A deep learning defect identification method for a PDC drill bit cladding process

Abstract

This application relates to the field of deep learning technology, and in particular to a deep learning defect identification method for the cladding process of PDC drill bits. The method includes: acquiring a sequence of process parameters during the cladding process and surface images after cladding; using classification labels generated from the surface images to train a defect identification model in the first stage, obtaining a first identification model; locating the defect time at the defect position in the process parameter sequence and calculating the saliency index of the defect position; constructing a time-series correction weight based on the defect time and saliency index to determine a second loss; and training the first identification model in the second stage based on the second loss. The resulting second identification model is used to map the process parameter sequence of the cladding process to the defect identification result. Through the technical solution of this application, the second identification model can accurately learn the true characteristics of defect formation, thereby improving the accuracy of PDC drill bit cladding defect identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video transmission, and in particular to a deep learning defect identification method for a PDC drill bit cladding process. BACKGROUND

[0002] The PDC drill bit is a tool for efficient drilling by relying on the diamond compact. Due to the excellent wear resistance and impact resistance of the diamond compact, it is often applied to the fields of geological exploration and oil and gas drilling. Laser cladding technology is a key process for enhancing the surface performance of the PDC drill bit, and the processing quality directly determines the service life and operation efficiency of the drill bit. In order to ensure the quality of the PDC drill bit, the cladding process needs to be monitored to determine whether the PDC drill bit after the cladding process has defects.

[0003] At present, the patent application file with the application publication number CN106841177A discloses a defect online diagnosis method in a laser cladding process, which comprises: pretreating a workpiece to be cladded; adjusting the position of a fiber probe; when the cladding process starts, collecting the photo-induced plasma spectrum signal through the fiber probe and sending it into a computer; observing the relative radiation intensity fluctuation of photo-induced plasma of different wavelengths with time and the distribution of element characteristic peaks on the computer, determining a plurality of characteristic element spectral lines as analysis objects in combination with the parameters corresponding to the element characteristic peaks; using the above spectral lines, determining the plasma temperature in real time according to the plasma temperature calculation method, and making a time-domain graph corresponding to the time; combining the above time-domain graph, judging whether the photo-induced plasma temperature in the cladding process has sharp fluctuations or changes; if yes, it means that there is a manufacturing defect; otherwise, it means that there is no manufacturing defect.

[0004] The above method relies on the fluctuation of the photo-induced plasma temperature for discrimination, while the laser cladding process is a complex process jointly influenced by multiple process parameters such as laser power, scanning speed, and protective gas flow. Simple temperature monitoring is difficult to capture complex defects caused by the joint action of multiple process parameters, and it is difficult to meet the demand for high accuracy of defect identification in the PDC drill bit cladding process. SUMMARY

[0005] In order to solve the technical problem of low accuracy of defect identification in the PDC drill bit cladding process, the present application provides a deep learning defect identification method for a PDC drill bit cladding process, which can enable the second identification model to accurately learn the real features of the defects, thereby improving the accuracy of PDC drill bit cladding defect identification.

[0006] In a first aspect, the present application provides a deep learning defect identification method for a PDC bit cladding process, the identification method comprising: collecting a process parameter sequence of the cladding process and a surface image after the cladding is completed; performing first stage training of a defect identification model using classification labels generated from the surface image, to obtain a first identification model; locating a defect time corresponding to a defect position in the process parameter sequence based on a mapping relationship between the defect position in the surface image and the process parameter sequence in time sequence, and calculating a saliency index of the defect position relative to a preset standard template; constructing a time sequence correction weight based on the defect time and the saliency index, the time sequence correction weight being used to represent a contribution degree of each processing time in the process parameter sequence to defect formation; performing weighted constraint of a first loss of the first stage training using the time sequence correction weight, to obtain a second loss, and performing second stage training of the first identification model according to the second loss, to obtain a second identification model, the second identification model being used to map the process parameter sequence of the cladding process to a defect identification result.

[0007] By collecting the process parameter sequence and the surface image, and constructing the time sequence correction weight based on the time sequence mapping relationship between the defect position in the surface image and the process parameter sequence, the contribution degree of each processing time in the process parameter sequence to defect formation can be quantified, and the focus of the model can be constrained at the processing time with high contribution degree through the second loss function, so that the second identification model obtained finally can learn the real characteristics of defect occurrence, and the recognition accuracy of the second identification model under complex cladding conditions is improved.

[0008] Preferably, the first stage training of the defect identification model using the classification labels generated from the surface image comprises: performing threshold segmentation and morphological filtering on a frame difference image of the surface image and a standard template, extracting a connected domain of a suspected defect region, if an area of the connected domain exceeds a preset threshold, marking the process parameter sequence as defective, otherwise marking the process parameter sequence as non-defective, to obtain the classification labels; using the process parameter sequence as input and the classification labels as supervision information, updating the defect identification model using a binary cross-entropy loss function.

[0009] By performing threshold segmentation and morphological filtering on the frame difference image of the surface image and the standard template to extract the connected domain of the suspected defect region, and generating the classification labels according to the area of the connected domain, automatic label making is realized, without manual annotation.

[0010] Preferably, the first identification model is obtained by: monitoring a loss value change rate of the defect identification model on a test set in real time; calculating a loss value reduction amount in a preset monitoring window; in response to the loss value reduction amount being less than a preset reduction amount for continuous multiple times, or a training round reaching a first maximum round, stopping the first stage training to obtain the first identification model.

[0011] By monitoring the loss value change rate of the defect identification model on the test set in real time, and stopping the first stage training when the loss value reduction amount is less than the preset value for a plurality of times in succession, adaptive control of the training process is realized, overfitting of the first stage is prevented, and meanwhile the first identification model learns the mapping relationship between the defect identification result and the process parameter sequence.

[0012] Preferably, the locating the defect moment corresponding to the defect position in the process parameter sequence comprises: recording a movement trajectory of the cladding head in the cladding process, the movement trajectory comprising a processing position at any processing moment, and querying the movement trajectory according to the defect position to obtain a processing moment corresponding to the defect position as the defect moment.

[0013] The correspondence between the defect position on the surface image and the defect moment in the process parameter sequence is established, so that the model can accurately locate the processing moment causing the defect at the spatial position, and an accurate time reference is provided for subsequent construction of high-precision time correction weights.

[0014] Preferably, the calculating the saliency index of the defect position relative to the preset standard template comprises: establishing a mapping relationship between a surface image coordinate system and a workpiece coordinate system, aligning the surface image with the standard template; determining a region of interest of the defect position, calculating the absolute value of the gray value difference between the surface image and the standard template in the region of interest; and normalizing the average value of the absolute value to obtain a saliency index representing the saliency of the defect, the region of interest being a rectangular region of a preset size centered on the defect position.

[0015] The saliency index reflects the saliency of the defect, provides data reference for the construction of the time correction weight, and ensures that a greater gradient penalty can be generated at the processing moment corresponding to the serious defect in the second stage training process, prompting the model to preferentially focus on the processing moment corresponding to the major defect.

[0016] Preferably, the constructing the time correction weight based on the defect moment and the saliency index comprises: constructing a Gaussian distribution function centered on any defect moment to obtain a Gaussian weight of each processing moment in a neighborhood period corresponding to the defect moment, multiplying the Gaussian weight and the saliency index of the defect position corresponding to the defect moment to obtain a correction weight of each processing moment in the neighborhood period corresponding to the defect moment; and calculating the sum of the correction weights of any processing moment in the process parameter sequence, and normalizing the sum to obtain the time correction weight of each processing moment.

[0017] The Gaussian distribution simulates the diffusion characteristics of the heat-affected zone in the cladding process, so that the time correction weight of any processing moment is no longer limited to a single moment, but covers all defect moments and their neighborhood periods, ensuring the accuracy of the time correction weight of each processing moment.

[0018] Preferably, before the second stage training of the first recognition model according to the second loss, the recognition method further comprises: inputting the process parameter sequence with the classification label of defects into the first recognition model, calculating the attention degree of each processing time by using the gradient class activation mapping algorithm; in response to the correlation between the attention degree and the timing correction weight being less than a correlation threshold, starting the second stage training; otherwise, taking the first recognition model as the second recognition model.

[0019] A quality gating mechanism is provided, which can automatically determine the necessity of the second stage training.

[0020] Preferably, the obtaining of the second loss comprises: obtaining the attention degree of each processing time of the process parameter sequence by the first recognition model; calculating the mean square error between the attention degree of each processing time and the timing correction weight as a consistency loss; and taking the weighted sum of the first loss and the consistency loss as the second loss.

[0021] Preferably, the defect recognition model is a recurrent neural network.

[0022] Preferably, the process parameters include at least one of laser power, scanning speed, molten pool temperature and protective gas flow.

[0023] The technical solution of the present application has the following beneficial technical effects:

[0024] By collecting the process parameter sequence of the cladding process and the surface image after the cladding is completed, and using the defect position in the surface image to deduce the defect time and the saliency index in the process parameter sequence, and then constructing the timing correction weight to constrain the second stage training of the defect recognition model, the timing correction weight can represent the contribution degree of each processing time in the process parameter sequence to the formation of defects, and then correct the attention distribution of the defect recognition model in the second stage training, forcing the trained second recognition model to pay attention to the processing time that really leads to the formation of defects, thereby solving the problem that the defect recognition model in the traditional training process easily learns false features or noise features irrelevant to defect recognition, and significantly improving the accuracy and robustness of the defect recognition of the PDC drill bit cladding process. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a flowchart of a deep learning defect recognition method for a PDC drill bit cladding process according to an embodiment of the present application.

[0026] Figure 2 is a loss value change curve in the first stage training and the second stage training process of the defect recognition model according to an embodiment of the present application.

[0027] Figure 3is a contrastive diagram of a timing correction weight distribution and an attention distribution according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] According to a first aspect of the present application, the present application provides a deep learning defect identification method for a PDC drill bit cladding process. Figure 1 is a flowchart of a deep learning defect identification method for a PDC drill bit cladding process according to an embodiment of the present application. As shown in Figure 1 , the deep learning defect identification method for a PDC drill bit cladding process includes steps S101 to S104, which are described in detail below.

[0029] S101, collect a process parameter sequence of the cladding process and a surface image after the cladding is completed, use a classification label generated by the surface image to perform first-stage training on a defect identification model, and obtain a first identification model.

[0030] In one embodiment, the process parameter sequence refers to device operation data recorded in chronological order during the PDC drill bit cladding process, the surface image refers to an appearance picture after the cladding is completed obtained by a visual sensor such as an industrial camera or a mobile phone, and the defect identification model is used to establish a neural network of the process parameter sequence and a defect identification result, and the defect identification model is a recurrent neural network such as LSTM or GRU.

[0031] On a laser cladding production line of a PDC drill bit, a process parameter sequence of the cladding process and a surface image after the cladding is completed are collected, wherein the process parameters include at least one of laser power, scanning speed, molten pool temperature, and protective gas flow rate, and in order to eliminate the influence of different physical dimensions, after the process parameter sequence is collected, the collected process parameter sequence also needs to be normalized, and at this time the process parameter sequence can reflect the change of the PDC drill bit cladding process at each process parameter.

[0032] In an embodiment of the present application, the defect identification model uses a long short-term memory network (LSTM network), and uses a gating mechanism in the long short-term memory network to capture long-time dependence features caused by heat accumulation effects in the cladding process.

[0033] One process parameter sequence corresponds to one surface image, the surface image is identified to determine whether there is a defect and the position of the defect, and the determination result of whether there is a defect is used as a classification label for training the defect identification model.

[0034] Specifically, the first-stage training of the defect recognition model using classification labels generated from surface images includes: thresholding and morphological filtering of the frame difference map between the surface image and the standard template to extract connected components of suspected defect regions. If the area of ​​the connected component exceeds a preset threshold, the process parameter sequence is marked as defective, and the suspected defect region is the defect location; otherwise, it is marked as defect-free, thus obtaining classification labels. The process parameter sequence is used as input, and the classification labels are used as supervision information. The defect recognition model is updated using a binary cross-entropy loss function. The binary cross-entropy loss function is the first loss.

[0035] The standard template is a standard appearance image of the PDC drill bit after cladding when there are no defects. By calculating the frame difference map between the surface image and the standard template, the interference of normal areas on defect identification can be accurately eliminated. The threshold segmentation can use the maximum inter-class variance method to locate the suspected defect areas in the frame difference map, and morphological filtering is used to remove minor noise interference to ensure the accuracy of classification labels.

[0036] In other embodiments, considering the influence of complex factory environments such as lighting, manual annotation can also be used to annotate defective areas in the surface image. If defective areas exist, the surface image is classified as defective.

[0037] After obtaining the classification labels of the process parameter sequence, the defect identification model undergoes its first stage of training. To prevent overfitting or undertraining, reasonable stopping conditions need to be set for the first stage of training. Obtaining the first identification model includes: real-time monitoring of the rate of change of the loss value of the defect identification model on the test set; calculating the reduction in the loss value within a preset monitoring window; and stopping the first stage of training in response to the loss value reduction being less than a preset reduction multiple times consecutively, or the training rounds reaching the first maximum round, thus obtaining the first identification model.

[0038] It should be noted that the preset monitoring window can be set to 3 training cycles, and the preset reduction amount can be set to 0.001; the value of the first maximum round is 50. When the loss value no longer decreases significantly, it indicates that the model has learned the basic mapping relationship between the process parameter sequence and the defect identification result.

[0039] Thus, by collecting the process parameter sequence of the cladding process and the surface image after cladding, and performing automated first-stage training, a basic mapping relationship between the process parameter sequence and the defect identification results was established, providing a data foundation for the subsequent second-stage training.

[0040] S102, based on the temporal mapping relationship between the defect location in the surface image and the process parameter sequence, locate the defect time corresponding to the defect location in the process parameter sequence, and calculate the significance index of the defect location relative to the preset standard template.

[0041] In one embodiment, locating the defect time corresponding to the defect location in the process parameter sequence includes: recording the movement trajectory of the cladding head during the cladding process, the movement trajectory including the processing position at any processing time, querying the movement trajectory based on the defect location, and obtaining the processing time corresponding to the defect location as the defect time.

[0042] Further quantifying the saliency of defects at each defect location, the calculation of the saliency index of the defect location relative to the preset standard template includes: aligning the surface image with the standard template; determining the region of interest at the defect location, calculating the absolute value of the grayscale difference between each pixel in the surface image and the standard template within the region of interest; normalizing the average value of the absolute values ​​to obtain a saliency index characterizing the saliency of the defect, wherein the region of interest is a rectangular area of ​​a preset size centered on the defect location.

[0043] The region of interest is a rectangular area of ​​5×5 or 7×7.

[0044] Defect location Significance index The following relationship must be satisfied:

[0045] ;

[0046] in, Defect location Region of interest Region of Interest The total number of pixels within, Coordinates within the region of interest grayscale value at that location For standard templates in coordinates The grayscale value at that location.

[0047] By calculating and normalizing the mean of grayscale differences within the region of interest, the significance of defects at each defect location is accurately quantified. The larger the value, the more significant the defect, providing a quantitative basis for the subsequent construction of the second loss.

[0048] S103, construct time-series correction weights based on defect time and significance index, wherein the time-series correction weights are used to characterize the contribution of each processing time in the process parameter sequence to defect formation.

[0049] In one embodiment, each processing moment in the process parameter sequence corresponds to a temporal correction weight, which characterizes the contribution of each processing moment in the process parameter sequence to defect formation. Specifically, constructing the temporal correction weight based on the defect moment and the significance index includes: constructing a Gaussian distribution function centered on any defect moment to obtain the Gaussian weight of each processing moment in the neighborhood time interval corresponding to the defect moment; multiplying the Gaussian weight by the significance index of the defect position corresponding to the defect moment to obtain the correction weight of each processing moment in the neighborhood time interval corresponding to the defect moment; calculating the sum of the correction weights of any processing moment in the process parameter sequence, and normalizing the summation result to obtain the temporal correction weight of each processing moment.

[0050] In the process parameter sequence, the process parameters at the defect time are the main cause of the defect. For any defect time, the larger the significance index at the defect location corresponding to that defect time, the more significant the defect caused by the process parameters at that defect time is. Therefore, the process parameters at that defect time contribute more to the formation of the defect.

[0051] Defect moment For example, the moment of defect Any processing time within the corresponding neighborhood time period Correction weights Satisfying the relation:

[0052] ;

[0053] in, For the moment of defect , For the moment of defect The significance index corresponding to the defect location, The thermal response constant is defined as the neighborhood time interval, which is the defect time in the process parameter sequence. Multiple processing times on the left and right sides; the thermal response constant is a preset non-zero constant, which can be set to 1 in this embodiment.

[0054] After obtaining the corrected weights for each processing time within the neighborhood time interval corresponding to all defect times, a processing time in the process parameter sequence may correspond to 0 corrected weights or one or more corrected weights. The processing time is then calculated. The summation result is obtained by summing all the corresponding corrected weights. Calculate processing time The summation result The ratio of this ratio to the sum of all processing times yields the processing time. The time-series correction weight.

[0055] In this way, the contribution of each defect moment in the process parameter sequence to defect formation can be deduced from the defect location in the surface image, indicating the key moment in the process parameter sequence that truly leads to the defect, and providing a clear optimization target for correcting the attention of the first recognition model in the second stage of training.

[0056] S104, the first loss of the first stage training is weighted and constrained by the time-series correction weight to obtain the second loss, and the first identification model is trained in the second stage based on the second loss to obtain the second identification model. The second identification model is used to map the process parameter sequence of the cladding process to the defect identification result.

[0057] In one embodiment, before performing the second stage of training, it is necessary to determine whether the second stage of training is necessary. Before performing the second stage of training on the first recognition model based on the second loss, the recognition method further includes: inputting a sequence of process parameters labeled as defective into the first recognition model, calculating the attention level at each processing moment using a gradient class activation mapping algorithm; initiating the second stage of training in response to the correlation between the attention level and the time-series correction weight being less than a correlation threshold; otherwise, using the first recognition model as the second recognition model.

[0058] It should be noted that after inputting the process parameter sequence labeled as defective into the first identification model, the gradient class activation mapping algorithm calculates the attention level by calculating the gradient magnitude of the first loss at each processing moment in the process parameter sequence. The larger the gradient magnitude of a processing moment, the greater the attention level of the first identification model to that processing moment. It should also be noted that the sum of the attention levels for each processing moment is 1.

[0059] The attention levels at each processing moment constitute an attention sequence, and the time-series corrected weights at each processing moment constitute a time-series corrected weight sequence. The absolute value of the Pearson correlation coefficient between the attention sequence and the time-series corrected weight sequence is calculated as the correlation between attention and time-series corrected weights. This correlation is used to measure the overlap between the processing moment that the first identification model focuses on when identifying defects and the processing moment that leads to the formation of defects. If the correlation is less than the correlation threshold, it indicates that the first identification model cannot accurately learn the real features that lead to the formation of defects, and a second stage of training is required for the first identification model. The correlation threshold is 0.6.

[0060] To enable the first identification model to focus on the processing moments that lead to defect formation during the second stage of training, a second loss containing consistency constraints needs to be constructed. Obtaining the second loss involves: acquiring the attention level of the first identification model to each processing moment in the process parameter sequence; calculating the mean square error between the attention level of each processing moment and the time-series correction weights as the consistency loss; and using the weighted sum of the first loss and the consistency loss as the second loss.

[0061] Specifically, the second loss Satisfying the relation:

[0062] ;

[0063] in, The first loss is the binary cross-entropy loss function. For balance coefficient, This represents the number of processing moments in the process parameter sequence. For the first identification model, the processing time attention, For processing time The time-series correction weights are calculated. The balance coefficient is set to 0.5.

[0064] Through consistency loss The attention distribution of the first recognition model is forced to approximate the physical truth distribution, which is the distribution of the temporal correction weights. If the first recognition model does not pay enough attention to the processing moment that leads to the formation of defects, the consistency loss will increase significantly. Thus, in backpropagation, gradient descent forces the first recognition model to adjust its parameters and shift its attention to the processing moment that leads to the formation of defects, so that the first recognition model can accurately learn the real features of defect formation and thus improve the accuracy of defect recognition.

[0065] The first recognition model is trained in the second stage based on the second loss until the value of the second loss is less than the preset loss or the number of training rounds reaches the second maximum number of rounds, at which point the trained second recognition model is obtained; the value of the second maximum number of rounds is 50.

[0066] Please see Figure 2 The graph shows the loss value change curves during the first and second training stages of the defect identification model according to embodiments of this application. The first training stage is stopped early when the loss value of the first loss decreases to approximately 0.3, resulting in the first identification model. When the second training stage begins, the first identification model continues to be trained, and under the constraint of the second loss, the loss value further decreases until the second identification model is obtained. After inputting the process parameter sequence labeled as defective into the second identification model, the time-series corrected weight distribution of the process parameter sequence and the attention distribution of the second identification model are obtained, as shown below. Figure 3 The comparison chart of the time-series corrected weight distribution and attention distribution shows that the attention of the second identification model can accurately fall on the processing moment that leads to the formation of defects.

[0067] During any PDC drill bit cladding process, the real-time sequence of process parameters during the cladding process is collected, normalized, and then input into the second identification model to obtain the defect identification results of this PDC drill bit cladding process.

[0068] Thus, by introducing consistency loss for the second stage of training, the final second identification model can accurately learn the real characteristics of defect formation, thereby improving the accuracy of defect identification and enhancing the robustness and accuracy of PDC drill bit cladding defect identification.

[0069] In other optional embodiments, in response to the defect identification result being defective, the attention of the second identification model to each moment in the real-time sequence of process parameters can be further obtained, and the moment with the attention greater than the average attention of each moment is taken as the target moment. The processing position corresponding to the target moment is the defect position in this PDC drill bit cladding process, thereby realizing defect localization.

[0070] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A deep learning-based defect identification method for the cladding process of PDC drill bits, characterized in that, The identification method includes: The process parameter sequence of the cladding process and the surface image after the cladding are collected. The classification labels generated from the surface images are used to train the defect recognition model in the first stage to obtain the first recognition model. Based on the temporal mapping relationship between the defect location in the surface image and the process parameter sequence, the defect time corresponding to the defect location is located in the process parameter sequence, and the significance index of the defect location relative to the preset standard template is calculated. A time-series correction weight is constructed based on the defect time and significance index. The time-series correction weight is used to characterize the contribution of each processing time in the process parameter sequence to the formation of defects. The first loss of the first stage training is weighted by time-series correction weights to obtain the second loss. The first identification model is then trained in the second stage based on the second loss to obtain the second identification model. The second identification model is used to map the process parameter sequence of the cladding process to the defect identification result.

2. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, The first stage of training the defect identification model using classification labels generated from surface images includes: Threshold segmentation and morphological filtering are performed on the frame difference map of the surface image and the standard template to extract the connected components of the suspected defect region. If the area of ​​the connected component exceeds the preset threshold, the process parameter sequence is marked as defective; otherwise, it is marked as defect-free, and a classification label is obtained. The process parameter sequence is used as input, and the classification label is used as supervision information. The defect recognition model is updated using the binary cross-entropy loss function.

3. The deep learning defect identification method for the PDC drill bit cladding process according to claim 2, characterized in that, The obtained first recognition model includes: Real-time monitoring of the rate of change of loss values ​​of the defect identification model on the test set; calculation of the reduction in loss values ​​within a preset monitoring window; If the reduction in the loss value is less than the preset reduction amount for several consecutive times, or if the training rounds reach the first maximum round, the first stage of training is stopped, and the first recognition model is obtained.

4. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, The step of locating the defect time corresponding to the defect position in the process parameter sequence includes: recording the movement trajectory of the cladding head during the cladding process, the movement trajectory including the processing position at any processing time, querying the movement trajectory based on the defect position, and obtaining the processing time corresponding to the defect position as the defect time.

5. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, The significance index of the calculated defect location relative to the preset standard template includes: Establish a mapping relationship between the surface image coordinate system and the workpiece coordinate system, and align the surface image with the standard template; determine the region of interest (ROI) at the defect location, and calculate the absolute value of the grayscale difference between each pixel in the ROI between the surface image and the standard template; normalize the average value of the absolute values ​​to obtain a saliency index characterizing the degree of defect significance, wherein the ROI is a rectangular area of ​​a preset size centered on the defect location.

6. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, The time-series correction weights constructed based on defect timing and significance indicators include: A Gaussian distribution function is constructed with any defect time as the center, and the Gaussian weights of each processing time in the neighborhood time corresponding to the defect time are obtained. The Gaussian weights are multiplied by the significance index of the defect position corresponding to the defect time to obtain the corrected weights of each processing time in the neighborhood time corresponding to the defect time. For each processing time t, calculate the sum of the correction weights for all defect times corresponding to that time, and normalize the summation result to obtain the time-series correction weights for each processing time.

7. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, Before performing a second-stage training of the first identification model based on the second loss, the identification method further includes: inputting a sequence of process parameters labeled as defective into the first identification model, and calculating the attention level at each processing moment using a gradient class activation mapping algorithm; in response to the correlation between the attention level and the time-series correction weight being less than a correlation threshold, initiating the second-stage training; otherwise, using the first identification model as the second identification model.

8. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 7, characterized in that, The second loss is obtained by: acquiring the attention of the first identification model to each processing moment in the process parameter sequence; calculating the mean square error between the attention of each processing moment and the time-series correction weight as the consistency loss; and taking the weighted sum of the first loss and the consistency loss as the second loss.

9. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, The defect identification model is a recurrent neural network.

10. The deep learning defect identification method for the cladding process of PDC drill bits according to claim 1, characterized in that, The process parameters include at least one of laser power, scanning speed, molten pool temperature, and protective gas flow rate.

Images