Defect detection method, device, equipment and medium for metal disk parts

Abstract

The present application relates to the technical field of metal defect detection, and specifically to a defect detection method, device, equipment and medium for metal disc parts. The method includes: obtaining an ultrasonic phased array signal of a target metal disc part and preprocessing it to obtain a preprocessed ultrasonic phased array signal, wherein the ultrasonic phased array signal includes a multi-channel sound wave signal collected at different angles from different detection surfaces of the metal disc part; using a pre-built CNN‑BiLSTM-based feature extraction model to obtain a multi-dimensional feature that integrates the time domain and the spatial domain, and using a classifier to determine the defect detection result. It solves the problem of being able to accurately identify and distinguish different types of defects in the scenario of processing duplicate defect combinations, achieves the purpose of improving the detection accuracy of deep or hidden defects, can provide accurate fault location and repair basis for operation and maintenance inspection personnel, effectively identify hidden defects of disc parts in mechanical devices, and avoid the occurrence of major safety accidents.

Description

Technical Field

[0001] The present application relates to the technical field of metal defect detection, and in particular to a defect detection method, device, equipment and medium for metal disc parts. Background Art

[0002] Metal discs are crucial components for ensuring stable train operation. During train operation, the discs are subjected to significant dynamic loads. The press-fitted parts of the discs, under various combined external loads, experience compression, deformation, and shear stresses, leading to fatigue cracks. These components are typical fatigue components. Once the discs are damaged, fatigue damage propagates under alternating stresses, potentially leading to derailment and significant loss of life and property.

[0003] The relevant defect detection technologies mainly have the following problems: First, in complex signal processing and defect feature identification, they rely on the inspector's operating experience and judgment. Due to the large number of factors of human intervention, this may lead to highly subjective detection results and prone to misjudgment or omission. In addition, this method is not only time-consuming, but also has bottlenecks in detection efficiency and cannot meet the needs of large-scale and rapid detection; second, it cannot handle recurring defect combinations, and it is difficult to accurately identify and distinguish different types of defects; third, the detection accuracy is poor when facing deep or hidden defects. Summary of the Invention

[0004] In view of this, one of the technical problems solved by the embodiments of the present application is to provide a defect detection method, device, electronic device and storage medium for metal disk parts to solve the problems of poor detection accuracy and high manual dependence in defect detection of metal disk parts.

[0005] In a first aspect of an embodiment of the present application, a defect detection method for a metal plate-type part is disclosed, the method comprising:

[0006] Acquiring and preprocessing the ultrasonic phased array signal of the target metal disk part to obtain a preprocessed ultrasonic phased array signal, wherein the ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk part;

[0007] The pre-built CNN-BiLSTM-based feature extraction model is used to extract features from the pre-processed ultrasonic phased array signal to obtain multi-dimensional features that fuse the time domain and spatial domain.

[0008] The multidimensional features are input into the classifier to obtain the defect detection results of the target metal plate parts.

[0009] In a second aspect of an embodiment of the present application, a defect detection device for metal plate parts is disclosed, the device comprising:

[0010] A data processing module is used to obtain and preprocess the ultrasonic phased array signal of the target metal disk part to obtain a preprocessed ultrasonic phased array signal. The ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk part;

[0011] The feature extraction module is used to extract features from the pre-processed ultrasonic phased array signal using a pre-built CNN-BiLSTM-based feature extraction model to obtain multi-dimensional features that fuse the time domain and spatial domain;

[0012] The defect detection module is used to input multi-dimensional features into the classifier to obtain defect detection results of target metal plate parts.

[0013] In a third aspect of an embodiment of the present application, an electronic device is disclosed. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0014] In a fourth aspect of the embodiments of the present application, a computer-readable storage medium is disclosed, which stores a computer program. When the computer program is executed by a processor, the steps of the above method are implemented.

[0015] The present application has the following advantages: by acquiring the ultrasonic phased array signal of the target metal disk-like part and preprocessing it, a preprocessed ultrasonic phased array signal is obtained, and the ultrasonic phased array signal includes multi-channel sound wave signals collected at different angles by different detection surfaces of the metal disk-like part. The pre-built CNN-BiLSTM-based feature extraction model is used to extract features from the preprocessed ultrasonic phased array signal to obtain multi-dimensional features that integrate the time domain and the spatial domain, and then the multi-dimensional features are input into the classifier to obtain the defect detection results of the target metal disk-like part. This combination of phased array ultrasonic technology and deep learning technology represents different defects in multi-dimensional terms according to space and time, which plays a role in accurately identifying and distinguishing different types of defects in the scenario of processing replicated defect combinations, thereby achieving the purpose of improving the detection accuracy of deep or hidden defects, and can provide accurate fault location and repair basis for operation and maintenance inspection personnel, effectively identify hidden defects of disk-like parts in mechanical devices, and avoid major safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic flow chart of a defect detection method for metal disk parts provided in one embodiment of the present application;

[0017] Figure 2 A schematic diagram of a training method for defect detection of metal disk parts provided in one embodiment of the present application;

[0018] Figure 3A schematic structural diagram of a defect detection device for metal disk parts provided in one embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0020] It should be noted that although the functional modules are divided in the device schematic and the logical order is shown in the flowchart, in some cases, the steps shown or described can be performed in a different order than the module division in the device or the order in the flowchart.

[0021] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0022] According to one embodiment of the present application, a defect detection method for metal disk parts is provided, such as Figure 1 As shown, the method includes step S101, step S102 and step S103.

[0023] Step S101: Acquire and preprocess the ultrasonic phased array signal of the target metal disk part to obtain a preprocessed ultrasonic phased array signal. The ultrasonic phased array signal includes multi-channel acoustic wave signals collected from different detection surfaces of the metal disk part at different angles.

[0024] Specifically, the automatic defect detection device acquires and preprocesses multi-channel ultrasonic phased array signals from metal disc-like parts, generating multi-channel preprocessed ultrasonic phased array signals. More specifically, the automatic defect detection device is typically a device (such as a PC) connected to an ultrasonic phased array signal acquisition device. During use, the ultrasonic phased array signal acquisition device can be controlled to generate functional acoustic waves toward the metal disc-like part, allowing the phased array signal acquisition device to capture the acoustic waves reflected from the metal disc-like part and perform preprocessing operations such as amplification, filtering, delay, and analog-to-digital conversion.

[0025] Specifically, the preprocessing of the ultrasonic phased array signal may include filtering, denoising, missing value processing, and the like.

[0026] Step S102: Using a pre-built CNN-BiLSTM-based feature extraction model to extract features from the pre-processed ultrasonic phased array signal, a multi-dimensional feature integrating the time domain and the spatial domain is obtained.

[0027] Step S103: input the multi-dimensional features into the classifier to obtain the defect detection results of the target metal plate part.

[0028] Specifically, the defect detection results may include the defect type (such as crack, fatigue, wear, spalling, pitting, etc.) and the specific location of the defect.

[0029] The embodiment of the present application obtains an ultrasonic phased array signal of a target metal disk-like part and preprocesses it to obtain a preprocessed ultrasonic phased array signal. The ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk-like part. A pre-built feature extraction model based on CNN-BiLSTM is used to extract features from the preprocessed ultrasonic phased array signal to obtain multi-dimensional features that fuse the time domain and spatial domain. The multi-dimensional features are then input into a classifier to obtain a defect detection result for the target metal disk-like part. This combination of phased array ultrasonic technology and deep learning technology represents different defects in a multi-dimensional manner according to space and time, which plays a role in accurately identifying and distinguishing different types of defects in processing replicated defect combination scenarios, thereby achieving the purpose of improving the detection accuracy of deep or hidden defects, and can provide operation and maintenance inspection personnel with accurate fault location and repair basis, effectively identify hidden defects of disk-like parts in mechanical devices, and avoid major safety accidents.

[0030] Furthermore, the feature extraction model includes a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and a fusion fully connected layer. The outputs of the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM are connected to the fusion fully connected layer through the CNN fully connected layer and the BiLSTM fully connected layer respectively. Step S102 further includes: inputting the preprocessed ultrasonic phased array signal into the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM for feature extraction, flattening the extracted features using the CNN fully connected layer and the BiLSTM fully connected layer respectively to obtain a one-dimensional spatial feature map and a time feature map, and splicing the spatial feature map and the time feature map using the fusion fully connected layer to obtain a multi-dimensional fusion feature map that integrates the time domain and the spatial domain. The convolutional neural network (CNN) performs convolution processing on the preprocessed ultrasonic phased array signal and then uses the CNN fully connected layer to perform spatial dimensionality reduction to reduce the amount of computation and the risk of overfitting. The bidirectional long short-term memory network (BiLSTM) integrates the extracted time series features into global features through the BiLSTM fully connected layer. Finally, the fusion fully connected layer is used to fuse the outputs of the CNN fully connected layer and the BiLSTM fully connected layer to obtain multidimensional features that combine space and time.

[0031] Specifically, a convolutional neural network CNN generally includes an input layer and a convolutional layer. The number of convolutional layers can be set as needed, and is generally set to 3 or 5 layers. When applied, the number of convolutional layers included in the convolutional neural network CNN can be determined based on the electronic device to which the present application is applied. When applied, the number of convolutional layers corresponding to each of the multiple device configurations can be provided to determine the number of convolutional layers corresponding to the target detection device configuration. By providing multiple convolutional neural networks CNN, the problem of computational cost and the impact on the operating speed of the device can be considered.

[0032] In some embodiments, before step S102, the method further includes:

[0033] The preprocessed ultrasonic phased array signal is converted into a phased array ultrasound image, so that the phased array ultrasound image is used as the input of the convolutional neural network (CNN) and the bidirectional long short-term memory network.

[0034] Specifically, the pre-processed ultrasonic phased array signal can be converted into an image through image reconstruction technology, and then the image can be convolved using a convolutional neural network (CNN). More specifically, after the phased array ultrasonic image is converted, candidate defect areas can be extracted from the phased array ultrasonic image. For example, the candidate areas can be determined by a user's marking operation, or after calculating the pixel distribution of the phased array ultrasonic image, the areas in the pixel distribution with pixel values ​​greater than a preset pixel value threshold are determined as candidate areas, and the convolutional neural network (CNN) is used to convolve the candidate defect areas.

[0035] Furthermore, the regression loss function loss formula used by the classifier is as follows:

[0036] ;

[0037] ;

[0038] ;

[0039] in, Adjust the parameters for the loss function, Predict values ​​for the model, is the true value of the data, is the Smooth L1 splicing parameter, which is used to adjust the shape of the Smooth L1 near-zero value area. n is the data length.

[0040] In some embodiments, before step S102, the method further includes:

[0041] Acquire sample data, the sample data including different metal plate parts and multi-channel acoustic wave signals collected at different angles from different detection surfaces corresponding to the different metal plate parts;

[0042] After defect labeling, filtering and denoising, data augmentation, and random perturbation processing on the sample data, the ultrasonic phased array signal to be trained is obtained;

[0043] The initial feature extraction model and the initial classifier constructed based on CNN-BiLSTM are trained using the ultrasonic phased array signal to be trained until the extracted multi-dimensional fusion feature map can enable the classifier to converge during the training process.

[0044] Specifically, after completing preprocessing procedures such as denoising and normalization, features are extracted and defects are identified. If the selected point is greater than 5 times the mean value of the feature value, it is judged as a defect.

[0045] Specifically, defect annotation can include defect type annotation and defect location annotation. Specifically, a pre-set filter can be used to perform filtering and denoising. More specifically, the pre-set filter can be obtained by superimposing a Gaussian filter with an FIR digital filter. When applied, the pre-set filter can be obtained by weighted summation of the Gaussian filter and the FIR digital filter.

[0046] Furthermore, the filtering denoising algorithm uses a Gaussian filter operator superimposed on a FIR digital filter operator, and the formula is:

[0047] ;

[0048] x is the original signal, is the filtered signal, is the standard deviation of the Gaussian distribution, is the filter coefficient, N is the FIR filter operator order, is an adaptive parameter related to the variance of the original signal, , The operator weights are adaptively adjusted as the variance of the original data changes.

[0049] Specifically, This method is used to adjust the contribution ratio of the FIR digital filter relative to the Gaussian filter. This application combines the smoothing properties of the Gaussian filter with the flexibility of the FIR digital filter, allowing the FIR digital filter to be tailored to the specific frequency response or signal processing task. This hybrid approach simultaneously achieves noise suppression (via Gaussian smoothing) and precise signal shaping (via FIR filtering) for ultrasonic phased array signals.

[0050] Furthermore, the random perturbation formula is:

[0051] ; Among them, rand(x) represents a function that generates random numbers in the range [-x,x], Represents the value of any channel after data augmentation processing, mean( ) represents the average value of multiple channels processed by data augmentation, The value of any channel after perturbation after data augmentation. By adding random perturbations to improve the diversity of the training dataset and the generalization of the trained model, and to address the differences and lack of diversity in the defect data generated by data augmentation, random perturbations are added to all replicated defect data, and each value in the replicated defect signal is randomly perturbed.

[0052] The following combination Figure 2 The present application is described. The diagram shows that the present application includes multiple links in deep learning, including data acquisition, data preprocessing, model selection, target identification, model training, model evaluation and optimization, simulation verification, and system integration. First, the data source comes from the phased array signal data acquisition device, that is, the phased array signal data acquisition device obtains array ultrasonic data of different metal disk parts, specifically including multi-channel acoustic wave signals collected by different detection surfaces at different angles. The collected phased array ultrasonic signals usually contain a large amount of noise and redundant information, so the data preprocessing work converts the data format, file format, etc., and completes data labeling, data augmentation, denoising, normalization and other processing work. Data labeling includes the labeling of different types of defects such as cracks, wear, and spalling. The present application uses two models based on convolutional neural network (CNN) and long short-term memory network (LSTM) to complete deep learning of data features. After training, the model can learn and identify different types of defects through supervised learning on a large amount of labeled data. The recognition process includes two parts: classification and positioning. The model is evaluated by calculating loss functions (such as cross-entropy loss and mean squared error) to assess the difference between predicted and true values, completing model evaluation and optimization. Simulations are then used to verify the model's performance under different defect types and signal noise conditions, ensuring its stability and robustness in real-world environments. Finally, the model is integrated into the inspection equipment's control system as needed to automatically detect and provide feedback on surface or internal defects.

[0053] One embodiment of the present application provides a defect detection device for metal disk parts, such as Figure 3 As shown, the device 30 includes: a data processing module 301 , a feature extraction module 302 and a defect detection module 303 .

[0054] The data processing module 301 is used to obtain and preprocess the ultrasonic phased array signal of the target metal disk part to obtain a preprocessed ultrasonic phased array signal. The ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk part.

[0055] A feature extraction module 302 is configured to extract features from the preprocessed ultrasonic phased array signal using a pre-built CNN-BiLSTM-based feature extraction model to obtain multi-dimensional features that fuse the time domain and the spatial domain;

[0056] The defect detection module 303 is used to input the multi-dimensional features into the classifier to obtain the defect detection results of the target metal plate part.

[0057] The embodiment of the present application obtains an ultrasonic phased array signal of a target metal disk-like part and preprocesses it to obtain a preprocessed ultrasonic phased array signal. The ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk-like part. A pre-built feature extraction model based on CNN-BiLSTM is used to extract features from the preprocessed ultrasonic phased array signal to obtain multi-dimensional features that fuse the time domain and spatial domain. The multi-dimensional features are then input into a classifier to obtain a defect detection result for the target metal disk-like part. This combination of phased array ultrasonic technology and deep learning technology represents different defects in a multi-dimensional manner according to space and time, which plays a role in accurately identifying and distinguishing different types of defects in processing replicated defect combination scenarios, thereby achieving the purpose of improving the detection accuracy of deep or hidden defects, and can provide operation and maintenance inspection personnel with accurate fault location and repair basis, effectively identify hidden defects of disk-like parts in mechanical devices, and avoid major safety accidents.

[0058] Furthermore, the feature extraction model includes a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and a fusion fully connected layer. The outputs of the convolutional neural network CNN and the bidirectional long short-term memory network BiLSTM are connected to the fusion fully connected layer through the CNN fully connected layer and the BiLSTM fully connected layer respectively. The feature extraction module is used to: input the preprocessed ultrasonic phased array signal into the convolutional neural network CNN and the bidirectional long short-term memory network for feature extraction respectively, flatten the extracted features respectively using the CNN fully connected layer and the BiLSTM fully connected layer to obtain a one-dimensional spatial feature map and a temporal feature map, and use the fusion fully connected layer to splice the spatial feature map and the temporal feature map to obtain a multi-dimensional fusion feature map that integrates the time domain and the spatial domain.

[0059] Furthermore, before the preprocessed ultrasonic phased array signal is input into the convolutional neural network CNN and the bidirectional long short-term memory network for feature extraction, the feature extraction module is also used to: convert the preprocessed ultrasonic phased array signal into a phased array ultrasonic image, so as to use the phased array ultrasonic image as input to the convolutional neural network CNN and the bidirectional long short-term memory network.

[0060] Furthermore, the regression loss function loss formula used by the classifier is as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] in, Adjust the parameters for the loss function, Predict values ​​for the model, is the true value of the data, is the Smooth L1 splicing parameter, which is used to adjust the shape of the Smooth L1 near-zero value area. n is the data length.

[0065] Furthermore, before the feature extraction module uses the pre-built CNN-BiLSTM-based feature extraction model to extract features from the pre-processed ultrasonic phased array signal to obtain multi-dimensional features that fuse the time domain and the spatial domain, it also includes:

[0066] The sample acquisition submodule is used to acquire sample data, which includes multi-channel acoustic wave signals collected at different angles from different metal plate parts and different detection surfaces corresponding to different metal plate parts;

[0067] The sample preprocessing submodule is used to perform defect labeling, filtering and denoising, data augmentation, and random perturbation processing on the sample data to obtain the ultrasonic phased array signal to be trained;

[0068] The model training submodule is used to train the initial feature extraction model and the initial classifier built based on CNN-BiLSTM using the ultrasonic phased array signal to be trained, until the extracted multi-dimensional fusion feature map can enable the classifier to converge during the training process.

[0069] Furthermore, the filtering denoising algorithm uses a Gaussian filter operator superimposed on a FIR digital filter operator, and the formula is:

[0070] ;

[0071] x is the original signal, is the filtered signal, is the standard deviation of the Gaussian distribution, is the filter coefficient, N is the FIR filter operator order, is an adaptive parameter related to the variance of the original signal, , The operator weights are adaptively adjusted as the variance of the original data changes.

[0072] Furthermore, the random perturbation formula is:

[0073] ; Among them, rand(x) represents a function that generates random numbers in the range [-x,x], Represents the value of any channel after data augmentation processing, mean( ) represents the average value of multiple channels processed by data augmentation, It is the value after perturbation and enhancement of any channel after data augmentation processing.

[0074] The defect detection device for metal disk-type parts of this embodiment can execute the defect detection method for metal disk-type parts provided in the embodiment of this application. The implementation principle is similar and will not be repeated here.

[0075] Yet another embodiment of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the above method.

[0076] Specifically, a processor may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, a transistor logic device, a hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. A processor may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0077] Specifically, the processor is connected to the memory via a bus. The bus may include a path for transmitting information. The bus may be a PCI bus or an EISA bus. The bus may be divided into an address bus, a data bus, a control bus, etc.

[0078] The memory may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an EEPROM, CD-ROM or other optical disk storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to these.

[0079] Optionally, the memory is used to store the code of the computer program that executes the solution of the present application, and the execution is controlled by the processor. The processor is used to execute the application program code stored in the memory to implement the actions of the device provided by the above embodiment.

[0080] Another embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the above method when executed by a processor.

[0081] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0082] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage, or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0083] The above is a specific description of the preferred implementation of the present application, but the present application is not limited to the above implementation mode. Technical personnel familiar with the field can also make various equivalent modifications or substitutions without violating the spirit of the present application. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present application.

Claims

1. A method for detecting defects in metal disk parts, characterized in that: include: Acquiring and preprocessing an ultrasonic phased array signal of a target metal disk-like part to obtain a preprocessed ultrasonic phased array signal, wherein the ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk-like part; Using a pre-built CNN-BiLSTM-based feature extraction model to extract features from the pre-processed ultrasonic phased array signal, a multi-dimensional feature that integrates time domain and spatial domain is obtained; The feature extraction model includes a convolutional neural network (CNN), a bidirectional long short-term memory (BiLSTM) network, and a fusion fully connected layer. The outputs of the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network are connected to the fusion fully connected layer through the CNN fully connected layer and the BiLSTM fully connected layer, respectively. The method uses a pre-built CNN-BiLSTM-based feature extraction model to extract features from the pre-processed ultrasonic phased array signal to obtain multi-dimensional features that fuse the time domain and the spatial domain, including: inputting the pre-processed ultrasonic phased array signal into the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network for feature extraction, flattening the extracted features using the CNN fully connected layer and the BiLSTM fully connected layer to obtain a one-dimensional spatial feature map and a temporal feature map, and splicing the spatial feature map and the temporal feature map using the fusion fully connected layer to obtain a multi-dimensional fusion feature map that fuses the time domain and the spatial domain. The multidimensional features are input into a pre-built classifier to obtain the defect detection results of the target metal plate parts; the regression loss function used by the classifier is the mean square error loss function And improved Smooth L1 loss function The combined formula is as follows: ； ； ； in, Adjust the parameters for the loss function, Predict values ​​for the model, is the true value of the data, is the SmoothL1 splicing parameter, which is used to adjust the shape of the Smooth L1 near-zero value area. n is the data length.

2. The method according to claim 1, characterized in that Before inputting the pre-processed ultrasonic phased array signal into the convolutional neural network (CNN) and the bidirectional long short-term memory (LSTM) network for feature extraction, the method further includes: The preprocessed ultrasonic phased array signal is converted into a phased array ultrasonic image, so that the phased array ultrasonic image is used as an input of the convolutional neural network (CNN) and the bidirectional long short-term memory network.

3. The method according to claim 1, characterized in that Before extracting features from the preprocessed ultrasonic phased array signal using the pre-built CNN-BiLSTM-based feature extraction model to obtain multi-dimensional features that fuse the time domain and the spatial domain, the method further includes: Acquiring sample data, the sample data including different metal disk parts and multi-channel acoustic wave signals collected at different angles by different detection surfaces corresponding to the different metal disk parts; After performing defect labeling, filtering and denoising, data augmentation, and random perturbation processing on the sample data, an ultrasonic phased array signal to be trained is obtained; The ultrasonic phased array signal to be trained is used to train an initial feature extraction model and an initial classifier constructed based on CNN-BiLSTM until the extracted multidimensional fusion feature map enables the classifier to converge during the training process.

4. The method according to claim 3, characterized in that The filtering denoising algorithm uses a Gaussian filter operator superimposed on a FIR digital filter operator. The formula is: ； x is the original signal, is the filtered signal, is the standard deviation of the Gaussian distribution, is the filter coefficient, N is the FIR filter operator order, is an adaptive parameter related to the variance of the original signal, , The operator weights are adaptively adjusted as the variance of the original data changes.

5. The method according to claim 3, characterized in that The random perturbation formula is: ; Among them, rand(x) represents a function that generates random numbers in the range [-x,x], Represents the value of any channel after data augmentation processing, mean( ) represents the average value of multiple channels processed by data augmentation, It is the value after perturbation and enhancement of any channel after data augmentation processing.

6. A defect detection device for metal plate parts, characterized in that: include: a data processing module, configured to acquire and preprocess the ultrasonic phased array signal of the target metal disk-like part to obtain a preprocessed ultrasonic phased array signal, wherein the ultrasonic phased array signal includes multi-channel acoustic wave signals collected at different angles from different detection surfaces of the metal disk-like part; A feature extraction module is used to extract features from the preprocessed ultrasonic phased array signal using a pre-built CNN-BiLSTM-based feature extraction model to obtain multi-dimensional features that fuse time domain and spatial domain; The feature extraction model includes a convolutional neural network (CNN), a bidirectional long short-term memory (BiLSTM) network, and a fusion fully connected layer. The outputs of the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network are connected to the fusion fully connected layer through the CNN fully connected layer and the BiLSTM fully connected layer, respectively. The feature extraction module is used to: input the preprocessed ultrasonic phased array signal into the convolutional neural network (CNN) and the bidirectional long short-term memory (BiLSTM) network for feature extraction, flatten the extracted features using the CNN fully connected layer and the BiLSTM fully connected layer, respectively, to obtain a one-dimensional spatial feature map and a temporal feature map, and then use the fusion fully connected layer to splice the spatial feature map and the temporal feature map to obtain a multidimensional fusion feature map that integrates the time domain and the spatial domain. The defect detection module is used to input the multidimensional features into the classifier to obtain the defect detection results of the target metal plate parts; the regression loss function loss formula used by the classifier is as follows: ； ； ； in, Adjust the parameters for the loss function, Predict values ​​for the model, is the true value of the data, It is the SmoothL1 splicing parameter, which is used to adjust the shape of the Smooth L1 near-zero value area. n is the data length.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images