Neuron Soundware's AI Progression

→ LOGIN

Exploring Neuron Soundware’s AI Progression: Highlights from the Past Year

As artificial intelligence continues to advance rapidly, Neuron Soundware has achieved significant milestones in algorithm development over the last year. We have focused on four key areas that greatly enhance the value of industrial data processing for our customers.

One-click training: Streamlining device monitoring at scale

Our first major innovation, “one-click training,” revolutionizes the process of training models (computational algorithms) for machines. This feature enables AI models to be trained efficiently, within 24 hours, and across a diverse range of machines. It operates automatically, removing the need for intervention by Machine Learning (ML) specialists.

Previously, training models and algorithms required manual intervention by ML specialists to adapt to specific machine behaviors, in order to ensure accurate results across different machines and to identify relevant changes in machine sound.

To illustrate, consider the scenario of hundreds of pumps within cement plants’ piping systems. Pump failures are common due to high loads, and each pump is configured slightly differently. This necessitates individual calibration of monitoring technology, despite the identical installation of IoT units and sensors.

Until now, calibrating this process for hundreds of devices would take weeks or months, causing delays in initiating the service for customers. With the automation of calibration, this process now takes only a few hours and is fully automated.

Moreover, the expertise needed to set up the monitoring process has been significantly reduced. While previously, a ML engineer was essential, now a project manager on the supplier or customer side is required to provide nominal data and train the model against it.

Data search: Enhancing Anomaly Detection with "Similarity Search"

We have Introduced our latest innovation, the “Similarity Search” module, which aims to streamline the identification of anomalous states by leveraging historical data. Previously, we relied heavily on customer feedback to refine and train our computational algorithms.

With the “Similarity Search” module, we have significantly reduced the need for extensive customer input when labeling specific processes or machine states. When an anomaly is detected, the algorithm autonomously searches for similar data within the machine’s behavior history. Users can then label this data, creating datasets used to train specialized models for fault detection. This process allows us to provide more precise insights to our customers, enhancing production and maintenance management.

This capability is particularly critical for timely maintenance tasks such as gearbox relubrication in transport or energy equipment. Maintenance operators no longer require disassembly to confirm the type of anomaly. With historical data guiding the identification process, operators can promptly address the issue armed with the necessary tools and materials, thus optimizing maintenance procedures.

Custom Classifier: Extending the Capabilities of Similarity Search

In addition to our latest innovation, the abovementioned Similarity Search module, we are expanding our efforts with the development of the “Custom Classifier.” This feature empowers machine operators to easily identify similar undesirable machine conditions and take timely corrective actions.

For instance, when detecting tool bluntness, we swiftly identify this condition as it arises, preventing further production of rejects. This is achieved by identifying and labeling data representing a specific fault, which is then used to train a model. Subsequently, we can effectively identify similar states of the machine as they occur.

Designed for non-programmers, particularly machine operators, this module enables continuous marking of undesirable machine states. The system then automatically searches for similar faults in the machine’s behavior, streamlining the intervention process.

In environments where thousands of metal parts are produced daily on a machine tool, such capabilities are indispensable. Algorithms evaluate these conditions in real-time, drawing on insights gained from previous process monitoring experiences.

Data Fusion: Enhancing Machine Status Identification through Multimodal Data Analysis

At our core, we employ data fusion to achieve more precise detection of machine and process states. Utilizing multi-modal models—algorithms capable of processing multiple data types simultaneously—we enhance our analysis. For instance, by integrating sound, temperature, and electrical current data from a single machine, we can effectively correlate and compare these values.

By amalgamating data from diverse sources, we mitigate the risk of false alerts and avoid unnecessary operator interventions in processes. This aspect of our technology empowers, for instance, electric motors to determine their operational status more accurately from a remote standpoint. Consequently, operators can better assess the severity of faults and devise appropriate strategies for machine operation adjustments.

Expanding beyond acoustic, temperature, and current data, we are actively exploring the integration of visual data from cameras, pressure measurements, and oil quality assessments into our analysis.

Machine Learning Industry Trends: What to Expect This Year

Our team is closely tracking the latest advancements in machine learning, also known as the “State of the Art” developments within the industry. We are harnessing cutting-edge techniques such as convolutional networks, originally designed for image processing but now applied to sound and other physical parameters. Additionally, similar to innovations seen with ChatGPT and OpenAI, we are utilizing transformers equipped with attention mechanisms.

These advancements are evident in one of our recent projects—a generative AI application tailored for automated report processing for our clients. By taking machine documentation as input, our AI recommends interventions based on the machine’s current state, utilizing pre-processed manuals. Once again, our goal is to streamline operations for operators while providing expertise akin to that of diagnosticians and mechanical engineers where it is most needed.