A defect detection method and system for numerical control blades

Abstract

The present invention provides a defect detection method and system for numerically controlled blades, which are applied to the field of data processing technology. The present application processes the use information of a target blade to generate a blade detection influencing factor; processes the blade image information of a target blade within a preset time period to generate target blade sequence information; processes the target blade sequence information based on the detection requirement information of a target user to generate target blade parameter information; processes the production information of a target blade to generate a blade prediction detection feature vector; processes a preset blade defect warning model based on a training sample set to generate a target blade defect warning model; processes the blade detection influencing factor, target blade parameter information and blade prediction detection feature vector based on the target blade defect warning model to generate blade defect warning information.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a defect detection method and system for a numerical control blade. Background Art

[0002] In the machining industry, CNC blades are key cutting tools, and their quality is directly related to the machining accuracy, efficiency and quality of the final product. In past production practices, traditional blade defect detection methods mainly focused on manual visual inspection and simple measuring tool measurement. During manual inspection, inspectors rely on their own experience and naked eyes to observe whether there are obvious scratches, cracks, wear and other defects on the surface of the blade, and use calipers, micrometers and other tools to measure the dimensional accuracy of the blade. However, this method has many disadvantages. On the one hand, manual inspection is highly subjective, and different inspectors may have different judgment criteria for the same defect, making it difficult to ensure the consistency of the inspection results. On the other hand, its efficiency is extremely low. In the scenario of large-scale production, manual inspection requires a lot of time and manpower costs, which seriously affects the production progress. In addition, due to the limitations and fatigue of the human eye, it is easy to miss inspections, and some minor defects may not be discovered in time, thus posing a potential risk to the processing quality.

[0003] With the rapid development of modern manufacturing, the precision of products is constantly improving, which puts more stringent requirements on the performance and reliability of CNC blades. For example, in high-end fields such as aerospace and automobile manufacturing, the processing accuracy often needs to reach the micron or even nanometer level, which requires the blade to have extremely high stability and precision retention. Traditional detection methods are obviously unable to meet the needs of such high-precision and large-scale production.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] The purpose of this application is to provide a defect detection method and system for CNC blades, which at least overcomes the problems existing in the prior art to a certain extent, by collecting information such as images of target blades in a preset period of time, usage purposes, user detection requirements, production information, early warning models and training sample sets. For usage purposes, the blade classification factors and weight values ​​are processed, and the detection deviation values ​​are calculated to determine the detection influencing factors; the image information is first normalized and filled to form a target specification image, and then the features are extracted and integrated into a sequence image. The scene information is generated according to user needs to process the sequence image to obtain the target parameters. The production information is also processed, normalized and screened to obtain a prediction vector. In the training sample set link, the features are obtained to determine the sampling ratio, and the sampling features are processed and combined into a data group to train the preset model, which is verified with the verification set, and the one that meets the conditions becomes the target model. Finally, the relevant factors, parameters and vectors are processed by the target model to generate correlation values ​​and abnormal vectors, and finally the early warning information that can accurately characterize the type and severity of blade defects is obtained, which effectively guarantees the quality detection of blades.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present invention.

[0007] According to one aspect of the present application, a defect detection method for a numerically controlled blade is provided, comprising: obtaining blade image information of a target blade within a preset time period, usage information of the target blade, detection requirement information of a target user, production information information of the target blade, a preset blade defect warning model and a training sample set; processing the usage information of the target blade to generate a blade detection influencing factor; processing the blade image information of the target blade within the preset time period to generate target blade sequence information; processing the target blade sequence information based on the detection requirement information of the target user to generate target blade parameter information; processing the production information information of the target blade to generate a blade prediction detection feature vector; processing the preset blade defect warning model based on the training sample set to generate a target blade defect warning model; processing the blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector based on the target blade defect warning model to generate blade defect warning information.

[0008] In another aspect of the present application, a defect detection device for a numerically controlled blade is characterized in that it includes: an acquisition module, which is used to acquire blade image information of a target blade within a preset time period, usage information of the target blade, detection requirement information of a target user, production information information of the target blade, a preset blade defect warning model and a training sample set; a processing module, which is used to process the usage information of the target blade to generate a blade detection influencing factor; process the blade image information of the target blade within the preset time period to generate target blade sequence information; process the target blade sequence information based on the detection requirement information of the target user to generate target blade parameter information; process the production information information of the target blade to generate a blade prediction detection feature vector; process the preset blade defect warning model based on the training sample set to generate a target blade defect warning model; process the blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector based on the target blade defect warning model to generate blade defect warning information.

[0009] According to another aspect of the present application, an electronic device is provided, characterized in that it comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-mentioned CNC blade defect detection method by executing the executable instructions.

[0010] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the above-mentioned defect detection method for a CNC blade is implemented.

[0011] The present application provides a method and system for defect detection of CNC blades. The server obtains image information of the target blade under different viewing angles and light intensities within a preset time period, its usage information (such as processing parameters, etc.), the target user's detection demand information on blade accuracy, life, etc., production information information (raw materials, processes, etc.), preset blade defect warning model and training sample set. For the usage information, the blade classification influencing factor and weight value are first processed and generated, and the detection deviation value is calculated in combination with the detection threshold of the same type of blades, and then the blade detection influencing factor is obtained. The blade image information is normalized and filled into a target scale specification image, and then the features under different viewing angles and lighting are extracted and sorted and integrated into sequence image information. Based on the user's detection demand information, the processing and application scenario information is generated, and the sequence image information is processed accordingly to obtain matching features, and the target blade parameter information is obtained through image segmentation positioning and further processing. The production information is processed to generate a feature set, and the predicted detection feature vector is obtained after normalization and screening.

[0012] For the training sample set, first obtain the data features and determine the sampling ratio according to the quantity, process the sampling features, combine and generate the data set to train the preset model, and verify it with the verification sample set. The trained model that meets the conditions becomes the target model. Finally, use the target model to process the blade detection influencing factors, parameter information and feature vectors, and generate the blade feature impact correlation value and abnormal feature vector in turn, and finally obtain the warning information that characterizes the specific type and severity of the blade defect quantitative indicators, providing an effective means for CNC blade quality inspection.

[0013] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flow chart showing a defect detection method for a CNC blade provided by an embodiment of the present application is shown;

[0015] Figure 2 A schematic structural diagram of a defect detection device for a CNC blade provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0016] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0017] Combine the following Figure 1 To describe the defect detection method of a CNC blade according to an exemplary embodiment of the present application. It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principle of the present application, and the embodiments of the present application are not limited in this regard. On the contrary, the embodiments of the present application are applied to any applicable scenario.

[0018] In one implementation, the present application also proposes a defect detection method and system for a CNC blade. Figure 1 The following schematically shows a flow chart of a defect detection method for a CNC blade according to an embodiment of the present application. Figure 1 As shown, the method is applied to a server, comprising:

[0019] S101, obtaining blade image information of a target blade within a preset time period, usage information of the target blade, detection requirement information of a target user, production information information of the target blade, a preset blade defect warning model and a training sample set.

[0020] In one embodiment, a batch of CNC blades is inspected in a machining workshop. Within a preset time period (e.g., the past week), the target blade is photographed from multiple fixed positions and angles (e.g., vertically above, 45 degrees to the side, etc.) by a high-resolution camera installed on the production line. A large number of blade images are obtained under different light intensities, including normal workshop lighting (about 500-800 lux), strong light simulating high temperature and high light environment during processing (1000-1200 lux), and weak light simulating night or low light environment (200-300 lux). These images clearly show the appearance of the blade under different conditions, such as the sharpness of the cutting edge, the flatness of the surface, and whether there are defects such as scratches and cracks. This batch of blades is mainly used for CNC milling of aluminum alloy parts. During this processing process, the blade needs to withstand certain cutting forces, rotation speeds, and feed speeds. For example, the cutting speed is usually maintained at 200-300 m / min, the feed speed is 0.1-0.2 mm / tooth, and the cutting depth is between 2-5 mm. These usage parameters and processing material information constitute the usage information of the target blade, which has important reference value for the subsequent analysis of blade wear and possible defect types.

[0021] As the target user, the quality control department of the machining workshop mainly focuses on ensuring that the accuracy and life of the blade can meet the production requirements. Specifically, they hope to accurately detect whether the blade edge wear is within the allowable tolerance range (for example, the wear of the blade radius must not exceed 0.05 mm), and whether the surface roughness of the blade meets the standard for processing aluminum alloy parts (the surface roughness must be less than 0.8 microns). At the same time, users also require timely detection of potential defects that may cause blade breakage, such as internal cracks, to avoid scrap and equipment damage during processing. These blades are produced by a well-known tool manufacturer. The raw materials are sourced from a specific supplier and use high-quality cemented carbide materials, of which the content of tungsten carbide is 85%, the content of cobalt is 10%, and other trace elements account for 5%. In terms of production process parameters, the forging temperature is controlled at 1400-1500 degrees Celsius, the sintering temperature is 1600-1700 degrees Celsius, and advanced precision grinding equipment is used in the processing process to ensure the dimensional accuracy and surface quality of the blade. The production batch information shows that this is the manufacturer’s 10th batch of products. Through the quality analysis of previous batches of products, it can provide certain reference and comparison for this inspection.

[0022] The preset blade defect warning model adopts a hybrid model architecture combining convolutional neural network (CNN) and random forest (RandomForest). The CNN layer is used to automatically extract complex features from blade image information, while the random forest layer conducts comprehensive analysis and decision-making on the extracted image features and other non-image data (such as usage information, production information, etc.). Among them, the convolutional neural network layer contains multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer uses convolution kernels of different sizes (such as 3x3, 5x5, etc.) to extract local features of the image, the pooling layer is used for downsampling to reduce the amount of calculation and maintain the main information of the features, and the fully connected layer integrates and maps the extracted features. For example, the first convolutional layer has 32 3x3 convolution kernels, a step size of 1, and a ReLU activation function; the pooling layer uses maximum pooling and a window size of 2x2. The random forest layer consists of multiple decision trees, and the number of decision trees can be set to 100. Each decision tree splits and makes decisions based on the comprehensive features of the input, and finally determines the defect type and severity of the blade through a voting mechanism. During the training process, the model performance is optimized by adjusting the depth of the decision tree, the selection criteria for split nodes (such as the Gini coefficient or information gain), and other parameters. The initial learning rate is set to 0.001. During the training process, the learning rate decay strategy can be adopted according to the learning curve and verification results. For example, the learning rate decays to 0.9 times the original after every 10 epochs. The regularization parameter uses L2 regularization, and the weight decay coefficient is set to 0.0001 to prevent the model from overfitting. The number of training rounds is set to 100 rounds, and the loss and accuracy on the verification set are observed to determine whether to stop training early.

[0023] Collect historical blade data from multiple machining companies, covering blade information of different brands, models, production batches and usage scenarios. At the same time, it also includes data obtained by testing blades under laboratory simulation of different working conditions. Each sample contains the input data required by the above model and the corresponding output data (known blade defect types and severity). For example, the blade image information in a sample is an image of a certain model of blade taken at a specific viewing angle and lighting, and its corresponding usage information is the parameters of the blade in a certain processing technology. The production information is a detailed record of its production process, and the output is the defect type determined by the blade after professional testing (such as the edge wear level is 3). The training sample set contains at least 5,000 samples to ensure that the model has enough data for learning and generalization. And it is divided into a training set and a validation set according to a certain ratio (such as 8:2), where the training set is used for model training, and the validation set is used to adjust model parameters and evaluate model performance to prevent overfitting.

[0024] The training sample set contains thousands of blade sample data from different production batches and different usage scenarios. These sample data cover various types of defects (such as wear, cracks, chipping, etc.) and normal blade information. Each sample contains the corresponding blade image features, usage parameters, production information data, and annotation information on whether there are defects. By learning and training these samples, the model can identify the potential relationship between different factors and blade defects, thereby providing accurate early warnings in the subsequent detection of target blades.

[0025] S102: Process the usage information of the target blade to generate a blade detection influencing factor.

[0026] In one embodiment, the use information of the target blade is processed to generate a blade classification influencing factor and a blade classification weight value. Assume that there is a batch of CNC blades, and the target blades are mainly used in the processing process of high-speed milling of stainless steel materials. When processing the use information of the target blade, first analyze its processing technology characteristics. Since it is high-speed milling of stainless steel, this processing process has high requirements on the hardness, wear resistance and heat resistance of the blade. According to these characteristics, the blade classification influencing factor and the blade classification weight value are generated. For example, the influence factor of hardness is set to 0.4, because hardness is crucial to resisting wear during stainless steel cutting; the influence factor of wear resistance is set to 0.35, and the influence factor of heat resistance is set to 0.25. The corresponding weight values ​​are allocated according to their relative importance in the overall performance. Assuming that in this processing scenario, after a large number of experiments and data analysis, it is determined that the weight value of hardness is 0.45, the weight value of wear resistance is 0.3, and the weight value of heat resistance is 0.25.

[0027] The detection threshold of the blade of the same type as the target blade is obtained, and the blade classification influencing factor and the blade classification weight value are processed based on the detection threshold of the blade of the same type as the target blade to generate the detection deviation value of the target blade. Next, the detection threshold of the blade of the same type as the target blade (the same type of blade also used for high-speed milling of stainless steel) is obtained. These thresholds are obtained by statistical analysis of a large number of detection data of this type of blades in normal working conditions and different wear stages. For example, for hardness, the hardness detection threshold range of the blade in normal working conditions is 850-950HV (Vickers hardness); for wear resistance, through a specific wear test method, its threshold is set to a wear amount of no more than 0.05 mm under a certain cutting length; for heat resistance, after long-term high-temperature work, the degree of change of the metallographic structure of the blade is within a certain standard range as the threshold. Then, based on these detection thresholds, the generated blade classification influencing factor and blade classification weight value are processed. Assume that the target blade is tested and its hardness is 830HV, which is deviated from the lower threshold of 850HV. According to the formula (actual value - lower limit of threshold) × influence factor × weight value, the deviation contribution value of hardness is calculated as (830-850) × 0.4 × 0.45 = -3.6. Similarly, similar calculations are performed for wear resistance and heat resistance, assuming that the deviation contribution value of wear resistance is 2.5 and the deviation contribution value of heat resistance is -1.8. Adding these deviation contribution values, the detection deviation value of the target blade is -3.6 + 2.5 - 1.8 = -2.9.

[0028] The detection deviation value of the target blade is processed to generate a blade detection influencing factor. The detection deviation value of the target blade, -2.9, is processed. The degree of its influence on blade detection can be determined by setting different interval ranges. For example, when the detection deviation value is between -5 and 5, it is considered to have a certain impact on the blade performance, but it is still within an acceptable range; if it exceeds this range, it may indicate that the blade has a greater defect risk. In this case, -2.9 is within the acceptable range, but close to the lower limit, indicating that the blade may have shown some performance degradation trends during use and requires further attention. After such processing, the detection deviation value is converted into a blade detection influencing factor. This influencing factor will be used as an important reference indicator in the subsequent blade defect warning model, and will be used together with other information to comprehensively judge the defect situation of the blade.

[0029] S103, processing the blade image information of the target blade within a preset time period to generate target blade sequence information.

[0030] In one embodiment, the blade image information of the target blade within a preset time period is normalized and filled to generate blade image information of the target scale specification. Assume that in a tool production and inspection workshop, a batch of CNC blades are quality monitored. The preset time period is set to the past week. During this week, the target blade is photographed from different angles (such as directly above, 30 degrees to the side, 60 degrees to the side, etc.) and under different light intensities (such as 500 lux normal lighting, 800 lux strong light to simulate the strong light environment during processing, and 200 lux weak light to simulate special working conditions) by multiple fixed-position high-resolution cameras, and a large amount of blade image information is obtained.

[0031] First, these images are normalized and filled. The normalization operation maps the pixel values ​​of the image to a specific range, such as normalizing the pixel values ​​of all images to between 0 and 1. This is achieved by calculating the difference ratio between the original value of each pixel and the maximum and minimum pixel values ​​in the image. For the filling process, if the image has incomplete or missing edges, an edge extension algorithm based on the image content is used for filling. The algorithm analyzes the pixel gradient information of the edge of the image to determine the most likely edge extension direction and pixel value. For example, for areas with blurred edges, the missing parts are filled by performing a weighted average calculation on the pixels of the adjacent clear areas to ensure that the image has a complete shape and size, thereby generating a blade image information of 512×512 pixels with the target scale specification.

[0032] The image information of the blade of the target scale is processed to generate the image feature information under different viewing angles and the image feature information under different light intensities. For images of different viewing angles, a feature extractor based on a convolutional neural network (CNN) is used for feature extraction. In the CNN model, multiple convolution layers and pooling layers are set. For example, the first convolution layer uses a 3×3 convolution kernel with a step size of 1 and 32 channels. The local features of the image are extracted through the convolution operation, and then downsampled through a 2×2 maximum pooling layer to reduce the amount of data while retaining the key features. From the image from the top view, after such processing, the overall contour features of the blade, the position of the central symmetry axis and other information may be extracted; from the image from the side view, the focus is on extracting the features such as the inclination angle of the blade and the thickness change. These features can be obtained by further analyzing and calculating the feature map output by the convolution layer. For images under different light intensities, in strong light images, since strong light will make tiny scratches or flaws on the blade surface more obvious, the grayscale co-occurrence matrix algorithm is used to extract the texture features and grayscale change features of these areas, such as calculating texture contrast, correlation and other parameters; in weak light images, more attention is paid to the clarity of the reflective area and overall shape of the blade, and the edge detection-based algorithm (such as Canny edge detection) is used to extract the corresponding features to determine the approximate outline of the blade and the boundaries of key parts.

[0033] The image feature information at different viewing angles and the image feature information at different light intensities are sorted and integrated to generate the target blade sequence information. The images are sorted in time order, viewing angle order and light intensity order. In terms of time order, the image shooting timestamp is used for sorting to ensure that the images of different viewing angles and light intensities taken at the same time are adjacent. Then the images are arranged in order from directly above to 30 degrees to the side and 60 degrees to the side. For each viewing angle, the images are integrated in the order of light intensity from weak to strong. In this way, the target blade sequence information is generated, which completely records the changes in the image features of the target blade at different times, different viewing angles and different light intensities within a week, providing a rich data basis for subsequent blade quality analysis and defect detection. For example, in the subsequent analysis, the development process of blade wear or damage can be observed by comparing the feature information at the same viewing angle and light intensity at different time points, or the performance status of the blade can be more comprehensively evaluated by analyzing the differences in features at different viewing angles and light intensities.

[0034] S104: Process the target blade sequence information based on the detection requirement information of the target user to generate target blade parameter information.

[0035] In one embodiment, the detection demand information of the target user is processed to generate the processing scenario information of the target blade and the application scenario information of the target blade. In an aviation parts manufacturing enterprise, the detection demand information of the target user indicates that the CNC blades used to process aircraft engine blades need to be detected. These aircraft engine blades are made of specific high-temperature alloys, and the processing technology is complex and the precision requirements are extremely high. High-speed milling (stable speed at 12,000 rpm), electrospark machining and other composite processes are usually used, and specially formulated cutting fluids are required during the processing. They have strict requirements on cutting temperature, lubricity, etc. to ensure the processing quality and tool life. Based on this, the processing scenario information of the target blade is generated as a high-precision, high-speed (12,000 rpm), complex curved surface and multi-axis linkage processing environment. The surface roughness needs to be controlled within Ra0.2-Ra0.4 microns, and the dimensional accuracy tolerance is within ±0.005 mm. The application scenario information is the aerospace field. In this field, the quality reliability of parts is related to flight safety. The blades must ensure the processing quality while running stably for a long time, and must pass the strict certification of aviation safety standards.

[0036] The target blade sequence information is processed based on the target blade processing scenario information and the target blade application scenario information to generate the target blade feature information, wherein the target blade feature information is the feature that matches the target blade processing scenario information and the target blade application scenario information. When analyzing the image, the high-resolution microscope image is used to focus on observing the microscopic wear characteristics of the blade edge when simulating high-speed cutting at 12,000 rpm, such as microscopic cracks, chipping, and uneven wear of the blade edge; for blades under complex surface processing paths (such as processing blade edges with complex curvatures), the three-dimensional reconstruction technology and stress analysis algorithm are used to study the stress distribution area characteristics of the blade surface, including the location, size, and direction of the stress concentration point, which are the target blade feature information that matches the processing and application scenarios.

[0037] Based on the target blade feature information, the target blade sequence information is segmented and positioned to generate the target blade image slice information, and the target blade image slice information is processed to generate the target blade parameter information. The improved U-Net network based on deep learning (with an attention mechanism added to better focus on key areas) is used to accurately segment the blade edge wear area, surface stress concentration area and other key areas from the entire image according to the unique texture and grayscale change pattern of the edge wear area and the characteristic pattern of the stress concentration area, and determine its coordinate position in the image, thereby generating the target blade image slice information. By performing pixel analysis and edge detection algorithms on the image of the edge wear area, combined with the geometric shape model of the wear area, the wear degree of the edge can be accurately quantified, such as the specific distribution of the wear width between 0.02-0.04 mm and the average wear depth of 0.008 mm. For the surface stress concentration area, the texture analysis method based on grayscale co-occurrence matrix and Fourier transform is used to analyze its texture changes and grayscale value distribution, and calculate the size and distribution of surface stress and other parameters, such as the maximum stress value of the surface stress concentration area is 850MPa, and the stress gradient is 50MPa / mm in a specific area. These quantified parameters, namely the target blade parameter information, can intuitively reflect the performance status of the blade in this specific processing and application scenario, and provide key data support for subsequent blade quality evaluation and defect judgment, so as to replace or maintain the blade in time.

[0038] S105, processing the production information of the target blade to generate a blade prediction detection feature vector.

[0039] In one embodiment, the production information of the target blade is processed to generate blade production feature set information, wherein the blade production feature set information is used to characterize the raw material source information of the blade, the composition information of the blade, the production process parameters of the blade, the precision and stability information of the production equipment of the blade, and the production batch information of the blade. Assume that there is a batch of CNC blades, and its production information is as follows: the raw materials of the blades come from a supplier A that is well-known for its high quality in the industry. The supplier strictly controls the purity and impurity content of the raw materials. The cemented carbide powder it provides has a high reputation in the global market and has a stable supply channel and quality assurance system. After testing and analysis, the main components of the blade are tungsten carbide (WC) with a content of about 88%, cobalt (Co) with a content of about 10%, and the rest are trace elements such as a small amount of titanium (Ti) and tantalum (Ta). The proportions of these components are carefully adjusted to meet the performance requirements of the blade in terms of hardness, toughness and wear resistance. During the production process, the forging temperature is strictly controlled between 1450-1550 degrees Celsius. This temperature range can ensure that the raw materials are fully integrated and the internal structure is uniform. The sintering temperature is maintained at 1650-1750 degrees Celsius, which enables the blade to obtain good density and mechanical properties. At the same time, in the grinding process, advanced CNC grinding equipment is used, and its processing accuracy can reach ±0.002 mm, ensuring the dimensional accuracy and surface finish of the blade.

[0040] Key equipment on the production line, such as the powder press, has a pressure control accuracy of ±0.5 tons, which can ensure the density consistency of the blade during the molding process. The temperature uniformity of the sintering furnace is within ±5 degrees Celsius, which effectively avoids blade performance fluctuations caused by temperature differences. In addition, these devices are regularly maintained and calibrated, have high stability, and the equipment failure rate is less than 1%. This batch of blades belongs to the manufacturer's 20th batch of products. Through quality tracing and statistical analysis of previous batches of products, it was found that the overall quality of the manufacturer's products in batches 15-19 showed a steady upward trend, and the product qualification rate increased from 90% to 95%, but in batch 18, there was a small amount of surface roughness exceeding the standard due to grinding process problems. The blade production feature set information is generated based on the above information. This feature set comprehensively covers the key attributes of the blade from raw materials to the production process, providing basic data for subsequent analysis.

[0041] The blade production feature set information is normalized to generate blade production features in the target format. For raw material source information, different suppliers can be quantitatively scored according to their industry rankings and quality reputations, and normalized to between 0 and 1; for composition information, the content of each element is normalized by dividing it by its maximum content under ideal conditions; production process parameters are linearly normalized according to their allowable range of values, such as forging temperature can be converted into the form of (actual temperature-minimum temperature) / (maximum temperature-minimum temperature); production equipment accuracy and stability information can be normalized by calculating the ratio of the actual accuracy of the equipment to the industry standard accuracy; production batch information can be encoded and normalized according to the batch sequence and the quality of previous batches. Through these normalization operations, the target format of blade production features is generated, making it uniform in dimension and comparable.

[0042] The blade production features in the target format are processed by feature screening to generate a blade prediction detection feature vector. Through methods such as principal component analysis (PCA), the features that have a greater impact on blade quality and defects are screened out. For example, it is found that features such as tungsten carbide content, forging temperature, and grinding equipment accuracy have a high contribution rate in explaining blade performance differences. These key features are combined into a blade prediction detection feature vector. This feature vector can highlight the core factors in the blade production process, provide effective input data for the subsequent defect prediction model based on machine learning, and help accurately judge the potential defect risks of the blade.

[0043] S106: Process the preset blade defect warning model based on the training sample set to generate a target blade defect warning model.

[0044] In one embodiment, any number of data features in the training sample set is obtained, and a sampling ratio is generated based on the number of each data feature in the training sample set. First, 5,000 data features are randomly selected from the 10,000 samples. When selecting image features, advanced image analysis algorithms are used to extract features such as curvature change features of the edge profile, roughness and directional features of the surface texture, etc. For the use characteristics, the mechanical performance parameters such as hardness and toughness of the processed material, as well as the processing parameters such as cutting speed, feed rate, and cutting depth, and even the type and performance parameters of the processing equipment are recorded in detail. In terms of production information features, in addition to the raw material supplier number, production process temperature parameters, and equipment accuracy level, it also covers the environmental parameters of the production workshop (such as temperature, humidity, cleanliness, etc.) and the skill level information of the operator. For image features, 2,000 different description dimensions are determined through dimensional analysis and clustering algorithms of feature vectors, and the correlation and importance distribution between these dimensions are further analyzed. For use characteristics and production information characteristics, classification statistics and data mining techniques are used to sort out 1,500 different categories and values. According to these quantity distributions, the sampling ratio is generated according to certain rules. For image features, the sampling ratio is set to 0.4 because of their large number and the fact that they are crucial to defect detection in previous research and practice; for usage features and production information features, the sampling ratios are set to 0.3 respectively, considering their comprehensive impact on blade performance.

[0045] The training sample set is processed based on the sampling ratio to generate a preset number of sampling features. 800 features are selected from the image features at a sampling ratio of 0.4. During the selection process, a combination of stratified sampling and random sampling is used to ensure that the selected features can cover different types of image features and are random. 450 features are selected from the usage features, and weighted sampling is performed according to the frequency of occurrence of different processing materials and process parameters, so that common and key usage features have a higher probability of being selected. 450 features are selected from the production information features, and sampling is performed based on the importance of production batches and equipment, giving priority to those production information features from key batches or equipment that have a greater impact on blade quality. In this way, a preset number of sampling features (such as 1,700) are generated.

[0046] Based on any data feature and each sampling feature, multiple data groups are generated, wherein each data group contains a preset number of data samples, and at least one data sample includes identification information. A specific blade profile shape feature is combined with the other 1699 sampling features. In the combination process, the permutation combination algorithm and data fusion technology are used to ensure that each combination can fully reflect the interaction between different features. After forming a set of data groups, this process is repeated until all data features are involved in the combination, and finally multiple data groups are generated. Each data group contains a preset number (such as 50) of data samples, and among these data samples, at least one data sample includes identification information, which indicates whether the blade corresponding to the sample has defects and the type of defects (such as wear, cracks, chipping, etc.). When generating data groups, the data samples are also randomly shuffled and grouped and balanced to avoid the order and distribution of the data from having adverse effects on model training.

[0047] The preset blade defect warning model is trained based on data samples in multiple data sets to generate a trained blade defect warning model. During the training process, the optimized training algorithm and parallel computing technology are used to continuously adjust the parameters of the model (such as kernel function type, penalty coefficient, etc.). For example, different kernel functions (linear kernel, polynomial kernel, Gaussian kernel, etc.) are tried, and the best kernel function and parameter combination are determined through cross-validation and performance evaluation indicators (such as accuracy, recall rate, F1 value, etc.). After multiple iterations of training, the performance of the model on the training set and validation set is evaluated and analyzed in each iteration, and the training strategy is adjusted according to the evaluation results, and finally the trained blade defect warning model is generated.

[0048] The trained blade defect warning model is processed based on the validation sample set to generate a validation result. If the data sample containing the identification information in the validation result is a risk factor that characterizes the quality of the finished blade, the trained blade defect warning model is used as the target blade defect warning model. The trained blade defect warning model is processed using an independent validation sample set (containing 2,000 samples). The validation sample set is input into the model to obtain the model's prediction results, which are compared with the actual identification information in the sample set. For example, it was found that the model's prediction results for 1,800 samples were consistent with the actual situation, with an accuracy rate of 90%. For the 200 samples with incorrect predictions, a detailed error analysis was performed. Through visualization technology and feature importance analysis, check whether there is a certain pattern or feature that is ignored by the model. For example, it was found that for blades of certain special materials, the model was insufficient in considering the synergy between processing technology and production information, resulting in misjudgment.

[0049] Finally, check the data samples containing identification information in the verification results. If these data samples are misjudged by the model and actually represent risk factors that affect the quality of the finished blade (such as predicting a cracked blade as defect-free), further analysis of the reasons may be that the weights of some features are set unreasonably or the complexity of the model is not enough. A series of optimization measures have been taken to address these problems. For example, readjust the feature engineering, re-extract and weight the image features and production process features related to cracks; increase the training data, collect more blade samples with crack defects, and annotate and analyze their features more carefully; improve the model structure, increase the number of layers or nodes of the model, and improve the model's expression and generalization capabilities. Through continuous adjustment and optimization, until the samples representing risk factors in the verification results can be accurately identified by the model. At this time, the trained blade defect warning model is used as the target blade defect warning model for subsequent defect detection and warning of the target blade.

[0050] S107: Process the blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector based on the target blade defect warning model to generate blade defect warning information.

[0051] In one embodiment, the blade detection influencing factor and the target blade parameter information are processed based on the target blade defect warning model to generate a blade feature impact association value. This batch of blades is mainly used for high-speed milling of aluminum alloy cylinder materials, and its designed standard cutting speed is 300 meters per minute and the feed rate is 0.15 mm per tooth. However, actual detection found that due to some performance fluctuations of the processing equipment, the current cutting speed is an average of 320 meters per minute and the feed rate is 0.17 mm per tooth. After calculation and processing, the generated blade detection influence factor is 0.7 (within the range of 0-1, the larger the value, the greater the deviation from the standard usage). This shows that the deviation in usage conditions may have a greater impact on the performance of the blade. Through image analysis and measurement, the average value of the edge wear width at different positions is 0.04 mm (the upper limit of the normal wear width of this type of blade is 0.03 mm), and the surface roughness is detected to be Ra0.5 microns (while this processing scenario requires the surface roughness to be less than Ra0.4 microns). In addition, a slight chipping phenomenon was observed on the edge in the high-resolution microscope image. Although the area is small, this is also a potential important defect feature.

[0052] First, the blade detection influencing factors and target blade parameter information are input into the target blade defect warning model. The model processes according to the pre-set algorithm and weight. For example, the blade detection influencing factors The value is 0.7, which is the value obtained after processing the target blade usage information. It reflects the degree of influence of the blade usage conditions deviating from the standard conditions on its performance. The value range is between 0 and 1. The larger the value, the greater the deviation from the standard usage.

[0053] include and , specifically, is the cutting speed deviation weight, whose value is 0.3, and is used to characterize the importance of the cutting speed deviation on the final correlation value. is the feed deviation weight, whose value is 0.2, and is used to characterize the importance of the feed deviation on the final correlation value.

[0054] include and , specifically, is the actual cutting speed, and its value is 320 (unit: m / min), which is the cutting speed measurement value obtained in the actual test. is the actual feed rate, the value is 0.17 (unit: mm / tooth), which is the feed rate measurement value obtained in the actual detection.

[0055] include and , specifically, It is the standard cutting speed, and its value is 300 (unit: m / min), which is the cutting speed value of the blade under the design or ideal state. The standard feed rate is 0.15 (unit: mm / tooth), which is the feed value of the blade under the design or ideal state.

[0056] include and , specifically, is the weight of the cutting edge wear width exceeding the standard value, and its value is 0.4. It is used to characterize the importance of the influence of the cutting edge wear width exceeding the standard value on the final correlation value. The weight of surface roughness exceeding the standard value is 0.1, which is used to characterize the importance of the influence of surface roughness exceeding the standard value on the final correlation value.

[0057] include and , specifically, is the actual edge wear width, and its value is 0.04 (unit: mm), which is the measurement result of the edge wear width in actual detection. It is the actual surface roughness, and its value is 0.5 (unit: micron), which is the measurement result of the surface roughness in actual detection.

[0058] include and , specifically, It is the upper limit of the standard cutting edge wear width, and its value is 0.03 (unit: mm), which is the standard upper limit of the cutting edge wear width. It is the standard surface roughness, and its value is 0.4 (unit: micron), which is the standard value of surface roughness.

[0059] The blade feature impact correlation value is calculated by the following formula, ;in, Represents the blade characteristics affecting the associated value; represents the blade detection influencing factor; represents the i-th parameter deviation weight, which is used to characterize the importance of the deviation of different parameters (such as cutting speed, feed rate, etc.) on the final correlation value; represents the actual value of the i-th parameter, which is the measured value of the blade-related parameter obtained in the actual detection; represents the standard value of the i-th parameter, which is used to characterize the value that the parameter should reach under the design or ideal state of the blade; Represents the weight of the i-th parameter that exceeds the standard value, which is used to characterize the importance of the influence of different parameters that exceed the standard value on the final correlation value. Different parameters that exceed the standard value correspond to different value; represents the i-th parameter that actually exceeds the standard value, and is used to characterize the measurement result of the i-th parameter that exceeds the standard value in the actual detection; It represents the standard value of the i-th corresponding parameter exceeding the standard value, that is, the standard value that the i-th parameter exceeding the standard value should reach under the design or ideal state, and n represents the number of parameter deviations involved in the calculation.

[0060] That is, the blade feature influences the correlation value to be 0.58.

[0061] The blade prediction detection feature vector is processed based on the blade feature impact association value to generate a blade abnormality feature vector. The blade prediction detection feature vector contains information such as the source of the blade's raw material (from supplier B, whose material has shown a certain risk of fluctuation in high-temperature performance in past data analysis), production process parameters (forging temperature is 1480 degrees Celsius, close to the upper limit of the standard forging temperature range of the material), and production equipment accuracy (the pressure control accuracy of the powder press is ±0.6 tons, slightly higher than the industry standard of ±0.5 tons). The model adjusts and analyzes these feature vector elements based on the association value. For example, if the association value is high and the raw material has a risk of performance fluctuation, the model will increase the focus on the raw material-related features and give a higher weight in the calculation to generate a blade abnormality feature vector. This vector highlights abnormalities in edge wear, raw material high-temperature performance, and production equipment accuracy.

[0062] The blade abnormal feature vector is processed to generate blade defect warning information, wherein the blade defect warning information is used to characterize the specific type of the target blade defect and the quantitative index of severity. The model determines that the blade has relatively serious defects, and the defect types are excessive wear of the cutting edge, potential material performance defects, and hidden dangers of equipment accuracy in the production process. It is quantified as follows: the defect type is coded as 03 (representing a combination of cutting edge wear, material and equipment problems), and the severity quantitative index is 4 (assuming a level of 1-5, 5 is the most serious). This means that the blade needs to be replaced immediately or further detailed inspection to avoid quality problems and equipment failures during the processing of automobile engine cylinder blocks.

[0063] In this application, the server obtains the image information of the target blade under different viewing angles and light intensities within a preset time period, its usage information (such as processing parameters, etc.), the target user's detection demand information for blade accuracy, life, etc., production information information (raw materials, processes, etc.), preset blade defect warning model and training sample set. For the usage information, first process and generate the blade classification influencing factor and weight value, and then calculate the detection deviation value in combination with the detection threshold of the same type of blades, and then obtain the blade detection influencing factor. The blade image information is normalized and filled into the target scale specification image, and then the features under different viewing angles and lighting are extracted and sorted and integrated into sequence image information. Based on the user's detection demand information, the processing and application scenario information is generated, and the sequence image information is processed accordingly to obtain the matching features, and the target blade parameter information is obtained through image segmentation positioning and further processing. The production information is processed to generate a feature set, and the predicted detection feature vector is obtained after normalization and screening.

[0064] For the training sample set, first obtain the data features and determine the sampling ratio according to the quantity, process the sampling features, combine and generate the data set to train the preset model, and verify it with the verification sample set. The trained model that meets the conditions becomes the target model. Finally, use the target model to process the blade detection influencing factors, parameter information and feature vectors, and generate the blade feature impact correlation value and abnormal feature vector in turn, and finally obtain the warning information that characterizes the specific type and severity of the blade defect quantitative indicators, providing an effective means for CNC blade quality inspection.

[0065] In one embodiment, if Figure 2 As shown, the present application also provides a defect detection device for a CNC blade, comprising:

[0066] An acquisition module 201 is used to acquire blade image information of a target blade within a preset time period, usage information of the target blade, detection requirement information of a target user, production information information of the target blade, a preset blade defect warning model and a training sample set;

[0067] The processing module 202 is used to process the usage information of the target blade to generate a blade detection influencing factor; process the blade image information of the target blade within a preset time period to generate target blade sequence information; process the target blade sequence information based on the detection requirement information of the target user to generate target blade parameter information; process the production information of the target blade to generate a blade prediction detection feature vector; process the preset blade defect warning model based on the training sample set to generate a target blade defect warning model; process the blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector based on the target blade defect warning model to generate blade defect warning information.

[0068] The computer-readable storage medium provided in the above-mentioned embodiments of the present application and the defect detection method for a CNC blade provided in the embodiments of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0069] Each embodiment in the present application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the defect detection method, electronic device, electronic device, and readable storage medium embodiment for evaluating CNC blades, since they are basically similar to the above-mentioned CNC blade defect detection method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the above-mentioned CNC blade defect detection method embodiment.

Claims

1. A method for detecting defects of a numerically controlled blade, characterized in that: include: Obtain blade image information of a target blade within a preset time period, usage information of the target blade, detection requirement information of a target user, production information information of the target blade, a preset blade defect warning model and a training sample set; Processing the usage information of the target blade to generate a blade detection influencing factor; Processing blade image information of the target blade within a preset time period to generate target blade sequence information; Processing the target blade sequence information based on the detection requirement information of the target user to generate target blade parameter information; Processing the production information of the target blade to generate a blade prediction detection feature vector; The preset blade defect warning model is processed based on the training sample set to generate a target blade defect warning model, including: Acquire any number of data features in a training sample set; generate a sampling ratio based on the number of each data feature in the training sample set; process the training sample set based on the sampling ratio to generate a preset number of sampling features; generate multiple data groups based on processing any data feature and each sampling feature, wherein each data group contains a preset number of data samples, and at least one data sample includes identification information; train the preset blade defect warning model based on the data samples in the multiple data groups to generate a trained blade defect warning model; process the trained blade defect warning model based on the verification sample set to generate a verification result; if the data sample containing the identification information in the verification result is a risk factor that represents an impact on the quality of the finished blade product, use the trained blade defect warning model as the target blade defect warning model; The blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector are processed based on the target blade defect warning model to generate blade defect warning information.

2. The method according to claim 1, characterized in that Processing the usage information of the target blade to generate a blade detection influencing factor includes: Processing the usage information of the target blade to generate a blade classification influencing factor and a blade classification weight value; Get the blade detection threshold of the same type as the target blade; Processing the blade classification influencing factor and the blade classification weight value based on a blade detection threshold of the same type as the target blade to generate a detection deviation value of the target blade; The detection deviation value of the target blade is processed to generate a blade detection influencing factor.

3. The method according to claim 1, characterized in that Processing the blade image information of the target blade within a preset time period to generate target blade sequence information includes: Normalizing and filling the blade image information of the target blade within a preset time period to generate blade image information of a target size specification; Processing the blade image information of the target size specification to generate image feature information under different viewing angles and image feature information under different light intensities; The image feature information under different viewing angles and the image feature information under different light intensities are sorted and integrated to generate target blade sequence information.

4. The method according to claim 3, characterized in that The target blade sequence information is processed based on the detection requirement information of the target user to generate target blade parameter information, including: Processing the detection requirement information of the target user to generate processing scenario information of the target blade and application scenario information of the target blade; Processing the target blade sequence information based on the processing scenario information of the target blade and the application scenario information of the target blade to generate target blade feature information, wherein the target blade feature information is a feature that matches the processing scenario information of the target blade and the application scenario information of the target blade; Performing image segmentation and positioning processing on the target blade sequence information based on the target blade feature information to generate target blade image slice information; The target blade image slice information is processed to generate target blade parameter information.

5. The method according to claim 1, characterized in that Processing the production information of the target blade to generate a blade prediction detection feature vector includes: Processing the production information of the target blade to generate blade production feature set information, wherein the blade production feature set information is used to characterize the raw material source information of the blade, the component information of the blade, the production process parameters of the blade, the precision and stability information of the production equipment of the blade, and the production batch information of the blade; Normalizing the blade production feature set information to generate blade production features in a target format; The blade production features of the target format are subjected to feature screening processing to generate a blade prediction detection feature vector.

6. The method according to claim 1, characterized in that The blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector are processed based on the target blade defect warning model to generate blade defect warning information, including: Processing the blade detection influencing factor and the target blade parameter information based on the target blade defect warning model to generate a blade feature impact association value; Processing the blade prediction detection feature vector based on the blade feature impact association value to generate a blade abnormality feature vector; The blade abnormality feature vector is processed to generate blade defect warning information, wherein the blade defect warning information is used to characterize a quantitative index of the specific type and severity of the target blade defect.

7. A defect detection device for a CNC blade, characterized in that: For implementing the defect detection method of the CNC blade according to claim 1, the device comprises: An acquisition module, used to acquire blade image information of a target blade within a preset time period, usage information of the target blade, detection requirement information of a target user, production information information of the target blade, a preset blade defect warning model and a training sample set; The processing module is used to process the usage information of the target blade to generate a blade detection influencing factor; process the blade image information of the target blade within a preset time period to generate target blade sequence information; process the target blade sequence information based on the detection requirement information of the target user to generate target blade parameter information; process the production information of the target blade to generate a blade prediction detection feature vector; process the preset blade defect warning model based on the training sample set to generate a target blade defect warning model; process the blade detection influencing factor, the target blade parameter information and the blade prediction detection feature vector based on the target blade defect warning model to generate blade defect warning information.

8. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the defect detection method for a CNC cutting blade according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the defect detection method for a CNC blade according to any one of claims 1 to 6 is implemented.

Images