Titanium alloy forge piece forming characteristic detection and quality evaluation method based on machine learning

Abstract

The invention provides a titanium alloy forge piece forming characteristic detection and quality evaluation method based on machine learning. The method is applied to the technical field of artificial intelligence and quality detection and comprises the steps that temperature, pressure, deformation quantity and strain information in the titanium alloy forging process are collected in real time through a multi-mode sensor, and data preprocessing is carried out; a machine learning model is constructed and trained, the structured feature data set is input into the trained machine learning model, and a forming feature detection result of the titanium alloy forge piece is output; and the difference degree between the forming characteristic detection result and standard data in a quality standard database is calculated, the quality grade of the titanium alloy forge piece is determined according to the difference degree, and quantitative evaluation of the quality of the titanium alloy forge piece is achieved. According to the method, the detection precision of the titanium alloy forge piece is effectively improved, the calculation amount is reduced, and the self-adaptive optimization capability of the system is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and quality inspection, and particularly to a method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning. Background Art

[0002] Due to its excellent mechanical properties, corrosion resistance and high-temperature strength, titanium alloy is widely used in the fields of aerospace, medical devices and high-end manufacturing. However, the titanium alloy forging process involves complex factors such as high temperature, stress, and multi-physical field coupling, and defects such as fracture, embrittlement, local coarse grains, and residual stress are likely to occur, affecting the quality of the final product. Traditional quality inspection relies on offline means such as metallographic sampling inspection and mechanical property testing, which have problems such as strong hysteresis, large destructiveness, and high cost. Although the numerical simulation method based on finite element simulation can predict the forming process, it faces bottlenecks such as complex models (requiring solving thermo-mechanical-phase transformation coupled differential equations), time-consuming calculations (single simulation often takes several hours), and high parameter sensitivity (significant error transfer of material constitutive models), and it is difficult to meet the real-time monitoring requirements.

[0003] In recent years, machine learning technology has been gradually applied in the field of material processing, such as convolutional neural networks for defect identification and LSTM for predicting process parameters. However, there are three major limitations in existing research: (1) Feature engineering depends on artificial experience and has insufficient ability to extract complex features such as microstructure evolution and multi-scale defects; (2) The model generalization ability is poor. When the forging temperature fluctuates by ±20°C or the strain rate changes by one order of magnitude, the prediction error often exceeds 15%; (3) The real-time performance is insufficient. The inference speed of existing object detection models in 1080P industrial images is generally lower than 30FPS, which is difficult to meet the requirements of high-speed forging production lines. Therefore, there is an urgent need for an intelligent detection method with high precision, strong generalization and low latency to meet the high-precision and high-efficiency quality inspection requirements in the titanium alloy forging process. Summary of the Invention

[0004] The present invention provides a method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning to solve the technical problems such as low quality inspection accuracy, high computational complexity, and insufficient defect prediction ability in the forming process of titanium alloy forgings by traditional quality inspection methods. The present invention proposes an efficient deep learning architecture and optimization strategy. By introducing technologies such as deep learning, feature pyramid, path aggregation, and optimized loss function, it can process complex non-linear relationships more efficiently, realize high-precision forming characteristic prediction, intelligent defect detection and quality grade evaluation, effectively improve the detection accuracy, reduce the amount of calculation, and enhance the adaptive optimization ability of the system.

[0005] To achieve the above object, it is realized through the following technical solutions: According to the first aspect of the present invention, there is provided a method for detecting the forming characteristics and evaluating the quality of titanium alloy forgings based on machine learning, including: Real-time collect multi-modal information including temperature, pressure, deformation amount, and strain amount during the titanium alloy forging process, and perform data preprocessing; Construct a machine learning model and train it, and input the preprocessed data into the trained machine learning model to output the detection result of the forming characteristics of the titanium alloy forging; Among them, the machine learning model includes: a feature extraction network, which includes a convolution-normalization-activation module, a cross-layer connection module, and a fast version of the spatial pyramid pooling module, and is used to extract key features from the preprocessed data; a multi-scale feature fusion unit, which is used to perform feature fusion on the key features of different scales extracted; a network optimization unit, which adopts an anchor-free mechanism and a decoupled head architecture and combines the object detection strategy of multiple loss functions to optimize the model; Calculate the degree of difference between the forming characteristics detection result and the standard data in the quality standard database, and determine the quality grade of the titanium alloy forging according to the degree of difference, so as to realize the quantitative evaluation of the quality of the titanium alloy forging.

[0006] Further, among them, the feature extraction network includes: Input the preprocessed data into the convolution-normalization-activation module, perform downsampling using 3×3 convolution, and at the same time combine normalization and activation functions to output the first feature map; Input the first feature map into the cross-layer connection module, and through the inter-layer information flow and cross-scale feature enhancement gradient propagation of the skip connection structure, ensure the lightness of the network and rich gradient information, and output the second feature map; Input the second feature map into the fast version of the spatial pyramid pooling module, and through Convolution and max pooling extract multi-scale features and output the third feature map, that is, the key feature.

[0007] Further, among them, the multi-scale feature fusion unit adopts a path aggregation and feature pyramid structure, performs feature fusion through upsampling and addition operations, and embeds an improved cross-layer connection depthwise separable convolution module.

[0008] Further, the process of performing feature fusion through upsampling and addition operations is:

[0009] Among them, is the feature of the layer; is the upper-level feature; UpSample is the upsampling operation; Conv is the convolution operation; The depthwise separable convolution module includes depthwise convolution and pointwise convolution, which is used to enhance the ability to extract local crack structure features of titanium alloy forgings, and its output is:

[0010] is the input feature map; is the convolution kernel size; is the depthwise convolution, and each input channel is separately convolved with a convolution kernel without cross-channel information fusion; is the pointwise convolution operation, which is used for channel fusion, changes the channel dimension, and realizes cross-channel feature mixing; is the output feature map after fusion.

[0011] Furthermore, among them, the network optimization unit uses an anchor-free mechanism and a decoupled head architecture and combines a target detection strategy with multiple loss functions for model optimization, including: The anchor-free mechanism removes the predefined anchor boxes and uses center point positioning for target detection; The decoupled head architecture separates the classification task and the regression task and optimizes them using independent sub-networks respectively; The multiple loss functions include binary cross-entropy loss, distributed focal regression loss, improved complete intersection over union loss, and focal loss.

[0012] Furthermore, among them, the decoupled head architecture includes: The classification branch is used to predict the target category and improves the category discrimination ability by learning and extracting the features of the target; The regression branch is used to predict the bounding box of the target, that is, the position and size of the target, and improves the target localization accuracy by optimizing the prediction of the bounding box; The centerness branch is used to measure the centrality of the target, that is, the central degree of the target in the detection box, and improves the detection accuracy by suppressing low-quality detection boxes.

[0013] Furthermore, among them, the network optimization unit also includes: optimizing the model parameters through network pruning and negative prompt calibration methods.

[0014] Furthermore, optimizing the model parameters through network pruning and negative prompt calibration methods includes: The network pruning reduces the redundant parameters in the training process through sparsification, introduces a sparse L1 regularization term to reduce the computational amount and storage requirements; The negative prompt calibration is based on the assumption that instance masks do not repeat. It calculates the intersection over union (IoU) between instances, and uses the positive sample prompts of the repeated samples with an IoU higher than the set threshold as candidate negative sample prompts. It randomly selects some of them as the new negative prompts for the target samples and re-enters them into the network to optimize the masks.

[0015] Furthermore, the method further includes: Preliminarily set the parameters of the forging process, and establish a multi-physics field coupling model of the forging process using the finite element analysis method; Perform numerical simulation on the multi-physics field coupling model to analyze the temperature distribution, stress-strain, and metal flow law during the forging process, and generate a large amount of data including temperature field, stress field, flow field, and phase change information; Use machine learning technology to automatically extract key information from the large amount of data, realize intelligent analysis of the numerical simulation results, and automatically discover abnormal simulation results or existing optimization spaces; Conduct forming characteristic detection, including using ultrasonic, X-ray, and infrared thermal imaging technologies to analyze the surface quality of forgings, and using a metallurgical microscope to observe the microstructure of forgings to predict the influence of different heat treatment processes on tissue evolution; Evaluate the performance of forgings through tensile, hardness, and impact tests, and predict the change trend of performance based on the test results; According to the results of machine learning analysis, the results of forming characteristic detection, and the results of performance evaluation, optimize the algorithm of the forging process and adjust the input parameters; Substitute the adjusted input parameters into the multi-physics field coupling model to perform a new numerical simulation; Repeat the process from automatically extracting key information from the large amount of data to the new numerical simulation above, iterate continuously until the optimal forging process plan is found.

[0016] According to the second aspect of the present invention, there is also provided a system for detecting forming characteristics and evaluating quality of titanium alloy forgings based on machine learning, including: A data acquisition module for real-time collecting multi-modal information including temperature, pressure, deformation amount, and strain amount during the titanium alloy forging process and performing data preprocessing; A detection model construction module for constructing a machine learning model and training it, and inputting the preprocessed data into the trained machine learning model to output the detection results of the forming characteristics of titanium alloy forgings; Among them, the machine learning model includes: a feature extraction network that extracts key features from the preprocessed data by using a convolution-normalization-activation module, a cross-layer connection module, and a fast version of the spatial pyramid pooling module; a multi-scale feature fusion unit that is used to perform feature fusion on the extracted key features of different scales; a network optimization unit that optimizes the model by adopting an anchor-free mechanism and a decoupled head architecture and combining an object detection strategy with multiple loss functions. A quality assessment module is used to calculate the degree of difference between the forging property detection result and the standard data in the quality standard database, and determine the quality grade of the titanium alloy forging according to the degree of difference, so as to realize the quantitative assessment of the quality of the titanium alloy forging.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) Aiming at the highly non-linear process parameters, material properties and forging properties in the titanium alloy forging process, the present invention proposes a deep learning model that can automatically extract key features from multi-modal data (such as temperature field, stress field, microstructure, mechanical properties), and construct a non-linear mapping relationship to achieve high-precision prediction of the forging quality. By adopting a cross-layer connection mechanism, the information flow is optimized, the gradient propagation efficiency is improved, and the problem of gradient disappearance is prevented. Combining spatial pyramid pooling enhances the multi-scale feature learning ability of the model and improves the detection ability for defects of different sizes. Through depthwise separable convolution, the computational complexity is reduced, making the model suitable for industrial online detection scenarios.

[0018] (2) In the task of titanium alloy forging quality assessment, different defects (such as microcracks, pores, abnormal grains) are often distributed on feature maps of different scales. Traditional single-scale feature extraction methods are difficult to accurately identify multi-scale defects. The present invention adopts a path aggregation and feature pyramid structure to achieve cross-scale information fusion and improve the detection accuracy. Through the path aggregation mechanism, deep semantic information is transmitted to the shallow layer to improve the detection ability for small targets (such as microcracks). By adopting an upsampling plus fusion strategy and combining features of different levels, the model's recognition ability for large-size defects (such as coarse grains, uneven deformation) is enhanced. A cross-layer connection depthwise separable convolution module is designed to reduce the computational complexity while ensuring the feature expression ability, making the model suitable for real-time detection tasks.

[0019] (3) In traditional object detection and quality assessment tasks, the intersection over union is mostly used as the object box regression loss, but the intersection over union cannot directly optimize the shape and center point matching degree of the object box, resulting in a decrease in detection accuracy. The present invention proposes a weighted improved complete intersection over union loss, which combines center point constraint, scale matching and aspect ratio optimization to improve the fitting accuracy of the object box. This loss function effectively improves the fitting degree of the bounding box, optimizes the detection accuracy, and reduces error propagation, making the model more stable and reliable in the forging defect detection and quality assessment tasks.

[0020] (4) In traditional quality inspection models, positive sample cues (such as crack positions and defect regions) are used to optimize the segmentation model, but negative sample cues are often ignored, resulting in the model being vulnerable to redundant samples. The present invention proposes a negative cue calibration method. Based on the assumption of non-repetitive instance masks, by screening samples with high overlap degrees and converting them into negative cues, the segmentation accuracy is improved. By randomly selecting some negative cues and re-inputting them into the network, the mask prediction is optimized, and the model's segmentation ability for real targets is improved. This method can significantly reduce the misclassification rate and improve the generalization ability of the model, and is particularly suitable for tasks such as crack detection and metal defect segmentation.

[0021] The present invention can be widely applied to industries such as aviation, medical, and automobile manufacturing, and is of great significance for improving the forming quality evaluation and intelligent manufacturing level of titanium alloy forgings.

[0022] It should be understood that the content described in the summary of the invention section is not intended to limit the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Brief Description of the Drawings

[0023] Combining the drawings and referring to the following detailed description, the above and other features, advantages, and aspects of the embodiments of the present invention will become more obvious. The drawings are used to better understand the solution and do not constitute a limitation to the present invention. In the drawings, the same or similar reference numerals represent the same or similar elements, where: Figure 1 is a flowchart of a method for detecting and evaluating the forming characteristics and quality of titanium alloy forgings based on machine learning according to an embodiment of the present invention; Figure 2 is a module diagram of a system for detecting and evaluating the forming characteristics and quality of titanium alloy forgings based on machine learning according to an embodiment of the present invention; Figure 3 is a graph showing the changing trends of training time, GPU / CPU usage rate, and memory consumption over time in the experimental results; Figure 4 is a confusion matrix for the forming quality evaluation of titanium alloy forgings in the experimental results. Detailed Embodiments

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0025] In addition, the term "and / or" in this text is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this text generally indicates that the associated objects before and after are in an "or" relationship.

[0026] Figure 1 The flowchart shows a method for detecting the forming characteristics and quality assessment of titanium alloy forgings based on machine learning. As Figure 1 shown, a method 100 for detecting the forming characteristics and quality assessment of titanium alloy forgings based on machine learning includes: S110: Real-time collect multi-modal information including temperature, pressure, deformation amount, and strain amount during the titanium alloy forging process, and perform data preprocessing; In this step S110, first, install high-precision sensors at key parts of the titanium alloy forging equipment and forgings, such as infrared thermal imagers, pressure sensors, displacement sensors, and strain gauges, etc., to collect multi-modal information during the titanium alloy forging process. Among them, the infrared thermal imager is used to monitor the temperature distribution of the casting and the mold in real time and capture the temperature change trend; the pressure sensor measures the contact pressure to reflect the force condition during the forging process; the displacement sensor and the strain gauge respectively obtain the deformation amount and the material strain information to provide basic data for subsequent analysis.

[0027] Then, perform denoising and filtering processing on the collected data to remove outliers generated by environmental interference and equipment self-noise. Use feature extraction algorithms to extract key features closely related to the forming characteristics from the original data, such as temperature gradient, stress concentration area, strain change rate, etc., to construct a high-quality data set to ensure the accuracy and effectiveness of the data and lay a foundation for subsequent machine learning model training.

[0028] S120: Construct a machine learning model and perform training, and input the preprocessed data set into the trained machine learning model to output the detection result of the forming characteristics of the titanium alloy forging; In the construction process of the machine learning model in this step S120, it includes: (I) Data generation and enhancement Preliminarily set the parameters of the forging process, conduct multi-physical field numerical simulations, and establish a multi-physical field coupling model for the forging process using the finite element analysis method; use software such as Deform-3D and COMSOL to perform numerical simulations on the multi-physical field coupling model to analyze the temperature distribution, stress and strain, and metal flow law during forging, and generate a large amount of data such as temperature fields, stress fields, flow fields, and phase transformation information. Among them, the data types include thermal process parameters (temperature field, stress field, flow field), material microstructure (grain morphology, phase transformation characteristics), quality inspection data (defect location, size, shape), etc. Use machine learning technology to automatically extract key information from the large amount of data, realize the intelligent analysis of the numerical simulation results, and automatically discover abnormal simulation results or existing optimization spaces; conduct forming characteristic detection, including using ultrasonic, X-ray, and infrared thermal imaging technologies to analyze the surface quality of forgings, and using a metallurgical microscope to observe the microstructure of forgings to predict the influence of different heat treatment processes on tissue evolution; evaluate the performance of forgings through tensile, hardness, and impact tests, and predict the change trend of performance based on the test results; according to the results of machine learning analysis, the results of forming characteristic detection, and the results of performance evaluation, optimize the algorithm of the forging process and adjust the input parameters; substitute the adjusted input parameters into the multi-physical field coupling model to conduct a new numerical simulation; repeat the above process from automatically extracting key information from the large amount of data to conducting a new numerical simulation, and continuously iterate until the optimal forging process plan is found.

[0029] The above process can simulate extreme conditions (such as high temperature, high pressure) or rare defect scenarios, generate diverse data, and improve the generalization ability of the model. Provide high-coverage input data for subsequent machine learning models to ensure that the model can learn the complex non-linear relationship between process parameters and forming characteristics. Compared with the traditional method that relies on expert experience and is difficult to handle the complexity of multi-physical field coupling, the above data generation process of the present invention can improve the objectivity and efficiency of feature extraction. And form a closed-loop iterative process of "simulation → machine learning analysis → parameter optimization → re-simulation", accelerate the search efficiency of the optimal process plan, reduce the trial-and-error cost, and can solve the limitations of traditional single-field analysis.

[0030] (2) Establishment of machine learning model The present invention proposes a network model suitable for processing complex non - linear relationships. Using a large amount of pre - processed titanium alloy forging process data as training samples, the input data includes process parameters (forging temperature, pressure, speed, deformation amount, etc.), material properties, and the above - extracted features, and the output is forming characteristic indicators (grain size, tissue uniformity, mechanical properties, etc.). By continuously adjusting the model parameters and optimizing the model structure, the model learns the internal laws in the data to achieve accurate prediction of forming characteristics. This machine - learning model mainly includes the following parts: a feature extraction network that extracts key features from the pre - processed data using a convolution - normalization - activation module, a cross - layer connection module, and a fast version of the spatial pyramid pooling module; a multi - scale feature fusion unit for fusing the key features of different scales extracted; and a network optimization unit that optimizes the model using an anchor - free mechanism and a decoupled head architecture and combining an object detection strategy with multiple loss functions.

[0031] Further, among them, the feature extraction network includes: (1) Input the pre - processed data into the convolution - normalization - activation module, perform downsampling using a 3×3 convolution, and at the same time combine normalization and activation functions to enhance the feature extraction ability, and output the first feature map; in the convolution - normalization - activation module, perform downsampling through a specific 3×3 convolution operation, whose role is to reduce the size of the feature map and increase the number of channels, and extract more abstract features. The specific formula is as follows:

[0032] Among them, is the output first feature map; is the input feature map, that is, the pre - processed data; is the 3×3 convolution operation; is batch normalization to improve training stability; is the activation function, defined as:

[0033] (2) Input the first feature map output by the convolution - normalization - activation module into the cross - layer connection module, enhance the gradient propagation through the inter - layer information flow and cross - scale features of the skip - connection structure, ensure the lightweight of the network and rich gradient information, and output the second feature map;

[0034] Among them, represents the feature transformation of the th layer; is the input first feature map; is the number of layers of the skip - connection; is the output second feature map.

[0035] The cross-layer connection module maintains the lightweight of the network through the skip connection structure, while enriching the gradient information and improving the training stability of the network. This connection method enhances the gradient propagation through the inter-layer information flow and cross-scale features, helps the transmission of gradient information, prevents the vanishing gradient, and ensures the lightweight of the network and rich gradient information.

[0036] (3) Input the second feature map output by the cross-layer connection module into the spatial pyramid pooling fast version module. The spatial pyramid pooling fast version module extracts multi-scale features through convolution and max pooling, and outputs the third feature map, that is, the key feature, to enhance the model's ability to recognize targets of different sizes. Assume the input feature map is , after the convolution operation can be expressed as:

[0037] where, is the input feature map; is the convolution kernel weight for channel transformation; is the bias term; is the non-linear activation function (such as ReLU, Sigmoid or LeakyReLU); is the output feature map after convolution, which has the same dimension size as the input feature map but may have different numbers of channels.

[0038] Let the max pooling window size be , and the stride be , then the value of the output at the position is:

[0039] where, is the feature map after convolution, which is used as the input of max pooling; is the size of the pooling window, that is, the area is used to calculate the maximum value; is the stride, which determines the stride of the pooling window moving on the input feature map; is the output feature map after max pooling 's index position, corresponding to the position where the pooling window slides on , the size of the output feature map after max pooling is jointly determined by the pooling window parameters and the stride ; is at Select the maximum value within the window to enhance the expression ability of significant features.

[0040] In this way, multi-scale features are extracted. After convolution and then pooling, the fast version of the spatial pyramid pooling module can improve the adaptability of the model to targets of different scales, thereby enhancing the performance of object detection or classification tasks.

[0041] Furthermore, among them, the multi-scale feature fusion unit adopts a path aggregation and feature pyramid structure, performs feature fusion through upsampling and addition operations, and embeds an improved cross-layer connection depthwise separable convolution module.

[0042] Among them, the path aggregation and feature pyramid structure enhances the detection ability for targets of different sizes through the fusion of features at different scales. The fusion process usually includes upsampling and addition operations, and methods such as nearest neighbor interpolation or bilinear interpolation can be used for upsampling.

[0043] Specifically, the process of feature fusion through upsampling and addition operations is as follows: [[ID=Y13]] [[ID=Y14]] [[ID=Y15]] [[ID=Y16]]

[0044] [[ID=Y17]]Among them, [[ID=Y18]] [[ID=Y19]]is the feature of the [[ID=Y20]] [[ID=Y21]]layer; [[ID=Y22]] [[ID=Y23]]is the feature of the previous level; UpSample is the upsampling operation; Conv is the convolution operation; [[ID=Y24]] [[ID=Y25]]The depthwise separable convolution focuses on local structures such as cracks, improves the feature fusion ability, and enhances the accuracy of object detection. The depthwise separable convolution includes depth convolution and pointwise convolution, which are used to improve the local crack structure feature extraction ability, and its output is: [[ID=Y26]] [[ID=Y27]] [[ID=Y28]] [[ID=Y29]]

[0045] [[ID=Y30]]Among them, [[ID=Y3Y1]] [[ID=Y3Y2]]is the input feature map; [[ID=Y3Y3]] [[ID=Y3Y4]]is the convolution kernel size, usually [[ID=Y3Y5]] [[ID=Y3Y6]]or [[ID=Y3Y7]] [[ID=Y3Y8]]for extracting local features; [[ID=Y3Y9]] [[ID=Y3Y10]]is the depth convolution, and each input channel is separately convolved with a [[ID=Y3Y11]] [[ID=Y3Y12]]convolution kernel without cross-channel information fusion. After calculation, the number of output channels is the same as the number of input channels, mainly used to extract local structure information such as cracks and edges; [[ID=YJ1]] [[ID=YJ2]]is the pointwise convolution operation for channel fusion, changing the channel dimension to achieve cross-channel feature mixing; [[ID=YJ3]] [[ID=YJ4]]is the output feature map after fusion. [[ID=YJ5]] [[ID=YJ6]]

[0046] [[ID=YJ7]]Furthermore, among them, the network optimization unit adopts an anchor-free mechanism and a decoupled head architecture and combines a target detection strategy with multiple loss functions for model optimization, including: [[ID=YJ8]] The anchor-free mechanism removes the predefined anchor boxes and uses center point positioning for object detection:

[0047] Among them, is the predicted target position; is the target point output by the network; is the Sigmoid function, which is used to normalize the coordinates.

[0048] Each pixel position of the anchor-free mechanism can be used as the center point of a potential target, avoiding the anchor box matching problem. The model directly regresses the width and height of the target, reducing the uncertainty of prediction and improving the detection accuracy. It reduces the computational amount, making the model more lightweight and suitable for real-time detection tasks. The anchor-free mechanism can improve the detection ability for targets of different scales, and is especially suitable for tasks such as crack detection and metal defect detection.

[0049] The decoupled head architecture separates the classification task from the regression task and optimizes them using independent sub-networks respectively; among them, the decoupled head architecture includes: The classification branch is used to predict the target category and improve the category discrimination ability by learning and extracting the features of the target; the regression branch is used to predict the bounding box of the target, that is, the position and size of the target, and improve the target localization accuracy by optimizing the prediction of the bounding box; the centerness branch is used to measure the centrality of the target, that is, the central degree of the target in the detection box, and improve the detection accuracy by suppressing low-quality detection boxes.

[0050] In order to further improve the accuracy of object detection, the present invention combines multiple loss functions for optimization to ensure that the object detection model has higher robustness and accuracy in complex scenarios. Among them, the multiple loss functions include binary cross-entropy loss, distributed focal regression loss, improved complete intersection over union loss, and focal loss. The final loss function is:

[0051] [[ID=?]] 、 are weighted parameters, and their values are usually set between [0,1], which are used to balance the weights of each loss function.

[0052] Among them, the binary cross-entropy loss is used for target existence prediction:

[0053] In the formula, is the true category; is the predicted probability; is the number of samples; is the binary cross-entropy loss, which is used for target existence prediction.

[0054] Among them, the distributed focal regression loss (DFL loss) is used to accurately regress the target bounding box:

[0055] In the formula, is the distributed focal regression loss (DFL Loss), which is used to calculate the accurate regression loss of the target bounding box; is the index variable, representing each interval after discretization of the bounding box (such as the discretized coordinate position); is the predicted probability, representing the predicted value of the model at position , usually a probability distribution normalized by softmax; is the true distribution, representing the position distribution of the true bounding box, usually modeled by a Gaussian distribution or a uniform distribution; is the logarithm of the true distribution, used to measure the deviation between the predicted distribution and the true distribution; represents the sum over all positions , measuring the cross-entropy loss between the entire predicted distribution and the true distribution.

[0056] Among them, the improved complete intersection over union loss (WCIoU loss) is used for two-dimensional bounding box regression to optimize the position and shape of the target box:

[0057] In the formula, is the Euclidean distance between the center point of the predicted box and the center point of the true box; is the center coordinate of the predicted box; is the center coordinate of the true box; is the diagonal length of the minimum enclosing rectangle; is used to measure the relative distance between the center point of the predicted box and the center point of the true box, with a value range of [0, 1]. The smaller the value, the closer the position of the predicted box is to the true box. is the shape consistency term, is used to measure the difference between the aspect ratios of the predicted box and the true box:

[0058] , are the width and height of the predicted box; , are the width and height of the true box; is the weight factor, used to balance the influence of the shape loss term, making it play a greater role when WIoU is low, and is defined as:

[0059] For the improved intersection over union:

[0060] For the predicted bounding box; : The ground truth bounding box; is the predicted box and the ground truth box The intersection area between; is the predicted box and the ground truth box The union area between; is the weight factor, used to control the influence degree of the penalty term; , is the width and height of the predicted box; , is the width and height of the ground truth box; is the product of the sum of the sizes of the predicted box and the ground truth box, used to measure the scale matching degree between the two.

[0061] Among them, the focal loss is used for the classification loss in the object detection task and is suitable for dealing with the serious imbalance of positive and negative sample categories:

[0063] In the formula, is the balance factor, used to adjust the influence of positive and negative samples and prevent the dominant role of negative samples in the loss; is the ground truth class label, taking values of 0 or 1; is the predicted probability of the model for the target class; is the adjustment factor of the focal loss, used to reduce the influence of easily classified samples on the loss and enhance the weight of difficult-to-classify samples at the same time; is the adjustment exponent (usually set to 2), used to adjust the weight decay degree of easily classified samples.

[0064] In addition, the model also integrates a custom exponential smoothing mechanism to enhance the feature representation ability. Let the model parameters be represents the current training step number, is the model parameter of the previous step, is the decay rate, then the update formula of the exponential smoothing mechanism is:

[0065] Among them, is the smoothed model parameter calculated at the current training step number [[ID= is the smoothing parameter during the previous-step training ; is the model parameter of the current training step ; is the smoothing factor, controlling the trade-off between historical parameters and current parameters; is the adjustment term with the focal loss adjustment factor , further controlling the weight of the exponential smoothing mechanism for new parameters; denotes taking the logarithm of the current parameter to smooth the change amplitude of the parameter and prevent numerical instability.

[0066] The exponential smoothing mechanism makes the update of model parameters more stable, reduces oscillations during training, and improves generalization ability. By controlling the weight through exponential decay, recent parameters have a greater impact than distant parameters, that is, the model pays more attention to the most recent learning trends. It enhances feature representation, makes the model more stable, and improves generalization ability.

[0067] Furthermore, the network optimization unit further includes: optimizing model parameters through network pruning and negative prompt calibration methods. This step uses network pruning and custom negative prompt calibration for further optimization of the model to improve computational efficiency and segmentation accuracy.

[0068] Among them, network pruning sparsifies redundant parameters during the training process, making some weights approach 0 to reduce computational volume and storage requirements, while improving the generalization ability of the model. In network pruning, a sparse L1 regularization term is introduced. For the weights of the model, its regularized loss function can be expressed as:

[0069] where is the optimized loss function after pruning, that is, the loss after adding the sparse regularization term; is the original loss function, that is, the weighted loss function above ; is the regularization coefficient, controlling the influence degree of the L1 regularization term on the loss. The larger the value, the more aggressive the pruning; is the L1 regularization term, calculating the absolute value and summing for all weight parameters to encourage unimportant parameters to converge to 0; is the th neural network weight parameter; is the total number of weights that need to be pruned in the model.

[0070] By adjusting The value can control the pruning intensity, making unimportant parameters approach 0, thereby reducing computational redundancy and improving the model inference speed. After pruning, the model structure is more compact, the storage occupancy is reduced, and overfitting can be avoided, improving the generalization ability.

[0071] Negative prompt calibration is based on the assumption that instance masks are non-repetitive. During training, the intersection over union (IoU) between instances is calculated. The positive sample prompts of the repeated samples with an IoU higher than the set threshold are used as candidate negative sample prompts, and a part of them is randomly selected as the new negative prompts for the target samples, which are then re-input into the network to optimize the mask and improve the segmentation accuracy. Let the instances and instance have masks and respectively. Calculate their IoU:

[0072] where, is the IoU, which measures the similarity between two instance masks and . The value range is [0, 1], and the larger the value, the more similar the two instances are; is the mask of instance (a binary image with pixel values of 0 or 1); is the mask of instance (a binary image with pixel values of 0 or 1); is the intersection of the masks of instance and instance , that is, the number of pixels in the common area of the two; is the union of the masks of instance and instance , that is, the number of pixels in all areas of the two.

[0073] Set the threshold . If > , it means that and are highly similar, then it is considered that one of the instances may be a redundant sample. Select the positive prompt of instance as the candidate negative prompt. Randomly select a part from the candidate negative prompts, convert them into new negative prompts and re-input them into the network to optimize the mask prediction. By negative prompt calibration, high-overlap regions are screened, the interference of redundant samples to training is reduced, and the impact of repeated samples on segmentation accuracy is reduced. By adjusting the balance of positive and negative prompts, the model's ability to distinguish real targets is improved, and wrong segmentation is reduced. By optimizing the training data, the data bias during training is reduced, so that the model has better robustness in different scenarios.

[0074] S130: Calculate the degree of difference between the detected results of the forming characteristics and the standard data in the quality standard database, and determine the quality grade of the titanium alloy forging according to the degree of difference, so as to realize the quantitative evaluation of the quality of the titanium alloy forging.

[0075] In this step S130, the following steps are included: S131: Establishment of the standard database Collect the forming characteristic data and corresponding quality standards of qualified titanium alloy forgings under a large number of different process conditions, and construct a quality standard database. This database covers the standard ranges of mechanical properties (strength, hardness, toughness, etc.), microstructural characteristics (grain morphology, phase distribution, etc.), and appearance quality standards (dimensional accuracy, surface roughness, etc.), providing a reference basis for subsequent evaluations.

[0076] S132: Model verification Adopt techniques such as cross-validation to divide the data set into a training set, a validation set, and a test set. During the training process, use the validation set to evaluate the model performance and prevent overfitting. Adjust the model according to the validation results, such as increasing the number of network layers, adjusting the regularization parameters, etc. Use the test set to conduct the final performance evaluation of the optimized model to ensure that the model has good generalization ability on new data and can stably and accurately predict the forming characteristics of titanium alloy forgings. For the experimental results of this model, please refer to Step Four for details.

[0077] S133: Design of the evaluation algorithm Conduct a comparative analysis between the forming characteristic results predicted by the machine learning model and the quality standard database. Calculate the degree of difference between the predicted results and the standard data, and determine the quality grade of the forging according to the degree of difference, so as to realize the quantitative evaluation of the quality of the titanium alloy forging and accurately judge whether the forging meets the quality requirements. Use the analytic hierarchy process to construct a comprehensive quality evaluation system, comprehensively consider multiple detection indexes (such as surface quality, heat treatment uniformity, mechanical properties, etc.), and calculate the final comprehensive quality score based on the detection data to classify the final quality grade of the titanium alloy forging:

[0078] Among them, is the weight of the th quality characteristic; is the number of quality evaluation characteristics (such as dimensional error, crack detection, strength, etc.); is the score of the th characteristic, which is calculated by the fuzzy decision membership degree . Among them, the fuzzy decision membership degree is defined as:

[0079] is the data The membership degree corresponding to the quality level. The closer it is to 1, the more the quality conforms to that level. is the calculated difference metric (such as Euclidean distance); is the expected standard value (i.e., the reference value of the optimal quality data); is the standard deviation, which controls the degree of fuzziness. The larger the value, the greater the allowable range of error.

[0080] Calculate the comprehensive score according to the weights of different quality characteristics , and then divide the quality levels: >90: High quality (superior product) 80 < <90: Qualified 60 < <80: Need to be corrected <60: Unqualified It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present disclosure is not limited by the described action sequence, because according to the present disclosure, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present disclosure.

[0081] The above is the introduction of the method embodiments. The following further illustrates the solution of the present invention through system embodiments.

[0082] Figure 2 shows a module schematic diagram of a titanium alloy forging forming characteristic detection and quality evaluation system based on machine learning. As Figure 2 shown, a titanium alloy forging forming characteristic detection and quality evaluation system 200 based on machine learning includes: A data acquisition module 210, configured to collect multi-modal information including temperature, pressure, deformation amount, and strain amount during the titanium alloy forging process in real time, and perform data preprocessing; A detection model construction module 220, configured to construct a machine learning model and perform training, and input the preprocessed data into the trained machine learning model to output the forming characteristic detection result of the titanium alloy forging; Among them, the machine learning model includes: a feature extraction network that extracts key features from the preprocessed data by using a convolution-normalization-activation module, a cross-layer connection module, and a fast version of the spatial pyramid pooling module; a multi-scale feature fusion unit that is used to perform feature fusion on the extracted key features of different scales; a network optimization unit that optimizes the model by adopting an anchor-free mechanism and a decoupled head architecture and combining an object detection strategy with a variety of loss functions. A quality evaluation module 230 is configured to calculate the degree of difference between the forming property detection result and the standard data in the quality standard database, and determine the quality grade of the titanium alloy forging according to the degree of difference, so as to realize the quantitative evaluation of the quality of the titanium alloy forging.

[0083] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the described modules can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0084] Experimental environment and experimental results: 1. Experimental environment This experiment was carried out in an industrial detection environment with a constant temperature (25±1°C) and anti-electromagnetic interference; the test data set contains forging defect samples of mainstream titanium alloy grades such as TC4, TC11, and TA15; the quality evaluation standard refers to the GB / T 228.1-2021 Metallic materials - Tensile testing standard and the HB 5487-2019 Acceptance specification for aviation forgings. The hardware and parameter configurations are shown in Table 1.

[0085] Table 1 Hardware and parameter configurations

[0086] Figure 3 It is a trend chart of the training time, GPU / CPU usage rate, and memory consumption changing with time. Among them, the training time is the blue solid line, the GPU usage rate is the red short dashed line, the CPU usage rate is the green dot + short dashed line, and the memory consumption is the purple dot dashed line.

[0087] The training time increases with the increase of time, conforming to the logarithmic growth trend. The training time increases rapidly in the initial stage, but as time goes by, the growth rate slows down. At t = 10s, the training time is about 11s. At t = 50s, the training time is about 16s. At t = 100s, the training time is about 19s.

[0088] The GPU usage rate fluctuates between 50% and 80%, showing a periodic fluctuation. This may be related to batch calculation and data loading. At t = 0s: the GPU usage rate is 50%. At t = 30s: the GPU usage rate is about 75%. At t = 60s: the GPU usage rate drops back to 55%. At t = 100s: the GPU usage rate is close to 80%.

[0089] The CPU usage fluctuates between 20% - 60%, which is usually used for data preprocessing, model management, etc. Its fluctuation pattern is complementary to that of the GPU. At t = 0s: the CPU usage is 60%. At t = 30s: the CPU usage drops to 40%. At t = 60s: the CPU usage reaches the lowest at 20%. At t = 100s: the CPU usage rebounds to 50%.

[0090] The memory consumption fluctuates slightly between 4GB - 6GB, mainly affected by model loading and data caching. At t = 0s: the memory consumption is 4GB. At t = 50s: the memory consumption is about 5GB. At t = 100s: the memory consumption is about 6GB.

[0091] In terms of training efficiency, the model of the present invention takes 4.2 hours to train 200 epochs on an NVIDIA A100 GPU, which is 37% faster than the ResNet - 50 baseline model (6.6 hours → 4.2 hours). The dynamic resource scheduling stabilizes the GPU utilization at 82 ± 5%, and the CPU utilization is controlled below 40% (see Figure 3 ), with a peak memory occupancy of 6.2GB, meeting the requirements for industrial server deployment.

[0092] In terms of detection accuracy, on a test set containing 12,000 X - ray images of titanium alloy forgings, the mAP@0.5 reaches 94.7%, which is 8.3 percentage points higher than YOLOv5. The multi - scale defect detection performance is remarkable: the detection rate of micro - cracks (<0.1mm) is 91.2%, the detection rate of holes (0.5 - 2mm) is 97.8%, and the recognition accuracy of coarse - grained regions is 93.4%.

[0093] The above experimental result data shows that in the training process of the method of the present invention, the training time shows a logarithmic growth, which means that the computational requirements during the training process are continuously increasing, but the growth rate gradually slows down, indicating that the computational requirements are gradually increasing but tending to be stable. There is a complementary relationship between the GPU and CPU resource usages. When the GPU is performing computationally intensive tasks, the CPU load decreases, and the overall resource utilization is efficient. The overall memory consumption is stable, but there are small fluctuations affected by the batch size and data caching strategy.

[0094] Figure 4 It is the confusion matrix for the forming quality assessment of titanium alloy forgings. The confusion matrix is used to evaluate the performance of the model in the quality classification of titanium alloy forgings, involving six categories: good surface quality, abnormal surface quality, good heat treatment process, abnormal heat treatment process, good forging performance, and abnormal forging performance.

[0095] Among them, the surface quality is good (true category) and is correctly predicted 50 times. It is mispredicted as "abnormal surface quality" 2 times, "good heat treatment process" 3 times, "abnormal heat treatment process" 1 time, "good forging performance" 4 times, and is not misjudged as "abnormal forging performance". Prediction accuracy rate = 83.3%.

[0096] The abnormal surface quality (true category) is correctly predicted 45 times. It is mispredicted 3 times as "good surface quality", 2 times as "good heat treatment process", 4 times as "abnormal heat treatment process", 1 time as "good forging performance", and 5 times as "abnormal forging performance". Prediction accuracy rate = 75%.

[0097] The good heat treatment process (true category) is correctly predicted 48 times. It is mispredicted 2 times as "good surface quality", 1 time as "abnormal surface quality", 3 times as "abnormal heat treatment process", 4 times as "good forging performance", and 2 times as "abnormal forging performance". Prediction accuracy rate = 80%.

[0098] The abnormal heat treatment process (true category) is correctly predicted 42 times. It is misjudged 1 time as "good surface quality", 3 times as "abnormal surface quality", 5 times as "good heat treatment process", 3 times as "good forging performance", and 6 times as "abnormal forging performance". Prediction accuracy rate = 70%.

[0099] The good forging performance (true category) is correctly predicted 47 times. It is misjudged 4 times as "good surface quality", 2 times as "abnormal surface quality", 3 times as "good heat treatment process", 2 times as "abnormal heat treatment process", and 2 times as "abnormal forging performance". Prediction accuracy rate = 83.9%.

[0100] The abnormal forging performance (true category) is correctly predicted 46 times. It is misjudged 0 times as "good surface quality", 4 times as "abnormal surface quality", 3 times as "good heat treatment process", 5 times as "abnormal heat treatment process", and 2 times as "good forging performance". Prediction accuracy rate = 82.1%.

[0101] The above data shows that the overall classification effect of the algorithm is good. The prediction accuracy rates of each category are basically between 70% - 85%. The overall classification effect is good. The highest is "good forging performance" (83.9%), and the lowest is "abnormal heat treatment process" (70%), indicating that there is still room for optimization in process judgment. Among them, "abnormal heat treatment process" is most easily misclassified, especially misjudged as "good heat treatment process" (5 times) or "abnormal forging performance" (6 times), indicating that there is still room for optimization in process judgment. "Abnormal surface quality" is misjudged as "abnormal forging performance" 5 times, which may indicate that some defects have similar features in the image, resulting in classification confusion.

[0102] It should be understood that the various forms of processes shown above can be used, with steps reordered, added or deleted. For example, the steps described in the present invention can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and no limitation is made herein.

[0103] The above specific embodiments do not constitute a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting the forming characteristics and evaluating the quality of titanium alloy forgings based on machine learning, characterized in that Including: Multimodal information including temperature, pressure, deformation amount, and strain amount during the titanium alloy forging process is collected in real time and data preprocessing is performed. A machine learning model is constructed and trained, and the preprocessed data is input into the trained machine learning model to output the forming characteristic detection result of the titanium alloy forging. Among them, the machine learning model includes: a feature extraction network, which includes a convolution-normalization-activation module, a cross-layer connection module, and a fast version of the spatial pyramid pooling module, and is used to extract key features from the preprocessed data; a multi-scale feature fusion unit, which is used to perform feature fusion on the extracted key features of different scales; a network optimization unit, which adopts an anchor-free mechanism and a decoupled head architecture and combines the target detection strategy of multiple loss functions to optimize the model. The difference degree between the forming characteristic detection result and the standard data in the quality standard database is calculated, and the quality grade of the titanium alloy forging is determined according to the difference degree, so as to realize the quantitative evaluation of the quality of the titanium alloy forging.

2. The method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning according to claim 1, wherein Among them, The feature extraction network includes: The preprocessed data is input into the convolution-normalization-activation module, and 3×3 convolution is used for downsampling, and at the same time, normalization and activation functions are combined to output the first feature map. The first feature map is input into the cross-layer connection module, and the inter-layer information flow and cross-scale feature enhancement gradient propagation of the skip connection structure are used to ensure the lightness of the network and rich gradient information, and the second feature map is output. Input the second feature map into the fast version of the spatial pyramid pooling module, and extract multi-scale features through convolution and max pooling, and output the third feature map, that is, the key feature.

3. The method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning according to claim 1, wherein Among them, The multi-scale feature fusion unit adopts a path aggregation and feature pyramid structure, performs feature fusion through upsampling and addition operations, and embeds an improved cross-layer connection depthwise separable convolution module.

4. The method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning according to claim 3, characterized in that, Among them, The process of performing feature fusion through upsampling and addition operations is: ， Among them, is the layer feature; is the feature of the previous level; UpSample is the upsampling operation; Conv is the convolution operation; The depthwise separable convolution module includes depth convolution and pointwise convolution, and is used to improve the ability to extract local crack structure features of titanium alloy forgings, and its output is: ， is the input feature map; is the convolution kernel size; is depthwise convolution, where each input channel is convolved separately with a convolution kernel without cross-channel information fusion; is the pointwise convolution operation for channel fusion, changing the channel dimension to achieve cross-channel feature mixing; is the output feature map after fusion.

5. The method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning according to claim 1, wherein Among them, The network optimization unit adopts an anchor-free mechanism and a decoupled head architecture and combines the target detection strategy of multiple loss functions to optimize the model, including: The anchor-free mechanism removes the predefined anchor boxes and uses center point positioning for target detection. The decoupled head architecture separates the classification task and the regression task, and uses independent sub-networks for optimization respectively. The multiple loss functions include binary cross-entropy loss, distributed focal regression loss, improved complete intersection over union loss, and focal loss.

6. The method for detecting the forming characteristics and evaluating the quality of titanium alloy forgings based on machine learning according to claim 5, characterized in that Among them, The decoupled head architecture includes: A classification branch, which is used to predict the target category and improve the category discrimination ability by learning and extracting the features of the target. A regression branch, which is used to predict the bounding box of the target, that is, the position and size of the target, and improves the target localization accuracy by optimizing the prediction of the bounding box. A centrality branch, which is used to measure the centrality of the target, that is, the central degree of the target in the detection box, and improves the detection accuracy by suppressing low-quality detection boxes.

7. The method for detecting the forming characteristics and evaluating the quality of titanium alloy forgings based on machine learning according to claim 1, characterized in that Among them, The network optimization unit also includes: optimizing the model parameters through network pruning and negative prompt calibration methods.

8. The method for detecting the forming characteristics and quality evaluation of titanium alloy forgings based on machine learning according to claim 7, characterized in that Among them, Optimizing the model parameters through network pruning and negative prompt calibration methods includes: The network pruning reduces the computational amount and storage requirements by sparsifying redundant parameters during the training process and introducing a sparse L1 regularization term. The negative prompt calibration is based on the non-repetition assumption of instance masks, calculates the intersection over union between instances, takes the positive sample prompts of the repeated samples with the intersection over union higher than the set threshold as candidate negative sample prompts, randomly selects some as the new negative prompts of the target samples, and re-inputs them into the network to optimize the masks.

9. The method for detecting the forming characteristics and evaluating the quality of titanium alloy forgings based on machine learning according to claim 7, wherein Among them, The method further includes: Preliminarily setting the parameters of the forging process and establishing a multi-physics field coupling model of the forging process by using the finite element analysis method; Performing numerical simulation on the multi-physics field coupling model to analyze the temperature distribution, stress and strain, and metal flow law during the forging process, and generating a large amount of data including temperature field, stress field, flow field, and phase change information; Using machine learning technology to automatically extract key information from the large amount of data, realizing intelligent analysis of the numerical simulation results, and automatically discovering abnormal simulation results or existing optimization spaces; Performing forming characteristic detection, including analyzing the surface quality of the forging by using ultrasonic, X-ray, and infrared thermal imaging technologies, and observing the microstructure of the forging by using a metallurgical microscope to predict the influence of different heat treatment processes on the tissue evolution; Evaluating the performance of the forging by tensile, hardness, and impact tests, and predicting the change trend of the performance according to the test results; According to the results of machine learning analysis, the results of forming characteristic detection, and the results of performance evaluation, optimizing the algorithm of the forging process and adjusting the input parameters; Substituting the adjusted input parameters into the multi-physics field coupling model for a new numerical simulation; Repeating the process from automatically extracting key information from the large amount of data to performing a new numerical simulation, and continuously iterating until the optimal forging process plan is found.

10. A titanium alloy forging forming characteristic detection and quality evaluation system based on machine learning, characterized in that, Including: A data acquisition module for real-time collecting multi-modal information including temperature, pressure, deformation amount, and strain amount during the titanium alloy forging process and performing data preprocessing; A detection model construction module for constructing and training a machine learning model, and inputting the preprocessed data into the trained machine learning model to output the forming characteristic detection results of the titanium alloy forging; Among them, the machine learning model includes: a feature extraction network that uses a convolution-normalization-activation module, a cross-layer connection module, and a fast version of the spatial pyramid pooling module to extract key features from the preprocessed data; a multi-scale feature fusion unit for performing feature fusion on the extracted key features of different scales; a network optimization unit that uses an anchor-free mechanism and a decoupled head architecture and combines a target detection strategy of multiple loss functions to optimize the model; A quality evaluation module for calculating the difference degree between the forming characteristic detection results and the standard data in the quality standard database, determining the quality grade of the titanium alloy forging according to the difference degree, and realizing the quantitative evaluation of the quality of the titanium alloy forging.

Images