Do you notice anything different about me today? — Anomaly detection with Sound AI
Have you ever felt a sudden sense of fear when you’re home alone and the refrigerator makes an unexpected noises? Initially, the sound might frighten you, leading to concerns about whether there’s a problem with your refrigerator. I’m not the only one experiencing confusion when confronted with unexpected machine noises, right? During such cases, when you reach out to the after-service center, they politely inquire about the type of noise you heard. Then, they provide information on whether it’s a normal occurrence or not. If the situation is abnormal, they offer recommendations on how to solve it.
In situations like these, individuals who have spent a significant amount of time in a specific environment and gained experience with various sounds can easily distinguish between normal and abnormal sounds. Furthermore, they develop a keen understanding of the unique traits of specific sounds, almost becoming the masters. Sound carries more information than one might initially think.
[What is anomaly detection?]
Anomaly detection involves identifying rare and abnormal events that deviate significantly from the majority of occurrences and do not align with typical scenarios. Abnormal events can be viewed as deviations from the general and normal patterns, encompassing situations that are faulty, fake, noisy, or erroneous. For example, a sudden and drastic increase in server traffic may indicate a ‘DDoS’ attack. Similarly, if a credit card is used at a convenience store in Seoul, and then, 30 minutes later, a large amount of money is spent either abroad or in a domestic location outside Seoul, the transaction may seem suspicious.
Consider the profession of someone who can discern engine sounds to detect any issues. In our daily lives, we frequently encounter situations when it becomes crucial to differentiate between what is normal and what is abnormal. Anomaly detection finds widespread application in various sectors, including finance, medicine, security, and manufacturing.
I’ll explain anomaly detection to a real life moment that feels more familiar to us. If you’re part of a couple and enjoy spending time together, you’ve either heard or posed a familiar question: “Do you notice anything different about me today?” Honestly and personally, this is a rather challenging question to answer. The term ‘anything different’ can be subjective and dependent on various standards. If you’re not familiar with the specific criteria in question, making an accurate guess becomes quite tricky. At the heart of inquiry lies the notion that, in comparison to typical days, there’s something unique or exceptional about the person. While changes like a new hairstyle or the addition of glasses are easily noticeable, subtle alterations such as a different lipstick color or a distinct belt design can be considerably more challenging to know at first glance.
The most difficult aspect of anomaly detection lies in its need to encompass every normal status across various situations, observe circumstances through multiple criteria, and respond to even the minutest details. This highlights the key technical distinction between classification and anomaly detection. In classification, there is often a clear, correct answer, such as ‘Distinguish between U and V.’ Conversely, anomaly detection involves analyzing data with the objective of identifying deviations from the common data distribution, framed as ‘Find the differences compared to the typical data distribution situation.’
[Why is anomaly detection so important?]
This significance of anomaly detection lies in its ability to prevent damage from potentially dangerous situations that may arise in our daily lives. By identifying abnormal patterns or behaviors in real-time, anomaly detection plays a crucial role in various sectors. In finance, it aids in detecting fraud or cheating, while in the IT sector, it helps prevent errors within systems. In the manufacturing industry, anomaly detection facilitates the separation of defective products. Ultimately, anomaly detection serves as a valuable tool benefiting both producers involved in product creation and the end-users of products or services.
People naturally refine their senses through both direct and indirect experiences. At times, we might perceive something as ‘weird’, and there’s a saying that this feeling is not just a mere sense but a form of big data built upon accumulated life experiences. When confronted with an abnormal situation, it’s challenging to provide a logical explanation, yet individuals often make quick decisions based on cumulative life experiences. This sense of anomaly detection is typically cultivated through personal experiences, often referred to as ‘embodied patterns’. In contrast, AI models approach anomaly detection through training. They pre-recognize variables based on various standards and develop an understanding of potential outcomes associated with each variables.
[Sound anomaly detection by Cochl]
In this way, anomaly detection can be achieved through the utilization of sound. In the past, particularly for identifying flaws in machines, human intuition or experience played an important role in flaw detection. However, with the advent of trained AI models, the accuracy of anomaly detection has significantly improved. So, why did Cochl decide to venture into anomaly detection using sound?
Cochl has established a foundational model for sound detection, offering this technology to users in the form of an API or SDK. When we assert that the model can generally detect various sounds, it implies that the model possesses a keen understanding of the subtle characteristics of each sound, enabling it to extract valuable traits from the audio with high accuracy. Consequently, when the model identifies a new sound that hasn’t been part of its training, it can quantitatively measure the degree of difference between the original sound and the newly detected sound.
As an example, let’s consider the case of Company A, which utilizes our anomaly detection technology. The company manufactures components integrated into automobiles and faces the challenge of identifying and removing defective products that emit anomalous sounds from the window. In order to enhance their Quality Assurance/Quality Control (QA/QC) processes, they opted for our solution. Previously, the company had constructed a soundproof room within the factory premises, where employees had to manually inspect each sound. However, this approach presented several issues, including:
- High construction costs for the soundproof room
- Ongoing labor costs associated with manual inspections
- Potential for human errors and dependency (e.g,. variations in fault coverage rates due to fatigue, differences in skill levels, and a lack of consistent working history)
By leveraging Cochl’s technology, Company A can significantly reduce labor costs and minimize the likelihood of human errors. AI can execute the same work processes consistently under unified conditions, leading to a reduction in fault coverage rates, decreased human intervention, and the opportunity for human employees to focus on more productive tasks rather than manually detecting anomalous sounds.
In the past, individuals with skilled experience were relied on to identify anomalous situations. However, AI provides a quicker and more efficient method for pinpointing the causes of errors. Nonetheless, there remains a unique capability exclusive to humans that AI cannot fully replace. Furthermore, AI serves as a supportive tool, alleviating the burden on humans.
If you have an interest in Cochl’s anomaly detection technology and the support it can offer, please don’t hesitate to reach out to us at contact@cochl.ai!
