Method for identifying internal defects of plywood based on sound features

Abstract

The invention provides a method for identifying internal defects of plywood based on sound characteristics. The method comprises the following steps: step 1, acquiring sound of the plywood: knocking the surface of the plywood by utilizing an array cylinder so as to generate sound; step 2, extracting sound characteristics of the plywood: collecting a sound wave signal through a pickup device, and converting the sound wave signal into a digital signal; and 3, identifying the internal defects of the plywood, namely processing the digital signals by adopting Fourier transform and a deep neural network, and identifying the internal defect state and the distribution rule of the plywood. The problems that traditional manual grading is large in error, low in efficiency and high in cost are solved, automatic, standardized and large-scale plywood sorting is achieved, and wide market application prospects are achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of wood processing, and in particular relates to a method for identifying internal defects of plywood based on sound characteristics. Background Art

[0002] Plywood has the advantages of high wood utilization, high strength, and strong tolerance. It can be used in many fields such as aircraft, ships, trains, automobiles, home decoration, construction, and containers. In recent years, with the improvement of people's living standards and the development of customized homes and fully decorated houses, people's requirements for the quality of plywood products have increased. During the production process of plywood, due to factors such as uneven glue application, cracking of wood chips, and surface live knots, degumming, bubbling, and debonding may occur inside the plywood, which not only affects the mechanical properties of the plywood, but also affects the service life of the plywood. Therefore, in order to ensure the quality and performance of plywood products, it is of great significance to design a method that can accurately detect internal defects of plywood.

[0003] At present, the traditional method of detecting internal defects of plywood mainly relies on manually knocking on the plywood and listening to the sound produced by the vibration of the plywood to identify the internal defects and distribution patterns of the plywood. For example, if the sound heard when knocking is relatively dull, it is believed that there are cavities inside the plywood; if the sound is low and short, it is often that the internal structure of the plywood is not tight enough and there may be local looseness; when irregular and abnormal noises are heard when knocking, there may be cracks or foreign objects inside the plywood. However, this method has many serious shortcomings and deficiencies. In actual operation, due to the different knocking strength and the concentration of workers, the quality of plywood may be misjudged or missed.

[0004] As for the defect detection method of plywood, there are existing records of the method of using computer vision to identify surface defects of eucalyptus chips. This method can identify and sort defects such as pits and cracks on the surface of eucalyptus chips. However, computer vision technology focuses on the image acquisition and analysis of the appearance of plywood. It uses a high-definition camera to obtain the image of the board and uses image processing algorithms to extract features such as color, texture, and shape. This method can only identify defects on the surface of eucalyptus chips, such as knots, wormholes, and cracks. However, it is powerless to identify defects such as delamination, delamination, and bulging inside the plywood. Summary of the invention

[0005] In view of this, the object of the present invention is to provide a method for identifying internal defects of plywood based on sound characteristics, so as to comprehensively detect the defects existing inside the plywood and realize the grading of plywood of different qualities;

[0006] The technical solution of the present invention is: a method for identifying internal defects of plywood based on sound characteristics, comprising:

[0007] Step 1: Collect the sound of plywood: Use the array cylinder to knock on the surface of the plywood to generate sound;

[0008] Step 2: Extract the sound characteristics of plywood: collect the sound wave signal through the sound pickup device and convert the sound wave signal into a digital signal;

[0009] Step 3: Identify internal defects of plywood: Use Fourier transform and deep neural network to process digital signals to identify the defect status and distribution pattern inside the plywood.

[0010] Preferably, in step 1, the array cylinder is controlled by a sound-generating device, wherein the sound-generating device comprises: a PLC control box, an array cylinder, a hammer, a bracket, a sliding device, and a material rack;

[0011] The bracket is a double-layer structure, the sliding device is arranged on the upper layer of the bracket, and the material rack is arranged on the lower layer of the bracket; the cylinder is installed at the lower part of the sliding device; the hammer is installed at the end of the piston rod of the cylinder; the PLC control box is respectively connected with the sliding device and the array cylinder signal;

[0012] The sliding device includes a slide, a cylinder bracket, a slide rail, a synchronous belt, a slide motor, a motor bracket, a slide block, a coupling, a driving wheel and a passive wheel;

[0013] The slide is installed on both sides of the upper layer of the bracket, the slide rails are set on the slides on both sides, the slide block is installed on the slide rails, the driving wheel and the passive wheel are installed at both ends of the slide rails respectively, the synchronous belt is fixed to the slide block, and the two ends of the synchronous belt are respectively connected to the driving wheel and the passive wheel, one end of the coupling is fixed to the rotating shaft of the driving wheel, and the other end of the coupling is fixed to the rotating shaft of the slide motor; the two ends of the cylinder bracket are fixed on the slide blocks on both sides;

[0014] There are three groups of sliding blocks on the slide table in parallel, and 9 array cylinders are provided. Each group of sliding blocks drives three array cylinders to move.

[0015] Preferably, in step 2, the sound pickup device includes a piezoelectric ceramic sensor, an amplifier circuit, and an analog-to-digital conversion element; the piezoelectric ceramic sensor is installed under the material rack;

[0016] The piezoelectric ceramic sensor collects the sound wave signal generated by the plywood when it is knocked and converts it into an electrical signal. The electrical signal is amplified by the amplifier circuit, then converted into a digital signal through analog-to-digital conversion, and transmitted to the industrial control computer through the signal acquisition card.

[0017] Preferably, step 3 comprises:

[0018] Set the BP neural network structure, with 60 input layer nodes, 60 hidden layer nodes, and 1 output layer node; input the collected digital signal into the BP neural network, perform network training, and obtain the optimal weights of each input layer, hidden layer, and output layer of the neural network, i.e., the neural network model;

[0019] The fast Fourier transform algorithm is used to convert the digital signal from the time domain to the frequency domain. 20 frequency intervals of vibration frequency ranging from 1000Hz to 20000Hz are set to extract the amplitude accumulation, peak frequency and total amplitude average. The sound feature data of plywood of different qualities are obtained and labeled as "good" or "bad", and the data are organized into a training data set for the neural network.

[0020] Preferably, the amplitude accumulation, peak frequency and amplitude sum average values ​​of the 20 frequency intervals of the vibration frequency of 1000 Hz to 20000 Hz are respectively: the amplitude accumulation value is S=A1+A2+A3+...+An; the peak frequency is retrieved by retrieving the frequency domain sound signal obtained by the previous fast Fourier transform processing; the amplitude sum average value AVG=S / n, A is the amplitude of a certain sampling point, n is the total number of sampling points, and each plywood is marked with a grade label of "good" or "bad".

[0021] Preferably, the BP neural network consists of an input layer with 60 nodes, a hidden layer with 60 nodes, and an output layer with 1 node. The structure of the neural network is described by the following equation:

[0022] Z1=Input*W1+b1(m,1);

[0023] A1=σ(Z1);

[0024] Z2=A1*W2+b2(m,1);

[0025] Output = σ(Z2);

[0026] In the formula: σ() is the activation function, defined as 1 / (1+1 / (e to the power of x)); Input is the input matrix, that is, the characteristic data of the average value of the accumulated amplitude values, peak frequency, and sum of amplitudes in 20 intervals of vibration frequency from 1000Hz to 20000Hz; Output is the output value, 1 for qualified boards and 0 for unqualified boards; W1 and W2 are weight matrices, which are composed of a two-dimensional array of 60 rows and 60 columns, and each array element uses a random value between [0,1] generated by a random function; b1 and b2 are bias matrices, which are composed of a two-dimensional array of 60 rows and 1 column, and each array element value of the bias matrix is ​​a random value between [0,1] generated by a random function.

[0027] Preferably, the training of the neural network includes: forward propagation and back propagation of the neural network to correct the weights and bias values ​​of each hidden layer and fully connected layer of the network; the forward propagation of the neural network uses the description equation of the neural network structure, inputs a training data (Input1), and calculates the predicted value (Output_prediction) under the training data;

[0028] Z1=Input1*W1+b1(m,1);

[0029] A1=σ(Z1);

[0030] Z2=A1*W2+b2(m,1);

[0031] Output_prediction = σ(Z2);

[0032] Back propagation of the neural network, calculate the difference between Output_prediction and the true value, Output1 is the true value of the training data Input1;

[0033] dZ2 = Output_prediction - Output1

[0034] The matrix dZ2 represents the error of the output layer, and each element of it represents the difference between the predicted value of the output unit and the true label;

[0035] Calculate the gradient dW2 of the output layer weight W2:

[0036] dW2=A1 T *dZ2

[0037] A1 T is the transposed matrix of the hidden layer activation value A1. Through this calculation, the gradient of the weight W2 on the output layer error can be obtained; the gradient db2 of the output layer bias b2 is obtained by summing the error matrix dZ2 in the column direction:

[0038] db2 = dZ2*colwise().sum()

[0039] In the formula, colwise().sum() is an operation to sum the columns of the matrix;

[0040] After calculating the error of the output layer, continue to propagate the error backward to the hidden layer, calculate the contribution of the hidden layer activation value A1 to the output layer error, and obtain dA1:

[0041] dA1=dZ2*W2 T

[0042] By multiplying the weight matrix W2 T , propagate the output layer error “backward” to the hidden layer;

[0043] The derivative of the Sigmoid activation function is used to correct the hidden layer error and obtain the hidden layer error:

[0044] dZ1=dA1*σ′(A1)

[0045] Among them, σ′(A1) represents the derivative of the Sigmoid activation function with respect to A1;

[0046] Calculate the gradient dW1 of the hidden layer weight W1:

[0047] dW1=X T *dZ1

[0048] X T is the transposed matrix of the input matrix X. Through this calculation, the gradient of the hidden layer weight on the error can be obtained;

[0049] The gradient db1 of the hidden layer bias b1 is obtained by summing dZ1 in the column direction:

[0050] db1 = dZ1 * colwise().sum()

[0051] After calculating all the gradients, the network weights and biases are updated using the gradient descent algorithm. The update method is as follows:

[0052] W1=W1-learning_rate*dW1

[0053] b1 = b1 - learning_rate * db1

[0054] W2=W2-learning_rate*dW2

[0055] b2 = b2 - learning_rate * db2

[0056] In the formula, learning_rate is the learning rate, and the value of the learning rate is between 0.0001 and 0.1. The iteration of input training data and output results makes the network weights and biases adjust in the direction of reducing errors. Repeat the above process until the predetermined number of training rounds is reached or the error is less than the set threshold. The number of training times is greater than 100,000 times, and the prediction error is less than 0.001.

[0057] Preferably, the method of converting the signal from the time domain to the frequency domain using the fast Fourier transform algorithm is: converting each sound signal of the sound collector from the time domain to the frequency domain using the fast Fourier transform algorithm, that is, the horizontal axis is time and the vertical axis is vibration frequency;

[0058] First, the number of sampling points of the sound signal is N, and the sound signal is represented as a time domain sequence x(n), where n ranges from 0 to N-1; first, the time domain sequence x(n) is divided into two groups according to odd and even numbers, and x(2r) is set as an even number sequence, and x(2r+1) is set as an odd number sequence, r = 0, 1, ..., N / 2-1;

[0059] Perform discrete Fourier transform calculations on these two groups respectively, assuming that X1(k) is the DFT of the even sequence and X2(k) is the DFT of the odd sequence; for the DFT calculation of each group, continue to subdivide according to odd and even numbers, and repeat this subdivision process until the smallest subdivision is obtained;

[0060] The results of the subdivision calculations are gradually combined through the butterfly operation, the core operation unit in the fast Fourier transform algorithm. n k is the rotation factor, where N is the total number of points, k is related to the current operation level and number of points. Let X(k) be the final frequency domain signal, X1(k) and X2(k) be the two sets of subdivided frequency domain signals, then the butterfly operation can be expressed as X(k)=X1(k)+W n k X2(k),X(k+N / 2)=X1(k)-W n k X2(k).

[0061] The present invention provides a method for identifying internal defects of plywood based on sound features. The method uses an array cylinder to knock on the plywood to make it sound, and a sound pickup device collects sound wave signals, and uses Fourier transform and deep neural network to identify the defect state and distribution law. The sound collection device is composed of a sound generation, sound pickup and control device, which can comprehensively detect internal defects of plywood. The sound feature extraction adopts a fast Fourier transform algorithm to analyze the signal amplitude and peak frequency within a specific frequency range. The BP neural network is trained by a large amount of data to construct a mapping relationship between the vibration sound wave characteristics of the plywood and the quality of the plywood, which has the characteristics of high recognition accuracy and fast recognition speed.

[0062] The present invention solves the problems of large error, low efficiency and high cost in traditional manual grading, and makes plywood sorting automated, standardized and large-scale, which has broad market application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 A schematic diagram of the overall structure of the sound collection component provided by the present invention;

[0066] Figure 2 A schematic diagram of the local structure of the sound-generating device provided by the present invention;

[0067] Figure 3 This is a schematic diagram of the overall structure of the sound-generating device provided by the present invention. DETAILED DESCRIPTION

[0068] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present invention. Instead, they are merely examples of systems consistent with some aspects of the present invention as detailed in the appended claims.

[0069] Through in-depth research on plywood, it is found that there are significant differences in the natural frequencies of plywood with different bonding properties. The natural frequencies of plywood with good internal bonding are higher, while the natural frequencies of plywood with internal defects such as debonding, bulging, and delamination are relatively low. The natural frequency of a material directly affects the response behavior of the material to external vibrations. For plywood with good bonding, the content of high-frequency components in the spectrum generated during its vibration is higher; while the content of high-frequency components in the spectrum generated during the vibration of plywood with poor bonding is lower. Based on this, a method for identifying internal defects of plywood based on characteristic sound of plywood is designed. The plywood is knocked by an array cylinder to make it produce sound. The sound wave signal is collected by a pickup device. Fourier transform and deep neural network are used to identify the internal defect state and distribution law of plywood.

[0070] A method for identifying internal defects of plywood based on sound features, using a hammer 9 installed on an array cylinder 23 to strike the plywood 8 to cause it to vibrate and generate sound, collecting sound wave signals through a sound pickup device, identifying the internal defects and distribution patterns of the plywood through Fourier transform and deep neural network analysis, and realizing automatic and efficient identification of cracked, hollow and bulging plywood. The automatic grading of plywood is realized specifically through the following three steps: (1) plywood sound collection, (2) plywood sound feature extraction and (3) efficient identification of plywood internal defects. The method solves the problems of large errors, low efficiency and high cost of traditional manual grading, and makes plywood sorting automated, standardized and large-scale, with broad market application prospects.

[0071] The sound collection is composed of a sound generating device, a sound pickup device and a control device. The sound generating device is composed of an array cylinder 23, a hammer 9, a bracket 6, a sliding device and a material rack 6; the bracket 21 is a double-layer structure, the sliding device is arranged on the upper layer of the bracket 21, and the material rack 6 is arranged on the lower layer of the bracket 21;

[0072] Furthermore, the sliding device includes a slide 2, a slide rail 14, a synchronous belt 13, a slide motor 17, a motor bracket 19, a slide block 12, a coupling 18, a driving wheel 11 and a driven wheel 15; both ends of the cylinder bracket 3 are fixed on the slide block 12, and the cylinder 23 is installed below the cylinder bracket 3; the hammer 9 is installed at the connection of the piston rod end of the cylinder. The slide block is provided with three groups, and each group drives three array cylinders to move.

[0073] The sound pickup device adopts a piezoelectric ceramic sensor 7. The electrical signal collected by the piezoelectric ceramic sensor 7 is amplified by an amplifier circuit, and the analog-to-digital conversion is performed to convert the analog-to-digital signal into a digital signal, which is transmitted to the industrial control computer 5 through a signal acquisition card.

[0074] The control device is mainly a PLC control box 1. The box contains a CPU, memory, and input and output interfaces. The CPU executes the preset program, and the memory stores relevant data. The input interface receives instructions and signals, and the output interface controls the slide motor, cylinder, etc. Parameters such as the slide speed and cylinder knocking frequency can be flexibly set, and the operating status can be monitored, and an alarm and protection can be given in time if an abnormality occurs.

[0075] The sound-generating device, the sound-collecting device and the control device are integrated into one device. The cylinder is controlled to push the hammer to knock on the plywood to make a sound. The slide rail of the slide table can transfer the cylinder to different positions to knock on the plywood, comprehensively detect the defects inside the plywood, and realize the grading of plywood of different qualities.

[0076] Specifically, the plywood sound collection includes the following steps:

[0077] Place the plywood in a stable environment without obvious external interference, and ensure that there is no strong airflow, vibration source and excessive noise around, so as to create good conditions for the accurate collection of acoustic wave signals. Use the 9 cylinders and hammers arrayed above the plywood to knock the plywood in a set order and force. Start the highly sensitive piezoelectric ceramic sensor close to the bottom of the plywood to collect the acoustic wave signal generated by the plywood when it is knocked. Piezoelectric ceramics will generate a charge signal when subjected to mechanical stress from sound vibration. This signal is usually weak and needs to be amplified by an amplifier circuit before subsequent processing. We connect the two electrodes of the piezoelectric ceramic to the circuit respectively and select a suitable operational amplifier chip, such as LM358. It has two independent amplifiers, and one of them can be selected for single-ended amplification or two amplifiers can be used for differential amplification as needed. Connect the inverting input of the operational amplifier to the output through a resistor to form a voltage follower or amplifier. The gain of the amplifier can be set by adjusting the resistance value of the feedback resistor. The gain calculation formula is: Gain=1+Rf / Rin, where Rf is the feedback resistor and Rin is the input resistor. The signal amplifier uses a professional electronic amplification circuit, which can amplify weak sound wave signals. The collected signal is first transmitted to the signal amplifier, which amplifies the weak sound wave signal to enhance the strength and stability of the signal. The collected signal is transmitted to the analog-to-digital converter. The analog-to-digital converter converts the analog signal into a digital signal, which includes two steps: sampling and quantization. Sampling is to take the value of the analog signal within a certain time interval to obtain a series of discrete sample points. Quantization is to convert the amplitude of each sample point into a digital code, usually in binary coding. Through analog-to-digital conversion, digital sound signals of 9 areas of plywood are obtained. These digital signals can be transmitted to the computer through data transmission lines (such as serial communication lines or Ethernet, etc.). On the computer side, signal acquisition software is installed. The software can receive and process digital signals from the analog-to-digital converter, and display the waveform, spectrum and other information of the signal in real time in an intuitive way. Technicians can clearly observe the changes and characteristics of the signal through the software interface, and further analyze and process the signal, calculate the characteristic parameters of the signal, perform pattern recognition, etc., so as to gain a deep understanding of the acoustic characteristics and quality of the plywood.

[0078] The sound feature extraction of plywood includes the following steps:

[0079] Each sound wave signal obtained by the sound collector is a continuous curve, with time as the horizontal axis and amplitude as the vertical axis. The fast Fourier transform (FFT) algorithm is used to convert each sound signal of the sound collector from the time domain to the frequency domain, that is, the horizontal axis is time and the vertical axis is the vibration frequency. First, the number of sampling points of the sound signal is N, and the sound signal is represented as a time domain sequence x(n), where n ranges from 0 to N-1. Taking the FFT algorithm as an example, considering the characteristics of the plywood sound signal, the time domain sequence x(n) is first divided into two groups according to odd and even numbers, and x(2r) is set as an even sequence and x(2r+1) is set as an odd sequence, r=0,1,...,N / 2-1. Discrete Fourier transform (DFT) calculations are performed on these two groups respectively, and X1(k) is set as the DFT of the even sequence and X2(k) is set as the DFT of the odd sequence. For each group of DFT calculations, continue to subdivide according to odd and even numbers, and repeat this subdivision process until the smallest subsequence is subdivided. Because the sound of plywood may have complex superposition and attenuation during propagation and reflection, this subdivision process helps to analyze the signal more accurately. The results of the subdivision calculations are gradually combined through the butterfly operation, the core computing unit in the fast Fourier transform (FFT) algorithm. Let W n k is the rotation factor, where N is the total number of points, k is related to the current operation level and number of points. Let X(k) be the final frequency domain signal, X1(k) and X2(k) be the two sets of subdivided frequency domain signals, then the butterfly operation can be expressed as X(k)=X1(k)+W n k X2(k),X(k+N / 2)=X1(k)-W n k In this process, the influence of the plywood material on sound propagation and the signal changes caused by the difference in the hammer's striking position and force are fully considered, and finally the complete frequency domain signal X(k) is obtained.

[0080] For the frequency domain signals obtained by converting each sound signal, the frequency range is set to 1000 Hz to 20000 Hz according to the acoustic characteristics of plywood, so the first interval is 1000-1950 Hz, the second interval is 1950-2900 Hz, and so on, until the 20th interval is 19050-20000 Hz.

[0081] For the first interval 1000-1950Hz, traverse the corresponding frequency points in the frequency domain signal X(k). Assuming the sampling frequency is Fs, calculate the actual frequency f of each frequency point according to the formula f=k*Fs / N. Perform the same operation for the other 19 intervals.

[0082] At the same time, the peak frequency is determined in each interval, which reflects the frequency at which the plywood vibrates or responds most strongly in this specific frequency interval when struck by a hammer. The peak frequency k_peak can be found by comparing the amplitudes of all frequency points in the interval and finding the frequency point with the largest amplitude, k_peak = argmax(|X(k)|), that is, k_peak is the k value that maximizes |X(k)|, which is the peak frequency of the interval.

[0083] After analyzing the sound signals of 9 areas of a piece of plywood, the 9 sound signal features are then comprehensively analyzed. The average value of the sum of the amplitudes of the 9 sound signals in each interval is calculated. The specific method is that for the first interval 1000-1950Hz, the sum of the amplitudes of the 9 signals in this interval is added, and then divided by 9 to obtain the average amplitude of this interval. The same calculation is performed for the other 19 intervals to obtain more representative intensity characteristics. The average value of the sum of amplitudes reflects the average intensity level of the sound signal of the plywood in this frequency interval. By comparing the average amplitudes in different intervals, the distribution pattern of the plywood at different frequencies can be obtained, which can then be used to analyze its structural characteristics, quality status, etc.

[0084] The strength of plywood is related to the tightness of the material and the stability of the structure. If the average strength is low, it may indicate that the material is loose or there are inconsistencies in the internal structure. The distribution pattern of peak frequencies can provide information about the local characteristics of the plywood. If multiple peak frequencies are clustered in certain intervals, it may indicate that the plywood has a specific resonant structure or changes in material properties in the corresponding area. A more dispersed peak frequency distribution may mean that the material and structure of the plywood are relatively uniform in different locations. Through such detailed and comprehensive analysis and processing, it is possible to have a more accurate and in-depth understanding of the acoustic characteristics of plywood, and use this information for effective quality assessment and classification.

[0085] Specifically, the method for identifying internal defects of plywood includes the following steps:

[0086] According to the sound feature recognition and extraction method of plywood, a large number of sound feature data of plywood with different qualities and defects are obtained, including feature data of the cumulative amplitude value, peak frequency, and average value of the sum of amplitudes in each interval, where the amplitude cumulative value of the sound signal in a certain interval is S=A1+A2+A3+...+An, the peak frequency can be obtained by searching the frequency domain sound signal obtained by the previous fast Fourier transform, the average value of the sum of amplitudes AVG=S / n, A is the amplitude of a certain sampling point, n is the total number of sampling points, and each piece of plywood is marked with a grade label of "good" or "bad". These data are organized into a training data set. The features to be analyzed by this method are the feature data of the average value of the cumulative amplitude value, peak frequency, and sum of amplitudes, and the analyzed interval is 20 intervals of vibration frequency from 1000Hz to 20000Hz, so the number of input layer nodes is 60; the classification of plywood is a binary classification problem of "good" and "bad", so the number of output layer nodes is set to 1.

[0087] The neural network constructed by the present invention is composed of an input layer with 60 nodes, a hidden layer with 60 nodes, and an output layer with 1 node. The structure of the neural network can be described by the following equation.

[0088] Z1=Input*W1+b1(m,1);

[0089] A1=σ(Z1);

[0090] Z2=A1*W2+b2(m,1);

[0091] Output = σ(Z2);

[0092] In the above formula: σ() is the activation function, which is defined as 1 / (1+1 / (e to the power of x)); Input is the input matrix, that is, the characteristic data of the average value of the accumulated amplitude values, peak frequency, and sum of amplitudes in 20 intervals of vibration frequency from 1000Hz to 20000Hz; Output is the output value, 1 for qualified boards and 0 for unqualified boards; W1 and W2 are weight matrices, which are composed of a two-dimensional array of 60 rows and 60 columns, and each array element uses a random value between [0,1] generated by a random function; b1 and b2 are bias matrices, which are composed of a two-dimensional array of 60 rows and 1 column, and each array element value of the bias matrix is ​​a random value between [0,1] generated by a random function.

[0093] The training of neural networks includes two parts: forward propagation and back propagation of the neural network, and back propagation corrects the weights and bias values ​​of each hidden layer and fully connected layer of the neural network.

[0094] 1. Create a training dataset

[0095] The above-mentioned array cylinder drives a hammer to strike the surface of the plywood, causing it to vibrate and produce sound. The sound wave signal is collected by a sound pickup device, and the time-frequency curve of the waveform is obtained by Fourier transform. In 20 intervals of vibration frequency from 1000Hz to 20000Hz, the characteristic data of the amplitude value accumulation, peak frequency, and average value of the total amplitude of the curve are analyzed, and the plywood is manually graded into two categories of "good" and "bad". Each curve can obtain 60 input data and 1 output result. The sample capacity of the neural network training set must be greater than 1000 sets of training data.

[0096] 2. Forward propagation and reverse correction of neural networks

[0097] Using the descriptive equation of the above neural network structure, input a training data (Input1) and calculate the predicted value (Output_prediction) under the training data.

[0098] Z1=Input1*W1+b1(m,1);

[0099] A1=σ(Z1);

[0100] Z2=A1*W2+b2(m,1);

[0101] Output_prediction = σ(Z2);

[0102] Calculate the difference between Output_prediction and the true value, where Output1 is the true value of the training data Input1.

[0103] dZ2 = Output_prediction - Output1

[0104] The matrix dZ2 represents the error of the output layer, and each element of it represents the difference between the predicted value of the output unit and the true label.

[0105] Calculate the gradient dW2 of the output layer weight W2:

[0106] dW2=A1 T *dZ2

[0107] A1 T is the transposed matrix of the hidden layer activation value A1. This calculation can be used to obtain the gradient of the weight W2 on the output layer error. The gradient db2 of the output layer bias b2 is obtained by summing the error matrix dZ2 in the column direction:

[0108] db2 = dZ2*colwise().sum()

[0109] In the formula, colwise().sum() is an operation to sum the columns of the matrix.

[0110] After calculating the error of the output layer, continue to propagate the error backward to the hidden layer. Calculate the contribution of the hidden layer activation value A1 to the output layer error and get dA1:

[0111] dA1=dZ2*W2 T

[0112] By multiplying the weight matrix W2 T , propagating the output layer error “backward” to the hidden layer.

[0113] The hidden layer error is corrected using the derivative of the Sigmoid activation function to obtain the hidden layer error dZ1:

[0114] dZ1=dA1*σ′(A1)

[0115] Among them, σ′(A1) represents the derivative of the Sigmoid activation function with respect to A1.

[0116] Calculate the gradient dW1 of the hidden layer weight W1:

[0117] dW1=X T *dZ1

[0118] X T is the transposed matrix of the input matrix X. Through this calculation, the gradient of the hidden layer weight on the error can be obtained.

[0119] The gradient db1 of the hidden layer bias b1 is obtained by summing dZ1 in the column direction:

[0120] db1 = dZ1 * colwise().sum()

[0121] After calculating all the gradients, the gradient descent algorithm is used to update the weights and biases of the network. The update method is as follows:

[0122] W1=W1-learning_rate*dW1

[0123] b1 = b1 - learning_rate * db1

[0124] W2=W2-learning_rate*dW2

[0125] b2 = b2 - learning_rate * db2

[0126] In the formula, learning_rate is the learning rate, and the value of the learning rate is between 0.0001 and 0.1. The iteration of inputting training data and outputting results makes the network weights and biases be adjusted in the direction of reducing the error.

[0127] 3. Repeat the above process until the predetermined number of training rounds is reached or the error is less than the set threshold. The number of training rounds is greater than 100,000 times and the prediction error is less than 0.001.

[0128] 4. Model testing and application: The above-mentioned array cylinder drives a hammer to knock on the surface of the plywood to make it vibrate and produce sound. The sound wave signal is collected by the sound pickup device, and the time-frequency curve of the waveform is obtained by Fourier transform. In the 20 intervals of vibration frequency from 1000Hz to 20000Hz, the characteristic data of the amplitude value accumulation, peak frequency, and average value of the sum of amplitudes of the curve are analyzed. The collected characteristic data is used as input data and input into the trained neural network model, and the grade of the plywood is obtained through the above-mentioned matrix operation.

[0129] Example 1

[0130] A method for identifying internal defects of plywood based on sound characteristics, the system includes a sound generating device, a sound pickup device, a control device and corresponding data processing and analysis modules. The sound generating device is composed of an array cylinder 23, a hammer 9, a bracket 21, a sliding device, and a material rack 6; the sound pickup device is composed of a piezoelectric ceramic sensor 7, an amplifier circuit, and an analog-to-digital conversion element; and the control device is mainly a PLC control box 1.

[0131] The sliding device includes a slide 2, a cylinder bracket 3, a slide rail 14, a synchronous belt 13, a slide motor 17, a motor bracket 19, a slide block 12, a coupling 18, a driving wheel 11 and a passive wheel 15; the bracket 21 is a double-layer structure, the sliding device is arranged on the upper layer of the bracket 21, and the material rack 6 is arranged on the lower layer of the bracket. The slide 2 is installed on the upper layer of the bracket 21, the slide rail 14 is arranged on the slide, the slide block 12 is installed on the slide rail, the driving wheel 11 and the passive wheel 15 are respectively installed at both ends of the slide rail 14, the synchronous belt 13 is fixed to the slide block 12, and the two ends of the synchronous belt 13 are respectively connected to the driving wheel 11 and the passive wheel 15, one end of the coupling 18 is fixed to the rotating shaft of the driving wheel 11, and the other end of the coupling 18 is fixed to the rotating shaft of the slide motor 17; there are 9 array cylinders.

[0132] 1. When collecting sound, place the plywood in a stable environment without obvious external interference to ensure that the collected sound signal is accurate and reliable. The PLC control box 1 precisely controls the sliding device. The sliding device is set on the upper layer of the bracket 21. When the slide motor 17 is started, the coupling 18 drives the driving wheel 11 to rotate, and then drives the synchronous belt 13 to move, so that the slide block 12 moves on the slide rail 14, thereby realizing the position adjustment of the components installed on the slide block 12.

[0133] Under the command of the PLC control box 1, the sliding device moves 9 striking units consisting of array cylinders 23 and hammers 9 to a predetermined position. The bracket 21 is a double-layer structure, and the material rack 6 is set at the lower layer of the bracket for placing plywood 8. The control box 1 issues a command to make the cylinders 23 strike the plywood 8 in a set order.

[0134] When the cylinder 23 hits the plywood 8, the piezoelectric ceramic sensor 7 collects the sound wave signal generated when the plywood 8 is hit. The collected signal is transmitted to the amplifier circuit for amplification to enhance the strength and stability of the signal. The amplified signal is transmitted to the analog-to-digital converter to obtain the sound signals of 9 different areas on the plywood. The fast Fourier transform algorithm is used to convert each pre-processed sound signal from the time domain to the frequency domain. According to the acoustic characteristics of the plywood, 20 frequency intervals are set, and the characteristic data of the average value of the amplitude value accumulation, peak frequency, and sum of amplitudes of the frequency points in each interval are obtained. And through manual identification, each piece of plywood is marked with a grade of "good" or "bad".

[0135] Determine the structure of the BP neural network, including the input layer, hidden layer and output layer. The number of nodes in the input layer is 60, the number of nodes in the hidden layer is 60, and the number of nodes in the output layer is 1.

[0136] Conduct network training: Initialize the weights and biases of the neural network. Input the training set data into the neural network and perform forward propagation to calculate the output. Based on the error between the output and the true label, use the backpropagation algorithm to calculate the gradient and update the weights and biases. Repeat the above process until the predetermined number of training rounds is reached or the error meets the requirements.

[0137] The application of neural network first uses an array cylinder to drive a hammer to knock on the surface of the plywood, causing it to vibrate and produce sound. The sound wave signal is collected by a sound pickup device, and the time-frequency curve of the waveform is obtained by Fourier transform. In the 20 intervals of vibration frequency from 1000Hz to 20000Hz, the characteristic data of the curve's amplitude value accumulation, peak frequency, and average value of the total amplitude are analyzed. The collected characteristic data is used as input data and input into the trained neural network model, and the grade of the plywood is obtained through the above-mentioned matrix operation.

[0138] Through the above-mentioned specific implementation steps, the current problem of difficult identification and sorting of internal defects in plywood can be effectively solved, the automation, standardization and scale of plywood sorting can be realized, and the production efficiency and product quality can be significantly improved.

[0139] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These changes and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying internal defects of plywood based on sound characteristics, characterized in that: include: Step 1: Collect the sound of plywood: Use the array cylinder to knock on the surface of the plywood to generate sound; Step 2: Extract the sound characteristics of plywood: collect the sound wave signal through the sound pickup device and convert the sound wave signal into a digital signal; Step 3: Identify internal defects of plywood: Use Fourier transform and deep neural network to process digital signals to identify the defect status and distribution pattern inside the plywood.

2. The method for identifying internal defects of plywood based on sound characteristics according to claim 1, characterized in that: In step 1, the array cylinder is controlled by a sound generating device, wherein the sound generating device comprises: a PLC control box (1), an array cylinder (23), a hammer (9), a bracket (21), a sliding device, and a material rack (6); The support (21) is a double-layer structure, the sliding device is arranged on the upper layer of the support (21), and the material rack (6) is arranged on the lower layer of the support (21); the cylinder (23) is installed at the lower part of the sliding device; the hammer (9) is installed at the end of the piston rod of the cylinder; the PLC control box (1) is respectively connected to the sliding device and the array cylinder (23) by signal; The sliding device comprises a slide table (2), a cylinder bracket (3), a slide rail (14), a synchronous belt (13), a slide table motor (17), a motor bracket (19), a slide block (12), a coupling (18), a driving wheel (11) and a driven wheel (15); The slide (2) is installed on both sides of the upper layer of the bracket (21), the slide rails (14) are arranged on the slides (2) on both sides, the slide sliders (12) are installed on the slide rails, the driving wheel (11) and the passive wheel (15) are respectively installed on both ends of the slide rails (14), the synchronous belt (13) is fixed to the slide slider (12), and the two ends of the synchronous belt (13) are respectively connected to the driving wheel (11) and the passive wheel (15), one end of the coupling (18) is fixed to the rotating shaft of the driving wheel (11), and the other end of the coupling (18) is fixed to the rotating shaft of the slide motor (17); the two ends of the cylinder bracket (3) are fixed to the slide sliders (12) on both sides; The slide block (12) is provided in three groups in parallel, and the array cylinders are provided in nine groups. Each group of the slide block (12) drives three array cylinders to move.

3. The method for identifying internal defects of plywood based on sound characteristics according to claim 1, characterized in that: In step 2, the sound pickup device comprises a piezoelectric ceramic sensor (7), an amplifying circuit, and an analog-to-digital conversion element; the piezoelectric ceramic sensor (7) is installed below the material rack (6); The piezoelectric ceramic sensor (7) collects the sound wave signal generated when the plywood is struck, and converts it into an electrical signal. The electrical signal is amplified by an amplifier circuit, and then converted into a digital signal through analog-to-digital conversion, and then transmitted to the industrial control computer (5) through a signal acquisition card.

4. The method for identifying internal defects of plywood based on sound characteristics according to claim 1, characterized in that: Step 3 includes: Set the BP neural network structure, with 60 input layer nodes, 60 hidden layer nodes, and 1 output layer node; input the collected digital signal into the BP neural network, perform network training, and obtain the optimal weights of each input layer, hidden layer, and output layer of the neural network, i.e., the neural network model; The fast Fourier transform algorithm is used to convert the digital signal from the time domain to the frequency domain. 20 frequency intervals of vibration frequency ranging from 1000Hz to 20000Hz are set to extract the amplitude accumulation, peak frequency and total amplitude average value. The sound feature data of plywood of different qualities are obtained and labeled as "good" or "bad", and then organized into data to obtain the training data set of the neural network.

5. The method for identifying internal defects of plywood based on sound characteristics according to claim 4, characterized in that: The amplitude accumulation, peak frequency and amplitude sum average values ​​of the 20 frequency intervals of the vibration frequency from 1000 Hz to 20000 Hz are respectively: the amplitude accumulation value is S=A1+A2+A3+...+An; the peak frequency is retrieved by retrieving the frequency domain sound signal obtained by the previous fast Fourier transform processing; the amplitude sum average value AVG=S / n, A is the amplitude of a certain sampling point, n is the total number of sampling points, and each plywood is marked with a grade label of "good" or "bad".

6. The method for identifying internal defects of plywood based on sound characteristics according to claim 4, characterized in that: The BP neural network consists of an input layer with 60 nodes, a hidden layer with 60 nodes, and an output layer with 1 node. The structure of the neural network is described by the following equation: Z1=Input*W1+b1(m,1); A1=σ(Z1); Z2=A1*W2+b2(m,1); Output = σ(Z2); In the formula: σ() is the activation function, defined as 1 / (1+1 / (e to the power of x)); Input is the input matrix, that is, the characteristic data of the average value of the accumulated amplitude values, peak frequency, and sum of amplitudes in 20 intervals of vibration frequency from 1000Hz to 20000Hz; Output is the output value, 1 for qualified boards and 0 for unqualified boards; W1 and W2 are weight matrices, which are composed of a two-dimensional array of 60 rows and 60 columns, and each array element uses a random value between [0,1] generated by a random function; b1 and b2 are bias matrices, which are composed of a two-dimensional array of 60 rows and 1 column, and each array element value of the bias matrix is ​​a random value between [0,1] generated by a random function.

7. The method for identifying internal defects of plywood based on sound characteristics as claimed in claim 5, characterized in that: The training of the neural network includes two parts: the forward propagation and back propagation of the neural network to correct the weights and bias values ​​of each hidden layer and fully connected layer of the network; the forward propagation of the neural network uses the description equation of the neural network structure, inputs a training data (Input1), and calculates the predicted value (Output_prediction) under the training data; Z1=Input1*W1+b1(m,1); A1=σ(Z1); Z2=A1*W2+b2(m,1); Output_prediction = σ(Z2); Back propagation of the neural network, calculate the difference between Output_prediction and the true value, Output1 is the true value of the training data Input1; dZ2 = Output_prediction - Output1 The matrix dZ2 represents the error of the output layer, and each element of it represents the difference between the predicted value of the output unit and the true label; Calculate the gradient dW2 of the output layer weight W2: dW2=A1 T *dZ2 A1 T is the transposed matrix of the hidden layer activation value A1. Through this calculation, the gradient of the weight W2 on the output layer error can be obtained; the gradient db2 of the output layer bias b2 is obtained by summing the error matrix dZ2 in the column direction: db2 = dZ2*colwise().sum() In the formula, colwise().sum() is an operation to sum the columns of the matrix; After calculating the error of the output layer, continue to propagate the error backward to the hidden layer, calculate the contribution of the hidden layer activation value A1 to the output layer error, and obtain dA1: dA1=dZ2*W2 T By multiplying the weight matrix W2 T , propagate the output layer error "backward" to the hidden layer; The derivative of the Sigmoid activation function is used to correct the hidden layer error and obtain the hidden layer error: dZ1=dA1*σ′(A1) Among them, σ′(A1) represents the derivative of the Sigmoid activation function with respect to A1; Calculate the gradient dW1 of the hidden layer weight W1: dW1=X T *dZ1 X T is the transposed matrix of the input matrix X. Through this calculation, the gradient of the hidden layer weight on the error can be obtained; The gradient db1 of the hidden layer bias b1 is obtained by summing dZ1 in the column direction: db1 = dZ1 * colwise().sum() After calculating all the gradients, the network weights and biases are updated using the gradient descent algorithm. The update method is as follows: W1=W1-learning_rate*dW1 b1 = b1 - learning_rate * db1 W2=W2-learning_rate*dW2 b2 = b2 - learning_rate * db2 In the formula, learning_rate is the learning rate, and the value of the learning rate is between 0.0001 and 0.

1. The iteration of input training data and output results makes the network weights and biases adjust in the direction of reducing errors. Repeat the above process until the predetermined number of training rounds is reached or the error is less than the set threshold. The number of training times is greater than 100,000 times, and the prediction error is less than 0.

001.

8. The method for identifying internal defects of plywood based on sound characteristics as claimed in claim 5, characterized in that: The method of converting the signal from the time domain to the frequency domain using the fast Fourier transform algorithm is as follows: using the fast Fourier transform algorithm to convert each sound signal of the sound collector from the time domain to the frequency domain, that is, the horizontal axis is time and the vertical axis is vibration frequency; First, the number of sampling points of the sound signal is N, and the sound signal is represented as a time domain sequence x(n), where n ranges from 0 to N-1; first, the time domain sequence x(n) is divided into two groups according to odd and even numbers, and x(2r) is set as an even number sequence, and x(2r+1) is set as an odd number sequence, r = 0, 1, ..., N / 2-1; Perform discrete Fourier transform calculations on these two groups respectively, assuming that X1(k) is the DFT of the even sequence and X2(k) is the DFT of the odd sequence; for the DFT calculation of each group, continue to subdivide according to odd and even numbers, and repeat this subdivision process until the smallest subdivision is obtained; The results of the subdivision calculations are gradually combined through the butterfly operation, the core operation unit in the fast Fourier transform algorithm. n k is the rotation factor, where N is the total number of points, k is related to the current operation level and number of points. Let X(k) be the final frequency domain signal, X1(k) and X2(k) be the two sets of subdivided frequency domain signals, then the butterfly operation can be expressed as X(k)=X1(k)+W n k X2(k),X(k+N / 2)=X1(k)-W n k X2(k).

Images