Foreign matter monitoring method and device for automobile stamping die and automatic stamping production equipment

Abstract

The present application relates to a kind of automobile stamping die foreign matter monitoring method, device and automatic stamping production equipment.The automobile stamping die foreign matter monitoring method described in the present application includes: obtaining the die area image before sheet metal enters die;And using moving window mechanism to segment its image, obtain the current die small area image;While using a foreign matter image classification model and foreign matter identification model to classify and identify the current die small area image, obtain the classification result and foreign matter confidence of current die small area image;Then judge whether the classification result and foreign matter confidence of current die small area image meet a foreign matter condition: if not, then sheet metal is transported to die for stamping, and after stamping sheet metal is transported to next process;If yes, then prewarning is carried out, and die is cleaned.The automobile stamping die foreign matter monitoring method described in the present application significantly reduces foreign matter false detection rate, and improves foreign matter identification accuracy.

Description

Technical Field

[0001] The present invention relates to the field of automobile manufacturing, and in particular to a foreign matter monitoring method and device for automobile stamping dies and automatic stamping production equipment. Background Art

[0002] During the automobile manufacturing process, production relies on four core processes: stamping, welding, painting, and assembly. Stamping, as the initial shaping process for automobiles, is used to process sheet metal into large-scale metal parts such as the body and chassis. This process uses high pressure to shape sheet metal through a die. This involves combining the die with a stamping machine, which squeezes the sheet metal against the die, shaping it into the shape of the die to create the automotive parts, which are then transported to the next process.

[0003] The stamping process typically involves multiple forming operations, each of which gradually shapes the metal sheet into the desired form. Consequently, when the mold structure is complex or the material deforms during the stamping process, scrap, waste, or residual material may be generated. For this reason, machining shops typically utilize scrap recovery mechanisms during the stamping process, such as automatic conveyor belts, scrap chutes, or high-speed wind devices.

[0004] However, when scraps, waste, or residual materials are irregular in shape or are caused by external factors such as vibration, improper operation, or even other unknown objects that fall, they may become trapped inside the mold. Due to their irregular shape or structure, they are difficult to remove by gravity or airflow, eventually forming foreign objects in the mold. These foreign objects may hinder the normal movement of the sheet metal in the mold, causing scratches, indentations, or irregular deformation of the sheet metal, thereby affecting the quality of the final product. In addition, foreign objects that remain for a long time may cause wear on the mold surface, leading to cracks or breakage in the mold, which in turn may cause production line shutdowns and seriously affect production efficiency.

[0005] Based on this, existing object detection models are used to monitor stamping dies in real time. Once foreign objects are detected, an alert is issued and the production line is paused for cleaning. However, due to the large size of the dies used in the automotive stamping process, the images fed into the object detection model exhibit a "large field of view, small target" characteristic. This can cause the object detection model to mistakenly identify irrelevant details or noise as foreign objects, triggering false alarms and unnecessary production line pauses, severely impacting production efficiency. Therefore, existing technologies in the automotive stamping process struggle to accurately detect foreign objects within the die. Summary of the Invention

[0006] Based on this, the object of the present invention is to provide a foreign matter monitoring method for automobile stamping dies.

[0007] A method for monitoring foreign matter in an automobile stamping die comprises the following steps:

[0008] S1: Acquire the image of the mold area before the sheet enters the mold;

[0009] S2: Use the moving window mechanism to segment the current mold area image to obtain the current mold small area image;

[0010] S3A: Using a foreign body image classification model to classify the current small area image of the mold, obtaining a classification result of the current small area image of the mold;

[0011] S3B: Use a foreign body recognition model to perform fine recognition on the current small area image of the mold to obtain the foreign body confidence level of the current small area image of the mold;

[0012] S4: Determine whether the classification result and foreign body confidence of the current mold small area image meet a foreign body condition: if not, execute step S5A; if so, execute step S5B;

[0013] S5A: Determine whether the current mold small area image is the last small area image: if not, execute step S2 to select the next small area image; if yes, execute step S6;

[0014] S5B: issuing an early warning based on the foreign body confidence level of the current area image and cleaning the mold; wherein, after the mold cleaning is completed in step S5B, step S6 is executed to continue the stamping operation;

[0015] S6: conveying the sheet material to the mold for stamping, and conveying the stamped sheet material to the next process; wherein, after the sheet material is stamped through the step S6, if the current mold is in an empty state, execute step S1.

[0016] Compared with the existing technology, the foreign body monitoring method for automobile stamping dies described in the present invention divides the mold area into small areas through a moving window mechanism, which can effectively avoid missing small foreign bodies due to the presence of small targets in a large field of view and reduce the problem of false detection caused by environmental interference; at the same time, the foreign body image classification model and the foreign body recognition model work together to ensure higher detection accuracy and minimize the risk of false detection and missed detection, thereby improving the production efficiency of automobile stamping and the safety of mold use while ensuring high-precision foreign body recognition.

[0017] Furthermore, the foreign body image classification model includes a feature extraction module and a foreign body image classification module;

[0018] The specific calculation of the feature extraction module is as follows:

[0019] Fi =Pool(ReLU(Conv 3×3 (F i-1 )))

[0020] F1=Pool(ReLU(Conv 3×3 (I small )))

[0021] Where, F i F represents the features of the small area image of the mold at layer i; i-1 Represents the output of the previous layer; I small The small area image of the mold extracted from the current window; Conv 3×3 Represents a convolution layer with a convolution kernel of 3×3; ReLU represents an activation function; Pool represents a pooling layer;

[0022] The specific calculation of the foreign body image classification module is as follows:

[0023] P result =σ(H n )

[0024] Where, P result is the classification result of the current mold small area image, that is, the probability value of whether the foreign matter exists; σ is the sigmoid function; H n It is expressed as the output of the nth fully connected layer, and its specific calculation is as follows:

[0025] H n =ReLU(FC n (H n-1 )))H1=ReLU(FC1(F i ))

[0026] Where H n Represents the output of the nth fully connected layer, whose nth fully connected layer FC n The number of neurons n The specific calculation is expressed as:

[0027]

[0028] Where n∈[2, N], the calculation of N is expressed as:

[0029]

[0030] Among them, neuron1 is the number of neurons in the first fully connected layer.

[0031] The application can effectively extract the detailed features, i.e. local features, in the small area image of the mold through the hierarchical convolution and pooling operation of the feature extraction module, thereby improving the recognition ability of foreign matters under complex scenes or environmental interference; at the same time, the number of neurons is gradually reduced layer by layer in the full connection layer of the foreign matter image classification module, so as to ensure that the model can more accurately judge the foreign matter situation in the small area image of the mold, aiming to reduce the problems of false detection and missed detection; accordingly, the application can more accurately detect foreign matters by combining the feature extraction module and the foreign matter classification module, so as to adapt to more complex working conditions, thereby reducing the false positive rate and the missed detection rate of foreign matter detection, and further improving the efficiency and safety of the entire automobile stamping production line.

[0032] Further, the foreign matter recognition model comprises a backbone network, a neck network and a head network.

[0033] The backbone network is used for multi-level feature extraction on the input small area image of the mold and outputs feature maps of several scales.

[0034] The specific calculation of the preliminary feature extraction is as follows:

[0035] Feature1=Focus(I small )

[0036] In the formula, the Focus module is used for slicing the input image to obtain the initial feature map Feature1.

[0037] For the i-th level feature extraction, the specific calculation is as follows:

[0038] Feature i =C3(CONV(Feature i-1 ) 2 ), i∈[2,n-1]

[0039] In the formula, CONV represents a convolution module, and the specific structure is as follows:

[0040] SILU(BN(Conv2d(F x )))

[0041] In the formula, F x is the input image, Conv2d is a two-dimensional convolution, BN is batch normalization, and SILU is an activation function.

[0042] C3 represents a residual block, and the specific structure is as follows:

[0043] CONV(Concat(Add(CONV(F x ) 3 , CONV(F x )), CONV(Fx ) 2 ))

[0044] In the formula, Add represents element-by-element addition, and Concat represents concatenation operation;

[0045] The specific representation of feature extraction at the final level is as follows:

[0046] Feature n =C3(SPP(CONV(Feature i-1 )))

[0047] In the formula, Feature n Represents the feature map of the final scale; SPP is spatial pyramid pooling;

[0048] The neck network is used to perform feature fusion on feature maps of several scales to obtain fused feature maps of several scales, which are specifically expressed as follows:

[0049] F fusion ={FPN({Feature i |i∈[2,n]})}

[0050] Where, F fusion Represents a fusion feature atlas of several scales; FPN is a feature pyramid;

[0051] The head network is used to predict the fused feature maps of several scales, generate several prediction boxes, and filter the several prediction boxes using non-maximum suppression to obtain the foreign body confidence of the current small area image of the mold.

[0052] The present invention introduces residual blocks that perform element-by-element addition and concatenation operations, thereby ensuring that the model can effectively fuse feature information at multiple scales, thereby improving the ability to perceive mold image details. At the same time, it enables the model to maintain high detection accuracy even when there is noise or detail loss in small area images of the mold, reducing the occurrence of missed detections and false detections.

[0053] In one embodiment, the foreign matter condition is specifically expressed as follows:

[0054] The specific expression of the foreign matter condition is as follows:

[0055]

[0056] Wherein, isExist represents the judgment result of whether there is a foreign object in the current mold small area image. When isExist indicates that there is no foreign object, step S5A is executed; when isExist indicates that there is foreign object, step S5B is executed; P result is the classification result of the current mold small area image; Ptarget is the foreign body confidence of the current small area image of the mold; max is the maximum confidence threshold.

[0057] The present invention uses static fusion conditions to ensure that the system can accurately judge foreign objects based on a fixed maximum execution threshold, thereby avoiding false detections and ensuring the simplicity, stability and efficiency of system decisions. It aims to reduce the risk of production stagnation in high-speed production environments and ensure efficient and safe operation of the production line.

[0058] In one embodiment, the foreign matter condition is specifically expressed as follows:

[0059]

[0060] Wherein, isExist represents the judgment result of whether there is a foreign object in the current mold small area image. When isExist indicates that there is no foreign object, step S5A is executed; when isExist indicates that there is foreign object, step S5B is executed; P result is the classification result of the current mold small area image; P target is the foreign body confidence of the current small area image of the mold; max and th min are the maximum confidence threshold and the minimum confidence threshold respectively; cf A and cf B They are respectively the classification result of the current mold small area image and the dynamic confidence threshold corresponding to the foreign body confidence. The specific calculation is as follows:

[0061] cf A =w A ×cf MAX +w B ×cf MIN +bias

[0062] cf B =w B ×cf MAX +w A ×cf MIN +bias

[0063] Where, cf MAX and cf MIN Respectively expressed as P result and P target The maximum and minimum confidence between them; bias is expressed as offset; w A and w B They are the dynamic weights of the classification results and foreign body confidence, and the specific calculation is as follows:

[0064]

[0065] Where Δp is expressed as P result and P target The absolute value difference between them.

[0066] Accordingly, the dynamic confidence threshold of the present invention dynamically adjusts the weight according to the difference between the classification result and the foreign body confidence, and combines the offset to achieve a flexible response to different confidence combinations, thereby ensuring that when the confidence of one model is higher, the requirements for the other model can be appropriately relaxed, thereby reducing the missed detection rate while ensuring high detection accuracy, and significantly improving the production quality of the automobile stamping production line.

[0067] Furthermore, the specific calculation of the offset is expressed as follows:

[0068]

[0069] Where α is the tolerance coefficient.

[0070] The present invention further introduces the calculation of the offset. When there is a high-confidence result, a certain degree of tolerance is given; when there is a low-confidence result, a more stringent judgment is given; if neither the high-confidence nor the low-confidence conditions are met, the default judgment standard is maintained, thereby significantly optimizing the balance of foreign body detection. It effectively reduces production stagnation caused by the misjudgment of a single high-confidence result, and reduces quality problems caused by the missed detection of a single low-confidence result, thereby improving the efficiency and reliability of the entire automobile stamping production line.

[0071] A foreign body monitoring device for an automobile stamping die, comprising a die image acquisition unit, an image region segmentation unit, a foreign body image classification unit, a foreign body recognition unit, a foreign body judgment unit, an iterative monitoring and judgment unit, a foreign body early warning unit, and a die stamping control unit;

[0072] The mold image acquisition unit is used to acquire an image of the mold area before the sheet material enters the mold;

[0073] The image region segmentation unit is used to segment the current mold region image using a moving window mechanism to obtain the current mold small region image;

[0074] The foreign body image classification unit is used to classify the current mold small area image using a foreign body image classification model to obtain a classification result of the current mold small area image;

[0075] The foreign body recognition unit is used to use a foreign body recognition model to perform fine recognition on the current small area image of the mold to obtain the foreign body confidence level of the current small area image of the mold;

[0076] The foreign body judgment unit is used to judge whether the classification result and foreign body confidence of the current small area image of the mold meet a foreign body condition: if not, calling the iterative monitoring judgment unit; if so, calling the foreign body early warning unit;

[0077] The iterative monitoring and judging unit is used to judge whether the current mold small area image is the last small area image; if not, the image area segmentation unit is called to select the next small area image; if so, the mold stamping control unit is called;

[0078] The foreign body warning unit is used to issue a warning based on the foreign body confidence level of the current area image and clean the mold; wherein, after the mold is cleaned by the foreign body warning unit, the mold stamping control unit is called to continue the stamping operation;

[0079] The mold stamping control unit is used to transport the sheet material to the mold for stamping, and transport the stamped sheet material to the next process; wherein, after the sheet material is stamped by the mold stamping control unit, if the current mold is in an empty state, the mold image acquisition unit is called.

[0080] Furthermore, the foreign body image classification model includes a feature extraction module and a foreign body image classification module;

[0081] The specific calculation of the feature extraction module is as follows:

[0082] F i =Pool(ReLU(Conv 3×3 (F i-1 )))

[0083] F1=Pool(ReLU(Conv 3×3 (I small )))

[0084] Where, F i F represents the features of the small area image of the mold at layer i; i-1 Represents the output of the previous layer; I small The small area image of the mold extracted from the current window; Conv 3×3 Represents a convolution layer with a convolution kernel of 3×3; ReLU represents an activation function; Pool represents a pooling layer;

[0085] The specific calculation of the foreign body image classification module is as follows:

[0086] P result =σ(H n )

[0087] Where, P resultis the classification result of the current mold small area image, that is, the probability value of whether the foreign matter exists; σ is the sigmoid function; H n It is expressed as the output of the nth fully connected layer, and its specific calculation is as follows:

[0088] H n =ReLU(FC n (H n-1 )))

[0089] H1=ReLU(FC1(F i ))

[0090] Where H n Represents the output of the nth fully connected layer, whose nth fully connected layer FC n The number of neurons n The specific calculation is expressed as:

[0091]

[0092] Where n∈[2, N], the calculation of N is expressed as:

[0093]

[0094] Among them, neuron1 is the number of neurons in the first fully connected layer;

[0095] The foreign body recognition model includes a backbone network, a neck network and a head network;

[0096] The backbone network is used to perform multi-level feature extraction on the input small area image of the mold and output feature maps of several scales;

[0097] Among them, the specific calculation of preliminary feature extraction is as follows:

[0098] Feature1=Focus(I small )

[0099] In the formula, the Focus module is used to slice the input image to obtain the initial feature map Feature1;

[0100] For the feature extraction of level i, the specific calculation is expressed as:

[0101] Feature i =C3(CONV(Feature i-1 ) 2 ), i∈[2,n-1]

[0102] In the formula, CONV represents the convolution module, and its specific structure is:

[0103] SILU(BN(Conv2d(F x )))

[0104] Where, F x is the input image, Conv2d is the two-dimensional convolution, BN is the batch normalization and SILU is the activation function;

[0105] C3 represents the residual block, and its specific structure is:

[0106] CONV(Concat(Add(CONV(F x ) 3 ,CONV(F x )), CONV(F x ) 2 ))

[0107] In the formula, Add represents element-by-element addition, and Concat represents concatenation operation;

[0108] The specific representation of feature extraction at the final level is as follows:

[0109] Feature n =C3(SPP(CONV(Feature i-1 )))

[0110] In the formula, Feature n Represents the feature map of the final scale; SPP is spatial pyramid pooling;

[0111] The neck network is used to perform feature fusion on feature maps of several scales to obtain fused feature maps of several scales, which are specifically expressed as follows:

[0112] F fusion ={FPN({Feature i |i∈[2,n]})}

[0113] Where, F fusion Represents a fusion feature atlas of several scales; FPN is a feature pyramid;

[0114] The head network is used to predict the fused feature maps of several scales, generate several prediction boxes, and filter the several prediction boxes using non-maximum suppression to obtain the foreign body confidence of the current small area image of the mold.

[0115] Furthermore, the foreign matter condition is specifically expressed as follows:

[0116]

[0117] Wherein, isExist represents the judgment result of whether there is a foreign object in the current mold small area image. When isExist indicates that there is no foreign object, step S5A is executed; when isExist indicates that there is foreign object, step S5B is executed; P result is the classification result of the current mold small area image; P target is the foreign body confidence of the current small area image of the mold; max and th min are the maximum confidence threshold and the minimum confidence threshold respectively; cf A and cf B They are respectively the classification result of the current mold small area image and the dynamic confidence threshold corresponding to the foreign body confidence. The specific calculation is as follows:

[0118] cf A =w A ×cf MAX +W B ×cf MIN +bias

[0119] cf B =w B ×cf MAX +w A ×cf MIN +bias

[0120] Where, cf MAX and cf MIN Respectively expressed as P result and P target The maximum and minimum confidence between them; bias is expressed as offset; w A and w B They are the dynamic weights of the classification results and foreign body confidence, and the specific calculation is as follows:

[0121]

[0122] Where ΔP is expressed as P result and P target The absolute value difference between

[0123] The specific calculation of the offset is as follows:

[0124]

[0125] Where α is the tolerance coefficient.

[0126] An automatic stamping production equipment for automobile parts, comprising a conveyor belt, a stamping machine, a camera device and a foreign body monitoring device;

[0127] The conveyor belt is used to convey the sheet material to the mold and convey the sheet material after being punched by the punching machine to the next process;

[0128] The punching machine includes an upper frame and a workbench, and is provided with a die;

[0129] The die comprises an upper die and a lower die, wherein the upper die is fixedly mounted on the upper frame of the punching machine and is provided with a punch; the lower die is provided with a forming cavity matching the upper die and is fixed to the workbench of the punching machine; the punching machine is used to apply pressure to the sheet material through the upper die and the lower die to form the sheet material into a component of a preset shape;

[0130] The camera device is used to photograph the idle mold, obtain an image of the mold area before the sheet metal enters the mold, and transmit the image to the foreign matter detection device;

[0131] The foreign body monitoring device is used to identify foreign bodies in the input image and obtain a foreign body identification result. If the foreign body identification result indicates the presence of foreign bodies, the production equipment is suspended and an early warning is issued, waiting for the mold to be cleaned; if the foreign body identification result indicates the absence of foreign bodies, the mold is monitored continuously.

[0132] Wherein, the foreign matter monitoring device is the foreign matter monitoring device for the automobile stamping die described above.

[0133] For better understanding and implementation, the present invention is described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0134] Figure 1 This is a simplified structural diagram of the foreign matter monitoring device for automobile stamping dies according to the present invention;

[0135] Figure 2 This is a simplified flow chart of the method for monitoring foreign matter in automobile stamping dies according to the present invention;

[0136] Figure 3 This is a simplified structural diagram of the foreign body image classification model;

[0137] Figure 4 This is a simplified structural diagram of the foreign body recognition model described in the present invention. DETAILED DESCRIPTION

[0138] In order to solve the problem of difficulty in accurately monitoring foreign matter in the existing technology in the automobile stamping process, the present invention obtains the regional image before the sheet metal enters the mold, and adopts a moving window mechanism to segment the current regional image to obtain the current mold small regional image of the mold; and simultaneously adopts a foreign matter image classification model and a foreign matter recognition model to classify and finely recognize the current mold small regional image to obtain the classification result and foreign matter confidence of the current mold small regional image; and judges whether the classification result and foreign matter confidence of the current mold small regional image meet a foreign matter condition: if so, an early warning is issued according to the foreign matter confidence of the current regional image, and the mold is cleaned; if not, it is judged whether the current mold small regional image is the last small regional image: if not, the next small regional image is selected for classification and fine recognition; if so, the sheet metal is transported to the mold for stamping, and the stamped sheet metal is transported to the next process. Based on this, the present invention significantly reduces the false detection rate of foreign objects and improves the accuracy of foreign object identification by performing graded processing on small area images separately and combining the classification results with the confidence judgment of finely identified foreign objects, thereby reducing production stagnation caused by mold jamming, wear and other problems, and improving the stability and continuous production capacity of the entire production line.

[0139] Based on the above-mentioned design of the foreign body identification scheme, the present invention proposes a foreign body monitoring method for automobile stamping dies, and based on this method, proposes a foreign body monitoring device for automobile stamping dies.

[0140] Please also see Figure 1 and Figure 2 , Figure 1 This is a simplified structural diagram of the foreign body monitoring device for automobile stamping dies according to the present invention. Figure 2 The figure is a simplified flow chart of the foreign matter monitoring method for automobile stamping dies according to the present invention.

[0141] The foreign body monitoring device for the automobile stamping die includes a die image acquisition unit 1, an image area segmentation unit 2, a foreign body image classification unit 3A, a foreign body identification unit 3B, a foreign body judgment unit 4, an iterative monitoring and judgment unit 5A, a foreign body early warning unit 5B and a die stamping control unit 6.

[0142] The mold image acquisition unit 1 is used to execute step S1: acquiring an image of the mold area before the sheet material enters the mold.

[0143] Specifically, a photographing device is used to photograph the mold to obtain an image of the mold area before the sheet material enters the mold.

[0144] The sheet material is used in the automobile stamping process to manufacture the car body, chassis, etc., and exists in the form of large-sized metal sheets before being cut, stamped and formed;

[0145] The mold includes an upper mold and a lower mold, which work together to complete the forming of the sheet material; the upper mold is installed on the slider or upper frame of the punching machine, and is usually designed with a forming cavity, a punch or other forming tools; the lower mold is fixed on the workbench of the punching machine, and usually includes a forming cavity for accommodating the sheet material and assisting the relative movement between the sheet material and the upper mold. The lower mold is usually also provided with a discharge device or a waste channel for removing the waste or sheet material after stamping; accordingly, the upper mold and the lower mold are used to apply pressure to the sheet material through the punching machine to deform it into the required component shape to complete the forming of the sheet material.

[0146] The present invention obtains images of the empty mold and performs real-time foreign matter monitoring on it to ensure that there is no foreign matter in the mold before each sheet enters the mold, thereby effectively avoiding the impact of foreign matter on the stamping process of the sheet, and preventing scratches, indentations, deformation of the sheet or the production of defective parts due to foreign matter inside the mold, thereby improving production efficiency and product quality.

[0147] The image region segmentation unit 2 is configured to execute step S2: segmenting the current mold region image using a moving window mechanism to obtain a current mold small region image.

[0148] Specifically, the moving window mechanism is specifically expressed as follows:

[0149] I small ={I[x,y]|x∈[x1,x1+w],y∈[y1,y1+h]}

[0150] Where, I small Represents the small mold area image extracted by the current window, whose size is w×h; I represents the current mold area image, that is, the image of the empty mold; x1 and y1 represent the starting coordinates of the current window in the image; w and h represent the width and height of the window respectively;

[0151] Based on this, a small mold area image I is extracted by sliding the window in the image. small , thereby extracting local areas from the entire mold image for foreign body detection, preventing global features from being affected by environmental interference or the large overall size of the mold, thereby ensuring more accurate foreign body detection and effectively reducing the risk of false detection or missed detection.

[0152] See also Figure 3 , Figure 3 This is a simplified structural diagram of the foreign body image classification model.

[0153] The foreign body image classification unit 3A is used to perform step S3A: classify the current small area image of the mold using a foreign body image classification model to obtain a classification result of the current small area image of the mold.

[0154] Specifically, the foreign body image classification model includes a feature extraction module and a foreign body image classification module. The feature extraction module is used to extract the features of the input small area image of the mold. The specific calculation expression is as follows:

[0155] F i =Pool(ReLU(Conv 3×3 (F i-1 )))

[0156] F1=Pool(ReLU(Conv 3×3 (I small )))

[0157] Where, F i Represents the features of the small area image of the mold at layer i, i∈[1,15]; F i-1 Represents the output of the previous layer, that is, the features extracted by the i-1 layer; Conv 3×3 ReLU represents a convolutional layer with a 3×3 convolution kernel, which is used to extract features from the input image or the feature map of the previous layer. ReLU represents an activation function, which is used to introduce nonlinear transformations and set the negative values ​​output by the convolutional layer to zero, thereby enhancing the significance of positive features. Pool represents a pooling layer, which is used to reduce the spatial dimension of the feature map, reduce the amount of computation, and retain the most important feature information.

[0158] The foreign body image classification module is used to classify the features of the input small area image of the mold to obtain the classification result of the current small area image of the mold. The specific calculation is as follows:

[0159] P result =σ(H n )

[0160] Where, P result is the classification result of the current mold small area image, that is, the probability value of whether foreign matter exists; σ is the sigmoid function, which is used to map the output to a probability value between 0 and 1, indicating the confidence level of the presence of foreign matter in the input feature; H n It is expressed as the output of the nth fully connected layer, and its specific calculation is as follows:

[0161] H n =ReLU(FC n (H n-1 )))

[0162] H1=ReLU(FC1(F i )), i=15

[0163] Where H n Represents the output of the nth fully connected layer, whose nth fully connected layer FC n The number of neuronsn The specific calculation is expressed as: And n∈[2,N], the calculation of N is expressed as:

[0164]

[0165] Among them, neuron1 is the number of neurons in the first fully connected layer, that is, the total number of neurons in the input layer, which is set to 2048.

[0166] Based on this, the present invention adopts a convolutional neural network as a feature extraction module to extract the features of the small area image of the mold extracted by the sliding window mechanism, and uses a fully connected neural network, that is, the foreign body image classification module, to refine the learning layer by layer to ensure high-precision separation, and uses the sigmoid function to output the confidence of the presence of foreign matter to achieve high-precision binary classification results, so as to reduce false detection and thereby improve the accuracy of foreign body detection in the mold during the automobile stamping process.

[0167] In order to increase the classification accuracy of the foreign body classification model, the present invention trains the foreign body classification model in the following manner, which specifically includes:

[0168] Data processing stage: Collect and annotate a large number of small area images of the mold, and ensure that all images are accurately labeled with foreign body labels, which are used to indicate whether foreign body exists;

[0169] Next, the small area image of the mold is combined with the corresponding label to form a classification data set. In order to improve the model's ability to classify foreign objects under different environments and conditions, a data enhancement method can be used to enhance the current classification data set, such as rotation, scaling, or cropping. The present invention does not specifically limit the data enhancement method.

[0170] Finally, the classification dataset is divided into training set, test set, and validation set to ensure the effectiveness of model training and the comprehensiveness of evaluation.

[0171] Training and optimization phase: Preprocess the small area images of the mold in the training set and input them into the current foreign body classification model to obtain the foreign body classification confidence;

[0172] Using binary cross entropy as the loss function, the loss value between the foreign object classification confidence and the corresponding label is calculated, and the gradient is calculated through the backpropagation algorithm. In combination with the optimizer (such as Adam), the current foreign object classification model weight is updated.

[0173] Next, the test set is fed into the model, and the accuracy of the output is calculated. If the accuracy reaches a preset threshold or the training round reaches the end condition, the model proceeds to the next stage; otherwise, the training and optimization stages are repeated to further improve the model performance.

[0174] Model Evaluation and Validation Phase: The validation set is fed into the trained object classification model. Metrics such as accuracy, recall, and F1 score are calculated and analyzed to comprehensively assess the model's classification accuracy and robustness. If the validation set's evaluation metrics do not meet the requirements, hyperparameters are adjusted and the training and optimization phase is repeated. If the validation set's metrics meet the requirements, a highly accurate object classification model is confirmed, completing the training process.

[0175] Based on this, the present invention ensures that the foreign body classification model has high accuracy and strong generalization ability through data processing, training optimization and model verification processes, significantly reduces the false detection and missed detection rates, and thus improves the accuracy and production efficiency of foreign body detection in the mold during the automobile stamping process.

[0176] See also Figure 4 , Figure 4 A simplified structural diagram of the foreign body recognition model described in the present invention

[0177] The foreign body recognition unit 3B is used to execute step S3B: using a foreign body recognition model to perform fine recognition on the current small area image of the mold to obtain the foreign body confidence of the current small area image of the mold.

[0178] Specifically, the foreign body recognition model includes a backbone network, a neck network, and a head network. The backbone network is used to perform multi-level feature extraction on the input small area image of the mold and output feature maps of several scales. The specific calculation of the preliminary feature extraction in the multi-level feature extraction is expressed as follows:

[0179] Feature1=Focus(I small )

[0180] In the formula, the Focus module is used to slice the input image. e ), that is, the input image is downsampled to concentrate the information into the channel space and obtain the initial feature map Feature1; for the feature extraction of the i-th level, the specific calculation is expressed as:

[0181] Feature i =C3(CONV(Feature i-1 ) 2 ), i∈[2,n-1]

[0182] In the formula, CONV represents the convolution module, and its specific structure is:

[0183] SILU(BN(Conv2d(F x )))

[0184] Where, F xis the input image, Conv2d is the two-dimensional convolution, BN is the batch normalization and SILU is the activation function;

[0185] C3 represents the residual block, and its specific structure is:

[0186] CONV(Concat(Add(CONV(F x ) 3 ,CONV(F x )), CONV(F x ) 2 ))

[0187] Wherein, Add represents element-by-element addition, and Concat represents concatenation. In the present invention, the stride of the convolution operation of the residual block is 1, and the convolution kernel size is 3×3.

[0188] The specific representation of feature extraction at the final level is as follows:

[0189] Feature n =C3(SPP(CONV(Feature i-1 )))

[0190] In the formula, Feature n Represents the feature map of the final scale, n is set to 5 by default; SPP is spatial pyramid pooling, which is used to enhance the ability to extract multi-scale features;

[0191] The neck network is used to perform feature fusion on feature maps of several scales to obtain fused feature maps of several scales, which are specifically expressed as follows:

[0192] F fusion ={FPN({Feature i |i∈[2,n]})}

[0193] Where, F fusion Represents a fusion feature atlas of several scales; FPN is a feature pyramid used to fuse features of multiple scales.

[0194] The head network is used to predict the fused feature maps of several scales, generate several prediction boxes, and filter the several prediction boxes using non-maximum suppression to obtain the foreign body confidence of the current small area image of the mold.

[0195] Among them, the predicted box includes the location of the target, that is, the bounding box coordinates, and the foreign object confidence.

[0196] Accordingly, the present invention introduces element-by-element addition and splicing operations in the residual block (C3). The element-by-element addition ensures the continuity and gradient flow of feature information, while the splicing operation helps to introduce more feature dimensions, thereby enhancing the model's recognition ability for fine-grained features (such as small objects or foreign matter). In addition, the present invention sets the convolution stride in the residual block (C3) to 1 to ensure the complete retention of detail information, thereby improving spatial resolution and reducing information loss in the detection of small objects or foreign matter.

[0197] In order to increase the detection accuracy of the foreign body recognition model, the invention trains the foreign body recognition model in the following manner, which specifically includes:

[0198] Data processing: Collect and annotate a large number of small area images of the mold to ensure that each image accurately identifies the location and shape of the foreign object and the corresponding label, which is used to indicate whether the foreign object exists;

[0199] Based on the location and shape of foreign matter in the small area image of the mold, a target frame is drawn, and the target frame is associated with the label to construct a target detection dataset;

[0200] In order to improve the model's ability to classify foreign objects under different environments and conditions, a data enhancement method may be used to enhance the current target detection dataset, such as rotation, scaling, or cropping. The present invention does not specifically limit the data enhancement method.

[0201] Finally, the object detection dataset is divided into training set, test set and validation set to ensure the effectiveness of model training and the comprehensiveness of evaluation.

[0202] Training and optimization phase: Preprocess the small area images of the mold in the training set and input them into the current foreign object recognition model to obtain the foreign object recognition confidence;

[0203] The binary cross entropy, IOU loss of the target box, and L1 loss are used as the combined loss function to calculate the loss value between the foreign object recognition confidence and the corresponding label and target box. The gradient is calculated through the backpropagation algorithm and combined with the optimizer (such as Adam) to update the current foreign object classification model weights.

[0204] Next, the test set is fed into the model, and the accuracy of the output is calculated. If the accuracy reaches a preset threshold or the training round reaches the end condition, the model proceeds to the next stage; otherwise, the training and optimization stages are repeated to further improve the model performance.

[0205] Model Evaluation and Validation Phase: The validation set is fed into the trained object recognition model. Metrics such as accuracy, recall, and F1 score are calculated and analyzed to comprehensively assess the model's classification accuracy and robustness. If the validation set's evaluation metrics do not meet the requirements, hyperparameters are adjusted and the training and optimization phase is repeated. If the validation set's metrics meet the requirements, a highly accurate object recognition model is confirmed, completing the training process.

[0206] During model training, the introduction of element-by-element addition and concatenation operations significantly enhances the model's learning capabilities. Through direct paths to lower-level features, they ensure that features at different levels are fully learned, thus avoiding performance degradation caused by poor information flow. Furthermore, the concatenation operation establishes richer connections between different levels and scales, enhancing the model's performance in handling foreign object detection tasks at multiple scales. This allows the model to automatically capture features at different scales during training, improving its expressiveness and generalization capabilities.

[0207] The foreign body judgment unit 4 is used to execute step S4: judging whether the classification result and foreign body confidence of the current small area image of the mold meet a foreign body condition: if not, calling the iterative monitoring judgment unit 5A; if so, calling the foreign body early warning unit 5B.

[0208] Specifically, the foreign matter condition is specifically expressed as follows:

[0209]

[0210] Wherein, isExist is used to indicate the judgment result of whether there is a foreign body in the current mold small area image. When isExist indicates that there is no foreign body, the iterative monitoring judgment unit 5A is called; when isExist indicates that there is a foreign body, the foreign body early warning unit 5B is called; result is the classification result of the current mold small area image; P target is the foreign body confidence of the current small area image of the mold; max is the maximum confidence threshold, which is set to 0.8 by default.

[0211] Accordingly, the present invention adopts a fixed threshold approach, thereby minimizing the misjudgment rate while ensuring system stability, ensuring that the production line will not be frequently paused due to false detections, thereby improving production efficiency and system reliability.

[0212] In another embodiment, the foreign matter condition is specifically expressed as follows:

[0213]

[0214] Wherein, isExist is used to indicate the judgment result of whether there is a foreign body in the current mold small area image. When isExist indicates that there is no foreign body, the iterative monitoring judgment unit 5A is called; when isExist indicates that there is a foreign body, the foreign body early warning unit 5B is called; result is the classification result of the current mold small area image; P target is the foreign body confidence of the current small area image of the mold; max and th min are the maximum confidence threshold and the minimum confidence threshold, respectively, and their default settings are 0.8 and 0.6 respectively; cf A and cf B They are respectively the classification result of the current mold small area image and the dynamic confidence threshold corresponding to the foreign body confidence. The specific calculation is as follows:

[0215] cf A =w A ×cf MAX +w B ×cf MσN +bias

[0216] cf B =w B ×cf MAx +w A ×cf MIN +bias

[0217] Where, cf MAX and cf MIN Respectively expressed as P result and P target The maximum and minimum confidence between A and w B They are the dynamic weights of the classification results and foreign body confidence, respectively. The specific calculation is as follows:

[0218]

[0219] Where ΔP is expressed as P result and P target The absolute value difference between them, namely |P target -P result |; Its bias is expressed as offset, and its specific calculation is as follows:

[0220]

[0221] Where α is the tolerance coefficient, and its default value is 0.05.

[0222] The present invention uses a fixed maximum confidence threshold th in the foreign body condition maxand the minimum confidence threshold th min To ensure that the presence of foreign matter can be accurately judged under clear high confidence or low confidence conditions; then, by dynamically adjusting the confidence threshold, the result and P target When there is a high confidence level in the target, the tolerance for the close confidence level is balanced. The tolerance coefficient of the offset in the dynamic confidence threshold can increase the tolerance for another target with lower confidence level when there is a high confidence level, thereby avoiding missed detection due to overly strict judgment.

[0223] Accordingly, the present invention ensures a certain degree of tolerant judgment in high-confidence situations through flexible adjustment of dynamic confidence thresholds, while maintaining strictness in low-confidence situations, significantly reducing the false positive rate and missed detection rate.

[0224] The iterative monitoring and judgment unit 5A is used to execute step S5A: determine whether the current mold small area image is the last small area image: if not, call the image area segmentation unit 2 to select the next small area image; if so, call the mold stamping control unit 6.

[0225] Based on this, the detection status of each small area image is iteratively detected, and whether the stamping task needs to be performed is determined based on whether all small area images have been detected, thereby avoiding missing any potential foreign object detection area.

[0226] The foreign body early warning unit 5B is used to execute step S5B: issuing an early warning according to the foreign body confidence level of the current area image and cleaning the mold.

[0227] Specifically, based on the foreign body confidence information in the prediction box annotation of the foreign body confidence of the current area image, that is, the foreign body position and size, an early warning signal is issued, so that the foreign body on the mold can be processed through cleaning equipment or other means.

[0228] When the cleaning is completed, the mold stamping control unit 6 is called to continue the stamping operation.

[0229] When foreign matter is detected, a timely warning is issued to the operator, preventing it from adversely affecting subsequent production processes or product quality. Precise cleaning instructions are also provided, improving mold management efficiency and production safety. Furthermore, stamping operations are quickly resumed after cleaning, avoiding long pauses in the production process, maintaining production rhythm, and improving production efficiency.

[0230] It is understandable that when the early warning signal is issued, the conveyor belt can be suspended from conveying the sheet material through a control system such as a PLC, thereby providing a certain amount of time for the operator to clean up foreign objects, thereby ensuring the safety of the production line during the cleaning of foreign objects.

[0231] The die stamping control unit 6 is configured to execute step S6: conveying the sheet material to the die for stamping, and conveying the stamped sheet material to the next process.

[0232] Specifically, by accurately conveying the sheet material to the lower die and sending a control signal to the control system, the control system, such as PLC, controls the punching machine to apply pressure to the upper die of the mold, thereby punching the sheet material to achieve the forming of the sheet material and obtain a part of a predetermined shape.

[0233] Then, the parts with predetermined shapes are transported to the next process for subsequent processing.

[0234] Among them, when the sheet metal stamping is completed and the stamped parts are transported to the next process, if the current mold is in an empty state, that is, there is no sheet metal remaining, the mold image acquisition unit 1 is called.

[0235] Compared with the existing technology, the present invention divides the mold area image into multiple small areas by adopting a moving window mechanism, thereby effectively avoiding the problem of small foreign objects being missed in a large field of view or false detection caused by environmental interference; at the same time, the collaborative analysis of the foreign object image classification model and the foreign object recognition model ensures the high accuracy of the foreign object detection results and significantly reduces the risk of false detection and missed detection; further, by introducing a dynamic confidence threshold adjustment mechanism, which combines the calculation of dynamic weights and offsets, it gives appropriate tolerance to lower confidence detection results under high confidence conditions, thereby enhancing the system's flexible response capabilities to different confidence combinations. Accordingly, the present invention significantly improves the efficiency and safety of automobile stamping production lines by combining multiple models with a dynamic threshold mechanism, ensuring the stable output of high-quality parts.

[0236] Based on the same inventive concept, the present application also provides an electronic device, which can be a terminal device such as a server, desktop computing device, or mobile computing device (e.g., laptop computing device, handheld computing device, tablet computer, netbook, etc.). The device includes one or more processors and a memory, wherein the processor is used to execute a program to implement the foreign matter monitoring method for an automotive stamping die according to an embodiment of the present invention; and the memory is used to store a computer program executable by the processor.

[0237] Based on the same inventive concept, the present application also provides a computer-readable storage medium, corresponding to the embodiment of the aforementioned method for monitoring foreign matter in an automobile stamping die. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the method for monitoring foreign matter in an automobile stamping die recorded in any of the aforementioned embodiments.

[0238] The present application may take the form of a computer program product implemented on one or more storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-usable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. The information can be computer-readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0239] The above-described embodiments merely represent several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, and the present invention is intended to encompass such modifications and variations.

Claims

1. A foreign body monitoring method for automobile stamping dies, characterized in that: The following steps are involved: S1: Acquire the image of the mold area before the sheet enters the mold; S2: Use the moving window mechanism to segment the current mold area image to obtain the current mold small area image; S3A: Using a foreign body image classification model to classify the current small area image of the mold, obtaining a classification result of the current small area image of the mold; S3B: Use a foreign body recognition model to perform fine recognition on the current small area image of the mold to obtain the foreign body confidence level of the current small area image of the mold; S4: Determine whether the classification result and foreign body confidence of the current mold small area image meet a foreign body condition: if not, execute step S5A; if so, execute step S5B; wherein the foreign body condition is specifically expressed as follows: Wherein, isExist represents the judgment result of whether there is a foreign object in the current mold small area image. When isExist indicates that there is no foreign object, step S5A is executed; when isExist indicates that there is foreign object, step S5B is executed; P result is the classification result of the current mold small area image; P target is the foreign body confidence of the current small area image of the mold; max and th min are the maximum confidence threshold and the minimum confidence threshold respectively; cf A and cf B They are respectively the classification result of the current mold small area image and the dynamic confidence threshold corresponding to the foreign body confidence. The specific calculation is as follows: cf A =w A ×cf MAX +w B ×cf MIN +bias cf B =w B ×cf MAX +w A ×cf MIN +bias Where, cf MAX and cf MIN Respectively expressed as P result and P target The maximum and minimum confidence between them; bias is expressed as offset; w A and w B They are the dynamic weights of the classification results and foreign body confidence, and the specific calculation is as follows: Where ΔP is expressed as P result and P target The absolute value difference between S5A: Determine whether the current mold small area image is the last small area image: if not, execute step S2 to select the next small area image; if yes, execute step S6; S5B: issuing an early warning based on the foreign body confidence level of the current area image and cleaning the mold; wherein, after the mold cleaning is completed in step S5B, step S6 is executed to continue the stamping operation; S6: conveying the sheet material to the mold for stamping, and conveying the stamped sheet material to the next process; wherein, after the sheet material is stamped through the step S6, if the current mold is in an empty state, execute step S1.

2. The foreign matter monitoring method for automobile stamping dies according to claim 1, characterized in that: The foreign body image classification model includes a feature extraction module and a foreign body image classification module; The specific calculation of the feature extraction module is as follows: F i =Pool(ReLU(Conv 3×3 (F i-1 ))) F1=Pool(ReLU(Conv 3×3 (I small ))) Where, F i F represents the features of the small area image of the mold at layer i; i-1 Represents the output of the previous layer; I small The small area image of the mold extracted for the current window; Conv 3×3 Represents a convolution layer with a convolution kernel of 3×3; ReLU represents an activation function; Pool represents a pooling layer; The specific calculation of the foreign body image classification module is as follows: P result =σ(H n ) Where, P result is the classification result of the current mold small area image, that is, the probability value of whether the foreign matter exists; σ is the sigmoid function; H n It is expressed as the output of the nth fully connected layer, and its specific calculation is as follows: H n =ReLU(FC n (H n-1 ))) H1=ReLU(FC1(F i )) Where H n Represents the output of the nth fully connected layer, whose nth fully connected layer FC n The number of neurons n The specific calculation is expressed as: Where n∈[2,N], the calculation of N is expressed as: Among them, neuron1 is the number of neurons in the first fully connected layer.

3. The foreign matter monitoring method for automobile stamping dies according to claim 2, characterized in that: The foreign body recognition model includes a backbone network, a neck network and a head network; The backbone network is used to perform multi-level feature extraction on the input small area image of the mold and output feature maps of several scales; Among them, the specific calculation of preliminary feature extraction is as follows: Feature1=Focus(I small ) In the formula, the Focus module is used to slice the input image to obtain the initial feature map Feature1; For the feature extraction of the i-th level, the specific calculation is expressed as: Feature i =C3(CONV(Feature i-1 ) 2 ),i∈[2,n-1] In the formula, CONV represents the convolution module, and its specific structure is: SILU(BN(Conv2d(F x ))) Where, F x is the input image, Conv2d is the two-dimensional convolution, BN is the batch normalization and SILU is the activation function; C3 represents the residual block, and its specific structure is: CONV(Concat(Add(CONV(F x ) 3 ,CONV(F x )),CONV(F x ) 2 )) In the formula, Add represents element-by-element addition, and Concat represents concatenation operation; The specific representation of feature extraction at the final level is as follows: Feature n =C3(SPP(CONV(Feature i-1 ))) In the formula, Feature n Represents the feature map of the final scale; SPP is spatial pyramid pooling; The neck network is used to perform feature fusion on feature maps of several scales to obtain fused feature maps of several scales, which are specifically expressed as follows: F fusion ={FPN({Feature i |i∈[2,n]})} Where, F fusion Represents a fusion feature atlas of several scales; FPN is a feature pyramid; The head network is used to predict the fused feature maps of several scales, generate several prediction boxes, and filter the several prediction boxes using non-maximum suppression to obtain the foreign body confidence of the current small area image of the mold.

4. The foreign matter monitoring method for automobile stamping dies according to claim 3, characterized in that: The specific calculation of the offset is as follows: Where α is the tolerance coefficient.

5. A foreign body monitoring device for automobile stamping dies, characterized in that: It includes a mold image acquisition unit, an image area segmentation unit, a foreign body image classification unit, a foreign body recognition unit, a foreign body judgment unit, an iterative monitoring and judgment unit, a foreign body early warning unit and a mold stamping control unit; The mold image acquisition unit is used to acquire an image of the mold area before the sheet material enters the mold; The image region segmentation unit is used to segment the current mold region image using a moving window mechanism to obtain the current mold small region image; The foreign body image classification unit is used to classify the current mold small area image using a foreign body image classification model to obtain a classification result of the current mold small area image; The foreign body recognition unit is used to use a foreign body recognition model to perform fine recognition on the current small area image of the mold to obtain the foreign body confidence of the current small area image of the mold; The foreign body judgment unit is used to judge whether the classification result and foreign body confidence of the current mold small area image meet a foreign body condition: if not, the iterative monitoring judgment unit is called; if so, the foreign body early warning unit is called; wherein the foreign body condition is specifically expressed as follows: Wherein, isExist represents the judgment result of whether there is a foreign body in the current mold small area image. When isExist indicates that there is no foreign body, the iterative monitoring judgment unit is called; when isExist indicates that there is a foreign body, the foreign body early warning unit is called; P result is the classification result of the current mold small area image; P target is the foreign body confidence of the current small area image of the mold; max and th min are the maximum confidence threshold and the minimum confidence threshold respectively; cf A and cf B They are respectively the classification result of the current mold small area image and the dynamic confidence threshold corresponding to the foreign body confidence. The specific calculation is as follows: cf A =w A ×cf MAX +w B ×cf MIN +bias cf B =w B ×cf MAX +w A ×cf MIN +bias Where, cf MAX and cf MIN Respectively expressed as P result and P target The maximum and minimum confidence between them; bias is expressed as offset; w A and w B They are the dynamic weights of the classification results and foreign body confidence, and the specific calculation is as follows: Where ΔP is expressed as P result and P target The absolute value difference between The iterative monitoring and judgment unit is used to judge whether the current mold small area image is the last small area image: if not, calling the image area segmentation unit to select the next small area image; if so, calling the mold stamping control unit; The foreign body warning unit is used to issue a warning based on the foreign body confidence level of the current area image and clean the mold; wherein, after the mold is cleaned by the foreign body warning unit, the mold stamping control unit is called to continue the stamping operation; The mold stamping control unit is used to transport the sheet material to the mold for stamping, and transport the stamped sheet material to the next process; wherein, after the sheet material is stamped by the mold stamping control unit, if the current mold is in an empty state, the mold image acquisition unit is called.

6. The foreign matter monitoring device for automobile stamping dies according to claim 5, characterized in that: The foreign body image classification model includes a feature extraction module and a foreign body image classification module; The specific calculation of the feature extraction module is as follows: F i =Pool(ReLU(Conv 3×3 (F i-1 ))) F1=Pool(ReLU(Conv 3×3 (I small ))) Where, F i F represents the features of the small area image of the mold at layer i; i-1 Represents the output of the previous layer; I small The small area image of the mold extracted for the current window; Conv 3×3 Represents a convolution layer with a convolution kernel of 3×3; ReLU represents an activation function; Pool represents a pooling layer; The specific calculation of the foreign body image classification module is as follows: P result =σ(H n ) Where, P result is the classification result of the current mold small area image, that is, the probability value of whether the foreign matter exists; σ is the sigmoid function; H n It is expressed as the output of the nth fully connected layer, and its specific calculation is as follows: H n =ReLU(FC n (H n-1 ))) H1=ReLU(FC1(F i )) Where H n Represents the output of the nth fully connected layer, whose nth fully connected layer FC n The number of neurons n The specific calculation is expressed as: Where n∈[2,N], the calculation of N is expressed as: Among them, neuron1 is the number of neurons in the first fully connected layer; The foreign body recognition model includes a backbone network, a neck network and a head network; The backbone network is used to perform multi-level feature extraction on the input small area image of the mold and output feature maps of several scales; Among them, the specific calculation of preliminary feature extraction is as follows: Feature1=Focus(I small ) In the formula, the Focus module is used to slice the input image to obtain the initial feature map Feature1; For the feature extraction of the i-th level, the specific calculation is expressed as: Feature i =C3(CONV(Feature i-1 ) 2 ),i∈[2,n-1] In the formula, CONV represents the convolution module, and its specific structure is: SILU(BN(Conv2d(F x ))) Where, F x is the input image, Conv2d is the two-dimensional convolution, BN is the batch normalization and SILU is the activation function; C3 represents the residual block, and its specific structure is: CONV(Concat(Add(CONV(F x ) 3 ,CONV(F x )),CONV(F x ) 2 )) In the formula, Add represents element-by-element addition, and Concat represents concatenation operation; The specific representation of feature extraction at the final level is as follows: Feature n =C3(SPP(CONV(Feature i-1 ))) In the formula, Feature n Represents the feature map of the final scale; SPP is spatial pyramid pooling; The neck network is used to perform feature fusion on feature maps of several scales to obtain fused feature maps of several scales, which are specifically expressed as follows: F fusion ={FPN({Feature i |i∈[2,n]})} Where, F fusion Represents a fusion feature atlas of several scales; FPN is a feature pyramid; The head network is used to predict the fused feature maps of several scales, generate several prediction boxes, and filter the several prediction boxes using non-maximum suppression to obtain the foreign body confidence of the current small area image of the mold.

7. The foreign matter monitoring device for automobile stamping dies according to claim 6, characterized in that: The specific calculation of the offset is as follows: Where α is the tolerance coefficient.

8. An automatic stamping production equipment for automobile parts, characterized in that: Includes conveyor belts, punching machines, camera equipment and foreign body monitoring devices; The conveyor belt is used to convey the sheet material to the mold and convey the sheet material after being punched by the punching machine to the next process; The punching machine includes an upper frame and a workbench, and is provided with a die; The die comprises an upper die and a lower die, wherein the upper die is fixedly mounted on the upper frame of the punching machine and is provided with a punch; the lower die is provided with a forming cavity matching the upper die and is fixed to the workbench of the punching machine; the punching machine is used to apply pressure to the sheet material through the upper die and the lower die to form the sheet material into a component of a preset shape; The photographing device is used to photograph the idle mold, obtain an image of the mold area before the sheet metal enters the mold, and transmit the image to the foreign matter detection device; The foreign body monitoring device is used to identify foreign bodies in the input image and obtain a foreign body identification result. If the foreign body identification result indicates the presence of foreign bodies, the production equipment is suspended and an early warning is issued, waiting for the mold to be cleaned; if the foreign body identification result indicates the absence of foreign bodies, the mold is monitored continuously. Wherein, the foreign matter monitoring device is the foreign matter monitoring device for an automobile stamping die according to any one of claims 5-7.

Images