A method and system for monitoring leakage of molten aluminum in a distributor based on sound recognition

Abstract

The application discloses a kind of based on sound recognition's shunt disc molten aluminum leakage monitoring method and system.The method is collected by sound signal acquisition device the sound of shunt disc and background environmental noise;Establishment sound dataset and logarithm mel spectrum chart dataset;Based on the lightweight neural network with picture feature extraction capability, construct sound recognizer and train;The sound of shunt disc collected is detected using the trained sound recognizer, to judge whether there is molten aluminum leakage.The application uses lightweight neural network, identifies the special sound emitted when molten aluminum leakage directly contacts water, realizes fast, accurate aluminum leakage detection, solves the problem that traditional convolutional neural network model is relatively large, and requires high hardware on site, and has important significance for safety production of deep well casting enterprises.

Description

Technical Field

[0001] This invention belongs to the technical field of safe production, specifically relating to a method and system for monitoring aluminum liquid leakage in a diversion plate based on sound recognition. Background Technology

[0002] The aluminum processing industry is a high-risk industry, with smelting and casting workshops posing a deadly explosion risk. In the past five years, there has been at least one major safety accident annually in China involving explosions in smelting and casting workshops with three or more employees, causing severe casualties and economic losses to enterprises. The main cause of these explosions is the contact between molten aluminum and water, forming a "water-in-aluminum" phenomenon. This causes the liquid water to rapidly vaporize and expand, resulting in a vapor explosion. Therefore, rapid detection and timely handling of molten aluminum contact with water are crucial for preventing explosions.

[0003] Currently, there is a lack of research on detecting molten aluminum leaks in China, and the main method still relies on manual experience, which requires a high level of expertise. Since the temperature of molten aluminum is around 700℃, when a leaking molten aluminum comes into direct contact with water, it will emit a distinctive "sizzling" sound, which can be used as a basis for judging whether there is a leak. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method and system for monitoring aluminum liquid leakage in a diversion plate based on sound recognition. It adopts a lightweight neural network to achieve rapid and accurate detection of aluminum leakage, which is of great significance to the safe production of deep well casting enterprises.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] One aspect of the present invention provides a method for monitoring aluminum molten metal leakage in a flow divider based on sound recognition, comprising the following steps:

[0007] The sound from the distributor plate and background ambient noise are collected using a sound signal acquisition device;

[0008] Establish a sound dataset and a log-Mel spectrogram dataset;

[0009] A voice recognizer was built and trained based on a lightweight neural network with image feature extraction capabilities;

[0010] The trained sound recognition device is used to detect the sound collected from the distribution plate to determine whether there is aluminum leakage.

[0011] As a preferred technical solution, the sound signal acquisition device includes multiple bidirectional microphones and a mixer; the bidirectional microphones are arranged around the bottom of the splitter plate and face the bottom surface of the splitter plate; the mixer combines the sound signals acquired by the multiple bidirectional microphones into one sound signal.

[0012] As a preferred technical solution, the establishment of the sound dataset and the log-Mel spectrogram dataset specifically involves:

[0013] The ambient sound and aluminum leakage sound collected by the sound signal acquisition device are mixed to obtain a variety of normal and abnormal sounds, thereby establishing a sound dataset; the normal sound contains only ambient sound, and the abnormal sound contains both ambient sound and aluminum leakage sound.

[0014] Extract the log-Mel spectrogram of each sound class from the sound dataset and label the log-Mel spectrogram;

[0015] The extraction of the log-Mel spectrum for each type of sound is specifically as follows:

[0016] For each type of sound, an audio sample with a duration of x∈[1-10] seconds is read each time, and read once every y seconds, where y≤x. Each audio sample is framed and windowed. The processed sound signal X(n) is subjected to a fast Fourier transform to obtain the signal X(k). The linear frequency of X(k) is converted to the Mel-scale frequency using a Mel filter, and a log-Mel spectrogram is generated. That is, the audio duration corresponding to one log-Mel spectrogram is x seconds, and one is generated every y seconds.

[0017] As a preferred technical solution, the step of performing frame segmentation and windowing processing on each sound sample specifically involves:

[0018] The original sound signal S(n) is divided into several frames, and the frame shift is t. m It must be smaller than the frame length t s Half of;

[0019] Each frame of signal is windowed using a Hamming window. The expression for the Hamming window is as follows:

[0020]

[0021] Where N is the number of samples in a frame, N = f s ·t s , where a is the coefficient and n is the number of samples in a frame.

[0022] As a preferred technical solution, the lightweight neural network based on image feature extraction capability adopts one of the following networks: EfficientNetV1, EfficientNetV2, MobileNetV2, MobileNetV3, and ShufflNet V2.

[0023] The EfficientNetV2 network consists of eight stages. Stage 0 is a standard 3×3 convolutional layer, BN layer, and SiLU activation function. Stages 1-3 are Fused-MBConv layers, stages 4-6 are MBConv layers, and stage 7 is a 1×1 convolutional layer, average pooling layer, and fully connected layer. SE attention mechanism is added to stages 1-6.

[0024] Specifically, the SE attention mechanism is replaced with a parameterless 3D attention mechanism in the MobileNetV3 network, and Dropout is added to the MobileNetV3 network.

[0025] As a preferred technical solution, the training steps for the voice recognition system are as follows:

[0026] 1) Divide the log-Mel spectrogram into a training set and a validation set;

[0027] 2) Initialize the network parameters and weights, and set the hyperparameter width coefficient to 1;

[0028] 3) Input the training set and validation set into the model for training, and plot the loss and Acc curves during the training process;

[0029] 4) Update the model's weights and biases using the Adam optimizer;

[0030] 5) Compare the classification accuracy Acc under each epoch, retain the best result of Acc, and save the parameters to the folder after training.

[0031] Another aspect of the present invention provides a sound recognition-based aluminum liquid leakage monitoring system for a diversion plate, including a sound signal acquisition device and an aluminum leakage sound event detection module;

[0032] The sound signal acquisition device is used to acquire the sound of the distributor plate and background environmental noise;

[0033] The aluminum leakage sound event detection module includes a log-Mel spectrum generator and a sound recognizer;

[0034] The log-Mel spectrogram generator is used to generate a log-Mel spectrogram based on the sound signal acquired by the sound signal acquisition device.

[0035] The sound recognizer is built and trained based on a lightweight neural network with image feature extraction capabilities, and is used to determine whether there is aluminum leakage in the corresponding sound signal based on the log-Mel spectrogram.

[0036] As a preferred technical solution, the sound signal acquisition device includes multiple bidirectional microphones and a mixer; the bidirectional microphones are arranged around the bottom of the splitter plate and face the bottom surface of the splitter plate; the mixer combines the sound signals acquired by the multiple bidirectional microphones into one sound signal.

[0037] As a preferred technical solution, a log-Mel spectrum is generated based on the sound signal acquired by the sound signal acquisition device, specifically as follows:

[0038] The ambient sound and aluminum leakage sound collected by the sound signal acquisition device are mixed to obtain a variety of normal and abnormal sounds, thereby establishing a sound dataset; the normal sound contains only ambient sound, and the abnormal sound contains both ambient sound and aluminum leakage sound.

[0039] Extract the log-Mel spectrogram of each sound class from the sound dataset and label the log-Mel spectrogram;

[0040] The extraction of the log-Mel spectrum for each type of sound is specifically as follows:

[0041] For each type of sound, an audio sample with a duration of x∈[1-10] seconds is read each time, and read once every y seconds, where y≤x. Each audio sample is framed and windowed. The processed sound signal X(n) is subjected to a fast Fourier transform to obtain the signal X(k). The linear frequency of X(k) is converted to the Mel-scale frequency using a Mel filter, and a log-Mel spectrogram is generated. That is, the audio duration corresponding to one log-Mel spectrogram is x seconds, and one is generated every y seconds.

[0042] As a preferred technical solution, the lightweight neural network based on image feature extraction capability adopts one of the following networks: EfficientNetV1, EfficientNetV2, MobileNetV2, MobileNetV3, and ShufflNetV2.

[0043] The EfficientNetV2 network consists of eight stages. Stage 0 is a standard 3×3 convolutional layer, BN layer, and SiLU activation function. Stages 1-3 are Fused-MBConv layers, stages 4-6 are MBConv layers, and stage 7 is a 1×1 convolutional layer, average pooling layer, and fully connected layer. SE attention mechanism is added to stages 1-6.

[0044] Specifically, the SE attention mechanism is replaced with a parameterless 3D attention mechanism in the MobileNetV3 network, and Dropout is added to the MobileNetV3 network.

[0045] The training steps for the voice recognition device are as follows:

[0046] 1) Divide the log-Mel spectrogram into a training set and a validation set;

[0047] 2) Initialize the network parameters and weights, and set the hyperparameter width coefficient to 1;

[0048] 3) Input the training set and validation set into the model for training, and plot the loss and Acc curves during the training process;

[0049] 4) Update the model's weights and biases using the Adam optimizer;

[0050] 5) Compare the classification accuracy Acc under each epoch, retain the best result of Acc, and save the parameters to the folder after training.

[0051] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0052] (1) The present invention provides a sound acquisition device that reduces the dead zone of sound pickup while reducing the interference of ambient sound;

[0053] (2) This invention uses lightweight neural networks with image feature extraction capabilities, such as EfficientNetV2, MobileNetV3 and ShuffleNet; it solves the problem that traditional convolutional neural network models are relatively large and have high requirements for on-site hardware.

[0054] (3) This invention adds an SE attention mechanism to the EfficientNetV2 network to increase the model’s ability to extract various sound features;

[0055] (4) This invention improves the ability of MobileNetV3 network to extract aluminum leakage sound features by using the SE attention mechanism and the parameterless 3D attention mechanism in the MobileNetV3 network without increasing the network parameters; and adds Dropout to the MobileNetV3 network to suppress the overfitting phenomenon of the model. Attached Figure Description

[0056] Figure 1 This is a flowchart of the aluminum leakage sound recognition process of the shunt plate according to an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the installation position of the sound acquisition device according to an embodiment of the present invention;

[0058] Figure 3 This is a log-Mel spectrum of the sound of water flow according to an embodiment of the present invention. Detailed Implementation

[0059] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0060] Example

[0061] like Figure 1 As shown, this embodiment provides a method for monitoring aluminum molten metal leakage in a flow divider based on sound recognition, including the following steps:

[0062] S1. Collect the sound of the distributor plate and background ambient noise through a sound signal acquisition device;

[0063] like Figure 2 As shown, the sound signal acquisition device includes multiple bidirectional microphones and a mixer; the bidirectional microphones are arranged around the bottom of the splitter plate and face the bottom surface of the splitter plate; the mixer combines the sound signals acquired by the multiple bidirectional microphones into one sound signal.

[0064] S2. Create a sound dataset and a log-Mel spectrogram dataset;

[0065] Various ambient sounds and aluminum leakage sounds were collected using a sound signal acquisition device. The ambient sounds included, but were not limited to, water flow sounds, sawing sounds, forklift sounds, and fan sounds. These ambient sounds were mixed with aluminum leakage sounds to obtain various normal and abnormal sounds, including, but not limited to (1) aluminum leakage sound + water flow sound, (2) aluminum leakage sound, (3) aluminum leakage sound + water flow sound + sawing sound, (4) water flow sound + sawing sound, etc., as shown in Table 1. These audios were classified and a sound dataset was established.

[0066]

[0067] Table 1. Types of Audio Datasets

[0068] Log-Mel spectrograms of each sound class are extracted from the sound dataset, and the log-Mel spectrograms are labeled according to the type of acoustic scene.

[0069] The extraction of the log-Mel spectrum for each type of sound is specifically as follows:

[0070] For each type of sound, an audio sample of duration x∈[1-10] seconds is read each time, with a reading every y seconds, where y≤x. Each audio sample is then processed by framing and windowing, specifically as follows:

[0071] The original sound signal S(n) is divided into several frames, and the frame shift is t. m It must be smaller than the frame length t s Half of;

[0072] Each frame of signal is windowed using a Hamming window. The expression for the Hamming window is as follows:

[0073]

[0074] Where N is the number of samples in a frame, N = f s ·t s , where a is the coefficient and n is the number of samples in a frame.

[0075] The processed audio signal X(n) is subjected to a Fast Fourier Transform (FFT) to obtain signal X(k). A Mel filter is then used to convert the linear frequency of X(k) to Mel-scale frequencies, generating a log-Mel spectrogram. One log-Mel spectrogram corresponds to an audio duration of x seconds, and one spectrogram is generated every y seconds. For example... Figure 3 The image shows the log-Mel spectrum of the sound of flowing water.

[0076] S3. Based on a lightweight neural network with image feature extraction capabilities, a voice recognizer is constructed and trained using one of the following networks: EfficientNetV1, EfficientNetV2, MobileNetV2, MobileNetV3, and ShufflNet V2.

[0077] The EfficientNetV2 network consists of eight stages. Stage 0 is a standard 3×3 convolutional layer, BN layer, and SiLU activation function. Stages 1-3 are Fused-MBConv layers, stages 4-6 are MBConv layers, and stage 7 is a 1×1 convolutional layer, average pooling layer, and fully connected layer. SE attention mechanism is added to stages 1-6.

[0078] Specifically, the SE attention mechanism in the MobileNetV3 network is replaced with a parameterless 3D attention mechanism, and Dropout is added to the MobileNetV3 network.

[0079] The training steps for a voice recognition system are as follows:

[0080] 1) Divide the log-Mel spectrogram into a training set and a validation set;

[0081] 2) Initialize the network parameters and weights, and set the hyperparameter width coefficient to 1;

[0082] 3) Input the training set and validation set into the model for training, and plot the loss and Acc curves during the training process;

[0083] 4) Update the model's weights and biases using the Adam optimizer;

[0084] 5) Compare the classification accuracy Acc under each epoch, retain the best result of Acc, and save the parameters to the folder after training.

[0085] S4. Use the trained sound recognition device to detect the sound of the collected distribution plate to determine whether there is aluminum liquid leakage.

[0086] In another embodiment of this application, a sound recognition-based aluminum liquid leakage monitoring system for a diversion plate is provided. The system includes a sound signal acquisition device and an aluminum leakage sound event detection module.

[0087] 1. The sound signal acquisition device is used to acquire the sound of the distributor plate and background environmental noise, specifically:

[0088] Multiple bidirectional microphones are arranged around the bottom of the splitter plate, facing towards the bottom surface of the plate, to minimize dead zones. The bidirectional microphones separately collect sound from the target direction and background ambient noise. A mixer is used to combine the multiple audio signals into a single audio signal. Their placement is as follows: Figure 2 As shown;

[0089] In actual casting processes, sometimes a single distribution plate produces as many as 70 or even 100 aluminum rods with a small diameter. When operators apply talc powder to the graphite rings in each crystallizer before production, some graphite rings may have missed areas or unevenly coated talc powder, easily leading to aluminum leakage at the start of casting. Furthermore, due to the large number of aluminum rods, insufficient heat exchange between the cooling water in the distribution plate and the crystallizer also easily causes aluminum leakage during casting. Since the temperature of the aluminum liquid is around 700℃, when leaking aluminum liquid comes into direct contact with water, it will emit a distinctive sound, which can be used as an indicator of aluminum leakage. Because the ambient noise, including sawing, fan noise, forklift noise, and water flow noise, usually affects the collection and analysis of aluminum leakage sounds, a directional microphone will be used as the primary sound acquisition device to reduce the impact of ambient noise on the target sound acquisition.

[0090] 2. The aluminum leakage sound event detection module includes a log-Mel spectrum generator and a sound recognizer;

[0091] (1) The log-Melbourne spectrogram generator is used to generate a log-Melbourne spectrogram based on the sound signal acquired by the sound signal acquisition device, specifically including the following steps:

[0092] (1.1) Establish a sound dataset.

[0093] Various ambient sounds and aluminum leakage sounds are collected using a sound signal acquisition device. The ambient sounds include, but are not limited to, water flow sounds, sawing sounds, forklift sounds, and fan sounds. These ambient sounds are mixed with aluminum leakage sounds to obtain various normal and abnormal sounds, including but not limited to (1) aluminum leakage sound + water flow sound, (2) aluminum leakage sound, (3) aluminum leakage sound + water flow sound + sawing sound, (4) water flow sound + sawing sound, etc., as shown in Table 1 of the above embodiment. These audios are classified and a sound dataset is established.

[0094] (1.2) Log-Mel spectrogram extraction. Log-Mel spectrograms for each sound class are extracted from the sound dataset, and the corresponding labels are assigned to the log-Mel spectrograms according to the type of acoustic scene.

[0095] First, each sound sample is framed and windowed. Then, the processed sound signal is processed using a Mel filter to generate the corresponding log-Mel spectrogram.

[0096] The extraction of the log-Mel spectrum for each type of sound is specifically as follows:

[0097] For each type of sound, an audio sample of duration x∈[1-10] seconds is read each time, with a reading every y seconds, where y≤x. Each audio sample is then processed by framing and windowing, specifically as follows:

[0098] The original sound signal S(n) is divided into several frames, and the frame shift is t. m It must be smaller than the frame length t s Half of;

[0099] Each frame of signal is windowed using a Hamming window. The expression for the Hamming window is as follows:

[0100]

[0101] Where N is the number of samples in a frame, N = f s ·t s , where a is the coefficient and n is the number of samples in a frame.

[0102] The processed audio signal X(n) is subjected to a Fast Fourier Transform (FFT) to obtain signal X(k). A Mel filter is then used to convert the linear frequency of X(k) to Mel-scale frequencies, generating a log-Mel spectrogram. One log-Mel spectrogram corresponds to an audio duration of x seconds, and one spectrogram is generated every y seconds. For example... Figure 3 The image shows the log-Mel spectrum of the sound of flowing water.

[0103] (2) The sound recognizer is based on a lightweight neural network with image feature extraction capability. It is constructed and trained using one of the following networks: EfficientNetV1, EfficientNetV2, MobileNetV2, MobileNetV3, and ShufflNet V2. It is used to determine whether the corresponding sound signal has aluminum leakage based on the log-Mel spectrogram.

[0104] (2.1) Establishing a voice recognizer. Since traditional convolutional neural network models are quite large and have high requirements for on-site hardware, a lightweight neural network with image feature extraction capabilities is used to establish a voice recognizer.

[0105] In this embodiment, the EfficientNetV2 network has a total of 8 stages. Stage 0 is a regular 3×3 convolutional layer, BN layer and SiLU activation function. Stages 1 to 3 are all Fused-MBConv layers. Stages 4 to 6 are all MBConv layers. Stage 7 is a 1×1 convolutional layer, average pooling layer and fully connected layer. In addition, SE attention mechanism is added to stages 1 to 6.

[0106] For MobileNetV3, the SE attention mechanism increases the number of network parameters, which slows down the model training speed. Therefore, in this embodiment, the SE attention mechanism in the MobileNetV3 network is replaced with a parameterless 3D attention mechanism to improve the MobileNetV3 network's ability to extract aluminum leakage sound features without increasing the network parameters. In addition, since binary classification models are prone to overfitting, Dropout is added to the MobileNetV3 network to suppress overfitting, and the extracted log-Melogram spectrum is input into the sound recognizer for training.

[0107] (2.2) The model training steps are as follows:

[0108] 1) Divide the log-Mel spectrogram into a training set and a validation set;

[0109] 2) Initialize the network parameters and weights, and set the hyperparameter width coefficient to 1;

[0110] 3) Input the training set and validation set into the model for training, and plot the loss and Acc curves during the training process;

[0111] 4) Update the model's weights and biases using the Adam optimizer;

[0112] 5) Compare the classification accuracy Acc under each epoch, retain the best result of Acc, and save the parameters to the folder after training.

[0113] (3) The aluminum leakage sound detection module based on log-Melogram spectrum is set up in the cloud server. The sound signal acquisition device collects sound signals from the production site and uploads them to the cloud server in real time. The aluminum leakage sound detection module reads and analyzes the detected sound signals in real time on the cloud server. When the aluminum leakage sound detection module detects aluminum leakage sound, it issues an aluminum leakage alarm signal. The alarm signal is transmitted from the cloud server back to the alarm device in the production site, realizing the alarm process.

[0114] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. This system is a sound recognition-based aluminum liquid leakage monitoring method for the diversion plate applied to the above embodiments.

[0115] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0116] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for monitoring leakage of molten aluminum in a distributor based on sound recognition, characterized by, Includes the following steps: The sound of the distribution plate and background ambient noise are collected by a sound signal acquisition device. The sound signal acquisition device includes multiple bidirectional microphones and a mixer. The bidirectional microphones are arranged around the bottom of the distribution plate and face the bottom surface of the distribution plate. The mixer combines the sound signals collected by the multiple bidirectional microphones into a single sound signal. The sound dataset and the log-Mel spectrogram dataset are established as follows: The ambient sound and aluminum leakage sound collected by the sound signal acquisition device are mixed to obtain a variety of normal and abnormal sounds, thereby establishing a sound dataset; the normal sound contains only ambient sound, and the abnormal sound contains both ambient sound and aluminum leakage sound. Extract the log-Mel spectrogram of each sound class from the sound dataset and label the log-Mel spectrogram; The extraction of the log-Mel spectrum for each type of sound is specifically as follows: For each type of sound, an audio sample with a duration of x∈[1-10] seconds is read each time, and read once every y seconds, where y≤x. Each audio sample is framed and windowed. The processed sound signal X(n) is subjected to a fast Fourier transform to obtain the signal X(k). The linear frequency of X(k) is converted to the Mel-scale frequency using a Mel filter, and a log-Mel spectrogram is generated. That is, the audio duration corresponding to one log-Mel spectrogram is x seconds, and one is generated every y seconds. A voice recognizer was built and trained based on a lightweight neural network with image feature extraction capabilities; The trained sound recognition device is used to detect the sound collected from the distribution plate to determine whether there is aluminum leakage; The lightweight neural network based on image feature extraction capability adopts one of the following networks: EfficientNetV1, EfficientNetV2, MobileNetV2, MobileNetV3, and ShufflNet V2. The EfficientNetV2 network consists of eight stages. Stage 0 is a standard 3×3 convolutional layer, BN layer, and SiLU activation function. Stages 1-3 are Fused-MBConv layers, stages 4-6 are MBConv layers, and stage 7 is a 1×1 convolutional layer, average pooling layer, and fully connected layer. SE attention mechanism is added to stages 1-6. Specifically, the SE attention mechanism is replaced with a parameterless 3D attention mechanism in the MobileNetV3 network, and Dropout is added to the MobileNetV3 network.

2. The method of claim 1, wherein the method is a method of monitoring leakage of molten aluminum in a distributor based on sound recognition. The process of framing and windowing each audio sample specifically involves: The original sound signal S(n) is divided into several frames, and the frame shift is less than half of the frame length . Each frame of signal is windowed using a Hamming window. The expression for the Hamming window is as follows: ； where N is the number of samples in a frame, a is a coefficient, N is the number of samples in a frame.

3. A sound recognition based monitoring system for leakage of molten aluminium from a distributor tray, characterised in that, Includes a sound signal acquisition device and an aluminum leakage sound event detection module; The sound signal acquisition device is used to acquire the sound of the distributor plate and background environmental noise; The aluminum leakage sound event detection module includes a log-Mel spectrum generator and a sound recognizer; The log-Mel spectrum generator is used to generate a log-Mel spectrum based on the sound signal acquired by the sound signal acquisition device, specifically: The ambient sound and aluminum leakage sound collected by the sound signal acquisition device are mixed to obtain various normal and abnormal sounds, thereby establishing a sound dataset; the normal sound contains only ambient sound, and the abnormal sound contains both ambient sound and aluminum leakage sound; the sound signal acquisition device includes multiple bidirectional microphones and a mixer; the bidirectional microphones are arranged around the bottom of the distribution plate and face the bottom surface of the distribution plate; the mixer combines the sound signals collected by the multiple bidirectional microphones into a single sound signal; Extract the log-Mel spectrogram of each sound class from the sound dataset and label the log-Mel spectrogram; The extraction of the log-Mel spectrum for each type of sound is specifically as follows: For each type of sound, an audio sample with a duration of x∈[1-10] seconds is read each time, and read once every y seconds, where y≤x. Each audio sample is framed and windowed. The processed sound signal X(n) is subjected to a fast Fourier transform to obtain the signal X(k). The linear frequency of X(k) is converted to the Mel-scale frequency using a Mel filter, and a log-Mel spectrogram is generated. That is, the audio duration corresponding to one log-Mel spectrogram is x seconds, and one is generated every y seconds. The sound recognizer is built and trained based on a lightweight neural network with image feature extraction capabilities, and is used to determine whether there is aluminum leakage in the corresponding sound signal based on the log-Mel spectrogram. The lightweight neural network based on image feature extraction capability adopts one of the following networks: EfficientNetV1, EfficientNetV2, MobileNetV2, MobileNetV3, and ShufflNetV2. The EfficientNetV2 network consists of eight stages. Stage 0 is a standard 3×3 convolutional layer, BN layer, and SiLU activation function. Stages 1-3 are Fused-MBConv layers, stages 4-6 are MBConv layers, and stage 7 is a 1×1 convolutional layer, average pooling layer, and fully connected layer. SE attention mechanism is added to stages 1-6. Specifically, the SE attention mechanism is replaced with a parameterless 3D attention mechanism in the MobileNetV3 network, and Dropout is added to the MobileNetV3 network.

Images