Industrial product quality online detection method based on heterogeneous model dynamic collaboration

Abstract

The invention relates to the technical field of intelligent manufacturing, and particularly discloses an industrial product quality online detection method based on heterogeneous model dynamic collaboration, and the method comprises the steps: obtaining a visible light image and a multispectral image of a to-be-detected industrial product; sequentially training and reducing the defect detection model to obtain a compact defect detection model with a high entropy value; inputting a visible light image and a multispectral image of an industrial product to be detected into the compact defect detection model for defect detection so as to output a defect detection result of the industrial product to be detected; transmitting the defect detection result of the industrial product to be detected from the end side to a service side; and generating a quality evaluation report of the to-be-detected industrial product on the service side according to the defect detection result of the to-be-detected industrial product. According to the industrial product quality online detection method based on heterogeneous model dynamic collaboration, the reliability of industrial product quality detection can be ensured.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and more specifically, to an online quality inspection method for industrial products based on dynamic collaboration of heterogeneous models. Background Technology

[0002] In the wave of intelligent transformation in the manufacturing industry, industrial product quality inspection, as a key link in ensuring product reliability and enhancing market competitiveness, is undergoing a profound transformation in its technological system, moving from traditional manual-dominated to intelligent, automated, and high-precision processes. Traditional industrial product quality inspection methods have long relied on manual visual observation, manual measuring tools, and offline sampling inspection. This not only suffers from low inspection efficiency and high labor costs but also makes it difficult to avoid the risks of missed or false detections due to human error. Especially when facing mass production, complex structural parts, or the inspection of minute defects, traditional technologies can no longer meet the real-time, comprehensive, and accurate quality control requirements of intelligent manufacturing. With the popularization of intelligent manufacturing technologies, the automation level of production lines has increased significantly, product updates and iterations have accelerated, and personalized customization production models are gradually emerging, placing higher-dimensional demands on quality inspection. On the one hand, it is necessary to achieve real-time linkage with the production line to complete closed-loop control of "online inspection - data feedback - process adjustment" to avoid the generation of batches of defective products. On the other hand, it is necessary to address the inspection flexibility in multi-variety, small-batch production scenarios, as well as the special requirements of emerging fields such as new energy vehicles, high-end equipment, and semiconductors for ultra-precision inspection and multi-parameter synchronous inspection.

[0003] Current industrial product quality inspection typically uses single-modal optical equipment to acquire product appearance characteristics and constructs a mapping model from image data to defect types and damage levels using prior-driven or data-driven methods. After obtaining the defect type and degree, a quality report is manually generated and published. This process not only fails to capture complete defect information of industrial products but also cannot achieve online quality report publication, which seriously affects the progress of subsequent process flows.

[0004] Therefore, how to construct an online inspection method for industrial product quality, deeply explore the mapping relationship between the appearance data and defects of industrial products through the end-side collection and analysis of multimodal data, and then transmit it to the server in a secure and efficient manner through the channel, and release accurate, credible and reliable industrial product quality assessment reports in real time is an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an online inspection method for industrial product quality based on heterogeneous model dynamic collaboration. This method ensures the reliability of industrial product quality inspection through accurate defect detection, secure data transmission, and reliable report generation.

[0006] As a first aspect of the present invention, an online quality inspection method for industrial products based on heterogeneous model dynamic collaboration is provided, comprising the following steps: Step S1: Acquire visible light and multispectral images of the industrial product to be inspected; Step S2: Train and reduce the defect detection model sequentially to obtain a compact defect detection model with high entropy. Step S3: Input the visible light image and multispectral image of the industrial product to be inspected into the compact defect detection model for defect detection, and output the defect detection result of the industrial product to be inspected; Step S4: Transmit the defect detection results of the industrial product to be inspected from the terminal side to the server side; Step S5: Generate a quality assessment report for the industrial product under test on the service side based on the defect detection results of the industrial product under test.

[0007] Further, step S1 includes: Spatially align the visible light image and multispectral image of the industrial product to be tested to obtain aligned visible light image and aligned multispectral image of the industrial product to be tested.

[0008] Further, step S2 includes: Acquire training data and construct the defect detection model, wherein the defect detection model includes multiple instance segmentation models; The training data is augmented using data augmentation, and the defect detection model is trained using error gradient backpropagation to obtain a pre-trained defect detection model. The pre-trained defect detection model is reduced using a multi-agent reinforcement learning method to obtain the high-entropy compact defect detection model.

[0009] Furthermore, the reduction of the pre-trained defect detection model using a multi-agent reinforcement learning method to obtain the high-entropy compact defect detection model includes: The architecture search agent dynamically evaluates and selects multiple predefined basic components in the model during the early stages of training. The dynamic weight fusion mechanism determines the best combination of basic components for each training stage, so that the model architecture gradually evolves towards optimal performance. The parameter quantization agent utilizes simulated quantization nodes to provide quantization simulation information during the forward and backward propagation of the model, thereby guiding the fine-tuning of model parameters; The performance evaluation agent monitors various performance metrics of the model on the validation set in real time, and dynamically controls the learning pace and exploration intensity of the architecture search agent and the parameter quantization agent to ensure that the model architecture search and parameter quantization training work together.

[0010] Furthermore, the process of dynamically evaluating and selecting multiple predefined basic components in the model during the early stages of training through an architecture search agent, and determining the optimal combination of basic components for each training stage through a dynamic weight fusion mechanism, includes: The predefined basic components include an external attention mechanism module, a window attention mechanism module, and a self-attention mechanism module. The output results of the three attention mechanism modules are fused through dynamic weights to obtain the fused output result. The architecture search agent continuously and dynamically optimizes the combination of the three attention mechanism modules based on the fused output to obtain the optimal model architecture.

[0011] Furthermore, the step of dynamically controlling the learning pace and exploration intensity of the architecture search agent and the parameter quantization agent by real-time monitoring of various performance indicators of the model on the validation set through a performance evaluation agent includes: After each training cycle, the performance evaluation agent determines the learning rate adjustment strategy of the architecture search agent and the parameter quantization agent for the next training cycle based on the model's evaluation metrics and in combination with the number of model parameters and the number of floating-point operations. If the detection accuracy of the current model is greater than the target threshold, the performance evaluation agent will increase the learning rate of the architecture search agent and the parameter quantization agent to enhance the exploration efforts; If the detection accuracy of the current model is lower than or equal to the target threshold, the performance evaluation agent will reduce the learning rate of the architecture search agent and the parameter quantization agent to consolidate the current model architecture.

[0012] Further, step S3 includes: The visible light image and multispectral image of the industrial product to be inspected are fused to obtain a fused image, and the fused image is input into the first instance segmentation model for defect detection to output the first defect detection result of the industrial product to be inspected. Feature extraction is performed on the visible light image and the multispectral image of the industrial product to be inspected, respectively. The extracted features of the visible light image and the multispectral image are then fused to obtain fused features. The fused features are then input into the second instance segmentation model for defect detection to output the second defect detection result of the industrial product to be inspected. The visible light image of the industrial product to be inspected is input into the third instance segmentation model for defect detection, so as to output the third defect detection result of the industrial product to be inspected. The multispectral image of the industrial product to be inspected is input into the fourth instance segmentation model for defect detection, so as to output the fourth defect detection result of the industrial product to be inspected. The first, second, third, and fourth defect detection results of the industrial product under test are fused together to output the final defect detection result of the industrial product under test; wherein, the final defect detection result of the industrial product under test includes the defect type and the degree of damage.

[0013] Furthermore, the final defect detection result of the industrial product to be tested o The decision fusion formula is as follows: ; In the formula, o This indicates the final defect detection result of the industrial product to be tested. c Indicates the confidence level of different defect detection results. These represent the confidence levels of the first defect detection result, the second defect detection result, the third defect detection result, and the fourth defect detection result, respectively.

[0014] Furthermore, step S4 also includes: On the terminal side, a preset binary encoder is used to encode the final defect detection result of the industrial product to be inspected, so as to obtain the encoded defect detection result of the industrial product to be inspected. The encoded defect detection result of the industrial product to be inspected is then transmitted to the service side through the channel. The encoding formula is as follows: ; In the formula, o This indicates the final defect detection result of the industrial product to be tested. Indicates a preset binary encoder. Indicates a preset time-varying encryption signal. This indicates the detection result of the coding defect of the industrial product to be tested; On the service side, the received encoded defect detection results of the industrial product to be inspected are decoded using a pair of preset binary decoders to obtain the decoded defect detection results of the industrial product to be inspected; wherein, the decoding formula is as follows: ; In the formula, Indicates and The corresponding preset binary decoder, This indicates the decoding defect detection result of the industrial product to be tested.

[0015] Further, step S5 includes: On the service side, the decoding defect detection results of the industrial product to be tested are... and domain knowledge base The data is input into a multimodal large language model to generate a quality assessment report for the industrial product under test. r The quality assessment report of the industrial product to be tested is saved as a PDF file. r The process is defined as follows: ; In the formula, Represents a multimodal large language model. Represents a domain knowledge base. r This refers to the quality assessment report of the industrial product to be tested.

[0016] The online industrial product quality inspection method based on heterogeneous model dynamic collaboration provided by this invention has the following advantages: Based on visible-spectral multimodal data, this invention accurately identifies product defect types and damage levels through a compact industrial product defect detection model deployed on the edge. Semantic communication is used to encrypt and encode product defect information, which is then transmitted to the server via a channel. A large language model is then used to generate a reliable product quality report online, providing reference and data support for guiding subsequent process flows. Furthermore, this invention can expand the number of modalities used by the defect detection model according to product type and characteristics, acquiring multimodal data including images, sound, intensity, and vibration, ensuring the accuracy of defect detection and possessing broad engineering application and promotional prospects. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof.

[0018] Figure 1 The flowchart shows the online quality inspection method for industrial products based on dynamic collaboration of heterogeneous models provided by this invention.

[0019] Figure 2 The flowchart illustrates the specific implementation of the online quality detection method for industrial products based on heterogeneous model dynamic collaboration provided by this invention.

[0020] Figure 3 This is an architecture diagram of the industrial product defect detection module based on multimodal joint embedding provided by the present invention.

[0021] Figure 4This is a schematic diagram of data transmission based on semantic communication provided by the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the publication of a product quality report based on search augmentation provided by the present invention. Detailed Implementation

[0023] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an online industrial product quality detection method based on heterogeneous model dynamic collaboration proposed according to the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] This embodiment provides an online quality inspection method for industrial products based on heterogeneous model dynamic collaboration, such as... Figure 1 As shown, an online quality inspection method for industrial products based on heterogeneous model dynamic collaboration includes the following steps: Step S1: Acquire visible light and multispectral images of the industrial product to be inspected; Preferably, such as Figure 2 As shown, step S1 includes: The affine matrix is ​​solved using a typical labeling method to spatially align the visible light image and multispectral image of the industrial product to be tested, so as to obtain the aligned visible light image and aligned multispectral image of the industrial product to be tested.

[0026] Step S2: Train and reduce the defect detection model sequentially to obtain a compact defect detection model with high entropy. Preferably, such as Figure 2 As shown, step S2 includes: Acquire training data and construct the defect detection model, wherein the defect detection model includes multiple instance segmentation models; The training data is augmented using data augmentation, and the defect detection model is trained using error gradient backpropagation to obtain a pre-trained defect detection model. The data augmentation refers to performing specificity calculations on reference samples and determining the augmentation factor based on the specificity values ​​of different samples. On this basis, a pre-trained diffusion model is used for data augmentation.

[0027] The pre-trained defect detection model is reduced using a multi-agent reinforcement learning method to obtain the high-entropy compact defect detection model.

[0028] Specifically, the reduction of the pre-trained defect detection model using a multi-agent reinforcement learning method to obtain the high-entropy compact defect detection model includes: The architecture search agent dynamically evaluates and selects multiple predefined basic components in the model during the early stages of training. The dynamic weight fusion mechanism determines the best combination of basic components for each training stage, so that the model architecture gradually evolves towards optimal performance. Furthermore, the process of dynamically evaluating and selecting multiple predefined basic components in the model during the early stages of training through an architecture search agent, and determining the optimal combination of basic components for each training stage through a dynamic weight fusion mechanism, includes: The predefined basic components include an external attention mechanism module, a window attention mechanism module, and a self-attention mechanism module. The output results of the three attention mechanism modules are fused through dynamic weights to obtain the fused output result. The architecture search agent continuously and dynamically optimizes the combination of the three attention mechanism modules based on the fused output to obtain the optimal model architecture.

[0029] Specifically, the architecture search agent is used to dynamically adjust the combination of basic component modules in the network. This embodiment employs three attention mechanisms: external attention, window attention, and self-attention. Each combined attention block within the network dynamically fuses the outputs of these three attention modules using weights. ; in, For the first i The dynamic weights of each attention mechanism module, For the first iThe output of each attention mechanism module The fused output is normalized using the Soft max function to determine the dynamic weights of each attention mechanism. : ; in, and All represent the weight vectors of different attention mechanism modules before SoftMax.

[0030] The parameter quantization agent utilizes simulated quantization nodes to provide quantization simulation information during the forward and backward propagation of the model, thereby guiding the fine-tuning of model parameters; In this embodiment of the invention, the parameter quantization agent introduces Quantization Aware Training (QAT). By inserting simulated quantization nodes (FakeQuant) inside the model, the actual quantization effect is simulated during training, thereby improving the performance stability of the model in the actual quantization deployment stage. The network model realizes the simulated quantization of weights and activation values ​​through simulated quantization nodes (FakeQuant). The quantization calculation is defined as: ; in, Represents the original weight parameters. This represents the truncation function. This represents the rounding function, where s is the quantization scale factor and z is the quantization zero. and These are the upper and lower limits for quantization. This embodiment uses INT8 quantization. and The values ​​are -128 and 127 respectively. The parameter quantization agent updates the quantization parameters through backpropagation, gradually approaching the optimal quantization configuration.

[0031] The performance evaluation agent monitors various performance metrics of the model on the validation set in real time, and dynamically controls the learning pace and exploration intensity of the architecture search agent and the parameter quantization agent to ensure that the model architecture search and parameter quantization training work together.

[0032] Furthermore, the step of dynamically controlling the learning pace and exploration intensity of the architecture search agent and the parameter quantization agent by real-time monitoring of various performance indicators of the model on the validation set through a performance evaluation agent includes: After each training cycle, the performance evaluation agent determines the learning rate adjustment strategy of the architecture search agent and the parameter quantization agent for the next training cycle based on the model's evaluation metrics and in combination with the number of model parameters (Params) and the number of floating-point operations (FLOPs). The performance feedback mechanism dynamically adjusts the learning rate according to the following rules: If the detection accuracy of the current model is greater than the target threshold, the performance evaluation agent will increase the learning rate of the architecture search agent and the parameter quantization agent to enhance the exploration effort: ; If the detection accuracy of the current model is lower than or equal to the target threshold, the performance evaluation agent will reduce the learning rate of the architecture search agent and the parameter quantization agent to consolidate the current model architecture. ; in, and These represent the new learning rate and the old learning rate, respectively. , Both involve adjusting the step size coefficient, thereby achieving closed-loop dynamic adjustment of training feedback through this mechanism.

[0033] In this invention, an architecture search agent, a parameter quantization agent, and a performance evaluation agent are used to reduce and compact the pre-trained instance segmentation model, thereby reducing the runtime load while maintaining the model's detection capabilities. The architecture search agent and parameter quantization agent aim to reduce storage and computational consumption, while the performance evaluation agent aims to reduce the complexity of the detection model while ensuring detection performance.

[0034] Step S3: Input the visible light image and multispectral image of the industrial product to be inspected into the compact defect detection model for defect detection, and output the defect detection result of the industrial product to be inspected; It should be noted that, through a compact instance segmentation model, the product defect type and damage level are accurately identified based on spatially aligned industrial visible-spectral multimodal data. The performance and efficiency of defect detection are guaranteed by complete sensing data and a high-entropy detection model.

[0035] Preferably, such as Figure 3 As shown, step S3 includes: The visible light image and multispectral image of the industrial product to be inspected are fused to obtain a fused image, and the fused image is input into the first instance segmentation model for defect detection to output the first defect detection result of the industrial product to be inspected. Feature extraction is performed on the visible light image and the multispectral image of the industrial product to be inspected, respectively. The extracted features of the visible light image and the multispectral image are then fused to obtain fused features. The fused features are then input into the second instance segmentation model for defect detection to output the second defect detection result of the industrial product to be inspected. The visible light image of the industrial product to be inspected is input into the third instance segmentation model for defect detection, so as to output the third defect detection result of the industrial product to be inspected. The multispectral image of the industrial product to be inspected is input into the fourth instance segmentation model for defect detection, so as to output the fourth defect detection result of the industrial product to be inspected. The first, second, third, and fourth defect detection results of the industrial product under test are fused together to output the final defect detection result of the industrial product under test; wherein, the final defect detection result of the industrial product under test includes the defect type and the degree of damage.

[0036] In this embodiment of the invention, a multi-level fusion scheme integrating image data-level fusion, feature-level fusion, and decision-level fusion is adopted to ensure the adaptability of the instance segmentation model to different lighting conditions in open scenes.

[0037] Specifically, during the model training phase, a cost function is used to optimize the network parameters. The formula for the cost function is as follows: ; In the formula, and Let represent the cost function of the first instance segmentation model and the cost function of the second instance segmentation model, respectively. and Let $\mathbf$ represent the cost functions of the third instance segmentation model and the fourth instance segmentation model, respectively.

[0038] Specifically, the final defect detection result of the industrial product to be inspected. o The decision fusion formula is as follows: ; In the formula, o This indicates the final defect detection result of the industrial product to be tested. c Indicates the confidence level of different defect detection results. These represent the confidence levels of the first defect detection result, the second defect detection result, the third defect detection result, and the fourth defect detection result, respectively. In other words, this invention uses a highest confidence strategy to make the final decision on the detection results.

[0039] Step S4: Transmit the defect detection results of the industrial product to be inspected from the terminal side to the server side; Preferably, such as Figure 4 As shown, step S4 further includes: On the terminal side, a preset binary encoder is used to encode the final defect detection result of the industrial product to be inspected, so as to obtain the encoded defect detection result of the industrial product to be inspected. The encoded defect detection result of the industrial product to be inspected is then transmitted to the service side through the channel. The encoding formula is as follows: ; In the formula, o This indicates the final defect detection result of the industrial product to be tested. Indicates a preset binary encoder. Indicates a preset time-varying encryption signal. This indicates the detection result of the coding defect of the industrial product to be tested; On the service side, the received encoded defect detection results of the industrial product to be inspected are decoded using a pair of preset binary decoders to obtain the decoded defect detection results of the industrial product to be inspected; wherein, the decoding formula is as follows: ; In the formula, Indicates and The corresponding preset binary decoder, This indicates the decoding defect detection result of the industrial product to be tested.

[0040] In this embodiment of the invention, the data interaction between the terminal and the server is carried out by using a binary encoding-semantic transmission-content decoding method, which can not only improve the data transmission efficiency, but also largely guarantee the security of the data transmission process.

[0041] It should be noted that a binary neural network is used to process the detection results, reducing the reconstruction data error caused by high-level floating-point operations. Secondly, pairs of pre-defined time-varying encryption signals are added during the encoding and decoding stages to further ensure the security of data transmission.

[0042] Step S5: Generate a quality assessment report for the industrial product under test on the service side based on the defect detection results of the industrial product under test.

[0043] Preferably, such as Figure 5 As shown, step S5 includes: On the service side, the decoding defect detection results of the industrial product to be tested are... and domain knowledge base The data is input into a multimodal large language model to generate a quality assessment report for the industrial product under test. r The quality assessment report of the industrial product to be tested is saved as a PDF file. r The process is defined as follows: ; In the formula, Represents a multimodal large language model. Represents a domain knowledge base. r This refers to the quality assessment report of the industrial product to be tested.

[0044] It should be noted that the features in the domain knowledge base are stored in the form of a binary tree, which can greatly reduce the amount of computation generated by retrieval augmentation, thereby improving the efficiency of publishing industrial product quality reports.

[0045] In this embodiment of the invention, based on the detected types and extent of industrial product defects, constrained by knowledge of the industrial product defect analysis domain, and driven by a large language model, a reliable and trustworthy product quality assessment report is generated in real time, providing accurate data support for subsequent process flows.

[0046] In embodiments of the present invention, such as Figure 2 As shown, the overall execution flow of this invention includes steps such as data preprocessing, edge-side industrial product defect detection, industrial product defect channel transmission, and product quality report generation. The core of data preprocessing is the spatial alignment of multi-source image data. Edge-side small-model defect detection includes the training, reduction, and inference of the instance segmentation model. The server-side large-model report generation model is mainly used to coordinate domain information and observation information to form a credible and reliable product quality information description.

[0047] This invention provides an online industrial product quality inspection method based on heterogeneous model dynamic collaboration, comprising product defect detection based on multimodal joint embedding, data transmission based on semantic communication, and product quality report publishing based on retrieval augmentation. The product defect detection based on multimodal joint embedding utilizes visible light and multispectral data, employing a compact instance segmentation model to assess the defect type and severity of industrial products. The data transmission based on semantic communication adopts a bandwidth-for-computation approach, reducing data transmission costs and improving data transmission security through end-side semantic encoding, channel semantic transmission, and server-side content decoding. The product quality report publishing based on retrieval augmentation forms a knowledge base based on historical data and domain knowledge, and uses an open-source large language model to summarize, organize, and conclude the product defect detection results, generating a product quality assessment report. This method achieves complete industrial product data acquisition, accurate defect detection, secure data transmission, and reliable assessment reports, providing important theoretical basis and technical support for online industrial product quality inspection.

[0048] The present invention provides an online industrial product quality inspection method based on dynamic collaboration of heterogeneous models. It takes into account multiple aspects such as the completeness of data perception, the security of information transmission, and the reliability of evaluation reports. It balances the efficiency of the small model on the edge side and the knowledge richness of the large model on the service side, while also considering the data security when the large and small models are dynamically collaborated. It is applicable to various industrial product quality inspection problems in the intelligent manufacturing industry chain and improves the accuracy, security and stability of online industrial product quality inspection.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An industrial product quality online detection method based on heterogeneous model dynamic collaboration, characterized in that, The method comprises the following steps: Step S1: obtaining visible light images and multispectral images of an industrial product to be detected; Step S2: sequentially training and reducing a defect detection model to obtain a compact defect detection model with high entropy value; Step S3: inputting the visible light images and the multispectral images of the industrial product to be detected into the compact defect detection model for defect detection to output a defect detection result of the industrial product to be detected; Step S4: transmitting the defect detection result of the industrial product to be detected from an end side to a service side; Step S5: generating a quality evaluation report of the industrial product to be detected on the service side according to the defect detection result of the industrial product to be detected.

2. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 1, characterized in that, In the step S1, the following steps are included: The visible light images and the multispectral images of the industrial product to be detected are spatially aligned to obtain aligned visible light images and aligned multispectral images of the industrial product to be detected.

3. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 1, characterized in that, In the step S2, the following steps are included: Training data are obtained, and the defect detection model is constructed, wherein the defect detection model comprises a plurality of instance segmentation models; The training data are expanded in a data augmentation manner, and the defect detection model is trained in an error gradient back propagation manner to obtain a pre-trained defect detection model; The pre-trained defect detection model is reduced using a multi-agent reinforcement learning method to obtain the compact defect detection model with high entropy value.

4. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 3, characterized in that, In the step of reducing the pre-trained defect detection model using the multi-agent reinforcement learning method to obtain the compact defect detection model with high entropy value, the following steps are included: A model architecture search agent dynamically evaluates and selects a plurality of predefined basic components in the model at an early training stage, determines the best basic component combination at each training stage through a dynamic weight fusion mechanism, and makes the model architecture gradually evolve towards the optimal performance direction; A parameter quantization agent provides quantization simulation information in model forward and backward propagation using a simulation quantization node, thereby guiding fine adjustment of model parameters; A performance evaluation agent monitors various performance indicators of the model on a validation set in real time, dynamically controls learning pace and exploration intensity of the architecture search agent and the parameter quantization agent, and ensures that model architecture search and parameter quantization training are cooperatively matched.

5. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 4, characterized in that, In the step of dynamically evaluating and selecting the plurality of predefined basic components in the model at the early training stage by the architecture search agent and determining the best basic component combination at each training stage through the dynamic weight fusion mechanism, the following steps are included: The plurality of predefined basic components include an external attention mechanism module, a window attention mechanism module, and a self-attention mechanism module, and the output results of the three attention mechanism modules are fused to obtain a fused output result; The architecture search agent continuously dynamically optimizes the combination of the three attention mechanism modules according to the fused output result to obtain a model architecture with optimal performance.

6. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 4, characterized in that, In the step of monitoring various performance indicators of the model on the validation set in real time by the performance evaluation agent and dynamically controlling learning pace and exploration intensity of the architecture search agent and the parameter quantization agent, the following steps are included: At the end of each training cycle, the performance evaluation agent determines the learning rate adjustment strategy of the architecture search agent and the parameter quantization agent in the next training cycle according to the evaluation index of the model, in combination with the model parameter quantity and the number of floating point operations; If the detection accuracy of the current model is greater than the target threshold, the performance evaluation agent will increase the learning rate of the architecture search agent and the parameter quantization agent to enhance the exploration intensity; If the detection accuracy of the current model is less than or equal to the target threshold, the performance evaluation agent will reduce the learning rate of the architecture search agent and the parameter quantization agent to consolidate the current model architecture.

7. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 3, characterized in that, The step S3 comprises: fuse the visible light image and the multispectral image of the industrial product to be detected to obtain a fused image, and input the fused image into a first instance segmentation model for defect detection to output a first defect detection result of the industrial product to be detected; extract features from the visible light image and the multispectral image of the industrial product to be detected respectively, fuse the extracted features of the visible light image and the multispectral image to obtain fused features, and then input the fused features into a second instance segmentation model for defect detection to output a second defect detection result of the industrial product to be detected; input the visible light image of the industrial product to be detected into a third instance segmentation model for defect detection to output a third defect detection result of the industrial product to be detected; input the multispectral image of the industrial product to be detected into a fourth instance segmentation model for defect detection to output a fourth defect detection result of the industrial product to be detected; fuse the first defect detection result, the second defect detection result, the third defect detection result and the fourth defect detection result of the industrial product to be detected to output a final defect detection result of the industrial product to be detected; wherein the final defect detection result of the industrial product to be detected includes defect type and damage degree.

8. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 7, characterized in that, The final defect detection result of the industrial product to be detected o The decision fusion formula is as follows: ； In the formula, o represents the final defect detection result of the industrial product to be detected, c represents the confidence of different defect detection results, respectively represent the confidence of the first defect detection result, the confidence of the second defect detection result, the confidence of the third defect detection result, and the confidence of the fourth defect detection result.

9. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 7, characterized in that, The step S4 further comprises: encode the final defect detection result of the industrial product to be detected using a preset binary encoder at the end side to obtain an encoded defect detection result of the industrial product to be detected, and transmit the encoded defect detection result of the industrial product to be detected to the service side through a channel; wherein the encoding formula is as follows: ； In the formula, o represents the final defect detection result of the industrial product to be detected, represents a preset binary encoder, represents a preset time-varying encryption signal, represents the encoded defect detection result of the industrial product to be detected; decode the received encoded defect detection result of the industrial product to be detected using a pair of preset binary decoders at the service side to obtain a decoded defect detection result of the industrial product to be detected; wherein the decoding formula is as follows: ； In the formula, indicates the preset binary decoder corresponding to indicates the preset binary decoder corresponding to indicates the decoding defect detection result of the industrial product to be detected.

10. The industrial product quality online detection method based on heterogeneous model dynamic collaboration according to claim 9, characterized in that, The step S5 comprises: On the service side, the decoding defect detection results of the industrial product to be tested are... and domain knowledge base The data is input into a multimodal large language model to generate a quality assessment report for the industrial product under test. r The quality assessment report of the industrial product to be tested is saved as a PDF file. r The process is defined as follows: ； In the formula, represents a multi-modal large language model, represents a domain knowledge base, r represents a quality evaluation report of the industrial product to be detected.

Images