Wine bottle crack detection method based on knocking device

Abstract

The invention discloses a winebottle crack detection method based on a knocking device, and the method comprises the steps: collecting the knocking audio data of various types of normal winebottles and winebottles with cracks, and constructing a sample database; training a deep learning algorithm model by using the collected sample data; in the equipment operation process, audio signals generated by knocking the wine bottle are collected in real time, the trained deep learning model is called to analyze the audio signals, and whether cracks exist in the wine bottle or not is judged. And an audio signal generated by knocking is analyzed and processed by adopting a deep learning algorithm, so that high-precision crack detection is realized.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to a method for detecting cracks in wine bottles based on a tapping device. Background Technology

[0002] In the wine bottle production process, traditional manual methods of inspecting for bottle cracks are inefficient and inaccurate, making it difficult to meet the needs of large-scale production. Existing automated inspection equipment suffers from low accuracy and a high rate of false rejections, resulting in persistently high defect rates and increased production costs for manufacturers.

[0003] Currently, some bottle crack detection equipment uses simple tapping devices and basic sound analysis methods. For example, some equipment uses manual tapping of the bottle, with operators relying on experience to judge the presence of cracks by listening to the sound. This method is extremely inefficient and easily affected by operator subjectivity, making it difficult to guarantee detection accuracy. Some automated equipment uses simple tapping devices with unstable tapping force and frequency, resulting in poor quality audio signals. In terms of sound analysis, simple threshold judgment methods are used, which cannot accurately identify complex crack features, leading to a high false rejection rate. For example, one existing device has a detection speed of approximately 3000 bottles / hour, a rejection rate of only 85%, and a false rejection rate as high as 2%, far from meeting the requirements of manufacturing enterprises for detection accuracy and efficiency.

[0004] In existing technologies, manual inspection methods suffer from low accuracy and low efficiency due to factors such as human auditory fatigue and differences in experience, making them unsuitable for large-scale production. In automated equipment, the instability of the tapping device means that the generated audio signal cannot accurately reflect the true condition of the bottle, and simple sound analysis methods cannot effectively distinguish between normal bottles and cracked bottles, resulting in a high false rejection rate and a low rejection rate. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a method for detecting bottle cracks based on a tapping device to overcome or at least partially solve the above problems.

[0006] According to one aspect of the present invention, a method for detecting cracks in a wine bottle based on a tapping device is provided, the method comprising:

[0007] Collect tapping audio data of various types of normal wine bottles and cracked wine bottles to build a sample database;

[0008] The collected sample data is used to train the deep learning algorithm model;

[0009] During equipment operation, the audio signals generated by tapping the wine bottle are collected in real time, and a trained deep learning model is used to analyze the audio signals to determine whether the wine bottle has cracks.

[0010] Optionally, the detection method further includes:

[0011] Perform statistical analysis on the data generated during equipment testing;

[0012] Analyze the data to promptly identify problems during equipment operation and optimize and adjust the equipment accordingly.

[0013] Optionally, the sample database specifically includes: various possible crack conditions and the audio characteristics of normal wine bottles.

[0014] Optionally, training the deep learning algorithm model using the collected sample data specifically includes:

[0015] The deep learning algorithm model adopts a deep learning model based on the CAM+ResNet module;

[0016] The deep learning algorithm model is an improved deep learning framework that combines the channel attention mechanism (CAM) and the ResNet residual network for feature extraction and classification tasks.

[0017] The deep learning algorithm model is divided into an input layer, a neighborhood feature extraction layer, a spectral-space improved residual learning module, a pooling layer, a fully connected layer, and an output layer.

[0018] Embedding the attention mechanism into the residual block enhances the learning ability of key features;

[0019] The system collects the audio signal of the wine bottle being tested by tapping in real time, performs preprocessing and feature extraction, and inputs it into a pre-trained deep learning algorithm model for classification and judgment.

[0020] Optionally, the step of collecting audio signals generated by tapping the bottle in real time during device operation, and calling a trained deep learning model to analyze the audio signals to determine whether the bottle has cracks specifically includes:

[0021] Multimodal feature extraction and fusion includes:

[0022] Multidimensional PNCC coefficients: Based on traditional PNCC, multidimensional PNCC introduces a multidimensional feature fusion mechanism, combining perceptual cepstral coefficients with multidimensional features.

[0023] Time-frequency joint domain features, spatial domain features, higher-order statistical features, and contextual information;

[0024] Multidimensional LPC coefficients: By extending the traditional LPC model to a multidimensional space;

[0025] Multidimensional LPC introduces a multidimensional feature fusion mechanism on the basis of traditional LPC, combining linear prediction coefficients with other multidimensional features to form a speech signal representation;

[0026] An improved spatial attention mechanism is employed, including:

[0027] Dual-path feature aggregation: Frequency-time domain joint attention weights are generated through a parallel structure of mean pooling and max pooling;

[0028] Dilated convolution design: Dilated convolutional layers are used to expand the receptive field to capture long-term dependencies;

[0029] Residual connection: Feature map dimension matching is achieved through 1×1 convolution.

[0030] Optionally, the residual connection: achieving feature map dimension matching through 1×1 convolution specifically includes:

[0031] A residual convolutional neural network architecture is adopted;

[0032] Feature extraction layer: Consists of two levels of residual modules, batch normalization and ReLU activation function, supporting deep feature learning;

[0033] Classification Header: Employs a combination of global average pooling and fully connected layers to output anomaly probability distribution.

[0034] This invention provides a method for detecting bottle cracks based on a tapping device. The method includes: collecting tapping audio data from various types of normal and cracked bottles to construct a sample database; training a deep learning algorithm model using the collected sample data; and during device operation, acquiring the audio signals generated by tapping the bottles in real time, and using the trained deep learning model to analyze the audio signals to determine whether the bottle has a crack. By employing advanced deep learning algorithms to analyze and process the tapping audio signals, high-precision crack detection is achieved.

[0035] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating a method for detecting bottle cracks based on a tapping device, provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the hammer assembly provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of the motor and cam provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the structure before being struck, provided in an embodiment of the present invention;

[0041] Figure 5 This is a schematic diagram of the structure during striking provided in an embodiment of the present invention;

[0042] Figure 6 A flowchart illustrating the deep learning model provided in an embodiment of the present invention;

[0043] Figure 7 This is a structural diagram of the device in operation provided in an embodiment of the present invention;

[0044] Figure 8 A flowchart of system operation provided in an embodiment of the present invention;

[0045] Figure 9 This is a flowchart of the shutdown operation provided in an embodiment of the present invention;

[0046] Figure 10 The debugging mode operation process provided in the embodiments of the present invention;

[0047] Figure 11 A flowchart illustrating the complete processing chain from raw signal input to classified output provided in this embodiment of the invention;

[0048] Figure 12 A schematic diagram of the spatial attention mechanism provided in an embodiment of the present invention. Detailed Implementation

[0049] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0050] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0052] like Figure 1 As shown, a method for detecting cracks in a wine bottle based on a tapping device is described, the method comprising:

[0053] Audio data from tapping various types of normal and cracked wine bottles are collected to construct a sample database. The collected sample data is used to train a deep learning algorithm model. During equipment operation, the audio signals generated by tapping the wine bottles are collected in real time, and the trained deep learning model is used to analyze the audio signals to determine whether the wine bottle has cracks.

[0054] The purpose of this invention is to provide a highly efficient and accurate bottle crack detection device based on a tapping mechanism. By optimizing the structure of the tapping mechanism, its tapping sensitivity is made high and its response is fast, enabling it to stably generate high-quality audio signals. Simultaneously, advanced deep learning algorithms are used to analyze and process the acquired audio signals, improving the accuracy of crack detection. The technical effects achievable by this invention are: detection capacity ≥ 5000 bottles / hour, rejection rate ≥ 96%, false rejection rate ≤ 0.1%, effectively improving product quality, reducing defect rate, reducing production costs, and enabling online detection, thus improving production efficiency.

[0055] This testing equipment mainly consists of a striking device, a sound acquisition unit, a main control computer, a sound AI inspection system, an automated control system, an operation display platform, a rejection device, an equipment housing, and a data acquisition device. The striking device includes a hammer assembly, a motor, a base, a cam, a spring, and a stop. The hammer assembly is fixed to the base via a rotating shaft and can rotate relative to the base around the axis of rotation. The motor and stop are fixed to the base, the cam is fixed to the motor output shaft, and the stop limits the swing angle range of the hammer assembly. During operation, the motor drives the cam to rotate. When the cam rotates to a certain angle, the hammer assembly begins to rotate under the action of the cam. After disengaging from the cam, it continues to rotate under inertia, striking the bottle. After completing the striking action, it resets under the reaction force of the impact and the tension of the spring. The sound acquisition unit collects the audio signal generated by striking the bottle and transmits it to the sound AI inspection system. The main control computer, as the core control unit of the entire equipment, runs deep learning algorithms to analyze and process the audio signal. The automated control system is responsible for controlling the overall operation of the equipment, such as controlling the conveying speed of the bottles and the working rhythm of the tapping device. The operation display platform is used by operators to operate and monitor the equipment, displaying information such as test results. The rejection device, according to instructions from the main control computer, removes defective bottles to the non-conforming collection platform box. The data acquisition device is responsible for collecting various data during equipment operation, such as the number of bottles tested and the number of defective bottles, for statistical analysis.

[0056] like Figure 3 The diagram shows an embodiment of the motor and cam connection method: the cam consists of a cam base 41 and a cam follower 42. The cam follower 42 is fixed to the cam base 41 to form a complete cam 4. The cam base 41 is fixed to the output shaft of the motor 2. When the output shaft of the motor 2 rotates, it drives the cam 4 to rotate.

[0057] like Figure 2 As shown, one embodiment of the hammer assembly is as follows: The hammer assembly 1 consists of a striking rod 11 and a hammer base 12. The striking rod 11 is fixed to the hammer base 12 to form the hammer assembly 1. The hammer base 12 has a shaft hole that mates with the rotating shaft 7, allowing it to rotate.

[0058] like Figure 4 and Figure 5 As shown in the two figures before and during the strike, the hammer assembly 1 is limited to its extreme position by the stop block 6.

[0059] The striking force is adjusted by regulating the motor speed. When a higher striking force is needed, the motor output speed is increased, allowing cam 4 to reach a higher speed. Cam 4 drives hammer assembly 1 to rotate at this high speed, resulting in higher rotational kinetic energy and a greater impact force when striking the target object, achieving a powerful striking effect. Conversely, reducing the motor output speed slows down cam 4's rotation, reducing the kinetic energy of hammer assembly 1, suitable for gentler striking needs. Precise control of the motor speed allows for flexible adjustment of the striking force to meet the needs of different working scenarios. During hammer assembly reset, the spring tension and reaction force work together to ensure accurate reset. The sound AI inspection system analyzes audio signals in real time to determine the quality of the bottles. Based on the analysis results, the main control computer directs the rejection device to efficiently separate defective products, ensuring both production efficiency and product quality.

[0060] A striking device for non-destructive testing includes: a hammer assembly 1, a motor 2, a base 3, a cam 4, a spring 5, a stop 6, a rotating shaft 7, a spring hanger 18, and a spring hanger 29. The hammer assembly is fixed to the base via the rotating shaft and can rotate relative to the base around the axis of the rotating shaft. The motor and the stop are fixed to the base, and the cam is fixed to the motor output shaft. The stop limits the swing angle range of the hammer assembly. During operation, when the motor drives the cam to rotate a certain angle, the hammer assembly begins to rotate under the action of the cam. After disengaging from the cam, it continues to rotate under inertia. After the striking action is completed, it resets under the reaction force of striking the container being tested and the spring tension, completing one striking action.

[0061] The advantages of this device are high tap sensitivity and fast response. It immediately detaches after tapping the container under test, which helps to improve the audio firing quality of the container under test.

[0062] Sample Acquisition Module: Before the equipment is put into use, a large amount of audio data from tapping different types of normal and cracked wine bottles is collected to build a sample database. This sample data covers various possible crack conditions and the audio characteristics of normal wine bottles.

[0063] like Figure 6 As shown, the deep learning training module uses collected sample data to train the deep learning algorithm model. By continuously adjusting the model's parameters, the model can accurately learn the feature differences between audio signals from normal and cracked wine bottles, thus enabling it to accurately classify unknown audio signals.

[0064] The model employs a deep learning framework based on CAM+ResNet modules: this model is an improved deep learning framework that combines the Channel Attention (CAM) mechanism with the ResNet residual network, primarily used for feature extraction and classification tasks. The model is divided into an input layer, a neighborhood feature extraction layer, a spectral-space improved residual learning module, a pooling layer, a fully connected layer, and an output layer. Its core innovation lies in embedding the attention mechanism into the residual blocks to enhance the learning ability of key features.

[0065] The model implementation process is shown in the figure below. The audio signal of the wine bottle being tested is collected in real time, preprocessed, and feature extracted. The signal is then input into the trained model for classification and judgment.

[0066] like Figure 7 As shown, the detection module: During equipment operation, the sound acquisition unit collects the audio signals generated by tapping the wine bottle in real time and transmits them to the main control computer. The detection module calls a trained deep learning model to analyze the audio signals and determine whether the wine bottle has cracks.

[0067] The system operation process is as follows: Figure 8 As shown, the shutdown procedure is as follows: Figure 9 As shown, the debug mode operation process is as follows: Figure 10 As shown.

[0068] The sound acquisition unit is positioned near the percussion device to ensure clear audio signal capture.

[0069] The detection process is as follows: detect when the bottle reaches the striking position → activate the striking device → collect audio signals → transmit to the main control computer → model analysis and judgment → output the detection results.

[0070] Production data statistics and analysis module: This module performs statistical analysis on data generated during equipment testing, such as the number of products tested, the number of defective products, the rejection rate, and the false rejection rate. By analyzing this data, problems in equipment operation can be identified in a timely manner, allowing for equipment optimization and adjustments to improve testing accuracy and stability.

[0071] The uniquely designed striking mechanism ensures stable and efficient striking of the bottle, generating high-quality audio signals and improving the quality of audio output.

[0072] Advanced deep learning algorithms are used to analyze and process the audio signals generated by the impact, achieving high-precision crack detection.

[0073] The entire testing equipment works in concert, including the precise control of the equipment's operation by the automated control system, and the data interaction and collaboration between the data acquisition devices, the sound AI inspection system, and the main control computer. Points to be protected:

[0074] The specific structural components of the striking device, their connection relationships, and their working methods, such as the coordination between the hammer assembly, motor, cam, spring, and stop.

[0075] The crack detection method based on deep learning algorithm and related software modules include the functions and implementation process of sample acquisition module, deep learning training module, detection module, and production data statistics and analysis module.

[0076] The system architecture of the entire testing equipment includes the connection relationships and collaborative working methods of the various hardware components, as well as the interaction methods between software and hardware.

[0077] Regarding the striking mechanism, an electromagnetically driven striking method could be considered to replace the motor-driven cam method. By controlling the magnitude and frequency of the electromagnetic force, precise control of the hammer can be achieved, producing a stable striking action.

[0078] In audio analysis algorithms, besides deep learning algorithms, other machine learning algorithms such as Support Vector Machines (SVM) can be used to extract features and classify audio signals. By comparing and optimizing different algorithms, the most suitable one can be selected to improve detection accuracy.

[0079] In terms of the overall architecture of the device, a distributed computing approach can be explored. Distributing the sound analysis task across multiple computing nodes improves data processing speed and further enhances the device's detection efficiency.

[0080] The model is an audio anomaly detection method and system based on multimodal feature fusion and spatial attention mechanism, aiming to achieve the following objectives: (1) improve the completeness of anomaly features through multimodal feature fusion; (2) enhance the time-frequency domain feature expression capability by using audio multidimensional feature extraction and spatial attention mechanism; and (3) improve the detection robustness of the system in noisy environments.

[0081] like Figure 11 As shown, this model includes:

[0082] 1. Multimodal feature extraction and fusion

[0083] (1) Multidimensional PNCC coefficients: Based on traditional PNCC, multidimensional PNCC introduces a multidimensional feature fusion mechanism to combine the perceptual cepstral coefficients with the following multidimensional features:

[0084] Features in the time-frequency joint domain (such as short-time Fourier transform and wavelet transform), spatial domain features (such as spatial information of microphone arrays), higher-order statistical features (such as spectral flatness and chromaticity features), and contextual information (such as LSTM output for time series modeling) are all included. Multidimensional PNCC significantly improves the robustness and accuracy of audio signal analysis by fusing perceptual cepstral features with multidimensional features. Its core lies in:

[0085] Perceptual modeling: Simulating the characteristics of human hearing;

[0086] Multidimensional complementarity: combining time-frequency, spatial, and higher-order statistical features;

[0087] Context modeling: Using time-series models to capture dynamic changes in audio.

[0088] (2) Multidimensional LPC Coefficients: Multidimensional LPC is an audio feature extraction method based on linear predictive coding. By extending the traditional LPC model to a multidimensional space (such as multi-channel, time-frequency domain, spatial domain, etc.), it enhances the ability to model speech signals, and is particularly suitable for speaker recognition, speech synthesis, and noise suppression tasks in complex scenarios. Multidimensional LPC introduces a multidimensional feature fusion mechanism on the basis of traditional LPC, combining linear predictive coefficients with other multidimensional features (such as spatial information, time-frequency joint domain features, contextual dependencies, etc.) to form a more comprehensive speech signal representation. Its core objective is to enhance robustness to environmental noise and other interferences through the complementarity of multidimensional features.

[0089] (3) Zero crossing rate: measures the number of times an audio signal crosses zero in a unit of time, reflecting the frequency and transient characteristics of the signal.

[0090] (4) Spectral characteristics: including spectral amplitude, spectral centroid, spectral bandwidth, and spectral roll-off, which intuitively reflect the frequency domain distribution.

[0091] like Figure 12 As shown, the improved spatial attention mechanism

[0092] (1) Dual-path feature aggregation: Through the parallel structure of mean pooling and maximum pooling, frequency domain-time domain joint attention weights are generated.

[0093] (2) Dential convolution design: Dential convolution layers are used to expand the receptive field to capture long-term dependencies.

[0094] (3) Residual connection: feature map dimension matching is achieved through 1×1 convolution, avoiding the gradient vanishing problem during the training process of the attention mechanism.

[0095] Residual convolutional neural network architecture

[0096] (1) Feature extraction layer: It consists of two levels of residual modules (ResBlock), batch normalization (BatchNorm) and ReLU activation function, which support deep feature learning.

[0097] (2) Classification head: The classification head is a combination of global average pooling (AdaptiveAvgPool2d) and a fully connected layer to output the probability distribution of anomalies.

[0098] Beneficial effects:

[0099] Regarding the striking mechanism, an electromagnetically driven striking method could be considered to replace the motor-driven cam method. By controlling the magnitude and frequency of the electromagnetic force, precise control of the hammer can be achieved, producing a stable striking action.

[0100] In audio analysis algorithms, besides deep learning algorithms, other machine learning algorithms such as Support Vector Machines (SVMs) can be used to extract features and classify audio signals. By comparing and optimizing different algorithms, the most suitable one can be selected to improve detection accuracy.

[0101] In terms of the overall architecture of the device, a distributed computing approach can be explored. Distributing the sound analysis task across multiple computing nodes improves data processing speed and further enhances the device's detection efficiency.

[0102] In practical applications of container crack detection, this invention achieves the following performance improvements compared to traditional solutions: detection accuracy is increased from 92.2% to 99.98%, and noise robustness is improved by ±50dB. Traditional algorithms have a detection rate of 30% to 72%, which is greatly affected by production noise, while this algorithm has a detection rate of 95% to 98%, demonstrating strong noise resistance and better meeting the actual factory production environment.

[0103] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting cracks in wine bottles based on a tapping device, characterized in that, The detection method includes: Collect tapping audio data of various types of normal wine bottles and cracked wine bottles to build a sample database; The collected sample data is used to train the deep learning algorithm model; During equipment operation, the audio signals generated by tapping the wine bottle are collected in real time, and a trained deep learning model is used to analyze the audio signals to determine whether the wine bottle has cracks.

2. The method for detecting bottle cracks based on a tapping device according to claim 1, characterized in that, The detection method further includes: Perform statistical analysis on the data generated during equipment testing; Analyze the data to promptly identify problems during equipment operation and optimize and adjust the equipment accordingly.

3. The method for detecting bottle cracks based on a tapping device according to claim 1, characterized in that, The sample database specifically includes: various possible crack conditions and the audio characteristics of normal wine bottles.

4. The method for detecting bottle cracks based on a tapping device according to claim 1, characterized in that, The process of training the deep learning algorithm model using the collected sample data specifically includes: The deep learning algorithm model adopts a deep learning model based on the CAM+ResNet module; The deep learning algorithm model is an improved deep learning framework that combines the channel attention mechanism (CAM) and the ResNet residual network for feature extraction and classification tasks. The deep learning algorithm model is divided into an input layer, a neighborhood feature extraction layer, a spectral-space improved residual learning module, a pooling layer, a fully connected layer, and an output layer. Embedding the attention mechanism into the residual block enhances the learning ability of key features; The system collects the audio signal of the wine bottle being tested by tapping in real time, performs preprocessing and feature extraction, and inputs it into a pre-trained deep learning algorithm model for classification and judgment.

5. The method for detecting bottle cracks based on a tapping device according to claim 1, characterized in that, The process of collecting audio signals generated by tapping the bottle in real time during equipment operation, and analyzing the audio signals using a trained deep learning model to determine whether the bottle has cracks specifically includes: Multimodal feature extraction and fusion includes: Multidimensional PNCC coefficients: Based on traditional PNCC, multidimensional PNCC introduces a multidimensional feature fusion mechanism, combining perceptual cepstral coefficients with multidimensional features. Time-frequency joint domain features, spatial domain features, higher-order statistical features, and contextual information; Multidimensional LPC coefficients: By extending the traditional LPC model to a multidimensional space; Multidimensional LPC introduces a multidimensional feature fusion mechanism on the basis of traditional LPC, combining linear prediction coefficients with other multidimensional features to form a speech signal representation; An improved spatial attention mechanism is employed, including: Dual-path feature aggregation: Frequency-time domain joint attention weights are generated through a parallel structure of mean pooling and max pooling; Dilated convolution design: Dilated convolutional layers are used to expand the receptive field to capture long-term dependencies; Residual connection: Feature map dimension matching is achieved through 1×1 convolution.

6. The method for detecting bottle cracks based on a tapping device according to claim 5, characterized in that, The residual connection: feature map dimension matching is achieved through 1×1 convolution, specifically including: A residual convolutional neural network architecture is adopted; Feature extraction layer: Consists of two levels of residual modules, batch normalization and ReLU activation function, supporting deep feature learning; Classification Header: Employs a combination of global average pooling and fully connected layers to output anomaly probability distribution.

Images