Industrial quality detection method and device, communication equipment and storage medium

Abstract

The invention relates to an industrial quality detection method and device, communication equipment and a storage medium. The method comprises the following steps: identifying first sample data of a to-be-detected product based on a small model to obtain a first quality inspection result; receiving a second quality inspection result sent by the cloud server; the second quality inspection result is obtained by identifying second sample data by the cloud server based on a large model; and obtaining a target quality inspection result of the to-be-detected product according to the first quality inspection result and the second quality inspection result. By adopting the method, fine defects, complex flaws and the like in the product can be identified, and the accuracy of industrial quality detection is improved.

Description

Technical Field

[0001] This application relates to the field of industrial internet technology, and in particular to an industrial quality inspection method, apparatus, communication equipment, and storage medium. Background Technology

[0002] Industrial quality inspection refers to a series of inspections, measurements, and evaluations conducted during the industrial production process to ensure that products meet predetermined quality standards and specifications.

[0003] In existing technologies, machine learning or deep learning methods can be used for industrial quality inspection. Specifically, images of the product to be inspected are first acquired, and feature extraction is performed on the acquired images to obtain image features. Based on a pre-trained model, the image features are then used to identify defects and flaws in the product. However, due to the use of a single model, the above-mentioned industrial quality inspection methods usually lack generalization ability and cannot accurately detect subtle defects and complex flaws in the product, resulting in low accuracy of the inspection results.

[0004] Therefore, current industrial quality testing technologies suffer from low accuracy in test results. Summary of the Invention

[0005] This application provides an industrial quality testing method, apparatus, communication equipment, storage medium, and computer program product, which can improve the accuracy of industrial quality testing results.

[0006] An industrial quality inspection method, the method being applied to production line end equipment, comprising:

[0007] The first sample data of the product to be tested is identified based on a small model to obtain the first quality inspection result;

[0008] Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0009] Based on the first quality inspection result and the second quality inspection result, the target quality inspection result of the product to be tested is obtained.

[0010] An industrial quality inspection method, the method being applied to a cloud server, includes:

[0011] The second sample data of the product to be tested is identified based on the large model to obtain the second quality inspection result;

[0012] The second quality inspection result is sent to the production line end equipment so that the production line end equipment determines the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

[0013] An industrial quality inspection device, applied to production line end equipment, comprising:

[0014] The first quality inspection module is used to identify the first sample data of the product to be inspected based on a small model and obtain the first quality inspection result.

[0015] The result receiving module is used to receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0016] The result determination module is used to obtain the target quality inspection result of the product to be inspected based on the first quality inspection result and the second quality inspection result.

[0017] An industrial quality inspection device, the device being used on a cloud server, comprising:

[0018] The second quality inspection module is used to identify the second sample data of the product to be tested based on the large model, and obtain the second quality inspection result.

[0019] The result sending module is used to send the second quality inspection result to the production line end equipment so that the production line end equipment can determine the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

[0020] A communication device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0021] The first sample data of the product to be tested is identified based on a small model to obtain the first quality inspection result;

[0022] Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0023] Based on the first quality inspection result and the second quality inspection result, the target quality inspection result of the product to be tested is obtained.

[0024] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0025] The first sample data of the product to be tested is identified based on a small model to obtain the first quality inspection result;

[0026] Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0027] Based on the first quality inspection result and the second quality inspection result, the target quality inspection result of the product to be tested is obtained.

[0028] A computer program product includes a computer program, characterized in that, when executed by a processor, the computer program implements the industrial quality inspection method provided in the embodiments of this application, the method being:

[0029] The first sample data of the product to be tested is identified based on a small model to obtain the first quality inspection result;

[0030] Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0031] Based on the first quality inspection result and the second quality inspection result, the target quality inspection result of the product to be tested is obtained.

[0032] The aforementioned industrial quality inspection methods, devices, communication equipment, storage media, and computer program products obtain a first quality inspection result by identifying the first sample data of the product to be inspected based on a small model. They then receive a second quality inspection result sent from a cloud server, which is obtained by the cloud server identifying the second sample data based on a large model. Based on the first and second quality inspection results, the target quality inspection result of the product to be inspected is obtained. Leveraging the strong reasoning capabilities of large models, sample data where the local small model detects anomalies but cannot detect specific anomalies are sent to the cloud for inspection by the large model in the cloud. This process identifies subtle defects and complex flaws in the product, improving the accuracy of industrial quality inspection. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating an industrial quality inspection method in one embodiment;

[0034] Figure 2 This is a flowchart illustrating an industrial quality inspection method in another embodiment;

[0035] Figure 3 This is a flowchart illustrating an industrial quality inspection process involving collaboration between large and small models on the cloud in one embodiment.

[0036] Figure 4This is a flowchart illustrating the model optimization and update process in a cloud-edge collaboration between large and small models in one embodiment.

[0037] Figure 5 This is a schematic diagram of an industrial quality inspection system that uses a large-scale model with cloud-edge collaboration in one embodiment.

[0038] Figure 6 This is an interactive flowchart of an industrial quality inspection method in one embodiment. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0040] The industrial quality inspection method provided in this application can be applied to industrial quality inspection scenarios, which may include a cloud server and one or more production line-end devices, each of which communicates with the cloud server. The cloud server can be a cloud server providing cloud computing services. The production line-end devices can be terminals deployed on the production line side, including but not limited to various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices.

[0041] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0042] In one embodiment, such as Figure 1 As shown, an industrial quality inspection method is provided. Taking the application of this method to production line equipment as an example, the method includes the following steps:

[0043] Step S102: Based on the small model, the first sample data of the product to be tested is identified to obtain the first quality inspection result.

[0044] Small models refer to machine learning or deep learning models with fewer parameters and simpler structures compared to large models. They are designed to run on limited computing resources or to achieve low latency and high energy efficiency inference on edge devices and mobile devices. The advantages of small models are faster training and inference speeds and lower memory usage, making them suitable for applications in scenarios with high real-time requirements or limited resources.

[0045] The product to be tested can be any product whose quality needs to be inspected, measured, and evaluated. The first sample data can be sample data of the product to be tested collected by equipment at the production line end. The first quality inspection result can be the preliminary quality inspection result obtained through a small model.

[0046] In practice, the equipment on the production line can collect the first sample data of the product to be tested, input the first sample data into a small model, and use the small model to identify the first sample data to obtain the first quality inspection result.

[0047] In practical applications, production line equipment can connect to various sensors deployed on the production line. The sensors collect raw sample data such as product appearance images, dimensions, and colors. The raw sample data is preprocessed to obtain the first sample data. The production line equipment can then perform low-complexity quality inspection on the first sample data, that is, use a small model to detect simple product defects, such as abnormal product size, surface scratches, cracks, spots, uneven color, etc. The first quality inspection result output by the small model can include product identification and the specific abnormality when the product is abnormal. The product identification can be normal, abnormal, or unidentifiable.

[0048] In this context, "Normal" means the mini-model identifies the product as normal. "Abnormal" means the mini-model identifies an anomaly in the product, specifying the exact nature of the anomaly, such as dimensional inconsistencies, surface scratches, cracks, spots, or uneven coloring. "Unidentifiable" means the mini-model identifies an anomaly but cannot pinpoint the specific nature of the anomaly.

[0049] Step S104: Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data.

[0050] Large-scale pre-trained models refer to deep learning models with strong representation and generalization capabilities obtained through unsupervised pre-training on massive amounts of data in fields such as natural language processing (NLP) and computer vision (CV). The advantage of large-scale models is that they can capture complex patterns and semantic relationships in the data, thus performing well in various downstream tasks and achieving good results even on tasks with smaller amounts of data.

[0051] The second sample data can be the first sample data that the small model identified as having anomalies, but which cannot describe the specific anomalies of the corresponding products. The second quality inspection result can be the quality inspection result further identified by the large model.

[0052] In practice, the production line equipment can extract the second sample data from the first sample data and send the second sample data to the cloud server. The cloud server uses a large model to identify the second sample data and sends the obtained second quality inspection result to the production line equipment. The production line equipment receives the second quality inspection result sent by the cloud server.

[0053] In practical applications, production line equipment can identify the first sample data as unidentifiable and use it as the second sample data. The second sample data is then sent to the cloud server for highly complex quality inspection. This involves using a large model to detect complex product defects, including but not limited to internal structural defects and surface micro-cracks. Similar to the small model, the second quality inspection result output by the large model can also include product identification and specific abnormalities when the product is abnormal.

[0054] Step S106: Based on the first quality inspection result and the second quality inspection result, obtain the target quality inspection result of the product to be inspected.

[0055] The target quality inspection result can be the final determined quality inspection result.

[0056] In practice, the equipment on the production line can combine the first and second quality inspection results to determine the final quality inspection result of the product to be inspected, and obtain the target quality inspection result.

[0057] For example, the first and second quality inspection results that identify the product as abnormal can be extracted as the target quality inspection results. In practical applications, the products corresponding to the target quality inspection results can be identified as defective products and transferred through sorting equipment.

[0058] The aforementioned industrial quality inspection method identifies a first sample of the product to be inspected based on a small model to obtain a first quality inspection result. It then receives a second quality inspection result from a cloud server, which is obtained by the cloud server identifying the second sample based on a large model. Based on the first and second quality inspection results, the target quality inspection result for the product to be inspected is obtained. This method leverages the strong reasoning capabilities of large models to send sample data where the local small model detects anomalies but cannot identify specific anomalies to the cloud. The large model in the cloud then performs the detection, thereby identifying subtle defects and complex flaws in the product, thus improving the accuracy of industrial quality inspection.

[0059] In one embodiment, the above-mentioned industrial quality inspection method may further include: identifying the first sample data that the small model cannot recognize as the second sample data; and sending the second sample data to a cloud server.

[0060] In practice, after identifying the first sample data of the product to be tested based on the small model and obtaining the first quality inspection result, the equipment at the production line end can identify the product in the first quality inspection result as the first sample data that cannot be identified, determine it as the second sample data, and send the second sample data to the cloud server.

[0061] In this embodiment, by identifying the first sample data that the small model cannot recognize as the second sample data and sending the second sample data to the cloud server, the powerful reasoning ability of the large model can be used to more accurately identify the subtle defects of the product when the reasoning ability of the small model is insufficient and cannot identify the quality inspection results, thereby reducing the misjudgment rate of industrial quality inspection.

[0062] In one embodiment, the step of sending the second sample data to the cloud server may specifically include: when the second sample data meets a first preset condition, sending the second sample data to the cloud server in real time; the first preset condition includes a real-time requirement exceeding a preset threshold; when the second sample data meets a second preset condition, packaging the second sample data and periodically sending the packaged second sample data to the cloud server; the second preset condition includes a real-time requirement not exceeding a preset threshold and / or limited bandwidth resources.

[0063] In practice, when the real-time requirement for the second sample data exceeds a preset threshold, the production line equipment can send the second sample data to the cloud server in real time. When the real-time requirement for the second sample data does not exceed the preset threshold, and / or the transmission bandwidth between the production line equipment and the cloud server is limited, the production line equipment can package the second sample data and periodically send the packaged second sample data to the cloud server.

[0064] For example, for quality inspection tasks with high real-time requirements, the second sample data can be transmitted to the cloud server in real time. For scenarios with low real-time requirements and / or limited bandwidth resources, multiple sets of second sample data can be packaged into a batch and transmitted to the cloud server periodically.

[0065] In this embodiment, when the second sample data meets the first preset condition, the second sample data is sent to the cloud server in real time. When the second sample data meets the second preset condition, the second sample data is packaged and periodically sent to the cloud server. Different transmission strategies can be adopted for different application scenarios to ensure the timeliness and reliability of industrial quality inspection.

[0066] In one embodiment, the above-mentioned industrial quality inspection method may further include: updating the small model based on the abnormal sample labels sent by the cloud server; the abnormal sample labels are obtained based on the abnormal results in the second quality inspection results.

[0067] Here, the abnormal result can be a second quality inspection result that identifies the product as abnormal. The abnormal sample label can be the label corresponding to the second sample data that the large model identifies as abnormal and can identify the specific abnormal situation. The abnormal sample label can include the data index of the second sample data and the specific abnormal situation corresponding to the second sample data.

[0068] In practice, the cloud server can add abnormal sample labels to the second sample data that the large model identifies as abnormal and the specific abnormal situation. The abnormal sample labels are then sent to the production line equipment. The production line equipment can use the received abnormal sample labels to determine the sample data that the small model cannot identify as the specific abnormal situation. Using this sample data and its corresponding specific abnormal situation, the small model is retrained and its parameters are updated.

[0069] For example, suppose a product associated with a sample data has an internal structural defect. The production line equipment detects the anomaly in the product corresponding to this sample data using a small model, but cannot identify the specific anomaly. Therefore, the sample data is sent to the cloud server. The cloud server identifies the product anomaly using a large model, and the specific anomaly is identified as an internal structural defect. It then adds an anomaly sample label to the sample data, which contains the index of the sample data and the corresponding specific anomaly. The cloud server sends the anomaly sample label to the production line equipment. Based on the received anomaly sample label, the production line equipment determines that the sample data and the corresponding specific anomaly are internal structural defects. Thus, the sample data is used as a new training sample, and the internal structural defect is used as the corresponding new sample label to retrain the small model.

[0070] In this embodiment, by updating the small model based on the abnormal sample labels sent by the cloud server, the recognition results of the large model can be used to update the small model, thereby improving the detection performance of the small model.

[0071] In one embodiment, the above-mentioned industrial quality inspection method may further include: updating the small model according to the new small model parameters issued by the cloud server; the new small model parameters are obtained by the cloud server updating the large model according to the abnormal results in the second quality inspection results, and the updated large model is subjected to lightweight processing, the lightweight processing including at least one of pruning, quantization, distillation, and low-rank decomposition.

[0072] Among them, the small model parameters can be the model parameters of the small model.

[0073] In practice, the cloud server can use the second sample data of anomalies and specific anomalies identified by the large model as new training samples, and the specific anomalies identified by the large model as corresponding new sample labels to retrain the large model and obtain an updated large model. The cloud server can also perform lightweight processing on the updated large model, including but not limited to pruning, quantization, distillation, low-rank decomposition, etc., to obtain new small model parameters. The new small model parameters are sent to the production line equipment, and the production line equipment can update the small model according to the received new small model parameters.

[0074] In this embodiment, by updating the small model according to the new small model parameters issued by the cloud server, the detection performance of the small model can be improved, and the update speed of the small model can also be improved.

[0075] In one embodiment, the above-mentioned industrial quality inspection method may further include: acquiring original sample data of the product to be inspected; preprocessing the original sample data to obtain first sample data; the preprocessing includes at least one of cleaning, compression, noise reduction and feature extraction.

[0076] The original sample data can be the original sample data collected.

[0077] In practice, production line equipment can collect raw sample data of the product to be tested through various sensors deployed on the production line. The raw sample data is then preprocessed, such as cleaning, compression, noise reduction, and feature extraction, to obtain the first sample data.

[0078] In this embodiment, by acquiring the original sample data of the product to be tested and preprocessing the original sample data to obtain the first sample data, the reliability of the first sample data can be improved, thereby improving the accuracy of industrial quality inspection.

[0079] In one embodiment, such as Figure 2 As shown, another industrial quality inspection method is provided, and the method is illustrated using a cloud server as an example. The method includes the following steps:

[0080] Step S202: Based on the large model, identify the second sample data of the product to be tested to obtain the second quality inspection result;

[0081] Step S204: Send the second quality inspection result to the production line end equipment so that the production line end equipment can determine the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

[0082] In practice, the production line equipment can collect the first sample data of the product to be tested, identify the first sample data using a small model, and obtain the first quality inspection result. The production line equipment can also extract the second sample data from the first sample data and send the second sample data to the cloud server. The cloud server uses a large model to identify the second sample data and sends the obtained second quality inspection result to the production line equipment. The production line equipment combines the first quality inspection result and the second quality inspection result to determine the target quality inspection result of the product to be tested.

[0083] Since the specific processing procedures of the cloud server have been described in detail in the foregoing embodiments, they will not be repeated here.

[0084] In this embodiment, a second quality inspection result is obtained by identifying the second sample data of the product to be inspected based on a large model, and then the second quality inspection result is sent to the equipment at the production line end. The advantage of the strong reasoning ability of the large model can be utilized to send sample data that the local small model detects as abnormal but cannot detect the specific abnormality to the cloud. The large model in the cloud can then perform the detection, thereby identifying subtle defects and complex flaws in the product, and improving the accuracy of industrial quality inspection.

[0085] In one embodiment, the above-mentioned industrial quality inspection method may further include: receiving second sample data sent by equipment at the production line end; the second sample data is first sample data that cannot be recognized by the small model.

[0086] In practice, the production line equipment can identify the first sample data, which is marked as unidentifiable in the first quality inspection result, as the second sample data and send it to the cloud server. The cloud server then receives the second sample data sent by the production line equipment.

[0087] In this embodiment, by receiving the second sample data sent by the production line equipment, when the reasoning ability of the small model is insufficient and cannot identify the quality inspection results, the powerful reasoning ability of the large model can be used to more accurately identify the subtle defects of the product and reduce the misjudgment rate of industrial quality inspection.

[0088] In one embodiment, the above-mentioned industrial quality inspection method may further include: obtaining an abnormal sample label based on the abnormal results in the second quality inspection result; and sending the abnormal sample label to the production line end equipment so that the production line end equipment updates the small model based on the abnormal sample label.

[0089] In practice, the cloud server can add abnormal sample labels to the second sample data that the large model identifies as abnormal and the specific abnormal situation. The abnormal sample labels are then sent to the production line equipment. The production line equipment can use the received abnormal sample labels to determine the sample data that the small model cannot identify as the specific abnormal situation. Using this sample data and its corresponding specific abnormal situation, the small model is retrained and its parameters are updated.

[0090] In this embodiment, by obtaining abnormal sample tags based on the abnormal results in the second quality inspection results and sending the abnormal sample tags to the equipment at the production line end, the identification results of the large model can be used to update the small model, thereby improving the detection performance of the small model.

[0091] In one embodiment, the above-mentioned industrial quality inspection method may further include: updating the large model based on the abnormal results in the second quality inspection results; performing lightweight processing on the updated large model to obtain new small model parameters; the lightweight processing includes at least one of pruning, quantization, distillation, and low-rank decomposition; and sending the new small model parameters to the production line end equipment so that the production line end equipment updates the small model according to the new small model parameters.

[0092] In practice, the cloud server can use the second sample data of anomalies and specific anomalies identified by the large model as new training samples, and the specific anomalies identified by the large model as corresponding new sample labels to retrain the large model and obtain an updated large model. The cloud server can also perform lightweight processing on the updated large model, including but not limited to pruning, quantization, distillation, low-rank decomposition, etc., to obtain new small model parameters. The new small model parameters are sent to the production line equipment, and the production line equipment can update the small model according to the received new small model parameters.

[0093] In this embodiment, by updating the large model based on the abnormal results in the second quality inspection results, and performing lightweight processing on the updated large model to obtain new small model parameters, the new small model parameters are sent to the equipment at the production line end. This can improve the detection performance of the small model and also increase the update speed of the small model.

[0094] In one embodiment, the step of updating the large model based on the abnormal results in the second quality inspection results may specifically include: counting the number of collaborative quality inspection tasks and / or the number of defects of the same type; when the number of collaborative quality inspection tasks exceeds a first preset threshold and / or the number of defects of the same type exceeds a second preset threshold, updating the large model based on the abnormal results in the second quality inspection results.

[0095] The number of collaborative quality inspection tasks can be the total number of tasks where large and small models collaborate on quality inspection. The first preset threshold can be the upper limit of the number of collaborative quality inspection tasks. The number of defects of the same type can be the total number of defects of a specified type. The second preset threshold can be the upper limit of the number of defects of the same type.

[0096] In practice, the cloud server can determine the number of collaborative quality inspection tasks and the number of defects of the same type based on the actual needs of industrial quality inspection. If the number of collaborative quality inspection tasks exceeds the first preset threshold and / or the number of defects of the same type exceeds the second preset threshold, the large model update process is initiated, and the large model is updated based on the abnormal results in the second quality inspection results.

[0097] In this embodiment, by counting the number of collaborative quality inspection tasks and / or the number of defects of the same type, when the number of collaborative quality inspection tasks exceeds the first preset threshold and / or the number of defects of the same type exceeds the second preset threshold, the large model is updated according to the abnormal results in the second quality inspection results. This can adapt to the specific needs of actual scenarios and improve the reliability of industrial quality inspection.

[0098] In one embodiment, step S202 may specifically include: when the second sample data corresponds to one production line end device, identifying the feature data of the second sample data to obtain a second quality inspection result; when the second sample data corresponds to at least two production line end devices, fusing the feature data of the second sample data, and obtaining a second quality inspection result based on the fused feature data.

[0099] In specific implementation, if the second sample data comes from a single production line device, for example, if the second sample data corresponds to a specific production stage of a specific product, the cloud server can directly extract features from the second sample data to obtain feature data, and then identify the feature data to obtain the second quality inspection result. If the second sample data comes from multiple production line devices, for example, if the second sample data corresponds to different production stages of the same product, the cloud server needs to first extract features from each second sample data to obtain feature data, then fuse the various feature data to obtain fused features, and then identify the fused features to obtain the second quality inspection result. The fusion methods include, but are not limited to, feature-level fusion, decision-level fusion, and statistical data integration.

[0100] In this embodiment, when the second sample data corresponds to one production line end device, the feature data of the second sample data is identified to obtain the second quality inspection result. When the second sample data corresponds to at least two production line end devices, the feature data of the second sample data is fused, and the second quality inspection result is obtained based on the fused feature data. This allows the cloud-based large model to determine the quality inspection result in different ways according to different sample data collection scenarios, further improving the accuracy of industrial quality inspection.

[0101] To facilitate a deeper understanding of the embodiments of this application by those skilled in the art, a specific example will be used for illustration below.

[0102] In the field of industrial quality inspection, small models generally lack generalization ability and cannot identify complex defects. While large models can perform accurate quality inspection, they cannot run locally on production line equipment and have high inference costs. To address these issues, this application proposes a quality inspection method that combines small and large models in a cloud-based collaborative manner. Based on the collaborative work of a small model on the production line and a large model in the cloud, the small model on the production line can quickly process real-time data and perform preliminary screening, reducing the business concurrency processing pressure on the large model in the cloud. The large model in the cloud has powerful inference capabilities and can more accurately identify subtle defects in parts, reducing the false positive rate. The combination of edge and cloud makes the system more adaptable and fault-tolerant. In the event of network interruptions or cloud server failures, the edge equipment can work independently, ensuring the normal operation of routine basic quality inspection on the production line. Edge-cloud collaboration improves the performance of the quality inspection model: on the one hand, the large model generates new sample labels based on real-time detection results, allowing the production line to update the small model in a timely manner; on the other hand, the large model periodically performs model self-optimization based on historical defect data and generates suitable parameters for the small model on the edge, updating the small model on the production line.

[0103] In one embodiment, a method is provided to improve industrial quality inspection performance through collaborative reasoning, optimization, and updating of small models on the production line and large models in the cloud. This method enhances the accuracy and efficiency of industrial quality inspection by leveraging collaborative reasoning, optimization, and updating of small models on the production line and large models in the cloud. The method comprises two aspects: collaborative model reasoning and collaborative model optimization and updating.

[0104] In quality inspection scenarios involving collaborative reasoning between small and large models, the small model on the production line can process the sample data to be inspected in real time and perform preliminary quality inspection, reducing the processing pressure of concurrent business of the large model in the cloud. The large model in the cloud has powerful reasoning capabilities, which can more accurately identify subtle defects in parts and discover previously unidentified product flaws, thereby reducing the false judgment rate.

[0105] refer to Figure 3 The implementation process of this method on the production line is as follows:

[0106] Step ES1, Data Acquisition and Preprocessing of Products to be Inspected: Using various sensors deployed on the production line, raw sample data such as the appearance image, size, and color of the product are collected for quality inspection. The raw sample data is then cleaned, compressed, denoised, and feature extracted to obtain the first sample data.

[0107] Step ES2, Low-complexity quality inspection: Based on the first sample data, a small model deployed at the production line is used to detect simple product defects, including abnormal product size, surface scratches, cracks, spots, uneven color, etc. The output quality inspection results include product identification (normal, abnormal, unidentifiable) and corresponding defect descriptions.

[0108] Step ES3, Initiation of a High-Complexity Quality Inspection Task: For sample data that is abnormal, unidentifiable, and for which detailed defect descriptions cannot be provided (second sample data), it is sent to the cloud for further inspection. The data transmission strategy may include:

[0109] (1) For quality inspection tasks with high real-time requirements, send them out promptly;

[0110] (2) For situations where real-time requirements are not high and current bandwidth resources are limited, multiple sets of data are packaged into a batch and sent periodically.

[0111] Step ES4, Quality Inspection Result Output: Based on the quality inspection results of step ES2 and the highly complex quality inspection results fed back from the cloud, a definite quality inspection result is given for the inspected product, and unqualified products are diverted to the production line.

[0112] refer to Figure 3 The implementation process of this method in the cloud is as follows:

[0113] Step CS1, Collaborative Quality Inspection Task Reception: Receive high-complexity quality inspection task requirements and data from the production line, including sample data (second sample data) after preprocessing of the product to be inspected and small model test result data.

[0114] Step CS2, High-Complexity Quality Inspection: Utilize a large cloud-based model to conduct in-depth quality inspection of the samples (second sample data), including but not limited to internal structural defects and surface micro-cracks, and output corresponding defect descriptions, which may include:

[0115] (1) Single-sample depth detection: Image depth detection is performed using image feature data from a single production line end;

[0116] (2) Multi-sample fusion analysis: The quality inspection results of the corresponding product samples at different production stages are retrieved from the global database and fused for analysis. The fusion methods include, but are not limited to, feature-level fusion, decision-level fusion and statistical data integration.

[0117] Step CS3, Quality Inspection Result Feedback: Send the defect description generated in step CS2 to the corresponding production line equipment.

[0118] Compared to existing technologies, the above processing method can improve the accuracy and efficiency of industrial quality inspection: Based on the collaborative work of the small model at the production line and the large model in the cloud, the small model at the production line can quickly process real-time data and perform preliminary screening, reducing the business concurrency processing pressure on the large model in the cloud; the large model in the cloud has powerful reasoning capabilities, which can more accurately identify subtle defects in individual parts. Moreover, since the cloud server can obtain sample quality inspection records of the same product at different production stages from multiple production line equipment, it can conduct fusion analysis of the same product at different production stages, further reducing the false judgment rate. Furthermore, it can also enhance the robustness of the quality inspection system: the combination of edge and cloud makes the system more adaptable and fault-tolerant. In the event of network interruption, cloud server failure, etc., the edge equipment can work independently, ensuring the normal operation of routine basic quality inspection on the production line.

[0119] In the application of collaborative model optimization and updating, on the one hand, the large model generates new sample labels based on the detection results, which are then used by the smaller models on the production line for local optimization. On the other hand, the large model iteratively optimizes itself as needed based on the frequency of received collaborative quality inspection tasks and historical defect data, and generates smaller model parameters suitable for the production equipment based on the optimized large model, which are then sent to the production equipment. This collaborative optimization method achieves both rapid updates to the smaller model and on-demand updates to the large model, and the optimized model can further improve the accuracy of subsequent quality inspections.

[0120] refer to Figure 4 The implementation process of this method on the production line is as follows:

[0121] Step ES5, Local Update of Small Model: Update the parameters of the local small model based on the new labels (abnormal sample labels) of the abnormal samples received from the cloud.

[0122] Step ES6, Small Model Synchronous Optimization: Receive optimization notifications and small model parameters sent from the cloud periodically, and update the local small model parameters.

[0123] The implementation process of this method on the production line is as follows:

[0124] Step CS4, New Label Generation for Quality Inspection Samples: Based on the results of high-complexity quality inspection, determine whether the abnormal sample is an unidentified type. If so, generate a new label for the abnormal sample (abnormal sample label).

[0125] Step CS5, Periodic Model Optimization and Synchronization: Based on accumulated historical data and defect types, optimize the large quality inspection model as needed, and generate smaller models for each production line based on the optimized large model. Send the optimized smaller model parameters to the production line. This may include the following steps:

[0126] Step CS51, Model Optimization Start: Based on the specific needs of the industrial scenario, set thresholds for the number of collaborative quality inspection tasks and the number of similar defects. If the frequency of receiving collaborative quality inspection tasks and the number of similar defects both exceed the preset thresholds, then start large model parameter optimization.

[0127] Step CS52, Model Lightweighting: Generate a smaller model suitable for production line equipment from the optimized large model. Lightweighting techniques include, but are not limited to, model pruning, quantization, distillation, and low-rank decomposition.

[0128] Compared to existing technologies, the above process achieves both rapid updates of small models and on-demand updates of large models. Moreover, the optimized model can further improve the accuracy of subsequent quality inspections: on the one hand, the large model generates new sample labels based on the detection results, allowing the small model on the production line to optimize locally in a timely manner; on the other hand, the large model iterates and optimizes the model as needed based on the frequency of received collaborative quality inspection tasks and historical defect data, and generates small model parameters suitable for production equipment based on the optimized large model, assisting the production equipment in updating the model as needed.

[0129] The large-scale model-cloud collaborative quality inspection method proposed in this application can be used in quality inspection scenarios in the field of intelligent manufacturing. The system architecture for implementing this method includes production line-end equipment and cloud servers, such as... Figure 5 As shown.

[0130] The production line equipment may include the following modules:

[0131] Data acquisition and preprocessing: Various images and sensors on the production line acquire sample data of the products to be inspected, and perform processing such as cleaning, compression, noise reduction and feature extraction on the sample data;

[0132] Low-complexity quality inspection: Simple quality inspection is carried out using small end-side models, including but not limited to product size abnormalities, surface scratches, cracks, spots, uneven color, etc., and corresponding defect descriptions are output;

[0133] Initiating a high-complexity quality inspection task: For products that are abnormal and for which an accurate description of the defect cannot be given, a collaborative quality inspection request is sent to the cloud server;

[0134] Model Update: Update the parameters of the local small model.

[0135] A cloud server may include the following modules:

[0136] Global database: Records historical quality inspection data of equipment on subordinate production lines and cloud servers;

[0137] High-complexity quality inspection: Based on the high-complexity quality inspection requirements sent from the production line (including sample data of the product to be inspected after preprocessing and small model detection results), we use cloud-based large models to carry out in-depth quality inspection, including but not limited to internal structural defects and surface micro-cracks, and output the corresponding defect descriptions and new labels for the sample.

[0138] Cloud-based large model updates: The quality inspection large model is optimized periodically based on accumulated historical data and defect types;

[0139] Production line small model generation and synchronization: Generate small models adapted to each production line based on the optimized large model, and send the optimized small models to the production line.

[0140] In one embodiment, for a quality inspection scenario in an automotive precision parts production line, an example of a collaborative quality inspection method between large and small models and the cloud is provided, which may specifically include the following steps:

[0141] Step 301: Production line equipment collects raw sample data of parts and components and performs data preprocessing: The production line equipment uses cameras and sensors such as temperature and sound to collect raw sample data for quality inspection, such as appearance images, size, color, and surface temperature of the products to be inspected from different angles, and performs cleaning, compression, noise reduction and feature extraction on the raw sample data.

[0142] Step 302: Low-complexity quality inspection is carried out on the production line equipment: Using the small model deployed on the production line equipment, and taking the original sample data collected in step 301 as input, simple product defect detection is carried out, including abnormal product size, surface scratches, cracks, spots, uneven color, etc. The detection results include product identification (normal, abnormal, unidentifiable), and the corresponding defect description is output for abnormal products.

[0143] Step 303: Production line equipment initiates a high-complexity quality inspection task: For samples whose inspection results are marked as abnormal or unidentifiable, and for which no detailed defect description can be provided, a data transmission strategy is formulated based on the characteristics of the quality inspection task and network bandwidth resources:

[0144] (1) For quality inspection tasks with high real-time requirements, such as when the part is an intermediate part and needs to go through other processes to generate the final product, the high-complexity quality inspection task should be sent in a timely manner.

[0145] (2) For situations where real-time requirements are not high and current bandwidth resources are limited, multiple sets of data are packaged into a batch and sent periodically.

[0146] Step 304: The cloud server receives the collaborative quality inspection task sent by the production line: The received data includes the pre-processed data of the products to be inspected by the production line equipment and the preliminary test results of the small model.

[0147] Step 305: The cloud server performs high-complexity quality inspection: Based on the sample data sent by the production line equipment, the cloud server uses a large cloud model to conduct in-depth quality inspection of complex defects, including but not limited to internal structural defects and surface micro-cracks, and outputs corresponding defect descriptions. In-depth quality inspection includes two aspects:

[0148] (1) Single-sample depth detection: Image depth detection is performed using image feature data from a single production line end;

[0149] (2) Multi-sample fusion analysis: The quality inspection results of the corresponding component products at different production stages are retrieved from the global database and fused for analysis. The fusion methods include, but are not limited to, feature-level fusion, decision-level fusion, and statistical data integration. Feature-level fusion can use an attention mechanism to fuse feature data from samples at different stages.

[0150] Step 306: The cloud server feeds back the highly complex quality inspection results to the production line equipment: The quality inspection results include the identification results of the samples in the task and the specific description of defects.

[0151] Step 307: Production line equipment summarizes quality inspection results for product sorting: Based on local low-complexity quality inspection results and cloud-based feedback of high-complexity quality inspection results, a definite classification and defect description are given to the inspected products, and sent to the sorting equipment for the production line to divert defective products.

[0152] In the quality inspection scenario of automotive precision parts production lines, the collaborative optimization and updating process of large and small models on the cloud and edge can include the following steps:

[0153] Step 401: The cloud server generates new labels for the component quality inspection samples: Based on the high-complexity quality inspection results and combined with the historical quality inspection data of the production line, the large model determines whether the abnormal sample is an unidentified type. If so, a new label is generated.

[0154] Step 402, Local Update of Small Model on Production Line: The production line equipment updates the local model parameters based on the new labels of abnormal samples sent from the cloud. The specific process includes:

[0155] Forward propagation: The sample data of the product is input into the model, and after a series of calculations (such as matrix multiplication, application of activation functions, etc.), the model's predicted output is obtained;

[0156] Calculate the loss: Use a loss function (such as mean squared error, cross-entropy loss, etc.) to compare the predicted output of the small model with the new label sent from the cloud to obtain a numerical value representing the prediction error;

[0157] Backpropagation: Calculates the gradient of the loss function with respect to each weight, determining the degree of influence of each weight on the total loss. This process starts from the model's output layer and proceeds layer by layer backward until the input layer.

[0158] Weight update: After obtaining gradient information, optimization algorithms (such as stochastic gradient descent SGD, adaptive moment estimation Adam, etc.) are used to update the model weights;

[0159] Iterative optimization: Repeat the above process multiple times until the model converges or reaches the predetermined number of iterations.

[0160] Step 403, Periodic Optimization and Synchronization of the Large Model on the Cloud Server: Based on accumulated historical data and defect types, the quality inspection large model is periodically optimized, and small models are generated for each production line based on the optimized large model. The optimized small models are then sent to the production line. This step is completed in two steps:

[0161] (1) Model optimization start: According to the specific needs of the industrial scenario, set the threshold for the number of collaborative quality inspection tasks and the number of similar defects respectively. If the frequency of receiving collaborative quality inspection tasks and the number of similar defects both exceed the preset threshold, then start the large model parameter optimization.

[0162] (2) Model lightweighting: The optimized large model is used to generate a small model suitable for production line equipment. Lightweighting techniques include, but are not limited to, model pruning, quantization, distillation, and low-rank decomposition.

[0163] Step 404, Synchronous optimization of small models on production line equipment: Update the local small model according to the small model update parameters sent from the cloud.

[0164] In one embodiment, a method and apparatus for enhancing industrial quality inspection performance through edge-cloud collaboration of large and small models are provided, wherein edge-cloud collaboration for industrial quality inspection may include the following steps:

[0165] Step 501, Data Acquisition and Preprocessing on the Production Line: The equipment at the production line end uses various sensors such as cameras to collect raw sample data (including appearance images, size, color, etc.) of the products to be inspected from different angles for quality inspection, and performs preprocessing on the sample data, such as cleaning, compression, noise reduction, feature extraction, etc.

[0166] Step 502, Low-complexity quality inspection on the production line: Using a small model deployed on the equipment at the end of the production line, simple product defect detection is performed based on the raw sample data collected in step 501. This includes detecting abnormal product dimensions, surface scratches, cracks, spots, uneven color, etc. The detection results can include product identification (normal, abnormal, unidentifiable), and output corresponding defect descriptions for abnormal products.

[0167] Step 503: The production line equipment initiates a high-complexity inspection task: For samples marked as abnormal or unidentifiable, and for which a detailed defect description cannot be provided, a data transmission strategy is formulated based on the characteristics of the inspection task and network bandwidth resources.

[0168] (1) For quality inspection tasks with high real-time requirements, data should be sent in a timely manner;

[0169] (2) For tasks with low real-time requirements and limited bandwidth resources, multiple sets of data are packaged into batches and sent periodically.

[0170] Step 504: The cloud server receives the collaborative inspection task: The received data includes preprocessed data from the production line equipment of the product to be inspected, as well as the preliminary inspection results of the small model.

[0171] Step 505, High-Complexity Quality Inspection on the Cloud Server: Based on the sample data sent by the equipment on the production line, the cloud server uses a large model to perform in-depth inspection of complex defects, including but not limited to internal structural defects and surface micro-cracks, and outputs corresponding defect descriptions. The in-depth inspection specifically includes:

[0172] (1) Single-sample depth detection: Image depth detection is performed using image feature data from a single production line end;

[0173] (2) Multi-sample fusion analysis: If multiple samples from different production stages of the same product are received from multiple production line end devices, the feature data of different stages are fused for analysis. Fusion methods include, but are not limited to, feature-level fusion, decision-level fusion and statistical data integration. Among them, feature-level fusion can use attention mechanism to integrate the feature data of samples from different stages.

[0174] Step 506: Feed back highly complex test results from the cloud server to the production line equipment: The test results include the identification results of the samples in the task and specific defect descriptions.

[0175] Step 507: Summarize the test results by production line equipment for product sorting: Based on local low-complexity test results and high-complexity test results fed back from the cloud, classify and describe the defects of the tested products, and send the results to the sorting equipment so that defective products can be transferred on the production line.

[0176] In one embodiment, such as Figure 6 As shown, an industrial quality inspection method is provided, including the following steps:

[0177] Step S601: The equipment at the production line end identifies the first sample data of the product to be inspected based on the small model and obtains the first quality inspection result;

[0178] In step S602, the production line equipment determines the second sample data based on the first quality inspection result and sends the second sample data to the cloud server.

[0179] Step S603: The cloud server identifies the second sample data based on the large model, obtains the second quality inspection result, and sends the second quality inspection result to the production line equipment.

[0180] In step S604, the production line equipment determines the target quality inspection result based on the first quality inspection result and the second quality inspection result.

[0181] In practice, the production line equipment can collect the first sample data of the product to be tested, identify the first sample data using a small model, and obtain the first quality inspection result. The production line equipment can also extract the second sample data from the first sample data based on the first quality inspection result and send the second sample data to the cloud server. The cloud server uses a large model to identify the second sample data and sends the obtained second quality inspection result to the production line equipment. The production line equipment combines the first and second quality inspection results to determine the target quality inspection result of the product to be tested.

[0182] Since the specific processing procedures of the production line equipment and cloud server have been described in detail in the aforementioned embodiments, they will not be repeated here.

[0183] In this embodiment, the production line equipment identifies the first sample data of the product to be inspected based on a small model to obtain a first quality inspection result. Based on the first quality inspection result, a second sample data is determined and sent to a cloud server. The cloud server identifies the second sample data based on a large model to obtain a second quality inspection result, which is then sent to the production line equipment. The production line equipment determines the target quality inspection result based on the first and second quality inspection results. This approach leverages the strong reasoning capabilities of large models to send sample data where the local small model detects anomalies but cannot identify specific anomalies to the cloud. The large model in the cloud then performs the detection, thereby identifying subtle defects and complex flaws in the product, improving the accuracy of industrial quality inspection.

[0184] It should be understood that, although Figure 1-4 The steps in flowchart 6 are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order in which these steps are performed; they can be executed in other orders. Furthermore, Figure 1-4At least some of the steps in 6 may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0185] In one embodiment, an industrial quality inspection device is provided, the device being applied to production line end equipment, comprising:

[0186] The first quality inspection module is used to identify the first sample data of the product to be inspected based on a small model and obtain the first quality inspection result.

[0187] The result receiving module is used to receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0188] The result determination module is used to obtain the target quality inspection result of the product to be inspected based on the first quality inspection result and the second quality inspection result.

[0189] In one embodiment, the above-mentioned industrial quality inspection device further includes a second sample data sending module, used to identify the first sample data that the small model cannot recognize as the second sample data; and send the second sample data to the cloud server.

[0190] In one embodiment, the second sample data sending module is further configured to send the second sample data to the cloud server in real time when the second sample data meets the first preset condition; the first preset condition includes a real-time requirement exceeding a preset threshold; when the second sample data meets the second preset condition, the second sample data is packaged and the packaged second sample data is periodically sent to the cloud server; the second preset condition includes a real-time requirement not exceeding a preset threshold and / or limited bandwidth resources.

[0191] In one embodiment, the above-mentioned industrial quality inspection device further includes a first small model update module, used to update the small model according to the abnormal sample tags sent by the cloud server; the abnormal sample tags are obtained based on the abnormal results in the second quality inspection results.

[0192] In one embodiment, the above-mentioned industrial quality inspection device further includes a second small model update module, used to update the small model according to the new small model parameters issued by the cloud server; the new small model parameters are obtained by the cloud server updating the large model according to the abnormal results in the second quality inspection results, and the updated large model is subjected to lightweight processing, the lightweight processing including at least one of pruning, quantization, distillation, and low-rank decomposition.

[0193] In one embodiment, the above-mentioned industrial quality inspection device further includes a preprocessing module for acquiring the original sample data of the product to be inspected; preprocessing the original sample data to obtain the first sample data; the preprocessing includes at least one of cleaning, compression, noise reduction and feature extraction.

[0194] In one embodiment, an industrial quality inspection device is provided, the device being applied to a cloud server, comprising:

[0195] The second quality inspection module is used to identify the second sample data of the product to be tested based on the large model, and obtain the second quality inspection result.

[0196] The result sending module is used to send the second quality inspection result to the production line end equipment so that the production line end equipment can determine the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

[0197] In one embodiment, the above-mentioned industrial quality inspection device further includes a second sample data receiving module for receiving the second sample data sent by the production line end equipment; the second sample data is the first sample data that the small model cannot recognize.

[0198] In one embodiment, the above-mentioned industrial quality inspection device further includes an abnormal sample tag sending module, which is used to obtain an abnormal sample tag based on the abnormal result in the second quality inspection result; and send the abnormal sample tag to the production line end equipment so that the production line end equipment updates the small model based on the abnormal sample tag.

[0199] In one embodiment, the above-mentioned industrial quality inspection device further includes a small model parameter sending module, used to update the large model according to the abnormal results in the second quality inspection result; perform lightweight processing on the updated large model to obtain new small model parameters; the lightweight processing includes at least one of pruning, quantization, distillation, and low-rank decomposition; and send the new small model parameters to the production line end equipment so that the production line end equipment updates the small model according to the new small model parameters.

[0200] In one embodiment, the small model parameter sending module is further used to count the number of collaborative quality inspection tasks and / or the number of defects of the same type; when the number of collaborative quality inspection tasks exceeds a first preset threshold and / or the number of defects of the same type exceeds a second preset threshold, the large model is updated according to the abnormal results in the second quality inspection results.

[0201] In one embodiment, the second quality inspection module is further configured to: identify the feature data of the second sample data when the second sample data corresponds to one production line end device, and obtain the second quality inspection result; and fuse the feature data of the second sample data when the second sample data corresponds to at least two production line end devices, and obtain the second quality inspection result based on the fused feature data.

[0202] Specific limitations regarding industrial quality testing devices can be found in the limitations of industrial quality testing methods described above, and will not be repeated here. Each module in the aforementioned industrial quality testing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0203] In one embodiment, a communication device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0204] The first sample data of the product to be tested is identified based on a small model to obtain the first quality inspection result;

[0205] Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data;

[0206] Based on the first quality inspection result and the second quality inspection result, the target quality inspection result of the product to be tested is obtained.

[0207] In one embodiment, when the processor executes the computer program, it further performs the following steps: identifying the first sample data that the small model cannot recognize as the second sample data; and sending the second sample data to the cloud server.

[0208] In one embodiment, when the processor executes the computer program, it further implements the following steps: when the second sample data meets a first preset condition, the second sample data is sent to the cloud server in real time; the first preset condition includes a real-time requirement exceeding a preset threshold; when the second sample data meets a second preset condition, the second sample data is packaged and the packaged second sample data is periodically sent to the cloud server; the second preset condition includes a real-time requirement not exceeding a preset threshold and / or limited bandwidth resources.

[0209] In one embodiment, when the processor executes the computer program, it further performs the following steps: updating the small model according to the abnormal sample labels sent by the cloud server; the abnormal sample labels are obtained based on the abnormal results in the second quality inspection results.

[0210] In one embodiment, when the processor executes the computer program, it further performs the following steps: updating the small model according to the new small model parameters issued by the cloud server; the new small model parameters are obtained by the cloud server updating the large model according to the abnormal results in the second quality inspection results, and performing lightweight processing on the updated large model, wherein the lightweight processing includes at least one of pruning, quantization, distillation, and low-rank decomposition.

[0211] In one embodiment, when the processor executes the computer program, it further performs the following steps: acquiring raw sample data of the product to be tested; preprocessing the raw sample data to obtain the first sample data; the preprocessing includes at least one of cleaning, compression, noise reduction, and feature extraction.

[0212] In one embodiment, a communication device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0213] The second sample data of the product to be tested is identified based on the large model to obtain the second quality inspection result;

[0214] The second quality inspection result is sent to the production line end equipment so that the production line end equipment determines the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

[0215] In one embodiment, when the processor executes the computer program, it further performs the following steps: receiving the second sample data sent by the production line end device; the second sample data is the first sample data that the small model cannot recognize.

[0216] In one embodiment, when the processor executes the computer program, it further performs the following steps: obtaining an abnormal sample label based on the abnormal result in the second quality inspection result; sending the abnormal sample label to the production line end device so that the production line end device updates the small model based on the abnormal sample label.

[0217] In one embodiment, when the processor executes the computer program, it further performs the following steps: updating the large model based on the abnormal results in the second quality inspection results; performing lightweight processing on the updated large model to obtain new small model parameters; the lightweight processing includes at least one of pruning, quantization, distillation, and low-rank decomposition; and sending the new small model parameters to the production line end device so that the production line end device updates the small model according to the new small model parameters.

[0218] In one embodiment, when the processor executes the computer program, it further performs the following steps: counting the number of collaborative quality inspection tasks and / or the number of defects of the same type; when the number of collaborative quality inspection tasks exceeds a first preset threshold and / or the number of defects of the same type exceeds a second preset threshold, updating the large model based on the abnormal results in the second quality inspection results.

[0219] In one embodiment, when the processor executes the computer program, it further implements the following steps: when the second sample data corresponds to one production line end device, the feature data of the second sample data is identified to obtain the second quality inspection result; when the second sample data corresponds to at least two production line end devices, the feature data of the second sample data is fused to obtain the second quality inspection result based on the fused feature data.

[0220] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0221] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps in the above-described method embodiments.

[0222] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0223] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0224] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An industrial quality inspection method, characterized in that, The method is applied to production line end equipment, including: The first sample data of the product to be tested is identified based on a small model to obtain the first quality inspection result; Receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data; Based on the first quality inspection result and the second quality inspection result, the target quality inspection result of the product to be tested is obtained.

2. The method according to claim 1, characterized in that, The method further includes: The first sample data that the small model could not recognize was identified as the second sample data; The second sample data is sent to the cloud server.

3. The method according to claim 2, characterized in that, Sending the second sample data to the cloud server includes: When the second sample data meets the first preset condition, the second sample data is sent to the cloud server in real time; the first preset condition includes a real-time requirement exceeding a preset threshold. When the second sample data meets the second preset conditions, the second sample data is packaged and periodically sent to the cloud server; the second preset conditions include real-time requirements not exceeding a preset threshold and / or limited bandwidth resources.

4. The method according to claim 1, characterized in that, The method further includes: The small model is updated based on the abnormal sample labels sent by the cloud server; the abnormal sample labels are obtained based on the abnormal results in the second quality inspection results.

5. The method according to claim 1, characterized in that, The method further includes: The small model is updated according to the new small model parameters issued by the cloud server; the new small model parameters are obtained by the cloud server updating the large model according to the abnormal results in the second quality inspection results, and the updated large model is subjected to lightweight processing, the lightweight processing including at least one of pruning, quantization, distillation, and low-rank decomposition.

6. The method according to claim 1, characterized in that, The method further includes: Obtain the original sample data of the product to be tested; The original sample data is preprocessed to obtain the first sample data; the preprocessing includes at least one of cleaning, compression, noise reduction and feature extraction.

7. An industrial quality inspection method, characterized in that, The method is applied to a cloud server and includes: The second sample data of the product to be tested is identified based on the large model to obtain the second quality inspection result; The second quality inspection result is sent to the production line end equipment so that the production line end equipment determines the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

8. The method according to claim 7, characterized in that, The method further includes: The system receives the second sample data sent by the production line end equipment; the second sample data is the first sample data that the small model cannot recognize.

9. The method according to claim 7, characterized in that, The method further includes: Based on the abnormal results in the second quality inspection results, the abnormal sample labels are obtained; The abnormal sample tag is sent to the production line end device so that the production line end device updates the small model based on the abnormal sample tag.

10. The method according to claim 7, characterized in that, The method further includes: The large model is updated based on the abnormal results in the second quality inspection results; The updated large model is lightweighted to obtain new small model parameters; the lightweighting process includes at least one of pruning, quantization, distillation, and low-rank decomposition. The new small model parameters are sent to the production line end equipment so that the production line end equipment updates the small model according to the new small model parameters.

11. The method according to claim 10, characterized in that, The step of updating the large model based on the abnormal results in the second quality inspection results includes: Count the number of collaborative quality inspection tasks and / or the number of defects of the same type; When the number of collaborative quality inspection tasks exceeds a first preset threshold and / or the number of defects of the same type exceeds a second preset threshold, the large model is updated based on the abnormal results in the second quality inspection results.

12. The method according to claim 7, characterized in that, The second quality inspection result is obtained by identifying the second sample data of the product to be tested based on the large model, including: When the second sample data corresponds to a production line end device, the feature data of the second sample data is identified to obtain the second quality inspection result; When the second sample data corresponds to at least two production line end devices, the feature data of the second sample data are fused, and the second quality inspection result is obtained based on the fused feature data.

13. An industrial quality inspection device, characterized in that, The device is applied to production line end equipment and includes: The first quality inspection module is used to identify the first sample data of the product to be inspected based on a small model and obtain the first quality inspection result. The result receiving module is used to receive the second quality inspection result sent by the cloud server; the second quality inspection result is obtained by the cloud server based on the large model to identify the second sample data; The result determination module is used to obtain the target quality inspection result of the product to be inspected based on the first quality inspection result and the second quality inspection result.

14. An industrial quality inspection device, characterized in that, The device is used on a cloud server and includes: The second quality inspection module is used to identify the second sample data of the product to be tested based on the large model, and obtain the second quality inspection result. The result sending module is used to send the second quality inspection result to the production line end equipment so that the production line end equipment can determine the target quality inspection result based on the first quality inspection result and the second quality inspection result; the first quality inspection result is obtained by the production line end equipment based on the first sample data of the product to be inspected using a small model.

15. A communication device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6 or 7 to 12.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6 or 7 to 12.

17. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 6 or 7 to 12.

Images