Equipment abnormal sound detection method based on deep learning
Through deep learning technology, multi-source heterogeneous data is integrated, and equipment anomaly detection is performed using multi-layer neural networks and attention mechanisms, the problems of low efficiency and insufficient robustness in the existing technology are solved, and efficient and intelligent equipment anomalies are achieved. Detection improves detection accuracy and user experience.
Patent Information
- Application Number
- CN202411088424.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-13
AI Technical Summary
The existing equipment abnormality detection technology is inefficient, lacks adaptability and robustness, and it is difficult to cope with complex and variable working conditions. The model-based method has limited ability to express data sources and is easily disturbed by noise, resulting in errors in judgment.
Using deep learning-based equipment anomaly sound detection method, by collecting and preprocessing heterogeneous data such as sound, vibration, and temperature, using multi-layer deep neural network fusion characteristics, combining attention mechanisms and bidirectional long and short-term memory networks for timing modeling. Combined with supervised and semi-supervised learning, a visual monitoring interface and a continuous optimization mechanism are provided.
It improves the accuracy and robustness of equipment abnormality detection, reduces dependence on large amounts of labeled data, enhances the interpretability and user experience of the system, and forms a closed-loop automated abnormality detection and diagnosis system, which can continuously improve detection accuracy.
Smart Images
Figure CN119128627A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning, and in particular to a method for detecting abnormal sound of a device based on deep learning. Background Art
[0002] Equipment anomaly detection technology refers to a method of determining whether a device has a fault or abnormal state by analyzing various data generated during the operation of the device. Timely and accurate detection of equipment anomalies is crucial to preventing major accidents, reducing maintenance costs, and improving production efficiency.
[0003] At present, there are mainly the following methods for detecting equipment anomaly: 1) Relying on human experience, discovering anomalies through human observation and judgment; 2) Rule-based methods, setting rules such as thresholds in advance, and judging it as abnormal when the detection data exceeds the rule range; 3) Model-based methods, building a model of the data when the equipment is operating normally, and judging it as abnormal when the new data deviates greatly from the model. However, these traditional methods have certain limitations.
[0004] Manual detection is inefficient and difficult to detect hidden faults in a timely manner; rule-based methods lack adaptability, often require a lot of experience and knowledge, and are difficult to cope with complex and changing working conditions; and model-based methods have certain requirements for model establishment and maintenance, and the model expression ability of a single data source is limited, and it is easy to be disturbed by noise and cause misjudgment. Therefore, a new intelligent and more robust anomaly detection method is urgently needed to solve the shortcomings of existing technologies. Summary of the invention
[0005] In order to solve the problems of the prior art, an embodiment of the present invention provides a method for detecting abnormal sound of a device based on deep learning. The technical solution is as follows:
[0006] On the one hand, a method for detecting abnormal sound of a device based on deep learning is provided, comprising:
[0007] Collecting sound data and at least one type of heterogeneous data during the operation of the device;
[0008] Preprocessing the collected sound data to obtain sound features, and preprocessing the collected heterogeneous data to obtain heterogeneous features;
[0009] Inputting the obtained sound features into a first deep neural network to learn a pattern representation of the sound features, and inputting the obtained heterogeneous features into a second deep neural network to learn a pattern representation of the heterogeneous features;
[0010] The attention fusion mechanism is used to fuse the pattern representation of sound features with the pattern representation of heterogeneous features to obtain a fused feature representation;
[0011] The fused feature representation is input into the third deep neural network, and the anomaly detection score is obtained through time series modeling;
[0012] The real-time collected sound data and heterogeneous data are input into the trained deep neural network model, and the abnormality detection score obtained is used to determine whether the equipment has an abnormality and output an abnormality alarm.
[0013] Furthermore, the step of preprocessing the sound data to obtain the sound features includes:
[0014] Perform pre-processing such as noise reduction and resampling on the original sound data;
[0015] The preprocessed sound data is converted into the time-frequency domain and sound features such as Mel-frequency cepstral coefficients are extracted.
[0016] Furthermore, the sound features also include mel loudness features and gammatone frequency cepstral coefficient features.
[0017] Furthermore, the heterogeneous data includes vibration data and temperature data during the operation of the equipment.
[0018] Furthermore, the method also includes a step of augmenting the sound data and heterogeneous data to expand abnormal data samples.
[0019] Furthermore, the first deep neural network adopts a convolution kernel structure adapted to time-frequency features, and uses a convolution block attention module to model the long-range dependency between features of different frequency bands.
[0020] Furthermore, the third deep neural network includes a bidirectional long short-term memory network for capturing temporal context information of features.
[0021] Furthermore, the training of the deep neural network model includes:
[0022] Supervised training based on labeled normal and abnormal data;
[0023] The model is further trained using semi-supervised learning methods combined with unlabeled data.
[0024] Furthermore, the semi-supervised learning method adopts a contrastive learning paradigm.
[0025] Furthermore, it also includes:
[0026] Display real-time sound waveforms, abnormal detection score change trends, possible fault causes and handling suggestions;
[0027] The deep neural network model is continuously optimized and updated according to actual maintenance feedback information.
[0028] The technical solution provided by the embodiment of the present invention has the following beneficial effects:
[0029] 1. By integrating multi-source heterogeneous data such as sound, vibration, and temperature, it can fully reflect the operating status of the equipment and improve the robustness and accuracy of detection.
[0030] 2. A convolution kernel structure that adapts to time-frequency features is proposed, and the attention mechanism is used to model the long-range dependencies between features of different frequency bands, effectively extracting rich local and global pattern information in sound signals.
[0031] 3. Combining the time series modeling capabilities of the bidirectional long short-term memory network, it can capture the evolution of abnormal behaviors in the time dimension and improve the ability to detect continuous anomalies.
[0032] 4. The use of multiple training strategies such as supervised, semi-supervised, and transfer learning reduces the dependence on large amounts of labeled data and reduces the cost and workload of data preparation.
[0033] 5. Provides an intelligent visual monitoring interface to intuitively present real-time data, abnormal trends and fault diagnosis information, greatly improving the system's interpretability and user experience.
[0034] 6. By continuously optimizing the model and combining it with actual maintenance feedback, a closed-loop automated anomaly detection and diagnosis system is formed, which can continuously improve detection accuracy.
[0035] 7. The algorithm has low deployment cost and only requires ordinary sound and vibration sensors. It can be widely used in factories, power and other fields to realize intelligent operation and maintenance and fault warning of industrial equipment.
[0036] 8. The innovation of this technical solution lies in the combination of multimodal deep learning and industrial big data, which overcomes many defects of traditional detection methods and proposes a new paradigm that is efficient, robust and intelligent, with important theoretical significance and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0038] Figure 1 It is a flowchart of a method for detecting abnormal sound of a device based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION
[0039] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0040] This paper proposes a method for detecting abnormal sound of equipment based on deep learning, aiming to improve the accuracy and intelligence level of abnormality detection. Figure 1 :.
[0041] The method of the present invention first collects sound data and at least one heterogeneous data during the operation of the equipment, such as vibration data, temperature data, etc. This multi-source heterogeneous data fusion method can fully reflect the operating status of the equipment and improve the robustness of the detection.
[0042] For sound data, preprocessing such as noise reduction and resampling is required, and then it is converted to the time-frequency domain to extract rich sound features such as Mel frequency cepstral coefficients (MFCC), Mel loudness features, and Gammatone frequency cepstral coefficients as neural network input. Compared with traditional MFCC, this feature can better describe the fine-grained pattern of sound.
[0043] For heterogeneous data such as vibration and temperature data, corresponding statistical features are extracted, such as vibration energy, temperature mean, etc.
[0044] The present invention also uses data augmentation technology to expand sound and heterogeneous data to increase the sample size of abnormal data and enhance the generalization ability of the model.
[0045] Next, the obtained sound features are input into the first deep neural network for modeling. This neural network uses a convolution kernel structure that adapts to time-frequency features, which can effectively capture the local pattern of the sound signal and use the convolution block attention module to model the long-range dependency of different frequency band features. At the same time, the heterogeneous features are input into the second deep neural network for modeling.
[0046] Then, the output features of the two networks are fused using an attention fusion mechanism to obtain a fused feature representation, which is then input into a third deep neural network. As described in claim 7, the network includes a bidirectional long short-term memory network, which can capture the temporal context information of the features and perform temporal modeling of abnormal behaviors.
[0047] The deep neural network eventually outputs an anomaly detection score, with a high score indicating a possible anomaly. The model inputs the real-time collected sound and heterogeneous data, and determines whether the device status is abnormal based on the output score. When the score exceeds the preset threshold, an abnormality alert is output.
[0048] The deep neural network model of the present invention adopts multiple training strategies. First, supervised training is performed using labeled normal and abnormal data; then semi-supervised learning methods such as contrastive learning are used to further train the model in combination with unlabeled data to reduce the need for data labeling.
[0049] Finally, the present invention designs a Web-based visual monitoring interface to intuitively display real-time sound waveforms, abnormal score change trends, possible fault causes and processing suggestions, etc., and continuously optimizes and updates the model based on actual maintenance feedback to form a closed-loop intelligent detection and diagnosis system.
[0050] Through the above implementation methods, the deep learning-based anomaly detection method proposed in the present invention can intelligently and efficiently detect the operating status of the equipment, discover anomalies in time, and reduce the risk of failure, which has important theoretical significance and application value.
[0051] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for detecting abnormal sound of equipment based on deep learning, characterized in that: include: Collecting sound data and at least one type of heterogeneous data during the operation of the device; Preprocessing the collected sound data to obtain sound features, and preprocessing the collected heterogeneous data to obtain heterogeneous features; Inputting the obtained sound features into a first deep neural network to learn a pattern representation of the sound features, and inputting the obtained heterogeneous features into a second deep neural network to learn a pattern representation of the heterogeneous features; The attention fusion mechanism is used to fuse the pattern representation of sound features with the pattern representation of heterogeneous features to obtain a fused feature representation; The fused feature representation is input into the third deep neural network, and the anomaly detection score is obtained through time series modeling; The real-time collected sound data and heterogeneous data are input into the trained deep neural network model, and the abnormality detection score obtained is used to determine whether the equipment has an abnormality and output an abnormality alarm.
2. The method according to claim 1, characterized in that: The step of preprocessing the sound data to obtain the sound features comprises: Pre-process the original sound data by denoising and resampling; The preprocessed sound data is converted into the time-frequency domain and the sound features of the Mel-frequency cepstral coefficients are extracted.
3. The method according to claim 2, characterized in that The sound features also include mel loudness features and gammatone frequency cepstral coefficient features.
4. The method according to claim 1, characterized in that: The heterogeneous data includes vibration data and temperature data during the operation of the equipment.
5. The method according to claim 1, characterized in that The method also includes a step of performing data augmentation on sound data and heterogeneous data to expand abnormal data samples.
6. The method according to claim 1, characterized in that The first deep neural network adopts a convolution kernel structure adapted to time-frequency features, and uses a convolution block attention module to model the long-range dependency between features of different frequency bands.
7. The method according to claim 1, characterized in that The third deep neural network includes a bidirectional long short-term memory network for capturing temporal context information of features.
8. The method according to claim 1, characterized in that The training of the deep neural network model includes: Supervised training based on labeled normal and abnormal data; The model is further trained using semi-supervised learning methods combined with unlabeled data.
9. The method according to claim 8, characterized in that The semi-supervised learning method adopts the paradigm of contrastive learning.
10. The method according to claim 1, characterized in that Also includes: Display real-time sound waveforms, abnormal detection score change trends, possible fault causes and handling suggestions; The deep neural network model is continuously optimized and updated according to actual maintenance feedback information.