Enhancing Operational Insights for Industrial Assets with AWS IoT AI/ML Solutions (Part 2)

More Info

In this continuation of our two-part blog series, we delve into an AWS IoT AI/ML solution designed to assist industrial clients in effectively monitoring their assets at scale. The first installment covered:

• Creating an asset simulator using AWS IoT SiteWise.
• Establishing a data pipeline to connect Amazon Lookout for Equipment with AWS IoT SiteWise.

In this post, we will further develop the solution initiated in Part 1. Ensure that you have configured AWS IoT SiteWise and SiteWise Monitor with the relevant industrial assets and established the data pipeline for sending information to Amazon Lookout for Equipment. If you haven’t completed these preliminary steps, we recommend reviewing Part 1, Steps 1 and 2 before continuing.

Steps 3 and 4 will guide you through the following processes:

• Training the Amazon Lookout for Equipment model with historical data and evaluating its performance.
• Utilizing Amazon Lookout for Equipment to configure an inference scheduler for asset anomaly predictions.
• Enhancing the dashboard created in Part 1 by incorporating Amazon Lookout for Equipment for anomaly alerts and remote monitoring.

Step 3: Train the Lookout for Equipment Model

Before building our model, let’s revisit what Amazon Lookout for Equipment entails and its operational framework. Amazon Lookout for Equipment leverages machine learning to identify abnormal behavior in your equipment and predict potential failures. Every piece of industrial equipment is deemed an industrial asset, or simply an asset. To monitor your asset using Lookout for Equipment, follow these steps:

1. Supply Lookout for Equipment with data from your asset. This data is gathered from sensors measuring various characteristics of your asset, such as temperature and pressure.
2. Initiate a training job in Amazon Lookout for Equipment to develop a custom ML model.
3. Configure an inference scheduler to monitor your asset for anomalies nearly continuously.

Although asset failures are infrequent, and even similar types of failures can exhibit unique data patterns, all detectable failures are preceded by behaviors or conditions that deviate from the norm. Lookout for Equipment is engineered to detect these patterns by training a model using sensor data to establish a baseline of normal asset behavior. In essence, it learns what constitutes standard functioning and identifies deviations as it monitors your equipment. To accentuate unusual behavior, Lookout for Equipment employs labeled data in the model training process. Labeled data consists of historical date ranges indicative of abnormal asset behavior. While providing this labeled data is optional, having it on hand can enhance the model’s training accuracy and efficiency.

The screenshot from the Amazon Lookout for Equipment service illustrates an example of labeled data highlighting periods of abnormal asset behavior.

After training your model, you can visualize its evaluation results on the Amazon Lookout for Equipment console. You’ll also be able to select each event, allowing Lookout for Equipment to unpack the sensor rankings and display the top sensors contributing to detected anomalies. This anomaly score ranking is beneficial for the operational technology (OT) team, enabling more efficient component checks or repairs by prioritizing sensors with elevated anomaly scores.

Once your model is ready, you can monitor your asset by scheduling the frequency at which Lookout for Equipment processes new sensor data through batch inference every 5 minutes. The screenshot of the Lookout for Equipment inference scheduler illustrates the inference history at this frequency.

Now that we’ve established a solid understanding of Amazon Lookout for Equipment and its functionality, let’s proceed with model development.

In Part 1, Step 1, we set up an AWS IoT SiteWise simulator using a CloudFormation template, and the UUIDs of two pump assets were provided as outputs. Navigate to the Outputs section and copy the AssetID values.

Next, go to the SageMaker console and find the notebook instance created by the template. Select Open JupyterLab.

Within JupyterLab, navigate to the l4e_notebooks folder, and (1) insert the first pump asset (FirstAssetId) in the AssetID within config.py; (2) add the BUCKET name (as shown in the previous screenshot) for the Amazon S3 bucket created in Step 2 for pump asset 1.

It’s worth noting that Amazon Lookout for Equipment will develop a unique model for each industrial asset, offering tailored insights based on the asset’s operational environment. To train a model for asset 2, update config.py with the new S3 path and UUID for that asset, then rerun all notebooks. You may also choose to train just one model at this stage. However, we will later discuss maximizing value from monitoring multiple similar assets.

Run each notebook in the l4e_notebooks subdirectory sequentially. While these notebooks contain detailed explanations for every step, here is a high-level overview of their functions:

1. In 1_data_preparation.ipynb, the notebook will execute the following tasks: (1) Download the sample dataset from the original S3 bucket; (2) Uncompress the contents into a local directory; (3) Load the data into the training bucket for Lookout for Equipment.
2. After completing the steps in 1_data_preparation.ipynb, proceed to 2_dataset_creation.ipynb, where we will create a data schema for our data and upload it into Lookout for Equipment using the CreateDataset and StartDataIngestionJob APIs.
3. In 3_model_training.ipynb, this notebook will train an ML model in Lookout for Equipment. It defines the training and evaluation date ranges, passes the S3 path to labels.csv containing known historical anomalies, and finally initiates a training job using the CreateModel API.
4. In 4_model_evaluation.ipynb, you can assess the trained model by extracting associated metrics with the DescribeModel API. This step is optional and does not commit any changes; it merely allows for manual analysis of training results.
5. Lastly, in 5_inference_scheduling.ipynb, the notebook deploys the model into production by invoking the CreateInferenceScheduler API.

Step 4: Create an AWS IoT SiteWise Monitor Dashboard

Once the Lookout for Equipment inference schedule is established, the data pipeline that connects to the dashboard will be enhanced with anomaly detection insights, facilitating more informed decision-making. For more information, check out this blog post that dives deeper into the topic. For expert insights, visit this authority on the subject. Additionally, if you’re curious about the experience of being an Amazon warehouse worker, this resource provides valuable insights.