Real-Time Audio-Based AI Diagnostics

Solution Overview

As the renewable energy sector rapidly evolves with advancements in AI and machine learning, new opportunities for efficiency gains and cost reductions are emerging. Companies are increasingly turning to predictive maintenance to increase efficiency and reduce costs. However, predictive maintenance systems face the following challenges:

• Integrating data with diverse formats and sources.

• Scaling the high volume of IoT signals that are generated.

• Performing real-time analysis of vast datasets, which can be resource-intensive.

• Acquiring and effectively utilizing unstructured data, which hinders the development of robust predictive maintenance models.

This solution uses MongoDB Atlas Vector Search to explore the application of AI in real-time anomaly detection using sound input. This approach has several benefits:

• Document data model: MongoDB's BSON (binary JSON) format stores diverse data types, including unstructured data, which simplifies maintenance and enables faster responses to changes.

• Time-series collections: MongoDB handles time-series data, which is crucial for real-time monitoring in predictive maintenance and ensures timely interventions.

• Real-time data processing: MongoDB enables immediate diagnostics and responses, which is vital for proactive maintenance to prevent costly repairs.

• Data aggregation: MongoDB's powerful aggregation capabilities provide comprehensive insights into fleet-wide performance trends.

• Atlas Vector Search: MongoDB Atlas allows you to search unstructured data, with features like vector indexing and retrieval to enable robust predictive maintenance solutions. To create your first index, visit the Atlas Vector Search Quick Start guide.

Reference Architectures

This solution is divided into two parts:

1. Audio Preparation

First, the solution captures the audio from the equipment in different situations, such as normal operation, high or low load, obstructed equipment, or not operating.

Once each sound is collected, use an embedding model to process the audio data and convert it into vector embeddings. By generating embeddings for each audio track, the system captures the unique characteristics of each sound.

Then, upload the vectors to MongoDB Atlas. After only a few examples of sounds have been added to the database, they are ready to be searched and compared with the sound emitted by equipment during real-time operations.

2. Real-time Audio Diagnosis

Next, put your equipment in normal operation and start capturing the sound it makes in real time. The preceding video demonstration captures one-second long clips of audio. Then, it takes audio clips and converts them to vector embeddings with the same embedding model used previously. This process happens in milliseconds, which allows you to monitor the audio in real time. The vector embeddings are then sent to MongoDB Atlas Vector Search, which searches for the most similar sound out of the ones that were recorded in the previous step. Vector Search returns the result with a percentage of similarity. The system performs this step every second by leveraging fast vector embedding and quick searches. This allows for real-time audio-based monitoring.

Figure 1. Real-time wind turbine diagnosis by analyzing the emitted audio to determine whether it's operating normally, stopped, or experiencing any issues

Data Model Approach

The data model for vectorized audio collection is simple. The solution uses a collection called sounds to store documents that represent the prepared audio. These documents include an audio label and a URL for the GIF that gets displayed on the solution UI. Once the system vectorizes the reference audio for each status, it adds the embedding to the document.

During the real-time audio diagnosis stage, the one-second audio clips that are recorded in real time are vectorized and sent to MongoDB Atlas Vector Search, where they are compared against embeddings from the sounds collection.

Build the Solution

This solution uses the Wind Turbine Diagnostics GitHub repository. For more detailed instructions, see the repository's README.

{
   "mappings": {
      "dynamic": true,
      "fields": {
         "emb": {
            "dimensions": 2048,
            "similarity": "cosine",
            "type": "knnVector"
         }
      }
   }
}

Key Learnings

• Understand the role of AI and machine learning in revolutionizing predictive maintenance in the renewable energy industry.

• Explore how MongoDB Atlas Vector Search facilitates real-time anomaly detection and addresses challenges faced by businesses and development teams.

• Create a Vector Search index in Atlas or on a local deployment.

Authors

• Ainhoa Múgica, MongoDB

• Arnaldo Vera, MongoDB

• Dr. Humza Akhtar, MongoDB

• Dr. Han Heloir, MongoDB

• Ralph Johnson, MongoDB