Steel structure welding seam quality detection method and device based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a steel structure welding seam quality detection method and device based on artificial intelligence. The method comprises the following steps: performing a mechanical property test of which the duration is a target duration on a target welding seam of which the quality is to be detected in a target steel structure body, so as to obtain target strain data of the target welding seam; 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 the target strain wave according to the target time-frequency matrix, and determines a quality detection result of the target welding seam according to the target propagation characteristic data; wherein the strain wave indicates that the deformation generated by the welding seam is propagated in the steel structure body in a wave form; according to the invention, the technical problems of limited hidden defect identification and existence of detection blind areas and artificial errors in the existing welding seam quality detection method can be solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular, to a method and device for detecting the quality of steel structure welds based on artificial intelligence. Background Art

[0002] In the welding process, after two steel structure components 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 tensile force, pressure, shear force, and bending moment, etc.; among them, 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, ray, and magnetic particle, etc. However, due to the characteristics of small size, diverse shapes, and complex evolution processes of the steel structure weld defects, the above-mentioned prior art has limited recognition of hidden defects such as early micro-cracks, lack of fusion, and pores, and the above-mentioned prior art is limited by the operating environment, the experience of the detection personnel, and the sensor distribution density, resulting in detection blind spots and human errors. Summary of the Invention

[0004] In view of this, the purpose of the present application is to provide a method and device for detecting the quality of steel structure welds based on artificial intelligence, so as to solve the technical problems of limited recognition of hidden defects and the existence of detection blind spots and human errors in the existing weld quality detection methods.

[0005] In a first aspect, the present application provides a method for detecting the quality of steel structure welds based on artificial intelligence, the method comprising: Performing a mechanical property test with a target duration on a target weld to be detected for quality in a target steel structure to obtain target strain data of the target weld; wherein the strain data indicates the deformation characteristics presented by the weld during the target duration; Performing preprocessing on 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 generated by the weld propagates in the steel structure in the form of a wave.

[0006] In a second aspect, the present application provides a device for detecting the quality of steel structure welds based on artificial intelligence, the device comprising: a data acquisition module and a quality detection module; The data acquisition module is used to perform a mechanical property test with a target duration on a target weld to be detected for quality in a target steel structure to obtain target strain data of the target weld; Among them, the strain data indicates the deformation characteristics presented by the weld during the target duration; The quality detection module is used to preprocess the target strain data to obtain a target time-frequency matrix; Input 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; Among them, the strain wave indicates that the deformation generated by the weld propagates in the steel structure in the form of a wave.

[0007] Beneficial effects: The present application provides a method for detecting the quality of steel structure welds based on artificial intelligence. The method includes: performing a mechanical property test with a target duration on a target weld to be detected for quality in a target steel structure to obtain target strain data of the target weld; among them, the strain data indicates the deformation characteristics presented by the weld during 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; among them, the strain wave indicates that the deformation generated by the weld propagates in the steel structure in the form of a wave; by constructing a method for detecting the quality of steel structure welds based on artificial intelligence, the present application can identify hidden defects in the target weld and avoid detection blind spots and manual errors through an automated identification process, improving the detection quality and detection efficiency. Description of the drawings

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required to be used in the embodiments of the present application. The following drawings only show some embodiments of the present application, so they should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained according to these drawings.

[0009] Figure 1 It is a schematic flow chart of a method for detecting the quality of steel structure welds based on artificial intelligence provided by an embodiment of the present application; Figure 2 It is a schematic data processing flow chart of a convolutional layer provided by an embodiment of the present application; Figure 3Schematic diagram of the data processing flow of the gated recurrent unit provided by the embodiment of the present application; Figure 4 Data processing flow chart of the regularization layer and the fully connected layer provided by the embodiment of the present application; Figure 5 Comparison diagram of the effects of different noise suppression methods provided by the embodiment of the present application; Figure 6 Comparison diagram of the defect detection rates of different noise suppression methods provided by the embodiment of the present application; Figure 7 Schematic diagram of the crack size detection error of different strain-stress conversion methods provided by the embodiment of the present application; Figure 8 Comparison diagram of time-frequency domain feature extraction methods provided by the embodiment of the present application; Figure 9 Comparison diagram of the ROC curves of different quality prediction models provided by the embodiment of the present application; Figure 10 Schematic diagram of the structure of the steel structure weld quality detection device based on artificial intelligence provided by the embodiment of the present application. Detailed implementation manners

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0011] First, the present application provides a steel structure weld quality detection method based on artificial intelligence, as Figure 1 shown, Figure 1 Schematic diagram of the flow of the steel structure weld quality detection method based on artificial intelligence provided by the embodiment of the present application. The method includes: S100~S300, details are as follows: S100: Perform a mechanical property test with a target duration on the target weld whose quality is to be detected in the target steel structure body to obtain the target strain data of the target weld; Among them, the strain data indicates the deformation characteristics presented by the weld during the target duration.

[0012] 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 there is at least one weld in the target steel structure; the "target weld" is the weld that needs to be subjected to quality inspection; the "mechanical property test" is a test for measuring and analyzing various mechanical responses of the weld under the action of force, and the strain data obtained through 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 inspection" refers to the monitoring for determining whether there are defects in the target weld.

[0013] In actual operation, after determining the target weld in the target steel structure, the mechanical property test can be carried out on the target weld to obtain the target strain data of the target weld; the target strain data indicates the deformation characteristics presented by the target weld due to the mechanical property test within the target time period. Since the deformation characteristics of the target weld are closely related to the internal structure and quality of the target weld, analyzing the strain data indicating the deformation characteristics can effectively detect whether there are defects in the target weld.

[0014] In one implementation, the quality inspection model includes: a convolutional layer; as Figure 2 shown, Figure 2 is the schematic diagram of the data processing flow of the convolutional layer provided by the embodiments of the present application. During the iterative training process of the quality inspection model, the current training process before reaching the training stop condition includes: S310~S340, details are as follows: S310: After inputting the sample time-frequency matrix into the quality inspection model, for each sample time-frequency matrix, perform propagation convolution on the sample time-frequency matrix through the convolutional kernel in the convolutional layer to obtain the direction feature maps corresponding to each simulated propagation direction.

[0015] Among them, 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 convolutional 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 performing a convolution operation based on the convolution kernel values determined by the simulated propagation speed and multiple simulated propagation directions of the sample strain wave in the sample steel structure.

[0016] Specifically, in the embodiments of the present application, a batch-by-batch iterative training strategy is adopted. All the sample strain data is divided into batches of a fixed size. Each time the forward propagation, loss calculation, and backward propagation of a batch are completed, it is counted as one iteration; traversing all the sample strain data once is counted as one training process, and the loss function value of each training process is continuously monitored. If the loss function values corresponding to multiple consecutive training processes do not decrease, the training process is forced to terminate and the current optimal model parameters are saved.

[0017] In the embodiments of the present application, before training the quality detection model, it is necessary to obtain a plurality of sample strain data; in actual operation, a high-precision resistance strain gauge array or a distributed optical fiber sensor network can be arranged in the typical stress areas and defect-free base material areas of the sample welds in the sample steel structure, such as the welding heat affected zone and the fusion line area, and the dynamic strain signals of the sample welds, that is, the sample strain data, are collected at a preset sampling frequency.

[0018] In actual operation, the collected sample strain data should be collected under multiple working conditions respectively. The types of working conditions can include mechanical property tests such as static load, fatigue cycle or impact load, etc.; the sample strain data of the sample welds in the sample steel structure collected under multiple working conditions cover all working environments that the welds may experience during actual use, and also cover various defects that the welds may appear, such as cracks, lack of fusion and pores, etc., so as to improve the detection accuracy of the quality detection model trained by the sample strain data; when the sample strain data is determined, the sample strain data is preprocessed to obtain a sample time-frequency matrix.

[0019] According to physical knowledge, a strain wave is a phenomenon in which when an object is subjected to an external force, due to the existence of mutual forces between various parts of the object, strain is generated inside the object, and this strain propagates in the object in the form of a wave; in actual applications, when the welds in a steel structure are subjected to an external force, the strain generated also propagates in the steel structure in the form of a wave. Therefore, in the embodiments of the present application, the propagation characteristics of the strain wave are explored to determine whether there are defects in the welds.

[0020] In actual applications, the propagation of the strain wave generated by the weld in the steel structure has direction specificity. For example, it preferentially propagates along the length direction of the weld. However, the convolution kernels in conventional convolutional neural networks do not consider the spatial propagation characteristics of the strain wave, resulting in the application dilemmas of slow convergence and easy capture of pseudo-features. To solve this application dilemma, in the embodiments of the present application, a physically guided convolution kernel is constructed using the wave equation to fully consider the spatial propagation characteristics of the strain wave.

[0021] In the embodiments of the present application, the convolution kernel values of each element in the convolution kernel are jointly determined by the product of a Gaussian attenuation function and a cosine function; the center frequency of the Gaussian attenuation function covers the lowest to the highest analysis frequencies of the cosine function, the simulated propagation direction of the strain wave covers the full angle, and the simulated propagation speed of the strain wave is calculated from the Young's modulus, Poisson's ratio, and the density of the steel structure body, so that the response mode of the convolution kernel matches the propagation characteristics of the strain wave in the steel structure body; wherein, the meaning of the simulated propagation direction is the propagation direction obtained by simulating the propagation direction of the strain wave in the steel structure body, and the meaning of the simulated propagation speed is the propagation speed obtained by simulating the propagation speed of the strain wave in the steel structure body; the formula for calculating the convolution kernel value is as follows: ; ; ; ; ; ; ; ; In the formula, represents the convolution kernel value required when performing convolution along the th simulated propagation direction among the simulated propagation directions; in the embodiments of the present application, the simulated propagation direction is the indexing direction of the elements in the convolution kernel during convolution, ; represents the horizontal component of the spatial coordinate of the convolution kernel during convolution, represents the vertical component of the spatial coordinate of the convolution kernel during convolution; represents the L2 norm; represents the position vector of the convolution kernel; represents the center position of the convolution kernel; represents the width of the convolution kernel; represents the height of the convolution kernel; represents the transpose of a vector; represents the center frequency of the Gaussian attenuation function; represents the lowest analysis frequency of the cosine function; represents the highest analysis frequency of the cosine function; represents the Gaussian attenuation coefficient; represents the minimum value function; represents the vector of the simulated propagation direction of the strain wave, covering to ; The angle representing the simulated propagation direction of the strain wave; The simulated propagation speed of the strain wave; The Young's modulus; The Poisson's ratio. In actual operation, the value of can be set to 0.3; The density of the steel structure. In actual operation, the value of can be set to 7850 .

[0022] According to the foregoing discussion, the strain wave generated by the weld has direction specificity in the steel structure, such as preferentially propagating along the length direction of the weld. However, the conventional convolutional layer equally processes the features in all directions, which will reduce the significance of the key direction features. The standard channel attention mechanism will ignore the direction specificity of the strain wave propagation and cannot enhance the direction features most relevant to the defect, resulting in the application dilemma that the obtained quality inspection results are inaccurate. To solve this application dilemma, the embodiments of the present application will determine the direction feature maps corresponding to each simulated propagation direction, and then determine the weights of each simulated propagation direction according to the above direction feature maps, so as to highlight the simulated propagation direction most relevant to the defect of the weld through the weights.

[0023] In actual operation, when the sample time-frequency matrix is input to the convolutional layer, the convolutional layer first determines the convolutional kernel values of each element in the convolutional kernel according to the sample time-frequency matrix. Specifically, the convolutional kernel values are determined based on the simulated propagation speed of the sample strain wave in the sample steel structure and multiple simulated propagation directions. For details, please refer to the "formula for calculating convolutional kernel values" recorded above; when the convolutional layer determines the convolutional kernel values of each element in the convolutional kernel, it performs propagation convolution on the sample time-frequency matrix according to the convolutional kernel values to obtain multiple direction feature maps, and each simulated propagation direction corresponds to a direction feature map; among them, the formula for determining the direction feature map is as follows: ; In the formula, represents the direction feature map corresponding to the th simulated propagation direction; represents the sample time-frequency matrix, that is, the time-frequency coefficient at time and the acquisition frequency ; represents the rectified linear unit activation function; represents the convolution operation.

[0024] S320: Through the convolutional layer, according to the direction feature maps corresponding to each simulated propagation direction, determine the energy flow vectors corresponding to each simulated propagation direction.

[0025] Among them, the energy flow vector indicates the energy propagation direction of the energy flow of the strain wave.

[0026] Specifically, in the embodiments of the present application, the sample time-frequency matrix is regarded as the initial source of "energy". When the sample time-frequency matrix is passed through the convolutional layer, the convolutional layer transforms the sample time-frequency matrix, which is similar to the transmission and conversion of energy in a medium.

[0027] 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 perform summation processing on the spatial gradient amplitude of the direction feature map along the horizontal and vertical directions respectively, which is used to characterize the main propagation direction of the energy flow. Among them, the formula for calculating the energy flow vector is as follows: ; In the formula, represents the energy flow vector; represents the th direction feature map corresponding to the simulated propagation direction in the axis direction, and the discrete implementation represents the Sobel operator; represents the th direction feature map corresponding to the simulated propagation direction in the axis direction; represents the number of width pixels of the direction feature map; represents the number of height pixels of the direction feature map.

[0028] S330: Determine the direction weights corresponding to each simulated propagation direction through the convolutional layer according to the energy flow vectors corresponding to each simulated propagation direction.

[0029] Specifically, in the embodiments of the present application, calculate the direction weights corresponding to each simulated propagation direction according to the energy flow vector and the adjustment based on the energy flow vector determined temperature coefficient. Among them, the formula for calculating the direction weight is as follows: ; ; In the formula, represents the weight corresponding to the th simulated propagation direction, ; represents the vector of the simulated propagation direction of the strain wave; represents transpose of; represents the temperature coefficient, which is used to adaptively adjust the weight distribution.

[0030] S340: Determine the sample propagation feature data of the sample strain wave through the convolutional layer based on the direction feature maps and direction weights corresponding to each simulated propagation direction.

[0031] Specifically, in the embodiments of the present application, after determining the direction weights corresponding to each simulated propagation direction among multiple simulated propagation directions, the sample propagation feature data of the sample strain wave can be determined according to the direction weights; wherein, the formula for calculating the sample propagation feature data is as follows: ; In the formula, represents the sample propagation feature data.

[0032] In one implementation manner, the quality detection model further includes: a gated recurrent unit; as Figure 3 shown, Figure 3 is a schematic diagram of the data processing flow of the gated recurrent unit provided by the embodiments of the present application. After S340, the method further includes: S350~S360, details are as follows: S350: For each sample propagation feature data, determine the strain accumulation factor through the gated recurrent unit according to the sample propagation feature data.

[0033] Among them, the strain accumulation factor characterizes the memory degree of the sample weld to the damage caused by being impacted.

[0034] S360: Through the gated recurrent unit, perform feature enhancement processing on the data representing the sudden strain event in the sample propagation feature data to obtain the sample enhanced feature data; Among them, the bias for calculating the update gate state value in the gated recurrent unit is the strain accumulation factor; the sudden strain event indicates the deformation generated after the weld receives an impact with a strength higher than the preset strength within a time shorter than the preset duration.

[0035] Specifically, in the embodiments of the present application, the sudden strain event indicates the deformation generated after the weld receives an impact with a strength higher than the preset strength within a time shorter than the preset duration; wherein, the preset duration and the preset strength can both be determined according to actual needs. The meaning of "a time shorter than the preset duration" is a shorter time, and the meaning of "an impact with a strength higher than the preset strength" is an impact with a relatively high strength. That is, the sudden strain event refers to the deformation generated after suffering an impact with a relatively high strength within a relatively short time. The sudden strain event is usually caused by external forces.

[0036] In practical applications, the evolution of weld damage has time-dependence. Some weld damages are caused by long-term impacts, while others are caused by sudden impacts. The damages caused by different impacts are also different. However, the conventional GRU (Gated Recurrent Unit) is not sensitive to sudden impacts, that is, sudden strain events. The Sigmoid gating function of the conventional GRU will saturate the gradient during sudden strain changes, resulting in the application dilemma of insufficient memory ability for sudden strain events. To solve this application dilemma, in the update gate of the gated recurrent unit in this application embodiment, a strain accumulation factor is adopted. The strain accumulation factor is constructed by the historical mean value of the L1 norm of the sample propagation feature data and the cumulative gain to enhance the long-term memory of historical damage. The reset gate and the calculation of the candidate hidden state remain unchanged. When updating the hidden state, the output of the update gate, the hidden state at the previous moment, and the candidate hidden state are combined to improve the modeling ability for the long-term dependence of the evolution of weld damage. Among them, the formula for calculating the sample enhanced feature data is as follows: ; ; ; ; ; ; In the formula, represents the output of the reset gate; represents the weight matrix of the reset gate; represents the hidden state at the th sampling moment; represents the data corresponding to the th sampling moment in the sample propagation feature data.In actual operation, the sample strain data is time-series data, including the strain values of the sample welds collected at multiple sampling moments included in the target duration respectively. The strain values at multiple sampling moments form the sample strain data of the sample weld to reflect the deformation characteristics of the sample weld over time. Correspondingly, is also time-series data; represents the output of the update gate; represents the Sigmoid activation function; represents the weight matrix of the update gate; represents the strain accumulation factor used to enhance the long-term memory of historical damage; represents the cumulative gain; represents the data corresponding to the th sampling moment in the sample propagation feature data; represents within the time The maximum value; Indicates the L1 norm; Indicates the candidate hidden state; Indicates the weight matrix of the candidate hidden state; Indicates the Hadamard product; Indicates the hidden state at the

[0037] th sampling moment, that is, the sample enhanced feature data. Figure 4 As shown in Figure 4 FIG. is the data processing flow chart of the regularization layer and the fully connected layer provided by the embodiment of the present application. After S360, the method further includes: S370~S400, the details are as follows: S370: For each sample time-frequency matrix, the regularization layer determines the strain acceleration standard deviation according to the sample time-frequency matrix; Among them, the strain acceleration standard deviation indicates the boundary of the physical rationality that the reference regularization data should conform to; the reference regularization data is the regularization data obtained by simulation without passing through the processing of the regularization layer; the physical rationality characterizes the physical constraint.

[0038] Specifically, in the embodiment of the present application, the "reference regularization data" refers to the regularization data obtained by regularizing the sample enhanced feature data in the embodiment of the present application by a conventional regularization method (that is, without passing through the regularization layer provided by the embodiment of the present application).

[0039] In practical applications, there are fuzzy boundaries in the classification of the quality of welds, such as microcracks and stress concentrations. However, conventional regularization methods will destroy the physical continuity of the strain of the weld, that is, conventional regularization methods do not consider the physical constraints of the strain of the weld. Randomly discarding features will result in invalid regularization data that violates the material constitutive relationship. To solve this application dilemma, the embodiment of the present application guides the regularization process of the regularization layer of the present application by determining whether the regularization data obtained by regularizing through a conventional regularization method conforms to the physical constraints.

[0040] In actual operation, it is not necessary to actually use other regularization methods to regularize the sample enhanced feature data to obtain the reference regularization data. In the embodiments of the present application, the reference regularization data is constructed to indirectly determine whether the regularization data obtained by regularizing the sample enhanced feature data through the conventional regularization method conforms to the physical constraints through the reference regularization data. Since the embodiments of the present application do not actually use other regularization methods to regularize the sample enhanced feature data to obtain the reference regularization data, the reference regularization data is called "simulated" regularization data.

[0041] In the embodiments of the present application, the boundary of physical rationality is characterized by the standard deviation of strain acceleration; among them, the formula for calculating the standard deviation of strain acceleration is as follows: ; In the formula, is the synthetic correction strain value corresponding to the th sampling moment determined according to the sample strain data; represents the number of sampling moments.

[0042] S380: Determine the mask function value through the regularization layer according to the acceleration threshold coefficient, the standard deviation of strain acceleration, and the reference regularization data; Among them, the acceleration threshold coefficient is determined according to the material properties of the sample steel structure body and 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; Specifically, in the embodiments of the present application, after determining the standard deviation of strain acceleration, the acceleration threshold coefficient is combined to obtain the mask function value; among them, the acceleration threshold coefficient is determined according to the material properties of the sample steel structure body and 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.

[0043] Among them, the mask function has the following formula: If , being equal to 0 indicates that the reference regularization data exceeds the boundary of physical rationality; If , being equal to 1 indicates that the reference regularization data does not exceed the boundary of physical rationality; In the formula, represents the reference regularization data, , represents the reversible feature mapping; Denotes the use of a 1×1 convolution operation to propagate the sample feature data The reference regularization data obtained by projecting the response into the strain space; Denotes the second-order derivative of the reference regularization data with respect to time, and the discrete calculation uses the central difference method to characterize the strain acceleration; Denotes the acceleration threshold coefficient.

[0044] S390: Determine the sample regularization data through the regularization layer based on the sample enhanced feature data and the mask function value.

[0045] Specifically, in the embodiment of the present application, after determining the mask function value, apply the mask function value to the hidden state, multiply each element of the mask vector and the hidden state (i.e., the sample enhanced feature data) element by element, and fill the mean value of the time dimension of the hidden state (i.e., the sample enhanced feature data) at the positions where the mask is 0 to obtain the regularized hidden state, that is, the sample regularization data; where the formula for calculating the sample regularization data is as follows: ; ; In the formula, Denotes the sample regularization data; Denotes the mask vector, Composed of The composition dimension is consistent with the time series, and the time series refers to the sequence of multiple sampling moments corresponding to the sample strain data; Denotes a vector of all 1s with the same dimension as ; Denotes the mean value of the hidden state (i.e., the sample enhanced feature data) in the time dimension.

[0046] In actual operation, the quality detection model further includes a fully connected layer for determining the quality prediction result corresponding to the sample strain data according to the sample regularization data.

[0047] S400: Determine the quality prediction result through the fully connected layer according to the sample regularization data.

[0048] Specifically, in the embodiment of the present application, both the quality prediction result and the quality detection result are binary classification probabilities, used to characterize defects or no defects.

[0049] In one implementation, before S310, the method further includes: step (1), details are as follows: Step (1): Determine the quality label corresponding to each sample strain data according to the sample strain data.

[0050] Among them, the types of quality labels include: defective and non-defective.

[0051] Specifically, in the embodiments of the present application, sample annotation is performed according to the non-destructive testing results of the sample steel structure; among them, the sample strain data with non-destructive testing results meeting the expectations is determined as positive sample strain data, and the sample strain data with non-destructive testing results not meeting the expectations is determined as negative sample strain data; in actual operation, the content of sample annotation may include, in addition to the type of quality label, the type, location and severity level of the defect, forming a strain-label paired data set with a timestamp.

[0052] In one implementation, before S400, the method further includes: steps (1) to (4), the details are as follows: Step (1): For the positive sample strain data with the type of non-defective for each quality label, determine the physical focusing parameter according to the sample regularization data.

[0053] Specifically, in actual applications, the defects characterized by the sample strain data of the sample weld are extremely unbalanced. For example, the qualified products are much larger than the cracked products. In addition, the strain responses of micro-cracks and noise are slightly different. As a result, the conventional cross-entropy loss function has insufficient discrimination ability for difficult-to-separate samples near the decision boundary, that is, the conventional cross-entropy loss function does not consider the absolute threshold of the strain amount. To solve this 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 focusing parameter related to the physical threshold, and for the negative sample strain data, the loss weight combines the class balance factor and the focusing parameter related to the confidence level. The final loss is the mean of the losses of all samples to enhance the sensitivity to micro-defects and sample imbalance scenarios; among them, the formula for determining the physical focusing parameter is as follows: ; ; In the formula, represents the physical focusing parameter; represents the first scaling factor. In actual operation, can be set to 2; represents the maximum reconstruction strain value corresponding to the th sample strain data among sample strain data, is a positive integer representing the number of sample strain data; represents the physical damage threshold. In actual operation, can be set to 0.005; represents the reconstruction strain function for mapping the hidden state (i.e., the sample enhanced data) to the strain value; is the The sample regularization data corresponding to the sample strain data; Indicates at the th sampling moment the maximum value.

[0054] Step (3): For the negative sample strain data with the type of quality label being defective, determine the confidence focusing parameter according to the quality prediction result.

[0055] Specifically, in the embodiment of the present application, the formula for determining the confidence focusing parameter is as follows: ; In the formula, represents the confidence focusing parameter; represents the second scaling factor. In actual operation, can be set to 3; represents the quality prediction result the binary information entropy of.

[0056] Step (4): Determine the loss function value of the current training process according to the number of sample strain data with the type of quality label being defective, the number of sample strain data with the type of quality label being non-defective, multiple quality prediction results, multiple physical focusing parameters, and multiple confidence focusing parameters; wherein, 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.

[0057] Specifically, in the embodiment of the present application, the formula of the loss function for calculating the loss function value is as follows: ; ; ; In the formula, represents the loss function; represents the total number of sample strain data; represents the indicator function; represents when the th sample strain data is positive sample strain data, takes the value , otherwise represents ; represents when the th sample strain data is negative sample strain data, takes the value , otherwise represents ; represents the class balance factor; represents the number of negative sample strain data, represents the number of positive sample strain data; Indicates the quality prediction result of the n-th sample strain data; Indicates the weight of the fully connected layer; Indicates the sample regularization data corresponding to the n-th sample strain data.

[0058] In actual operation, the gradients of all trainable network parameters in the quality detection model are calculated according to the loss function value, and the network parameters are updated according to the calculated gradients; among them, the gradient calculation is implemented by automatic differentiation technology, and the error signal is reversely transmitted from the output layer to the input layer along the computational graph; the parameter update is immediately executed after each training process to ensure that the iterative optimization direction of the model always points towards minimizing the loss.

[0059] In one implementation, the sample strain data includes the original strain values collected at multiple sampling times; before S400, the method further includes: steps (5) to (9), details are as follows: Step (5): For each sample strain data, determine the adaptive time window length corresponding to the first sampling time according to the preset sampling frequency, the original strain values corresponding to the adjacent first sampling time and the second sampling time in the sample strain data, and the strain change rate threshold.

[0060] Among them, 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 the critical value of whether the original strain value is a mutant strain; the value of the adaptive time window length indicates the response speed to the mutant strain.

[0061] Specifically, in the embodiments of the present application, the sample strain data includes the original strain values collected at multiple sampling times; 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.

[0062] In practical applications, the multiple original strain values included in the sample strain data of the sample weld usually have baseline offset caused by equipment drift and high-frequency electromagnetic noise in the industrial environment, which is manifested as the superposition of a non-stationary low-frequency trend term and impulse-type high-frequency interference, making conventional normalization amplify the noise and mask the characteristics of small defects. Since traditional moving average filtering will smooth the mutant characteristics and wavelet threshold denoising requires preset basis functions, it is difficult to adapt to the application dilemma of the mixed mode of sudden plastic deformation and elastic recovery in weld strain. To solve this application dilemma, the embodiments of the present application process the original strain values through a dual-channel adaptive filtering mechanism to eliminate the baseline offset and high-frequency electromagnetic noise caused by equipment drift and retain the characteristics of small defects.

[0063] In actual operation, if it is necessary to eliminate the baseline offset and high-frequency electromagnetic noise caused by device drift and retain the characteristics of minute defects, it is first necessary to calculate the length of the adaptive time window corresponding to the first sampling moment; among them, the formula for calculating the length of the adaptive time window is as follows: ; In the formula, represents the length of the adaptive time window corresponding to the first sampling moment, and is used to dynamically adjust the integration interval to ensure a rapid response to sudden strain; represents the preset sampling frequency, which characterizes the acquisition rate of strain values. In actual operation, can be set to a value of 10 kHz; represents the strain change rate threshold, and the strain change rate threshold is used to determine the critical value of whether the original strain value is a sudden strain. In actual operation, can be set to .

[0064] represents the first sampling moment, that is, the original strain value corresponding to the moment; represents the second sampling moment, that is, the original strain value corresponding to the moment; represents rounding up.

[0065] Step (6): Determine the baseline value corresponding to the first sampling moment according to the strain rate threshold, the length of the adaptive time window, and the original strain value corresponding to the third sampling moment.

[0066] Among them, the baseline value is the value remaining after removing the high-frequency interference data from the original strain value corresponding to the first sampling moment.

[0067] Specifically, in the embodiment of the present application, after determining the length of the adaptive time window corresponding to the first sampling moment, the baseline drift can be separated by using the strain rate threshold, and only the strain values with a strain rate lower than the given strain rate threshold are integrated to obtain a low-frequency baseline value; among them, the formula for calculating the limit value is as follows: Baseline value = ; In the formula, is the integral of the strain value in the interval , which characterizes the time window integral; represents the third sampling moment, that is, the original strain value at the moment; represents the indicator function. In the term, characterizes the strain rate. When When, the value of the indicating function is 1 to retain the baseline; when When, the value of the indicating function is 0 to filter out the valid signal; represents the strain rate threshold, which can take the maximum strain rate in the elastic stage of the material; represents the derivative of the strain value with respect to the third sampling moment, that is, At the moment, it is calculated by discrete difference approximation and characterizes the strain rate.

[0068] Step (7): Determine the baseline-removed strain value corresponding to the first sampling moment according to the original strain value and the baseline value corresponding to the first sampling moment.

[0069] Specifically, in the embodiment of the present application, after determining the baseline value, the baseline drift is removed according to the baseline value to obtain the baseline-removed strain value to retain the high-frequency valid signal; among them, the formula for calculating the baseline strain value is as follows: ; In the formula, represents the first sampling moment, that is, The baseline-removed strain value corresponding to the moment.

[0070] Step (8): Determine the noise-suppressed strain value according to the baseline-removed strain value corresponding to the first sampling moment.

[0071] Among them, the noise-suppressed strain value is the remaining value after suppressing the noise of the original strain value corresponding to the first sampling moment.

[0072] Specifically, in the embodiment of the present application, impulse noise is suppressed by non-linear compression, and a compression operation combining a sign function and a natural logarithm is applied to the baseline-removed strain value. The compression coefficient is updated in real time by the mean value of the absolute values of the baseline-removed strain values within the time window to reduce the impulse amplitude; among them, the formula for calculating the noise-suppressed strain value is as follows: Noise-suppressed strain value = ; ; In the formula, represents the sign function; in the term, if , then the value of the sign function is 1, if , then the value of the sign function is 0; if , then the value of the sign function is -1; represents the impulse compression coefficient, which updates the noise level in real time; represents the time window The number of original strain values within; represents the natural logarithm function.

[0073] Step (9): Add the baseline value and the noise-suppressed strain value to obtain the composite corrected strain value corresponding to the first sampling moment.

[0074] Specifically, in the embodiment of the present application, after adding the baseline value and the noise-suppressed strain value corresponding to the first sampling moment, the composite corrected strain value corresponding to the first sampling moment can be obtained. In actual operation, other sampling moments included in the sample strain data can be sequentially determined as the first sampling moment to determine the composite corrected strain values corresponding to each sampling moment.

[0075] In one implementation, after step (9), the method further includes: steps (10) to (12), details are as follows: Step (10): Determine the stress value corresponding to the first sampling moment according to the material yield strain, material yield stress, Young's modulus in the elastic section, strain hardening modulus, hardening index, and the composite corrected strain value corresponding to the first sampling moment. Among them, the stress value indicates the stress capacity of the elastic section and / or the hardening section in the weld.

[0076] Specifically, in the embodiment of the present application, the defects in the sample weld are manifested as local micro-strain 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 linearly elastic and cannot handle the non-linear hardening effect in the heat-affected zone of the weld, resulting in distorted defect stress field estimation. 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.

[0077] 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 composite corrected strain value does not exceed the material yield strain, the Young's modulus in the elastic section is used to calculate the stress; if the composite corrected strain value exceeds the material yield strain, the plastic hardening section stress is calculated based on the yield stress, hardening modulus, and hardening index to distinguish the elastic section from the plastic hardening section; among them, the formula for determining the stress value corresponding to the first sampling moment is as follows: If , is equal to ; If , is equal to ; In the formula, represents the first sampling moment, that is, the stress value corresponding to the moment; represents the first sampling moment, that is the synthetic correction strain value corresponding to the moment; represents the material yield strain, which is determined by the material structure of the steel structure body. In actual operation, the value of can be set to 0.0017; represents the material yield stress, satisfying ; In actual operation, the value of can be 345×10 6 Pa; represents the Young's modulus in the elastic section, and the value is based on the steel standard. In actual operation, the value of can be set to 210×10 9 Pa; represents the strain hardening modulus, which is calibrated through a material tensile test. In actual operation, the value of can be set to 3; represents the hardening index, which characterizes the plastic deformation ability. In actual operation, the value of can be set to 0.2.

[0078] Step (11): Determine the normalized strain acceleration corresponding to the first sampling moment according to the synthetic correction strain value corresponding to the first sampling moment.

[0079] Specifically, in the embodiment of the present application, by calculating the normalized strain acceleration, the absolute value of the second-order derivative of the correction strain value with respect to time is divided by the larger value of the absolute value of the first-order derivative of time and the anti-zero constant to eliminate the influence of dimensions; among them, the formula for calculating the normalized strain acceleration is as follows: ; In the formula, represents the first sampling moment, that is the normalized strain acceleration corresponding to the moment; represents the second-order derivative of the synthetic correction strain value with respect to time. The discrete calculation uses the central difference method to characterize the strain change rate; represents the first-order derivative of the synthetic correction strain value with respect to time. The discrete calculation uses the forward difference method to characterize the strain rate; represents the anti-zero constant, which is used to avoid the denominator being zero. In actual operation, the value of can be set to 10 6 ; represents the function of taking the larger value of the two.

[0080] Step (12): Determine the damage feature vector corresponding to the first sampling moment according to the stress value and the normalized strain acceleration corresponding to the first sampling moment.

[0081] Specifically, in the embodiments of the present application, the fusion stress value and the normalized strain acceleration are combined, and a characteristic enhancement factor is constructed by combining the damage gain coefficient, the hyperbolic tangent function, and the time-domain mean value of the absolute value of the strain acceleration, and then a damage feature vector is constructed. The formula for calculating the damage feature vector is as follows: ; ; In the formula, represents the first sampling moment, that is, the damage feature vector corresponding to the moment; represents the damage gain coefficient, which controls the contribution weight of the strain acceleration. In actual operation, can be set to 0.3; represents the hyperbolic tangent function; represents the characteristic enhancement factor, which is defined as the time-domain mean value of the absolute value of the strain acceleration; represents the number of sampling moments corresponding to the sample strain data; represents the differential of;

[0082] In one implementation, after step (12), the method further includes steps (13) to (18), which are as follows: Step (13): Perform a short-time Fourier transform on the sample strain data to obtain a frequency point set including multiple frequency points.

[0083] Specifically, in actual applications, when a crack occurs in a weld seam, it will excite acoustic emission signals in a specific frequency band. However, this feature in the sample strain data is submerged by the main load response. It is necessary to focus on the sensitive frequency band in the time-frequency domain to highlight the frequency band related to damage. Traditional fixed-window time-frequency analysis methods such as the short-time Fourier transform cannot balance the resolution requirements for high-frequency transient events such as crack bursts and low-frequency continuous damage such as fatigue propagation, resulting in key features being blurred or omitted. To solve this application dilemma, the embodiments of the present application focus on the damage-sensitive frequency band through adaptive bandwidth transformation to adaptively optimize the time-frequency resolution and highlight the energy of the frequency band related to damage.

[0084] In actual operation, it is necessary to perform a short-time Fourier transform on the sample strain data to obtain a frequency point set including multiple frequency points.

[0085] Step (14): Determine the normalized power spectral density at the first frequency point at the first sampling moment according to the first frequency point and the second frequency point corresponding to the first sampling moment in the frequency point set.

[0086] Specifically, in the embodiments of the present application, the normalized power spectral density is calculated based on the energy ratio of the short-time Fourier transform coefficients; wherein, the formula for calculating the normalized power spectral density is as follows: ; ; In the formula, represents the normalized power spectral density at the first sampling moment and the th frequency point , and its value range represents , characterizing the proportion of the energy at this frequency point in the total energy; represents the output of the short-time Fourier transform; represents the complex-valued time-frequency coefficient at the first sampling moment, that is, the moment and the th frequency point ; represents the complex-valued time-frequency coefficient at the first sampling moment, that is, the moment and the th frequency point ; represents the th frequency point, and its value range represents to , covering the typical acoustic emission frequency band of weld cracks; represents the th frequency point, and its value range represents to , used for frequency point summation; represents a positive integer; represents the number of frequency points; represents rounding down; represents the logarithmic function, with the default base being 10.

[0087] Step (15): Determine the time-varying spectral entropy corresponding to the first sampling moment according to the normalized power spectral density corresponding to the first sampling moment.

[0088] Among them, the value of the time-varying spectral entropy indicates the signal complexity.

[0089] Specifically, in the embodiments of the present application, the time-varying spectral entropy is calculated using the normalized power spectral density to quantify the signal complexity; wherein, the formula for calculating the time-varying spectral entropy is as follows: ; In the formula, represents the time-varying spectral entropy at the first sampling moment, that is, the moment, and its value range is , the lower the value, the lower the signal complexity; represents a positive integer.

[0090] Step (16): Determine 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.

[0091] Among them, 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. Specifically, in the embodiment of the present application, the window width related to the frequency is dynamically calculated based on the time-varying spectral entropy and the window width adjustment factor. A narrow window is used for high-frequency analysis, and a wide window is used for low-frequency analysis; among them, the formula for calculating the window width parameter is as follows: ; In the formula, represents the window width parameter related to the frequency. The larger the value, the wider the window; represents the analysis frequency, and the value range is to , that is to ; represents the window width adjustment factor, which controls the overall scaling ratio of the window width. In actual operation, can be set to 0.35.

[0092] Step (17): Determine the window function output corresponding to the first sampling moment according to the window width parameter corresponding to the first sampling moment.

[0093] Specifically, in the embodiment of the present application, a Gaussian window function is constructed, and the window width is controlled by the window width parameter related to the frequency; among them, the window function for calculating the window function output is as follows: ; In the formula, represents the window function output, and the value range is , which is used to weight the local signal segment in the time-frequency analysis; represents the time difference, which represents the offset between the current time point and the integration variable, that is ; represents the integration variable; represents the natural exponential function; represents pi, approximately taken as 3.1416.

[0094] Step (18): Determine the time-frequency matrix at the first sampling moment and at the analysis frequency according to the damage feature vector corresponding to the first sampling moment and the window function output.

[0095] Specifically, in the embodiments of the present application, an adaptive bandwidth transformation is performed, and the damage feature vector is combined with a dynamic window function to generate a time-frequency matrix; wherein, the formula for generating the time-frequency matrix is as follows: ; In the formula, represents the time-frequency matrix, and represents at the first sampling moment, that is, at the moment and frequency the time-frequency coefficient at, and its amplitude characterizes the signal energy; represents the damage feature vector at the moment; represents the imaginary unit.

[0096] S200: Preprocess the target strain data to obtain a target time-frequency matrix.

[0097] Specifically, the steps of preprocessing the target strain data to obtain a target time-frequency matrix can refer to the steps of "determining the sample time-frequency matrix according to the sample strain data" mentioned above, and will not be elaborated here.

[0098] S300: Input the target time-frequency matrix into a pre-trained quality detection model, so that the quality detection model determines the target propagation feature data of the target strain wave according to the target time-frequency matrix, and determines the 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 steel structure in the form of a wave.

[0099] Specifically, in the embodiments of the present application, the quality detection result is a binary classification probability. If the quality detection result is greater than a preset threshold, it is considered that there are defects in the target weld and a warning is issued, otherwise it is considered that there are no defects in the target weld.

[0100] Now, to evaluate the processing effects of different noise suppression methods on the strain data of steel structure welds, by comparing the processing results of the original signal, traditional moving average filtering, and the dual-channel adaptive filtering method of the embodiments of the present application, the ability of each method to suppress baseline drift and impulse noise while retaining the characteristics of small crack mutation defects is emphasized. As Figure 5 shown, Figure 5 is a comparison diagram of the effects of different noise suppression methods provided by the embodiments 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 burrs. Although the traditional moving average filtering smooths the noise, it seriously weakens the crack mutation characteristics at 300 milliseconds, while the embodiments of the present application (i.e., Figure 5The technical method in [specific context] not only effectively eliminates baseline drift but also completely preserves key defect features. The red curve in the figure shows a steep peak at the crack position, and at the same time, it significantly suppresses random pulse interference. The abnormal peaks in the green curve are eliminated, demonstrating the advantage of the embodiment of this application in extracting weak damage features in a complex industrial noise environment.

[0101] Now, to evaluate the influence of different noise suppression techniques on the detection rate of weld defects, in the form of a box plot combined with a scatter plot, the performance of moving average filtering, wavelet threshold denoising, Kalman filtering, and the embodiment of this application under the same test conditions is compared, as Figure 6 shown. Figure 6 This is a comparison chart of the defect detection rates of different noise suppression methods provided by the embodiment of this application. The height of the box reflects the stability of the detection results, and the scatter distribution shows the results of multiple experiments. The embodiment of this application (i.e., Figure 6 the technical method in [specific context]) shows the highest median detection rate and the most concentrated high-value distribution. Its box position is significantly higher than other methods, and the scatter points are densely distributed in the top area, proving that the dual-channel adaptive filtering mechanism adopted by the embodiment of this application can effectively eliminate equipment drift and industrial electromagnetic noise, improve detection reliability while retaining tiny defect features, and solve the defects of traditional methods that either amplify noise or smooth mutation features.

[0102] Now, to analyze the influence of different strain-stress conversion methods on the detection accuracy of crack size, the changing trends of the detection errors of Hooke's law of linear elasticity, non-linear hardening model, and the embodiment of this application under different actual crack sizes are compared, as Figure 7 shown. Figure 7 This is the crack size detection error of different strain-stress conversion methods provided by the embodiment of this application. The shaded area emphasizes the error range. The error curve of the embodiment of this application (i.e., Figure 7 the technical method in [specific context]) is always at the bottom, and it maintains a stable low error as the crack size increases, especially showing obvious advantages in the tiny crack interval, verifying the effectiveness of the segmented stress mapping and damage feature enhancement mechanism of the embodiment of this application. By fusing stress values and normalized strain acceleration, it overcomes the limitations of traditional linear elastic assumptions in the heat-affected zone of the weld and significantly improves the recognition accuracy of tiny defects.

[0103] Now, the feature extraction effects of conventional short-time Fourier transform and the adaptive bandwidth transform of the embodiment of this application are visually compared through a time-frequency heat map, as Figure 8 shown. Figure 8 This is a comparison chart of the time-frequency domain feature extraction methods provided by the embodiment of this application. The upper part of the figure is the traditional short-time Fourier transform (STFT), whose frequency axis and time axis resolutions are fixed. The lower part of the figure is the embodiment of this application (i.e., Figure 8As a result of the present technology herein, a dynamically adjusted time-frequency window is adopted. The color intensity represents the signal energy, and the bright area corresponds to the damage feature. It can be seen from the experiment that the impact features at 100 Hz (a typical frequency band for crack generation) at 300 milliseconds and the features at 200 Hz (a frequency band for crack propagation) at 700 milliseconds presented by the traditional method show a blurred and diffused state, and the background noise interference is obvious. However, the energy of the feature points at the same position in the embodiment of the present application is concentrated and the boundary is clear, and the 150 Hz continuous damage feature (the yellow strip in the figure) completely submerged in the traditional method is separated, proving that the embodiment of the present application effectively solves the contradiction of the fixed resolution method in simultaneously capturing high-frequency transient events and low-frequency continuous damage through the frequency-related window width adjustment mechanism.

[0104] Currently, the receiver operating characteristic curve is used to evaluate the comprehensive performance of different detection models. As Figure 9 shown, Figure 9 This is a comparison chart of the ROC curves of different quality prediction models provided by the embodiment of the present application. 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 binary classification models 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 guessing benchmark. The curve of the model of the embodiment of the present application (i.e., Figure 9 the present technology herein) is closest to the upper left corner, still maintaining a high true positive rate in the low false positive rate region, and the area under the curve is the largest, reflecting the synergistic advantages of the physically guided direction-sensitive convolutional kernel, the gated recurrent unit enhanced by the strain accumulation factor, and the uncertainty regularization in the embodiment of the present application. By fusing the material constitutive relationship and the wave propagation characteristics, the ability to identify micro-cracks and fuzzy boundaries is strengthened under the condition of unbalanced samples, which is significantly superior to the traditional data-driven models.

[0105] Second, the present application provides a steel structure weld quality detection device based on artificial intelligence. As Figure 10 shown, Figure 10 This is a schematic structural diagram of the steel structure weld quality detection device based on artificial intelligence provided by the embodiment of the present application. The device includes: a data acquisition module 400 and a quality detection module 500; The data acquisition module 400 is used to perform a mechanical property test with a target duration on a target weld with a quality to be detected in a target steel structure body to obtain target strain data of the target weld; Among them, the strain data indicates the deformation characteristics of the weld within the target time; The quality detection module 500 is used to pre-process 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 the target strain wave according to the target time-frequency matrix, and determines the 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.

[0106] In one implementation, the quality detection model includes: a convolution layer; the apparatus further 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 a sample time-frequency matrix into the quality detection model, perform propagation convolution on the sample time-frequency matrix using a convolution kernel in the convolution layer for each sample time-frequency matrix 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; A training module is used 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 convolution layer; Among them, the energy flow vector indicates the energy propagation direction of the energy flow of the strain wave; A training module is used 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 used to determine the sample propagation characteristic data of the sample strain wave according to the directional characteristic map and directional weight corresponding to each simulated propagation direction through the convolution layer.

[0107] In one implementation, the quality detection model further includes: a gated recurrent unit; a training module for determining a strain accumulation factor based on the sample propagation characteristic data for each sample through the gated recurrent unit; Among them, the strain accumulation factor represents the memory degree of the sample weld to the damage caused by the impact; A training module, configured to perform feature enhancement processing on data representing sudden strain events in the sample propagation feature data according to a strain accumulation factor through a gated recurrent unit to obtain sample enhanced feature data; Wherein, the bias for calculating the update gate state value in the gated recurrent unit is the strain accumulation factor; the sudden strain event indicates the deformation generated after the weld receives an impact higher than a preset intensity within a time shorter than a preset duration.

[0108] In one implementation, the quality detection model further includes: a regularization layer and a fully connected layer; the training module is configured to, for each sample time-frequency matrix, determine the strain acceleration standard deviation according to the sample time-frequency matrix through the regularization layer; Wherein, the strain acceleration standard deviation indicates the boundary of the physical rationality that the reference regularization data should conform to; the reference regularization data is the regularization data obtained by simulation without passing through the processing of the regularization layer; the physical rationality represents physical constraints; The training module is configured to determine a mask function value according to an acceleration threshold coefficient, the strain acceleration standard deviation, and the reference regularization data through the regularization layer; Wherein, the acceleration threshold coefficient is determined according to the material properties of the sample steel structure body and is used to control the influence degree of the physical rationality on the reference regularization data; the mask function value indicates whether the reference regularization data exceeds the boundary of the physical rationality; The training module is configured to determine sample regularization data according to the sample enhanced feature data and the mask function value through the regularization layer; The training module is configured to determine a quality prediction result according to the sample regularization data through the fully connected layer.

[0109] In one implementation, the device further includes: an annotation module; The annotation module is configured to determine a quality label corresponding to each sample strain data according to the sample strain data; Wherein, the types of the quality labels include: defective and non-defective; The training module is further configured to, for each positive sample strain data with the quality label type of non-defective, determine a physical focusing parameter according to the sample regularization data; The training module is further configured to, for each negative sample strain data with the quality label type of defective, determine a confidence focusing parameter according to the quality prediction result; The training module is further configured to determine the loss function value of the current training process according to the number of sample strain data with the quality label type of defective, the number of sample strain data with the quality label type of non-defective, multiple quality prediction results, multiple physical focusing parameters, and multiple confidence focusing parameters; Among them, 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.

[0110] In one implementation, the sample strain data includes the original strain values respectively collected at multiple sampling moments; the training module is further configured to, for each sample strain data, determine the adaptive time window length corresponding to the first sampling moment according to the preset sampling frequency, the original strain values respectively corresponding to the adjacent first sampling moment and the second sampling moment in the sample strain data, and the strain change rate threshold. Wherein, the preset sampling frequency is the sampling frequency when collecting the sample strain data; the first sampling moment is later than the second sampling moment; the strain change rate threshold is used to determine the critical value of whether the original strain value is a mutation strain; the value of the adaptive time window length indicates the response speed to the mutation strain. The training module is further configured to determine the baseline value corresponding to the first sampling moment according to the strain rate threshold, the adaptive time window length, and the original strain value corresponding to the third sampling moment. Wherein, the baseline value is the value remaining after removing the high-frequency interference data from the original strain value corresponding to the first sampling moment. The training module is further configured to determine the baseline-removed strain value corresponding to the first sampling moment according to the original strain value and the baseline value corresponding to the first sampling moment. The training module is further configured to determine the noise-suppressed strain value according to the baseline-removed strain value corresponding to the first sampling moment; wherein, the noise-suppressed strain value is the value remaining after suppressing the noise from the original strain value corresponding to the first sampling moment. The training module is further configured to add the baseline value and the noise-suppressed strain value to obtain the synthesized corrected strain value corresponding to the first sampling moment.

[0111] In one implementation, the training module is further configured to determine the stress value corresponding to the first sampling moment according to the material yield strain, the material yield stress, the Young's modulus in the elastic section, the strain hardening modulus, the hardening index, and the synthesized corrected strain value corresponding to the first sampling moment. Wherein, the stress value indicates the stress capacity in the elastic section and / or the hardening section in the weld. The training module is further configured to determine the normalized strain acceleration corresponding to the first sampling moment according to the synthesized corrected strain value corresponding to the first sampling moment. The training module is further configured to determine the damage feature vector corresponding to the first sampling moment according to the stress value and the normalized strain acceleration corresponding to the first sampling moment.

[0112] In one implementation, the training module is further configured to perform a short-time Fourier transform on the sample strain data to obtain a frequency point set including multiple frequency points. The training module is further configured to determine, according to a first frequency point and a second frequency point corresponding to a first sampling moment in the frequency point set, a normalized power spectral density at the first frequency point at the first sampling moment; The training module is further configured to determine, according to the normalized power spectral density corresponding to the first sampling moment, a time-varying spectral entropy corresponding to the first sampling moment; Wherein, the value of the time-varying spectral entropy indicates the signal complexity; 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 moment, a window width parameter corresponding to the first sampling moment; Wherein, 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; The training module is further configured to determine, according to the window width parameter corresponding to the first sampling moment, a window function output corresponding to the first sampling moment; The training module is further configured to determine, according to the damage feature vector corresponding to the first sampling moment and the window function output, a time-frequency matrix at the first sampling moment and at the analysis frequency.

[0113] Thirdly, the present application further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of S100~S300 provided in the above embodiments are implemented.

[0114] Fourthly, the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program stored on the computer-readable medium is run by a processor, the steps of S100~S300 in the above embodiments are executed.

[0115] Fifthly, the computer program product provided by the present application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the foregoing method embodiments. For specific implementation, reference can be made to the steps of S100~S300 in the method embodiments, which will not be elaborated herein.

[0116] In the embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces. The indirect coupling or communication connection of the devices or units may be in an electrical, mechanical or other form.

[0117] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0118] Furthermore, in each embodiment of the present application, the various functional modules may be integrated together to form an independent part, or each module may exist alone, or two or more modules may be integrated to form an independent part.

[0119] It should be noted that if a function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may 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: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0120] In this document, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0121] The above description is only for the embodiments of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An artificial intelligence-based steel structure weld quality inspection method, characterized in that, The method includes: Performing a mechanical property test with a target duration on a target weld to be detected for quality in a target steel structure to obtain target strain data of the target weld; Wherein, the strain data indicates the deformation characteristics presented by the weld during 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 generated by the weld propagates in the steel structure in the form of a wave.

2. The method according to claim 1, wherein 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 inputting a sample time-frequency matrix into the quality detection model, for each sample time-frequency matrix, performing propagation convolution on the sample time-frequency matrix through a convolutional kernel in the convolutional layer to obtain direction feature maps corresponding to respective simulated propagation directions; Wherein, the sample time-frequency matrix is obtained by performing the preprocessing on sample strain data; the sample strain data is collected based on a sample weld in a sample steel structure; the convolutional 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 performing a convolution operation based on convolution kernel values determined by the simulated propagation speed of the sample strain wave in the sample steel structure and a plurality of the simulated propagation directions; Determining, by the convolutional layer, energy flow vectors corresponding to respective simulated propagation directions according to the direction feature maps corresponding to respective simulated propagation directions; Wherein, the energy flow vector indicates the energy propagation direction of the energy flow of the strain wave; Determining, by the convolutional layer, direction weights corresponding to respective simulated propagation directions according to the energy flow vectors corresponding to respective simulated propagation directions; Determining, by the convolutional layer, sample propagation characteristic data of the sample strain wave according to the direction feature maps and the direction weights corresponding to respective simulated propagation directions.

3. The method according to claim 2, wherein The quality detection model further includes: a gated recurrent 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 by the gated recurrent unit according to the sample propagation characteristic data; Wherein, the strain accumulation factor characterizes the memory degree of the sample weld for the damage caused by being impacted; Performing feature enhancement processing on data in the sample propagation characteristic data that characterizes a sudden strain event by the gated recurrent unit according to the strain accumulation factor to obtain sample enhanced feature data; Wherein, the bias for calculating the update gate state value in the gated recurrent unit is the strain accumulation factor; the sudden strain event indicates the deformation generated by the weld after receiving an impact with a strength higher than a preset strength within a time shorter than a preset duration.

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

5. The method according to claim 3, wherein Before performing iterative training on the quality detection model, the method further includes: Determine the quality label corresponding to each sample strain data according to the sample strain data; Wherein, the types of the quality labels include: defective and non-defective; After the fully connected layer determines the quality prediction result according to the sample regularization data, the method includes: For each positive sample strain data with the type of the quality label being non-defective, determine the physical focusing parameter according to the sample regularization data; For each negative sample strain data with the type of the quality label being defective, determine the confidence focusing parameter according to the quality prediction result; Determine the loss function value of the current training process according to the number of the sample strain data with the type of the quality label being defective, the number of the sample strain data with the type of the quality label being non-defective, multiple quality prediction results, multiple physical focusing parameters, and multiple physical focusing parameters; Wherein, 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 the original strain values respectively collected at multiple sampling moments; Before performing iterative training on the quality detection model, the method further includes: For each sample strain data, determine the adaptive time window length corresponding to the first sampling moment according to the preset sampling frequency, the original strain values respectively corresponding to the adjacent first sampling moment and the second sampling moment in the sample strain data, and the strain change rate threshold; Wherein, the preset sampling frequency is the sampling frequency when collecting the sample strain data; the first sampling moment is later than the second sampling moment; the strain change rate threshold is used to determine the critical value of whether the original strain value is a mutant strain; the value of the adaptive time window length indicates the response speed to the mutant strain; Determine the baseline value corresponding to the first sampling moment according to the strain rate threshold, the adaptive time window length, and the original strain value corresponding to the third sampling moment; Wherein, the baseline value is the value remaining after removing the high-frequency interference data from the original strain value corresponding to the first sampling moment; Determine the baseline-removed strain value corresponding to the first sampling moment according to the original strain value and the baseline value corresponding to the first sampling moment; Determine the noise-suppressed strain value according to the baseline-removed strain value corresponding to the first sampling moment; wherein, the noise-suppressed strain value is the value remaining after suppressing the noise from the original strain value corresponding to the first sampling moment; Add the baseline value and the noise-suppressed strain value to obtain the synthetic corrected strain value corresponding to the first sampling moment.

7. The method according to claim 6, wherein After adding the baseline value and the noise-suppressed strain value to obtain the synthetic corrected strain value corresponding to the first sampling moment, the method further includes: Determine the stress value corresponding to the first sampling moment according to the material yield strain, the material yield stress, the Young's modulus in the elastic section, the strain hardening modulus, the hardening index, and the synthetic corrected strain value corresponding to the first sampling moment; Wherein, the stress value indicates the stress capacity in the elastic section and / or the hardening section in the weld; Determine the normalized strain acceleration corresponding to the first sampling moment according to the synthetic corrected strain value corresponding to the first sampling moment; Determine the damage feature vector corresponding to the first sampling moment according to the stress value and the normalized strain acceleration corresponding to the first sampling moment.

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: Perform short-time Fourier transform on the sample strain data to obtain a frequency point set including a plurality of frequency points; Determine the normalized power spectral density at the first frequency point corresponding to the first sampling moment in the frequency point set according to the first frequency point and the second frequency point corresponding to the first sampling moment; Determine the time-varying spectral entropy corresponding to the first sampling moment according to the normalized power spectral density corresponding to the first sampling moment; Wherein, the value of the time-varying spectral entropy indicates the signal complexity; Determine 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; Wherein, 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; Determine the window function output corresponding to the first sampling moment according to the window width parameter corresponding to the first sampling moment; Determine the time-frequency matrix at the first sampling moment and at the analysis frequency according to the damage feature vector corresponding to the first sampling moment and the output of the window function.

9. An artificial intelligence-based steel structure weld quality detection device, characterized in that, The device includes: a data acquisition module and a quality detection module; The data acquisition module is configured to perform a mechanical property test with a target duration on a target weld to be detected for quality in a target steel structure to obtain target strain data of the target weld; Wherein, the strain data indicates the deformation characteristics presented by the weld during the target duration; The quality detection module is configured to preprocess the target strain data to obtain a target time-frequency matrix; Input 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 steel structure in the form of a wave.

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 the iterative training of the quality detection model, in the current training process before reaching the training stop condition, after inputting a sample time-frequency matrix into the quality detection model, for each sample time-frequency matrix, perform propagation convolution on the sample time-frequency matrix through a convolutional kernel in the convolutional layer to obtain direction feature maps corresponding to each simulated propagation direction; Wherein, 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 convolutional 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; propagation convolution indicates performing a convolution operation based on the convolution kernel values determined by 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 an energy flow vector corresponding to each simulated propagation direction through the convolutional layer according to the direction feature maps corresponding to each simulated propagation direction; Wherein, 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 a direction weight corresponding to each simulated propagation direction through the convolutional layer according to the energy flow vectors corresponding to each simulated propagation direction; The training module is further configured to determine sample propagation feature data of the sample strain wave through the convolutional layer according to the direction feature maps and direction weights corresponding to each simulated propagation direction.

Images