Porcelain damage detection method based on sound

Abstract

The invention discloses a porcelain damage detection method based on sound. The porcelain damage detection method comprises the following steps: step 1, establishing a lossless and damaged porcelain sound database; step 2, performing normalization preprocessing on the sound signals; step 3, first-stage main phase feature extraction; 4, performing second-stage main phase feature extraction; step 5, splicing the first-stage main phase feature and the second-stage main phase feature to generate a sound signal phase feature; step 6, SVM abnormal sound classifier learning is carried out; step 7, acquiring knocking sound signals of the porcelain with unknown damage conditions, performing normalization preprocessing, extracting phase features, and determining whether the porcelain is damaged or not according to SVM classification; the method has the characteristics of simple and easy operation and high detection precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of sound signal processing, and in particular relates to a porcelain damage detection method based on audio information. Background Art

[0002] Porcelain is no longer just a daily necessity; it's a cultural work of art. However, its inherent fragility, determined by its material and manufacturing process, makes it susceptible to breakage from external forces during production, transportation, and use. Therefore, rapid and accurate porcelain breakage detection is crucial for ensuring quality and minimizing economic losses, and represents a highly worthy research challenge.

[0003] At present, the existing methods for detecting porcelain damage are mainly divided into two categories: manual detection and automatic detection:

[0004] Manual inspection method: workers mainly inspect the porcelain surface with their naked eyes or with the help of simple tools (such as magnifying glasses, measuring instruments, etc.), and judge whether there are cracks, chips or other damage defects based on experience. However, this method has the following significant drawbacks:

[0005] The detection efficiency is low, unable to meet the needs of large-scale production scenarios, and the labor cost is high;

[0006] The test results are greatly affected by workers' experience, subjective judgment, and fatigue level, which can easily lead to misjudgment or missed detection;

[0007] For defects such as minor damage or internal cracks, manual identification is difficult and detection accuracy is difficult to guarantee.

[0008] Automatic detection method: The current mainstream automatic detection technology is based on visual information analysis. Its typical process is: a camera captures an image of a standard intact porcelain and an image of the porcelain to be inspected, and compares the difference between the two images pixel by pixel. If the difference exceeds a preset threshold, the porcelain is judged to be damaged. Although this method reduces manual intervention to a certain extent, it still has the following technical bottlenecks:

[0009] The image acquisition equipment has stringent requirements for accuracy and stability, and images must be acquired under strictly controlled lighting and angle conditions, which are difficult to meet in actual production environments;

[0010] The algorithm is sensitive to image noise and is prone to misjudgment due to uneven lighting, reflections, or dust interference, requiring a significant amount of time for image preprocessing.

[0011] The computational complexity of pixel-by-pixel comparison is high, the real-time detection performance is poor, and it is difficult to adapt to the needs of high-speed production lines;

[0012] It can only detect visible damage on the surface, but cannot identify hidden damage to the internal structure of the porcelain.

[0013] In addition, neither of the above two methods can simultaneously strike a balance between detection accuracy, efficiency and cost. It is urgent to propose a new detection solution to overcome the inherent defects of existing technologies. Summary of the Invention

[0014] In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide a porcelain breakage detection method based on sound information, which has the characteristics of being simple and easy to operate and having high detection accuracy.

[0015] To achieve the above object, the technical solution adopted by the present invention is: a sound-based porcelain damage detection method, comprising the following steps:

[0016] Step 1: Create a sound database of intact and damaged porcelain

[0017] Prepare intact and damaged porcelain samples, use a microphone to collect the vibration sound signals generated by knocking on the surface of the porcelain samples, and save them into an audio file in wav format. The file name contains damaged or lossless category information, which indicates the intact sound sample and the damaged sound sample;

[0018] Step 2: Normalize and preprocess the sound signal

[0019] The maximum and minimum normalization method is used to linearly transform the amplitude of the sound signal to between [0,1];

[0020] Step 3: Extract the main phase features in the first stage. The specific steps of this step are:

[0021] Step 3-1: Divide the sound signal into frames. The frame length and frame shift of the first stage are represented by fl1 and fm1 respectively, in milliseconds.

[0022] Step 3-2: Perform a short-time discrete Fourier transform on each frame of the sound signal to transform the time domain information x into frequency domain information X:

[0023] X(m,k)=STFT[x(m,n)] (1)

[0024] In formula (1), x(m,n) refers to the nth value of the mth frame of the sound signal, and k is the frequency index;

[0025] Step 3-3, let the imaginary part and real part of each frame short-time Fourier transform result X(m,k) be represented as X r 、X i , the phase characteristics of each frame of Fourier transform are:

[0026] θ=arctan2(X r ,X i ) (2)

[0027] In formula (2), θ represents the phase characteristic;

[0028] Step 3-4: Define the width of the sliding window as wl1 points, and traverse the short-time Fourier transform phase image with a step size of ww1 points. During the traversal process, perform SVD decomposition on the phase feature θ of each sliding window of the phase image.

[0029] θ=USV T (3)

[0030] In formula (3), U and V are orthogonal matrices, the elements s1≥s2≥s3≥… on the diagonal of the diagonal matrix S are the singular values of the phase feature θ, and the maximum singular value s1 is taken as the main phase feature value of the sliding window. The main phase features of all sliding windows are sequentially spliced to generate the main phase feature of the first stage;

[0031] Step 4: Second stage main phase feature extraction

[0032] The second stage of main phase feature extraction takes the first stage main phase feature as input and performs the same processing on the input as the first stage. The difference from the first stage is the different values of the frame length and frame shift, sliding window step size and width parameters.

[0033] First, divide the frame. Here, the frame length and frame shift are fl2 and fm2 points respectively.

[0034] Secondly, perform short-time discrete Fourier transform on each frame.

[0035] Then find the Fourier transform phase characteristics,

[0036] Finally, the phase feature is decomposed by SVD using a sliding window with a width of wl2 points and a step size of wm2 points. All the maximum singular values of the SVD decomposition are sequentially spliced to form the second stage main phase feature.

[0037] Step 5: Sound signal phase characteristics

[0038] The main phase characteristics of the first stage and the main phase characteristics of the second stage are sequentially spliced to form the final phase characteristics of the sound signal;

[0039] Step 6: SVM abnormal sound classifier learning

[0040] The lossless and lossy sound samples collected in step 1 are set as positive samples and negative samples respectively. The SVM classifier is learned using the ten-fold cross-validation method. The parameters that need to be learned or adjusted during the learning process include the main phase feature parameters of the first stage: fl1, fm1, wl1, ww1, the main phase feature parameters of the second stage: fl2, fm2, wl2, ww2, and related parameters of the SVM classifier;

[0041] Step 7: For unknown damaged porcelain, first collect the sound signal generated by knocking, then normalize and preprocess it, and finally extract the phase features according to the optimal parameters determined in step 6 and use the SVM classifier to classify the sound signal. When the classification result is a positive sample, it is intact porcelain, otherwise it is damaged porcelain.

[0042] The audio file has a sampling frequency of 44100 Hz, a collection time of 3 seconds, and a collection accuracy of 16 bits.

[0043] The beneficial effects of the present invention are:

[0044] 1) The present invention utilizes the different characteristics of the knocking sound of broken porcelain to extract features from the collected sound and perform porcelain breakage detection, thereby making up for the shortcomings of manual or visual detection methods.

[0045] 2) The present invention proposes a sound phase feature extraction method that is simple and easy to understand.

[0046] 3) The present invention utilizes the SVM classifier based on the sound phase feature to detect porcelain damage with high detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure is a flow chart of a porcelain damage detection method based on sound of the present invention. DETAILED DESCRIPTION

[0048] The specific implementation methods for realizing the invention are described in detail (if there are drawings, the explanation should be made with reference to the drawings, but the drawings cannot replace the text description).

[0049] In the embodiments, not only the various technical features involved in the technical solution should be described in detail, but also the relevant contents that help to understand the invention should be described in detail (for example: the preparation process and equipment of the product, the source of raw materials, the molding state, the scope of application, the method of use, etc.; the implementation equipment and scope of application of the method, etc.).

[0050] The number of embodiments plays a decisive role in defining the scope of protection of the technical solution of the invention. If a wider scope of protection is desired, multiple embodiments should be provided (for example, if a component structure can take multiple forms, the names, structures, and replacement methods of multiple alternatives should be provided; if the component content is regional, at least the two endpoints of the region and the specific formula of the content at a certain point within the region and the other components should be provided).

[0051] The embodiment of the present invention taps porcelain lightly according to certain rules and collects signals; pre-processes the sound signals and extracts the phase characteristics of the signals; and uses a classifier to determine whether the porcelain is damaged.

[0052] A method for detecting porcelain damage based on sound, comprising the following steps:

[0053] Step 1: Create a sound database of intact and damaged porcelain

[0054] Prepare intact and damaged porcelain samples. Use a microphone to capture the vibration sound signals generated by tapping the porcelain sample surface and save them as an audio file in WAV format. The file name contains information about the damaged or intact category, indicating the lossless and damaged sound samples. Each audio file has a sampling frequency of 44100 Hz, a capture duration of 3 seconds, and a capture accuracy of 16 bits. The file is saved in WAV format.

[0055] Step 2: Preprocess the sound signal

[0056] Normalization processing: The amplitude of the speech signal varies greatly. Normalization processing can improve the robustness of signal processing and the stability of the algorithm. The maximum and minimum value normalization method is used to linearly map the amplitude of the sound signal to the range [0, 1].

[0057] Step 3: Extract the main phase features in the first stage. The specific steps of this step are:

[0058] Step 3-1: Divide the sound signal into frames. The frame length and frame shift of the first stage are represented by fl1 and fm1 respectively, in milliseconds.

[0059] Step 3-2: Perform a short-time discrete Fourier transform on each frame of the sound signal to transform the time domain information x into frequency domain information X:

[0060] X(m,k)=STFT[x(m,n)] (1)

[0061] In formula (1), x(m,n) refers to the nth value of the mth frame of the sound signal, and k is the frequency index;

[0062] Step 3-3, let the imaginary part and real part of each frame short-time Fourier transform result X(m,k) be represented as X r 、X i , the phase characteristics of each frame of Fourier transform are:

[0063] θ=arctan2(X r ,X i ) (2)

[0064] In formula (2), θ represents the phase characteristic;

[0065] Step 3-4: Define the width of the sliding window as wl1 points, and traverse the short-time Fourier transform phase image with a step size of ww1 points. During the traversal process, perform SVD decomposition on the phase feature θ of each sliding window of the phase image.

[0066] θ=USV T (3)

[0067] In formula (3), U and V are orthogonal matrices, the elements s1≥s2≥s3≥… on the diagonal of the diagonal matrix S are the singular values of the phase feature θ, and the maximum singular value s1 is taken as the main phase feature value of the sliding window. The main phase features of all sliding windows are sequentially spliced to generate the main phase feature of the first stage;

[0068] Step 4: Second stage main phase feature extraction

[0069] The second stage of main phase feature extraction takes the first stage main phase feature as input and performs the same processing on the input as the first stage. The difference from the first stage is the different values of the frame length and frame shift, sliding window step size and width parameters.

[0070] First, divide the frame. Here, the frame length and frame shift are fl2 and fm2 points respectively.

[0071] Secondly, perform short-time discrete Fourier transform on each frame.

[0072] Then find the Fourier transform phase characteristics,

[0073] Finally, the phase feature is decomposed by SVD using a sliding window with a width of wl2 points and a step size of wm2 points. All the maximum singular values of the SVD decomposition are sequentially spliced to form the second stage main phase feature.

[0074] Step 5: Sound signal phase characteristics

[0075] The main phase characteristics of the first stage and the main phase characteristics of the second stage are sequentially spliced to form the final phase characteristics of the sound signal;

[0076] Step 6: SVM abnormal sound classifier learning

[0077] The lossless and lossy sound samples collected in step 1 are set as positive samples and negative samples respectively. The SVM classifier is learned using the ten-fold cross-validation method. The parameters that need to be learned or adjusted during the learning process include the main phase feature parameters of the first stage: fl1, fm1, wl1, ww1, the main phase feature parameters of the second stage: fl2, fm2, wl2, ww2, and related parameters of the SVM classifier;

[0078] Step 7: For unknown damaged porcelain, first collect the sound signal generated by knocking, then normalize and preprocess it, and finally extract the phase features according to the optimal parameters determined in step 6 and use the SVM classifier to classify the sound signal. When the classification result is a positive sample, it is intact porcelain, otherwise it is damaged porcelain.

[0079] The audio file has a sampling frequency of 44100 Hz, a collection time of 3 seconds, and a collection accuracy of 16 bits.

[0080] This example collects percussion sound data from 110 porcelain vases of varying degrees of damage and 120 intact porcelain vases. The sounds of intact and damaged porcelain vases are designated as positive and negative samples, respectively. In the experiment, the main phase feature parameters for the first phase are fl1 = 10, fm1 = 5, wl1 = 32, and ww1 = 12; the main phase feature parameters for the second phase are fl2 = 512, fm2 = 100, wl2 = 48, and ww2 = 24. The SVM model uses the RBF kernel function K(x1, x2) = exp(-||x1-x2|| 2 / 2σ 2 ), σ = 10, penalty coefficient C = 62, relaxation factor ξ = 0.66.

[0081] Using a ten-fold cross-validation method, the experimental results of comparing the method of this embodiment with five existing classic sound feature extraction methods are shown in Table 1 below:

[0082] Table 1 Results of porcelain damage detection and classification experiments

[0083]

[0084] Table 1 shows that the feature extraction method of the present invention significantly improves accuracy and outperforms existing feature extraction methods in classification. The above experimental results demonstrate that the feature extraction method of the present invention can improve the representation of porcelain breakage sound information and significantly enhance porcelain breakage detection performance.

Claims

1. A method for detecting porcelain damage based on sound, characterized in that: The following steps are involved: Step 1: Create a sound database of intact and damaged porcelain Prepare intact and damaged porcelain samples, use a microphone to collect the vibration sound signals generated by knocking on the surface of the porcelain samples, and save them into an audio file in wav format. The file name contains damaged or lossless category information, which indicates the intact sound sample and the damaged sound sample; Step 2: Normalize and preprocess the sound signal The maximum and minimum normalization method is used to linearly transform the amplitude of the sound signal to between [0,1]; Step 3: Extract the main phase features in the first stage. The specific steps of this step are: Step 3-1: Divide the sound signal into frames. The frame length and frame shift of the first stage are represented by fl1 and fm1 respectively, in milliseconds. Step 3-2: Perform a short-time discrete Fourier transform on each frame of the sound signal to transform the time domain information x into frequency domain information X: X(m,k)=STFT[x(m,n)] (1) In formula (1), x(m,n) refers to the nth value of the mth frame of the sound signal, and k is the frequency index; Step 3-3, let the imaginary part and real part of each frame short-time Fourier transform result X(m,k) be represented as X r 、X i , the phase characteristics of each frame of Fourier transform are: θ=arctan2(X r ,X i ) (2) In formula (2), θ represents the phase characteristic; Step 3-4: Define the width of the sliding window as wl1 points, and traverse the short-time Fourier transform phase image with a step size of ww1 points. During the traversal process, perform SVD decomposition on the phase feature θ of each sliding window of the phase image. θ=USV T (3) In formula (3), U and V are orthogonal matrices, the elements s1≥s2≥s3≥… on the diagonal of the diagonal matrix S are the singular values of the phase feature θ, and the maximum singular value s1 is taken as the main phase feature value of the sliding window. The main phase features of all sliding windows are sequentially spliced to generate the main phase feature of the first stage; Step 4: Second stage main phase feature extraction The second stage of main phase feature extraction takes the first stage main phase feature as input and performs the same processing on the input as the first stage. The difference from the first stage is the different values of the frame length and frame shift, sliding window step size and width parameters. First, divide the frame. Here, the frame length and frame shift are fl2 and fm2 points respectively. Secondly, perform short-time discrete Fourier transform on each frame. Then find the Fourier transform phase characteristics, Finally, the phase feature is decomposed by SVD using a sliding window with a width of wl2 points and a step size of wm2 points. All the maximum singular values of the SVD decomposition are sequentially spliced to form the second stage main phase feature. Step 5: Sound signal phase characteristics The main phase characteristics of the first stage and the main phase characteristics of the second stage are sequentially spliced to form the final phase characteristics of the sound signal; Step 6: SVM abnormal sound classifier learning The lossless sound and lossy sound samples collected in step 1 are set as positive samples and negative samples respectively, and the SVM classifier is learned using the ten-fold cross-validation method. The parameters that need to be learned or adjusted during the learning process are the main phase feature parameters of the first stage: fl1, fm1, wl1, ww1, The main phase feature parameters of the second stage: fl2, fm2, wl2, ww2, and SVM classifier related parameters; Step 7: For unknown damaged porcelain, first collect the sound signal generated by knocking, then normalize and preprocess it, and finally extract the phase features according to the optimal parameters determined in step 6 and use the SVM classifier to classify the sound signal. When the classification result is a positive sample, it is intact porcelain, otherwise it is damaged porcelain.

2. The method for detecting porcelain damage based on sound according to claim 1, characterized in that: The audio file has a sampling frequency of 44100 Hz, a collection time of 3 seconds, and a collection accuracy of 16 bits.

Images