Artificial intelligence quality detection system and method for stationery production line

Abstract

The invention discloses an artificial intelligence quality detection system and method for a stationery production line, and relates to the field of production line quality inspection. A parameter determination module obtains production parameters, sets quality inspection parameters based on a minimum safety criterion, and replaces low-efficiency full inspection with efficient and accurate sampling inspection; the control module initializes or adjusts a quality inspection control table in real time based on the quality inspection parameters or time period result labels so as to ensure the casual inspection credibility; the acquisition module acquires an external image and projection data of the stationery in a quality inspection period; the detection module adopts a defect recognition network to perform surface defect recognition, performs internal defect recognition based on an FBP algorithm and a Marking Cubes algorithm, and accurately judges whether the stationery is qualified or not based on internal and external double view angles of the stationery, so that the quality inspection accuracy is improved; and the statistical analysis module counts sample quality inspection results in a single quality inspection time period, generates time period result labels, and autonomously judges whether to stop the production of the batch and stop loss in time, so that high-accuracy, high-reliability, high-efficiency and low-cost automatic quality inspection of the stationery production line is realized.

Description

Technical Field

[0001] The present invention relates to the field of production line quality inspection, and in particular to an artificial intelligence quality inspection system and method for a stationery production line. Background Art

[0002] In today's stationery manufacturing industry, quality inspection is a key link in ensuring product quality and meeting consumer demand. With the intensification of market competition and the improvement of consumer demand for product quality, it is of great significance to implement batch quality inspection on stationery production lines.

[0003] Currently, quality inspection for stationery products relies primarily on manual inspection and machine vision. However, both methods have significant drawbacks. Manual inspection requires significant manpower and is slow, failing to meet the demands of rapid production. Furthermore, due to human error, misjudgments and missed detections frequently occur, impacting product quality stability. Machine vision inspection, based on deep learning and artificial intelligence, analyzes high-definition images of stationery to detect macro-defects such as dimensional discrepancies and surface damage. This enables precise identification of macro-defects, significantly improving inspection efficiency.

[0004] However, with the large-scale application of automated production lines, the production volume and precision of stationery have increased dramatically. Full inspection often wastes time and increases meaningless quality inspection costs. At the same time, existing quality inspection solutions cannot effectively detect damage to the internal structure of stationery, and there is still room for improvement in quality inspection accuracy.

[0005] Therefore, studying an artificial intelligence quality inspection system and method for stationery production lines is of great significance for improving product quality, reducing production costs and improving production efficiency. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes an artificial intelligence quality inspection system and method for stationery production lines, which provides a high-efficiency sampling solution for both surface defects and internal defects of stationery from two perspectives, significantly reducing quality inspection costs and improving quality inspection efficiency while ensuring detection accuracy.

[0007] The technical solutions for achieving the purpose of the present invention are: An artificial intelligence quality inspection system for a stationery production line, comprising a parameter determination module, a control module, an acquisition module, a detection module, and a statistical analysis module; The parameter setting module sets the quality inspection parameters based on the minimum safety criteria and the production parameters of this batch, and uploads them to the control module; The control module initializes the quality inspection control table based on the quality inspection parameters Or adjust the quality control table in real time based on the time period result label , and sent to the acquisition module; Acquisition module based on quality control table Distinguish between the quality inspection period and the silent period, obtain the external image and projection data of the sample during the quality inspection period, add the period mark and upload it to the detection module; The detection module uses a defect recognition network to identify surface defects and uses the FBP algorithm and Marching Cubes algorithm to identify internal defects. It assigns sample result labels and adds time period tags before reporting them to the statistical analysis module. The statistical analysis module counts all sample result labels within a single quality inspection period, generates a period result label and feeds back to the control module to independently decide whether to stop the production of this batch.

[0008] Furthermore, the parameter determination module determines the quality inspection parameters including the following specific steps: Get the production parameters of this batch; Build a quality inspection scale model , based on the total production and product risk level Determine the quality inspection ratio , the specific formula is as follows: , in, The highest product risk level. and is the lower limit ratio and the upper limit ratio; Set the total number of quality inspection cycles , set the quality inspection start time Set to the production start time Follow a quality inspection cycle; A single quality inspection cycle is further divided into time periods, and the unit duration corresponding to each time period is , The total number of time periods.

[0009] Furthermore, the control module includes an initialization unit and a real-time adjustment unit; Initialization unit obtains quality inspection parameters and generates quality inspection control table , sent to the data acquisition module; The real-time adjustment unit obtains the period result label of a single period and updates the quality inspection control table based on the risk addition strategy And send it to the data acquisition module.

[0010] Furthermore, the initialization unit generates a quality control table The specific steps include: Generate runtime vector based on quality inspection parameters ; Initialize and generate a quality inspection identification vector of all 0s , select the first period of each quality inspection cycle as the quality inspection period, and set the starting time of each quality inspection period in the quality inspection identification vector The identifier of the corresponding position in is set to 1; Corresponding splicing running time vector and quality inspection logo vector Generate quality control table .

[0011] Furthermore, the real-time adjustment unit updates the quality control table based on the risk supplement strategy The specific steps include: Get the Time period The results label for the period , ; Judgment period result label Equal to 0 or 1, if the period result label If it is equal to 0, no adjustment is made; If the period result label If it is equal to 1, the product risk level is obtained. and the total number of remaining periods ; Product risk level and the total number of remaining periods The smaller value is used as the adjustment period number , will Time period After All time periods are adjusted to quality inspection periods, and the quality inspection control table is updated The identifier of the corresponding position in is 1.

[0012] Furthermore, the data acquisition module includes a switch unit, a shooting unit, a scanning unit and a timing unit; Switching unit based on quality control table The identifier in determines the quality inspection period and the silent period after the current period to turn on or off the shooting unit and the scanning unit; The photographing unit photographs an external image of the sample based on infrared triggering; The scanning unit includes an X-ray scanning device, which scans and obtains projection data of the sample at multiple angles during the quality inspection period; The timing unit distinguishes time periods based on the system time, adds time period tags to external images and projection data, and uploads them to the detection module.

[0013] Furthermore, the shooting unit includes an infrared emitting device, a receiving trigger device and a shooting device; The infrared emitting device continuously emits infrared light in a fixed direction toward the receiving trigger device; The receiving trigger device sends a start command to the shooting device when the receiving trigger device does not receive the infrared light or receives the infrared light that is partially blocked; After receiving the start command, the photographing device photographs the external image of the sample in real time.

[0014] Furthermore, the detection module includes a pre-processing unit, an external analysis unit, an internal analysis unit and a decision unit; The pre-processing unit uses a Gaussian filter to denoise the external image and projection data; The external analysis unit identifies the surface defects of the sample based on the defect analysis network, assigns an external decision label and sends it to the decision unit; The internal analysis unit generates a reconstructed 3D model using the FBP algorithm and the Marching Cubes algorithm, compares the model with the standard model, assigns an internal decision label, and sends it to the decision unit. The decision unit determines whether the sample is qualified based on whether the external decision tag and the internal decision tag are all 0, assigns the sample result tag and adds a time period mark to report to the statistical analysis module.

[0015] Furthermore, the external analysis unit assigns an external decision label based on the defect analysis network, including the following specific steps: Get the External images of samples , using Fast R-CNN to remove external images The background area in the image is used to obtain the sample image. ; Pass in sequence and Convolution kernel is used to generate 16-channel potential features ; Aggregate latent features using average pooling and maximum pooling The spatial information of the average feature map is generated and the maximum feature map , and generate dual-view feature maps by combining nonlinear modulation ; Ghost convolution is used to map the dual-view feature maps of 16 channels The ghost features are transformed into ghost features through different linear operations, and the ghost features of 16 channels are spliced to generate a ghost feature map. ; Two linear layers are used to reduce the dimension of the ghost feature map And further use the Sigmoid activation function to generate the Surface defect probability of a sample , based on the surface defect probability Whether it is greater than or equal to 0.5 sets the external decision label to 1 or 0.

[0016] Furthermore, the internal analysis unit generates an internal decision label including the following specific steps: Get the projection data of the cth sample , assuming the projection data Including projection vectors at different angles, is the total number of angles; Select the spatial filter of the FBP algorithm, multiply all projection vectors by the spatial filter after FFT transformation, and further use IFFT transformation and back projection to reconstruct the internal image; Tomography is used to stack internal images into 3D slices, and the Marching Cubes algorithm is used to generate a reconstructed 3D model; Import the reconstructed 3D model and the standard model into 3dsMax software, calculate the overlap ratio and perform overlap comparison; Determine whether the overlap ratio is greater than or equal to the ratio threshold. If so, The internal decision label of the sample is set to 0, and if it is less than , it is set to 1.

[0017] Furthermore, the statistical analysis module includes a statistical unit and a suspension unit; The statistics unit obtains the sample result label and identifies the added time period mark, assuming that it is the first time period, calculate the The qualified rate of each period is determined to be greater than or equal to the qualified threshold. If it is greater than or equal to, the result label is set to 0, if it is less than, it is set to 1, and the result label of the period is fed back. To the control module; The suspension unit performs statistical judgment, counts the total number of time periods that fail quality inspection, and determines whether it is greater than the total number of quality inspection cycles. If it is greater than, stop the production of this batch.

[0018] An artificial intelligence quality inspection method for a stationery production line, used to implement an artificial intelligence quality inspection system for a stationery production line, includes the following specific steps: Get the production parameters of this batch, set the quality inspection parameters based on the minimum safety criteria, and initialize the quality inspection control table ; Based on the quality control table Acquire the external image and projection data of the sample during the quality inspection period and add a time period mark; Use defect recognition network to identify surface defects, use FBP algorithm and Marching Cubes algorithm to identify internal defects, generate sample result labels and add time period marks; Based on the sample result labels within a single quality inspection period, the period qualification rate is calculated, the period qualification label is assigned, and statistical judgment is performed to decide whether to stop the production of this batch. If not, further judgment is made whether to adjust the quality inspection control table in real time. .

[0019] Furthermore, setting quality inspection parameters based on minimum safety criteria includes the following specific steps: Get the production parameters of this batch; Build a quality inspection scale model , based on the total production and product risk level Determine the quality inspection ratio , the specific formula is as follows: , in, The highest product risk level. and is the lower limit ratio and the upper limit ratio; Set the total number of quality inspection cycles , set the quality inspection start time Set to the production start time Follow a quality inspection cycle; A single quality inspection cycle is further divided into time periods, and the unit duration corresponding to each time period is , The total number of time periods.

[0020] Furthermore, initialize the quality control table The specific steps include: Generate runtime vector based on quality inspection parameters ; Initialize and generate a quality inspection identification vector of all 0s , select the first period of each quality inspection cycle as the quality inspection period, and set the starting time of each quality inspection period in the quality inspection identification vector The identifier of the corresponding position in is set to 1; Corresponding splicing running time vector and quality inspection logo vector Generate quality control table .

[0021] Furthermore, obtaining the external image and projection data of the sample and adding time period marks includes the following specific steps: Get the quality control sheet and the current period, determining a quality inspection period and a silent period following the current period based on an identifier in the quality inspection control table; During the quality inspection period, a directional infrared light is emitted. If the infrared light is not received or is partially blocked, the external image and projection data of the sample are obtained through the camera and X-ray scanning device; Add the current period mark to external images and projection data; During quiet periods, turn off the camera and X-ray scanner.

[0022] Furthermore, the use of a defect recognition network to identify surface defects includes the following specific steps: Get the External images of samples , using Fast R-CNN to remove external images The background area in the image is used to obtain the sample image. ; Pass in sequence and Convolution kernel is used to generate 16-channel potential features ; Aggregate latent features using average pooling and maximum pooling The spatial information of the average feature map is generated and the maximum feature map , and generate dual-view feature maps by combining nonlinear modulation ; Ghost convolution is used to map the dual-view feature maps of 16 channels The ghost features are transformed into ghost features through different linear operations, and the ghost features of 16 channels are spliced to generate a ghost feature map. ; Two linear layers are used to reduce the dimension of the ghost feature map And further use the Sigmoid activation function to generate the Surface defect probability of a sample , based on the surface defect probability Whether it is greater than or equal to 0.5, the external decision label is set to 0 or 1.

[0023] Furthermore, the use of the FBP algorithm and the Marching Cubes algorithm to identify internal defects includes the following specific steps: Get the Projection data of samples , assuming the projection data Including projection vectors at different angles, is the total number of angles; Select the spatial filter of the FBP algorithm, multiply all projection vectors by the spatial filter after FFT transformation, and further use IFFT transformation and back projection to reconstruct the internal image; Tomography is used to stack internal images into 3D slices, and the Marching Cubes algorithm is used to generate a reconstructed 3D model; Import the reconstructed 3D model and the standard model into 3dsMax software, calculate the overlap ratio and perform overlap comparison; Determine whether the overlap ratio is greater than or equal to the ratio threshold. If so, The internal and external judgment labels of the samples are set to 0, otherwise they are set to 1.

[0024] Furthermore, the sample result label is used to determine whether the sample is qualified based on whether the external decision label and the internal decision label of the sample are all 0. If they are all 0, the sample result label is set to 0, otherwise it is set to 1.

[0025] Furthermore, the autonomous determination of whether to stop production of this batch includes the following specific steps: Obtain all sample result labels for a single period and calculate the qualified rate for that period; Determine whether the qualified rate of the time period is greater than or equal to the qualified threshold. If so, set the qualified flag of the time period to 0; If it is less than, the qualified label of the time period is set to 1, and the total number of time periods that fail the quality inspection is increased by 1; Further determine whether the total number of time periods that failed quality inspection is greater than the total number of quality inspection cycles If it is greater than, stop the production of this batch; otherwise, determine whether to adjust the quality control table in real time. .

[0026] Furthermore, determine whether to adjust the quality control table in real time The specific steps include: Get the Time period The results label for the period , ; Judgment period result label Equal to 0 or 1, if the period result label If it is equal to 0, no adjustment is made; If the period result label If it is equal to 1, the product risk level and the total number of remaining periods The smaller value is used as the adjustment period number ; The first Time period After All time periods are adjusted to quality inspection periods, and the quality inspection control table is updated The identifier of the corresponding position in is 1.

[0027] Compared with the prior art, the present invention has the following significant advantages: 1. Design a parameter determination module, using minimum safety criteria to determine quality inspection parameters. Replace full inspections with efficient and accurate spot checks based on the quality inspection control table. Design a control module to adjust the quality inspection control table in real time based on time period result labels to ensure the credibility of spot checks. Design a statistical analysis module to coordinate the spot check process, promptly stop the production of substandard stationery, and stop losses in a timely manner. 2. Design a linkage analysis between the acquisition module and the detection module, use a defect recognition network to analyze external images to identify surface defects, and use the FBP algorithm and Marching Cubes algorithm to generate a reconstructed 3D model to determine whether there are defects in the internal structure. This combination of internal and external methods can more accurately detect whether stationery is qualified, greatly improving the accuracy of quality inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic diagram of an artificial intelligence quality inspection system for a stationery production line; Figure 2 This is a flow chart of the control module initialization or real-time adjustment of the quality inspection control table in the present invention; Figure 3 Schematic diagram of the defect analysis network model in the present invention; Figure 4 This is a flow chart of assigning internal decision labels in the present invention; Figure 5 The following is a flow chart of an artificial intelligence quality inspection method for stationery production lines. DETAILED DESCRIPTION

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Example

[0030] like Figure 1 As shown, a specific embodiment of the present invention discloses an artificial intelligence quality inspection system for a stationery production line, including a parameter setting module, a control module, an acquisition module, a detection module and a statistical analysis module; The parameter setting module refers to the minimum safety criteria, sets the quality inspection parameters based on the production parameters of this batch, and uploads the quality inspection parameters to the control module; The control module obtains quality inspection parameters and initializes the quality inspection control table , obtain the time period result label and adjust the quality control table in real time , send quality control form To get the module; Acquisition module based on quality control table The identifier in the image distinguishes the quality inspection period from the silent period, and only obtains the external image and projection data of the sample during the quality inspection period, and adds the period mark to upload to the detection module; The inspection module uses a defect recognition network to analyze external images for surface defect recognition, and further identifies internal defects based on the FBP algorithm and Marching Cubes algorithm. It assigns sample result labels and adds time period tags to report to the statistical analysis module. The statistical analysis module counts the sample result labels of all samples in a single quality inspection period and calculates the period qualification rate, assigns the period result label and feeds back to the control module, and independently decides whether to stop the production of this batch.

[0031] Furthermore, the minimum safety criterion determines the quality inspection parameters based on the production parameters, which is used to balance the quality inspection cost and quality inspection credibility, while ensuring the credibility of the quality inspection results and reducing the number of quality inspections as much as possible to improve efficiency and reduce costs. The production parameters include the total number of production , production speed , Production start time and product risk level , quality inspection parameters include quality inspection ratio , Quality inspection cycle , Quality inspection start time and unit quality inspection time .

[0032] Furthermore, the parameter determination module determines the quality inspection parameters including the following specific steps: Get the total number of production , production speed , Production start time and product risk level ; Build a quality inspection scale model , based on the total production and product risk level Determine the quality inspection ratio , the specific formula is as follows: , in, The highest product risk level. and is the lower limit ratio and upper limit ratio, quality inspection ratio model Indicates the quality inspection ratio Within a certain range, as the product risk level Improvement and production totals grows larger with the increase of Set the total number of quality inspection cycles , quality inspection starts time Production start time One quality inspection cycle, i.e. the time when quality inspection starts , quality inspection must be carried out in each quality inspection cycle; A single quality inspection cycle is further divided into time periods, and the unit duration corresponding to each time period is , The total number of time periods. In each quality inspection cycle, as long as at least one time period is used as a quality inspection period for quality inspection, the quality inspection ratio can be met. .

[0033] Furthermore, the control module includes an initialization unit and a real-time adjustment unit; Initialization unit obtains quality inspection parameters and initializes the generated dimensions to Quality control table , sent to the acquisition module, where and They are respectively the running time vector and the quality inspection identification vector, the quality inspection control table To distinguish The quality inspection period and silent period in a quality inspection cycle, is the quality inspection ratio, is the total number of quality inspection cycles; The real-time adjustment unit obtains the period result label of a single period, determines the number of adjustment periods based on the risk-added strategy, and updates the quality inspection control table. And send it to the acquisition module.

[0034] like Figure 2 As shown in (a), further, the initialization unit initializes and generates a quality control table The specific steps include: Generate runtime vector ,in, The time when quality inspection starts. is the total number of time periods, The unit quality inspection duration; Initialization Generation The quality inspection identification vector with all identifiers being 0 , the first period of each quality inspection cycle is used as the quality inspection period, and the rest of each quality inspection cycle is The period is used as the silent period, and the start time of the quality inspection period is determined at the running time vector The quality inspection mark vector The identifier of the corresponding position in is set to 1. 、 For example, the runtime vector and quality inspection logo vector The details are as follows:

[0035]

[0036] That is, within 2 quality inspection cycles, only the period and time period As a quality inspection period; Corresponding splicing running time vector and quality inspection logo vector Generate quality control table .

[0037] like Figure 2 As shown in (b), further, the real-time adjustment unit updates the quality control table based on the risk addition strategy The specific steps include: Get the result label of a single period, assuming it is Time period Period Results Label , ; Judgment period result label Equal to 0 or 1, if the period result label If it is equal to 0, it means that Time period The quality inspection has passed and no adjustment is made; If the period result label If it is equal to 1, it means that Time period If the product fails the quality inspection, obtain the product risk level and the total number of remaining periods ; Product risk level and the total number of remaining periods The smaller value is used as the adjustment period number , will Time period After All time periods are adjusted to quality inspection time periods. If they are already quality inspection time periods, they remain unchanged and the quality inspection control table is updated. The identifier of the corresponding position in is 1, that is, Period to The identifiers of the positions corresponding to the start times of each time period are all set to 1.

[0038] Furthermore, the acquisition module includes a switch unit, a shooting unit, a scanning unit and a timing unit; Switch unit obtains quality inspection control sheet and the current period, determining a quality inspection period and a silent period after the current period based on the identifier, turning on the shooting unit and the scanning unit during the quality inspection period, and turning off the shooting unit and the scanning unit during the silent period; The photographing unit photographs an external image of the sample based on infrared triggering; The scanning unit includes a rotatable X-ray scanning device, which is continuously turned on during the quality inspection period and scans at multiple angles to obtain projection data of the sample; The timing unit distinguishes time periods based on the system time, adds time period marks to external images and projection data, and uploads them to the detection module. The value of the time period mark is equal to the vector of the current time period at the runtime. Position in time period For example, the time period is marked as .

[0039] Furthermore, the shooting unit includes an infrared emitting device, a receiving trigger device and a shooting device; The infrared emitting device continuously emits infrared light in a fixed direction toward the receiving trigger device; The receiving trigger device sends a start command to the shooting device when the receiving trigger device does not receive the infrared light or receives the infrared light that is partially blocked; After receiving the start command, the photographing device photographs the external image of the sample in real time.

[0040] Furthermore, the detection module includes a pre-processing unit, an external analysis unit, an internal analysis unit and a decision unit; The pre-processing unit uses a Gaussian filter to denoise the external image and projection data; The external analysis unit is implemented based on the defect analysis network. It determines whether the sample has surface defects based on the external image, assigns an external judgment label and sends it to the judgment unit. The internal analysis unit uses the FBP algorithm to process the projection data and uses the Marching Cubes algorithm to generate a reconstructed 3D model. It performs a coincidence comparison with the standard model, assigns an internal decision label, and sends it to the decision unit. The judgment unit determines whether the sample is qualified based on whether the external judgment label and the internal judgment label of the sample are all 0. If they are all 0, the sample result label is set to 0, indicating that it is a qualified sample. If the internal judgment label is not all 0, the sample result label is set to 1, indicating that it is an unqualified sample, and a time period mark is added to report to the statistical analysis module.

[0041] like Figure 3 As shown, further, the external analysis unit assigns an external decision label based on the defect analysis network, including the following specific steps: Get the External images of samples , use Fast R-CNN for target detection and remove external images The background area in the image is used to obtain the sample image. ; Pass in sequence and Convolution kernel is used to fully explore the sample image Generate 16-channel latent features ; Aggregate latent features using average pooling and maximum pooling The spatial information of the surface defects is enhanced and highlighted from both global and local perspectives to generate an average feature map. and the maximum feature map and combines the average feature map through nonlinear modulation and the maximum feature map Generate dual-view feature maps ,in, and is a learnable linear weight; Ghost convolution is used instead of conventional convolution to further aggregate the 16-channel dual-view feature map , Ghost convolution maps the dual-view feature maps of 16 channels The ghost features are transformed into ghost features through different linear operations respectively, and the ghost features of 16 channels are spliced to generate a ghost feature map. , Ghost convolution can be used to avoid generating redundant features in the convolutional learning process of convolutional neural networks; Two linear layers are used to reduce the dimension of the ghost feature map And further use the Sigmoid activation function to generate the Surface defect probability of a sample , if the surface defect probability If it is greater than or equal to 0.5, The external decision label of the sample is set to 1, indicating that There are surface defects in the samples. If the surface defect probability If it is less than 0.5, it is set to 0, indicating that The samples are in good appearance.

[0042] like Figure 4 As shown, further, the internal analysis unit assigns the internal decision label including the following specific steps: Get the Projection data of samples , assuming the projection data Including projection vectors at different angles, where The projection vector of the angle is recorded as , is the total number of angles; The Hanning filter is selected as the spatial filter of the FBP algorithm. All projection vectors are transformed into frequency domain through FFT, multiplied by the Hanning filter, and then transformed back into time domain through IFFT. The internal image is reconstructed by back-projection. Tomography is used to stack internal images into 3D slices. The Marching Cubes algorithm is used to extract cubes from the 3D slices, separating the cubes that intersect with the isosurface. The intersection points of the isosurface and the cube edges are calculated using interpolation. Based on the relative positions of the cube vertices and the isosurface, the intersection points are connected in a certain manner. Smoothing is then performed to eliminate jagged edges and unevenness, generating a reconstructed 3D model. Import the reconstructed 3D model and the standard model into 3dsMax software in STL file format and calculate the overlap ratio, which refers to the ratio of the number of overlapping triangles in the reconstructed 3D model and the standard model to the number of triangles in the standard model. Perform overlap comparison to determine whether the overlap ratio is greater than or equal to the ratio threshold; If the overlap ratio is greater than or equal to the ratio threshold, the The internal decision label of the sample is set to 0, indicating that The internal structure of each sample is accurate; If the overlap ratio is less than the ratio threshold, it is set to 1, indicating that the The samples had internal defects.

[0043] Furthermore, the statistical analysis module includes a statistical unit and a suspension unit; The statistics unit obtains the sample result label and identifies the added time period mark, assuming that it is the first Time periods, respectively, count the The number of qualified samples and unqualified samples in each period, the number of qualified samples and the number of unqualified samples in the first period, The ratio of the total number of samples inspected in each period is taken as the The qualified rate of each period is used to determine the Is the qualified rate of each period greater than or equal to the qualified threshold? If the qualified rate of each period is greater than or equal to the qualified threshold, the period result label Set to 0, indicating the If the quality inspection of each period is passed, if the qualified rate of the period is less than the qualified threshold, it is set to 1, indicating that the The quality inspection of the time period fails, and the result label of the time period is fed back To the control module; The suspend unit performs statistical judgment and counts the total number of time periods that fail the quality inspection. When the total number of time periods that fail the quality inspection is greater than the total number of quality inspection cycles, If the test result is 0, all products in this batch will be judged as unqualified and the production of this batch will be stopped immediately.

[0044] Example 2 like Figure 5 As shown, the present invention also discloses an artificial intelligence quality detection method for a stationery production line, which is used to implement an artificial intelligence quality detection system for a stationery production line, including the following specific steps: Obtain the production parameters of this batch, set the quality inspection parameters based on the minimum safety criteria, and further initialize the quality inspection control table ; Based on the quality control table Acquire the external image and projection data of the sample during the quality inspection period and add a time period mark; Use defect recognition network to identify surface defects, use FBP algorithm and Marching Cubes algorithm to identify internal defects, assign sample result labels and add time period marks; Based on the sample result labels of all samples in a single quality inspection period, the period qualification rate is calculated and the period qualification label is assigned. Statistical judgment is performed to decide whether to stop the production of this batch. If not, the quality inspection control table is adjusted in real time based on the risk supplementation strategy. .

[0045] Furthermore, the quality inspection parameters include the quality inspection ratio , Quality inspection cycle , Quality inspection start time and unit quality inspection time , setting quality inspection parameters based on minimum safety criteria includes the following specific steps: Get the production parameters of this batch, including the total number of production , production speed , Production start time and product risk level ; Build a quality inspection scale model , based on the total production and product risk level Determine the quality inspection ratio , the specific formula is as follows: , in, The highest product risk level. and is the lower limit ratio and upper limit ratio, quality inspection ratio model Indicates the quality inspection ratio Within a certain range, as the product risk level Improvement and production totals grows larger with the increase of Set the total number of quality inspection cycles , quality inspection starts Production start time One quality inspection cycle, i.e. the time when quality inspection starts , quality inspection must be carried out in each quality inspection cycle; A single quality inspection cycle is further divided into time periods, and the unit duration corresponding to each time period is , The total number of time periods. In each quality inspection cycle, as long as at least one time period is used as a quality inspection period for quality inspection, the quality inspection ratio can be met. .

[0046] Furthermore, initialize the quality control table The specific steps include: Generate runtime vector ,in, The time when quality inspection starts. is the total number of time periods, The unit quality inspection duration; initialization The quality inspection identification vector is initialized to all 0 , the first period of each quality inspection cycle is used as the quality inspection period, and the rest of each quality inspection cycle is The period is used as the silent period, and the start time of the quality inspection period is determined at the running time vector The quality inspection mark vector The identifier of the corresponding position in is set to 1. 、 For example, the runtime vector and quality inspection logo vector The details are as follows: , , That is, within 2 quality inspection cycles, only the period and time period As a quality inspection period; Corresponding splicing running time vector and quality inspection logo vector Generate quality control table .

[0047] Furthermore, obtaining the external image and projection data of the sample and adding time period marks includes the following specific steps: Get the quality control sheet and the current period, determining a quality inspection period and a silent period following the current period based on an identifier in the quality inspection control table; During the quality inspection period, infrared light is continuously emitted in the specified direction. If the infrared light is not received or is partially blocked, it indicates that a sample has passed; An external image of the sample is obtained by using a photographing device, and projection data of the sample is obtained by using an X-ray scanning device; Add the current period mark to external images and projection data; During quiet periods, the camera and X-ray scanner are turned off to reduce system energy consumption.

[0048] Furthermore, the use of a defect recognition network to identify surface defects includes the following specific steps: Get the External images of samples , use Fast R-CNN for target detection and remove external images The background area in the image is used to obtain the sample image. ; Pass in sequence and Convolution kernel is used to fully explore the sample image Generate 16-channel latent features ; Aggregate latent features using average pooling and maximum pooling The spatial information of the surface defects is enhanced and highlighted from both global and local perspectives to generate an average feature map. and the maximum feature map and combines the average feature map through nonlinear modulation and the maximum feature map Generate dual-view feature maps ,in, and is a learnable linear weight; Ghost convolution is used instead of conventional convolution to further aggregate the 16-channel dual-view feature map , Ghost convolution maps the dual-view feature maps of 16 channels The ghost features are transformed into ghost features through different linear operations respectively, and the ghost features of 16 channels are spliced to generate a ghost feature map. , Ghost convolution can be used to avoid generating redundant features in the convolutional learning process of convolutional neural networks; Two linear layers are used to reduce the dimension of the ghost feature map And further use the Sigmoid activation function to generate the Surface defect probability of a sample , if the surface defect probability If it is greater than or equal to 0.5, The external decision label of the sample is set to 1, indicating that There are surface defects in the first sample, otherwise it is set to 0, indicating that the The samples are in good appearance.

[0049] Furthermore, the use of the FBP algorithm and the Marching Cubes algorithm to identify internal defects includes the following specific steps: Get the Projection data of samples , assuming the projection data Including projection vectors at different angles, where The projection vector of the angle is recorded as , is the total number of angles; The Hanning filter is selected as the spatial filter of the FBP algorithm. All projection vectors are transformed into frequency domain representation through FFT, multiplied by the Hanning filter, converted back to time domain representation through IFFT, and the internal image is reconstructed by back projection. Tomography is used to stack internal images into 3D slices. The Marching Cubes algorithm is used to extract cubes from the 3D slices, separating the cubes that intersect with the isosurface. The intersection points of the isosurface and the cube edges are calculated using interpolation. Based on the relative positions of the cube vertices and the isosurface, the intersection points are connected in a certain manner. Smoothing is then performed to eliminate jagged edges and unevenness, generating a reconstructed 3D model. Import the reconstructed 3D model and the standard model into 3dsMax software in STL file format and calculate the overlap ratio, which refers to the ratio of the number of overlapping triangles in the reconstructed 3D model and the standard model to the number of triangles in the standard model. Perform overlap comparison to determine whether the overlap ratio is greater than or equal to the ratio threshold. If so, The internal decision label of the sample is set to 0, indicating that The internal structure of the sample is accurate, otherwise it is set to 1, indicating that the The samples had internal defects.

[0050] Furthermore, the sample result label is used to determine whether the sample is qualified based on whether the external judgment label and the internal judgment label of the sample are all 0. If they are all 0, the sample result label is set to 0, indicating that it is a qualified sample. Otherwise, it is set to 1, indicating that it is an unqualified sample.

[0051] Furthermore, performing statistical discrimination includes the following specific steps: Based on the time period mark identification Time period , get the Time period All sample result labels; Count the first The number of qualified samples and unqualified samples in each period, the number of qualified samples and the number of unqualified samples in the first period, The ratio of the total number of samples inspected in each period is taken as the The pass rate of each period; Judge the Is the qualified rate of the time period greater than or equal to the qualified threshold? If so, the qualified label of the time period is set to 0, indicating that the first If the quality inspection of the time period is passed, otherwise it is set to 1, indicating that the Failed quality inspection during certain time periods; Jordi If the quality inspection of a period fails, the total number of time periods that fail the quality inspection is increased by 1, and the total number of time periods that fail the quality inspection is determined to be greater than the total number of quality inspection cycles. If it is greater than, then all products in this batch are judged to be unqualified and production is stopped immediately. Otherwise, the quality control table is adjusted in real time based on the risk supplement strategy. .

[0052] Furthermore, the quality control table can be adjusted in real time based on the risk-addition strategy. The specific steps include: Get the result label of a single period, assuming it is Time period The results label for the period , ; Judgment period result label Equal to 0 or 1, if the period result label If it is equal to 0, it means that Time period The quality inspection has passed and no adjustment is made; If the period result label If it is equal to 1, it means that Time period If the product fails the quality inspection, obtain the product risk level and the total number of remaining periods ; Product risk level and the total number of remaining periods The smaller value is used as the adjustment period number , will Time period After All time periods are adjusted to quality inspection time periods. If they are already quality inspection time periods, they remain unchanged and the quality inspection control table is updated. The identifier of the corresponding position in is 1, that is, Period to The identifiers of the positions corresponding to the start times of each time period are all set to 1.

[0053] The present invention discloses an artificial intelligence quality inspection system and method for a stationery production line. The parameter setting module obtains production parameters, sets quality inspection parameters based on minimum safety criteria, and replaces inefficient full inspection with efficient and accurate spot inspection. The control module initializes or adjusts the quality inspection control table in real time based on quality inspection parameters or time period result labels to ensure the credibility of spot inspection. The acquisition module obtains external images and projection data of stationery during the quality inspection period. The detection module uses a defect recognition network to identify surface defects, and uses the FBP algorithm and the Marching Cubes algorithm to identify internal defects, and accurately determines whether the stationery is qualified based on the internal and external dual perspectives of the stationery, thereby improving the quality inspection accuracy. The statistical analysis module counts the quality inspection results of samples within a single quality inspection period, generates a time period result label, and independently determines whether to stop the production of this batch and stop losses in time, thereby realizing high-accuracy, high-reliability, high-efficiency and low-cost automated quality inspection of the stationery production line.

[0054] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An artificial intelligence quality inspection system for stationery production lines, characterized in that: It includes parameter setting module, control module, acquisition module, detection module and statistical analysis module; The parameter setting module sets the quality inspection parameters based on the minimum safety criteria and the production parameters of this batch; The control module initializes the quality inspection control table based on the quality inspection parameters or adjusts the quality inspection control table in real time based on the time period result label; The acquisition module identifies the quality inspection period based on the quality inspection control table, acquires external images and projection data during the quality inspection period, adds a period mark and uploads it to the detection module; The detection module uses a defect recognition network to identify surface defects, and uses the FBP algorithm and Marching Cubes algorithm to identify internal defects, assigning sample result labels and adding time period marks; The statistical analysis module counts all sample result labels of a single quality inspection period, generates a period result label and feeds back to the control module, which independently decides whether to stop the production of this batch.

2. The artificial intelligence quality inspection system for a stationery production line according to claim 1, characterized in that: The control module includes an initialization unit and a real-time adjustment unit; The initialization unit obtains quality inspection parameters, initializes and generates a quality inspection control table, and sends it to the data acquisition module; The real-time adjustment unit obtains the period result label of a single period, updates the quality inspection control table based on the risk addition strategy, and sends the updated quality inspection control table to the data acquisition module.

3. The artificial intelligence quality inspection system for a stationery production line according to claim 2, characterized in that: The initialization unit generates the quality inspection control table by initialization, which includes the following specific steps: Generate a run time vector based on quality inspection parameters; Initialize and generate a quality inspection identification vector of all 0s, select a quality inspection period, and set the identifier of the corresponding position in the quality inspection identification vector at the start time of each quality inspection period to 1; A quality inspection control table is generated by correspondingly concatenating the running time vector and the quality inspection identification vector.

4. The artificial intelligence quality inspection system for a stationery production line according to claim 2, characterized in that: The real-time adjustment unit updates the quality inspection control table based on the risk addition strategy, including the following specific steps: Get the period result tag for a single period and decide whether to adjust based on whether the period result tag is equal to 0 or 1; If the period result label is equal to 1, the smaller value of the product risk level and the total number of remaining periods is used as the adjustment period number; After adjusting a single period, the period whose quantity is equal to the number of adjusted periods is the quality inspection period, and the identifier of the corresponding position in the quality inspection control table is updated to 1.

5. The artificial intelligence quality inspection system for a stationery production line according to claim 1, characterized in that: The detection module includes a pre-processing unit, an external analysis unit, an internal analysis unit and a decision unit; The pre-processing unit uses a Gaussian filter to perform denoising on the external image and projection data; The external analysis unit assigns an external decision label based on the defect analysis network; The internal analysis unit generates a reconstructed three-dimensional model using the FBP algorithm and the Marching Cubes algorithm, and performs a coincidence comparison with the standard model to assign an internal decision label; The decision unit determines whether the sample is qualified based on the external decision label and the internal decision label, assigns a sample result label and adds a time period mark to report to the statistical analysis module.

6. The artificial intelligence quality inspection system for a stationery production line according to claim 5, characterized in that: The external analysis unit assigns an external decision label based on the defect analysis network, including the following specific steps: Obtain an external image of a single sample and use Fast R-CNN to obtain the sample image; Generate 16-channel potential features through two convolutions; Average pooling and maximum pooling are used to aggregate potential features to generate average feature maps and maximum feature maps, which are then combined through nonlinear modulation to generate a 16-channel dual-view feature map. Ghost convolution is used to transform the dual-view feature maps of each channel into ghost features, and then spliced to generate ghost feature maps; Two linear layers are used to reduce the dimension of the ghost feature map and the Sigmoid activation function is used to generate the surface defect probability of a single sample. The external decision label is assigned as 0 or 1 based on the surface defect probability.

7. The artificial intelligence quality inspection system for a stationery production line according to claim 5, characterized in that: The internal analysis unit generates an internal decision label including the following specific steps: Acquire projection data of a single sample and select the spatial filter of the FBP algorithm; All projection vectors in the projection data are FFT-transformed and multiplied by the spatial filter, and the internal image is reconstructed using IFFT transformation and back-projection; Tomography is used to stack internal images into 3D slices, and the Marching Cubes algorithm is used to generate a reconstructed 3D model, and the overlap ratio with the standard model is calculated; Perform overlap comparison and assign the internal decision label of a single sample to 0 or 1 based on the relationship between the overlap ratio and the ratio threshold.

8. The artificial intelligence quality inspection system for a stationery production line according to claim 1, characterized in that: The parameter determination module determines the quality inspection parameters including the following specific steps: Get the production parameters of this batch; Construct a quality inspection ratio model to describe the relationship between total production volume and product risk level and quality inspection ratio, and determine the quality inspection ratio; Set the total number of quality inspection cycles and the start time of quality inspection; Set the total number of time periods to further divide a single quality inspection cycle into multiple time periods.

9. An artificial intelligence quality inspection method for a stationery production line, used to implement an artificial intelligence quality inspection system for a stationery production line as claimed in any one of claims 1 to 8, characterized in that: The specific steps include: Obtain the production parameters for this batch, set quality inspection parameters based on minimum safety criteria, and initialize the quality inspection control table; Based on the quality inspection control table, obtain the external image and projection data of the sample during the quality inspection period and add the time period mark; Use defect recognition network to identify surface defects, use FBP algorithm and Marching Cubes algorithm to identify internal defects, assign sample result labels and add time period marks; Based on the sample result labels within a single quality inspection period, the period qualification rate is calculated, the period qualification label is assigned, and statistical judgment is performed to decide whether to stop the production of this batch or further determine whether to adjust the quality inspection control table in real time.

10. The artificial intelligence quality inspection method for a stationery production line according to claim 9, characterized in that: The method of obtaining the external image and projection data of the sample and adding time period marks includes the following specific steps: Determine the quality inspection period and silent period after the current period based on the identifier in the quality inspection control table; During the quality inspection period, infrared light is emitted, and based on the infrared light reception result, a decision is made as to whether to activate the photographing device and the X-ray scanning device to obtain external images and projection data of the sample; Add the current period mark to external images and projection data; During quiet periods, turn off the camera and X-ray scanner; The sample result label is based on whether the external decision label and the internal decision label of the sample are all 0 to determine whether the sample is qualified. If all are 0, the sample result label is set to 0; if not all are 0, it is set to 1. Further, performing statistical discrimination includes the following specific steps: Obtain all sample result labels for a single period and calculate the qualified rate for that period; The qualified label of a time period is set to 0 or 1 based on the relationship between the qualified rate of the time period and the qualified threshold; If set to 1, the total number of time periods that failed quality inspection is updated and further compared with the total number of quality inspection cycles; If it is greater than the total number of quality inspection cycles, production will be stopped. If it is less than or equal to the total number of quality inspection cycles, it will be determined whether to adjust the quality inspection control table in real time.

Images