Artificial intelligence-based steel structure weld quality detection method and device

Abstract

The application relates to the technical field of artificial intelligence, in particular to a steel structure weld quality detection method and device based on artificial intelligence; the method comprises the following steps: performing mechanical property testing on a target weld with to-be-detected quality in a target steel structure body for a target duration, so as to obtain target strain data of the target weld; preprocessing the target strain data to obtain a target time-frequency matrix; inputting the target time-frequency matrix into a pre-trained quality detection model, so that the quality detection model determines target propagation feature data of a target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target weld according to the target propagation feature data; wherein the strain wave indicates that the deformation generated by the weld propagates in the form of a wave in the steel structure body; the application can solve the technical problems that the existing weld quality detection method has limited identification of hidden defects and has detection blind spots and artificial errors.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a steel structure weld quality detection method and device based on artificial intelligence. BACKGROUND

[0002] In a welding process, after two steel structure members are firmly connected together through the welding process, a weld will be generated between the two steel structures; the welded steel structures can jointly bear various loads such as tension, pressure, shear and bending moment; wherein the quality of the weld directly affects the overall performance, safety and service life of the steel structure.

[0003] In the prior art, the quality of the steel structure weld can be detected by methods such as ultrasonic, X-ray and magnetic powder, but due to the characteristics of small size, various forms and complex evolution process of steel structure weld defects, the above-mentioned prior art has limited recognition of early micro-cracks, incomplete fusion, porosity and other hidden defects, and the above-mentioned prior art is limited by operating environment, detection personnel experience and sensor distribution density, and has detection blind area and artificial error. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a steel structure weld quality detection method and device based on artificial intelligence to solve the technical problems of limited recognition of hidden defects and detection blind area and artificial error in the existing weld quality detection method.

[0005] In a first aspect, the present application provides a steel structure weld quality detection method based on artificial intelligence, which comprises:

[0006] Performing a mechanical property test on a target weld of a target steel structure to be detected for a target duration to obtain target strain data of the target weld;

[0007] Wherein, the strain data indicates the deformation characteristics of the weld within the target duration;

[0008] Preprocessing the target strain data to obtain a target time-frequency matrix;

[0009] Inputting the target time-frequency matrix into a pre-trained quality detection model, so that the quality detection model determines target propagation feature data of a target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target weld according to the target propagation feature data;

[0010] Wherein, the strain wave indicates that the deformation of the weld propagates in the form of a wave in the steel structure.

[0011] In a second aspect, the application provides a steel structure weld quality detection device based on artificial intelligence, which comprises a data acquisition module and a quality detection module.

[0012] The data acquisition module is configured to perform a mechanical property test on a target weld in a target steel structure to be detected for a target duration, so as to obtain target strain data of the target weld.

[0013] The strain data indicates the deformation characteristics of the weld within the target duration.

[0014] The quality detection module is configured to preprocess the target strain data to obtain a target time-frequency matrix.

[0015] The target time-frequency matrix is input into a pre-trained quality detection model, so that the quality detection model determines target propagation characteristic data of a target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target weld according to the target propagation characteristic data.

[0016] The strain wave indicates that the deformation of the weld propagates in the form of a wave in the steel structure.

[0017] Advantages:

[0018] The application provides a steel structure weld quality detection method based on artificial intelligence, which comprises: performing a mechanical property test on a target weld in a target steel structure to be detected for a target duration, so as to obtain target strain data of the target weld; wherein the strain data indicates the deformation characteristics of the weld within the target duration; preprocessing the target strain data to obtain a target time-frequency matrix; inputting the target time-frequency matrix into a pre-trained quality detection model, so that the quality detection model determines target propagation characteristic data of a target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target weld according to the target propagation characteristic data; wherein the strain wave indicates that the deformation of the weld propagates in the form of a wave in the steel structure; by constructing the steel structure weld quality detection method based on artificial intelligence, the application can identify the hidden defects in the target weld and avoid detection blind spots and human errors through an automatic identification process, thereby improving the detection quality and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments of the application. The following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0020] Figure 1 A flowchart of the steel structure weld quality detection method based on artificial intelligence provided by the embodiment of the present application is shown in the figure.

[0021] Figure 2 A data processing flowchart of the convolution layer provided by the embodiment of the present application is shown in the figure.

[0022] Figure 3 A data processing flowchart of the gating recurrent unit provided by the embodiment of the present application is shown in the figure.

[0023] Figure 4 A data processing flowchart of the regularization layer and the fully connected layer provided by the embodiment of the present application is shown in the figure.

[0024] Figure 5 A comparison chart of effects of different noise suppression methods provided by the embodiment of the present application is shown in the figure.

[0025] Figure 6 A comparison chart of defect detection rates of different noise suppression methods provided by the embodiment of the present application is shown in the figure.

[0026] Figure 7 A crack size detection error schematic diagram of different strain-stress conversion methods provided by the embodiment of the present application is shown in the figure.

[0027] Figure 8 A comparison chart of time-frequency domain feature extraction methods provided by the embodiment of the present application is shown in the figure.

[0028] Figure 9 A ROC curve comparison chart of different quality prediction models provided by the embodiment of the present application is shown in the figure.

[0029] Figure 10 A structure schematic diagram of the steel structure weld quality detection device based on artificial intelligence provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0030] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0031] First, the present application provides a steel structure weld quality detection method based on artificial intelligence, as shown in the figure. Figure 1 Figure 1 A flowchart of the steel structure weld quality detection method based on artificial intelligence provided by the embodiment of the present application is shown in the figure. ​

[0032] S100: Perform a mechanical property test on a target weld seam of a target steel structure to be detected for a target duration to obtain target strain data of the target weld seam.

[0033] The strain data indicates the deformation characteristics of the weld seam within the target duration.

[0034] Specifically, in the embodiments of the present application, the "target steel structure" is a steel structure obtained by welding at least two steel structures, and at least one weld seam exists in the target steel structure; the "target weld seam" is a weld seam that needs to be detected for quality; the "mechanical property test" is a test of measuring and analyzing various mechanical responses of the weld seam under stress, and the strain data obtained by the mechanical property test can be collected by a high-precision resistance strain gauge array or a distributed optical fiber sensor network installed in the target steel structure; the "quality detection" refers to the monitoring of whether there is a defect in the target weld seam.

[0035] In actual operation, after the target weld seam is determined in the target steel structure, a mechanical property test can be performed on the target weld seam to obtain target strain data of the target weld seam. The target strain data indicates the deformation characteristics of the target weld seam within the target duration due to the mechanical property test. Because the deformation characteristics of the target weld seam are closely related to the internal structure and quality of the target weld seam, the strain data indicating the deformation characteristics can effectively detect whether there is a defect in the target weld seam.

[0036] In one implementation, the quality detection model includes: a convolution layer; as shown in Figure 2 The data processing flowchart of the convolution layer provided by the embodiments of the present application is shown in the following figure: Figure 2 In the process of iteratively training the quality detection model, the current training process before reaching the training stop condition includes: S310-S340, details are shown as follows:

[0037] S310: After inputting the sample time-frequency matrix into the quality detection model, for each sample time-frequency matrix, the convolution kernel in the convolution layer is used to perform propagation convolution on the sample time-frequency matrix to obtain a direction feature map corresponding to each simulated propagation direction.

[0038] The sample time-frequency matrix is obtained by preprocessing the sample strain data; the sample strain data is collected based on the sample weld seam in the sample steel structure; the convolution kernel is constructed according to the propagation characteristics of the strain wave in the steel structure, and the propagation characteristics include: propagation speed and propagation direction; the propagation convolution indicates that the convolution kernel value is determined based on the simulated propagation speed and multiple simulated propagation directions of the sample strain wave in the sample steel structure.

[0039] Specifically, in the embodiment of the present application, a batch-by-batch iterative training strategy is adopted, all sample strain data is divided into batches of fixed size, and each iteration is completed after forward propagation, loss calculation and back propagation of each batch; one training process is counted as one traversal of all sample strain data, the loss function value of each training process is continuously monitored, if the loss function values corresponding to multiple consecutive training processes are not decreased, the training process is forcibly terminated and the current optimal model parameters are saved.

[0040] In the embodiment of the present application, multiple sample strain data need to be obtained before training the quality detection model; in actual operation, high-precision resistance strain gauge arrays or distributed optical fiber sensor networks can be arranged in typical stress regions and defect-free base material regions of the sample weld in the sample steel structure, such as the weld heat affected zone and the fusion line region, to collect dynamic strain signals of the sample weld, i.e. sample strain data, at a preset sampling frequency.

[0041] In actual operation, the collected sample strain data should be collected under multiple working conditions, and the types of working conditions can include static load, fatigue cycle or impact load, etc. mechanical property test; the sample strain data of the sample weld in the sample steel structure collected under multiple working conditions not only covers all working environments that the weld may experience in actual use, but also covers various defects that the weld may have, such as cracks, incomplete fusion and porosity, etc., so as to improve the detection accuracy of the quality detection model trained by the sample strain data; after determining the sample strain data, the sample strain data is preprocessed to obtain a sample time-frequency matrix.

[0042] According to physical knowledge, strain wave is a phenomenon that when an object is subjected to external force, due to the interaction between the internal parts of the object, strain is generated in the object, and the strain propagates in the object in the form of a wave; in actual application, the strain generated by the weld in the steel structure under the action of external force also propagates in the steel structure in the form of a wave, therefore, the embodiment of the present application determines whether the weld has defects by exploring the propagation characteristics of the strain wave.

[0043] In actual application, the propagation of the strain wave generated by the weld in the steel structure has direction specificity, such as preferentially propagating along the length direction of the weld, but the convolution kernel in the conventional convolutional neural network does not consider the spatial propagation characteristics of the strain wave, resulting in the application dilemma of slow convergence and easy to capture false features, in order to solve this application dilemma, the embodiment of the present application uses wave equation to construct a physically guided convolution kernel to fully consider the spatial propagation characteristics of the strain wave.

[0044] In an embodiment of the present application, the convolution kernel value of each element in the convolution kernel is jointly determined by the product of the Gaussian attenuation function and the cosine function; the center frequency of the Gaussian attenuation function covers the lowest to highest analysis frequency of the cosine function, the simulated propagation direction of the strain wave covers the entire angle, and the simulated propagation velocity of the strain wave is calculated by Young's modulus, Poisson's ratio, and the density of the steel structure, so that the response mode of the convolution kernel matches the propagation characteristics of the strain wave in the steel structure; wherein, the simulated propagation direction means the propagation direction obtained by simulating the propagation direction of the strain wave in the steel structure, and the simulated propagation velocity means the propagation velocity obtained by simulating the propagation velocity of the strain wave in the steel structure; the formula for calculating the convolution kernel value is as follows:

[0045] ;

[0046] ; ; ;

[0047] ; ;

[0048] ; ;

[0049] Where, Indicates along The first of the simulated propagation directions The convolution kernel value required for convolution in the simulated propagation direction; in the embodiment of the present application, the simulated propagation direction is the index direction of the elements in the convolution kernel during convolution, ;

[0050] Represents the horizontal component of the spatial coordinates of the convolution kernel during convolution, Represents the vertical component of the spatial coordinate of the convolution kernel during convolution;

[0051] represents the L2 norm; Represents the position vector of the convolution kernel; Indicates the center position of the convolution kernel; Indicates the width of the convolution kernel; Indicates the height of the convolution kernel; Represents the transpose of a vector;

[0052] represents the center frequency of the Gaussian decay function; Represents the lowest analytical frequency of the cosine function; Represents the highest analytical frequency of the cosine function;

[0053] represents the Gaussian attenuation coefficient; Represents the minimum function;

[0054] A vector representing the simulated propagation direction of the strain wave, covering to ; Angle representing the simulated propagation direction of the strain wave;

[0055] represents the simulated propagation velocity of the strain wave; represents Young's modulus; Represents Poisson's ratio. In actual operation, The value of is set to 0.3; Indicates the density of the steel structure. In actual operation, The value can be set to 7850 .

[0056] According to the above discussion, the strain wave generated by the weld has directional specificity in the propagation of the steel structure. For example, it preferentially propagates along the length direction of the weld. However, the conventional convolutional layer treats each directional feature equally, which will reduce the significance of the key directional features. The standard channel attention mechanism will ignore the directional specificity of the strain wave propagation and cannot enhance the directional features most relevant to the defects, resulting in the application dilemma of inaccurate quality inspection results. In order to solve this application dilemma, the embodiment of the present application will determine the directional feature map corresponding to each simulated propagation direction, and then determine the weight of each simulated propagation direction based on the above directional feature map, so as to highlight the simulated propagation direction most relevant to the weld defect through the weight.

[0057] In actual operation, after the sample time-frequency matrix is ​​input into the convolution layer, the convolution layer first determines the convolution kernel value of each element in the convolution kernel based on the sample time-frequency matrix. Specifically, the convolution kernel value is determined based on the simulated propagation velocity of the sample strain wave in the sample steel structure and multiple simulated propagation directions. For details, please refer to the "Formula for Calculating Convolution Kernel Values" recorded above. After the convolution layer determines the convolution kernel value of each element in the convolution kernel, it performs propagation convolution based on the convolution kernel value sample time-frequency matrix to obtain multiple directional characteristic maps. Each simulated propagation direction corresponds to a directional characteristic map. The formula for determining the directional characteristic map is as follows:

[0058] ;

[0059] Where, Indicates the Directional characteristic diagram corresponding to the simulated propagation direction; Represents the sample time-frequency matrix, that is, the moment and the acquisition frequency time-frequency coefficients at the position; denotes a rectified linear unit activation function; denotes a convolution operation.

[0060] S320: determining, by the convolution layer, an energy flow vector corresponding to each simulated propagation direction according to a direction feature map corresponding to the simulated propagation direction.

[0061] The energy flow vector indicates an energy propagation direction of the energy flow of the strain wave.

[0062] Specifically, in the embodiments of the present application, the sample time-frequency matrix is regarded as an initial source of "energy", and when the sample time-frequency matrix is transmitted in the convolution layer, the convolution layer transforms the sample time-frequency matrix, which is similar to the transmission and conversion of energy in a medium.

[0063] In the embodiments of the present application, the idea of determining the energy flow vector of the energy flow of the strain wave is to sum the spatial gradient amplitudes of the direction feature map along the horizontal and vertical directions respectively, which is used to represent the main direction of propagation of the energy flow; wherein the formula for calculating the energy flow vector is as follows:

[0064] ;

[0065] In the formula, denotes an energy flow vector; denotes a direction feature map corresponding to the i-th simulated propagation direction The partial derivative in the x-axis direction is discretely implemented to represent a Sobel operator. denotes a direction feature map corresponding to the i-th simulated propagation direction The partial derivative in the x-axis direction is discretely implemented to represent a Sobel operator. denotes a direction feature map corresponding to the i-th simulated propagation direction The partial derivative in the x-axis direction is discretely implemented to represent a Sobel operator. denotes the number of width pixels of the direction feature map; denotes the number of height pixels of the direction feature map.

[0066] S330: determining, by the convolution layer, a direction weight corresponding to each simulated propagation direction according to an energy flow vector corresponding to each simulated propagation direction.

[0067] Specifically, in the embodiments of the present application, the direction weight corresponding to each simulated propagation direction is calculated according to the energy flow vector and the temperature coefficient adjustment determined based on the energy flow vector; wherein the formula for calculating the direction weight is as follows:

[0068] ; ;

[0069] In the formula,​​ represents the weight corresponding to the i-th simulated propagation direction,

[0070] represents a vector of the simulated propagation direction of the strain wave; represents the transpose of

[0071] represents a temperature coefficient for adaptively adjusting the weight distribution.

[0072] S340: determining, by the convolution layer, sample propagation feature data of the sample strain wave according to the direction feature map corresponding to each simulated propagation direction and the direction weight.

[0073] Specifically, in the embodiments of the present application, when the direction weight corresponding to each simulated propagation direction in the plurality of simulated propagation directions is determined, the sample propagation feature data of the sample strain wave can be determined according to the direction weight; wherein the formula for calculating the sample propagation feature data is as follows:

[0074]

[0075] In the formula, represents the sample propagation feature data.

[0076] In an implementation manner, the quality detection model further comprises a gated recurrent unit; as shown in Figure 3 Figure 3 is a data processing flowchart of the gated recurrent unit provided by the embodiments of the present application, after S340, the method further comprises S350-S360, details are as follows:

[0077] S350: for each sample propagation feature data, determining, by the gated recurrent unit, a strain accumulation factor according to the sample propagation feature data.

[0078] Wherein, the strain accumulation factor represents the memory degree of the sample weld to the damage caused by the impact.

[0079] S360: performing feature enhancement processing on the data representing the burst strain event in the sample propagation feature data by the gated recurrent unit according to the strain accumulation factor to obtain sample enhanced feature data;

[0080] Wherein, the bias for calculating the updated gate state value in the gated recurrent unit is the strain accumulation factor; the burst strain event refers to the deformation of the weld after receiving an impact higher than the preset intensity in less than the preset time.

[0081] ​​​​​​Specifically, in an embodiment of the present application, a sudden strain event indicates deformation of a weld caused by an impact higher than a preset intensity within a period shorter than a preset time; wherein the preset time and the preset intensity can be determined according to actual needs, "a period shorter than the preset time" means a shorter time, and "an impact higher than a preset intensity" means an impact with a relatively high intensity, that is, a sudden strain event refers to deformation caused by an impact with a relatively high intensity within a relatively short period of time, and sudden strain events are usually caused by external forces.

[0082] In practical applications, the evolution of weld damage is time-dependent. Some weld damage is caused by long-term impacts, while others are caused by sudden impacts. Different impacts also cause different damages. However, conventional GRUs (Gated Recurrent Units) are not sensitive to sudden impacts, i.e., sudden strain events. The Sigmoid gating function of conventional GRUs will experience gradient saturation when the strain suddenly changes, resulting in an application dilemma of insufficient memory for sudden strain events. To address this application dilemma, the embodiment of the present application uses a strain accumulation factor in the update gate of the gated recurrent unit. The strain accumulation factor is constructed by the L1 norm historical mean and cumulative gain of the sample propagation feature data to enhance the long-term memory of historical damage. The reset gate and candidate hidden state calculations remain unchanged. When the hidden state is updated, the update gate output, the previous hidden state, and the candidate hidden state are combined to improve the modeling capability of the long-term dependence of the evolution of weld damage. The formula for calculating the sample enhanced feature data is as follows:

[0083] ;

[0084] ;

[0085] ; ;

[0086] ;

[0087] ;

[0088] Where, Represents the reset gate output; Represents the weight matrix of the reset gate; Indicates the The hidden state at each sampling moment; Indicates the sample propagation feature data corresponding to In actual operation, the sample strain data is time series data, including the strain values ​​of the sample weld collected at multiple sampling moments included in the target time. The strain values ​​at multiple sampling moments constitute the sample strain data of the sample weld to reflect the deformation characteristics of the sample weld over time. Accordingly, It is also time series data;

[0089] represents the update gate output; Represents the Sigmoid activation function; represents the weight matrix of the update gate; represents the strain accumulation factor, which is used to enhance the long-term memory of historical injuries; represents the cumulative gain; Indicates the sample propagation feature data corresponding to Data at each sampling moment; express within the time The maximum value of represents the L1 norm;

[0090] represents the candidate hidden state; The weight matrix representing the candidate hidden state; represents the Hadamard product; Indicates the The hidden state at each sampling moment is the sample enhanced feature data.

[0091] In one implementation, the quality detection model further includes: a regularization layer and a fully connected layer; Figure 4 As shown, Figure 4 The data processing flow chart of the regularization layer and the fully connected layer provided in the embodiment of the present application, after S360, the method further includes: S370 to S400, the details of which are as follows:

[0092] S370: For each sample time-frequency matrix, determine the strain acceleration standard deviation according to the sample time-frequency matrix through a regularization layer;

[0093] Among them, the standard deviation of strain acceleration indicates the boundary of physical rationality that the reference regularized data should comply with; the reference regularized data is the regularized data obtained by simulation without processing through the regularization layer; and the physical rationality represents the physical constraints.

[0094] Specifically, in the embodiment of the present application, "reference regularized data" refers to the regularized data obtained by regularizing the sample enhanced feature data in the embodiment of the present application through a conventional regularization method (i.e., without passing through the regularization layer provided in the embodiment of the present application).

[0095] In practical applications, there are fuzzy boundaries in the classification of the quality of the weld, such as micro-cracks and stress concentration, but the conventional regularization method will destroy the physical continuity of the strain of the weld, that is, the conventional regularization method does not consider the physical constraint of the strain of the weld, and randomly discarding features will lead to the generation of invalid regularization data that violates the material constitutive relationship, in order to solve the application dilemma, the embodiment of the application determines whether the regularization data obtained by the regularization processing of the conventional regularization method meets the physical constraint, to guide the process of the regularization processing of the regularization layer of the embodiment of the application.

[0096] In actual operation, it is not necessary to really use other regularization methods to perform regularization processing on the sample enhanced feature data to obtain reference regularization data, the embodiment of the application constructs reference regularization data, which is used to indirectly determine whether the regularization data obtained by the regularization processing of the sample enhanced feature data by the conventional regularization method meets the physical constraint, since the embodiment of the application does not need to really use other regularization methods to perform regularization processing on the sample enhanced feature data to obtain reference regularization data, the reference regularization data is called "simulated" regularization data.

[0097] In the embodiment of the application, the strain acceleration standard deviation is used to represent the boundary of physical rationality; wherein the formula for calculating the strain acceleration standard deviation is as follows:

[0098] ;

[0099] In the formula, is the synthesized correction strain value corresponding to the i-th sampling time determined according to the sample strain data; represents the number of sampling times.

[0100] S380: determining the mask function value by the regularization layer according to the acceleration threshold coefficient, the strain acceleration standard deviation and the reference regularization data;

[0101] Wherein, the acceleration threshold coefficient is determined according to the material properties of the sample steel structure, which is used to control the influence degree of physical rationality on the reference regularization data; the mask function value indicates whether the reference regularization data exceeds the boundary of physical rationality;

[0102] Specifically, in the embodiment of the application, after determining the strain acceleration standard deviation, the acceleration threshold coefficient is combined to obtain the mask function value; wherein the acceleration threshold coefficient is determined according to the material properties of the sample steel structure, which is used to control the influence degree of physical rationality on the reference regularization data, in actual operation, the value of the acceleration threshold coefficient can be 2.5; the mask function value indicates whether the reference regularization data exceeds the boundary of physical rationality.​

[0103] wherein the mask function is formulated as follows:

[0104] If , the value of is equal to 0, indicating that the reference regularization data exceeds the boundary of physical rationality;

[0105] If , the value of is equal to 1, indicating that the reference regularization data does not exceed the boundary of physical rationality;

[0106] wherein, denotes the reference regularization data, , denotes the reversible feature mapping; denotes the reference regularization data obtained by projecting the sample propagation feature data back to the deformation space using the 1x1 convolution operation;

[0107] denotes the second-order derivative of the reference regularization data with respect to time, which is calculated using the central difference method and is used to represent the strain acceleration; denotes the acceleration threshold coefficient.

[0108] S390: determining sample regularization data by the regularization layer according to the sample enhanced feature data and the mask function value.

[0109] Specifically, in the embodiments of the present application, after determining the mask function value, the mask function value is applied to the hidden state, the mask vector is multiplied element by element with each element in the hidden state (i.e., the sample enhanced feature data), and the time dimension mean of the hidden state (i.e., the sample enhanced feature data) is filled in the position where the mask is 0, to obtain the regularization hidden state, i.e., the sample regularization data; wherein the formula for calculating the sample regularization data is as follows:

[0110] ;

[0111] ;

[0112] wherein, denotes the sample regularization data; denotes the mask vector, which is composed of , and the constituent dimension is consistent with the time sequence, which refers to the sequence of a plurality of sampling time points corresponding to the sample strain data; denotes a full 1 vector with the same dimension as ;

[0113] represents the mean value of the hidden state (i.e., the sample enhanced feature data) in the time dimension.

[0114] In actual operation, the quality detection model further includes a full connection layer configured to determine a quality prediction result corresponding to the sample strain data according to the sample regularization data.

[0115] S400: determining the quality prediction result according to the sample regularization data through the full connection layer.

[0116] Specifically, in the embodiments of the present application, the quality prediction result and the quality detection result are both binary classification probabilities, which are used to represent defects or no defects.

[0117] In one implementation, before S310, the method further includes step (1), details as follows:

[0118] Step (1): determining quality labels corresponding to the sample strain data according to the sample strain data.

[0119] The type of the quality label includes defects and no defects.

[0120] Specifically, in the embodiments of the present application, the sample is labeled according to the nondestructive testing result of the sample steel structure; wherein, the sample strain data whose nondestructive testing result meets the expectation is determined as positive sample strain data, and the sample strain data whose nondestructive testing result does not meet the expectation is determined as negative sample strain data; in actual operation, the content of the sample labeling includes not only the type of the quality label, but also the defect type, position, and severity level, forming a strain-label paired data set with a time stamp.

[0121] In one implementation, before S400, the method further includes steps (1)-(4), details as follows:

[0122] Step (1): for each positive sample strain data with the quality label type of no defects, determining the physical focusing parameter according to the sample regularization data.

[0123] Specifically, in actual application, the defects represented by the sample strain data of the sample weld are extremely unbalanced, such as the qualified products are much larger than the cracked products, and the strain response difference between the micro-cracks and the noise is small, thereby causing the conventional cross-entropy loss function to have insufficient distinguishing ability for the difficult-to-distinguish samples close to the decision boundary, that is, the conventional cross-entropy loss function does not consider the absolute threshold of the strain value, in order to solve the technical problem, in the embodiments of the present application, when calculating the loss function value, for the positive sample strain data, the loss weight combines the class balance factor and the physical focus parameter related to the physical threshold value, for the negative sample strain data, the loss weight combines the class balance factor and the focus parameter related to the confidence, and the final loss is the mean value of the loss of all samples, so as to enhance the sensitivity to the small defects and the unbalanced sample scene; wherein the formula for determining the physical focus parameter is as follows:

[0124] ;

[0125] ;

[0126] In the formula, denotes the physical focus parameter; denotes the first scaling factor, in actual operation, the value of can be set to 2; denotes sample strain data corresponding to the maximum reconstructed strain value of the sample strain data, is a positive integer, representing the number of sample strain data; denotes the physical damage threshold, in actual operation, the value of can be set to 0.005; denotes the reconstructed strain function, which is used to map the hidden state (i.e. sample enhanced data) to the strain value; is the sample regularization data corresponding to the sample strain data; denotes the maximum value of at the sampling time.

[0127] Step (3): For each quality label type of the negative sample strain data of the defect, according to the quality prediction result, the confidence focus parameter is determined.

[0128] Specifically, in the embodiments of the present application, the formula for determining the confidence focus parameter is as follows:

[0129] ;

[0130] In the formula, denotes the confidence focus parameter; denotes a second scaling factor, in actual operation, a value of the second scaling factor can be set to 3; denotes a second scaling factor, in actual operation, a value of the second scaling factor can be set to 3; denotes a binary information entropy of the quality prediction result .

[0131] Step (4): determining a loss function value of the current training process according to the number of sample strain data of the defect sample, the number of sample strain data of the non-defect sample, the plurality of quality prediction results, the plurality of physical focus parameters and the plurality of confidence focus parameters according to the type of the quality label;

[0132] The loss function value is used to indicate a prediction loss of the current training process; the prediction loss indicates a difference between the quality label and the quality prediction result.

[0133] Specifically, in the embodiment of the present application, a formula of the loss function used to calculate the loss function value is as follows:

[0134] .

[0135] . .

[0136] In the formula, loss denotes the loss function; N denotes a total number of sample strain data; I denotes an indicator function; y denotes a value of the quality prediction result when the i-th sample strain data is positive sample strain data, otherwise y denotes 0; y denotes a value of the quality prediction result when the i-th sample strain data is negative sample strain data, otherwise y denotes 0; C denotes a class balance factor; N denotes a number of negative sample strain data; P denotes a number of positive sample strain data; y denotes a quality prediction result of the i-th sample strain data; W denotes a weight of the full connection layer; x denotes sample regularization data corresponding to the i-th sample strain data.

[0137] ​​​​​​​​​​​​​​​​​​​In actual operation, the gradient of all trainable network parameters in the quality detection model is calculated according to the loss function value, and the network parameters are updated according to the calculated gradient; wherein the gradient calculation is realized by automatic differentiation technology, and the error signal is reversely transmitted along the calculation graph from the output layer to the input layer; the parameter update is executed immediately after the end of each training process, ensuring that the model iteration optimization direction is always towards loss minimization.

[0138] In an implementation manner, the sample strain data includes original strain values respectively collected at a plurality of sampling time points; before the step S400, the method further includes steps (5) to (9), details of which are shown as follows:

[0139] The step (5) is to determine, for each sample strain data, an adaptive time window length corresponding to the first sampling time point according to a preset sampling frequency, original strain values respectively corresponding to the first sampling time point and the second sampling time point in the sample strain data, and a strain change rate threshold.

[0140] The preset sampling frequency is the sampling frequency when the sample strain data is collected; the first sampling time point is later than the second sampling time point; the strain change rate threshold is a critical value for determining whether the original strain value is a sudden strain; and the value of the adaptive time window length indicates a response speed to the sudden strain.

[0141] Specifically, in the embodiment of the present application, the sample strain data includes original strain values respectively collected at a plurality of sampling time points; the preset sampling frequency is the sampling frequency when the sample strain data is collected; and the first sampling time point is later than the second sampling time point.

[0142] In actual application, the plurality of original strain values included in the sample strain data of the sample weld usually have baseline deviation caused by equipment drift and high-frequency electromagnetic noise in the industrial environment, which are superimposed with non-stationary low-frequency trend items and pulse-type high-frequency interference, so that the conventional normalization will amplify the noise and mask the characteristics of the tiny defects. The traditional moving average filtering will smooth the sudden characteristics, and the wavelet threshold denoising needs to preset the base function, thus leading to the application dilemma that it is difficult to adapt to the mixed modal of sudden plastic deformation and elastic recovery in the weld strain. In order to solve the application dilemma, the original strain value is processed by a double-channel adaptive filtering mechanism in the embodiment of the present application to eliminate the baseline deviation caused by the equipment drift and the high-frequency electromagnetic noise and retain the characteristics of the tiny defects.

[0143] In actual operation, in order to eliminate the baseline deviation caused by the equipment drift and the high-frequency electromagnetic noise and retain the characteristics of the tiny defects, it is necessary to first calculate the adaptive time window length corresponding to the first sampling time point; wherein the formula for calculating the adaptive time window length is as follows:

[0144] ;

[0145] wherein, denotes the adaptive time window length corresponding to the first sampling time, used for dynamically adjusting the integration interval to ensure fast response to sudden strain; denotes the preset sampling frequency, representing the collection rate of the strain value, in actual operation, the value of may be set to 10 kHz;

[0146] denotes the strain rate threshold, which is used to determine whether the original strain value is a critical value of sudden strain, in actual operation, the value of may be set to .

[0147] denotes the original strain value corresponding to the first sampling time, i.e. ; wherein, denotes the original strain value corresponding to the second sampling time, i.e. ; wherein, denotes the upward rounding.

[0148] Step (6): determining the baseline value corresponding to the first sampling time according to the strain rate threshold, the adaptive time window length and the original strain value corresponding to the third sampling time.

[0149] wherein, the baseline value is the original strain value corresponding to the first sampling time, and the value remaining after removing the high-frequency interference data.

[0150] Specifically, in the embodiments of the present application, after determining the adaptive time window length corresponding to the first sampling time, the strain rate threshold can be used to separate the baseline drift, and only integrate the strain values with strain rate lower than the given strain rate threshold to obtain the low-frequency baseline value; wherein, the formula for calculating the limit value is as follows:

[0151] Baseline value= .

[0152] wherein, is used to represent the integration of strain values in the interval , representing the time window integration; denotes the original strain value at the third sampling time, i.e. ; wherein,

[0153] denotes the indicator function, in the term , the strain rate is represented by ; when , the value of the indicator function is 1 to retain the baseline; when , the value of the indicator function is 0 to filter out the effective signal;

[0154] represents the strain rate threshold, which can take the maximum strain rate in the elastic stage of the material;

[0155] represents the derivative of the strain value with respect to the third sampling time, i.e. , which is calculated by a discrete difference approximation and represents the strain rate.

[0156] Step (7): determining the baseline-removed strain value corresponding to the first sampling time according to the original strain value and the baseline value corresponding to the first sampling time.

[0157] Specifically, in the embodiments of the present application, after the baseline value is determined, the baseline-removed strain value is obtained by removing the baseline drift according to the baseline value to retain the high-frequency effective signal; wherein the baseline strain value used for calculation is as shown below:

[0158]

[0159] wherein, represents the baseline-removed strain value corresponding to the first sampling time, i.e. .

[0160] Step (8): determining the noise-suppressed strain value according to the baseline-removed strain value corresponding to the first sampling time.

[0161] wherein the noise-suppressed strain value is the value remaining after the original strain value corresponding to the first sampling time is suppressed.

[0162] Specifically, in the embodiments of the present application, the non-linear compression is used to suppress the pulse noise, and the compression operation of the combination of the sign function and the natural pair function is applied to the baseline-removed strain value, and the compression coefficient is updated in real time by the absolute value mean of the baseline-removed strain value in the time window to reduce the pulse amplitude; wherein the formula used for calculating the noise-suppressed strain value is as shown below:

[0163] Noise-suppressed strain value=

[0164]

[0165] wherein, represents the sign function; in , if , the value of the sign function is 1, if , the value of the sign function is 0, and if , the value of the sign function is -1; represents the pulse compression coefficient, which is updated in real time by the noise level; represents the number of original strain values in the time window .​​​ Represents the natural logarithm function.

[0166] Step (9): Add the baseline value and the noise suppression strain value to obtain the synthetic correction strain value corresponding to the first sampling moment.

[0167] Specifically, in the embodiment of the present application, the baseline value corresponding to the first sampling moment and the noise suppression strain value are added together to obtain the synthetic correction strain value corresponding to the first sampling moment. In actual operation, the other sampling moments included in the sample strain data can be sequentially determined as the first sampling moments to determine the synthetic correction strain value corresponding to each sampling moment.

[0168] In one implementation, after step (9), the method further includes: steps (10) to (12), the details of which are as follows:

[0169] Step (10): determining the stress value corresponding to the first sampling moment according to the material yield strain, the material yield stress, the elastic section Young's modulus, the strain hardening modulus, the hardening exponent and the synthetic corrected strain value corresponding to the first sampling moment;

[0170] The stress value indicates the stress capacity of the elastic section and / or hardened section in the weld.

[0171] Specifically, in the embodiment of the present application, the defects of the sample weld are manifested as local microstrain concentration in the sample strain data, but the strain value is significantly affected by the boundary conditions. The traditional Hooke's law assumes that the material behavior is linear elastic, which cannot handle the nonlinear hardening effect of the heat-affected zone of the weld, resulting in distortion of the defect stress field estimation. In order to solve this application dilemma, the embodiment of the present application performs strain-stress conversion and damage feature enhancement to overcome the limitations of the traditional linear elastic assumption and highlight the local strain mutation caused by damage.

[0172] In actual operation, during the process of damage feature enhancement, the corrected strain value is converted into a stress value through segmented stress mapping. If the synthetic corrected strain value does not exceed the material yield strain, the stress is calculated using the Young's modulus of the elastic segment. If the synthetic corrected strain value exceeds the material yield strain, the stress of the plastic hardening segment is calculated based on the yield stress, hardening modulus, and hardening exponent to distinguish between the elastic segment and the plastic hardening segment. The formula for determining the stress value corresponding to the first sampling moment is as follows:

[0173] like , equal ;

[0174] like , equal ;

[0175] wherein, denotes the first sampling time, i.e. the stress value corresponding to the time;

[0176] denotes the first sampling time, i.e. the synthetic corrected strain value corresponding to the time;

[0177] denotes the material yield strain, determined by the material structure of the steel structure, in actual operation, the value of may be set to 0.0017; denotes the material yield stress, satisfying ; in actual operation, the value of may be 345x10 6 Pa; denotes the Young's modulus of the elastic segment, the value is determined according to the steel standard, in actual operation, the value of may be set to 210x10 9 Pa; denotes the strain hardening modulus, calibrated by the material tensile test, in actual operation, the value of may be set to 3; denotes the hardening index, representing the plastic deformation capacity, in actual operation, the value of may be set to 0.2.

[0178] Step (11): determining the normalized strain acceleration corresponding to the first sampling time according to the synthetic corrected strain value corresponding to the first sampling time.

[0179] Specifically, in the embodiments of the present application, by calculating the normalized strain acceleration, the absolute value of the time second order derivative of the corrected strain value is divided by the larger value of the absolute value of the time first order derivative and the anti-zero constant, to eliminate the dimensional influence; wherein, the formula for calculating the normalized strain acceleration is as follows:

[0180] ;

[0181] wherein, denotes the first sampling time, i.e. the normalized strain acceleration corresponding to the time; denotes the second order derivative of the synthetic corrected strain value with respect to time, the central difference method is used for discrete calculation, representing the strain change rate; denotes the first order derivative of the synthetic corrected strain value with respect to time, the forward difference method is used for discrete calculation, representing the strain rate; denotes the constant for preventing zero, to avoid the denominator being zero, in actual operation, the value of may be set to 106 ; denotes a function taking the larger value of the two.

[0182] Step (12): determining a damage feature vector corresponding to the first sampling time according to the stress value corresponding to the first sampling time and the normalized strain acceleration.

[0183] Specifically, in the embodiments of the present application, the stress value and the normalized strain acceleration are fused, a feature enhancement factor constructed by combining a damage gain coefficient, a hyperbolic tangent function and a time domain mean value of the absolute value of the strain acceleration is used, and then a damage feature vector is constructed. The formula for calculating the damage feature vector is as follows:

[0184] ;

[0185] ;

[0186] In the formula, t denotes the first sampling time, i.e. the damage feature vector corresponding to the first sampling time;

[0187] denotes a damage gain coefficient, which controls the contribution weight of the strain acceleration. In actual operation, the value of the damage gain coefficient can be set to 0.3. denotes a hyperbolic tangent function.

[0188] denotes a feature enhancement factor, which is defined as a time domain mean value of the absolute value of the strain acceleration.

[0189] denotes the number of sampling times corresponding to the sample strain data. denotes the differential of .

[0190] In one implementation manner, after step (12), the method further includes steps (13) to (18), the details of which are as follows:

[0191] Step (13): performing a short-time Fourier transform on the sample strain data to obtain a frequency point set including a plurality of frequency points.

[0192] ​​Specifically, in actual application, the crack of the weld seam generates acoustic emission signals of a specific frequency band, but the feature in the sample strain data is submerged by the main load response, and the sensitive frequency band needs to be focused in the time-frequency domain to highlight the frequency band related to the damage. The traditional fixed window time-frequency analysis method such as short-time Fourier transform cannot meet the resolution requirements of crack burst and other high-frequency transient events and fatigue expansion and other low-frequency continuous damage, resulting in that the key features are blurred or missed. In order to solve the application dilemma, the embodiment of the present application focuses on the damage-sensitive frequency band by adaptive bandwidth transformation to adaptively optimize the time-frequency resolution and highlight the energy of the damage-related frequency band.

[0193] In actual operation, short-time Fourier transform needs to be performed on the relative sample strain data to obtain a frequency point set including a plurality of frequency points.

[0194] Step (14): determining the normalized power spectral density at the first frequency point at the first sampling time according to the first frequency point and the second frequency point in the frequency point set corresponding to the first sampling time.

[0195] Specifically, in the embodiment of the present application, the normalized power spectral density is calculated based on the energy proportion of the short-time Fourier transform coefficient; wherein the formula for calculating the normalized power spectral density is as follows:

[0196] ; ;

[0197] In the formula, represents the normalized power spectral density at the first sampling time and the first frequency point , the value domain of which represents , indicating that the proportion of the energy of the frequency point to the total energy;

[0198] represents the output of the short-time Fourier transform; represents the complex-valued time-frequency coefficient at the first sampling time, that is, the time and the first frequency point ; represents the complex-valued time-frequency coefficient at the first sampling time, that is, the time and the first frequency point ;

[0199] represents the first frequency point, the value range of which represents to , covering the typical acoustic emission frequency band of the weld crack; represents the first frequency point, the value range of which represents to , for frequency point summation; represents a positive integer;

[0200] represents the number of frequency points; represents the floor function; represents the logarithmic function, and the default base is 10.

[0201] Step (15): determining the time-varying spectral entropy corresponding to the first sampling moment according to the normalized power spectral density corresponding to the first sampling moment.

[0202] wherein the value of the time-varying spectral entropy indicates the signal complexity.

[0203] Specifically, in the embodiments of the present application, the time-varying spectral entropy is calculated by using the normalized power spectral density to quantify the signal complexity; wherein the formula for calculating the time-varying spectral entropy is as follows:

[0204] ;

[0205] In the formula, represents the time-varying spectral entropy at the first sampling moment, i.e. , and the value range is , and the lower the value, the lower the signal complexity; represents a positive integer.

[0206] Step (16): determining the window width parameter corresponding to the first sampling moment according to the preset analysis frequency, the preset window width adjustment factor and the time-varying spectral entropy corresponding to the first sampling moment.

[0207] wherein the window width adjustment factor is used to control the scaling ratio of the adaptive bandwidth; and the value of the window width parameter indicates the adaptive bandwidth.

[0208] Specifically, in the embodiments of the present application, the frequency-dependent window width is dynamically calculated based on the time-varying spectral entropy and the window width adjustment factor, and a narrow window is used for high-frequency analysis and a wide window is used for low-frequency analysis; wherein the formula for calculating the window width parameter is as follows:

[0209] ;

[0210] In the formula, represents the frequency-dependent window width parameter, and the larger the value, the wider the window;

[0211] represents the analysis frequency, and the value range is to , i.e. to ;

[0212] denotes a window width adjustment factor, controls the overall scaling of the window width, in actual operation, the value of may be set to 0.35.

[0213] Step (17): determining the window function output corresponding to the first sampling time according to the window width parameter corresponding to the first sampling time.

[0214] Specifically, in the embodiment of the present application, a Gaussian window function is constructed, and the window width is controlled by the frequency-dependent window width parameter; wherein the window function used to calculate the window function output is as follows:

[0215] ;

[0216] In the formula, w denotes the window function output, the value range of w is , and w is used to weight the local signal segment in the time-frequency analysis; denotes the time difference, denotes the offset of the current time point and the integral variable, that is, ; denotes the integral variable; denotes the natural exponential function; denotes the natural exponential function; denotes the natural exponential function.

[0217] Step (18): determining the time-frequency matrix at the first sampling time and at the analysis frequency according to the damage feature vector corresponding to the first sampling time and the window function output.

[0218] Specifically, in the embodiment of the present application, the adaptive bandwidth transform is performed, the damage feature vector is combined with the dynamic window function to generate the time-frequency matrix; wherein the formula used to generate the time-frequency matrix is as follows:

[0219] ;

[0220] In the formula, X denotes the time-frequency matrix, denotes the time-frequency coefficient at the first sampling time, that is, at time and frequency , the amplitude value of the time-frequency coefficient represents the signal energy; denotes the damage feature vector at time ;denotes the imaginary unit. S200: pre-processing the target strain data to obtain a target time-frequency matrix.

[0221] Specifically, the step of pre-processing the target strain data to obtain a target time-frequency matrix can refer to the aforementioned steps about “determining a sample time-frequency matrix according to sample strain data”, which will not be repeated here.

[0222] Specifically, the step of pre-processing the target strain data to obtain a target time-frequency matrix can refer to the aforementioned steps about “determining a sample time-frequency matrix according to sample strain data”, which will not be repeated here.

[0223] S300: input the target time-frequency matrix into the pre-trained quality detection model, so that the quality detection model determines target propagation characteristic data of the target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target weld according to the target propagation characteristic data;

[0224] The strain wave indicates deformation of the weld in the form of a wave propagating in the steel structure.

[0225] Specifically, in the embodiment of the present application, the quality detection result is a binary classification probability. If the quality detection result is greater than a preset threshold value, it is considered that there is a defect in the target weld and a warning is given. Otherwise, it is considered that there is no defect in the target weld.

[0226] To evaluate the processing effect of different noise suppression methods on the strain data of the steel structure weld, the processing results of the original signal, the traditional sliding average filtering and the double-channel adaptive filtering method of the present application are compared, and the ability of each method to retain the characteristics of crack mutation and small defects while suppressing baseline drift and pulse noise is investigated, as shown in Figure 5 . Figure 5 The effect comparison chart of different noise suppression methods provided by the embodiment of the present application, the time axis shows the signal change process, and the strain value axis reflects the signal amplitude characteristics. It can be observed that the original signal contains obvious low-frequency drift and high-frequency burr. Although the traditional sliding average filtering smooths the noise, it severely weakens the crack mutation characteristics at 300 milliseconds. The present technical method in the embodiment of the present application (i.e. Figure 5 ) not only effectively eliminates the baseline drift, but also completely retains the key defect characteristics. The red curve in the figure presents a steep peak at the crack position, while the random pulse interference is significantly suppressed. The abnormal peak in the green curve is eliminated, proving the advantage of the present application in extracting weak damage characteristics in a complex industrial noise environment.

[0227] To evaluate the influence of different noise suppression techniques on the detection rate of weld defects, the performance of sliding average filtering, wavelet threshold denoising, Kalman filtering and the present application under the same test conditions is compared through box plots combined with scatter plots, as shown in Figure 6 . Figure 6 The defect detection rate comparison chart of different noise suppression methods provided by the embodiment of the present application, the box height reflects the stability of the detection result, and the scatter distribution shows the results of multiple experiments. The present application (i.e. Figure 6The present technical method) shows the highest median detection rate and the most concentrated high-value distribution. Its box position is significantly higher than other methods, and the scattered points are densely distributed in the top area, proving that the dual-channel adaptive filtering mechanism adopted in the embodiment of the present application can effectively eliminate equipment drift and industrial electromagnetic noise, while retaining the characteristics of minor defects, improving detection reliability, and solving the defects of traditional methods of either amplifying noise or smoothing mutation characteristics.

[0228] In order to analyze the influence of different strain-stress conversion methods on the crack size detection accuracy, the detection error variation trends of the linear Hooke's law, the nonlinear hardening model and the embodiment of the present application under different real crack sizes are compared. Figure 7 As shown, Figure 7 The crack size detection error of different strain-stress conversion methods provided in the embodiment of the present application is shown in the shadow area, which emphasizes the error range. Figure 7 The error curve of the present technical method) is always at the bottom, and maintains a stable low error as the crack size increases, especially in the small crack range, which verifies the effectiveness of the segmented stress mapping and damage feature enhancement mechanism of the embodiment of the present application. By fusing stress values ​​and normalized strain acceleration, it overcomes the limitations of the traditional linear elastic assumption in the heat-affected zone of the weld, and significantly improves the recognition accuracy of small defects.

[0229] The feature extraction effects of conventional short-time Fourier transform and adaptive bandwidth transform in the embodiment of the present application are now visually compared through the time-frequency heat map. Figure 8 As shown, Figure 8 The comparison diagram of the time-frequency domain feature extraction method provided in the embodiment of the present application is shown in the upper part. The upper part is the traditional short-time Fourier transform (STFT), whose frequency axis and time axis resolution are fixed. The lower part is the embodiment of the present application (i.e. Figure 8 The results of the present technology in this application are obtained using a dynamically adjusted time-frequency window. Color intensity represents signal energy, and bright areas correspond to damage features. Experiments show that the traditional method's 100 Hz impact features at 300 milliseconds (a typical frequency band for crack generation) and 200 Hz features at 700 milliseconds (a frequency band for crack propagation) are fuzzy and diffuse, with significant background noise interference. However, the feature points at the same location in the present application have concentrated energy and clear boundaries, separating the 150 Hz persistent damage feature (yellow strip in the figure) that was completely obliterated by the traditional method. This demonstrates that the present application's frequency-dependent window width adjustment mechanism effectively resolves the contradiction between fixed-resolution methods in simultaneously capturing high-frequency transient events and low-frequency persistent damage.

[0230] The receiver operating characteristic curve is now used to evaluate the comprehensive performance of different detection models, such as Figure 9 As shown, Figure 9The ROC curve comparison diagram of different quality prediction models provided by the embodiment of the application is shown in FIG. 1. The network structures of other quality detection models can be SVM (Support Vector Machine), random forest and CNN (Convolutional Neural Network). The ROC curve (Receiver Operating Characteristic Curve) is an important tool for evaluating the performance of a binary classification model in machine learning and statistics. The horizontal axis of the curve is the false positive rate (FPR), and the vertical axis is the true positive rate (TPR). The diagonal line represents the random guess benchmark. The curve of the model in the embodiment of the application (i.e. the technology in the embodiment of the application) is closest to the upper left corner, still maintains a high true positive rate in the low false positive rate area, and has the maximum area under the curve, which reflects the synergistic advantages of the physically guided direction-sensitive convolution kernel, the strain accumulation factor enhanced gating recurrent unit and the uncertainty regularization. Through the fusion of the material constitutive relation and the wave propagation characteristics, the discrimination ability for the micro cracks and the fuzzy boundaries is strengthened under the condition of sample imbalance, which is significantly better than the traditional data-driven model. Figure 9

[0231] Secondly, the application provides a steel structure weld quality detection device based on artificial intelligence, as shown in FIG. 2. Figure 10 Figure 10 The structure diagram of the steel structure weld quality detection device based on artificial intelligence provided by the embodiment of the application is shown in FIG. 2. The device comprises a data acquisition module 400 and a quality detection module 500.

[0232] The data acquisition module 400 is used for performing mechanical property testing on the target weld with to-be-detected quality in the target steel structure body for a target duration to obtain target strain data of the target weld.

[0233] The strain data indicates the deformation characteristics of the weld in the target duration.

[0234] The quality detection module 500 is used for pre-processing the target strain data to obtain a target time-frequency matrix.

[0235] The target time-frequency matrix is input into a pre-trained quality detection model, so that the quality detection model determines target propagation characteristic data of a target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target weld according to the target propagation characteristic data.

[0236] The strain wave indicates that the deformation of the weld propagates in the form of a wave in the steel structure body.

[0237] ​​In an implementation manner, the quality detection model comprises a convolution layer; the device further comprises a training module; the training module is configured to, in a current training process before a training stop condition is reached in an iterative training process for the quality detection model, perform, for each sample time-frequency matrix, propagation convolution on the sample time-frequency matrix by using a convolution kernel in the convolution layer to obtain a directional feature map corresponding to each simulated propagation direction after the sample time-frequency matrix is input into the quality detection model.

[0238] The sample time-frequency matrix is obtained by preprocessing sample strain data; the sample strain data is collected based on a sample weld in a sample steel structure; the convolution kernel is constructed according to propagation characteristics of a strain wave in the steel structure, and the propagation characteristics include a propagation speed and a propagation direction; the propagation convolution indicates that the convolution operation is performed based on a value of the convolution kernel determined according to a simulated propagation speed and a plurality of simulated propagation directions of the sample strain wave in the sample steel structure.

[0239] The training module is configured to determine, by using the convolution layer, an energy flow vector corresponding to each simulated propagation direction according to the directional feature map corresponding to each simulated propagation direction.

[0240] The energy flow vector indicates an energy propagation direction of an energy flow of the strain wave.

[0241] The training module is configured to determine, by using the convolution layer, a directional weight corresponding to each simulated propagation direction according to the energy flow vector corresponding to each simulated propagation direction.

[0242] The training module is configured to determine, by using the convolution layer, sample propagation characteristic data of the sample strain wave according to the directional feature map corresponding to each simulated propagation direction and the directional weight.

[0243] In an implementation manner, the quality detection model further comprises a gated recurrent unit; the training module is configured to determine, by using the gated recurrent unit, a strain accumulation factor according to the sample propagation characteristic data for each sample propagation characteristic data.

[0244] The strain accumulation factor represents a memory degree of the sample weld to damage caused by the impact.

[0245] The training module is configured to perform feature enhancement processing on data representing a sudden strain event in the sample propagation characteristic data by using the gated recurrent unit according to the strain accumulation factor to obtain sample enhanced feature data.

[0246] The bias used to calculate the updated gate state value in the gated recurrent unit is the strain accumulation factor; the sudden strain event indicates deformation of the weld after the weld receives an impact with a strength higher than a preset strength within a time shorter than a preset time.

[0247] In an implementation manner, the quality detection model further comprises: a regularization layer and a full connection layer; and the training module is configured to determine, for each sample time-frequency matrix, a strain acceleration standard deviation by the regularization layer according to the sample time-frequency matrix.

[0248] The strain acceleration standard deviation indicates a boundary of physical rationality that the reference regularization data should meet.

[0249] The training module is configured to determine, by the regularization layer, a mask function value according to the acceleration threshold coefficient, the strain acceleration standard deviation and the reference regularization data.

[0250] The acceleration threshold coefficient is determined according to material properties of the sample steel structure body and is used to control an influence degree of the physical rationality on the reference regularization data.

[0251] The training module is configured to determine, by the regularization layer, sample regularization data according to the sample enhanced feature data and the mask function value.

[0252] The training module is configured to determine, by the full connection layer, the quality prediction result according to the sample regularization data.

[0253] In an implementation manner, the device further comprises: a labeling module.

[0254] The labeling module is configured to determine, according to the sample strain data, a quality label corresponding to each sample strain data.

[0255] The type of the quality label includes: defect and no defect.

[0256] The training module is further configured to determine, for each positive sample strain data of which the type of the quality label is no defect, a physical focus parameter according to the sample regularization data.

[0257] The training module is further configured to determine, for each negative sample strain data of which the type of the quality label is defect, a confidence focus parameter according to the quality prediction result.

[0258] The training module is further configured to determine, according to a quantity of sample strain data of which the type of the quality label is defect, a quantity of sample strain data of which the type of the quality label is no defect, a plurality of quality prediction results, a plurality of physical focus parameters and a plurality of confidence focus parameters, a loss function value of a current training process.

[0259] The loss function value is used to indicate a prediction loss of the current training process. The prediction loss indicates a difference between the quality label and the quality prediction result.

[0260] In an implementation manner, the sample strain data includes original strain values respectively collected at a plurality of sampling time points; the training module is further configured to, for each sample strain data, determine an adaptive time window length corresponding to a first sampling time point according to a preset sampling frequency, original strain values respectively corresponding to the first sampling time point and a second sampling time point adjacent to the first sampling time point in the sample strain data, and a strain change rate threshold value;

[0261] The preset sampling frequency is a sampling frequency when the sample strain data is collected; the first sampling time point is later than the second sampling time point; the strain change rate threshold value is a critical value for determining whether the original strain value is a sudden strain; and the value of the adaptive time window length indicates a response speed to the sudden strain.

[0262] The training module is further configured to determine a baseline value corresponding to the first sampling time point according to the strain rate threshold value, the adaptive time window length, and an original strain value corresponding to a third sampling time point.

[0263] The baseline value is a value of the original strain value corresponding to the first sampling time point after high-frequency interference data is removed.

[0264] The training module is further configured to determine a baseline strain value corresponding to the first sampling time point according to the original strain value corresponding to the first sampling time point and the baseline value.

[0265] The training module is further configured to determine a noise suppression strain value corresponding to the first sampling time point according to the baseline strain value corresponding to the first sampling time point; the noise suppression strain value is a value of the original strain value corresponding to the first sampling time point after noise is suppressed.

[0266] The training module is further configured to add the baseline value and the noise suppression strain value to obtain a synthetic correction strain value corresponding to the first sampling time point.

[0267] In an implementation manner, the training module is further configured to determine a stress value corresponding to the first sampling time point according to a material yield strain, a material yield stress, an elastic segment Young's modulus, a strain hardening modulus, a hardening index, and the synthetic correction strain value corresponding to the first sampling time point.

[0268] The stress value indicates a stress capacity of an elastic segment and / or a hardening segment in the weld.

[0269] The training module is further configured to determine a normalized strain acceleration corresponding to the first sampling time point according to the synthetic correction strain value corresponding to the first sampling time point.

[0270] The training module is further configured to determine a damage feature vector corresponding to the first sampling time point according to the stress value corresponding to the first sampling time point and the normalized strain acceleration.

[0271] In an implementation manner, the training module is further configured to perform a short-time Fourier transform on the sample strain data to obtain a frequency point set comprising a plurality of frequency points;

[0272] The training module is further configured to determine, according to a first frequency point and a second frequency point in the frequency point set corresponding to the first sampling time, a normalized power spectral density at the first frequency point at the first sampling time;

[0273] The training module is further configured to determine, according to the normalized power spectral density corresponding to the first sampling time, a time-varying spectral entropy corresponding to the first sampling time;

[0274] The value of the time-varying spectral entropy indicates a signal complexity;

[0275] The training module is further configured to determine, according to a preset analysis frequency, a preset window width adjustment factor, and the time-varying spectral entropy corresponding to the first sampling time, a window width parameter corresponding to the first sampling time;

[0276] The window width adjustment factor is used to control a scaling ratio of the adaptive bandwidth, and the value of the window width parameter indicates the adaptive bandwidth;

[0277] The training module is further configured to determine, according to the window width parameter corresponding to the first sampling time, a window function output corresponding to the first sampling time;

[0278] The training module is further configured to determine, according to the damage feature vector corresponding to the first sampling time and the window function output, a time-frequency matrix at the first sampling time and at the analysis frequency.

[0279] Thirdly, the application further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the steps of S100-S300 provided by the above-mentioned embodiments.

[0280] Fourthly, the application further provides a computer readable storage medium, and the computer readable storage medium stores a computer program, the computer readable medium stores a computer program, and the computer program is executed by the processor to execute the steps of S100-S300 of the above-mentioned embodiments.

[0281] Fifthly, the computer program product provided by the application comprises a computer readable storage medium storing program codes, the instructions included in the program codes are used to execute the method in the above-mentioned method embodiments, and the specific implementation can refer to the steps of S100-S300 of the method embodiments, which will not be described herein.

[0282] In the embodiments of the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. The embodiments described above are merely exemplary, for example, the division of the units is only a logical function division, and there can be another division manner in actual implementation; for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0283] In addition, the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0284] In addition, the functional modules in the various embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0285] It should be noted that if the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various other media that can store program codes.

[0286] In this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations.

[0287] The above merely provides an example of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A steel structure weld quality detection method based on artificial intelligence, characterized in that: The method comprises: Performing a mechanical property test for a target duration on a target weld to be tested in a target steel structure to obtain target strain data of the target weld; The strain data indicates the deformation characteristics of the weld during the target time period. Preprocessing the target strain data to obtain a target time-frequency matrix; Inputting the target time-frequency matrix into a pre-trained quality inspection model, so that the quality inspection model determines target propagation characteristic data of a target strain wave according to the target time-frequency matrix, and determines a quality inspection result of the target weld according to the target propagation characteristic data; Among them, the strain wave indicates that the deformation generated by the weld propagates in the steel structure in the form of waves; The quality detection model includes a convolutional layer and a gated recurrent unit; The convolution layer performs propagation convolution on the sample time-frequency matrix through the convolution kernel to obtain the directional feature map corresponding to each simulated propagation direction; The convolution layer determines the energy flow vector corresponding to each of the simulated propagation directions according to each of the directional feature maps; The convolution layer determines the directional weight corresponding to each of the simulated propagation directions according to each of the energy flow vectors; The convolution layer determines the sample propagation characteristic data of the sample strain wave according to each of the directional characteristic maps and the directional weights; The gated loop unit performs feature enhancement processing on the data representing the sudden strain event in the sample propagation feature data to obtain sample enhanced feature data.

2. The method according to claim 1, characterized in that The quality detection model includes: a convolutional layer; during the iterative training of the quality detection model, the current training process before reaching the training stop condition includes: After the sample time-frequency matrix is ​​input into the quality detection model, for each sample time-frequency matrix, the sample time-frequency matrix is ​​propagated and convolved by the convolution kernel in the convolution layer to obtain a directional characteristic map corresponding to each simulated propagation direction; The sample time-frequency matrix is ​​obtained by preprocessing the sample strain data; the sample strain data is collected based on the sample weld in the sample steel structure; the convolution kernel is constructed based on the propagation characteristics of the strain wave in the steel structure, and the propagation characteristics include: propagation speed and propagation direction; the propagation convolution indicates that a convolution operation is performed based on a convolution kernel value determined based on a simulated propagation speed of the sample strain wave in the sample steel structure and a plurality of simulated propagation directions; Determining, by the convolution layer, the energy flow vector corresponding to each simulated propagation direction according to the directional characteristic map corresponding to each simulated propagation direction; Wherein, the energy flow vector indicates the energy propagation direction of the energy flow of the strain wave; Determining, by the convolution layer, a directional weight corresponding to each simulated propagation direction according to the energy flow vector corresponding to each simulated propagation direction; The sample propagation characteristic data of the sample strain wave is determined by the convolution layer according to the directional characteristic map and the directional weight corresponding to each of the simulated propagation directions.

3. The method according to claim 2, characterized in that The quality detection model further includes: a gated cycle unit; after determining the sample propagation characteristic data of the sample strain wave, the method further includes: For each of the sample propagation characteristic data, determining a strain accumulation factor according to the sample propagation characteristic data by the gated cycle unit; The strain accumulation factor represents the memory degree of the sample weld to the damage caused by the impact; Performing feature enhancement processing on the data representing the sudden strain event in the sample propagation feature data according to the strain accumulation factor by the gated cycle unit to obtain sample enhanced feature data; The bias used in the gated cycle unit to calculate and update the gate state value is the strain accumulation factor; the sudden strain event indicates the deformation of the weld caused by receiving an impact higher than a preset intensity within a shorter than preset time.

4. The method according to claim 3, characterized in that The quality detection model further includes: a regularization layer and a fully connected layer; after performing feature enhancement processing on the data representing the sudden strain event in the sample propagation feature data according to the strain accumulation factor by the gated recurrent unit to obtain sample enhanced feature data, the method further includes: For each sample time-frequency matrix, determining the strain acceleration standard deviation according to the sample time-frequency matrix through the regularization layer; The strain acceleration standard deviation indicates the boundary of physical rationality that the reference regularized data should comply with; the reference regularized data is regularized data obtained by simulation without being processed by the regularization layer; and the physical rationality represents a physical constraint. Determining a mask function value through the regularization layer according to an acceleration threshold coefficient, the strain acceleration standard deviation, and the reference regularization data; The acceleration threshold coefficient is determined according to the material properties of the sample steel structure and is used to control the degree of influence of the physical rationality on the reference regularized data; the mask function value indicates whether the reference regularized data exceeds the boundary of the physical rationality; Determining sample regularization data according to the sample enhancement feature data and the mask function value through the regularization layer; The quality prediction result is determined according to the sample regularization data through the fully connected layer.

5. The method according to claim 4, characterized in that Before performing iterative training on the quality detection model, the method further includes: determining, according to the sample strain data, a quality label corresponding to each sample strain data; The types of the quality labels include: defective and non-defective; After determining the quality prediction result according to the sample regularization data through the fully connected layer, the method includes: For each of the positive sample strain data whose quality label is defect-free, determining a physical focusing parameter according to the sample regularization data; For each of the negative sample strain data whose quality label is the defect, determining a confidence focus parameter according to the quality prediction result; determining a loss function value of a current training process according to the number of the sample strain data whose quality label is the defect, the number of the sample strain data whose quality label is the non-defective, a plurality of the quality prediction results, a plurality of the physical focusing parameters, and a plurality of physical focusing parameters; The loss function value is used to indicate the prediction loss of the current training process; the prediction loss indicates the difference between the quality label and the quality prediction result.

6. The method according to claim 2, characterized in that The sample strain data includes original strain values ​​collected at multiple sampling moments; Before performing iterative training on the quality detection model, the method further includes: For each of the sample strain data, determine the adaptive time window length corresponding to the first sampling moment according to a preset sampling frequency, the original strain values ​​corresponding to the adjacent first sampling moment and the second sampling moment in the sample strain data, and a strain change rate threshold; The preset sampling frequency is the sampling frequency when collecting the sample strain data; the first sampling time is later than the second sampling time; the strain change rate threshold is used to determine whether the original strain value is a critical value of a sudden change strain; the value of the adaptive time window length indicates the response speed to the sudden change strain; determining a baseline value corresponding to the first sampling moment according to a strain rate threshold, the adaptive time window length, and the original strain value corresponding to the third sampling moment; The baseline value is the original strain value corresponding to the first sampling moment, which is the value remaining after removing high-frequency interference data; determining a de-baseline strain value corresponding to the first sampling moment according to the original strain value and the baseline value corresponding to the first sampling moment; Determining a noise-suppressed strain value based on the baseline-removed strain value corresponding to the first sampling moment; wherein the noise-suppressed strain value is the value remaining after noise is suppressed from the original strain value corresponding to the first sampling moment; The baseline value and the noise suppression strain value are added to obtain a synthetic corrected strain value corresponding to the first sampling moment.

7. The method according to claim 6, characterized in that After adding the baseline value and the noise suppression strain value to obtain a synthetic corrected strain value corresponding to the first sampling moment, the method further includes: Determining a stress value corresponding to the first sampling moment according to the material yield strain, the material yield stress, the elastic section Young's modulus, the strain hardening modulus, the hardening exponent, and the synthetic corrected strain value corresponding to the first sampling moment; wherein the stress value indicates the stress capacity of the elastic section and / or the hardened section in the weld; determining a normalized strain acceleration corresponding to the first sampling moment according to the synthetic corrected strain value corresponding to the first sampling moment; A damage characteristic vector corresponding to the first sampling moment is determined according to the stress value corresponding to the first sampling moment and the normalized strain acceleration.

8. The method according to claim 7, characterized in that After determining the damage feature vector corresponding to the first sampling moment, the method further includes: Performing a short-time Fourier transform on the sample strain data to obtain a frequency point set including a plurality of frequency points; Determining a normalized power spectral density at the first frequency point at the first sampling moment according to a first frequency point and a second frequency point in the frequency point set corresponding to the first sampling moment; determining a time-varying spectral entropy corresponding to the first sampling moment according to a normalized power spectral density corresponding to the first sampling moment; Wherein, the value of the time-varying spectral entropy indicates signal complexity; Determining a window width parameter corresponding to the first sampling moment according to a preset analysis frequency, a preset window width adjustment factor, and a time-varying spectral entropy corresponding to the first sampling moment; The window width adjustment factor is used to control the scaling ratio of the adaptive bandwidth; the value of the window width parameter indicates the adaptive bandwidth; determining a window function output corresponding to the first sampling moment according to a window width parameter corresponding to the first sampling moment; A time-frequency matrix at the first sampling moment and at the analysis frequency is determined according to the damage feature vector corresponding to the first sampling moment and the window function output.

9. An artificial intelligence-based steel structure weld quality detection device, characterized in that: Used to execute the steel structure weld quality detection method based on artificial intelligence according to any one of claims 1 to 8; the device comprises: a data acquisition module and a quality detection module; The data acquisition module is used to perform a mechanical property test for a target weld to be tested in a target steel structure for a target duration to obtain target strain data of the target weld; The strain data indicates the deformation characteristics of the weld during the target time period. The quality detection module is used to preprocess the target strain data to obtain a target time-frequency matrix; Inputting the target time-frequency matrix into a pre-trained quality inspection model, so that the quality inspection model determines target propagation characteristic data of a target strain wave according to the target time-frequency matrix, and determines a quality inspection result of the target weld according to the target propagation characteristic data; Among them, the strain wave indicates that the deformation generated by the weld propagates in the steel structure in the form of waves.

10. The device according to claim 9, characterized in that The quality detection model includes: a convolutional layer; the device includes: a training module; The training module is configured to, during iterative training of the quality detection model, in a current training process before a training stop condition is reached, after inputting the sample time-frequency matrix into the quality detection model, perform propagation convolution on the sample time-frequency matrix using a convolution kernel in a convolution layer to obtain a directional feature map corresponding to each simulated propagation direction; The sample time-frequency matrix is ​​obtained by preprocessing the sample strain data; the sample strain data is collected based on the sample weld in the sample steel structure; the convolution kernel is constructed based on the propagation characteristics of the strain wave in the steel structure, and the propagation characteristics include: propagation speed and propagation direction; the propagation convolution indicates the convolution operation is performed based on the convolution kernel value determined based on the simulated propagation speed and multiple simulated propagation directions of the sample strain wave in the sample steel structure; The training module is further configured to determine the energy flow vector corresponding to each simulated propagation direction based on the directional feature map corresponding to each simulated propagation direction through a convolutional layer; Among them, the energy flow vector indicates the energy propagation direction of the energy flow of the strain wave; The training module is further configured to determine the directional weight corresponding to each simulated propagation direction according to the energy flow vector corresponding to each simulated propagation direction through a convolutional layer; The training module is further configured to determine sample propagation characteristic data of the sample strain wave according to the directional characteristic graphs and directional weights corresponding to each simulated propagation direction through a convolution layer.

Images