CN114861736A - Internal defect positioning model and internal defect positioning method based on GIALDN (generic identifier distribution network) - Google Patents

Abstract

The invention discloses a GIALDN network-based internal defect localization model, which comprises the following steps: a remolding module: obtaining input signal characteristics; a depoling pretreatment module: obtaining the characteristics of the preprocessed signals; lightweight signal denoising module: the device is used for carrying out soft threshold denoising processing on the preprocessed signal characteristics to obtain denoised signal characteristics; global interactive attention module: connecting each data point in the de-noising signal characteristic with each other, and establishing a relation among each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristic; a multilayer convolution module: obtaining a depth signal characteristic; a result output module: and obtaining a defect positioning result from the depth signal characteristic by using a softmax function. The invention also discloses a GIALDN network-based internal defect positioning method.

Description

Internal defect positioning model and internal defect positioning method based on GIALDN network

Technical Field

The invention belongs to the technical field of defect detection, and particularly relates to an internal defect positioning model and an internal defect positioning method based on a GIALDN network.

Background

The carbon fiber reinforced resin matrix Composite (CFRP) is formed by compounding carbon fibers and a resin matrix, has superior performances of high specific strength, high specific stiffness, high temperature resistance, corrosion resistance, good designability and the like, and is widely applied to the fields of aerospace, automobiles, construction, sports and the like. Particularly in the field of aerospace, the use ratio of the composite material on an aircraft structure is continuously increased, and the use amount of the composite material becomes one of important indexes for judging the advancement of the aircraft. However, due to the special molding process and internal structure of the composite material, internal defects are easily generated during molding, processing and service, such as: internal fat rich areas are created due to the unreasonable distribution of fibres; in the molding process, air is wrapped into the flowing front edge of the resin to form bubble defects; internal delamination and the like due to external forces or impacts during processing, assembly and service. These internal defects can greatly reduce the service performance of the CFRP, posing a fatal threat to the safe operation of the aircraft. Therefore, the detection of internal defects is of central importance in both CFRP factory testing and in the periodic inspection of aircraft.

At present, the common nondestructive detection method for the internal defects of the composite material mainly comprises an ultrasonic detection technology, an infrared thermography detection technology, an X-ray detection technology, a terahertz detection technology and the like. The ultrasonic detection utilizes the principle that when the ultrasonic waves encounter internal defects in the process of penetrating through the composite material, a part of ultrasonic waves can return to the interface of the composite material, and the defect detection is realized by analyzing reflected waves; however, the propagation characteristics of ultrasound waves inside composite materials and the acoustic characteristics of different materials are not currently studied in depth, which limits the use of ultrasound detection. The infrared thermal image detection technology detects defects by detecting the surface temperature change of an object according to the infrared radiation principle; the defects are that the sensitivity is low, the surface of a workpiece is required to have good heat absorption rate, and the infrared thermography technology has insufficient detection depth due to long infrared wave wavelength and weak penetrability. The X-ray detection technology detects defects by using the principle that loss of rays is different when the rays penetrate through the defect positions and normal components, although the sensitivity is high and most of the defects can be detected, ionizing radiation of the X-rays can damage human bodies, certain pollution is caused to the environment, the detection cost is high, and the adaptability is poor. The terahertz technology is a new detection technology, is currently in a starting stage, and has the problems of high detection cost and narrow application range. It can be seen that all the common nondestructive testing methods have the defects of low testing efficiency, high cost, poor adaptability and the like, and cannot well meet the testing requirements of the internal defects of the CFRP.

With the rapid development of the deep learning technology, a CFRP internal defect detection and positioning method combining a traditional detection method and the deep learning technology gradually appears. The deep learning technology effectively improves the processing precision and efficiency of data such as ultrasonic waves and infrared thermal imaging, but the problems of low efficiency, insufficient adaptability and the like of the technologies cannot be fundamentally solved. In recent years, techniques for processing vibration signals using machine learning methods have been developed and are also beginning to be applied to the detection of internal defects in composite materials. The data-driven CFRP internal defect detection method can solve the defects of the previous method, but the research still has the following defects: (1) the existing method only uses a shallow neural network, cannot extract abundant defect characteristics in signals, and has the defects of insufficient generalization and insufficient mobility; (2) due to the complex and random internal structure of the composite material, a large amount of noise exists in the acquired vibration signal, which is very unfavorable for defect positioning, but the factor is not considered in the current research, so that the detection precision is inevitably influenced; (3) in the aspect of neural network signal feature extraction, a quick and effective measure for extracting long-term information and cross-channel connection is lacked, which is important for defect positioning; (4) the current defect positioning model does not realize very high defect positioning precision (less than 90%), is far away from engineering practicality and has insufficient reliability. Therefore, it is very important to design a deep neural network which has a noise suppression and elimination function and can extract the communication between the channels in the signal to perform CFRP internal defect localization based on the excitation response signal.

Disclosure of Invention

In view of the above, an object of the present invention is to provide an internal defect localization model and an internal defect localization method based on a GIALDN network, which can suppress and eliminate noise signal interference, interact data in a signal, effectively extract a characteristic with identifiability from acquired signal data, and finally achieve high-precision localization of an internal defect.

In order to achieve the purpose, the invention provides the following technical scheme:

the invention firstly provides an internal defect positioning model based on a GIALDN network, which comprises the following steps:

a remolding module: the device is used for remodeling the collected vibration signal data to obtain input signal characteristics;

a deflashing pretreatment module: extracting vibration signal data rich in original information from the input signal characteristics by using the depoling preprocessing convolution layer to obtain preprocessing signal characteristics;

lightweight signal denoising module: the device is used for carrying out soft threshold denoising processing on the preprocessed signal characteristics so as to strip the noise signals and obtain denoised signal characteristics;

global interactive attention module: connecting each data point in the de-noised signal characteristics with each other, and establishing a relation between each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the distinguishing power of the characteristics;

a multilayer convolution module: the depth feature extraction module is used for carrying out depth feature extraction on the signal features sequentially processed by the lightweight signal denoising module and the global interaction attention module to obtain depth signal features;

a result output module: and obtaining a defect positioning result from the depth signal characteristic by using a softmax function.

Further, the method for the lightweight signal denoising module to perform soft threshold denoising processing is as follows:

wherein, y _i Representing the characteristic of a denoised signal of the ith channel; x is the number of _i Representing the preprocessed signal features of the ith channel; tau is _i Representing the ith channel soft threshold.

Further, the lightweight signal denoising module comprises a global mixing pooling layer and a one-dimensional convolution layer;

the global mixed pooling layer is used for simultaneously carrying out global maximum pooling and global average pooling on the preprocessed signal features to obtain mixed signal features;

the one-dimensional convolutional layer is used for extracting channel characteristics of mixed signal characteristics, and the learned coefficients are scaled to the range of [0,1] by using a sigmoid activation function to obtain a soft threshold coefficient sigma of each channel; then:

τ _i ＝σ _i ×|x|

wherein, tau _i Representing a soft threshold corresponding to the ith channel; sigma _i Representing the soft threshold coefficient corresponding to the ith channel; and | x | represents the preprocessed signal features after absolute value processing.

Further, the method for processing the de-noised signal features by the global interaction attention module is as follows:

(1) interpolating the denoised signal characteristic X by adopting an interplate function, and interpolating the characteristic XReducing the dimension of the feature map from h number pair to (h/2) × (w/2) to obtain the feature map X _i ；

(2) Feature map X is mapped via a convolutional layer with batch standardized BN layer, Relu activation function _i The number of channels is changed from c to (h/2) X (w/2), and a characteristic diagram X is obtained _c ；

(3) For feature X _c Performing global interaction to generate an attention diagram A _a ；

(4) Attention-seeking diagram A _a And feature map X _i Carrying out point-by-point multiplication to obtain a global interactive attention feature map A _c ；

(5) Interpolate function pair global interactive attention feature map A _c Performing interpolation operation to obtain a global interactive attention feature map A _c Reducing the characteristic diagram A into the same shape as the characteristic X of the de-noised signal _i ；

(6) Using the residual block to obtain the final output characteristic diagram as follows:

Y＝X+A _i 。

further, in the step (3), the feature X is processed _c Global interaction to generate an attention map A _a The method comprises the following steps:

(31) will be characterized by X _c Generating a data point feature map with the size of 1 multiplied by 1 and the number of channels of (h/2) multiplied by (w/2) in all channels for each data point, and then reshaping the data feature map into a single-channel feature map with the size of (h/2) multiplied by (w/2) and the number of channels of 1;

(32) aggregating the single-channel characteristic graphs of all the data points to obtain an intermediate characteristic graph;

(33) generating an attention map A from the intermediate feature map by using a softmax function _a 。

The invention also provides a GIALDN network-based internal defect positioning method, which comprises the following steps:

the method comprises the following steps: collecting data: driving the detected object to vibrate, converting vibration signals of different areas of the detected object into electric signals by using a plurality of sensors, and converting the electric signals into digital signals by using an A/D converter to obtain vibration signal data;

step two: remodeling: reshaping vibration signal data to obtain input signal characteristics;

step three: removing a pool for pretreatment: extracting vibration signal data rich in original information from the input signal characteristics by using the depoling preprocessing convolution layer to obtain preprocessing signal characteristics;

step four: denoising lightweight signals: carrying out soft threshold denoising processing on the preprocessed signal characteristics to strip the noise signals to obtain denoised signal characteristics;

step five: global interactive attention processing: connecting each data point in the de-noising signal characteristic with each other, and establishing a relation among each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristic;

step six: multilayer convolution processing: carrying out depth feature extraction on the signal features subjected to the lightweight signal denoising processing and the global interaction attention processing in sequence to obtain depth signal features;

step seven: and outputting a result: and obtaining a defect positioning result from the depth signal characteristic by using a softmax function.

Further, in the fourth step, the method for denoising the lightweight signal includes:

41) performing global maximum pooling and global average pooling on the preprocessed signal features by using a global mixed pooling layer to obtain mixed signal features;

42) extracting channel characteristics of mixed signal characteristics by using the one-dimensional convolutional layer, and adopting a sigmoid activation function to scale the learned coefficient to the range of [0,1] to obtain a soft threshold coefficient sigma of each channel; then:

τ _i ＝σ _i ×|x _i |

wherein, tau _i Representing a soft threshold corresponding to the ith channel; sigma _i Representing a soft threshold coefficient corresponding to the ith channel; | x _i L represents the preprocessed signal characteristic of the ith channel after absolute value conversion;

43) carrying out soft threshold denoising processing on the preprocessed signal characteristics:

wherein, y _i Representing the characteristic of a denoised signal of the ith channel; x is the number of _i Representing the preprocessed signal features of the ith channel; tau is _i Representing the ith channel soft threshold.

Further, in the fifth step, the method for processing global interaction attention is as follows:

(1) interpolating the denoised signal characteristic X by adopting an interplate function, reducing the dimensionality of the characteristic graph from h number pair to (h/2) X (w/2), and obtaining the characteristic graph X _i ；

(2) Feature map X is represented by a convolution layer with batch normalized BN layer, Relu activation function _i The number of channels is changed from c to (h/2) X (w/2), and a characteristic diagram X is obtained _c ；

(3) For feature X _c Performing global interaction to generate an attention diagram A _a ；

(4) Attention-seeking diagram A _a And feature map X _i Carrying out point-by-point multiplication to obtain a global interactive attention feature map A _c ；

(5) Interpolate function pair global interactive attention feature map A _c Performing interpolation operation to obtain a global interactive attention feature map A _c Reducing the characteristic diagram A into the same shape as the characteristic X of the de-noised signal _i ；

(6) Using the residual block to obtain the final output characteristic diagram as follows:

Y＝X+A _i 。

further, in the step (3), the feature X is processed _c Global interaction to generate an attention map A _a The method comprises the following steps:

(31) will be characterized by X _c Generating a data point feature map with the size of 1 multiplied by 1 and the number of channels of (h/2) multiplied by (w/2) in all channels for each data point, and then reshaping the data feature map into a single-channel feature map with the size of (h/2) multiplied by (w/2) and the number of channels of 1;

(32) aggregating the single-channel characteristic graphs of all the data points to obtain an intermediate characteristic graph;

(33) generating an attention map A from the intermediate feature map by using a softmax function _a 。

Further, the object to be detected is a CFRP laminated plate, the sensors are two groups with the same number, and the two groups of sensors are respectively and uniformly distributed on the inner layer and the outer layer of the CFRP laminated plate; in the first step, the data acquisition method comprises the following steps:

the sinusoidal digital signal generated by the signal generator is converted into an electric signal by a D/A converter, the amplitude is controlled by a power amplifier, and finally the electric signal is transmitted to a vibration exciter to control the generation of physical vibration;

the vibration of the vibration exciter acts on the CFRP laminated plate through the vibration exciting rod to enable the CFRP laminated plate to vibrate;

the vibration signals of different areas of the detected object are converted into electric signals by using the two groups of sensors, and the electric signals are converted into digital signals by using the A/D converter, so that vibration signal data are obtained.

The invention has the beneficial effects that:

the internal defect positioning model based on the GIALDN network has the beneficial effects that:

1) by using a threshold denoising method in a signal processing technology, a lightweight signal denoising module is designed to inhibit and eliminate noise-related characteristics and improve signal data quality;

2) each data point in the signal is connected with each other through a global interaction attention module, and a relation is also established among each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristics;

3) constructing a GIALDN by taking the pooled convolution layer and the multilayer convolution as a backbone and combining a lightweight signal denoising module and a global interaction attention module, so as to realize the feature extraction of the excitation response signal and the positioning of internal defects; an excitation response signal data set is collected by building a laboratory table, the performance of the GIALDN is tested, and the positioning precision of the GIALDN is higher than that of the existing common model.

Drawings

In order to make the object, technical scheme and beneficial effect of the invention more clear, the invention provides the following drawings for explanation:

FIG. 1 is a schematic diagram of a data acquisition system;

FIG. 2 is a schematic diagram of the structure of the internal defect localization model of the GIALDN network of the present invention;

FIG. 3 is a schematic structural diagram of a lightweight signal denoising module;

FIG. 4 is a schematic diagram of the structure of a global interactive attention module;

FIG. 5 is a schematic diagram of a global interaction process;

FIG. 6 is a pictorial view of the vibration testing apparatus;

FIG. 7 is a graph of model accuracy comparison;

FIG. 8 is a graph comparing model loss values;

figure 9 is a graph of training and test loss for GIALDN.

Detailed Description

The present invention is further described with reference to the following drawings and specific examples so that those skilled in the art can better understand the present invention and can practice the present invention, but the examples are not intended to limit the present invention.

As shown in fig. 1, the data acquisition system includes a vibration exciter, an object to be detected, a sensor, an a/D converter, a power amplifier, a signal acquisition card, and signal acquisition software on a computer. The object to be detected in this embodiment is a composite material plate, that is, a CFRP laminate. Firstly, the excitation signal is a sinusoidal digital signal generated by a signal generator, the digital signal is converted into an electric signal through a D/A converter, the amplitude is controlled through a power amplifier, and finally the electric signal is transmitted to an exciter to control the generation of physical vibration. The vibration of the vibration exciter acts on the composite material plate through the vibration exciting rod, so that the composite material plate vibrates. Meanwhile, the vibration response signal with the characteristics of the CFRP laminated plate is acquired by the sensor and converted into an electric signal, the electric signal is sampled and quantized by the A/D converter to be converted into a digital signal, and the digital signal is observed, stored and recorded in a signal acquisition system. In the experiment, 8 sensors are used in total, and an inner layer and an outer layer are uniformly distributed in four areas of the composite plate so as to simultaneously acquire signal data of 8 channels. According to the signal time domain sampling theorem f _ s which is more than or equal to 3-5f _ max, a proper sampling frequency is set, so that the change condition of the original data can be truly reflected by the acquired signal while the frequency aliasing error is avoided.

Fig. 2 is a schematic diagram of the structure of the internal defect localization model of the GIALDN network according to the present invention. The internal defect location model based on the GIALDN network in the embodiment comprises the following steps:

a remolding module: for reshaping the acquired vibration signal data to obtain input signal characteristics. Namely: and (4) remodeling the 8-path excitation response signal data acquired from the sensor to obtain the input signal characteristics of the whole model.

A depoling pretreatment module: and extracting vibration signal data rich in original information from the input signal characteristics by using the de-pooling pre-processing convolution layer to obtain pre-processing signal characteristics. It should be noted that the traditional pooling operation is abandoned in the convolutional layer. Although the pooling operation can compress features, reduce dimensions, simplify network complexity, and reduce computation, a great deal of valuable information is inevitably lost due to the fixed down-sampling rule of the pooling operation, and the association between the whole and the part is ignored. This is very disadvantageous for the task of defect localization using multi-channel vibration signals, since accurate defect location related information and small differences between the signal data are required for localization. Therefore, to preserve the exact detailed characteristics of the signal data, only the deconvolution layer consisting of convolution, batch normalization, Relu activation, and discard operations is used.

Lightweight Denoised Module (LMD): the method is used for carrying out soft threshold denoising processing on the preprocessed signal characteristics so as to strip the noise signals and obtain the denoised signal characteristics.

Global Interaction Module (GIAM): connecting each data point in the de-noising signal characteristic with each other, and establishing a relation among each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristic;

a multilayer convolution module: the depth feature extraction module is used for carrying out depth feature extraction on the signal features sequentially processed by the lightweight signal denoising module and the global interaction attention module to obtain depth signal features; the multi-layer convolution module of the present embodiment employs an 11-layer convolutional neural network with residual blocks.

A result output module: and obtaining a defect positioning result from the depth signal characteristics by using a softmax function.

Due to the complexity and randomness of the internal structure of the composite material, the excitation response signal collected from the composite material slab also contains a significant amount of noise. The noise can obscure signal characteristics and reduce signal quality, so that the neural network cannot extract useful characteristics related to fault and defect positioning from the vibration signal, and the output layer cannot make accurate judgment due to the fact that the learned characteristics lack sufficient identification, and important adverse effects are generated on a defect positioning task. Therefore, in the work of utilizing the vibration signal to locate the defect, it is necessary to design a signal denoising processing module to suppress and eliminate the noise and improve the signal-to-noise ratio of the data.

The signal denoising module of this embodiment employs a threshold denoising method to eliminate features related to noise. The module first uses global hybrid pooling to achieve feature aggregation and dimensionality reduction; secondly, 1DCNN is adopted to realize the interaction of local cross channels, capture the relation between the channels and generate a soft threshold coefficient; finally, a soft threshold denoising operation is performed on the input signal, as shown in fig. 3:

that is, the lightweight signal denoising module of this embodiment includes a global mixture pooling layer and a one-dimensional convolution layer.

And the global mixed pooling layer is used for simultaneously carrying out global maximum pooling and global average pooling on the preprocessed signal features to obtain mixed signal features.

Unlike SEnet and DRSN, in the stage of compressing the spatial dimension of the input to obtain the global characteristics of each channel, Global Average Pooling (GAP) is not adopted separately, but Global Maximum Pooling (GMP) and Global Average Pooling (GAP) are considered together to aggregate the global characteristics of the signals, so as to better endow each channel with a reasonable soft threshold.

GAP is the most common operation for aggregating global features, and although global average is a very important feature, it should be observed that it is only one of the global features, and cannot be completely expressed for the global features of the data. GMP extracts global saliency features, another important feature representation of the input data, which is important for setting reasonable soft thresholds for each channel. The simultaneous consideration of GAP and GMP has also proved to greatly improve the representation capability of the network.

Different from the full connection structure of the SEnet module adopting squeeze and excite operations, the embodiment adopts 1DCNN to extract the channel features. The advantages of using 1DCNN over SEnet are: (1) although the dimension reduction operation in SEnet reduces the model complexity and the calculation amount to a certain extent, the direct corresponding relation between the channels and the weights thereof is damaged, which is important for the generation of the soft threshold of each channel; (2) because the 1DCNN has the characteristic of parameter sharing, compared with SEnet, the parameter quantity and the model complexity are greatly reduced; (3)1DCNN enables efficient local cross-channel interaction, rather than unnecessary global channel interaction, which is important for learning channel soft thresholds. The network using 1DCNN has fewer parameters and higher accuracy than the fully-connected network using SE structure.

The one-dimensional convolutional layer is used for extracting channel characteristics of mixed signal characteristics, and the learned coefficients are scaled to the range of [0,1] by using a sigmoid activation function to obtain a soft threshold coefficient sigma of each channel; then:

τ _i ＝σ _i ×|x|

wherein, tau _i Representing a soft threshold corresponding to the ith channel; sigma _i Representing the soft threshold coefficient corresponding to the ith channel; and | x | represents the preprocessed signal features after absolute value processing.

The method for the lightweight signal denoising module to perform soft threshold denoising processing is as follows:

wherein, y _i Representing the characteristic of a denoised signal of the ith channel; x is the number of _i Representing the preprocessed signal features of the ith channel; tau is _i Representing the ith channel soft threshold.

According to the formula, the method for denoising by using the soft threshold is to filter the input features in the range of [ - τ, τ ] by setting the input features to be 0, and besides, the soft threshold is added to the negative features, and the soft threshold is subtracted from the positive features, so that the noise of the input features is suppressed to the maximum extent, and simultaneously, the effective signals are kept as complete as possible. Meanwhile, the soft threshold coefficient is obtained through global mixed pooling and 1DCNN layer learning, and each channel has a corresponding coefficient, so that the denoising capability of the module is continuously improved.

For the global interaction attention module, a cross-channel and global interaction attention mechanism is adopted, and the remote context of the input signal characteristics and the relation between channels can be flexibly aggregated; by generating a stacked attention map for the data of each position in the feature map, and connecting each position with other positions, various effective information of short and far positions of the signal features is fully acquired to enhance the model representation capability.

The composite material sheet collected in this example has some of the following characteristics:

a. the data collected are 8 channels of vibration signals collected by 8 sensors distributed on the composite board, and at each defect position, 8 channels of data about the defect can be collected by the sensors, and each channel of data has a unique characteristic about the defect. And, because of being located on the same board, there is a close relation between these 8 signal data;

b. the vibration signal data is one kind of time sequence signal, has the characteristics of strong tandem relation and long-time dependence of the time sequence data, and is realized by capturing the tandem relation or long-time dependence in the signal in the general time sequence prediction or fault diagnosis task. In the defect positioning task, the characteristic of the vibration signal can also help to position;

due to the characteristics of the multi-channel vibration signal, the embodiment provides a global interaction concept so as to effectively capture implicit characteristics such as the relation between 8-channel data and the dependency between each point and other points in the signal, and by aggregating the characteristic relations, the representation capability of the signal can be effectively improved, so that a model learns more characteristic representations, and the defect positioning accuracy is improved.

In the task of defect positioning, defects appearing on the composite material plate need to be accurately positioned. In the positioning task, the positioning of each defect position is influenced by other positions on the characteristic diagram, and data in each channel are also related, which is very important; therefore, this effect can be utilized to assist in achieving positioning. The specific operation is as follows: each position on the feature map acquires the influence of other positions on the position, namely the degree of the features of other positions can assist in prediction; meanwhile, own characteristics can also provide help for prediction of other positions under the interaction. In such a mode, the prediction of each position can be assisted by other positions, and the prediction of each position can also be assisted by the prediction of each position, so that the positioning of other positions is assisted, and a bidirectional information interaction is formed. The bidirectional information interaction can fully utilize hidden information among signal data to capture internal connection of each signal, so that the connection of each position is tight, the neural network can learn more comprehensive representation characteristics, and the defect positioning accuracy is improved.

Specifically, the method for processing the denoised signal feature by the global interaction attention module of the embodiment is as follows:

(1) interpolating the denoised signal characteristic X by adopting an interplate function, reducing the dimensionality of the characteristic graph from h number pair to (h/2) X (w/2), and obtaining the characteristic graph X _i (ii) a The calculation amount in the global interaction stage is reduced, and extra calculation overhead caused by the overlarge number of the characteristic channels is avoided.

(2) Feature map X is represented by a convolution layer with batch normalized BN layer, Relu activation function _i The number of channels is changed from c to (h/2) X (w/2), and a characteristic diagram X is obtained _c ；

(3) For feature X _c Performing global interaction to generate an attention diagram A _a ；

(4) Attention panel A _a And feature map X _i Carrying out point-by-point multiplication to obtain a global interaction attention feature map A _c ；

(5) Interpolate function pair global interactive attention feature map A _c Performing interpolation operation to obtain a global interactive attention feature map A _c Reducing the characteristic diagram A into the same shape as the characteristic X of the de-noised signal _i ；

(6) In order to prevent the feature response value from decreasing after the feature map undergoes global interaction, the embodiment uses a residual block to obtain a final output feature map as follows:

Y＝X+A _i 。

in the global interaction stage, the "global" interaction of the embodiment is that interaction is performed simultaneously in two aspects of space and channel, and the information of the acquired multi-channel vibration signal data is fully utilized while considering the relation between all points in the signal and the mutual influence and help between the channels. The specific global interactive process is shown in fig. 5. For feature X _c Global interaction to generate an attention map A _a The method comprises the following steps:

(31) will be characterized by X _c Generating a data point feature map with the size of 1 multiplied by 1 and the number of channels of (h/2) multiplied by (w/2) in all channels for each data point, and then reshaping the data feature map into a single-channel feature map with the size of (h/2) multiplied by (w/2) and the number of channels of 1;

(32) aggregating the single-channel characteristic graphs of all the data points to obtain an intermediate characteristic graph;

(33) generating an attention map A from the intermediate feature map by using a softmax function _a 。

Similarly, the specific implementation manner of the internal defect location method based on the giadn network proposed in this embodiment is as follows:

the internal defect positioning method based on the GIALDN network comprises the following steps:

the method comprises the following steps: collecting data: driving the detected object to vibrate, converting vibration signals of different areas of the detected object into electric signals by using a plurality of sensors, and converting the electric signals into digital signals by using an A/D converter to obtain vibration signal data;

the object to be detected in the embodiment is a CFRP laminated board, the number of sensors is two, each group of sensors includes 4, and the two groups of sensors are respectively and uniformly distributed in four areas of an inner layer and an outer layer of the CFRP laminated board; specifically, the method for acquiring data in this embodiment is as follows:

the sinusoidal digital signal generated by the signal generator is converted into an electric signal by a D/A converter, the amplitude is controlled by a power amplifier, and finally the electric signal is transmitted to a vibration exciter to control the generation of physical vibration;

the vibration of the vibration exciter acts on the CFRP laminated plate through the vibration exciting rod to enable the CFRP laminated plate to vibrate;

the vibration signals of different areas of the detected object are converted into electric signals by using the two groups of sensors, and the electric signals are converted into digital signals by using the A/D converter, so that vibration signal data are obtained.

Step two: remodeling: reshaping vibration signal data to obtain input signal characteristics;

step three: removing a pool for pretreatment: extracting vibration signal data rich in original information from the input signal characteristics by using the depoling preprocessing convolution layer to obtain preprocessing signal characteristics;

step four: denoising lightweight signals: carrying out soft threshold denoising processing on the preprocessed signal characteristics to strip the noise signals to obtain denoised signal characteristics; the method for denoising the lightweight signal comprises the following steps:

41) performing global maximum pooling and global average pooling on the preprocessed signal features by using a global mixed pooling layer to obtain mixed signal features;

42) extracting channel characteristics of mixed signal characteristics by using the one-dimensional convolutional layer, and adopting a sigmoid activation function to scale the learned coefficient to the range of [0,1] to obtain a soft threshold coefficient sigma of each channel; then:

τ _i ＝σ _i ×|x _i |

wherein, tau _i Representing a soft threshold corresponding to the ith channel; sigma _i Representing the soft threshold coefficient corresponding to the ith channel; | x _i L represents the preprocessed signal characteristic of the ith channel after absolute value conversion;

43) carrying out soft threshold denoising processing on the preprocessed signal characteristics:

wherein, y _i Representing the characteristic of a denoised signal of the ith channel; x is the number of _i Representing the preprocessed signal features of the ith channel; tau is _i Representing the ith channel soft threshold.

Step five: global interactive attention processing: connecting each data point in the de-noising signal characteristic with each other, and establishing a relation among each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristic; the method of global interactive attention processing is as follows:

(1) interpolating the denoised signal characteristic X by adopting an interplate function, reducing the dimensionality of the characteristic graph from h number pair to (h/2) X (w/2), and obtaining the characteristic graph X _i ；

(2) Feature map X is represented by a convolution layer with batch normalized BN layer, Relu activation function _i The number of channels is changed from c to (h/2) X (w/2), and a characteristic diagram X is obtained _c ；

(3) For feature X _c Performing global interaction to generate an attention diagram A _a (ii) a For feature X _c Global interaction to generate an attention map A _a The method comprises the following steps:

(31) will be characterized by X _c Generating a data point feature map with the size of 1 multiplied by 1 and the number of channels of (h/2) multiplied by (w/2) in all channels by each data point, and then reshaping the data feature map into a single-channel feature map with the size of (h/2) multiplied by (w/2) and the number of channels of 1;

(32) aggregating the single-channel characteristic graphs of all the data points to obtain an intermediate characteristic graph;

(33) generating an attention map A from the intermediate feature map by using a softmax function _a ；

(4) Attention-seeking diagram A _a And feature map X _i Carrying out point-by-point multiplication to obtain a global interactive attention feature map A _c ；

(5) Interpolate function pair global interactive attention feature map A _c Performing interpolation operation to obtain a global interactive attention feature map A _c Reducing the characteristic diagram A into the same shape as the characteristic X of the de-noised signal _i ；

(6) Using the residual block to obtain the final output characteristic diagram as follows:

Y＝X+A _i 。

step six: multilayer convolution processing: carrying out depth feature extraction on the signal features subjected to the lightweight signal denoising processing and the global interaction attention processing in sequence to obtain depth signal features;

step seven: and outputting a result: and obtaining a defect positioning result from the depth signal characteristic by using a softmax function.

The following describes specific embodiments of the internal defect localization model and the internal defect localization method based on the gialn network according to this embodiment based on examples.

In order to verify the effectiveness of the model and the method provided by the embodiment, a CFRP laminated plate excitation response test experimental platform is established, a data set is collected, and the GIALDN model is compared with other commonly used classical network models by utilizing the data set.

First, experimental data set

The vibration test platform designed by the theory mainly comprises a vibration exciter, an A/D converter, a power amplifier, a sensor, a vibration test bench, a data acquisition card and matched data acquisition software, wherein specific experimental equipment is shown in figure 6, and specific equipment models are shown in table 1. The A/D converter has the A/D, D/A bidirectional conversion function of analog signals and digital signals, and realizes signal type conversion; the power amplifier can directly control the excitation of the vibration exciter, such as amplitude and the like; the vibration excitation device can receive the analog electric signal from the power amplifier to generate vibration; the sensor can capture the vibration signal of the CFRP laminated plate, convert the vibration signal into an analog signal and transmit the analog signal to the A/D converter; the vibration experiment bench is used for fixing the CFRP laminated plate and the vibration exciter, and signal distortion caused by relative movement is avoided; the data acquisition card and the matched software can generate digital signals with corresponding frequency and amplitude, such as sinusoidal signals, square wave signals and the like, according to the setting requirement, and can observe and store and record the acquired signals in real time.

TABLE 1 Experimental Equipment

In the experiment, the excitation frequency is set to be 120Hz and the sampling frequency is set to be 1000Hz on the signal acquisition software, so that the acquired signals can truly reflect the vibration information of the composite board; the amplitude is set to be 1V, and the amplitude is stable through the power amplifier, so that the sensor can be firmly and stably fixed on the composite board in the vibration process. In the defect position setting, the defect positions were set at intervals of 10mm for a composite material plate area of 240 × 240mm, and vibration signals of 552 defect positions were collected in total. In the model training, each signal takes 0.4 second of data, namely each sample channel has 400 data values, and the defect position of each sample channel is located. This localization of defects using short signals is very challenging and can also better verify the performance of the proposed model. To avoid overfitting during training, several continuous signals were taken from each of the 552 vibration signals collected to increase the data set. And finally, the composite plate defect vibration signal data set contains 11040 samples, and a training set and a testing set are divided according to a proportion of 0.2.

It is also observed that the noise signal persistence exists due to the inherent errors of the sensor and other devices, the surrounding environment noise, etc. during the whole process, which poses a challenge to the denoising capability of the model.

Second, model hyper-parameter setting

And (3) directly analyzing the composite material vibration data set by using a GIALDN model in a verification experiment, and finally judging the performance of the model according to the accuracy of the area where the defect position is located. In addition, the hyper-parameter settings related to the proposed network model are introduced below.

TABLE 2 Hyperparameter settings for GIALDN

The hyper-parameters associated with the designed GIALDN network structure, such as number of convolutional layer input channels, number of convolutional kernels, size, sliding step size, etc., are summarized in Table 2. Wherein, the parameters in the De-posing Conv2d sequentially represent the number of input and output channels of the convolutional layer, the size of a convolutional kernel, the sliding step length, the filling, the activation function and the Dropout proportion; LDM parameter represents the number of input and output channels; the parameter of the GIAM indicates the number of input channels, the number of interactive branches, whether compact or not, and the like. Multi-LayersConv2d is used as a convolution layer in the ResNet architecture, and its parameters indicate the number of convolution layers per ResBlock. The output data shapes of the different layers or modules are shown in the first column of the table, representing the number of channels x length x width, respectively. And using GAP operation on the final output layer of the Multi-Conv2d to reduce the output dimension, and obtaining the possibility of judging each position through Softmax, wherein the defect position needing to be positioned is the position with the largest numerical value.

In terms of training optimization of the whole network, because the Adam optimizer combines the advantages of Adagrad and RMSProp, the same learning rate is used for each parameter, the Adam optimizer can be independently adapted along with the learning, and the historical information of the gradient is used for optimizing the model, so that the Adam optimizer is selected and the learning rate is set to be 0.001 to ensure the training speed of the model. The batchsize setting of the present embodiment is 128 according to the principle that the larger the equipment performance and the better the batch size. In the training process, after each generation of training is finished, the model is verified by using a verification set so as to find problems occurring in the training process in advance and avoid time waste.

The proposed depth model is compiled on the basis of python version 3.8 using a pytorch deep learning framework, and all validation experiments are performed on a Linux server of the Nvidia GTX GPU 3080.

Third, Performance test

Two sets of experimental tests were set up in this example: the first set of experiments was used to compare the effectiveness of further improvements of the method and model proposed in this example; the second set of experiments was a comparison of the performance of the convolution structure set forth for the model of this example with other classical CNN networks.

3.1 composition component Performance test of model

The validity of each module of the GIALDN was experimentally verified in this example, and the specific results are shown in table 3. In the model, LDM and GIAM are used as two main modules of the model, which greatly affects the performance of the model, and secondly, the strategy of de-pooling is also effective for improving the performance, and the functions of interpolation and conv1d are mainly expressed in reducing the parameter quantity of the model.

TABLE 3 model component Performance test

It can be seen that although the debasing strategy increases the number of model parameters and increases the complexity of the model, it also increases the accuracy of the model. In fact, compared with the parameter number of the whole model, the increased calculation amount caused by the deflashing strategy is negligible, and the influence on the calculation efficiency and the training time of the whole model is little. Therefore, it is also desirable that the deballasting strategy be valuable as a whole.

In LDM, the value representing the characteristics of each channel obtained after global pooling is the basis for generating a soft threshold, and the denoising effect of the model is directly influenced. It can be seen from the table that the model accuracy using the global average pooling alone or the global maximum pooling alone is significantly lower than the hybrid pooling, which indicates that the channel information obtained using the two pooling approaches alone is incomplete and cannot represent most of the data in one channel, and only using them simultaneously can realize the complete representation of the characteristics of the whole channel. In the subsequent inter-channel feature extraction process, compared with a form in which all channels are connected by the full connection layer FC to obtain relationships between all channels, in this embodiment, a mode of extracting local information between the channels by using the Conv1d operation is adopted, so that an effective inter-channel relationship can be obtained, the channel information extracted by the FC is too redundant and inefficient, and meanwhile, the number of parameters is also significantly reduced, and the calculation efficiency is significantly improved. It should be noted that after the LDM is removed, the model accuracy is reduced by 1.02%, which fully indicates that the denoising operation is necessary and effective for this kind of high-noise vibration data.

In the modification experiment of the GIAM, the effect of the interpolation dimension reduction method on reducing the calculation amount and the model complexity is remarkable, and the method mainly represents two aspects: the size of the feature map and the realizability of the number of channels (the number of channels is within a reasonable range) are adopted. Compared with a network model adopting a GIAM module, after the GIAM module is removed, the model precision is reduced by 0.66%, obviously, the GIAM module is effective in extracting the data interaction relation and the inter-channel relation in the channel, so that useful data and more effective channels can pay more attention to the network model, and effective characteristic quantity extracted by the model is increased, which is important for improving the model performance.

In summary, the proposed GIALDN model shows good performance on a composite plate vibration signal dataset, which means that the data features are well learned.

3.2 comparison with other commonly used models

Because all the proposed network models are built by adopting convolutional neural networks, five common classical networks such as FaultNet, ResNet18, SEresnet18, VGGnet11 and DenseNet121 based on CNN are selected for comparison in the embodiment in order to verify the excellent performance of the GIALDN network, and the model performance is comprehensively evaluated by comparing the positioning accuracy, the model complexity, the training speed and the loss value of each model. In order to achieve a stable accuracy state for all 6 network models, the present embodiment has been trained for 1000 generations, although the network of the present embodiment has achieved very high accuracy already at 300 generations.

Under the same training strategy, the positioning accuracy of the test set of the five networks on the vibration signal data set of the composite plate is shown in fig. 7; meanwhile, the model performance, the parameter number, and the training time are counted in table 4. It can be seen that ResNet18, serenet 18 and DenseNet121 achieve good results, the accuracy rate exceeds 95%, which indicates that it is effective to overlap the number of layers of the convolutional network, and the model of this embodiment also integrates this advantage. Moreover, due to a series of innovative improvements provided by the embodiment, such as denoising and global interaction, the precision of the GIALDN is different from that of the three networks, the precision can reach more than 98.5%, and the network complexity and the training speed are improved, so that the provided threshold method can eliminate the characteristics related to noise in the signal data, the global interaction method can establish internal connection of the denoised signal data and dig out more useful characteristics, and the discrimination of the high-level characteristics of the output layer is stronger. Then, it can be seen in the figure that the accuracy of FauNet and VGGnet are both maintained at 80%, which indicates that they cannot effectively extract signal features for defect localization, and due to this, FauNet is that the number of network layers is too low, and VGGnet adopts a traditional and complex network architecture, which fully indicates that the number of network layers can improve feature extraction efficiency and effectiveness to a certain extent, but an excessively complex network can be against this, which is contrary to the idea of this embodiment, that is, while ensuring model feature extraction capability, other methods are used to enhance effective feature extraction, and finally the goal of network model lightweight and high accuracy is achieved. This is also reflected in the training speed, the GIALDN of this embodiment also has obvious advantages compared with other deep networks, the training time of the whole 1000 generation iteration process is less than 10 hours, and other deep networks are all over 20 hours, which makes its practicability relatively stronger.

TABLE 4 comparison of model Properties

To better compare the training efficiency and generalization performance of the models, test set cross entropy loss values for resnet18, serenet 18, densenet121, giandn are shown on fig. 8. It can be seen that although the first three achieve higher positioning accuracy, the resnet and serenet are lower than densenet in terms of the reduction of the loss value, which indicates that the network structure of residual connection can promote the parameter optimization of the deep network, so that the network structure has great advantage in data feature extraction. More importantly, the proposed GIALDN of this embodiment is superior to both resnet and serenet in terms of loss drop rate and eventually plateau loss value compared to classical resnet. Undoubtedly, in the training process, the denoising module and the global interaction module play roles, the denoising module eliminates noise related features and improves information discrimination, and the global interaction module deeply excavates useful information hidden in a multi-channel signal and increases feature quantity for positioning, so that the final high-grade features have discrimination, and the training efficiency of the whole model is improved.

In fig. 9 is a loss value curve for the proposed training and testing procedure of the giadn. It can be seen that the training loss curve is highly coincident with the test loss curve, with average losses at the last 150 stationary stages of 0.120 and 0.117, respectively. The fact that the generalization of the model is very strong and the overfitting phenomenon that the training set effect is good and the test set effect is far away does not occur is shown, on one hand, because the proposed network belongs to a lightweight network and the parameter quantity is small, the situation that the network can only accurately position the trained historical data and can not effectively position the newly-appeared defect position is avoided; on the other hand, the effectiveness of the network modules LDM and GIAM is also reflected, so that the network has strong information extraction capability. Of course, the present embodiment also improves the model generalization to some extent by providing the dropout layer to keep some neurons in a flexible state. In summary, GIALDN achieves better training results.

The above-mentioned embodiments are merely preferred embodiments for fully illustrating the present invention, and the scope of the present invention is not limited thereto. The equivalent substitution or change made by the technical personnel in the technical field on the basis of the invention is all within the protection scope of the invention. The protection scope of the invention is subject to the claims.

Claims (10)

1. An internal defect localization model based on a GIALDN network is characterized in that: the method comprises the following steps:

a remolding module: the device is used for remodeling the collected vibration signal data to obtain input signal characteristics;

a depoling pretreatment module: extracting vibration signal data rich in original information from the input signal characteristics by using the depoling preprocessing convolution layer to obtain preprocessing signal characteristics;

lightweight signal denoising module: the device is used for carrying out soft threshold denoising processing on the preprocessed signal characteristics to strip the noise signals to obtain denoised signal characteristics;

global interactive attention module: connecting each data point in the de-noising signal characteristic with each other, and establishing a relation between each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristic;

a multilayer convolution module: the depth feature extraction module is used for carrying out depth feature extraction on the signal features sequentially processed by the lightweight signal denoising module and the global interaction attention module to obtain depth signal features;

a result output module: and obtaining a defect positioning result from the depth signal characteristic by using a softmax function.

2. The GIALDN network-based internal defect localization model of claim 1, wherein: the method for the lightweight signal denoising module to perform soft threshold denoising processing is as follows:

wherein, y _i Representing the characteristic of a denoised signal of the ith channel; x is the number of _i Representing the preprocessed signal features of the ith channel; tau is _i Representing the ith channel soft threshold.

3. The GIALDN network-based internal defect localization model of claim 2, wherein: the lightweight signal denoising module comprises a global mixing pooling layer and a one-dimensional convolution layer;

the global mixed pooling layer is used for simultaneously carrying out global maximum pooling and global average pooling on the preprocessed signal features to obtain mixed signal features;

the one-dimensional convolutional layer is used for extracting channel characteristics of mixed signal characteristics, and the learned coefficients are scaled to the range of [0,1] by using a sigmoid activation function to obtain a soft threshold coefficient sigma of each channel; then:

τ _i ＝σ _i ×|x|

wherein, tau _i Representing a soft threshold corresponding to the ith channel; sigma _i Representing the soft threshold coefficient corresponding to the ith channel; and | x | represents the preprocessed signal features after absolute value processing.

4. The GIALDN network-based internal defect localization model of claim 1, wherein: the method for processing the de-noised signal features by the global interaction attention module is as follows:

(1) interpolating the denoised signal characteristic X by adopting an interplate function, reducing the dimensionality of the characteristic graph from h number pair to (h/2) X (w/2), and obtaining the characteristic graph X _i ；

(2) Feature map X is represented by a convolution layer with batch normalized BN layer, Relu activation function _i The number of channels is changed from c to (h/2) X (w/2), and a characteristic diagram X is obtained _c ；

(3) For feature X _c Performing global interaction to generate an attention diagram A _a ；

(4) Attention-seeking diagram A _a And feature map X _i Carrying out point-by-point multiplication to obtain a global interaction attention feature map A _c ；

(5) Interpolate function pair global interactive attention feature map A _c Performing interpolation operation to obtain a global interactive attention feature map A _c Reducing the characteristic diagram A into the same shape as the characteristic X of the de-noised signal _i ；

(6) Using the residual block to obtain the final output characteristic diagram as follows:

Y＝X+A _i 。

5. the GIALDN network-based internal defect localization model of claim 4, wherein: in the step (3), the characteristic X is measured _c Global interaction to generate an attention map A _a The method comprises the following steps:

(31) will be characterized by X _c Generating a data point feature map with the size of 1 multiplied by 1 and the number of channels of (h/2) multiplied by (w/2) in all channels by each data point, and then reshaping the data feature map into a single-channel feature map with the size of (h/2) multiplied by (w/2) and the number of channels of 1;

(32) aggregating the single-channel characteristic graphs of all the data points to obtain an intermediate characteristic graph;

(33) generating an attention map A from the intermediate feature map by using a softmax function _a 。

6. A GIALDN network-based internal defect location method is characterized in that: the method comprises the following steps:

the method comprises the following steps: collecting data: driving the detected object to vibrate, converting vibration signals of different areas of the detected object into electric signals by using a plurality of sensors, and converting the electric signals into digital signals by using an A/D converter to obtain vibration signal data;

step two: remodeling: reshaping vibration signal data to obtain input signal characteristics;

step three: removing a pool for pretreatment: extracting vibration signal data rich in original information from the input signal characteristics by using the depoling preprocessing convolution layer to obtain preprocessing signal characteristics;

step four: denoising lightweight signals: carrying out soft threshold denoising processing on the preprocessed signal characteristics to strip a noise signal to obtain denoised signal characteristics;

step five: global interactive attention processing: connecting each data point in the de-noising signal characteristic with each other, and establishing a relation among each channel so as to extract a potential relation between a long-distance relation and cross-channel data in the signal and improve the discrimination of the characteristic;

step six: multilayer convolution processing: carrying out depth feature extraction on the signal features subjected to the lightweight signal denoising processing and the global interaction attention processing in sequence to obtain depth signal features;

step seven: and outputting a result: and obtaining a defect positioning result from the depth signal characteristic by using a softmax function.

7. The GIALDN network-based internal defect location method of claim 6, wherein: in the fourth step, the method for denoising the lightweight signal comprises the following steps:

41) performing global maximum pooling and global average pooling on the preprocessed signal features by using a global mixed pooling layer to obtain mixed signal features;

42) extracting channel characteristics of mixed signal characteristics by using the one-dimensional convolutional layer, and adopting a sigmoid activation function to scale the learned coefficient to the range of [0,1] to obtain a soft threshold coefficient sigma of each channel; then:

τ _i ＝σ _i ×|x _i |

wherein, tau _i Representing a soft threshold corresponding to the ith channel; sigma _i Representing the soft threshold coefficient corresponding to the ith channel; | x _i L represents the preprocessed signal characteristic of the ith channel after absolute value conversion;

43) carrying out soft threshold denoising processing on the preprocessed signal characteristics:

wherein, y _i Representing the characteristic of a denoised signal of the ith channel; x is the number of _i Representing the preprocessed signal features of the ith channel; tau is _i Representing the ith channel soft threshold.

8. The GIALDN network-based internal defect location method of claim 6, wherein: in the fifth step, the method for processing the global interactive attention is as follows:

(1) interpolating the denoised signal characteristic X by adopting an interplate function, reducing the dimensionality of the characteristic graph from h number pair to (h/2) X (w/2), and obtaining the characteristic graph X _i ；

(2) Feature map X is represented by a convolution layer with batch normalized BN layer, Relu activation function _i The number of channels is changed from c to (h/2) X (w/2), and a characteristic diagram X is obtained _c ；

(3) For feature X _c Performing global interaction to generate an attention diagram A _a ；

(4) Attention-seeking diagram A _a And feature map X _i Carrying out point-by-point multiplication to obtain a global interactive attention feature map A _c ；

(5) Interpolate function pair global interactive attention feature map A _c Performing interpolation operation to obtain a global interactive attention feature map A _c Reducing the characteristic diagram A into the same shape as the characteristic X of the de-noised signal _i ；

(6) Using the residual block to obtain the final output characteristic diagram as follows:

Y＝X+A _i 。

9. the method of claim 8, wherein the method comprises: in the step (3), the feature X is subjected to _c Global interaction to generate an attention map A _a The method comprises the following steps:

(31) will be characterized by X _c Generating a data point feature map with the size of 1 multiplied by 1 and the number of channels of (h/2) multiplied by (w/2) in all channels for each data point, and then reshaping the data feature map into a single-channel feature map with the size of (h/2) multiplied by (w/2) and the number of channels of 1;

(32) aggregating the single-channel characteristic graphs of all the data points to obtain an intermediate characteristic graph;

(33) generating an attention map A from the intermediate feature map by using a softmax function _a 。

10. The GIALDN network-based internal defect location method of claim 6, wherein: the object to be detected is a CFRP laminated plate, the sensors are two groups with the same number, and the two groups of sensors are respectively and uniformly distributed on the inner layer and the outer layer of the CFRP laminated plate; in the first step, the data acquisition method comprises the following steps:

the sinusoidal digital signal generated by the signal generator is converted into an electric signal by a D/A converter, the amplitude is controlled by a power amplifier, and finally the electric signal is transmitted to a vibration exciter to control the generation of physical vibration;

the vibration of the vibration exciter acts on the CFRP laminated plate through the vibration exciting rod to enable the CFRP laminated plate to vibrate;

the vibration signals of different areas of the detected object are converted into electric signals by using the two groups of sensors, and the electric signals are converted into digital signals by using the A/D converter, so that vibration signal data are obtained.

Images