and they also consume precious technician resources.