AI (Artificial Intelligence) abnormal sound detection is an intelligent detection method based on machine learning theory. Different from traditional manual detection methods, it operates on an AI system composed of computers and software. Through learning and training with a large number of samples, the system automatically acquires criteria for judging abnormal sounds, thereby realizing the automatic detection of abnormal sounds. The following sections explain AI abnormal sound detection in terms of its principles, implementation, and detection effects.

I. Machine Learning

The criteria for judging abnormal sounds in AI-based abnormal sound detection are derived through machine learning methods. Based on a large number of samples, data-driven models and corresponding model parameters are obtained via automatic training, and these together constitute the judgment criteria for abnormal sound detection. Machine learning methods are the core of AI abnormal sound detection. Below, we will understand the basic principles of machine learning by comparing it with the human learning process.

The human learning process: Humans learn from materials and information, then generalize, organize, and summarize what they have learned. After repeated review and practice, this knowledge is eventually internalized as knowledge and experience. When encountering similar problems or situations, humans can then make correct responses and complete tasks.

Based on a large number of training samples (such as audio files of home appliances), a home appliance abnormal sound detection model is obtained using specific machine learning algorithms. This model is a data-driven model, similar to the knowledge and experience acquired by humans. Then, the model is tested and verified with test samples, and its parameters are optimized until the test results meet the expectations. Finally, the trained abnormal sound detection model is deployed to the production line to realize real-time online abnormal sound detection.

II. Exploration of Abnormal Sound Detection

In the research and exploration process of abnormal sound detection, three main methods have been developed, namely the most direct manual method, the early threshold method, and the current artificial intelligence method. The characteristics of these three methods are briefly introduced below.

• Manual Method: It relies on human ears for listening and manual judgment. Most sound detection workstations mainly adopt this method, which is in urgent need of intelligent upgrading to form a complete closed loop of intelligent manufacturing together with other automatic modules on the production line.

• Threshold Method: It analyzes and compares the differences between abnormal sounds and normal sounds, and establishes one or a set of parameters as the threshold for the normal range. Machines with abnormal sounds are screened out through this threshold. Practical application shows that this method is difficult to reach the accuracy of manual judgment and thus hard to realize true replacement of human labor.

• Artificial Intelligence Method: Based on a large number of audio samples, it conducts automatic training to obtain a data-driven abnormal sound judgment model. From the perspective of actually completed cases, this method has already reached or even exceeded the accuracy of manual judgment, can truly realize unmanned automatic detection, and can form a complete closed loop of intelligent manufacturing with other automatic modules on the production line.

III. Implementation Process

As can be seen from the machine learning process, AI-based abnormal sound detection consists of three major parts. Here we take compressor abnormal sound detection as an example to illustrate:

• Sample Collection: Vibration signals of compressors are acquired using acceleration sensors and data acquisition cards. The signal samples undergo data cleaning to remove those that do not meet requirements, and the remaining samples serve as the dataset for machine learning.

• Machine Learning: An appropriate machine learning model is selected, initial training parameters are set, and training is performed on the dataset to obtain an abnormal sound detection model. The model undergoes testing and verification, with parameters adjusted and repeated training conducted according to specific objectives until the test results meet expectations.

• Real-time Detection: The abnormal sound detection model is deployed to the production line. Real-time collected signal samples are input into the detection model, and judgment results are obtained within one production line cycle to complete online real-time detection.

IV. System Composition

The AI-based abnormal sound detection system can automatically determine the product status. It uses an AI abnormal sound detection model to replace the ears and judgment of professional technicians. The system can be divided into two parts: hardware and software, as detailed below:

Hardware

1. Acceleration Sensor: Converts the vibration acceleration of the compressor housing into an electrical signal and outputs it.

2. Data Acquisition Card: Converts the electrical signal output by the acceleration sensor into a digital signal that can be stored and analyzed by a computer, then outputs the digital signal.

3. Computer: Stores and analyzes the digital signals output by the data acquisition card.

Software

1. Sample Collection Program: Labels, records, cleans, and stores signal samples.

2. Machine Learning Program: Trains the abnormal sound detection model and tests/verifies the model's accuracy.

3. Abnormal Sound Detection Program: Applies the abnormal sound detection model to realize online abnormal sound detection.

V. Advantages

Compared with the manual method and the threshold method, the advantages of the AI-based abnormal sound detection are as follows.