DX-AI CMMS Roadmap
This presentation introduces a future-oriented equipment diagnostics system based on on-device AI and Large Action Models. We will explore the enhancement strategy and implementation plan for an iDSB-based CMMS (Computerized Maintenance Management System).
DX-AI iDSB Advanced CMMS Strategy
We introduce a future-oriented equipment diagnostic system based on on-device AI and Large Action Model. This system combines real-time response on-site and intelligent analysis to maximize the efficiency of equipment maintenance.
On-device AI
Lightweight AI models that can make judgments immediately on-site, enabling real-time response without network latency
Large Action Model
An advanced AI system that goes beyond simple text generation to predict and execute actual actions
Future-oriented Design
Scalable architecture that supports continuous functional improvements and system enhancements
Overview and Objectives of CMMS
The iDSB CMMS is a system that detects abnormalities in field equipment, manages failure records, and systematically manages response actions. The core objective is to improve the efficiency of equipment management by supporting maintenance-centered on the operator.

3

Operator-Centric Maintenance
Improve the work efficiency of field operators
Systematic Record Management
Databasing of failure and response action records
3
Abnormality Detection System
Early detection and response to equipment abnormalities
Current Configuration Summary (UI-based)
The current CMMS system has a user-centric interface, with the main pages consisting of work log entry, image gallery and detailed information, similar case search, and equipment analysis report. Each page is designed with an intuitive UI, allowing workers to easily access and utilize them.
Work Log Entry
This page is for recording equipment malfunctions and corrective actions that occur in the field.
Image Gallery and Detailed Information
This image-centric page allows visual confirmation of equipment status.
3
Similar Case Search
This page allows searching for and referencing past similar failure cases.
Equipment Analysis Report
This page provides a report analyzing equipment status and failure patterns.
Summary of Current Features and Advantages
The current CMMS system is improving the efficiency of equipment management by visualizing work information and systematizing records. It is possible to respond proactively through failure history analysis, and the intuitive UI has strengthened the accessibility of workers.
Visualization and Systematization of Work Information
By visually representing the equipment status and work history, it helps intuitive understanding, and through systematic record management, it supports data-driven decision making.
Proactive Response Possible through Failure History Analysis
By analyzing past failure history, you can discover similar patterns and take preventive measures before problems occur.
Strengthened Accessibility for Workers through Intuitive UI
The user-friendly interface design allows field workers to use the system effectively without complex training.
Overview of Expansion Direction - Why AI?
Current CMMS systems are strong in data collection and management, but have limitations in real-time response and intelligent analysis. Real-time field operations and intelligent analysis are needed, which can be solved through on-device AI and Large Action Model (LAM).
Enhance Real-time Capabilities
Immediate judgment and response is needed in the field, but currently there are delays
Need for Intelligent Capabilities
Data is collected sufficiently, but judgment and execution still depend on manual processes
Adopt On-device AI
Execute AI models directly on field devices to enable immediate judgment without network delays
Integrate LAM
Build an intelligent system that goes beyond simple analysis and connects to actual actions
What is On-Device AI?
On-device AI is a technology where the AI model runs directly on the user's device, without sending data to the cloud (server). This allows for immediate response without network latency, and also benefits personal data protection.
Cloud AI
  • Processed on servers
  • Requires fast computation
  • Data transfer needed
On-Device AI
  • Processed on the device itself
  • Works without a network
  • Protects personal data locally
Why is on-device AI important in the field?
In industrial settings, on-device AI has the advantages of being able to make judgments immediately without network latency, operating in offline environments, enhancing the security of sensitive data, and being highly energy-efficient for continuous operation.
Immediate response
It is possible to respond quickly by making judgments directly on-site without network delays.
Operates offline
It can be used continuously even if the network connection in the factory or equipment is unstable or interrupted.
Excellent security
Data security is enhanced as sensitive images or sensor data are not transmitted to external servers.
Power efficiency
It is possible to operate continuously without battery concerns by using lightweight models.
How will it be technically implemented?
On-device AI is implemented on mobile devices or edge devices through lightweight models and optimization technologies. Lightweight models such as Gemma 1B, TinyLLaMA, and DistilBERT, as well as optimization technologies such as quantization and pruning, are used, and they are implemented through various execution environments and frameworks.
What can on-device AI do in CMMS?
On-device AI can perform various functions in the CMMS system. It can immediately detect anomalies in photos taken on-site, automatically summarize the failure symptoms entered by the operator, provide advance notice of frequently occurring warning messages, and provide real-time support on-site through an AI chatbot.
Immediately detect anomalies when taking photos
As soon as equipment is photographed on-site, AI will judge whether there are any anomalies and immediately provide results such as "Anomaly detected".
AI automatically summarizes failure symptoms
When the operator enters a long explanation, AI will summarize and organize the key content to support efficient record management.
Provide advance notice of frequently occurring warning messages
By analyzing patterns such as "345 degrees often occurs → Inspection required", it will provide preventive guidance.
AI chatbot responds on-site
Real-time FAQ support such as "Do this when this happens" helps the operator solve problems.
Connecting LLMs and On-Device AI
The lightweight version of the LLM is executed on the device, and when necessary, it is linked to the server's LAM to delegate complex decisions. For example, when a worker uploads a photo, the on-device AI assesses the possibility of a defect, and the server's LAM analyzes the type of defect to propose a maintenance schedule.
Lightweight On-Device LLM
Execution of lightweight models such as Gemma-1B, Phi-2, and TinyLLaMA
Field Data Collection
Acquisition of field information such as photos, text, and sensor data
Immediate Judgment
Simple analysis and judgment are processed immediately on-site
Server LAM Integration
Complex decisions are delegated to the server's LAM for processing
Summary: On-device AI is the "Smart Assistant at the Equipment"
On-device AI acts as a smart assistant that supports workers next to the equipment. It is a small and intelligent AI model that can make quick decisions, respond immediately on-site, operate even in network-less environments, and protect privacy.

Quick Decision Making
Provides immediate analysis without network delays
On-site Response
Provides real-time support and guidance next to the equipment
Works Without Network
Maintains continuous functionality even in offline environments
Privacy Protection
Sensitive data does not leak outside
Definition of LAM - Differences from Simple LLM
Unlike a simple Large Language Model (LLM) that only generates text, a Large Action Model (LAM) is a model that predicts and executes actual actions. LAM can process various inputs such as text, images, sensor data, and structured databases, and output actionable instructions or system commands.
LAM's Structural Components
LAM is composed of a complex structure that processes various input data, recognizes situations through multimodal processing and decision-making structures, and generates appropriate action outputs. This allows it to perform concrete actions such as automatically generating work orders, sending messages, and calling system control APIs.
Input Data
  • Text (work logs, natural language questions)
  • Images (equipment photos, sensor images)
  • Numerical data (temperature, pressure, IoT)
  • Ontology/history data (maintenance records, etc.)
Multimodal Processing & Decision-making Structure
  • Situation recognition
  • Scenario branching (condition/state-based)
  • Action selection (task issuance, alerts, reports)
Execution Output
  • Automatic work order generation
  • Message sending (to workers, managers, etc.)
  • System control API calls (valves, sensors, etc.)
Key Technical Elements of LAM
LAM consists of technical elements such as multimodal input understanding, reinforcement learning or rule-based planning, ontology-based reasoning, and real-time feedback loops. These technologies are combined to implement an intelligent system that can understand complex situations and make appropriate decisions.
Multimodal Input Understanding (Multimodal Fusion)
This technology comprehensively analyzes text, images, and sensor data to accurately understand the situation. By integrating various data formats, it can make more accurate judgments.
Reinforcement Learning (RL) or Rule-based Planning
This technology learns and optimizes the results of actions, allowing for better decision-making based on past experience. It is combined with rule-based systems to increase stability.
Ontology-based Reasoning
By understanding causal relationships through knowledge-based connections, such as "this failure is caused by this equipment-part-condition relationship", it can identify problem-solving methods.
Real-time Feedback Loop
This system reflects the effectiveness of the actions taken and influences the next decision, enabling continuous learning and improvement.
CMMS Application Scenarios
LAM can be applied to CMMS in various ways. It is utilized in various scenarios, such as inferring the cause of equipment failure and automatically issuing work orders when a failure is detected, proactive response to recurring failures, detection of foreign objects and tracing the cause, and optimization of maintenance personnel deployment.
Equipment Failure Detection and Response
When the heater temperature exceeds 345 degrees, LAM infers the cause, maintains the alarm, and automatically issues a work order. This enables a rapid response.
Recurring Failure Prevention
If a failure occurs repeatedly at the same location, LAM will make a "proactive replacement proposal" and automatically schedule it after approval. This enables preventive maintenance.
Foreign Object Detection and Cause Tracing
When a foreign object is detected in the image, LAM will trace the cause of the foreign object to the location and request a response from the worker. This allows identifying and resolving the root cause of the problem.
Maintenance Optimization
If there is a large deviation in maintenance for different equipment, LAM will propose an optimization of personnel deployment and change the priority. This allows efficient utilization of resources.
Reasons for Applying LAM to CMMS
By applying LAM to CMMS, you can incorporate maintenance experience into the decision-making structure of AI, have the system act as the brain of an automated system that leads to actual work, and build an AI system focused on execution rather than just simple FAQs or document-based LLMs.
1
Utilizing Maintenance Experience
You can incorporate the experience and know-how of skilled maintenance personnel into the decision-making structure of AI, preserving and expanding knowledge.
2
Automated System
The system can act as the brain of an automated system that leads to actual work, maximizing efficiency.
3
Execution-Focused AI
You can build an AI system focused on actual actions and execution, moving beyond simple FAQs or document-based LLMs.
Example of Collaboration between LLM and LAM
LLM and LAM play complementary roles and collaborate. When a worker asks about equipment failure, LLM explains the cause, and LAM executes the necessary actions. This collaborative structure organically connects explanation and execution.
Worker's Question
"This equipment is broken, why is that?"
LLM Response
Cause Explanation: "There is a history of the heater temperature rising rapidly. Suspected power overload"
LAM Execution
Action Taken: "Automatically registered a work order to inspect the heater + sent a message + assigned a work number"
Problem Solved
The worker understands the cause and the necessary actions are automatically taken
Detailed Functional Expansion Strategy - Work Log Registration
The current work log registration page operates by manually recording photos and text. In the future, it will evolve to use on-device AI to automatically recognize photos and infer the cause, LLM to summarize and classify the text, and LAM to assess past similar failures and recommend work instructions.
Current Features
  • Manual photo upload
  • Text-based failure recording
  • Direct input by workers
  • Basic save and view functionality
Future Expanded Features
  • On-device AI photo recognition and automatic cause inference
  • LLM-based text summarization and classification
  • LAM-based similar failure assessment
  • Automatic work instruction recommendation and generation
Expansion Strategy (Detailed) - Image Gallery
The current image gallery page is primarily focused on visualization. In the future, it will evolve to incorporate an embedded foreign object detection model for edge devices, and LAM will analyze the similarity to previous cases to provide automated warning functionality.
The image gallery is evolving from a simple photo repository to an intelligent analysis tool. On-device AI will analyze images in real-time to detect foreign objects, and LAM will compare to past cases to discover similar patterns, automatically generating warnings. This will enable early problem detection and rapid response.
Expansion Strategy (Detailed) - Similar Case Search
The current similar case search page is operated using a manual search method based on similarity. In the future, LAM will automatically recognize the situation and provide the optimal response guide, and offer interactions such as "Would you like to perform this action as the next step?"
Automatic Similar Case Analysis
LAM will automatically analyze past cases similar to the current situation and prioritize presenting the most relevant cases.
Optimal Response Guide
Based on the successful resolution methods of past cases, it will propose a step-by-step response guide optimized for the current situation.
Interactive Work Support
Through interactions such as "Would you like to perform this action as the next step?", it will support the decision-making of the worker and guide the workflow.
Detailed Expansion Strategy - Statistical Analysis
The current statistical analysis page focuses primarily on visualizing simple failure frequency. In the future, it will evolve to incorporate an LLM-based automatic tagging system that automatically classifies the cause, impact, and responsible equipment, and LAM will generate automatic maintenance schedules that include workers, time, and materials.
The LLM-based automatic tagging system automatically classifies failure data by cause, impact, and responsible equipment, enabling more accurate analysis. LAM then uses these analysis results to automatically generate optimal maintenance schedules that consider workers, time, and required materials, maximizing maintenance efficiency.
Integrated Architecture & Collaboration Structure
The integrated architecture of the CMMS consists of a worker, edge devices with on-device AI, a server-based LAM, and an automated Work Order management system. The system is designed to utilize the Gemma 1B/4B model on the edge devices and collaborate with the 12B model on the server when necessary.
Worker
Inspects equipment status and reports issues in the field
Edge Device + On-device AI
Immediate analysis using the Gemma 1B/4B model
Server-based LAM
Advanced analysis and decision-making using the 12B model
Automated Work Order
Automatic generation and assignment of necessary work orders
Summary of Expected Effects
By enhancing the CMMS with AI, we can expect to see a reduction in equipment downtime, a decrease in maintenance costs, a decrease in worker burden, and the establishment of a foundation for future responsiveness through maintenance automation. This will significantly contribute to improving the company's operational efficiency and competitiveness.
32%
Reduced Downtime
Significant reduction in equipment downtime through AI-based predictive maintenance
45%
Reduced Maintenance Costs
Cost savings through efficient resource utilization and preventive maintenance
60%
Reduced Worker Burden
Alleviation of worker workload through automation of repetitive tasks
85%
Improved Future Responsiveness
Flexible adaptation to future changes through an automated maintenance system
Conclusion and Recommendations
Based on the PoC (Proof of Concept), we propose a realistic step-by-step approach to enhance the CMMS. The AI capabilities can be modularized and introduced gradually, and lightweight models based on Streamlit and Ollama can be used to enable immediate demonstrations.
Phase 1: Implement On-Device AI
Deploy lightweight AI models on edge devices to implement basic image recognition and text analysis functionalities. Leverage Streamlit and Ollama to develop and demonstrate prototypes quickly.
Phase 2: Integrate LLM
Introduce server-based LLM to add advanced text analysis and summarization capabilities. Establish a collaborative structure between on-device AI and LLM to enhance the analysis capabilities.
Phase 3: Implement LAM
Build an LAM system with action execution capabilities. Implement features such as automatic task generation, notification sending, and system control to achieve full automation.
Phase 4: Continuous Improvement
Improve model performance and add new features based on actual usage data. Optimize the system by incorporating user feedback to drive continuous development.
Large Action Model
The Large Action Model (LAM) is an AI model that can perform actual actions, expanding on existing Large Language Models (LLMs).
Execution Capability
LAM can go beyond simple text generation to control systems and perform physical tasks.
System Integration
It integrates with existing systems like CMMS to support automated decision-making and task execution.
On-Device Implementation
It has a lightweight structure to operate on edge devices for immediate response in the field.
User Interaction
It understands worker commands and automates complex tasks to maximize field work efficiency.
1. Concept and Role of LAM
  • Definition
    LAM is a model that learns "when and how to perform a certain action", and adds the following four functions on top of the language understanding and generation capabilities of LLM.
  1. Action Design: Extracting specific action steps (e.g., "Create Work Order", "Execute DB Query") from text instructions
  1. Action Execution: Calling external system interfaces such as equipment control, API calls, and DB updates
  1. State Management: Recording and retrieving field conditions such as session, event, and IoT data in memory (DB/in-memory cache)
  1. Feedback Loop: Receiving the execution results and generating re-planning and supplementary actions
    DX-AI-CMMS
  • LLM vs. LAM
  • LLM (Large Language Model): "When the user describes a problem, generate a text-based answer"
  • LAM (Large Action Model): "Based on the answer from LLM, plan and execute actions that will actually operate in the system" DX-AI-CMMS
2. 5-Step Workflow for LAM Implementation
3. Example Technology Stack
4. Considerations for the on-device environment
  • Model Optimization: Reduce the size to 200–400 MB by applying INT8 quantization and pruning
  • Inference Runtime: Optimize using ONNX Runtime / PyTorch Mobile
  • Memory Budget: Allocate 2–4 GB for LAM state buffers and embedding caches
  • Latency: Maintain the total latency for each action (e.g., DB query 20 ms, API call 50 ms) within 200 ms
5. Expected Effects
  • Complete Local (on-premise) CMMS: Automates on-site control and decision-making even in unstable network environments
  • Real-time Response: Immediately designs and executes actions upon detecting abnormal events to minimize equipment downtime
  • Operational Efficiency: LAM automates repetitive tasks and significantly reduces the burden on operators