Metering detection intelligent glasses system integrating artificial intelligence visual identification and cloud platform

Abstract

The invention discloses a metering detection intelligent glasses system integrating artificial intelligence visual identification and a cloud platform, and relates to the technical field of computer software, the metering detection intelligent glasses system comprises an intelligent glasses driving module and adaptive hardware, and has the functions of dynamic scheduling, multi-thread processing, communication self-adaption and non-lagging AR rendering; the AI visual identification module supports batch identification, fine defect detection, off-line operation and incremental learning by using a lightweight model; the cloud platform software framework module adopts a distributed micro-service framework to realize data partition storage, real-time parallel analysis and the like and elastic capacity expansion and contraction; the interaction control module develops a multi-mode program and supports batch instructions and the like; and the system collaboration module realizes multi-module scheduling based on event driving, and ensures efficient collaboration of the whole detection process. According to the method, the precision, efficiency and environmental adaptability of measurement detection are remarkably improved, the batch processing capacity and the network fluctuation coping capacity are greatly enhanced, and efficient and reliable software support is provided for measurement detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer software, and particularly relates to a metrological detection intelligent glasses system integrating artificial intelligence visual recognition and a cloud platform. BACKGROUND

[0002] In the field of metrological detection, the adaptability and synergy of the software system of the intelligent glasses as a portable detection device directly affect the detection efficiency and accuracy. The software driver of the existing intelligent glasses is designed with fixed parameters, which cannot dynamically adjust the allocation of hardware resources according to the task load, resulting in high CPU occupation and causing lag during high-resolution image acquisition, and resource waste during low-complexity recognition tasks. At the same time, the single-thread mechanism is adopted for sensor data processing, and the multi-dimensional data acquisition such as ambient light intensity and temperature is not synchronized, which easily causes distortion of the input data of the subsequent recognition model. The wireless communication driver lacks intelligent switching and bandwidth optimization capabilities, and the signal is unstable in complex environments, resulting in frequent problems of data transmission interruption or high delay. Due to the defects of the rendering mechanism, the recognition result and the real-time picture are often overlaid with delay and lag, which affects the operation and judgment of the detection personnel.

[0003] The limitations of artificial intelligence visual recognition software are also prominent. The existing recognition model mostly adopts a single network architecture and is not optimized for the multi-size and multi-angle characteristics of measuring instruments, resulting in a significant decrease in recognition accuracy for small-size instruments or complex postures. The model training relies on a large amount of labeled data, and the entire model needs to be retrained during incremental updates, which is time-consuming and labor-intensive and cannot adapt to new types of instruments in field detection. The appearance defect detection is mostly based on traditional image processing algorithms, which can only identify obvious defects and cannot accurately locate subtle scratches and slight deformations. Moreover, it lacks defect semantic segmentation capability and cannot provide detailed defect information for subsequent analysis. In addition, the model relies on cloud computing, and the recognition capability is limited in offline state, making it difficult to meet the detection needs in outdoor or network-free environments.

[0004] The cloud platform software architecture and system synergy have obvious shortcomings. The existing platform mostly adopts a monolithic architecture and cannot dynamically scale according to the concurrency, which easily causes response delay when facing large-scale detection data. The data storage does not adopt a targeted time sequence management strategy, and the historical data archiving and retrieval efficiency is low. The analysis function is limited to simple statistics, and lacks precise prediction of device performance trends. The multi-module synergy relies on synchronous calling mechanism, and the failure of a module easily leads to interruption of the entire process. Moreover, it lacks distributed transaction coordination, and the data consistency is difficult to guarantee. At the same time, the multi-modal interaction has instruction conflict problems, and the response logic is not optimized according to user usage habits, resulting in a high rate of misoperation. The data security only relies on a single encryption method, and the detection data is easy to be tampered with or leaked. The low-power management lacks task priority differentiation, and the core detection task and non-core task compete for resources, resulting in insufficient endurance and difficulty in meeting the long-time outdoor detection needs. SUMMARY

[0005] The application provides a metering detection intelligent glasses system integrating artificial intelligence visual recognition and a cloud platform to solve the problems in the prior art.

[0006] To achieve the above-mentioned purpose, the application adopts the following technical scheme: a metering detection intelligent glasses system integrating artificial intelligence visual recognition and a cloud platform, comprising the following modules: An intelligent glasses software driving module develops a driving program suitable for a camera, a sensor array, wireless communication and an AR display unit; a dynamic resource scheduling sub-module is included, which can adjust the occupancy rate and memory allocation; the camera supports dynamic switching; the sensor processes light intensity, temperature and three-dimensional coordinate data in multiple threads; the wireless communication can monitor network quality, switch to Wi-Fi when the 5G signal is weak and enable bandwidth aggregation; the AR display adopts double-buffer rendering; An artificial intelligence visual recognition module develops and improves sample recognition and defect detection programs; sample recognition is performed through a model distillation compression network, a built-in metrological instrument feature library and incremental learning to update weights; defect detection can locate scratches, deformation and missing nameplates; A cloud platform software architecture deploys distributed microservices on a metrology institute private cloud, including data storage, analysis, API gateway and container orchestration services; the data storage uses a time series and object database to support life cycle management; the analysis service integrates a real-time computing engine to count data and generate a health score; the API gateway has a flow limiting mechanism and provides a dual interface; An interactive control module develops multi-modal interaction programs such as voice, gesture and touch control; the voice supports offline recognition of instructions and noise adaptation; the touch control supports AR virtual control operation; the priority of interaction is voice > gesture > touch; A system coordination module develops an event-driven scheduling program, defines a six-stage process and decouples each stage through a message queue; a distributed transaction coordination sub-module is included.

[0007] Further, an AI model optimization software module is further included, which develops an online reinforcement learning optimization program, and the optimization formula is wherein is a model parameter matrix, is a learning rate, is an optimization period, is a recognition result reward value at the moment, is a cross-entropy loss function at the moment, and the program adds regularization pruning when incrementally training the model.

[0008] Further, a data security software module is further included, which develops an encryption program based on the national encryption SM4 algorithm and a consortium chain storage program; the encryption program realizes the formula software execution, wherein is the encrypted data block, is a 256-bit dynamic key, is the original detection data, is a 32-round nonlinear transformation function based on the SM4 algorithm, and the program performs double encryption on the transmission data and the locally stored data.

[0009] Further, a scene-adaptive image anti-shake program is developed in the intelligent glasses software driving module. The program collects motion data through an acceleration sensor and a gyroscope, generates a motion state matrix by using a Kalman filtering algorithm, and the algorithm formula is wherein is the anti-shake image, is the original image, is a 3*3 motion transformation matrix, and the program has a built-in scene recognition sub-module. The scene recognition sub-module distinguishes between static scenes and dynamic scenes by using the image feature point change rate. The strong anti-shake mode is used for static scenes, and the weak anti-shake mode is used for dynamic scenes.

[0010] Further, an attention-enhanced multi-scale feature fusion program is developed in the artificial intelligence vision recognition software module. The program realizes the software processing of the formula wherein is a 1 / 8 resolution feature map, is a 1 / 4 resolution feature map, is a 1 / 2 resolution feature map, is a channel dimension splicing operation, is a feature fusion operation based on attention weights, and the program has a feature enhancement sub-module developed for small-size measuring instruments.

[0011] Further, a working condition perception time series prediction program is developed in the cloud platform software architecture module. The program completes the algorithm execution of the formula wherein is a device error prediction value at time t+1, is detection data at the last time points, is a time decay weight, and b is a bias term.

[0012] Further, a user-adaptive multi-modal attention decision program is developed in the interactive control software module. The program completes the software operation of the formula wherein is an interactive modal attention score, is a weight of the th interactive modal, is a feature function of the th interactive modal, =3 is the total number of interactive modes, and the program is embedded with a user habit learning sub-module.

[0013] Further, the system cooperates with the software module to develop an intelligent preloading edge and a cloud computing switching program to complete the software calculation of formula , wherein is a switching decision integral value, is a network delay at the moment, is a delay threshold value, is a decision period.

[0014] Further, the intelligent glasses software driving module develops a task priority low-power consumption management program, adopts a power consumption prediction algorithm , wherein is a total predicted power consumption, is real-time power consumption of the nth hardware module, is a module predicted working time, and the program divides tasks into three levels.

[0015] Further, the intelligent warehouse alignment software module develops a dynamic alignment control program cooperated with AI vision and laser ranging to complete the software operation of formula , wherein , are displacement amounts of samples in a warehouse coordinate system, , are real-time moving speeds in the x and y directions, , are initial position coordinates, and the program integrates an obstacle detection sub-module to identify obstacles between shelves by using a depth camera. Compared with the prior art, the present application has the following advantages: Through the all-round software innovation design, the present application effectively solves the core pain points of the existing measurement and detection intelligent glasses system, and the overall performance and practicability of the software are significantly improved. The dynamic resource scheduling mechanism of the software driving module realizes the on-demand allocation of hardware resources, avoids high-load lag and low-load resource waste, the intelligent switching and bandwidth aggregation technology of wireless communication guarantees stable connection in a complex environment, the AR double-buffer rendering mechanism realizes the lag-free superposition of recognition results and real-time pictures, and the smoothness of detection operation is greatly improved. The multi-threaded synchronous processing of sensors ensures the accuracy and integrity of data acquisition, and provides reliable input for subsequent recognition analysis.

[0016]

[0017] ​​​​The innovative design of the AI-powered visual recognition software significantly enhances recognition capabilities and adaptability. The improved network architecture, combined with model distillation technology, reduces model size while maintaining recognition accuracy. Multi-scale feature fusion and attention enhancement mechanisms effectively solve the recognition challenges of measuring instruments of different sizes and orientations, particularly improving the recognition of small instruments and those with subtle defects. Incremental learning and offline encrypted storage capabilities enable the model to quickly adapt to new types of instruments and support independent detection in offline environments, broadening the system's application scenarios. Reinforcement learning optimization continuously improves the model's robustness under complex lighting and angular conditions, ensuring the stability and reliability of detection results.

[0018] The cloud platform's distributed microservice architecture boasts flexible scaling capabilities, easily handling varying levels of concurrency in testing. The combination of time-series databases and object storage optimizes data storage and retrieval efficiency. Real-time computing engines and trend prediction programs not only rapidly process large-scale testing data but also provide early warnings of equipment performance degradation, offering scientific decision support for metrology instrument management. The system's collaborative event-driven architecture and distributed transaction coordination mechanism achieve efficient decoupling of multiple modules and ensure data consistency. Breakpoint resume technology guarantees the reliability of large-scale data transmission.

[0019] The multimodal interaction program significantly reduces command conflicts and error rates through attention-based decision-making and user habit learning, improving the convenience of testing operations. The data security module's dual encryption and consortium blockchain evidence storage mechanism construct a comprehensive security protection system, ensuring the immutability and traceability of testing data. The low-power management program significantly optimizes battery life through task priority differentiation and dynamic voltage adjustment, meeting the needs of long-term outdoor testing. The intelligent warehousing alignment software improves the automation level and accuracy of sample warehousing, further enhancing the software support capabilities of the entire metrology and testing process. Overall, the software system of this invention achieves qualitative improvements in adaptability, recognition accuracy, collaborative efficiency, security, and ease of use, providing strong software support for metrology and testing work. Attached Figure Description

[0020] Figure 1 This is a schematic block diagram of the intelligent glasses system for metrology and testing that integrates artificial intelligence visual recognition and cloud platform proposed in this invention. Figure 2 Line graph comparing sample recognition accuracy under different lighting conditions; Figure 3 A bar chart comparing the detection time of a single group under different batch sizes; Figure 4 A pie chart showing the system power consumption distribution under different task priorities. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0024] Reference Figures 1 to 4 A smart glasses system for metrology and testing integrating artificial intelligence visual recognition and a cloud platform, comprising the following modules: The smart glasses software driver module develops drivers adapted to high-resolution cameras, miniature sensor arrays, wireless communication modules, and AR display units. This module has a built-in dynamic resource scheduling submodule, which can automatically adjust the CPU core utilization (range 10%-80%) and memory allocation strategy according to the real-time task load (such as image acquisition resolution and recognition complexity). The camera driver supports dynamic switching between 4K / 2K / 1080P resolution (with adaptive frame rate of 30-60fps). The sensor driver uses a multi-threaded synchronization mechanism to process ambient light intensity (sampling rate 10Hz), sample surface temperature (error ±0.5℃), and three-dimensional spatial coordinate data. The wireless communication driver integrates a network quality monitoring submodule, which automatically switches to Wi-Fi and enables bandwidth aggregation technology when the 5G signal strength is <-100dBm. The AR display driver develops a double-buffered rendering mechanism to achieve smooth overlay of recognition results and real-time images (latency <20ms). The AI ​​visual recognition software module develops a sample recognition program based on an improved YOLOv8 architecture and an appearance defect detection program based on the U-Net++ architecture. The sample recognition program uses model distillation technology to compress a 128-layer deep network to 32 layers (reducing volume by 60%). It has a built-in multi-view feature library of 500 measuring instruments (weights, scales, gauges, etc.) and updates feature weights through a sliding window incremental learning algorithm (triggers model fine-tuning every 100 new data points). The recognition accuracy is ≥99.5% and the single-frame processing time is <50ms. The appearance defect detection program develops a defect semantic segmentation submodule, which can perform pixel-level localization of scratches (minimum recognition length 0.5mm), deformation (accuracy ±0.01mm), and nameplate missingness. The defect recognition accuracy is ≥98%. The program supports offline model encrypted storage (AES-256 encryption) and online differential updates (only transmitting the incremental part of the model). The cloud platform software architecture module is developed and deployed on the Institute of Metrology's private cloud using a distributed microservice architecture. It includes data storage microservices, analytics microservices, API gateway microservices, and container orchestration services. The data storage microservice uses a time-series database (InfluxDB) to store metrology and testing data (each record contains 28 fields including timestamp, device ID, and test value), combined with object storage (MinIO) to save raw images (compression ratio 1:5), and supports data lifecycle management (automatically archiving over 3 years of historical data). The analytics microservice integrates the Spark Streaming real-time computing engine, capable of performing statistical analysis on 1000+ test data points per second (such as pass rate and error distribution), and generating device health scores (0-100 points). The API gateway microservice uses a token bucket rate limiting algorithm (processing 1000 requests per second), provides both RESTful and gRPC interfaces, and has a data upload response time ≤100ms. The container orchestration service (Kubernetes) supports automatic scaling based on concurrency (dynamically adjusting from 5 to 50 container instances). The interactive control software module develops a multimodal interactive program integrating voice, gesture, and touch. The voice recognition program uses Mel-frequency feature extraction and a bidirectional LSTM network to support offline recognition of 200+ metrology commands (such as "calibrate range" and "record error value") with an accuracy of ≥95%. It also develops a noise adaptive submodule (adjusting the recognition threshold according to the ambient noise intensity from 0-80dB). The gesture recognition program collects key hand points (21 feature points) through a camera and uses a dynamic time warping algorithm to recognize 6 operation gestures (such as fist confirmation and swipe page turning) with a response time of <300ms. The touch interaction program develops an AR virtual button control library that supports single-finger click, two-finger zoom (zoom range 1-5x), and drag annotation. The three interaction methods are coordinated through a priority arbitration submodule (voice > gesture > touch) to avoid command conflicts. The system's collaborative software module develops a collaborative scheduler based on an event-driven architecture, defining a six-stage execution flow of "acquisition-preprocessing-recognition-upload-analysis-feedback". Each stage is decoupled through a message queue (RabbitMQ). When the smart glasses acquire sample images, a preprocessing event (image denoising and distortion correction) is triggered. After processing, the image is pushed to the AI ​​recognition module. After the recognition result is generated, an upload event is triggered (using a breakpoint resume mechanism, supporting data transmission of more than 2GB). After the cloud platform completes analysis, a feedback event is pushed to the AR display module via WebSocket. At the same time, a distributed transaction coordination submodule is developed, using the TCC (Try-Confirm-Cancel) mode to ensure data consistency across multiple modules (transaction success rate ≥99.9%), and recording a full-process audit log (including module name, timestamp, and processing result) for fault tracing.

[0025] This invention also includes an AI model optimization software module, which develops an online reinforcement learning optimization program, and the optimization formula is as follows: ,in For the model parameter matrix, The learning rate (adaptively adjusted from 0.001 to 0.01) To optimize the cycle (default 1 hour). for The reward value for the recognition result at each moment (1 for correct recognition, -0.5 for incorrect recognition). for The cross-entropy loss function at time step [time] is added to the program during incremental model training. Regularized pruning (removing neurons with absolute weights < 0.001) triggers global optimization every 5000 accumulated recognition data points, improving the model's recognition accuracy by ≥ 5% in complex environments such as backlighting (light intensity > 10000 lux) and sidelighting (angle 30°-60°). Simultaneously, model quantization (converting from 32-bit floating-point to 16-bit fixed-point) reduces computational resource consumption, ensuring real-time performance on embedded systems. This invention also includes a data security software module, which develops an encryption program and a consortium blockchain evidence storage program based on the national cryptographic algorithm SM4. The encryption program implementation formula is as follows: Software-based execution, in which For the encrypted data block, It is a 256-bit dynamic key (updated every 24 hours via a key exchange protocol). This is the original test data. The program employs a 32-round nonlinear transformation round function based on the SM4 algorithm. It performs double encryption on both the transmitted data (using the TLS 1.3 protocol) and the locally stored data (encrypted and stored in an embedded security chip). The consortium blockchain evidence storage program is developed based on Hyperledger Fabric. It packages the hash value of the detection data (SHA-256 calculation) and the device's unique identifier into a transaction. After consensus among more than 3 nodes, the transaction is written into a block with a block generation interval of ≤10 seconds, thus achieving the immutability and full traceability of the detection data (supporting reverse lookup from the detection result to the original image).

[0026] In this invention, a scene-adaptive image stabilization program is developed in the smart glasses software driver module. This program collects motion data through an accelerometer (sampling rate 100Hz) and a gyroscope (sampling rate 200Hz), and uses a Kalman filter algorithm to fuse and generate a motion state matrix. The algorithm formula is as follows: ,in For the image after image stabilization, For the original image, The program uses a 3×3 motion transformation matrix (including translation, rotation, and scaling parameters). It has a built-in scene recognition submodule that distinguishes between static scenes (such as laboratory scales) and dynamic scenes (such as assembly line instruments) by the rate of change of image feature points. Static scenes use a strong anti-shake mode (compensation range ±10 pixels), while dynamic scenes use a weak anti-shake mode (compensation range ±3 pixels) to avoid image distortion caused by excessive anti-shake and ensure the clarity of the acquired images in different scenes (blur score <0.2).

[0027] In this invention, an attention-enhanced multi-scale feature fusion program is developed within the artificial intelligence visual recognition software module. This program implements the formula... Software-based processing, among which A 1 / 8 resolution feature map (capturing global contours). This is a 1 / 4 resolution feature map (for extracting local details). It is a 1 / 2 resolution feature map (preserving texture information). This is a channel-level splicing operation. For attention-weighted feature fusion operations (weights are dynamically generated using the Sigmoid function), the program develops a feature enhancement submodule for small-sized measuring instruments (such as gauges with a diameter <5mm). By upsampling interpolation (magnification of 2-4 times) and edge sharpening algorithms, subtle features are highlighted, improving the recognition accuracy of small-sized samples by ≥8%. At the same time, the non-maximum suppression algorithm is optimized (IOU threshold adaptively set to 0.3-0.7) to solve the problem of misidentification of densely arranged samples.

[0028] In this invention, a working condition awareness time series prediction program is developed in the cloud platform software architecture module to implement the formula. The algorithm execution, its This represents the predicted equipment error value at time t+1. For history Detection data at each time point ( =7, taking the values ​​from the last 7 tests). Time decay weights (recent data has higher weights, such as...) =0.3, =0.25,…, =0.05), b is the bias term (based on the equipment model preset). This program integrates equipment usage data (such as cumulative usage time and ambient temperature fluctuations), optimizes the weight coefficients through the random forest algorithm, and has a prediction error of ≤3%. It can provide early warning of equipment performance degradation trends 1-3 months in advance (the warning is triggered when the prediction error exceeds 80% of the allowable threshold), and provides data support for the dynamic adjustment of the calibration cycle of measuring instruments (such as shortening it from 12 months to 9 months).

[0029] In this invention, a user-adaptive multimodal attention decision-making program is developed in the interactive control software module to implement the formula. The software calculation, among which For interaction modality attention score, For the first Weights of various interaction modalities ( =1 represents voice. =2 represents a gesture. =3 is for touch control). For the first The feature functions of the interaction modalities (speech is the recognition confidence, gesture is the feature point matching degree, and touch is the click pressure value). =3 represents the total number of interaction modalities. The program has a built-in user habit learning submodule that adjusts the weighting coefficients by analyzing the last 100 interaction records (e.g., users who frequently use gestures). Automatically increases by 0.2), when multimodal commands are triggered simultaneously (such as voice "confirm" and gesture "clench fist"), select... The instruction with the highest value is executed, reducing the instruction conflict rate to <1%. At the same time, a submodule for correcting misoperations was developed to avoid critical operation errors through contextual semantic analysis (such as the need for secondary confirmation for the "delete" instruction).

[0030] In this invention, an intelligent pre-loading edge-to-cloud computing switching program is developed in the system collaborative software module to implement the formula. The software calculation, in which To switch decision integral values, for Network latency at any given time (sampling interval 100ms). This is the delay threshold (default 200ms). For the decision cycle (5 seconds), when Edge computing mode is triggered when the response time is >500ms. Popular model pre-loading sub-modules are used to pre-cache 5-10 frequently used recognition models (such as electronic scales and calipers) based on user historical detection records (data from the last 30 days). In edge mode, the local recognition response time is <100ms, and the response time resumes after network recovery. Automatically switch to cloud computing mode within 100ms (seconds) and synchronize cached data at the edge (using incremental synchronization, only transmitting the difference), ensuring that the system can still work normally in weak network (latency > 500ms) or network outage environments (core function availability ≥ 90%).

[0031] In this invention, a low-power management program for task priority is developed in the smart glasses software driver module, employing a power prediction algorithm. ,in To predict total power consumption (in mAh). For the first Real-time power consumption of each hardware module (200mA for camera in 4K mode, 100mA for AR display, etc.). To estimate the module's working time, the program divides tasks into three levels (emergency detection > data upload > log synchronization). When the predicted remaining battery power is <20%, it automatically reduces the hardware performance of secondary tasks (e.g., the camera is reduced to 2K resolution and the sampling rate is reduced by 50%), while turning off unnecessary sensors (e.g., ambient light sensor). It further reduces power consumption through dynamic voltage regulation (CPU core voltage is reduced by 0.1V) to ensure that the battery life of the emergency detection task is ≥2 hours, and the overall battery life is improved by ≥30% (from 8 hours to 10.5 hours).

[0032] This invention also includes an intelligent warehouse alignment software module, which develops a dynamic alignment control program that combines AI vision and laser ranging to achieve the formula. The software calculation, among which , The displacement of the sample in the storage coordinate system (unit: mm). , for , Real-time movement speed in the direction (obtained by the laser rangefinder sampling 10 times per second). , The program uses the initial position coordinates (obtained via QR code positioning) and integrates an obstacle detection submodule. It uses a depth camera (90° field of view) to identify obstacles between shelves (such as protruding metal frames). When an obstacle is detected at a distance of <50mm, the program automatically plans a detour path (deviation ≤10mm). The program uses a PID control algorithm (proportional coefficient 0.5, integral time 0.2s, derivative time 0.1s) to adjust the moving speed in real time, ensuring that the sample placement accuracy is ≤2mm and improving storage efficiency by ≥40% (single sample placement time reduced from 30 seconds to 18 seconds).

[0033] The following two examples further illustrate the specific implementation of this system: Example 1: System Application in Laboratory Metrology and Testing Scenarios This embodiment is used for the routine testing of precision measuring instruments such as weights and electronic scales within the laboratory of a metrology institute. The specific operation mode of each module of the system is as follows: After the smart glasses software driver module starts, it automatically detects the hardware configuration and loads the appropriate driver. The dynamic resource scheduling submodule monitors the task load in real time. When the inspector starts 4K image acquisition for the appearance inspection of 0.1mg weights, the CPU core utilization automatically increases from the default 30% to 70%, and the memory allocation priority is increased by 2 levels. If switching to 1080P acquisition for the inspection of regular electronic scales, the CPU utilization drops to 40%, and the freed resources are used for background data preprocessing. The sensor driver adopts a three-thread synchronous mechanism. The ambient light intensity sensor collects laboratory illumination data ranging from 300 to 500 lux at a sampling rate of 10Hz. The temperature sensor records the surface temperature of the weights every 2 seconds, with an error controlled within ±0.5℃. The three-dimensional coordinate sensor outputs the spatial position of the instrument on the inspection table in real time. The wireless communication driver connects to the laboratory's 5G private network by default. When the signal strength is detected to drop to -105dBm, such as when near a metal shielding cabinet, it automatically switches to Wi-Fi and enables dual-band aggregation, increasing the bandwidth from 50Mbps to 120Mbps. The AR display driver uses double-buffered rendering to overlay the recognition results onto the real-time screen as a semi-transparent green frame, with the overlay latency remaining stable within 15ms and no stuttering.

[0034] During initialization, the AI ​​visual recognition software module loads a 32-layer sample recognition model (after model distillation) and a U-Net++ defect detection model. The models are stored in an embedded security chip and encrypted using AES-256. When detecting a 1kg standard weight, the sample recognition program extracts the metrological markings on the weight's surface using a sliding window algorithm and compares them with a built-in feature library containing multi-view features of 500 weights. The recognition process takes 45ms with an accuracy of 99.7%. When a new type of weight, such as a special weight with temperature compensation, is detected, the incremental learning algorithm is triggered. For every 100 newly added labeled data points, the model weights are fine-tuned while retaining 95% of the original weights. After three rounds of fine-tuning, the accuracy for recognizing new types of weights increases to 99.2%. The appearance defect detection program performs pixel-level scanning of the weight surface, identifying minute scratches as small as 0.5mm long, marked as red areas using semantic segmentation, and a deformation of 0.01mm. The defect location deviation is ≤0.1mm, and the detection results show 98.5% consistency with manual microscope detection.

[0035] The cloud platform software architecture modules are deployed on 8 server nodes in the Institute of Metrology's private cloud, using Kubernetes container orchestration. The data storage microservice writes each piece of test data, including 28 fields such as timestamp, weight number, and error value, into the InfluxDB time-series database. The original images are compressed using JPEG at a 1:5 ratio and then stored in MinIO object storage. Historical data older than 3 years is automatically migrated to cold storage nodes. The analysis microservice processes 500 test data entries per second using Spark Streaming, calculates the weight pass rate according to JJG99-2006, and generates a health score out of 100. Scores below 80 are marked as requiring calibration. The API gateway microservice uses a token bucket algorithm to limit requests to 800 per second, interacting with the laboratory management system via a RESTful interface, with a stable data upload response time of 80ms. When testing personnel submit a concentrated number of testing tasks, the container instances automatically scale up from 10 to 30, and shrink back to their initial state within 3 minutes after the peak task period.

[0036] The interactive control software module supports voice, gesture, and touch-based collaborative operation. When the inspector says "identify weights," the voice recognition program achieves a 96% accuracy rate under 60dB laboratory background noise through Mel-frequency feature extraction and bidirectional LSTM network analysis. When a fist gesture is made with a 92% matching rate of 21 feature points, the gesture recognition program responds with a "confirm inspection result" operation within 250ms. The AR virtual button allows for single-index annotation of defect locations, and two-finger zoom for 5x detail observation. The priority arbitration submodule defaults to a voice command weight of 0.6, gesture weight of 0.3, and touch weight of 0.1. When both a voice "cancel" and a gesture confirmation are triggered simultaneously, the system executes the voice command, keeping the conflict rate below 0.5%. Attention decisions are made according to the formula... Calculation, where To score attention, For the first Modal weights, It is the characteristic function.

[0037] The system's collaborative software modules are scheduled according to the "collection-preprocessing-identification-upload-analysis-feedback" process, following the formula... Calculate the edge-cloud switching decision value, where Set to 5 seconds. After the inspector aligns the weights, the camera captures an image, triggering a preprocessing event. Noise denoising is achieved using Gaussian filtering, and distortion correction is based on the camera's intrinsic parameter matrix. After processing, the image is pushed to the AI ​​recognition module via a RabbitMQ message queue. Once the recognition result is generated, an upload event is triggered, using a fragmented upload method with each fragment being 500KB. A breakpoint resume mechanism ensures that the upload can resume from the breakpoint even when the network fluctuates. After the cloud platform completes its analysis, a "qualified" result is pushed to the AR display module via WebSocket. Simultaneously, the TCC transaction coordination submodule confirms the data consistency between the smart glasses' local storage, the cloud platform database, and the management system, achieving a transaction success rate of 99.95%. The entire process audit log records the processing time of each module to the millisecond.

[0038] The AI ​​model optimization software module performs optimization once per hour, taking... , , The value is 1 for correct identification and -0.5 for incorrect identification, determined by the formula. Optimize model parameters, where For the model parameter matrix, Let be the gradient of the loss function at time t. After optimization of the recognition data under backlighting of 12000 lux, the model accuracy increased from 92% to 97%, while L1 regularization pruning eliminated 15% of redundant neurons, reducing the model size by 12%.

[0039] The data security software module encrypts transmitted data using the SM4 algorithm, according to the formula... Perform encryption, where For the encrypted data block, The 256-bit key is updated daily at 3 AM via the ECDH protocol. This is the original test data. It is a 32-round nonlinear transformation round function. The encrypted data block is transmitted via the TLS 1.3 protocol. The detection data hash value is calculated using SHA-256 and uniquely encoded with the weight to package into a transaction. After consensus by 3 nodes, it is written into the consortium chain block. A block is generated every 8 seconds. On-chain data can be traced back to the original image through the hash value. The tamper detection success rate is 100%.

[0040] The scene-adaptive image stabilization program enables a strong stabilization mode in static laboratory scenes. It uses Kalman filtering to fuse 100Hz accelerometer and 200Hz gyroscope data to generate a motion transformation matrix. The algorithm formula is as follows: ,in For the image after image stabilization, For the original image, With a 3×3 motion transformation matrix containing translation, rotation, and scaling parameters, the image blur score decreased from 0.35 to 0.15 after image stabilization.

[0041] Table 1 Comparison of System Performance Indicators in Laboratory Scenarios Index Existing system Inventive system Sample identification accuracy 95.2% 99.5% Defect detection minimum size 1.0 mm 0.5 mm Data upload response time 200 ms 80 ms Multi-modal instruction conflict rate 3.8% 0.5% Endurance time (continuous detection) 6 hours 10.5 hours Table 1 shows that the performance of this invention in laboratory scenarios is significantly superior to existing systems. The recognition accuracy is improved by 4.3 percentage points, thanks to multi-scale feature fusion and incremental learning algorithms, which can accurately match the features of different types of weights; the defect detection capability is doubled, capable of identifying finer scratches and deformations, meeting high-precision metrology requirements; the data upload speed is improved by 60%, attributed to the microservice architecture and bandwidth optimization of the cloud platform; the command conflict rate is reduced by 87%, achieving orderly collaboration of multimodal interactions through a priority arbitration mechanism; and the battery life is extended by 75%, with the low-power management program showing significant effect in resource control for non-core tasks, fully covering the laboratory's all-day testing needs.

[0042] Example 2: System Application in the Batch Verification of Precision Instruments by the Metrology Institute This embodiment is used by metrology institutes for centralized verification of precision instruments such as batch gauge blocks and standard resistors. The system is optimized for multi-batch, high-precision testing needs as follows: After the smart glasses software driver module starts, it automatically switches to batch inspection mode. The dynamic resource scheduling submodule detects that 20 sets of gauge blocks need to be processed simultaneously in parallel inspection, stabilizing the CPU core utilization at 75% and allocating 50% of memory to multi-threaded data processing. The camera driver automatically matches the resolution according to the type of instrument: 4K resolution at 40fps and macro mode is enabled for gauge block inspection; 2K resolution at 30fps is used for standard resistor inspection to ensure clear pin details. The sensor array synchronously collects data on the laboratory constant temperature environment (temperature controlled at 20±0.5℃, humidity 45% to 55%) and the flatness of the instrument placement data at a sampling rate of 20Hz. The driver program processes the flatness data through mean filtering, with the error controlled within ±0.001mm. The wireless communication driver prioritizes connection to the 10 Gigabit local area network within the metrology institute. When the signal strength drops to -95dBm near the shielded laboratory, it automatically switches to the backup fiber optic link to maintain a bandwidth of over 1Gbps. The AR display driver adopts a split-screen mode, displaying the real-time image on the left and historical test data on the right. The brightness is adjusted to 300 nits to ensure fatigue during long-term observation, and the superposition latency is stabilized at 12ms.

[0043] The AI ​​visual recognition software module loads an optimized model supporting batch recognition, compressed to 10MB. It also loads a block feature library containing length and grade information for 1000 blocks, and a standard resistor feature library containing nominal values ​​and accuracy grades for 500 resistors. When performing batch inspection of grade 0 blocks, the sample recognition program utilizes multi-scale feature fusion. To achieve feature enhancement, among which Capture the overall outline of the block for a 1 / 8 resolution feature map. Extract block end face details from 1 / 4 resolution feature maps. Identify the nominal values ​​of the blocks in the 1 / 2 resolution feature map. This is a channel-level splicing operation. For the attention-weighted fusion operation with contour 0.2, detail 0.6, and nominal value 0.2, the single-group gauge block recognition time is 55ms, and the batch recognition accuracy is 99.6%. The appearance defect detection program focuses on identifying bumps and knocks on the gauge block end faces with a minimum recognition area of ​​0.01mm. 2 With a minimum scratch length of 0.3mm, morphological operations are used to enhance defect features, resulting in a defect false negative rate of less than 0.3% in batch inspection. The model supports offline batch processing and can cache 500 sets of inspection data for synchronization when the network is idle.

[0044] The cloud platform software architecture module employs a distributed computing cluster for batch data processing, consisting of a computing resource pool of 10 server nodes. The data storage microservice partitions and stores batch inspection data by appliance type. Block data is written to a dedicated table containing 15 fields, including length deviation and flatness. Standard resistance data contains 12 fields, including resistance value and temperature coefficient. Original images are stored using lossless compression and are automatically backed up daily at 2 AM. The analysis microservice uses the Flink stream processing engine to process 1000 sets of inspection data per second in parallel, calculating the length deviation trend of the blocks according to a formula. Calculate the predicted value, where This is the predicted length deviation at time t+1. The length deviation is from the last 7 tests. to Using values ​​of 0.3, 0.25, 0.2, 0.1, 0.08, 0.05, and 0.02 respectively, and b being the material compensation coefficient for the gauge block (0.002mm for steel), the prediction deviation is ≤0.0005mm, providing an early warning of gauge block stability changes up to 3 months in advance. The API interface supports batch data import and export, using the protobuf protocol for transmission efficiency that is 50% higher than JSON.

[0045] The interactive control software module is adapted to the high-efficiency operation requirements of batch testing. The voice recognition program supports batch command recognition such as "batch start" and "skip current" with an accuracy of 97% under 45dB background noise in the laboratory. The gesture recognition program is optimized for fast and continuous operation recognition. After confirming with a fist, a swipe within 0.5 seconds can trigger the "next group" operation with a response time of 200ms. The touch interaction supports batch selection, allowing multiple instruments within a selected area to be marked with their testing status simultaneously. The user habit learning submodule records the preferred voice batch commands of the testing personnel and assigns voice weights to them. The value was adjusted from 0.6 to 0.7 to further reduce multimodal conflict, and attention decision-making was performed according to the formula. Calculate, where A is the attention score, For the first Modal weights, This is a characteristic function to ensure reasonable instruction priority.

[0046] The system's collaborative software module employs pipeline-style task scheduling during batch testing, according to the formula... Calculate the edge-to-cloud handover decision value, where T is set to 10 seconds. When bulk data uploads cause network latency When it rises to 150ms and lasts for 10 seconds, Triggering edge computing mode, the pre-loaded batch recognition model on the edge nodes takes over 50% of the recognition task, with local processing latency controlled within 80ms; after network recovery... Reduced to 80ms Automatically switch back to cloud computing mode; edge data saves 70% bandwidth by transmitting only the differences in data through incremental synchronization.

[0047] The AI ​​model optimization software module addresses the diversity of instruments in batch testing, taking into account... , When correctly identifying multiple sets of the same model of equipment, the reward value increases by 0.3, according to the formula. Optimize model parameters, where For the model parameter matrix, for The gradient of the loss function at each time step improves the recognition speed of batches of devices of the same model by 20%, while keeping the difference in recognition accuracy between different models within 1%. The model... After regularization pruning, redundant parameters are reduced by 25%, and memory usage is reduced by 30% during batch processing.

[0048] The data security software module encrypts batch detection data using the SM4 algorithm, according to the formula... Perform encryption, where For the encrypted data block, The 256-bit key is updated every 8 hours through the key distribution center. The 32-round round function includes nonlinear and linear transformations, and the encrypted data blocks are packaged and stored in batches. After each batch of testing is completed, the hash value of key data is stored as evidence through the metrology institute's internal consortium blockchain. After consensus confirmation by three audit nodes, it is written into the block generation interval of 5 seconds to ensure the integrity and traceability of batch data.

[0049] The intelligent warehouse alignment software module acquires real-time velocity in the x and y directions via a 10Hz laser rangefinder sensor when the bulk blocks are being stored. ,according to Calculate the displacement, where For displacement, , The initial position of the gauge blocks is determined by the warehousing system with an error ≤0.5mm. During the alignment process, if a shelf grid deviation >2mm is detected, the PID control algorithm automatically adjusts the moving speed proportional coefficient to 0.6, the integral time to 0.3 seconds, and the derivative time to 0.1 seconds, ultimately achieving an alignment accuracy of 1mm. This improves batch warehousing efficiency by 50%, reducing the warehousing time for a single batch of 50 gauge blocks from 15 minutes to 7.5 minutes.

[0050] Table 2 Comparison of System Performance Indicators in Batch Testing Scenarios of the National Institute of Metrology Index Existing system Inventive system Batch identification accuracy 96.1% 99.6% Single group detection time consumption 120 ms 55 ms Data transmission flow consumption 100% 30% Batch storage efficiency 50 groups / 15 minutes 50 groups / 7.5 minutes Model memory occupation 100% 70% Table 2 shows that the present invention has significant advantages in batch testing scenarios at the Institute of Metrology. The 3.5 percentage point improvement in batch identification accuracy is attributed to the multi-scale feature fusion and batch optimization model, which can accurately match subtle differences between different batches of the same model. The 54% reduction in single-group testing time is attributed to the model pruning and parallel processing mechanisms, which significantly improve batch processing efficiency. Data transmission traffic is reduced by 70%; edge-cloud collaboration and incremental synchronization effectively reduce the pressure of large-scale data transmission. Batch warehousing efficiency is improved by 100%; intelligent alignment algorithms and PID control ensure accurate and rapid operation in high-density warehousing. Model memory usage is reduced by 30%; the lightweight design allows the system to simultaneously load multiple instrument feature libraries to meet diverse batch testing needs. Overall, the system comprehensively improves the efficiency and accuracy of batch precision instrument verification at the Institute of Metrology, and is adapted to centralized, high-frequency testing work modes.

[0051] Reference Figure 2 This figure visually demonstrates the recognition stability of this invention under different lighting conditions. Existing systems show a significant decrease in accuracy with increasing light intensity, dropping to 75% under 10,000 lux of strong light, due to a lack of optimization for feature extraction in response to changes in lighting. In contrast, this invention, through multi-scale feature fusion and AI model optimization algorithms, maintains an accuracy rate of over 97% across the entire lighting range, with a decrease of only 1.7 percentage points under strong light. This is attributed to the attention mechanism's focus on key features under strong light and the reinforcement learning's dynamic compensation for lighting interference, ensuring reliable operation of metrological testing in complex lighting environments such as laboratories and outdoors, and reducing misjudgments caused by lighting conditions.

[0052] Reference Figure 3This figure clearly demonstrates the efficiency advantages of this invention in batch testing. Existing systems experience a significant increase in processing time as the batch size increases, reaching 165ms for 50 batches due to the lack of parallel processing and lightweight model design. This invention, through a distributed computing cluster and pruning optimization model, achieves a processing time of only 70ms for 50 batches, with a gradual increase in processing time. This stems from the efficient allocation of CPU and memory through dynamic resource scheduling, and the shared computation of repetitive features in the batch identification algorithm. This allows metrology institutes to ensure testing accuracy while significantly improving throughput when processing large-scale instrument verification, meeting the high-efficiency requirements of centralized verification.

[0053] Reference Figure 4 This diagram illustrates the intelligent power management strategy of this invention. Existing systems distribute power evenly, leading to frequent delays in emergency testing tasks due to insufficient resources. This invention, through a task priority algorithm, allocates 60% of power to emergency testing and compresses log synchronization to 10%. This stems from the low-power management program's dynamic adjustment of hardware parameters, prioritizing core testing functions when power is low, thus extending the runtime of emergency tasks. For metrology personnel conducting external inspections, this avoids testing interruptions caused by equipment power failures, significantly improving equipment reliability, especially in remote areas without charging facilities.

[0054] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart glasses system for metrology and testing integrating artificial intelligence visual recognition and a cloud platform, characterized in that, Includes the following modules: The smart glasses software driver module develops drivers adapted to cameras, sensor arrays, wireless communication, and AR display units; it includes a dynamic resource scheduling submodule that can adjust resource utilization and memory allocation; the camera supports dynamic switching; and the sensors process light intensity, temperature, and 3D coordinate data in a multi-threaded manner. Wireless communication can monitor network quality, switching to Wi-Fi and enabling bandwidth aggregation when 5G signal is weak; AR display uses double-buffered rendering; The AI ​​visual recognition module develops and improves sample recognition and defect detection programs; sample recognition uses a model distillation compression network with a built-in metrology instrument feature library and incremental learning to update weights. Defect detection can locate scratches, deformations, and missing nameplates; The cloud platform software architecture deploys distributed microservices on the Institute of Metrology's private cloud, including data storage, analysis, API gateway, and container orchestration services; Data storage uses time-series and object databases, supporting lifecycle management; the analytics service integrates a real-time computing engine, which can collect statistical data and generate health scores. The API gateway has a flow-limiting mechanism and provides dual interfaces; The interactive control module develops multimodal interactive programs using voice, gestures, and touch. Voice commands support offline recognition with noise adaptation. Touch controls support AR virtual control operations. Interactive voice > gestures > touch priority is coordinated. The system collaboration module develops an event-driven scheduler, defines a six-stage process, and decouples each stage through a message queue; it also includes a distributed transaction coordination submodule.

2. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform as described in claim 1, characterized in that, It also includes an AI model optimization software module, developing an online reinforcement learning optimization program, with optimization formulas as follows: ,in For the model parameter matrix, For learning rate, To optimize the cycle, for The reward value for the recognition result at each moment. for The cross-entropy loss function at time step [time] is added to the program during incremental model training. Regularized pruning.

3. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform as described in claim 1, characterized in that, It also includes a data security software module, developing encryption programs and consortium blockchain evidence storage programs based on the national cryptographic algorithm SM4, and the encryption program implementation formula. Software-based execution, in which For the encrypted data block, It is a 256-bit dynamic key. This is the original test data. This program is a 32-round nonlinear transformation round function based on the SM4 algorithm, and it performs double encryption on both the transmitted data and the locally stored data.

4. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, The smart glasses software driver module develops a scene-adaptive image stabilization program. This program collects motion data through an accelerometer and gyroscope, and uses a Kalman filter algorithm to fuse and generate a motion state matrix. The algorithm formula is as follows: ,in For the image after image stabilization, For the original image, The program uses a 3×3 motion transformation matrix and has a built-in scene recognition submodule. It distinguishes between static and dynamic scenes by the rate of change of image feature points. Static scenes use a strong anti-shake mode, while dynamic scenes use a weak anti-shake mode.

5. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, An attention-enhanced multi-scale feature fusion program was developed within the AI ​​visual recognition software module. This program implements the formula... Software-based processing, among which It is a 1 / 8 resolution feature map. A 1 / 4 resolution feature map. For 1 / 2 resolution feature maps, This is a channel-level splicing operation. For attention-weighted feature fusion operations, the program develops a feature enhancement submodule for small-sized measuring instruments.

6. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, Develop a condition-aware time series prediction program within the cloud platform software architecture module, and complete the formula. The algorithm execution, its This represents the predicted equipment error value at time t+1. For history The detection data at each moment, is the time decay weight, and b is the bias term.

7. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, Develop a user-adaptive multimodal attention decision-making program in the interactive control software module and complete the formula. The software calculation, among which For interaction modality attention score, For the first Weights of different interaction modalities For the first Characteristic functions of various interaction modes =3 represents the total number of interaction modalities, and the program has a built-in user habit learning submodule.

8. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, Develop an intelligent preloading program for switching between edge computing and cloud computing within the system's collaborative software module, and complete the formula. The software calculation, in which To switch decision integral values, for Real-time network latency, The delay threshold, This refers to the decision-making cycle.

9. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, The smart glasses software driver module includes a task priority low-power management program that employs a power prediction algorithm. ,in To predict total power consumption, For the first Real-time power consumption of each hardware module To estimate the working time for each module, the program divides tasks into three levels.

10. The intelligent glasses system for metrology and testing integrating artificial intelligence visual recognition and cloud platform according to claim 1, characterized in that, It also includes an intelligent warehouse alignment software module, developing a dynamic alignment control program that combines AI vision and laser ranging, and completing the formula. The software calculation, among which , This represents the displacement of the sample in the storage coordinate system. , for , Real-time movement speed in direction, , Using the initial position coordinates, the program integrates an obstacle detection submodule, which uses a depth camera to identify obstacles between shelves.

Images