Magnetic shoe internal defect detection device and method based on acoustic emission signal

Abstract

The invention discloses a magnetic shoe internal defect detection device and method based on acoustic emission signals, and relates to the field of magnetic shoe internal defect detection.The detection device / method capable of being directly matched with a magnetic shoe production line is built, collected collision acoustic vibration signals are preprocessed and converted into digital signals, and then the digital signals are transmitted to the magnetic shoe production line. The method comprises the following steps of: decomposing into three component signals by using an improved clustering symplectic geometric mode decomposition method, selecting an optimal component signal as a feature signal through a minimum decomposition evaluation index, classifying the input feature signals by using an IITN network, and obtaining a feature signal classification result, namely, obtaining a defect detection result of the magnetic shoe. Therefore, a quick and effective magnetic shoe defect detection scheme which can be directly used for a magnetic shoe production line is formed.

Description

Technical Field

[0001] The present invention relates to the field of internal defect detection of magnetic tiles, and particularly to a device and method for detecting internal defects of magnetic tiles based on acoustic emission signals. Background Art

[0002] Magnetic tiles are key components widely used in electromagnetic devices such as motors and generators. They are usually made of permanent magnetic materials and can provide a stable magnetic field. The magnetic properties, mechanical strength, and integrity of the internal structure of magnetic tiles directly affect the efficiency, lifespan, and reliability of motors. However, the manufacturing process of magnetic tiles is highly complex. Factors such as die pressure control during pressing and sintering temperature management, as well as various other uncertain factors, may cause defects in magnetic tiles. Among the defects of magnetic tiles, internal defects account for more than 60%. Moreover, the positions, sizes, and shapes of the defects are diverse, which undoubtedly poses great challenges and difficulties to the detection of internal defects.

[0003] Currently, the detection of internal defects during the production process of magnetic tiles still largely relies on manual inspection. Inspectors analyze the sounds generated when magnetic tiles impact the metal surface to evaluate the presence of internal defects. This traditional method highly depends on the auditory perception and subjective judgment of inspectors, resulting in reduced operational efficiency, vague detection standards, and inconsistent accuracy. It is difficult for the detection accuracy and speed to meet the detection requirements. In view of the problems existing in the manual detection of internal defects of magnetic tiles, selecting an acoustic detection method with high speed and low cost as the detection means can meet the requirements of rapid and low-cost magnetic tile quality detection. However, the signal form for detecting internal defects of magnetic tiles is single, and the inherent nonlinearity and non-stationarity of the signal are not conducive to feature extraction and identification. Therefore, it is of great significance to develop a fast, efficient, and novel device and method for detecting internal defects of magnetic tiles. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, a device and method for detecting internal defects of magnetic tiles based on acoustic emission signals provided by the present invention solve the problems of low detection efficiency and large influence of human factors in the detection of internal defects of magnetic tiles in the magnetic tile production line.

[0005] To achieve the above invention purpose, the technical solution adopted by the present invention is as follows: Provide a device for detecting internal defects of magnetic tiles based on acoustic emission signals, which includes a defect detection component, an excitation inclined plane, and an acoustic emission sensor; the acoustic emission sensor is arranged on the excitation inclined plane; The acoustic emission sensor is used to capture the collision acoustic vibration signal generated when the magnetic tile falls on the excitation inclined plane; The defect detection component is used to receive the collision acoustic vibration signal captured by the acoustic emission sensor and detect the internal defects of the magnetic tile according to the collision acoustic vibration signal.

[0006] The beneficial effects of this device are as follows: The present invention detects the defects of magnetic tiles through the collision sound vibration signals generated by the magnetic tiles falling on the excitation inclined plane, which is convenient for application in the production line of magnetic tiles to realize the defect detection of magnetic tiles on the production line.

[0007] Furthermore, the defect detection component includes a signal amplifier, a signal acquisition card, a signal decomposition module, and a defect identification module, where: The signal amplifier is used to amplify and filter the collision sound vibration signals output by the acoustic emission sensor to obtain a preprocessed analog electrical signal; The signal acquisition card is used to convert the preprocessed analog electrical signal into a digital signal; The signal decomposition module is deployed with an improved clustering symplectic geometric mode decomposition method, which is used to decompose the digital signal into 3 component signals, and selects the component signal corresponding to the minimum decomposition evaluation index as the characteristic signal output by it; The defect identification module is deployed with a pre-trained IITN network, which is used to take the characteristic signal output by the signal decomposition module as the input of the pre-trained IITN network, classify the input characteristic signal through the IITN network, obtain the classification result of the characteristic signal, that is, obtain the defect detection result of the magnetic tile.

[0008] The beneficial effects of adopting the above further solution are as follows: After the collected collision sound vibration signals are preprocessed and converted into digital signals, they are decomposed into 3 component signals by the improved clustering symplectic geometric mode decomposition method, and the optimal (with instantaneous impact components) component signal is selected as the characteristic signal through the minimum decomposition evaluation index. The IITN network is used to classify the input characteristic signal to obtain the classification result of the characteristic signal, that is, obtain the defect detection result of the magnetic tile, thereby forming a set of fast and effective magnetic tile defect detection solutions.

[0009] Furthermore, the IITN network includes a first Inception module, a second Inception module, and a third Inception module connected in sequence, and a global average pooling layer, a flattening layer, and a fully connected layer connected in sequence; among them, the input of the first Inception module is processed by the CBAM attention mechanism and then connected to the output of the second Inception module in a residual connection, and the result of the residual connection is the input of the third Inception module; the output of the third Inception module is the input of the global average pooling layer; the input of the first Inception module is the input of the IITN network; the output of the fully connected layer is the output of the IITN network.

[0010] The beneficial effects of adopting the above further solution are as follows: The IITN network removes the redundant modules in the InceptionTime network architecture. A single InceptionTime module is composed of three Inception modules and a residual connection with the CBAM attention mechanism. Combining with the global average pooling layer, the flattening layer, and the fully connected layer, it can efficiently and quickly classify the characteristic signals corresponding to the magnetic tiles to obtain whether there are internal defects in the magnetic tiles.

[0011] Further, a first bracket and a second bracket are respectively arranged on both sides of the upper end of the excitation inclined plane; a photoelectric emitter is arranged on the first bracket, and a photoelectric receiver is arranged on the second bracket; the connection line between the photoelectric emitter and the photoelectric receiver intersects the falling path of the magnetic tile, so that the magnetic tile blocks the photoelectric receiver from receiving the signal from the photoelectric emitter before falling onto the excitation inclined plane, and the time point when the photoelectric receiver does not receive the signal from the photoelectric emitter is used as the signal acquisition starting point of the acoustic emission sensor.

[0012] Further, a first conveyor belt is arranged above the excitation inclined plane and beside the photoelectric receiver, and a second conveyor belt is arranged at the lower end of the excitation inclined plane; a third conveyor belt is arranged on one side of the second conveyor belt, and a pusher facing the third conveyor belt is arranged on the other side of the second conveyor belt; The first conveyor belt is used to convey the magnetic tile above the excitation inclined plane, so that the magnetic tile passes through the connection line between the photoelectric emitter and the photoelectric receiver and falls onto the excitation inclined plane; The second conveyor belt is used to convey the fallen magnetic tiles detected as defect-free towards the qualified magnetic tile discharge port; The pusher is used to push the fallen magnetic tiles detected as defective from the second conveyor belt onto the third conveyor belt; The third conveyor belt is used to convey the magnetic tiles detected as defective towards the waste recycling area.

[0013] Further, the magnetic tile is conveyed on the first conveyor belt in a standing posture, and the convex surface of the magnetic tile faces the conveying direction; the falling height of the magnetic tile is 20 mm.

[0014] The beneficial effects of adopting the above further solution are as follows: The present invention forms a signal for starting measurement through the photoelectric emitter and the photoelectric receiver, uses the first conveyor belt to make the magnetic tile fall on the excitation inclined plane, and when the magnetic tile slides to the bottom of the excitation inclined plane, it is conveyed by the second conveyor belt. During the conveying process, whether to start the pusher to classify the magnetic tile is determined according to the defect detection result, and the magnetic tile is conveyed to different links according to the classification result. The entire device can be directly arranged in the production line of the magnetic tile, and the defect detection and automatic classification and conveying of the magnetic tile can be completed without manual intervention, greatly improving the production efficiency.

[0015] Provided is a method for detecting internal defects of magnetic tiles based on acoustic emission signals, which includes the following steps: Let the magnetic tile fall on the excitation inclined plane, and capture the collision acoustic vibration signal generated when the magnetic tile falls on the excitation inclined plane through an acoustic emission sensor; Receive the collision acoustic vibration signal captured by the acoustic emission sensor through the defect detection component, and detect the internal defects of the magnetic tile according to the collision acoustic vibration signal.

[0016] Further, the specific method of receiving the collision acoustic vibration signal captured by the acoustic emission sensor through the defect detection component and detecting the internal defects of the magnetic tile according to the collision acoustic vibration signal includes the following steps: Use a signal amplifier to amplify and filter the collision acoustic vibration signal output by the acoustic emission sensor to obtain a preprocessed analog electrical signal; Use a signal acquisition card to convert the preprocessed analog electrical signal into a digital signal; Use a signal decomposition module deployed with an improved clustering symplectic geometric mode decomposition method to decompose the digital signal into 3 component signals, and select the component signal corresponding to the minimum decomposition evaluation index as the characteristic signal output; Use a defect recognition module deployed with a pre-trained IITN network to use the characteristic signal output by the signal decomposition module as the input of the pre-trained IITN network, classify the input characteristic signal through the IITN network, obtain the characteristic signal classification result, that is, obtain the defect detection result of the magnetic tile.

[0017] Further, the specific method of decomposing the digital signal into 3 component signals includes the following steps: Construct a one-dimensional time series using the time series topological equivalence method for the digital signal and form a trajectory matrix :

[0018] where represents the embedding dimension of the time series topological equivalence method; represents the delay time of the time series topological equivalence method, , represents the data length of the digital signal; represents at the delay time value; represents the trajectory matrix in the row and first column element; Perform autocorrelation analysis on the trajectory matrix to obtain the covariance matrix , and then obtain the matrix ; Obtain the matrix Eigenvectors of eigenvalues , using the eigenvectors and the trajectory matrix to construct the coefficient matrix , and then obtain the initial single-component reconstruction matrix ; where ; Adjust the elements in each initial single-component reconstruction matrix to obtain the adjusted initial single-component reconstruction matrix: For the element in the th row and the m -1th column of the initial single-component reconstruction matrix d , if , let ; for other elements in the initial single-component reconstruction matrix , let ; where is the element in the th row and the a th column of the initial single-component reconstruction matrix b ; is the element in the th row and the b th column of the initial single-component reconstruction matrix a ; Convert the th adjusted initial single-component reconstruction matrix into a one-dimensional time series , and then obtain the one-dimensional time series ; where the expression of the k th element in

[0019] where is the element in the th row and the th column of the adjusted initial single-component reconstruction matrix ; , min means taking the minimum value; , max means taking the maximum value; Recombine the one-dimensional time series into 3 component signals by clustering through the agglomerative hierarchical clustering algorithm.

[0020] Furthermore, the calculation expression of the minimum decomposition evaluation index is:

[0021] where represents the single-component signal and the digital signal The minimum decomposition evaluation index; Represents time; Represents the data length of the digital signal, that is, the number of data points in a single component signal.

[0022] The beneficial effects of this method are as follows: After preprocessing the collected collision acoustic vibration signals and converting them into digital signals, they are decomposed into 3 component signals by the improved clustering symplectic geometric mode decomposition method, and the optimal (with instantaneous impact components) component signal is selected as the feature signal through the minimum decomposition evaluation index. The IITN network is used to classify the input feature signal to obtain the classification result of the feature signal, that is, the defect detection result of the magnetic tile, thus forming a set of fast and effective magnetic tile defect detection solutions. Description of the drawings

[0023] Figure 1 Is the structural schematic diagram of this device; Figure 2 Are the time-domain and frequency-domain waveform diagrams of the acoustic emission signals of qualified and defective magnetic tiles; where (a) is the time-domain diagram of the acoustic emission signal of the qualified magnetic tile, (b) is the frequency-domain diagram of the acoustic emission signal of the qualified magnetic tile, (c) is the time-domain diagram of the acoustic emission signal of the defective magnetic tile, and (d) is the frequency-domain diagram of the acoustic emission signal of the defective magnetic tile; Figure 3 Is the decomposition effect of the clustering symplectic geometric components on a random sample; Figure 4 Is the structural schematic diagram of the IITN network; Figure 5 Is the comparison schematic diagram of the loss function convergence curve and the accuracy convergence curve of the original InceptionTime network and the IITN network; where (a) is the loss function convergence curve of the original InceptionTime network; (b) is the loss function convergence curve of the IITN network; (c) is the accuracy convergence curve of the original InceptionTime network; (d) is the accuracy convergence curve of the IITN network.

[0024] Among them: 1. Excitation inclined plane; 2. First bracket; 3. Second bracket; 4. First conveyor belt; 5. Second conveyor belt; 6. Third conveyor belt; 7. Pusher; 8. Magnetic tile; 9. Acoustic emission sensor; 10. Photoelectric receiver. Specific implementation manners

[0025] The specific implementation manners of the present invention will be described below to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation manners. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0026] As Figure 1 shown, the internal defect detection device of the magnetic tile based on the acoustic emission signal includes a defect detection component, an excitation inclined plane 1, and an acoustic emission sensor 9; the acoustic emission sensor 9 is arranged on the excitation inclined plane 1; The acoustic emission sensor 9 is used to capture the collision acoustic vibration signal generated when the magnetic tile 8 falls on the excitation inclined plane 1; The defect detection component is used to receive the collision acoustic vibration signal captured by the acoustic emission sensor 9, and detect the internal defect of the magnetic tile according to the collision acoustic vibration signal.

[0027] A time-domain and frequency-domain waveform diagram of the acoustic emission signals of a random qualified and defective magnetic tile is as Figure 2 shown. It can be seen that there is a certain similarity between the time-domain and frequency-domain waveforms of the qualified and defective samples, and there is no obvious standard to measure the difference between the two. If the collision acoustic vibration signal is directly used as the basis for defect detection, the defect detection accuracy may be low. Therefore, an effective signal decomposition method is needed to extract the detailed feature information.

[0028] In this embodiment, the defect detection component includes a signal amplifier, a signal acquisition card, a signal decomposition module, and a defect recognition module, where: The signal amplifier is used to amplify and filter the collision acoustic vibration signal output by the acoustic emission sensor 9 to obtain a preprocessed analog electrical signal; The signal acquisition card is used to convert the preprocessed analog electrical signal into a digital signal at a sampling frequency of 120 kHz; The signal decomposition module is deployed with an improved clustering symplectic geometry mode decomposition method (Symplectic geometry mode decomposition, SGMD) to decompose the digital signal into three component signals, and select the component signal corresponding to the minimum decomposition evaluation index (Evaluation index of decomposition, EIDC) as its output feature signal; where the decomposition evaluation index calculates the error value between the signal before decomposition and the signal after decomposition, and the lower the EIDC value, the higher the similarity between the component and the signal; The defect recognition module is deployed with a pre-trained IITN network, which is used to take the feature signal output by the signal decomposition module as the input of the pre-trained IITN network, classify the input feature signal through the IITN network, and obtain the feature signal classification result, that is, obtain the defect detection result of the magnetic tile 8.

[0029] Figure 3The time-domain and frequency-domain waveform diagrams of a random qualified magnetic tile original signal and three component signals are shown. It can be seen that among the three resulting component signals obtained by the improved clustering symplectic geometric mode decomposition of the signal, component signal 1 represents the instantaneous impact frequency, and component signals 2 and 3 represent meaningless low-frequency and high-frequency components. Among them, component signal 1 is the most contributive characteristic component (i.e., the characteristic signal), and component signal 1 is selected as the input of the IITN network through EIDC.

[0030] In this embodiment, as Figure 4 shown, the IITN network includes a first Inception module, a second Inception module, and a third Inception module connected in sequence, and a global average pooling layer, a flattening layer, and a fully connected layer connected in sequence; among them, the input of the first Inception module is processed by the CBAM attention mechanism and then undergoes a residual connection with the output of the second Inception module, and the result of the residual connection is the input of the third Inception module; the output of the third Inception module is the input of the global average pooling layer; the input of the first Inception module is the input of the IITN network; the output of the fully connected layer is the output of the IITN network.

[0031] Different from the conventional Inception network that uses 3×3 convolutions, the existing InceptionTime network captures features at different time scales through multi-scale convolutional kernels with parallel time steps of 10, 20, and 40. The conventional Inception network relies on the pooling layer to compress the time dimension, while the existing InceptionTime network stacks 6 Inception modules to gradually expand the receptive field. In addition, the existing InceptionTime network forcibly uses residual connections, directly adding the output of the Inception module to the input, effectively alleviating the problem of gradient disappearance.

[0032] The IITN network is improved based on the existing InceptionTime network. On the one hand, since the signal decomposition method has extracted the feature components, the processing of the 6-layer Inception module in the existing InceptionTime network not only causes excessive consumption of resources but also has the problem of long processing time. Therefore, the redundant InceptionTime modules (three Inception modules and the residual connections therein) in the existing InceptionTime network are removed. This architecture simplification strategy results in a reduction in the depth and breadth of feature extraction by the network, thereby affecting the network's ability to capture complex features. For this reason, in this embodiment, the CBAM attention mechanism is added to the residual connection. The CBAM attention mechanism enhances the accuracy of feature expression at different scales and spatial positions by combining the channel attention and spatial attention mechanisms, thereby improving the network recognition accuracy and reducing the data processing time.

[0033] Figure 5 The loss function convergence curve and accuracy convergence curve of the original InceptionTime network and the IITN network are shown. It can be seen that due to the 6-layer Inception module in the original InceptionTime network, the convergence speed of the loss function and accuracy is relatively slow, while the IITN network has a simple architecture and a relatively fast convergence speed. The addition of the CBAM attention mechanism also improves the model accuracy.

[0034] In the Inception module, for a time series with a dimension of first passes through a bottleneck layer, which performs convolution with a stride of 1 to reduce the input dimension to . At the same time, three filters with sizes of 10, 20, and 40 are applied to the time series to extract features at different scales and introduce parallel MaxPooling layers to maintain the robustness of the model. The outputs of each independent parallel convolution and max pooling are connected to form the output of the Inception module. After 3 Inception modules, the data is classified by global average pooling and a fully connected layer.

[0035] In this embodiment, first brackets 2 and second brackets 3 are respectively arranged on both sides of the upper end of the excitation inclined plane 1; a photoelectric emitter is arranged on the first bracket 2, and a photoelectric receiver 10 is arranged on the second bracket 3; the connection line between the photoelectric emitter and the photoelectric receiver 10 intersects the falling path of the magnetic tile 8, so that the magnetic tile 8 blocks the photoelectric receiver 10 from receiving the signal from the photoelectric emitter before falling onto the excitation inclined plane 1, and the time point when the photoelectric receiver 10 does not receive the signal from the photoelectric emitter is used as the signal acquisition starting point of the acoustic emission sensor 9.

[0036] Above the vibration inclined plane 1 and beside the photoelectric receiver 10, a first conveyor belt 4 is provided. At the lower end of the vibration inclined plane 1, a second conveyor belt 5 is provided. On one side of the second conveyor belt 5, a third conveyor belt 6 is provided. On the other side of the second conveyor belt 5, a pusher 7 facing the third conveyor belt 6 is provided. The first conveyor belt 4 is used to convey the magnetic tiles 8 above the vibration inclined plane 1, so that the magnetic tiles 8 pass through the connection line between the photoelectric emitter and the photoelectric receiver 10 and fall onto the vibration inclined plane 1. The second conveyor belt 5 is used to convey the fallen magnetic tiles 8 detected as defect-free in the direction of the qualified magnetic tile discharge port. The pusher 7 is used to push the fallen magnetic tiles 8 detected as defective from the second conveyor belt 5 onto the third conveyor belt 6. The third conveyor belt 6 is used to convey the magnetic tiles 8 detected as defective in the direction of the waste recycling area.

[0037] The magnetic tiles 8 are conveyed on the first conveyor belt 4 in a standing posture, and the convex surface of the magnetic tiles 8 faces the conveying direction. The falling height of the magnetic tiles 8 is 20 mm. This setting method can not only make the magnetic tiles 8 generate effective signals, but also avoid the magnetic tiles 8 from being broken.

[0038] Corresponding to the device, this embodiment also provides a method for detecting internal defects of magnetic tiles based on acoustic emission signals, which includes the following steps: S1. Make the magnetic tiles 8 fall on the vibration inclined plane 1, and capture the collision acoustic vibration signals generated when the magnetic tiles 8 fall on the vibration inclined plane 1 through the acoustic emission sensor 9. S2. Receive the collision acoustic vibration signals captured by the acoustic emission sensor 9 through the defect detection component, and detect the internal defects of the magnetic tiles according to the collision acoustic vibration signals.

[0039] The specific method for receiving the collision acoustic vibration signals captured by the acoustic emission sensor 9 through the defect detection component in step S2 and detecting the internal defects of the magnetic tiles according to the collision acoustic vibration signals includes the following steps: S2-1. Use a signal amplifier to amplify and filter the collision acoustic vibration signals output by the acoustic emission sensor 9 to obtain a preprocessed analog electrical signal. S2-2. Use a signal acquisition card to convert the preprocessed analog electrical signal into a digital signal. S2-3. Use a signal decomposition module deployed with an improved clustering symplectic geometric mode decomposition method to decompose the digital signal into 3 component signals, and select the component signal corresponding to the minimum decomposition evaluation index as the characteristic signal output. S2-4. Use the defect recognition module deployed with the pre-trained IITN network to take the feature signal output by the signal decomposition module as the input of the pre-trained IITN network, classify the input feature signal through the IITN network, obtain the feature signal classification result, that is, obtain the defect detection result of the magnetic tile 8.

[0040] In this embodiment, the specific method of decomposing the digital signal into 3 component signals in step S2-3 includes the following steps: S2-3-1. Apply the time series topological equivalence method to the digital signal to construct a one-dimensional time series and form a trajectory matrix :

[0041] where represents the embedding dimension of the time series topological equivalence method; represents the delay time of the time series topological equivalence method, , represents the data length of the digital signal; represents the value at the delay time; represents the element in the first column of the row of the trajectory matrix S2-3-2. Perform autocorrelation analysis on the trajectory matrix to obtain the covariance matrix , and then obtain the matrix S2-3-3. Obtain the eigenvectors of the eigenvalues of the matrix , and use the eigenvectors and the trajectory matrix to construct the coefficient matrix , and then obtain the initial single-component reconstruction matrix ; where ; S2-3-4. Adjust the elements in each initial single-component reconstruction matrix to obtain the adjusted initial single-component reconstruction matrix: For the element in the row and the m -1 column within the d row of the initial single-component reconstruction matrix , if , let ; for the other elements in the initial single-component reconstruction matrix , let be the element in the row and the a column of the initial single-component reconstruction matrix bElements of the column; is the initial single-component reconstruction matrix in the b row and a column element; S2-3-5. Convert the th adjusted initial single-component reconstruction matrix into a one-dimensional time series , and then obtain the one-dimensional time series ; where in the k th element expression is:

[0042] where is the adjusted initial single-component reconstruction matrix in the row and column element; , min represents taking the minimum value; , max represents taking the maximum value; S2-3-6. Through the agglomerative hierarchical clustering algorithm, the one-dimensional time series is reorganized into 3 component signals by clustering.

[0043] The calculation expression of the minimum decomposition evaluation index in step S2-3 is:

[0044] where represents the minimum decomposition evaluation index of the single-component signal and the digital signal ; represents time; represents the data length of the digital signal, that is, the number of data points in the single-component signal.

[0045] In step S2-3-6, each sample in the initial single-component set is regarded as an independent cluster, forming initial clusters. Using the Euclidean distance metric, calculate the distance between clusters to establish their relative spatial relationship. The inter-cluster distance between cluster and cluster is mathematically expressed as follows:

[0046] Merge the two clusters with the shortest inter-cluster distance, thereby reducing the total number of clusters in the current iteration. Update the distance matrix by recalculating the distance between the newly formed cluster and all remaining clusters. When the number of clusters reaches the specified target When this condition is met, the iterative process will terminate; otherwise, the inter-cluster distance calculation, cluster merging, and distance matrix update will be repeated. Among them , . That is, the set is reconstructed through agglomerative hierarchical clustering, with a value of 3, resulting in three clustering symplectic geometric components, namely three component signals.

[0047] In an embodiment of the present invention, the experimental platform is Win10, NVIDIA GeForce GTX 3060Ti GPU, and cuda11.3. The number of magnetic tile samples is 2000, with the same number of qualified and defective samples. The sampling frequency is 120 kHz, and the number of sampling points is 20000. Its size information is shown in Table 1.

[0048] Table 1: Basic Information of Magnetic Tile Samples

[0049] The feature signals are divided into training samples and test samples. The training samples are input into the InceptionTime pruning model (IITN network) with the CBAM attention mechanism for training, and the trained IITN network is tested with the test samples. The training loss is 0.0029, and the validation loss is 0.0458. The accuracy of the present invention on the test set reaches 99.17%, the precision is 99.65%, and the recall rate is 98.62%, meeting the enterprise's requirement for the recognition rate of qualified magnetic tiles not less than 95%.

[0050] The performance comparisons of the original InceptionTime network, InceptionTime with three Inception modules (3IM-ITN), and the IITN network proposed in the present invention are shown in Table 2, including four parameters: the number of parameters, accuracy, precision, and recall rate. The original InceptionTime network achieved an accuracy of 98.83%. To reduce the computational cost, the number of Inception modules was reduced to three, and the accuracy dropped to 98.17%. In order to enhance the feature extraction ability of the three-layer Inception module network in the present invention, the CBAM mechanism was added. CBAM adaptively adjusts the importance of each channel according to the global information by combining channel attention and spatial attention mechanisms, and captures the detailed information at key positions in the feature map, and the accuracy is improved to 99.17%, proving the effectiveness of combining the three-layer Inception module with the CBAM attention mechanism in enhancing the network robustness and overall classification performance, and the effect is also better than the original InceptionTime network.

[0051] Table 2: Comparison of Three Network Configurations

[0052] In summary, the present invention constructs a detection device / method that can be directly adapted to the magnetic tile production line. After preprocessing the collected collision sound and vibration signals and converting them into digital signals, the improved clustering symplectic geometric mode decomposition method is used to decompose the signals into three component signals. Then, the optimal component signal is selected as the feature signal through the minimum decomposition evaluation index. The IITN network is used to classify the input feature signal to obtain the classification result of the feature signal, that is, the defect detection result of the magnetic tile, thus forming a set of fast, effective, and directly applicable magnetic tile defect detection solutions for the magnetic tile production line.

Claims

1. A magnetic tile internal defect detection device based on acoustic emission signals, characterized in that It includes a defect detection component, an excitation inclined plane (1) and an acoustic emission sensor (9); the acoustic emission sensor (9) is arranged on the excitation inclined plane (1); The acoustic emission sensor (9) is used to capture the collision acoustic vibration signal generated when the magnetic tile (8) falls on the excitation inclined plane (1); The defect detection component is used to receive the collision acoustic vibration signal captured by the acoustic emission sensor (9) and detect the internal defects of the magnetic tile according to the collision acoustic vibration signal.

2. The internal defect detection device of a magnetic tile based on an acoustic emission signal according to claim 1, wherein, The defect detection component includes a signal amplifier, a signal acquisition card, a signal decomposition module and a defect identification module, where: The signal amplifier is used to amplify and filter the collision acoustic vibration signal output by the acoustic emission sensor (9) to obtain a preprocessed analog electrical signal; The signal acquisition card is used to convert the preprocessed analog electrical signal into a digital signal; The signal decomposition module is deployed with an improved clustering symplectic geometric mode decomposition method, which is used to decompose the digital signal into 3 component signals and select the component signal corresponding to the minimum decomposition evaluation index as its output feature signal; The defect identification module is deployed with a pre-trained IITN network, which is used to take the feature signal output by the signal decomposition module as the input of the pre-trained IITN network, classify the input feature signal through the IITN network, obtain the feature signal classification result, that is, obtain the defect detection result of the magnetic tile (8).

3. The internal defect detection device for magnetic tiles based on acoustic emission signals according to claim 2, characterized in that, The IITN network includes a first Inception module, a second Inception module and a third Inception module connected in sequence, and a global average pooling layer, a flattening layer and a fully connected layer connected in sequence; among them, the input of the first Inception module is processed by the CBAM attention mechanism and then connected with the output of the second Inception module in a residual connection, and the residual connection result is the input of the third Inception module; the output of the third Inception module is the input of the global average pooling layer; the input of the first Inception module is the input of the IITN network; the output of the fully connected layer is the output of the IITN network.

4. The internal defect detection device for magnetic tiles based on acoustic emission signals according to claim 1, characterized in that, On both sides of the upper end of the excitation inclined plane (1), a first bracket (2) and a second bracket (3) are respectively arranged; a photoelectric emitter is arranged on the first bracket (2), and a photoelectric receiver (10) is arranged on the second bracket (3); the connection line between the photoelectric emitter and the photoelectric receiver (10) intersects the falling path of the magnetic tile (8), so that the magnetic tile (8) blocks the photoelectric receiver (10) from receiving the signal from the photoelectric emitter before falling on the excitation inclined plane (1), and the time point when the photoelectric receiver (10) does not receive the signal from the photoelectric emitter is used as the signal acquisition starting point of the acoustic emission sensor (9).

5. The internal defect detection device for magnetic tiles based on acoustic emission signals according to claim 4, characterized in that, Above the excitation inclined plane (1) and beside the photoelectric receiver (10), a first conveyor belt (4) is arranged, and at the lower end of the excitation inclined plane (1), a second conveyor belt (5) is arranged; on one side of the second conveyor belt (5), a third conveyor belt (6) is arranged, and on the other side of the second conveyor belt (5), a pusher (7) facing the third conveyor belt (6) is arranged; The first conveyor belt (4) is used to convey the magnetic tiles (8) above the vibrating inclined plane (1), so that the magnetic tiles (8) pass through the connection line between the photoelectric emitter and the photoelectric receiver (10) and fall onto the vibrating inclined plane (1); The second conveyor belt (5) is used to convey the fallen magnetic tiles (8) detected as defect-free towards the direction of the qualified magnetic tile discharge port; The pusher (7) is used to push the fallen magnetic tiles (8) detected as defective from the second conveyor belt (5) onto the third conveyor belt (6); The third conveyor belt (6) is used to convey the magnetic tiles (8) detected as defective towards the direction of the waste recycling area.

6. The internal defect detection device for magnetic tiles based on acoustic emission signals according to claim 5, characterized in that, The magnetic tiles (8) are conveyed in a standing posture on the first conveyor belt (4), and the convex surface of the magnetic tiles (8) faces the conveying direction; the falling height of the magnetic tiles (8) is 20 mm.

7. A method for detecting internal defects of magnetic tiles based on acoustic emission signals, which is based on the device for detecting internal defects of magnetic tiles based on acoustic emission signals according to any one of claims 1 to 6, characterized in that, It includes the following steps: Let the magnetic tiles (8) fall onto the vibrating inclined plane (1), and capture the collision acoustic vibration signal generated when the magnetic tiles (8) fall onto the vibrating inclined plane (1) through the acoustic emission sensor (9); The defect detection component receives the collision acoustic vibration signal captured by the acoustic emission sensor (9), and detects the internal defects of the magnetic tiles according to the collision acoustic vibration signal.

8. A method for detecting internal defects of magnetic tiles based on acoustic emission signals according to claim 7, characterized in that, The specific method for the defect detection component to receive the collision acoustic vibration signal captured by the acoustic emission sensor (9) and detect the internal defects of the magnetic tiles according to the collision acoustic vibration signal includes the following steps: The signal amplifier is used to amplify and filter the collision acoustic vibration signal output by the acoustic emission sensor (9) to obtain a preprocessed analog electrical signal; The signal acquisition card is used to convert the preprocessed analog electrical signal into a digital signal; The signal decomposition module deployed with the improved clustering symplectic geometric mode decomposition method decomposes the digital signal into 3 component signals, and selects the component signal corresponding to the minimum decomposition evaluation index as the characteristic signal output; The defect recognition module deployed with the pre-trained IITN network uses the characteristic signal output by the signal decomposition module as the input of the pre-trained IITN network, classifies the input characteristic signal through the IITN network, obtains the classification result of the characteristic signal, that is, obtains the defect detection result of the magnetic tiles (8).

9. The internal defect detection method of magnetic tiles based on acoustic emission signals according to claim 8, characterized in that, The specific method for decomposing the digital signal into 3 component signals includes the following steps: Construct a one-dimensional time series using the time series topological equivalence method for digital signals and form a trajectory matrix : Among them represents the embedding dimension of the time series topological equivalence method; represents the delay time of the time series topological equivalence method, , represents the data length of the digital signal; represents the value after a delay of time; represents the element in the first column of the first row of the trajectory matrix; Perform autocorrelation analysis on the trajectory matrix to obtain the covariance matrix , and then obtain the matrix ; Obtain a matrix of the eigenvectors corresponding to the eigenvalues , and use the eigenvectors and the trajectory matrix to construct a coefficient matrix , and then obtain an initial single-component reconstruction matrix ; where ; Adjust the elements in each initial single-component reconstruction matrix to obtain the adjusted initial single-component reconstruction matrix: For the initial single-component reconstruction matrix in the m th row and within the d -1th column, if , let ; For the other elements in the initial single-component reconstruction matrix , let ; where is the element in the th row and a th column of the initial single-component reconstruction matrix b ; is the element in the th row and b th column of the initial single-component reconstruction matrix a ; Convert the th adjusted initial single-component reconstruction matrix into a one-dimensional time series , thereby obtaining the one-dimensional time series ; where the expression of the k th element in is: Among them is the adjusted initial single-component reconstruction matrix in the th row and th column element; , min means taking the minimum value; , max means taking the maximum value; Reorganize the one-dimensional time series through the agglomerative hierarchical clustering algorithm Reorganize it into three component signals by clustering 10. The internal defect detection method of magnetic tiles based on acoustic emission signals according to claim 8, characterized in that, The calculation expression of the minimum decomposition evaluation index is: wherein represents a single component signal and the digital signal of the minimum decomposition evaluation index; represents time; represents the data length of the digital signal, that is, the number of data points in a single component signal.

Images