Glass bottle crack detection device
By designing an automatic bottle knocking device and an AI recognition system, combined with a sound defect detection model, the problems of low crack detection efficiency and low manual detection rate in the prior art are solved, and efficient and accurate crack detection of glass bottles are achieved.
Patent Information
- Application Number
- CN202411314423.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2024-11-15
AI Technical Summary
During the production process of existing glass bottles, visual inspection solutions cannot effectively identify crack defects in glass bottles with patterns or paints, and the manual bottle knocking method is labor-intensive and the standards are not uniform, resulting in a low detection rate.
A glass bottle crack detection device is designed, including a chain plate conveyor line, a bottle knocking device, an AI recognition system, a sound collection device, an encoder and a bottle kicking device. The servo drive device and a rotating module are used to realize automatic bottle knocking, and through The AI module and the sound defect detection model are automatically identified to eliminate defect bottles.
It improves the efficiency and accuracy of crack detection in glass bottles, reduces the intensity of manual labor, ensures the stability of detection results, and reduces missed and missed inspections.
Smart Images

Figure CN118961726A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of spray-coated bottle and can production, and in particular to a glass bottle crack detection device. Background Art
[0002] In the process of glass bottle production, crack defects in glass bottles need to be detected. Currently, the industry mostly adopts visual inspection solutions. Due to the imaging principle, this solution cannot effectively detect glass bottles with patterns and paints. Or it uses manual knocking of bottles and manual listening to make judgments. On the one hand, this method has high labor intensity. On the other hand, manual recognition standards are not unified, and attention, energy, and judgment will gradually decrease with the extension of working hours, resulting in missed detections, false detections, and low detection rates. Summary of the invention
[0003] In order to solve the above technical problems, the present invention proposes a glass bottle crack detection device, in order to improve the efficiency and accuracy of crack defect detection in glass bottles, thereby solving the problem that visual inspection solutions cannot effectively identify cracks and the problem that the detection rate of manual bottle knocking is low, and providing a new solution for glass bottle crack identification.
[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:
[0005] The glass crack detection device of the present invention is characterized in that it comprises: a chain plate conveyor line, a bottle knocking device, a bottle kicking device, an AI recognition system, a sound collecting device, an encoder, a collecting device, a soundproof cover, and a frame;
[0006] A frame is installed on the chain conveyor line, the soundproof cover is installed on the frame, and a bottle knocking device and a sound collecting device are arranged in the soundproof cover;
[0007] The bottle kicking device and the collecting device are respectively installed on both sides of the chain conveyor line behind the frame;
[0008] The encoder is installed on the power shaft of the chain conveyor line;
[0009] The bottle knocking device comprises: a servo drive device, a rotating module and a bottle knocking rod;
[0010] A rotating module is installed on the output shaft of the servo drive device, and a bottle knocking rod is installed on the rotating module. One side of one end of the bottle knocking rod is connected to one end of a tension spring, and the other end of the tension spring is fixed on the spring support. A limit column is arranged on the other side of one end of the bottle knocking rod;
[0011] The AI recognition system includes an AI module and a control system;
[0012] After receiving the bottle arrival signal sent by the photoelectric switch on the chain conveyor line, the control system sends a forward drive signal to the servo drive device to drive the rotating module to drive the bottle knocking rod to knock the current bottle on the chain conveyor line under the action of the tension spring. Then, driven by the reverse drive signal, the bottle knocking rod rotates back to its position under the limitation of the limit column.
[0013] The sound collection device located above the bottle knocking point collects the current knocking sound and sends it to the AI module;
[0014] The encoder obtains the position of the current bottle when it is hit and sends it to the AI module;
[0015] The AI module matches the current position of the bottle with the current knocking sound, and inputs the current knocking sound into the trained sound defect detection model for identification, and obtains the type and confidence of the knocking sound of the current bottle to determine whether the current bottle has a defect; if so, the position of the current bottle when it reaches the bottle kicking device is calculated according to the current position of the bottle, so as to send a rejection signal to the bottle kicking device; otherwise, the identification of the next bottle continues;
[0016] If the bottle-kicking device receives a rejection signal, when the current bottle reaches the bottle-kicking device, the bottle will be rejected from the chain conveyor line into the collecting device; otherwise, the current bottle will not be rejected.
[0017] The glass crack detection device of the present invention is also characterized in that the sound defect detection model is trained according to the following steps:
[0018] Step 1: Collect bottle sound sample set ,in, Indicates Bottle sound samples, The label is recorded as ,like =0, indicating For a non-defective bottle, if =1, indicating is a defective bottle; n represents the total number of samples;
[0019] Step 2: Construct a sound defect detection network, including: data preprocessing module, feature extraction module, and CNN training module;
[0020] Step 2.1: The data preprocessing module After standardization and removal of the silent part, cropping and padding are performed to obtain the preprocessed Bottle sound samples ;
[0021] Step 2.2: The feature extraction module After MFCC feature extraction, we get MFCC features , then After Delta transformation, the i-th transformation feature is obtained ; thus and After fusion, the i-th fusion feature is obtained ;
[0022] Step 2.3: The CNN training module Process and get the predicted category i ;
[0023] Step 2.4: Use formula (1) to construct the loss function :
[0024] (1)
[0025] Step 2.5: Use the Adam optimization algorithm to train the sound defect detection network and calculate the loss function To update the network parameters until The training process is continued until convergence, thereby obtaining the trained sound defect detection model.
[0026] Furthermore, the CNN training module in step 2.3 includes: an initial mixed convolutional layer, a feature fusion module, a residual module, and a classification module;
[0027] Step 2.3.1, the initial mixed convolution layer includes a mixed convolution layer and a feature fusion layer;
[0028] The hybrid convolutional layer consists of 2D spatial blocks, where any A 2D spatial block consists of a convolution kernel × A convolutional layer, a BN layer and a ReLU activation function are connected in series; and is an odd number greater than or equal to 3;
[0029] Enter the After processing in 2D space blocks, we get Feature Map ,in, ≤ ;
[0030] The feature fusion layer extracts the Feature Map , …, …, After splicing in the channel dimension, the i-th convolution fusion feature map is obtained ;
[0031] Step 2.3.2, the feature fusion module includes two branches, each branch includes Depth-wise convolution, BN normalization, Point-wise convolution, BN normalization and ReLU activation operations;
[0032] After being processed by the two branches respectively, the i-th first branch fusion feature map is obtained And the ith second branch fusion feature map , and then and After connecting in the channel dimension, the i-th branch fusion feature map is obtained ;
[0033] Step 2.3.3, the residual module includes a short-circuited branch and a fusion branch, wherein the fusion branch is formed by serially connecting a hybrid convolutional layer, a BN layer and a ReLU activation function;
[0034] After processing in the short-circuit branch, the i-th short-circuit feature map is obtained. ;
[0035] After being processed in the input fusion branch, the i-th fusion feature map is obtained ;
[0036] Will and After the connection operation in the channel dimension, the i-th concatenated feature map is obtained ;
[0037] Step 2.3.4, the classification module consists of a global average pooling layer, a fully connected layer and a Sigmoid function, and After processing, the predicted category is obtained i .
[0038] Compared with the prior art, the present invention has the following beneficial effects:
[0039] The present invention can solve the problem that the visual solution cannot be identified by designing a set of automatic bottle knocking devices, and can also realize full automation of automatic bottle knocking, effectively solving the problem of high labor intensity of manual bottle knocking and saving labor. At the same time, the present invention can replace manual automatic judgment by designing a sound defect detection model based on a neural network, solving the problem that the energy and judgment of manual labor decrease as the working time increases, and can effectively ensure the stability of crack defect detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 This is a front view of the bottle knocking device according to an embodiment of the present application;
[0041] Figure 2 This is a cross-sectional view of the bottle knocking device 2 according to an embodiment of the present application;
[0042] Figure 3 This is an axonometric diagram of the bottle knocking device 2 according to an embodiment of the present application;
[0043] Figure 4 A flow chart of the sound defect detection model for this application;
[0044] Figure 5 This is the structural diagram of the initial mixed convolutional layer of this application;
[0045] Figure 6 This is a structural diagram of the feature fusion module and residual module of this application;
[0046] Numbers in the figure: 1 chain conveyor line; 2 bottle knocking device; 3 rejection device; 4 AI recognition system; 5 sound collection device; 6 encoder; 7 collection device; 8 sound insulation cover; 9 rack; 21 servo drive device; 22 connecting plate; 23 rotating module; 24 bottle knocking rod; 25 tension spring. DETAILED DESCRIPTION
[0047] In this embodiment, a glass crack detection device is provided. Figure 1 As shown, it includes: a chain conveyor line 1, a bottle knocking device 2, a bottle kicking device 3, an AI recognition system 4, a sound collecting device 5, an encoder 6, a collecting device 7, a soundproof cover 8, and a frame 9;
[0048] A frame 9 is installed on the chain conveyor line 1, a soundproof cover 8 is installed on the frame 9, and a bottle knocking device 2 and a sound collecting device 5 are arranged in the soundproof cover 8;
[0049] A bottle kicking device 3 and a collecting device 7 are respectively installed on both sides of the chain conveyor line 1 behind the frame 9;
[0050] An encoder 6 is installed on the power shaft of the chain conveyor line 1;
[0051] like Figure 2As shown, the bottle knocking device 2 includes: a servo drive device 21, a rotating module 23 and a bottle knocking rod 24;
[0052] like Figure 3 As shown, a rotating module 23 is installed on the output shaft of the servo drive device 21, and a bottle knocking rod 24 is installed on the rotating module 23. One side of one end of the bottle knocking rod 24 is connected to one end of a tension spring 25, and the other end of the tension spring 25 is fixed on a spring support 26. A limiting column 27 is provided on the other side of one end of the bottle knocking rod 24;
[0053] The AI recognition system includes an AI module and a control system;
[0054] After receiving the bottle arrival signal sent by the photoelectric switch on the chain conveyor line 1, the control system sends a positive drive signal to the servo drive device 21 to drive the rotating module 23 to drive the bottle knocking rod 24 to knock the current bottle on the chain conveyor line 1 under the action of the tension spring 25. Then, driven by the reverse drive signal, the bottle knocking rod 24 rotates back to its original position under the limitation of the limit column 27;
[0055] The sound collecting device 5 located above the bottle knocking point collects the current knocking sound and sends it to the AI module;
[0056] The encoder obtains the position of the current bottle when it is hit and sends it to the AI module;
[0057] The AI module matches the current position of the bottle with the current knocking sound, and inputs the current knocking sound into the trained sound defect detection model for identification, and obtains the type and confidence of the knocking sound of the current bottle to determine whether the current bottle has a defect; if so, the position of the current bottle when it reaches the bottle-kicking device 3 is calculated according to the current position of the bottle, and a rejection signal is sent to the bottle-kicking device 3; otherwise, the identification of the next bottle continues;
[0058] If the bottle-kicking device 3 receives a rejection signal, when the current bottle reaches the bottle-kicking device 3 , the bottle will be rejected from the chain conveyor line 1 into the collecting device 7 ; otherwise, the current bottle will not be rejected.
[0059] In this embodiment, Figure 4 As shown in the figure, the sound defect detection model is trained according to the following steps:
[0060] Step 1: Collect bottle sound sample set ,in, Indicates Bottle sound samples, The label is recorded as ,like =0, indicating For a non-defective bottle, if =1, indicating is a defective bottle; n represents the total number of samples; in this example, for better training effect, a set of defective and non-defective sound samples are collected at different times and locations at the work site.
[0061] Step 2: Construct a sound defect detection network, including: data preprocessing module, feature extraction module, and CNN training module;
[0062] Step 2.1: Data preprocessing module After standardization and removal of the silent part, cropping and padding are performed to obtain the preprocessed Bottle sound samples ; In order to ensure the consistency of all audio samples before feature extraction and model training, in this example, The frequency is normalized at 48KHz, silence is removed, and then the video is cropped and padded with a target length of 3 seconds.
[0063] Step 2.2: Feature extraction module After MFCC feature extraction, we get MFCC features , then After Delta transformation, the i-th transformation feature is obtained ; thus and After fusion, the i-th fusion feature is obtained In this example, in addition to extracting the spectral characteristics of the sound MFCC, the feature Delta of the audio changing over time is also extracted, and the two are fused to improve the feature differentiation ability and enhance the robustness of the model under different environments and recording conditions.
[0064] Step 2.3: Construct a CNN training module including: initial mixed convolution layer, feature fusion module, residual module, and classification module.
[0065] Step 2.3.1, such as Figure 5 As shown, the initial mixed convolution layer contains a mixed convolution layer and a feature fusion layer;
[0066] The mixed convolutional layer consists of 2D spatial blocks, where any A 2D spatial block consists of a convolution kernel × A convolutional layer, a BN layer and a ReLU activation function are connected in series; and is an odd number greater than or equal to 3;
[0067] Will Enter the After processing the 2D space blocks, we get Feature Map ,in ≤ ; The feature fusion layer extracts different convolution kernels Feature Map , …, …, }Splice in the channel dimension to get the i-th convolution fusion feature map .
[0068] In this example, in order to capture the time dimension information of different scales, three different 2D spatial blocks are set, namely 3×1, 5×1, and 7×1 convolution kernels with 32-channel outputs. At the same time, in order to improve the expression ability and Luban characteristics, the features processed by the above three 2D spatial blocks are processed in parallel to a 96-channel feature map.
[0069] Step 2.3.2, such as Figure 6 As shown in the figure, the feature fusion module contains two branches, each of which contains Depth-wise convolution, BN normalization, Point-wise convolution, BN normalization and ReLU activation operations;
[0070] After being processed by the two branches respectively, the i-th first branch fusion feature map is obtained And the ith second branch fusion feature map , then and After connecting in the channel dimension, the i-th branch fusion feature map is obtained .
[0071] In order to extract sound signal features at different scales, the Depth-wise convolution layers in the two branches of the feature fusion module use convolution kernels of different sizes. In this example, the input and output channels of the first branch Depth-wise convolution layer are both 96, the convolution kernel size is 3×1, padding=1, and the input and output channels of the first branch Point-wise convolution layer are 96. The input and output channels of the second branch Depth-wise convolution layer are both 96, the convolution kernel size is 5×1, padding=2, and the input and output channels of the second branch Point-wise convolution layer are 96.
[0072] Step 2.3.3, the residual module includes a short-circuited branch and a fusion branch; the fusion branch is composed of a mixed convolution, a BN layer and a ReLU activation function connected in series.
[0073] After processing in the short-circuit branch, the i-th short-circuit feature map is obtained. ;
[0074] After being processed in the input fusion branch, the i-th fusion feature map is obtained ;
[0075] and After the connection operation in the channel dimension, the i-th concatenated feature map is obtained .
[0076] In order to solve the gradient vanishing / exploding problem in deep networks and reduce the model complexity, in this example, three different convolution kernels 3×1, 5×1, and 7×1 are set for the mixed convolution.
[0077] Step 2.3.4, the classification module consists of a global average pooling layer, a fully connected layer and a Sigmoid function;
[0078] The input is sent to the classification module, and after being processed by the global average pooling layer and the fully connected layer, the predicted category is obtained through the Sigmoid function. i ;
[0079] Step 2.4: Use formula (1) to construct the loss function :
[0080] (1)
[0081] Step 2.5: Use the Adam optimization algorithm to train the sound defect detection network and calculate the loss function To update the network parameters until The training process is continued until convergence, thereby obtaining the trained sound defect detection model.
[0082] The present invention judges the quality of the bottle according to the sound made by the bottle by knocking on the bottle, thereby successfully solving the industry pain points of the current crack detection, greatly reducing the product defects of glass bottles, improving the production quality of the enterprise, and can well improve the efficiency of the enterprise.
Claims
1. A glass crack detection device, characterized in that: include: Chain conveyor line (1), bottle knocking device (2), bottle kicking device (3), AI recognition system (4), sound collection device (5), encoder (6), collection device (7), soundproof cover (8), rack (9); A frame (9) is installed on the chain conveyor line (1), the soundproof cover (8) is installed on the frame (9), and a bottle knocking device (2) and a sound collecting device (5) are arranged in the soundproof cover (8); The bottle-kicking device (3) and the collecting device (7) are respectively installed on both sides of the chain conveyor line (1) behind the frame (9); The encoder (6) is installed on the power shaft of the chain conveyor line (1); The bottle knocking device (2) comprises: a servo drive device (21), a rotating module (23) and a bottle knocking rod (24); A rotating module (23) is installed on the output shaft of the servo drive device (21), and a bottle knocking rod (24) is installed on the rotating module (23). One side of one end of the bottle knocking rod (24) is connected to one end of a tension spring (25), and the other end of the tension spring (25) is fixed on a spring support (26). A limiting column (27) is arranged on the other side of one end of the bottle knocking rod (24); The AI recognition system includes an AI module and a control system; After receiving the bottle arrival signal sent by the photoelectric switch on the chain conveyor line (1), the control system sends a positive drive signal to the servo drive device (21) to drive the rotating module (23) to drive the bottle knocking rod (24) to knock the current bottle on the chain conveyor line (1) under the action of the tension spring (25). Then, driven by the reverse drive signal, the bottle knocking rod (24) rotates back to its original position under the limitation of the limit column (27); The sound collecting device (5) located above the bottle knocking point collects the current knocking sound and sends it to the AI module; The encoder obtains the position of the current bottle when it is hit and sends it to the AI module; The AI module matches the current position of the bottle with the current knocking sound, and inputs the current knocking sound into the trained sound defect detection model for identification, thereby obtaining the type and confidence of the knocking sound of the current bottle, so as to determine whether the current bottle has a defect; if so, the position of the current bottle when it reaches the bottle-kicking device (3) is calculated according to the current position of the bottle, thereby sending a rejection signal to the bottle-kicking device (3); otherwise, the identification of the next bottle is continued; If the bottle-kicking device (3) receives a rejection signal, then when the current bottle reaches the bottle-kicking device (3), the bottle is rejected from the chain conveyor line (1) into the collecting device (7); otherwise, the current bottle is not rejected.
2. A glass crack detection device according to claim 1, characterized in that: The sound defect detection model is trained according to the following steps: Step 1: Collect bottle sound sample set ,in, Indicates Bottle sound samples, The label is recorded as ,like =0, indicating For a non-defective bottle, if =1, indicating is a defective bottle; n represents the total number of samples; Step 2: Construct a sound defect detection network, including: data preprocessing module, feature extraction module, and CNN training module; Step 2.1: The data preprocessing module After standardization and removal of the silent part, cropping and padding are performed to obtain the preprocessed Bottle sound samples ; Step 2.2: The feature extraction module After MFCC feature extraction, we get MFCC features , then After Delta transformation, the i-th transformation feature is obtained ; thus and After fusion, the i-th fusion feature is obtained ; Step 2.3: The CNN training module Process and get the predicted category i ; Step 2.4: Use formula (1) to construct the loss function : (1) Step 2.5: Use the Adam optimization algorithm to train the sound defect detection network and calculate the loss function To update the network parameters until The training process is continued until convergence, thereby obtaining the trained sound defect detection model.
3. A glass crack detection device according to claim 2, characterized in that: The CNN training module in step 2.3 includes: an initial mixed convolution layer, a feature fusion module, a residual module, and a classification module; Step 2.3.1, the initial mixed convolution layer includes a mixed convolution layer and a feature fusion layer; The hybrid convolutional layer consists of 2D spatial blocks, where any A 2D spatial block consists of a convolution kernel × A convolutional layer, a BN layer and a ReLU activation function are connected in series; and is an odd number greater than or equal to 3; Enter the After processing in 2D space blocks, we get Feature Map ,in, ≤ ; The feature fusion layer extracts the Feature Map , …, …, After splicing in the channel dimension, the i-th convolution fusion feature map is obtained ; Step 2.3.2, the feature fusion module includes two branches, each branch includes Depth-wise convolution, BN normalization, Point-wise convolution, BN normalization and ReLU activation operations; After being processed by the two branches respectively, the i-th first branch fusion feature map is obtained And the ith second branch fusion feature map , and then and After connecting in the channel dimension, the i-th branch fusion feature map is obtained ; Step 2.3.3, the residual module includes a short-circuited branch and a fusion branch, wherein the fusion branch is formed by serially connecting a hybrid convolutional layer, a BN layer and a ReLU activation function; After processing in the short-circuit branch, the i-th short-circuit feature map is obtained. ; After being processed in the input fusion branch, the i-th fusion feature map is obtained ; Will and After the connection operation in the channel dimension, the i-th concatenated feature map is obtained ; Step 2.3.4, the classification module consists of a global average pooling layer, a fully connected layer and a Sigmoid function, and After processing, the predicted category is obtained i .
