Automobile part abnormity monitoring method and system based on AI visual inspection

Abstract

The invention provides an automobile part abnormity monitoring method and system based on AI visual inspection, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring an original part production video of a target automobile part, segmenting the original part production video according to a corresponding production link, and outputting a plurality of segmented part production sub-videos corresponding to the target automobile part; performing video frame optimization on each segmented part production sub-video, and outputting a plurality of optimized part production sub-videos corresponding to the target automobile part; and performing anomaly identification on each optimized part production sub-video by using the updated target neural network, and outputting a plurality of part anomaly data corresponding to the target automobile part. Based on the above content, the reliability of abnormal monitoring of the automobile parts can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for monitoring abnormalities of automobile parts based on AI visual detection. Background Art

[0002] AI visual inspection means that after the video data of the monitored object is determined based on visual inspection technology, AI (artificial intelligence) technology can be used to analyze the video data. For example, in a specific application scenario, AI technology can be used to perform abnormal analysis on the production video of automotive parts to obtain corresponding abnormal data. However, in the existing technology, there is a problem that the reliability of abnormal monitoring of automotive parts is relatively low. Summary of the invention

[0003] In view of this, an object of the present invention is to provide a method and system for abnormality monitoring of automobile parts based on AI visual detection, so as to improve the reliability of abnormality monitoring of automobile parts.

[0004] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: A method for monitoring abnormality of automobile parts based on AI visual detection, the method comprising: Acquire an original component production video of a target automotive component, and segment the original component production video according to corresponding production links, and output a plurality of segmented component production sub-videos corresponding to the target automotive component, wherein the plurality of original component production video frames included in the original component production video are obtained by monitoring a production line of the target automotive component; Performing video frame optimization on each of the segmented parts production sub-videos respectively, and outputting a plurality of optimized parts production sub-videos corresponding to the target automotive parts, wherein the video frame optimization includes optimizing at least one original parts production video frame of a subsequent segmented parts production sub-video based on an image connected domain according to at least one original parts production video frame of a previous segmented parts production sub-video, to form at least one optimized parts production sub-video frame; Using the updated target neural network, each of the optimized component production sub-videos is respectively identified for abnormality, and a plurality of component abnormality data corresponding to the target automobile component is output, wherein the component abnormality data is used to reflect the abnormal type of the target automobile component.

[0005] In some preferred embodiments, in the above-mentioned auto parts abnormality monitoring method based on AI visual inspection, the step of optimizing the video frames of each of the segmented parts production sub-videos and outputting a plurality of optimized parts production sub-videos corresponding to the target auto parts includes: For any one of the plurality of segmented component production sub-videos, the following processing is performed: For each original component production video frame in the segmented component production sub-video, the image connected domain of the original component production video frame is identified to obtain a set of image connected domains corresponding to the original component production video frame, and the number of image connected domains in the set of image connected domains corresponding to each original component production video frame in the segmented component production sub-video is discretely calculated to output a distribution parameter of the connected domain corresponding to the segmented component production sub-video; Based on the connected domain distribution parameter, a target number of original component production video frames ranked first are selected from the segmented component production sub-videos, and there is a negative correlation between the target number and the connected domain distribution parameter; In the segmented parts production sub-video, based on the inter-frame change information between multiple original parts production video frames in the previous segmented parts production sub-video of the segmented parts production sub-video, video frame interpolation is performed on the target number of original parts production video frames to output an optimized parts production sub-video corresponding to the target automotive part.

[0006] In some preferred embodiments, in the above-mentioned auto parts abnormality monitoring method based on AI visual inspection, the step of using the target neural network after the update operation to perform abnormality identification on each of the optimized parts production sub-videos and outputting a plurality of parts abnormality data corresponding to the target auto parts includes: For any one of the plurality of optimized component production sub-videos, the following processing is performed: A typical component production video and a plurality of pending component abnormality data are extracted, wherein the typical component production video has a plurality of typical component abnormality data, and the typical component abnormality data is used to reflect the component abnormality information actually possessed by the typical component production video; Determine the abnormal features of the pending parts of each of the abnormal data of the pending parts by using a feature space mapping network; Based on the abnormal features of the pending parts, the plurality of abnormal data of the pending parts are classified to form a plurality of corresponding abnormal data classification sets of the parts, and then the classification type data of each abnormal data classification set of the parts is obtained according to the typical abnormal data of the parts; The typical parts production video is loaded into a parts abnormality analysis network, and a parts abnormality data target set is analyzed from the multiple parts abnormality data classification sets by using the parts abnormality analysis network, and pending parts abnormality data matching the typical parts production video is screened out from the parts abnormality data target set; Based on the possibility distribution data of the target set of component abnormality data and the classification type data, calculate and output the first dimension Loss value; Based on the possibility distribution data of the abnormal data of the pending parts matched with the typical parts production video and the abnormal data of the typical parts, a Loss value of the second dimension is calculated and outputted; Based on the Loss value of the first dimension and the Loss value of the second dimension, the feature space mapping network and the component abnormality analysis network are updated, and the abnormal features of the pending component are adjusted during the network update process; The optimized parts production sub-video is loaded into the updated parts anomaly analysis network, and the updated parts anomaly analysis network is used to analyze the parts anomaly data of the optimized parts production sub-video. The updated parts anomaly analysis network belongs to the target neural network.

[0007] In some preferred embodiments, in the above-mentioned method for monitoring abnormality of automobile parts based on AI visual inspection, the step of using a feature space mapping network to determine the abnormal features of the pending parts of each of the abnormal data of the pending parts includes: According to a pre-configured data splitting rule, the pending component abnormality data is split, and at least one corresponding component abnormality data segment is output; Loading the component abnormal data fragment into a feature space mapping network to determine the component abnormal fragment feature corresponding to the component abnormal data fragment; For any of the pending component abnormal data, all of the component abnormal fragment features corresponding to the pending component abnormal data are feature fused to output the pending component abnormal features corresponding to the pending component abnormal data.

[0008] In some preferred embodiments, in the above-mentioned automobile parts abnormality monitoring method based on AI visual inspection, the parts abnormality analysis network includes a feature mining subnetwork and an abnormality analysis subnetwork; The steps of loading the typical parts production video into the parts abnormality analysis network, analyzing the parts abnormality data target set from the multiple parts abnormality data classification sets using the parts abnormality analysis network, and screening out the pending parts abnormality data that matches the typical parts production video from the parts abnormality data target set include: Loading the typical parts production video into the feature mining sub-network to mine typical parts production features corresponding to the typical parts production video; Loading the typical component production features into the abnormality analysis subnetwork, and analyzing the first matching parameters of each component abnormality data classification set; Perform parameter interval mapping operations on each of the first matching parameters, and output a first classification possibility of each of the component abnormal data classification sets; If the first classification possibility exceeds the pre-configured first reference classification possibility, marking the component abnormality data classification set corresponding to the first classification possibility so as to be marked as the component abnormality data target set; In the target set of abnormal component data, the undetermined abnormal component data matching the typical component production video is screened out.

[0009] In some preferred embodiments, in the above-mentioned method for monitoring abnormalities of automobile parts based on AI visual inspection, the step of loading the typical parts production video into the feature mining subnetwork to mine the typical parts production features corresponding to the typical parts production video includes: Segmenting the typical parts production video to form a corresponding parts production video segment set, wherein the parts production video segment set includes a plurality of parts production video segments; The front position of the component production video clip set is configured with a start symbol, and the last position of the component production video clip set is configured with an end symbol, so as to form a corresponding symbol production video clip set; Mapping the symbol production video segment set to a feature space of the segment, and outputting a corresponding production video segment feature set; According to the associated focused feature analysis rules, the production video clip feature set is subjected to focused feature analysis through the feature mining sub-network, and a corresponding focused production video clip feature set is output, wherein the focused production video clip feature set includes focused production video clip features of each component production video clip in the symbol production video clip set; According to the focus feature analysis rules for itself, the features of each of the focus production video clips are aggregated to output the typical parts production features corresponding to the typical parts production video.

[0010] In some preferred embodiments, in the above-mentioned method for monitoring abnormalities of automobile parts based on AI visual inspection, the step of aggregating the features of each of the focused production video segments according to the focus feature analysis rules for itself and outputting the typical parts production features corresponding to the typical parts production video includes: Analyze the focus evaluation parameters corresponding to each of the focus production video segment features according to the focus feature analysis rules for itself and the focus production video segment features; Analyzing the focus importance parameters of each of the focus production video segment features based on the pre-configured parameter interval mapping rules and the focus evaluation parameters; According to the corresponding focus importance parameters, weighted superposition operations are performed on the features of each focus production video segment to output typical component production features corresponding to the typical component production video.

[0011] In some preferred embodiments, in the above-mentioned automobile parts abnormality monitoring method based on AI visual inspection, the possibility distribution data of the target set of abnormal data of the parts includes the first classification possibility of each classification set of the abnormal data of the parts; The step of calculating and outputting the first dimension Loss value based on the possibility distribution data of the component abnormality data target set and the classification type data comprises: Analyzing the target type possibility of each of the component abnormal data classification sets based on the classification type data of each of the component abnormal data classification sets; Determine the Loss value between each target type possibility and the corresponding first classification possibility, and output corresponding multiple type analysis Loss values; The sum of all the Loss values ​​of the type analysis is calculated to obtain the Loss value of the first dimension.

[0012] In some preferred embodiments, in the above-mentioned method for monitoring abnormalities of automobile parts based on AI visual inspection, the step of screening out the abnormal data of pending parts that match the typical parts production video in the target set of abnormal data of parts includes: Polling each of the pending component abnormality data in the component abnormality data target set, and analyzing feature-related parameters between the typical component production features and the pending component abnormality features of the pending component abnormality data; Perform parameter interval mapping operation on the feature-related parameters, and output the second classification possibility corresponding to the abnormal data of the pending component; If the second classification possibility exceeds the pre-configured second reference classification possibility, marking the pending component abnormality data corresponding to the second classification possibility so that the pending component abnormality data marked as matching the typical component production video; Furthermore, the possibility distribution data of the pending component abnormality data matched with the typical component production video includes the second classification possibility of each of the pending component abnormality data in the component abnormality data target set; The step of calculating and outputting a second dimension Loss value based on the possibility distribution data of the pending component abnormality data matched with the typical component production video and the typical component abnormality data comprises: Based on the typical component abnormality data, analyzing the target matching possibility of each of the pending component abnormality data in the component abnormality data target set; Determine a Loss value between each of the target matching possibilities and the corresponding second classification possibility, and output a plurality of corresponding matching analysis Loss values; The sum of all the matching analysis Loss values ​​is calculated to obtain the second dimension Loss value.

[0013] An embodiment of the present invention also provides an automobile parts abnormality monitoring system based on AI visual inspection, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above-mentioned automobile parts abnormality monitoring method based on AI visual inspection.

[0014] The method and system for monitoring abnormality of automobile parts based on AI visual detection provided by the embodiment of the present invention can first obtain the original production video of the target automobile parts, and, according to the corresponding production link, segment the original production video, and output multiple segmented production sub-videos corresponding to the target automobile parts; optimize the video frame of each segmented production sub-video, and output multiple optimized production sub-videos corresponding to the target automobile parts; use the target neural network after the update operation to identify the abnormality of each optimized production sub-video, and output multiple abnormal data of parts corresponding to the target automobile parts. Based on the above content, since the multiple sub-videos formed by the segmentation will be processed when the abnormality is identified, the granularity of the identification will be finer, thereby improving the reliability of abnormal monitoring of automobile parts to a certain extent, that is, improving the problem of low reliability of abnormal monitoring caused by redundant information interference between different production links; in addition, since at least one original production video frame of the previous segmented production sub-video is also used, at least one original production video frame of the next segmented production sub-video is associated and optimized, so that the information content of the sub-video of abnormal identification is richer, and the reliability of abnormal monitoring of automobile parts can be improved to a certain extent.

[0015] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a structural block diagram of an automotive parts abnormality monitoring system based on AI visual inspection provided by an embodiment of the present invention.

[0017] Figure 2 A flowchart of the steps of the method for abnormality monitoring of automobile parts based on AI visual inspection provided in an embodiment of the present invention.

[0018] Figure 3 A schematic diagram of various modules included in the automotive parts abnormality monitoring device based on AI visual detection provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, an embodiment of the present invention provides an automobile parts abnormality monitoring system based on AI visual detection. The automobile parts abnormality monitoring system based on AI visual detection may include a memory and a processor.

[0022] In detail, the memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, they can be electrically connected to each other through one or more communication buses or signal lines. The memory can store at least one software function module (computer program) that can exist in the form of software or firmware. The processor can be used to execute the executable computer program stored in the memory, thereby realizing the abnormal monitoring method of automobile parts based on AI visual detection provided by an embodiment of the present invention.

[0023] For example, in some embodiments, the memory may be, but not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. For example, in some embodiments, the processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), a system on chip (SoC), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0024] For example, in some implementations, the automotive parts abnormality monitoring system based on AI visual inspection can be a server with data processing capabilities.

[0025] Combination Figure 2 The embodiment of the present invention also provides an abnormality monitoring method for automobile parts based on AI visual detection, which can be applied to the abnormality monitoring system for automobile parts based on AI visual detection. The method steps defined in the process related to the abnormality monitoring method for automobile parts based on AI visual detection can be implemented by the abnormality monitoring system for automobile parts based on AI visual detection.

[0026] The following will Figure 2 The specific process shown is explained in detail.

[0027] Step S110, obtaining an original component production video of a target automobile component, and segmenting the original component production video according to corresponding production links, and outputting a plurality of segmented component production sub-videos corresponding to the target automobile component.

[0028] In an embodiment of the present invention, the automobile parts abnormality monitoring system based on AI visual detection can obtain the original parts production video of the target automobile parts, and, according to the corresponding production links, segment the original parts production video, and output multiple segmented parts production sub-videos corresponding to the target automobile parts. Exemplarily, one segmented parts production sub-video can correspond to one production link. The multiple original parts production video frames included in the original parts production video are obtained by monitoring the production line of the target automobile parts. For example, the production line of the target automobile parts can be monitored by video monitoring equipment. The target automobile parts can be automobile wheels, automobile bearings and other parts. For different automobile parts, the corresponding production links can be different. For example, for automobile wheels, its production links can be simply summarized as pouring aluminum alloy solution into a mold to make a blank, cooling to room temperature, lathe processing and polishing (also including surface cleaning, etc.), painting, etc.

[0029] Step S120, performing video frame optimization on each of the segmented component production sub-videos, and outputting a plurality of optimized component production sub-videos corresponding to the target automotive component.

[0030] In an embodiment of the present invention, the automobile parts abnormality monitoring system based on AI visual detection can optimize the video frames of each of the split parts production sub-videos, and output multiple optimized parts production sub-videos corresponding to the target automobile parts, wherein the video frame optimization includes optimizing at least one original parts production video frame of the previous split parts production sub-video based on the image connected domain to form at least one optimized parts production sub-video frame. Exemplarily, the front-to-back relationship between the split parts production sub-videos can refer to the front-to-back relationship between the corresponding production links, or the front-to-back relationship between the corresponding video formation times.

[0031] Step S130, using the target neural network that has undergone the update operation, respectively identify abnormalities for each of the optimized component production sub-videos, and output a plurality of component abnormality data corresponding to the target automotive component.

[0032] In an embodiment of the present invention, the automobile parts anomaly monitoring system based on AI visual inspection can use the target neural network that has undergone an update operation to perform anomaly identification on each of the optimized parts production sub-videos, and output multiple parts anomaly data corresponding to the target automobile parts. The parts anomaly data is used to reflect the abnormal type of the target automobile parts; illustratively, for the above-mentioned "polished by lathe", the anomaly type may be burrs and scratches on the polished surface, and for the above-mentioned "spray painting", the anomaly type may be different paint thickness, paint leakage, etc. In addition, since the target neural network is formed through an update operation, that is, during the update operation, it can learn the mapping relationship between different videos and corresponding anomaly types, therefore, anomaly identification can be performed in the application based on the mapping relationship.

[0033] Based on the foregoing content, since the multiple sub-videos formed by segmentation will be processed during anomaly recognition, the granularity of recognition will be finer, thereby improving the reliability of abnormality monitoring of automotive parts to a certain extent, that is, improving the problem of low reliability of abnormality monitoring caused by redundant information interference between different production links; in addition, since at least one original part production video frame of a subsequent segmented part production sub-video will be associated and optimized based on at least one original part production video frame of a previous segmented part production sub-video, the amount of information in the sub-video for abnormality recognition is richer (that is, the association of different production links is realized), the reliability of abnormality monitoring of automotive parts can also be improved to a certain extent.

[0034] For example, in some implementations, the step of optimizing the video frames of each of the segmented component production sub-videos and outputting a plurality of optimized component production sub-videos corresponding to the target automotive component, i.e., the above-mentioned step S120, may include: For any one of the plurality of segmented component production sub-videos, the following processing is performed (each of the segmented component production sub-videos can be processed in parallel or in sequence): For each original component production video frame in the segmented component production sub-video, the image connected domain of the original component production video frame is identified to obtain the image connected domain set corresponding to the original component production video frame, and the number of image connected domains in the image connected domain set corresponding to each original component production video frame in the segmented component production sub-video is discretely calculated, and the connected domain distribution parameter corresponding to the segmented component production sub-video is output. The discrete calculation process may include first calculating the mean of the number of image connected domains in the image connected domain set corresponding to each original component production video frame, then calculating the absolute difference between the number of image connected domains in the image connected domain set corresponding to each original component production video frame and the mean, and performing sum calculation on each absolute difference to obtain a target sum value. Finally, the ratio between the target sum value and the number of original component production video frames can be calculated. In this way, the connected domain distribution parameter corresponding to the segmented component production sub-video can be obtained. Therefore, the connected domain distribution parameter can characterize the distribution discreteness of the image connected domain corresponding to each original component production video frame. The larger the value, the greater the information amount of the sub-video as a whole (the difference between video frames is large). Based on the connected domain distribution parameter, a target number of original component production video frames ranked first are selected from the segmented component production sub-videos, and there is a negative correlation between the target number and the connected domain distribution parameter; In the segmented parts production sub-video, based on the inter-frame change information between multiple original parts production video frames in the previous segmented parts production sub-video of the segmented parts production sub-video, the target number of original parts production video frames are interpolated to output an optimized parts production sub-video corresponding to the target automotive parts; illustratively, the mean of the inter-frame difference can be calculated for the target number of original parts production video frames sorted later in the previous segmented parts production sub-video of the segmented parts production sub-video to obtain the mean inter-frame difference. The larger the mean inter-frame difference, the larger the number of video frames that can be inserted in the process of video frame interpolation; the smaller the mean inter-frame difference, the smaller the number of video frames that can be inserted in the process of video frame interpolation. The specific interpolation method can be based on relevant existing technologies and is not specifically limited here; in addition, the inter-frame difference can refer to the average value of the absolute difference of pixel values ​​corresponding to pixel points between two adjacent video frames.

[0035] For example, in some embodiments, the step of using the updated target neural network to perform abnormality identification on each of the optimized component production sub-videos and outputting a plurality of component abnormality data corresponding to the target automobile component, i.e., the above step S130, may include: For any one of the plurality of optimized component production sub-videos, the following processing is performed: A typical component production video and a plurality of pending component abnormality data are extracted, wherein the typical component production video has a plurality of typical component abnormality data, and the typical component abnormality data is used to reflect the component abnormality information actually possessed by the typical component production video, and the component abnormality information actually possessed by the typical component production video may be formed by manual annotation, and the typical component production video may be a historical component production video, which serves as a basis for network update; Determine the abnormal features of the pending parts of each of the abnormal data of the pending parts by using the feature space mapping network; that is, the abnormal data of the pending parts can be mapped to the feature space to obtain the abnormal features of the pending parts represented by vectors; Based on the abnormal features of the pending parts, the multiple abnormal data of the pending parts are classified to form multiple corresponding classification sets of abnormal data of the parts, and then the classification type data of each classification set of abnormal data of the parts is obtained based on the typical abnormal data of the parts; illustratively, the abnormal data of the pending parts with a smaller characteristic distance (such as cosine distance) between the abnormal features of the corresponding pending parts can be allocated to the same classification set of abnormal data of the parts, and the abnormal data of the pending parts with a larger characteristic distance (such as cosine distance) between the abnormal features of the corresponding pending parts can be allocated to different classification sets of abnormal data of the parts. In the specific classification process, the relevant prior art can be referred to. In this way, multiple corresponding classification sets of abnormal data of the parts can be formed. For example, the abnormal data in different production links can be allocated to different abnormal data of the parts. A common data classification set, abnormal data in the same production link can be assigned to the same component abnormal data classification set; wherein the classification type data is used to determine the correlation between the component abnormal data classification set and the typical component production video, therefore, for each component abnormal data classification set, the corresponding classification type data can be determined based on the matching relationship between the pending component abnormal data included in the component abnormal data classification set and the typical component abnormal data (such as whether they represent the same abnormal type); illustratively, the classification type data of the component abnormal data classification set that matches any typical component abnormal data can be configured as a first value, and the classification type data of the component abnormal data classification set that does not match any typical component abnormal data can be configured as a second value, and the second value is different from the first value; The typical component production video is loaded into the component anomaly analysis network, and the component anomaly analysis network is used to analyze the component anomaly data target set in the multiple component anomaly data classification sets, and the pending component anomaly data matching the typical component production video is screened out in the component anomaly data target set; that is, matching can be performed first in the classification set dimension, and then matching can be performed on the specific component anomaly data, that is, two-level matching including coarse-grained and fine-grained matching, so that the reliability of classification is higher; that is, the component anomaly data target set can be determined in all the component anomaly data classification sets obtained by classification according to the coarse-grained distribution of the component anomaly analysis network, and specifically, the component anomaly data classification set with a high correlation with the typical component production video can be determined in all the component anomaly data classification sets, and then the pending component anomaly data matching the typical component production video is determined in each pending component anomaly data in the component anomaly data target set according to the output of the component anomaly analysis network, so that the hierarchical classification of the component anomaly analysis network can be realized, and the complexity of the component anomaly analysis network can be reduced, thereby improving the operation efficiency of the component anomaly analysis network; Based on the possibility distribution data of the target set of component abnormal data and the classification type data, the first dimension Loss value is calculated and output, that is, the error of the first level is calculated; Based on the possibility distribution data of the abnormal data of the undetermined parts matched with the typical parts production video and the abnormal data of the typical parts, the output second dimension Loss value is calculated, that is, the error of the second level is calculated; Based on the Loss value of the first dimension and the Loss value of the second dimension, the feature space mapping network and the component abnormality analysis network are updated, and the abnormal features of the pending component are adjusted during the network update process; illustratively, the Loss value of the first dimension and the Loss value of the second dimension can be weighted and summed to obtain a target Loss value, and then, based on the goal of reducing the target Loss value, the feature space mapping network and the component abnormality analysis network can be updated, that is, the included network parameters are updated; in addition, by updating the weighted coefficient between the Loss value of the first dimension and the Loss value of the second dimension, the balance of the classification of the two levels can be adjusted, thereby improving the fitting effect of the feature space mapping network and the component abnormality analysis network; in addition, by adjusting the abnormal features of the pending component during the network update process, the reliability of the abnormal features of the pending component is higher, and it can be directly used in the subsequent process, thereby improving the calculation efficiency; The optimized parts production sub-video is loaded into the updated parts anomaly analysis network, and the updated parts anomaly analysis network is used to analyze the parts anomaly data of the optimized parts production sub-video. The updated parts anomaly analysis network belongs to the target neural network. Since the updated parts anomaly analysis network has a higher precision input-output mapping relationship, the optimized parts production sub-video can be reliably mapped to obtain reliable parts anomaly data.

[0036] For example, in some implementations, the step of using a feature space mapping network to determine the abnormal features of the pending components of each of the abnormal data of the pending components may include: According to the pre-configured data splitting rule, the pending component abnormality data is split and at least one corresponding component abnormality data segment is output; illustratively, the pending component abnormality data may be a description statement of the component abnormality, and therefore, the words of the description statement may be segmented and extracted to obtain at least one component abnormality description word, and each component abnormality description word may be used as a component abnormality data segment, based on which at least one component abnormality data segment corresponding to the pending component abnormality data may be obtained; Loading the component abnormality data fragment into a feature space mapping network to determine the component abnormality fragment feature corresponding to the component abnormality data fragment, wherein the feature space mapping network may be an Embedding network, which is used to convert discrete variables into continuous vectors; For any of the pending component abnormal data, all of the component abnormal fragment features corresponding to the pending component abnormal data are feature fused to output the pending component abnormal features corresponding to the pending component abnormal data; illustratively, all of the component abnormal fragment features corresponding to the pending component abnormal data can be spliced, superimposed or averaged to obtain the pending component abnormal features corresponding to the pending component abnormal data.

[0037] For example, in some implementations, the component anomaly analysis network includes a feature mining subnetwork and an anomaly analysis subnetwork. Based on this, the steps of loading the typical component production video into the component anomaly analysis network, analyzing a component anomaly data target set from the multiple component anomaly data classification sets using the component anomaly analysis network, and screening out pending component anomaly data that matches the typical component production video from the component anomaly data target set may include: The typical parts production video is loaded into the feature mining sub-network to mine the typical parts production features corresponding to the typical parts production video; that is, the features in the typical parts production video can be mined and represented in the form of vectors to obtain the corresponding typical parts production features, which can facilitate subsequent further processing; The typical component production features are loaded into the abnormality analysis subnetwork, and the first matching parameters of each component abnormality data classification set are analyzed. For example, the central features (such as cluster centers or mean features) of the undetermined component abnormality features corresponding to the undetermined component abnormal data included in each component abnormality data classification set can be calculated respectively. Then, the central features of each component abnormality data classification set and the typical component production features can be calculated for feature similarity, such as quantitative product, to obtain the corresponding first matching parameters. Perform parameter interval mapping operations on each of the first matching parameters, and output the first classification possibility of each of the component abnormal data classification sets; illustratively, the sum of each of the first matching parameters can be calculated, and then, for each component abnormal data classification set, the ratio between the first matching parameter corresponding to the component abnormal data classification set and the sum is calculated, so as to obtain the first classification possibility of the component abnormal data classification set; If the first classification possibility exceeds the pre-configured first reference classification possibility, the component abnormality data classification set corresponding to the first classification possibility is marked so as to be marked as the component abnormality data target set; illustratively, the specific value of the first reference classification possibility is not limited, for example, it can be 0.5, 0.6, 0.65, etc., or the component abnormality data classification set with the largest corresponding first classification possibility can also be marked so as to be marked as the component abnormality data target set; In the target set of component abnormality data, the pending component abnormality data that matches the typical component production video is screened out; illustratively, the feature similarity between the pending component abnormality feature corresponding to each pending component abnormality data in the target set of component abnormality data and the typical component production feature of the typical component production video can be calculated respectively, and then, the pending component abnormality data with the largest feature similarity can be determined as the pending component abnormality data that matches the typical component production video.

[0038] For example, in some implementations, the step of loading the typical parts production video into the feature mining subnetwork to mine typical parts production features corresponding to the typical parts production video may include: The typical parts production video is segmented to form a corresponding parts production video clip set, wherein the parts production video clip set includes a plurality of parts production video clips. Exemplarily, the specific manner of segmenting the video is not limited. For example, segmentation may be performed based on similarity between video frames, such as allocating two adjacent video frames with smaller similarity between video frames to two adjacent parts production video clips. Alternatively, segmentation may be performed according to a preset number of video frames, that is, a corresponding parts production video clip is formed based on each adjacent preset number of video frames. The front position of the component production video clip set is configured with a start symbol, and the last position of the component production video clip set is configured with an end symbol, so as to form a corresponding symbol production video clip set, wherein the start symbol and the end symbol may not have any semantic meaning and are only used to distinguish the start and end of the component production video clip; The symbol production video clip set is subjected to feature space mapping of the clips, and a corresponding production video clip feature set is output; that is, each component production video clip in the symbol production video clip set can be subjected to feature space mapping, and represented in the form of a vector, that is, the corresponding production video clip features are obtained, and the production video clip feature set is formed by combining; According to the associated focused feature analysis rules, the production video clip feature set is subjected to focused feature analysis through the feature mining sub-network, and a corresponding focused production video clip feature set is output, wherein the focused production video clip feature set includes focused production video clip features of each component production video clip in the symbol production video clip set; that is, based on other associated production video clip features in the production video clip feature set, associated focused feature analysis can be performed on each production video clip feature, so that the formed focused production video clip feature has higher information representation capability and integrates the semantics of other associated production video clip features; According to the focus feature analysis rules for itself, the features of each of the focused production video clips are aggregated to output the typical parts production features corresponding to the typical parts production video. According to the focus feature analysis rules for itself, its own features can be further mined, so that the characterization ability of the obtained typical parts production features is better.

[0039] For example, in some implementations, the step of performing focused feature analysis on the production video segment feature set through the feature mining sub-network based on the associated focused feature analysis rule and outputting the corresponding focused production video segment feature set may include: The production video clip feature set is serialized to form a production video clip feature sequence, in which the production video clip features are arranged in a sequential relationship between the corresponding parts production video clips; Based on the order in the production video segment feature sequence, polling each production video segment feature to form a currently polled production video segment feature; When the production video clip feature currently polled belongs to the first production video clip feature in the production video clip feature sequence, a linear mapping operation is performed on the production video clip feature currently polled (such as multiplying it with a weighting parameter, adding it with a superposition parameter, normalizing it, etc.), and the corresponding linear mapping feature is output, and a scalar product is calculated for the transposed feature of the production video clip feature currently polled and the linear mapping feature, and then based on the calculated scalar product, the production video clip feature currently polled is weighted to obtain an initial focused production video clip feature corresponding to the production video clip feature currently polled (or it can also be directly used as the focused production video clip feature); When the production video clip feature currently polled does not belong to the first production video clip feature in the production video clip feature sequence, a linear mapping operation is performed on the production video clip feature currently polled, and the corresponding linear mapping feature is output, and a transposed feature of the initial focused production video clip feature corresponding to the previous production video clip feature and the linear mapping feature are calculated for a scalar product, and then based on the calculated scalar product, the production video clip feature currently polled is weighted to obtain the initial focused production video clip feature corresponding to the production video clip feature currently polled (or it can also be directly used as the focused production video clip feature, so that the focused feature analysis can be completed at this time); According to the reverse order in the production video clip feature sequence, the initial focused production video clip features corresponding to the production video clip features are combined to form a corresponding initial focused production video clip feature sequence, and then based on the order in the initial focused production video clip feature sequence, each initial focused production video clip feature is polled to form the currently polled initial focused production video clip feature, and the currently polled initial focused production video clip feature is processed as in the above two steps (i.e. the currently polled initial focused production video clip feature is processed accordingly as the currently polled production video clip feature) to obtain the final focused production video clip feature, so that the front and back associated focusing can be achieved, and thus, a focused production video clip feature set is formed based on the combination of each final focused production video clip feature.

[0040] For example, in some implementations, the step of aggregating the features of each of the focused production video segments according to the focused feature analysis rule and outputting the typical component production features corresponding to the typical component production video may include: Analyze the focus evaluation parameters corresponding to each of the focus production video segment features according to the focus feature analysis rules for itself and the focus production video segment features; Based on the pre-configured parameter interval mapping rule and the focus evaluation parameter, the focus importance parameter of each of the focused production video clip features is analyzed; illustratively, the sum of the focus evaluation parameters corresponding to each of the focused production video clip features may be calculated first, and then, for each focused production video clip feature, the ratio between the focus evaluation parameter corresponding to the focused production video clip feature and the sum may be calculated to obtain the corresponding focus importance parameter; According to the corresponding focus importance parameter, each of the focus production video segment features is weighted and superimposed, that is, the focus importance parameter is used as the corresponding weighting coefficient to output the typical component production features corresponding to the typical component production video.

[0041] For example, in some implementations, the step of analyzing the focus evaluation parameters corresponding to each of the focus production video segment features based on the focus feature analysis rules for itself and the focus production video segment features may include: Based on the first parameter matrix, performing a cross multiplication operation on the focused production video segment feature and outputting a first cross multiplication result; and, based on the second parameter, performing a superposition operation on the first cross multiplication result and outputting a corresponding first superposition result; The first superposition result is subjected to excitation processing (which can be achieved through a corresponding excitation function), and the corresponding first excitation parameter is output. In addition, the transposed result of the third parameter matrix and the first excitation function are scalar-multiplied to obtain the focus evaluation parameters corresponding to the focus production video clip features. The first parameter matrix, the second parameter matrix and the third parameter matrix can all be used as parameters of a neural network and optimized and updated during the network update process.

[0042] For example, in some implementations, the possibility distribution data of the component abnormality data target set includes the first classification possibility of each component abnormality data classification set. Based on this, the step of calculating and outputting the first dimension Loss value based on the possibility distribution data of the component abnormality data target set and the classification type data may include: Based on the classification type data of each of the component abnormal data classification sets, the target type possibility of each of the component abnormal data classification sets is analyzed; the Loss value between each of the target type possibilities and the corresponding first classification possibility is determined, and the corresponding multiple type analysis Loss values ​​are output; the sum of all the type analysis Loss values ​​is calculated to obtain the first dimension Loss value; illustratively, the possibility distribution data of the component abnormal data target set can be a vector, each vector parameter corresponds to the first classification possibility of a component abnormal data classification set, and the classification type data is used to determine the target type possibility of the component abnormal data classification set. For example, the target type possibility can be configured to 1 or 0. When the target type possibility is 0, it reflects that the component abnormal data classification set and the typical component production video do not match. When the target type possibility is 1, it reflects that the component abnormal data classification set and the typical component production video match. For example, for 4 component abnormal data classification sets, the first The first classification possibility of one component abnormal data classification set is 0.2, the first classification possibility of the second component abnormal data classification set is 0.8, the first classification possibility of the third component abnormal data classification set is 0.2, and the first classification possibility of the fourth component abnormal data classification set is 0.7, that is, the possibility distribution data of the component abnormal data target set is {0.2, 0.8, 0.2, 0.7}, the target type possibility of the first component abnormal data classification set is 0, the target type possibility of the second component abnormal data classification set is 1, the target type possibility of the third component abnormal data classification set is 0, and the target type possibility of the fourth component abnormal data classification set is 1. In this way, the target type possibility of the component abnormal data classification set and the corresponding first classification possibility can be subjected to difference analysis (such as vector distance, etc.) to obtain the corresponding type analysis Loss value, and then the first dimension Loss value is determined by calculating the sum of the type analysis Loss values ​​of the four component abnormal data classification sets.

[0043] For example, in some implementations, the step of screening out the pending component abnormality data that matches the typical component production video from the component abnormality data target set may include: Polling each of the pending component abnormality data in the component abnormality data target set, and analyzing feature-related parameters, such as quantity product, between the typical component production features and the pending component abnormality features of the pending component abnormality data; Perform parameter interval mapping operation on the feature-related parameters, as described above, and output the second classification possibility corresponding to the abnormal data of the pending component; If the second classification possibility exceeds the pre-configured second reference classification possibility, the pending component abnormality data corresponding to the second classification possibility is marked, so that the pending component abnormality data marked as matching the typical component production video, the specific value of the second reference classification possibility is not restricted, such as 0.5, 0.6 and other values, or the corresponding pending component abnormality data with the maximum value of the second classification possibility can be marked as the pending component abnormality data matching the typical component production video.

[0044] For example, in some implementations, the possibility distribution data of the pending component abnormal data matched with the typical component production video includes the second classification possibility of each of the pending component abnormal data in the component abnormal data target set. Based on this, the step of calculating and outputting the second dimension Loss value based on the possibility distribution data of the pending component abnormal data matched with the typical component production video and the typical component abnormal data may include: Based on the typical component abnormal data, the target matching possibility of each of the pending component abnormal data in the component abnormal data target set is analyzed; the Loss value between each of the target matching possibilities and the corresponding second classification possibility is determined, and the corresponding multiple matching analysis Loss values ​​are output; the sum of all the matching analysis Loss values ​​is calculated to obtain the second dimension Loss value; illustratively, the possibility distribution data of the pending component abnormal data can be a vector, each vector parameter corresponds to the second classification possibility of a pending component abnormal data, and the typical component abnormal data is used to determine the target matching possibility of the pending component abnormal data. For example, the target matching possibility can be configured to 1 or 0. When the target matching possibility is 0, it reflects that the pending component abnormal data and the typical component production video do not match. When the target matching possibility is 1, it reflects that the pending component abnormal data and the typical component production video match. Matching, for example, a target set of component abnormal data includes three pending component abnormal data, the second classification possibility of the first pending component abnormal data is 0.88, the second classification possibility of the second pending component abnormal data is 0.35, and the second classification possibility of the third pending component abnormal data is 0.6. In this way, the possibility distribution data of the pending component abnormal data can be determined to be [0.88, 0.35, 0.6], and the target matching possibility of the first pending component abnormal data is 1, the target matching possibility of the second pending component abnormal data is 0, and the target matching possibility of the third pending component abnormal data is 1. The target matching possibility of the pending component abnormal data and the corresponding second classification possibility are analyzed for difference to obtain the matching analysis Loss value of the pending component abnormal data, and then the second dimension Loss value can be obtained by calculating the sum of the matching analysis Loss values ​​of the three pending component abnormal data.

[0045] Combination Figure 3 The embodiment of the present invention also provides an abnormal monitoring device for automobile parts based on AI visual detection, which can be applied to the above-mentioned abnormal monitoring system for automobile parts based on AI visual detection. The abnormal monitoring device for automobile parts based on AI visual detection can include: A production video segmentation module is used to obtain an original component production video of a target automobile component, and segment the original component production video according to corresponding production links, and output a plurality of segmented component production sub-videos corresponding to the target automobile component, wherein the plurality of original component production video frames included in the original component production video are obtained by monitoring the production line of the target automobile component; A video frame optimization module is used to optimize the video frames of each of the split parts production sub-videos, and output a plurality of optimized parts production sub-videos corresponding to the target automotive parts, wherein the video frame optimization includes optimizing at least one original parts production video frame of a subsequent split parts production sub-video based on at least one original parts production video frame of a previous split parts production sub-video, to form at least one optimized parts production sub-video frame; The component anomaly recognition module is used to use the target neural network that has undergone an updated operation to perform anomaly recognition on each of the optimized component production sub-videos, and output a plurality of component anomaly data corresponding to the target automobile component, wherein the component anomaly data is used to reflect the anomaly type of the target automobile component.

[0046] In summary, the method and system for monitoring abnormality of automobile parts based on AI visual detection provided by the present invention can first obtain the original production video of the target automobile parts, and, according to the corresponding production link, segment the original production video, and output multiple segmented production sub-videos corresponding to the target automobile parts; optimize the video frame of each segmented production sub-video, and output multiple optimized production sub-videos corresponding to the target automobile parts; use the target neural network after the update operation to identify the abnormality of each optimized production sub-video, and output multiple abnormal data of parts corresponding to the target automobile parts. Based on the above content, since the multiple sub-videos formed by the segmentation will be processed when the abnormality is identified, the granularity of the identification will be finer, thereby improving the reliability of abnormal monitoring of automobile parts to a certain extent, that is, improving the problem of low reliability of abnormal monitoring caused by redundant information interference between different production links; in addition, since at least one original production video frame of the previous segmented production sub-video is also used, at least one original production video frame of the next segmented production sub-video is associated and optimized, so that the information content of the sub-video of abnormal identification is richer, and the reliability of abnormal monitoring of automobile parts can be improved to a certain extent.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for monitoring abnormality of automobile parts based on AI visual inspection, characterized in that: The automobile parts abnormality monitoring method comprises: Acquire an original component production video of a target automotive component, and segment the original component production video according to corresponding production links, and output a plurality of segmented component production sub-videos corresponding to the target automotive component, wherein the plurality of original component production video frames included in the original component production video are obtained by monitoring a production line of the target automotive component; Performing video frame optimization on each of the segmented parts production sub-videos respectively, and outputting a plurality of optimized parts production sub-videos corresponding to the target automotive parts, wherein the video frame optimization includes optimizing at least one original parts production video frame of a subsequent segmented parts production sub-video based on an image connected domain according to at least one original parts production video frame of a previous segmented parts production sub-video, to form at least one optimized parts production sub-video frame; Using the updated target neural network, each of the optimized component production sub-videos is respectively identified for abnormality, and a plurality of component abnormality data corresponding to the target automobile component is output, wherein the component abnormality data is used to reflect the abnormal type of the target automobile component.

2. The method for abnormality monitoring of automobile parts based on AI visual inspection according to claim 1, characterized in that: The step of optimizing the video frames of each of the segmented parts production sub-videos and outputting a plurality of optimized parts production sub-videos corresponding to the target automotive parts comprises: For any one of the plurality of segmented component production sub-videos, the following processing is performed: For each original component production video frame in the segmented component production sub-video, the image connected domain of the original component production video frame is identified to obtain a set of image connected domains corresponding to the original component production video frame, and the number of image connected domains in the set of image connected domains corresponding to each original component production video frame in the segmented component production sub-video is discretely calculated to output a distribution parameter of the connected domain corresponding to the segmented component production sub-video; Based on the connected domain distribution parameter, a target number of original component production video frames ranked first are selected from the segmented component production sub-videos, and there is a negative correlation between the target number and the connected domain distribution parameter; In the segmented parts production sub-video, based on the inter-frame change information between multiple original parts production video frames in the previous segmented parts production sub-video of the segmented parts production sub-video, video frame interpolation is performed on the target number of original parts production video frames to output an optimized parts production sub-video corresponding to the target automotive part.

3. The method for abnormality monitoring of automobile parts based on AI visual inspection according to claim 1 or 2, characterized in that: The step of using the updated target neural network to identify anomalies for each of the optimized component production sub-videos and outputting a plurality of component anomaly data corresponding to the target automotive component comprises: For any one of the plurality of optimized component production sub-videos, the following processing is performed: A typical component production video and a plurality of pending component abnormality data are extracted, wherein the typical component production video has a plurality of typical component abnormality data, and the typical component abnormality data is used to reflect the component abnormality information actually possessed by the typical component production video; Determine the abnormal features of the pending parts of each of the abnormal data of the pending parts by using a feature space mapping network; Based on the abnormal features of the pending parts, the plurality of abnormal data of the pending parts are classified to form a plurality of corresponding abnormal data classification sets of the parts, and then the classification type data of each abnormal data classification set of the parts is obtained according to the typical abnormal data of the parts; The typical parts production video is loaded into a parts abnormality analysis network, and a parts abnormality data target set is analyzed from the multiple parts abnormality data classification sets by using the parts abnormality analysis network, and pending parts abnormality data matching the typical parts production video is screened out from the parts abnormality data target set; Based on the possibility distribution data of the target set of component abnormal data and the classification type data, calculate and output the first dimension Loss value; Based on the possibility distribution data of the abnormal data of the pending parts matched with the typical parts production video and the abnormal data of the typical parts, a Loss value of the second dimension is calculated and outputted; Based on the Loss value of the first dimension and the Loss value of the second dimension, the feature space mapping network and the component abnormality analysis network are updated, and the abnormal features of the pending component are adjusted during the network update process; The optimized parts production sub-video is loaded into the updated parts anomaly analysis network, and the updated parts anomaly analysis network is used to analyze the parts anomaly data of the optimized parts production sub-video. The updated parts anomaly analysis network belongs to the target neural network.

4. The method for abnormality monitoring of automobile parts based on AI visual inspection as claimed in claim 3, characterized in that: The step of using the feature space mapping network to determine the abnormal features of the pending parts of each of the abnormal data of the pending parts includes: According to a pre-configured data splitting rule, the pending component abnormality data is split, and at least one corresponding component abnormality data segment is output; Loading the component abnormal data fragment into a feature space mapping network to determine the component abnormal fragment feature corresponding to the component abnormal data fragment; For any of the pending component abnormal data, all of the component abnormal fragment features corresponding to the pending component abnormal data are feature fused to output the pending component abnormal features corresponding to the pending component abnormal data.

5. The method for abnormality monitoring of automobile parts based on AI visual inspection as claimed in claim 3, characterized in that: The component anomaly analysis network includes a feature mining subnetwork and an anomaly analysis subnetwork; The steps of loading the typical parts production video into the parts abnormality analysis network, analyzing the parts abnormality data target set from the multiple parts abnormality data classification sets using the parts abnormality analysis network, and screening out the pending parts abnormality data that matches the typical parts production video from the parts abnormality data target set include: Loading the typical parts production video into the feature mining sub-network to mine typical parts production features corresponding to the typical parts production video; Loading the typical component production features into the abnormality analysis subnetwork, and analyzing the first matching parameters of each component abnormality data classification set; Perform parameter interval mapping operations on each of the first matching parameters, and output a first classification possibility of each of the component abnormal data classification sets; If the first classification possibility exceeds the pre-configured first reference classification possibility, marking the component abnormality data classification set corresponding to the first classification possibility so as to be marked as the component abnormality data target set; In the target set of abnormal component data, the abnormal component data to be determined that matches the typical component production video is screened out.

6. The method for abnormality monitoring of automobile parts based on AI visual inspection according to claim 5, characterized in that: The step of loading the typical parts production video into the feature mining sub-network to mine typical parts production features corresponding to the typical parts production video includes: Segmenting the typical parts production video to form a corresponding parts production video segment set, wherein the parts production video segment set includes a plurality of parts production video segments; The front position of the component production video clip set is configured with a start symbol, and the last position of the component production video clip set is configured with an end symbol, so as to form a corresponding symbol production video clip set; Mapping the symbol production video segment set to a feature space of the segment, and outputting a corresponding production video segment feature set; According to the associated focused feature analysis rules, the production video clip feature set is subjected to focused feature analysis through the feature mining sub-network, and a corresponding focused production video clip feature set is output, wherein the focused production video clip feature set includes focused production video clip features of each component production video clip in the symbol production video clip set; According to the focus feature analysis rules for itself, the features of each of the focus production video clips are aggregated to output the typical parts production features corresponding to the typical parts production video.

7. The method for abnormality monitoring of automobile parts based on AI visual inspection according to claim 6, characterized in that: The step of aggregating the features of each of the focused production video segments according to the focused feature analysis rule and outputting the typical parts production features corresponding to the typical parts production video comprises: Analyze the focus evaluation parameters corresponding to each of the focus production video segment features according to the focus feature analysis rules for itself and the focus production video segment features; Analyzing the focus importance parameters of each of the focus production video segment features based on the pre-configured parameter interval mapping rules and the focus evaluation parameters; According to the corresponding focus importance parameters, weighted superposition operations are performed on the features of each focus production video segment to output typical component production features corresponding to the typical component production video.

8. The method for abnormality monitoring of automobile parts based on AI visual inspection as claimed in claim 5, characterized in that: The possibility distribution data of the target set of component abnormal data includes the first classification possibility of each of the classification sets of component abnormal data; The step of calculating and outputting the first dimension Loss value based on the possibility distribution data of the component abnormality data target set and the classification type data comprises: Analyzing the target type possibility of each of the component abnormal data classification sets based on the classification type data of each of the component abnormal data classification sets; Determine the Loss value between each target type possibility and the corresponding first classification possibility, and output corresponding multiple type analysis Loss values; The sum of all the Loss values ​​of the type analysis is calculated to obtain the Loss value of the first dimension.

9. The method for abnormality monitoring of automobile parts based on AI visual inspection according to claim 5, characterized in that: The step of screening out the pending component abnormality data matching the typical component production video from the component abnormality data target set comprises: Polling each of the pending component abnormality data in the component abnormality data target set, and analyzing feature-related parameters between the typical component production features and the pending component abnormality features of the pending component abnormality data; Perform parameter interval mapping operation on the feature-related parameters, and output the second classification possibility corresponding to the abnormal data of the pending component; If the second classification possibility exceeds the pre-configured second reference classification possibility, marking the pending component abnormality data corresponding to the second classification possibility so that the pending component abnormality data marked as matching the typical component production video; Furthermore, the possibility distribution data of the pending component abnormality data matched with the typical component production video includes the second classification possibility of each of the pending component abnormality data in the component abnormality data target set; The step of calculating and outputting a second dimension Loss value based on the possibility distribution data of the pending component abnormality data matched with the typical component production video and the typical component abnormality data comprises: Based on the typical component abnormality data, analyzing the target matching possibility of each of the pending component abnormality data in the component abnormality data target set; Determine a Loss value between each of the target matching possibilities and the corresponding second classification possibility, and output a plurality of corresponding matching analysis Loss values; The sum of all the matching analysis Loss values ​​is calculated to obtain the second dimension Loss value.

10. An abnormality monitoring system for automobile parts based on AI visual inspection, characterized in that: It includes a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the automobile parts abnormality monitoring method based on AI visual inspection as described in any one of claims 1 to 9.

Images