AI quality inspection system based on cloud upgrade

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an AI quality inspection system based on cloud upgrade, and the system comprises a mobile robot data collection module which is used for collecting data on an industrial production line; the edge node intelligent preprocessing module is used for receiving data and carrying out de-noising, feature extraction and screening; the cloud incremental learning and database self-iteration module is used for executing incremental learning on the data; the containerization deployment and cross-platform synchronization module is used for packaging the updated optimized AI quality inspection model into a container mirror image and performing cross-platform deployment; the isolation test and self-adaptive verification module is used for verifying the updated optimized AI quality inspection model; and an intelligent increment synchronous pushing module. According to the technical scheme of the intelligent incremental synchronous pushing module, bandwidth consumption between the cloud end and the equipment end is effectively reduced through incremental updating and differential parameter transmission modes, and the synchronization efficiency from the cloud end to the edge equipment is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI quality inspection system based on cloud upgrades. Background Art

[0002] Currently, traditional AI quality inspection systems generally use full synchronization for cloud-based model updates. For each update, the entire model file must be pushed from the cloud to all edge nodes. This not only consumes significant network bandwidth, but also tends to be slow, significantly increasing update time in large-scale deployments. Existing technologies often waste bandwidth resources and strain device storage, making it difficult to meet the demands of efficient and frequent updates.

[0003] Existing solutions for these technologies often require retransmitting the entire AI model to the device during updates, increasing computational and storage pressures. Each update requires full loading and verification on the device, increasing system response time and potentially causing system interruptions or crashes. This is especially true in demanding scenarios like production line monitoring and real-time testing, where system stability and real-time performance cannot be guaranteed during updates.

[0004] Currently, most cloud synchronization solutions don't consider the actual device load and network conditions when scheduling push tasks. The push process is typically static and lacks flexible adjustment mechanisms. Consequently, push speeds can drop significantly in the face of network congestion or heavy device loads, impacting the overall efficiency and stability of the system. In large-scale AI quality inspection systems, static push strategies lead to delays and instability in system push, making it difficult to cope with complex network environments and dynamic load demands.

[0005] Furthermore, existing technologies often interrupt synchronization when a network outage or device failure occurs, resulting in data loss. In these cases, the device must restart full synchronization, wasting bandwidth, time, and computing resources. This approach cannot guarantee stability and reliability in unstable environments, especially during large-scale system upgrades, where retransmissions caused by interruptions are unavoidable. Summary of the Invention

[0006] The purpose of this invention is to provide an AI quality inspection system based on cloud upgrade, which solves the problems of excessive bandwidth consumption, low synchronization efficiency, heavy system burden and inability to guarantee synchronization stability in unstable environments caused by full synchronous updates in existing AI quality inspection systems.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: an AI quality inspection system based on cloud upgrade, comprising: Mobile robot data acquisition module, used to collect data on industrial production lines, including product image data, environmental parameters and robot motion state parameters; The edge node intelligent pre-processing module is connected to the mobile robot data acquisition module to perform denoising, feature extraction and screening on the received data, and send the processed data to the cloud; The cloud-based incremental learning and database self-iteration module is connected to the edge node intelligent pre-processing module to perform incremental learning on the data, optimize the AI ​​quality inspection model, and self-update the database; The containerized deployment and cross-platform synchronization module is connected to the cloud-based incremental learning and database self-iteration module to encapsulate the updated optimized AI quality inspection model into a container image and deploy it across platforms. The isolation test and adaptive verification module is connected to the cloud-based incremental learning and database self-iteration module to verify the updated optimized AI quality inspection model and ensure its performance meets the requirements; The intelligent incremental synchronization push module is connected to the cloud-based incremental learning and database self-iteration module to push the incrementally updated optimized AI quality inspection model to the device side, transmitting only the differences.

[0008] Preferably, the data acquisition module collects data through a multimodal sensor, which includes a camera, a temperature and humidity sensor, and a lidar. The collected data of the sensor includes product surface image, light intensity, ambient temperature, humidity, and robot position.

[0009] Preferably, the edge node intelligent preprocessing module includes the following units: Data denoising unit, used to eliminate noise in collected data and smooth the data; Convolutional neural network feature extraction unit, used to extract important features from image data; The feature screening unit is used to screen out valid features according to the set feature selection criteria.

[0010] Preferably, the feature screening unit screens the extracted features based on a sparse optimization algorithm, specifically by solving the following optimization problem: Among them, w represents the feature weight vector, F represents the feature matrix, y represents the label or target value vector of the sample, and λ represents the regularization coefficient. Indicates the optimization of variable w.

[0011] Preferably, the cloud-based incremental learning and database self-iteration module includes: Incremental learning unit, used to fine-tune and optimize AI models based on newly added data sets; The database self-iteration unit is used to update the model version and quality inspection results in the database, and to self-optimize the system.

[0012] Preferably, the incremental learning adopts a small batch gradient update strategy, and the optimization objective function is defined as follows: Among them, θ represents the model parameter set, N represents the number of samples in the current mini-batch, and x i Represents the input feature vector of the i-th sample, y i represents the true label corresponding to the i-th sample, f(x i ;θ) represents the model with parameter θ on input x i The predicted output on Represents the loss function, which measures the difference between the predicted value and the true value, λ represents the regularization coefficient, and R(θ) represents the regularization term.

[0013] Preferably, the containerized deployment and cross-platform synchronization module includes: Containerized packaging unit, used to package the updated AI quality inspection model and its dependent environment into a container image; The image synchronization unit is used to synchronize container images to the cloud or edge nodes and select images for deployment based on device hardware.

[0014] Preferably, the isolation test and adaptive verification module includes: The accuracy verification unit is used to evaluate the prediction accuracy of the new model by calculating the classification accuracy and recall indicators; Among them, TP is the number of samples predicted to be positive and actually positive, TN is the number of samples predicted to be negative and actually negative, FP is the number of samples predicted to be positive but actually negative, and FN is the number of samples predicted to be negative but actually positive; stability verification unit, used to evaluate the stability of the new model under different environments; Latency testing unit, used to test the inference latency of new models.

[0015] Preferably, the intelligent incremental synchronization push module includes: Incremental update detection unit, used to compare the difference between the new and old model parameters; The synchronous push unit pushes differential parameter data to the device end to reduce the amount of data transmission.

[0016] Preferably, the differential parameter pushing algorithm is based on the following formula: Δθ=θ new -θ old ; Among them, θ new represents the model parameters of the new version, θ old represents the model parameters of the old version, Δθ represents the parameter change, and represents the difference between the new and old model parameters.

[0017] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This invention utilizes an intelligent incremental synchronization push module technology solution, effectively reducing bandwidth consumption between the cloud and the device through incremental updates and differential parameter transmission. Compared to traditional full updates, this invention only synchronizes the changed data, significantly improving synchronization efficiency from the cloud to the edge device. This enables AI quality inspection systems to maintain efficient data updates and resource conservation when deployed on a large scale.

[0018] 2. The present invention improves the real-time response capability of the AI ​​quality inspection system through the difference detection unit and incremental packaging technology. During the cloud upgrade process, only the changed model parameters will be pushed to the edge node, rather than full synchronization, ensuring that the AI ​​quality inspection system can complete the model update in the shortest possible time. Compared with the full update in the existing technology that requires a large amount of computing and storage resources, the present invention reduces the system burden and improves the real-time performance and accuracy of the AI ​​quality inspection system. In particular, in production line monitoring and real-time detection scenarios, it can effectively avoid system interruptions.

[0019] 3. This invention utilizes intelligent scheduling unit technology to optimize the push task scheduling of each node in the AI ​​quality inspection system based on network status and device load. Compared to traditional synchronization solutions, this invention dynamically selects the optimal node for data synchronization, reducing latency issues caused by network congestion or node overload, thereby improving the efficiency and reliability of the entire AI quality inspection system. This intelligent scheduling strategy is particularly suitable for distributed AI quality inspection systems and can effectively support large-scale, low-latency cloud-based upgrades and synchronization.

[0020] 4. This invention introduces a breakpoint-resume mechanism, making the cloud upgrade process of the AI ​​quality inspection system more stable and reliable. When a network interruption or device failure occurs during the synchronization process, the system can resume the synchronization update from the interruption point without having to re-push the entire data. Compared with existing solutions that cannot resume synchronization, this invention ensures the stability and high availability of the AI ​​quality inspection system in complex environments, ensuring continuity and seamless system upgrades during large-scale cloud upgrades. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the system framework of the present invention; Figure 2 This is a schematic diagram of the edge node intelligent preprocessing module of the present invention; Figure 3 This is a schematic diagram of the cloud-based incremental learning and database self-iteration module of the present invention; Figure 4 This is a schematic diagram of the containerized deployment and cross-platform synchronization module of the present invention; Figure 5 Schematic diagram of the isolation test and adaptive verification module of the present invention; Figure 6 This is a schematic diagram of the intelligent incremental synchronization push module of the present invention. DETAILED DESCRIPTION

[0022] The following is combined with Figure 1 -Attached Figure 6 , the present invention is described in further detail.

[0023] The present invention provides an AI quality inspection system based on cloud upgrade, such as Figure 1 -Attached Figure 6 As shown, the cloud-based upgraded AI quality inspection system includes: Mobile robot data acquisition module, used to collect data on industrial production lines, including product image data, environmental parameters and robot motion state parameters; Specifically, the mobile robot data acquisition module in this embodiment includes a data acquisition unit, a sensor fusion processing unit and a data encapsulation and transmission unit. The data acquisition unit is connected to a multimodal sensor group, which includes a high-definition industrial camera, an environmental temperature and humidity sensor and a lidar unit.

[0024] The data acquisition unit is used to obtain product surface image data, ambient temperature and humidity parameters, light intensity information and robot position information on the industrial production line.

[0025] The high-definition industrial camera is installed at the front end or top of the mobile robot to capture high-resolution images of products on the production line. The image resolution is preferably set to 3840×2160 pixels or above to ensure image clarity for small defect detection.

[0026] The ambient temperature and humidity sensor is fixed to the robot body to monitor the temperature and humidity of the production environment in real time. The temperature measurement range is -40°C to 85°C, and the humidity measurement range is 0%-100% RH, ensuring the reliability of quality inspection data under different environmental conditions.

[0027] LiDAR units are installed around the robot to achieve positioning and obstacle avoidance through 360-degree scanning, and work with the Inertial Navigation Unit (IMU) to provide accurate position information. The position information is expressed in coordinate form as follows: P(t)=(x(t),y(t),z(t)); Where P(t) is the spatial position of the robot at time t, and x(t), y(t), and z(t) represent the position coordinates in three-dimensional space.

[0028] The sensor fusion processing unit is used to synchronize time and space data from multiple sources, ensuring consistency among image data, environmental data, and location data through a timestamp alignment mechanism. Specific methods include: ΔT=|T image -T env |<∈; Among them, ΔT is the timestamp difference, which means the absolute difference between the two data collection times, T image is the image acquisition timestamp, T env is the timestamp for collecting environmental parameters, and ∈ is the maximum allowable synchronization error, which is generally less than 50 milliseconds.

[0029] Spatial registration uses the fusion of lidar and inertial navigation information to form continuous trajectory data. The trajectory curve is expressed as a sequence of discrete points: P={P(t0),P(t1),…,P(t n )}; Among them, P represents a set of trajectory points, that is, the point set of the trajectory, t0, t1,…, t n Indicates the time points of continuous sampling, from t0 to t n The sampling time between, P(t): represents the trajectory point corresponding to the time t.

[0030] The data encapsulation and transmission unit is used to uniformly package the above-mentioned synchronized data and encapsulate them into data frames according to the set protocol standards.

[0031] The encapsulated data is sent to the edge node intelligent pre-processing module via 5G wireless network or Gigabit Ethernet connection.

[0032] The edge node intelligent pre-processing module is connected to the mobile robot data acquisition module to perform denoising, feature extraction and screening on the received data, and send the processed data to the cloud; Specifically, in this embodiment, the edge node intelligent preprocessing module directly connects to the mobile robot's data acquisition module, receiving synchronized, encapsulated data frames as a data stream. Subsequently, through an orderly and efficient data processing process, the training data subsequently received by the cloud module is ensured to be highly effective and accurate.

[0033] The edge node intelligent preprocessing module includes a data denoising unit, a convolutional neural network feature extraction unit, and a feature screening unit. These units are logically connected in sequence and physically interconnected via a high-speed data bus to ensure the timeliness and consistency of data processing.

[0034] In some embodiments, the data denoising unit is used to eliminate noise components in the collected data and smooth the image data and environmental parameter data. Specifically, for the image data, a Gaussian filtering method is preferably applied, and the filtering process is defined as follows: Among them, I ′ (x,y) represents the pixel value of the filtered image at the coordinate (x,y), that is, the pixel value of the output image, I(x+i,y+j) represents the pixel value of the original image at the coordinate (x,y), that is, the pixel value of the input image, G(i,j) is the weight value of the Gaussian kernel at the position (i,j), also known as the Gaussian kernel function value, usually generated by a Gaussian function, (x,y) represents the position coordinates of a pixel point in the image, (i,j) represents the local neighborhood offset relative to the current point (x,y), i and j are the horizontal and vertical offsets respectively, k represents the Gaussian kernel radius, which defines half the size of the filter window (the total kernel size is (2k+1)×(2k+1)).

[0035] In this embodiment, the convolutional neural network feature extraction unit receives the denoised image data and uses deep convolution operations to extract multi-scale texture and structural features. The basic calculation definition of the convolution operation is as follows: Among them, F(i,j) represents the pixel value of the convolution output image at position (i,j), I ′ (i+m,j+n) represents the pixel value of the input image after the offset (m,n), K(m,n) represents the weight of the convolution kernel at the position (m,n), m,n represent the offset within the convolution window, and k represents the radius of the convolution kernel (the kernel size is 2k+1).

[0036] Alternatively, a convolutional neural network may include multiple convolutional layers and pooling layers stacked alternately to extract image feature information at different scales and semantic levels.

[0037] The feature screening unit is used to effectively screen the high-dimensional features output by the convolutional feature extraction unit to reduce redundant information and improve the efficiency of subsequent model learning. Specifically, feature screening is implemented based on a sparse optimization algorithm by solving the following optimization problem: in, Indicates that when the model parameter is θ, in the incremental data set The loss function value calculated above is Represents the new data set introduced in the incremental learning stage, θ0 represents the parameters of the original model, θ represents the new model parameters updated after incremental learning, and β represents the regularization coefficient.

[0038] In some embodiments, the feature screening unit may further screen out features whose weight amplitude is lower than a certain value based on a set threshold, thereby reducing the data dimension and optimizing subsequent processing performance while maintaining accuracy.

[0039] In this embodiment, the edge node intelligent preprocessing module ultimately packages the filtered feature data into a standard format data frame and sends it to the cloud-based incremental learning and database self-iteration module. The data packaging format can use Protobuf, ROS messages, or a custom lightweight protocol to adapt to different network bandwidth and latency conditions.

[0040] The cloud-based incremental learning and database self-iteration module is connected to the edge node intelligent pre-processing module to perform incremental learning on the data, optimize the AI ​​quality inspection model, and self-update the database; Specifically, after data screening and feature extraction by the edge node's intelligent preprocessing module, valid feature data is sent to the cloud in a standard format. To ensure the model's continuous evolution and adaptive adaptation to product and environmental changes, the system incorporates a cloud-based incremental learning and database self-iteration module. Typically, this module receives data streams from multiple edge nodes, performs dynamic incremental learning, and simultaneously updates the storage database, continuously optimizing the overall quality inspection model and improving detection accuracy and robustness.

[0041] In this embodiment, the cloud-based incremental learning and database self-iteration module includes a data receiving unit, an incremental learning unit, a data labeling and correction unit, and a database self-iteration unit. The units are connected in a logical order of the process and work together.

[0042] In some embodiments, the data receiving unit is configured to receive feature data frames uploaded from multiple edge nodes. Specifically, the received data includes image feature vectors, environmental parameter features, and robot motion state features, each of which is packaged in a unified format for subsequent processing.

[0043] Generally, data needs to be cached after being received. The cache can be implemented using high-concurrency middleware such as Redis or Kafka to buffer the peak data volume uploaded by edge nodes.

[0044] In this embodiment, the incremental learning unit updates the model based on the received data. Specifically, incremental learning adopts a small batch gradient update strategy, and the optimization objective function is defined as follows: Among them, θ represents the model parameter set, N represents the number of samples in the current mini-batch, and x i Represents the input feature vector of the i-th sample, y i represents the true label corresponding to the i-th sample, f(x i ;θ) represents the model with parameter θ on input x i The predicted output on Represents the loss function, which measures the difference between the predicted value and the true value, λ represents the regularization coefficient, and R(θ) represents the regularization term.

[0045] As an option, the loss function It can be a cross entropy loss function, a mean square error loss function, or can be set according to different task objectives.

[0046] In one possible implementation, the incremental learning unit uses a ReplayBuffer strategy to retain a portion of historical samples to mitigate catastrophic forgetting. Historical samples and new samples are mixed in a certain ratio for training, ensuring the model's ability to recognize old categories.

[0047] In this embodiment, the data labeling and correction unit is used to correct or supplement the label information in the data uploaded by the edge node. Specifically, a semi-supervised learning method based on confidence screening can be adopted. When the model's prediction confidence for an unlabeled sample exceeds a set threshold, the prediction result is used as a pseudo-label for incremental training.

[0048] The pseudo-label screening criteria are defined as follows; Among them, p(y|x) represents the category probability distribution predicted by the model under the input sample x, that is, the conditional probability of each category y, max(p(y|x)) represents the maximum value of all category prediction probabilities, that is, the probability of the most confident category, τ represents the confidence threshold, and the prediction result is accepted only when the maximum prediction probability is greater than τ. Indicates the pseudo-label category finally selected, that is, the category with the highest probability, argmax y p(y|x) represents the category y that maximizes p(y|x), then the logical connective.

[0049] Generally speaking, to ensure the reliability of label correction, a manual review mechanism can be introduced to conduct random inspections and verifications on pseudo-label results, and dynamically adjust the confidence threshold based on feedback.

[0050] In this embodiment, the database self-iteration unit is responsible for dynamically updating the feature database based on the model training progress and data changes. Specifically, after each incremental learning cycle, the system writes the newly added feature data and label data to the database and removes redundant, invalid, or erroneous data samples.

[0051] The containerized deployment and cross-platform synchronization module is connected to the cloud-based incremental learning and database self-iteration module to encapsulate the updated optimized AI quality inspection model into a container image and deploy it across platforms. Specifically, in this embodiment, after completing the optimization and update of the AI ​​quality inspection model by the cloud-based incremental learning and database self-iteration modules, in order to achieve rapid deployment and efficient synchronization of the model on multiple hardware platforms, the present invention further includes a containerized deployment and cross-platform synchronization module. Generally, this module is used to encapsulate the optimized AI quality inspection model and its operating environment into a standardized container image, and distribute it between the cloud and each edge node through an image synchronization mechanism, ensuring consistency and high availability during the system upgrade process.

[0052] In this embodiment, the containerized deployment and cross-platform synchronization module includes a containerized packaging unit and an image synchronization unit, which work together to complete the entire process from model packaging to image synchronization.

[0053] In this embodiment, the containerized packaging unit is used to package the updated AI quality inspection model and its required operating environment together to generate a container image. Specifically, the containerized packaging process includes the steps of base image selection, dependent environment configuration, model file packaging, and image construction.

[0054] Generally, the base image can choose a lightweight operating system image, such as Alpine or Ubuntu, to reduce the image size and increase the deployment speed.

[0055] As an option, the container image building process can be formally defined as follows: Among them, Image represents the final generated container image, Base is the selected basic image, Dep i is the i-th dependent component, n is the total number of dependencies, and Model is the optimized AI quality inspection model file.

[0056] In one possible implementation, the containerized packaging unit can also introduce multi-stage build technology (Multi-stageBuild) during the build process to further eliminate intermediate redundant files and optimize the image structure.

[0057] In this embodiment, the image synchronization unit is used to synchronize the packaged container image to the cloud warehouse or the local warehouse of the edge node, and select the adapted image version for deployment according to the hardware architecture differences of the device end.

[0058] Specifically, the image synchronization unit supports a cross-platform synchronization mechanism and can pull and deploy images according to the CPU architecture of the target device (such as x86, ARM).

[0059] Generally speaking, the image synchronization process involves three stages: image version detection, difference comparison, and synchronization push. The image difference detection standard can be defined as follows: Among them, Δ I mage represents the hash difference between the local image and the cloud image, H(·) is the hash calculation function, Represents XOR operation, Image local Represents a local area in the image, Image cloud Represents the cloud-related portion of an image or data originating from the cloud.

[0060] As an option, the hash calculation function can use the SHA-256 algorithm to ensure the reliability and security of image consistency verification.

[0061] In a possible implementation, to reduce network transmission load, the mirror synchronization unit may support layer-based synchronization, pushing only the changed layer, thereby effectively improving synchronization efficiency.

[0062] In this embodiment, the image synchronization unit uses a secure transmission protocol (such as HTTPS or TLS channel) when pushing the image to the edge node to prevent data leakage or tampering during the synchronization process.

[0063] Generally speaking, containerized deployment and cross-platform synchronization modules can ensure that the optimized AI quality inspection model can be quickly started and efficiently run on different edge nodes, and support smooth transition and zero-downtime updates when the system undergoes version iterations.

[0064] The isolation test and adaptive verification module is connected to the cloud-based incremental learning and database self-iteration module to verify the updated optimized AI quality inspection model and ensure its performance meets the requirements; Specifically, this example verifies the updated, optimized AI quality inspection model to ensure its accuracy, stability, and real-time reasoning capabilities in practical applications. Close integration with cloud-based incremental learning and database self-iteration modules ensures the model maintains efficient operation during automatic updates and optimizations, and consistently performs as expected across various production environments.

[0065] In this embodiment, the isolation test and adaptive verification module mainly includes an accuracy verification unit, a stability verification unit, and a delay test unit. These units work together to comprehensively test and verify the new model, ensuring its usability and reliability on industrial production lines.

[0066] The accuracy verification unit is used to evaluate the prediction accuracy of the optimized model. Specifically, the verification unit determines the effectiveness of the model by comparing the prediction results of the model on different datasets and calculating key performance indicators (such as classification accuracy and recall). Generally, the accuracy verification process includes the following steps: Collect and prepare a test dataset containing samples that the new model has not seen.

[0067] Input the test data into the updated model to obtain the model's prediction results.

[0068] Calculate performance indicators such as prediction accuracy and recall.

[0069] In one possible implementation, the calculation formula for accuracy verification can be defined as follows: Among them, TP is the number of samples predicted to be positive and actually are positive, TN is the number of samples predicted to be negative and actually are negative, FP is the number of samples predicted to be positive but actually are negative, and FN is the number of samples predicted to be negative but actually are positive. In general, the accuracy and recall rate indicators calculated by the above formula can fully reflect the predictive ability of the model in practical applications.

[0070] The stability verification unit is mainly used to evaluate the stability of the new optimization model in different environments to ensure that it can maintain consistent performance under different operating conditions. Stability verification includes the following tests: The model is tested under different environmental conditions (such as temperature, humidity, etc.).

[0071] Compare the fluctuations in the model's output during the test and evaluate its stability.

[0072] Specifically, the stability verification process can use the output fluctuation of the model as an evaluation indicator. For example, the model stability can be calculated in the following way: Among them, Std(y new ) represents the new model output y new The standard deviation, Std(y original ) represents the original model output y originalStabilityIndex represents the stability index, which measures the difference in stability between the output of the new model and the original model. The smaller the stability index, the more stable the model As an option, stability verification can further enhance the reliability of the model through continuous testing under different equipment or production environment conditions.

[0073] The latency testing unit is used to evaluate the inference latency of new models. This is particularly important when AI quality inspection models are applied to real-time industrial inspections, as inference latency directly impacts production efficiency. This unit evaluates the real-time responsiveness of new models by measuring the time from data input to model output.

[0074] The delay test standard can be calculated by the following formula: Latency = T output -T input Among them, Latency means delay, which refers to the time difference between data input and output, T output Indicates the timestamp (output time) when the model generates the output result, T input Indicates the timestamp when the input data is fed into the model (input time).

[0075] In some embodiments, to reduce latency, optimizing the inference engine or using hardware acceleration (such as GPU, TPU, etc.) may be an effective method for latency optimization.

[0076] In this embodiment, the functionality and structure of the isolated testing and adaptive verification module ensure that the accuracy, stability, and latency performance of each AI quality inspection model update remain within acceptable ranges and remain highly consistent throughout the model inference process. By comprehensively verifying the new model, this module can effectively reduce quality issues that may arise from model updates, thereby ensuring the normal operation of industrial production lines.

[0077] The intelligent incremental synchronization push module is connected to the cloud-based incremental learning and database self-iteration module to push the incrementally updated optimized AI quality inspection model to the device side, transmitting only the differences; Specifically, this embodiment's intelligent incremental synchronization push module directly connects to the cloud-based incremental learning and database self-iteration module, integrating model optimization results processed by the containerized deployment and cross-platform synchronization modules. Through incremental update detection and differential synchronization push mechanisms, this effectively reduces data transmission, improves edge node model update efficiency, and ensures rapid consistency across different nodes.

[0078] In this embodiment, the intelligent incremental synchronous push module includes an incremental update detection unit and a synchronous push unit, which are respectively used for difference detection between new and old model parameters and synchronous push of differential data.

[0079] In this embodiment, the incremental update detection unit is used to compare the differences between the optimized new version model parameters and the existing old version model parameters on the device. Specifically, this unit compares the new and old model parameters element by element, extracts the changed parameters, and generates a differential data set for subsequent push.

[0080] In general, the difference calculation of the incremental update detection unit follows the following formula: Δθ=θ new -θ old ; Among them, θ new Represents the parameter set of the new version model, θ old represents the parameter set of the old version of the model, and Δθ represents the parameter change, that is, the element-by-element difference between the parameters of the new and old models.

[0081] As an option, to improve detection accuracy, the incremental update detection unit can adopt a floating-point error tolerance mechanism and set a small change threshold. When the parameter change is lower than the threshold, it is considered as no significant change, avoiding invalid synchronization.

[0082] In some embodiments, to further reduce the computational load during the difference detection process, the incremental update detection unit may perform comparison and extraction only within a specific weight block or an important parameter block based on a block-level difference method.

[0083] In this embodiment, the synchronization push unit is used to push the differential parameter data generated by the incremental update detection unit to the edge device. Specifically, the unit transmits the differential parameter data in a compressed format through the network channel based on the content of the differential data, thereby reducing bandwidth consumption during the synchronization process.

[0084] In a possible implementation, the synchronous push unit compresses the differential parameters before pushing them, using a lossless compression algorithm such as Zstandard or LZ4 to further reduce the size of the data packet.

[0085] Generally, during the synchronous push process, data integrity verification information of the differential parameters will be attached, such as the use of SHA-256 verification code, to ensure that the pushed data has not been tampered with or lost during the transmission process.

[0086] As an option, after the synchronous push unit completes data transmission at the receiving end, it applies the received differential component Δθ\Delta\thetaΔθ to the existing model through the parameter reorganization mechanism, updating the parameters of the device-side AI quality inspection model to synchronize it with the latest optimized version. The reorganization update formula is as follows: θ updated =θ old +Δθ; Among them, θ updated represents the updated model parameter set, θ old represents the model parameter set before the update (old version), and Δθ represents the parameter update amount, that is, the change from the old version to the new version.

[0087] In some embodiments, in order to improve the update flexibility of the system, the synchronization push unit may first send a pre-synchronization instruction to the device end before pushing synchronization. The device end decides whether to apply the differential update immediately based on the current task status to avoid interfering with the real-time detection task.

[0088] In one possible implementation, in order to improve the system's exception handling capabilities, the synchronization push unit can support breakpoint resumption. When the transmission is interrupted, the differential data can be synchronized from the interruption point to ensure the integrity of the update process.

[0089] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Cloud-based AI quality inspection system, characterized by: include; Mobile robot data acquisition module, used to collect data on industrial production lines, including product image data, environmental parameters and robot motion state parameters; The edge node intelligent pre-processing module is connected to the mobile robot data acquisition module to perform denoising, feature extraction and screening on the received data, and send the processed data to the cloud; The cloud-based incremental learning and database self-iteration module is connected to the edge node intelligent pre-processing module to perform incremental learning on the data, optimize the AI ​​quality inspection model, and self-update the database; The containerized deployment and cross-platform synchronization module is connected to the cloud-based incremental learning and database self-iteration module to encapsulate the updated optimized AI quality inspection model into a container image and deploy it across platforms. The isolation test and adaptive verification module is connected to the cloud-based incremental learning and database self-iteration module to verify the updated optimized AI quality inspection model and ensure its performance meets the requirements; The intelligent incremental synchronization push module is connected to the cloud-based incremental learning and database self-iteration module to push the incrementally updated optimized AI quality inspection model to the device side, transmitting only the differences.

2. The cloud-based AI quality inspection system according to claim 1, characterized in that: The data acquisition module collects data through a multimodal sensor, which includes a camera, a temperature and humidity sensor, and a lidar. The collected data of the sensor includes product surface image, light intensity, ambient temperature, humidity, and robot position.

3. The cloud-based upgrade AI quality inspection system according to claim 1 is characterized in that: The edge node intelligent preprocessing module includes the following units: Data denoising unit, used to eliminate noise in collected data and smooth the data; Convolutional neural network feature extraction unit, used to extract important features from image data; The feature screening unit is used to screen out valid features according to the set feature selection criteria.

4. The cloud-based upgrade AI quality inspection system according to claim 3 is characterized in that: The feature screening unit screens the extracted features based on a sparse optimization algorithm, specifically by solving the following optimization problem: Among them, w represents the feature weight vector, F represents the feature matrix, y represents the label or target value vector of the sample, and λ represents the regularization coefficient. Indicates the optimization of variable w.

5. The cloud-upgraded AI quality inspection system according to claim 1, characterized in that: The cloud-based incremental learning and database self-iteration module includes: Incremental learning unit, used to fine-tune and optimize AI models based on newly added data sets; The database self-iteration unit is used to update the model version and quality inspection results in the database, and to self-optimize the system.

6. The cloud-based upgrade AI quality inspection system according to claim 5, characterized in that: The incremental learning adopts a small batch gradient update strategy, and the optimization objective function is defined as follows: Among them, θ represents the model parameter set, N represents the number of samples in the current mini-batch, and x i Represents the input feature vector of the i-th sample, y i represents the true label corresponding to the i-th sample, f(x i ;θ) represents the model with parameter θ on input x i The predicted output on , L(·) represents the loss function, which measures the difference between the predicted value and the true value, λ represents the regularization coefficient, and R(θ) represents the regularization term.

7. The cloud-based upgrade AI quality inspection system according to claim 1, characterized in that: The containerized deployment and cross-platform synchronization module includes: Containerized packaging unit, used to package the updated AI quality inspection model and its dependent environment into a container image; The image synchronization unit is used to synchronize container images to the cloud or edge nodes and select images for deployment based on device hardware.

8. The cloud-upgraded AI quality inspection system according to claim 1, characterized in that: The isolation test and adaptive verification module includes: The accuracy verification unit is used to evaluate the prediction accuracy of the new model by calculating the classification accuracy and recall indicators; Among them, TP is the number of samples predicted to be positive and actually positive, TN is the number of samples predicted to be negative and actually negative, FP is the number of samples predicted to be positive but actually negative, and FN is the number of samples predicted to be negative but actually positive; stability verification unit, used to evaluate the stability of the new model under different environments; Latency testing unit, used to test the inference latency of new models.

9. The cloud-upgraded AI quality inspection system according to claim 1, characterized in that: The intelligent incremental synchronization push module includes: Incremental update detection unit, used to compare the difference between the new and old model parameters; The synchronous push unit pushes differential parameter data to the device end to reduce the amount of data transmission.

10. The cloud-upgraded AI quality inspection system according to claim 9, characterized in that: The differential parameter push algorithm is based on the following formula: Δθ=θ new -θ old ; Among them, θ new represents the model parameters of the new version, θ old represents the model parameters of the old version, Δθ represents the parameter change, and represents the difference between the new and old model parameters.

Images