HAM-LET E-Catalog

Download PDF Catalog

on-Line College
Technical inquiry form
On-line COT Service

Integrating Artificial Intelligence into the Heart of Flow Systems

In the age of Industry 4.0, data has become increasingly valuable, and companies throughout industries are now developing innovative ways to capture data more efficiently and accurately.
In flow systems, flow data such as pressure, temperature, humidity, as well as valve data (i.e. the opening/closing of a valve) are crucial to ensure proper monitoring and maintenance.  
However, most of the methods used to capture these data require using ‘wet’ sensors (i.e. placing sensors in physical contact with the fluid). Although accurate, designing these sensors into the flow systems increase the number of possible leakage points. With the advancement in Machine Learning algorithm, an alternative way to obtain flow data is through audio analysis of the flow system. Audio sensors can be attached on the outer side of valves, fittings or pipes to obtain raw audio sounds, which can be later analyzed using Machine Learning Algorithms. 
Machine Learning in general requires substantial development time and expertise. Hence, the team at Ham-Let’s Innovation Lab decided to choose a specific use case to kick off this development. We also partnered with three Artificial Intelligence companies to develop prototypes of models which uses acoustic data to predict if there is a flow through a valve. This post describes our experience in developing the models with the AI companies and summarizes what to look out for when developing solutions that harness the power of Artificial Intelligence through Machine Learning.
1. Defining the Objectives & Scope
Like any project in general, the first steps of developing an AI solution is to define its objective. In this case, the objective is the information you want to obtain from the solution (e.g. presence of flow through a valve). For this post, let’s name this piece of information the “Objective”. 
Once that is defined, the next step is to list down the types of data that correlate with the Objective. These types of data are also known as “Features”. Start with conventional features, but be open to other types of data that may correlate with the Objective. The beauty of machine learning is that the machine may be able to detect relationships and patterns that humans may not be able to.
Figure 1: Setup for Audio Data Collection
2. Shortlisting the Features 
Once the possible features are listed, shortlist the features based on the constraints of data collection. In the case of Ham-Let’s project, the fluid’s flow rate and pressure were ruled out despite having an obvious relation with the presence of a flow in the valve. This was because we had set the constraints of data collection to be a ‘dry’ one, meaning that our sensors had to be attached to the outer surface of the valve, with no direct contact with the fluid. Eventually, we decided to use acoustic data since a microphone can be easily mounted on the outer surface of the valve, and we hypothesize that the presence of sound strongly correlates with the presence of flow in the valve.
3. Obtaining and Recording the Data
Once the feature(s) has been chosen, the next step is to design a method to collect and record the data. Important things to note when designing the method are:
i. How much data is required?
ii. What format must the data be stored in?
iii. Can the data be collected automatically? (This is important for Step 6)
iv. Will the training use supervised or unsupervised learning? Supervised learning requires collection of data of the independent variable (i.e. the Objective). 
The answers to the above questions should be answered by the AI Company/Vendor if the machine learning is outsourced. In Ham-Let’s case, it was important for us to know what format the audio data was in, and how many audio samples were needed for the training of the model. Moreover, of the three AI companies we partnered with, two used supervised learning, while the third used unsupervised learning. Hence, we had to collect a set of data that matched the audio samples with the actual open/close state of the valve. We also designed a system where the audio data can be logged automatically into our database. 
4. Developing the Model
There are many ways to develop the model, with each method (e.g. Linear Regression, Logistic Regression, Convoluted Neural Networks, etc.) having its advantages in different applications. In Ham-Let’s case, this was handled by the three AI companies, with very impressive results. To learn more about the different algorithms and an overview of Machine Learning, I would strongly recommend Andrew Ng’s “Machine Learning” course that can be found on Coursera. 
For those outsourcing the developing of the model, here are some points to take note:
i. Model Package: It is important to tie down how the AI Company/Vendor is exporting its model in. Some AI Companies require the model to be hosted on their own servers and accessed through APIs, while others package the model in a code base that can be embedded in your devices or systems themselves. Based on the application requirements, the exported model format varies accordingly. For example, audio data should be analyzed and processed locally as it is expensive to send audio data to the cloud. AI Companies specialized in audio analyses, like Otosense, have actually developed a optimized system where it is able to send only a small percentage of audio samples to the cloud for continuous training of the model, while the rest are analyzed locally. 
ii. Dashboard for Model Management: AI Companies/Vendors that provide a dashboard allow clients to upload and manage their data more easily. This is especially crucial when scaling to production, where more models (and even more data) are required for more applications. 
Figure 2: Otosense Dashboard
iii. Graphical Presentation of Results: Apart from the exported model, having a graphic presentation of the results and model is a bonus, allowing clients to understand the model more clearly as well as to easily validate the model.
Figure 3: Otosense Data Presentation with Tags
5. Validating the Model…with more Data
Once a model has been developed, it must be validated. AI Companies would have reserved part of the data given for validation. However, it is also important to test if the model works on new data sets. In our case, we collected more data from the same valve and ran them through the model to test if the predicted output of the model matches with the true value. Further validation tests would include testing on different valves of the same type. Again, this requires more data collection.
Here is an example of the validation we did of the models developed by Otosense and Realilty.AI
Figure 4: Validating model developed by Otosense
In this case, their models worked to a large extend. However, in cases where the model is invalidated, rethink the algorithm used for training, as well as the features used for training. If partnering with a AI company, work with them to think of how the model can be improved.
6. Designing a System to Continuously Train the Model
No developed model is ever perfect, and continuous training of the model is crucial. This can be done manually or automatically, depending on the system designed. As mentioned earlier, Otosense has developed an optimized system to do this automatically. 
7. Scaling to Production
Once you are satisfied with the model, it is time to scale it to production. This requires further design thinking on how the model can be calibrated to different environments and applications. 
In summary, developing a solution that harnesses the technology of Machine Learning is an intensive process. Unless you have a specialist in-house, it is recommended that you partner with companies that specializes in Machine Learning, especially when developing a solution for the industries, where reliability and robustness is key. On top of that, Machine Learning also enables Predictive Maintenance, a concept that has been gaining traction lately. Our team at Ham-Let’s innovation lab is looking to further develop our solution with real use cases. We are also constantly looking to partner with more companies to harness the power of Artificial Intelligence to develop smarter valves for smarter flow systems all around the world.
If you are interested to partner with us to develop a solution, whether as an AI company or a Flow Systems company, drop me a PM. If you have anything to add, comment below!


About us:

Ham-Let Group was established in 1950.

Traded on the Tel Aviv stock market since 1994.

A worldwide company, Ham-Let Group has 14 branches and 4 subsidiaries around the globe and is constantly growing.

We provide a wide variety of innovative valves, fittings, and hoses for many industries including: Oil & Gas, Semiconfductor, Power Generation.