A camera-assisted detection instrument field intelligent calibration method, system, storage medium and program product

Abstract

The present application relates to the technical field of instrument calibration, and particularly relates to a camera-assisted detection instrument field intelligent calibration method and system, a storage medium and a program product. The present application collects multi-view video streams through a camera array, analyzes instrument readings using a convolutional neural network-OCR model, combines environmental sensor data with an LSTM model for drift compensation, and generates calibration errors in real time. The system automatically records the entire process data on the chain, and provides correction instructions through an AR terminal in abnormal situations. Finally, a structured calibration report containing maintenance suggestions and blockchain hashes is generated, ensuring that the calibration results are reliable, traceable and automated.

Description

Technical Field

[0001] This invention relates to the field of instrument and equipment calibration technology, and in particular to a camera-assisted on-site intelligent calibration method, system, storage medium, and program product for testing instruments. Background Technology

[0002] Testing instruments (such as pressure, temperature, flow, spectral, and component analysis equipment) play a core quality benchmark role in highly sensitive industries such as pharmaceuticals, energy, food, and semiconductor manufacturing. More and more companies are required to perform periodic calibrations on-site to avoid downtime losses and secondary errors caused by disassembly and transportation. However, traditional on-site calibration still heavily relies on manual operation and paper / Excel records: calibration personnel manually transcribe readings, take photos, fill out reports, and then enter the data into the system back in the office. This process is cumbersome, subjective, and prone to errors, and the calibration traceability chain lacks objective evidence, making it difficult to meet the comprehensive requirements of digital quality management for real-time performance, traceability, and non-repudiation.

[0003] The current mainstream approach is to generate a unique QR code for each instrument in the metrology management platform, combining this with functions such as online commissioning by verification institutions and certificate downloading to achieve electronic file management. For example, Chinese invention patent CN106934630A proposes to link QR code labels with a cloud database to achieve a "one item, one code" traceability method for basic instrument information and verification conclusions. This solution improves the efficiency of certificate circulation, but still relies on manual entry of results, lacks "process-level" supervision of the calibration process itself, and cannot prevent errors in reading transcription or falsification.

[0004] As the price of industrial vision hardware decreases, some companies have begun installing cameras on-site to photograph and document key steps, manually comparing the data with instrument readings. However, most systems only perform "post-event review" and do not automatically extract data. Chinese invention patent CN105122303A explores "camera calibration using feature recognition" technology, which extracts image features to calculate camera parameters to ensure imaging accuracy; Chinese invention patent CN117119113B further proposes user-assisted camera self-calibration to improve OCR recognition accuracy. However, both focus on camera self-calibration and do not address the industry pain point of "simultaneous acquisition and real-time analysis of standard / calibrated instrument readings from multiple perspectives."

[0005] Currently, some manufacturers are attempting to introduce OCR into on-site calibration: using an industrial camera to photograph the calibration meter head and extracting numbers using general character recognition algorithms. However, the false recognition rate increases significantly in scenarios with strong backlighting, reflections, and irregular fonts; furthermore, a single camera angle is easily obstructed by people. While multi-camera complementary acquisition can increase the success rate, it requires complex external parameter calibration and time synchronization; otherwise, it is difficult to achieve millisecond-level data alignment, affecting the accuracy of the one-to-one correspondence between readings and standard values.

[0006] Artificial intelligence has been widely used in safety monitoring for recognizing behaviors such as posture, falls, and intrusions. The "AI Action Anomaly Analysis and Early Warning System," based on YOLOv5 and PoseEstimation, can determine the compliance of personnel movements in real time and detect unauthorized clothing or boundary violations. Multimodal large-scale models have achieved accident early warning scores exceeding 80 points in high-risk industries such as civil explosives and chemicals. However, publicly available information shows that these models are primarily geared towards safe production, focusing on the abnormal behavior itself, and have not yet provided a complete logical chain for whether ruler / instrument readings are correctly acquired or whether steps are executed in the correct order of verification procedures.

[0007] Blockchain inherently possesses tamper-proof and traceable characteristics, and has already been used in judicial video recording, agricultural traceability, and logistics delivery confirmation. Chinese invention patent CN116150234A discloses a "data storage method based on blockchain," which can write target data and its hash onto the blockchain after a user pays and return a storage certificate. In the field of metrology, some discussions have proposed writing PDF certificate hashes onto the blockchain, but the large volume of video streams, environmental parameters, and real-time differential data generated in this process is often constrained by issues such as "inability to store on the blockchain" or "only retaining the hash and lacking an online verification entry point," making it difficult to truly form a strong chain of evidence.

[0008] To address this, some instrument suppliers in Europe and the United States have already provided remote optimization and calibration services, such as SCIEX's "Remote Instrument Optimization and Calibration," which allows technical engineers to adjust and control parameters online. Meanwhile, AR Remote Assistance technology, combining real-time video streaming and virtual annotation, enables remote experts to guide on-site operators in completing complex repairs or adjustments. However, based on publicly available case studies, AR is mostly used in repair or installation scenarios. There are no readily available plugins for error calculations specific to metrology and calibration, or for synchronous comparisons between standards and calibrated instruments. Furthermore, issues such as unstable industrial wireless networks, privacy compliance, and data storage remain unresolved.

[0009] Meanwhile, in laboratory metrology, temperature, humidity, atmospheric pressure, magnetic fields, and even vibration can cause significant drift in high-precision instruments. The traditional approach is to place thermometers, hygrometers, and barometers on-site, with calibration personnel recording environmental values ​​and manually calculating correction coefficients. However, in outdoor or high-temperature workshops, environmental fluctuations are drastic, making it difficult to obtain continuous compensation curves through "point measurements." While IoT micro-sensor networks (such as LoRa-WAN data acquisition nodes) can provide second-level data, they lack real-time fusion algorithms with calibration error models, causing compensation strategies to remain at the "offline regression" stage.

[0010] In summary, while existing technologies have made progress in image recognition, AI behavior analysis, blockchain evidence storage, remote collaboration, and environmental IoT, a systematic solution that fully integrates multi-camera automatic reading, AI anomaly detection, real-time environmental compensation, and on-chain evidence storage is still lacking. This is particularly true in the niche scenario of on-site calibration of testing instruments, where extremely high precision, timing, and the integrity of the evidence chain are required, where significant technological gaps remain. Summary of the Invention

[0011] This invention addresses key pain points in existing on-site calibration processes for testing instruments, such as difficulty in verifying personnel identity, susceptibility to distortion in manual records, lack of objective process-level evidence for calibration steps, inability to compensate for environmental interference in real time, and susceptibility to data tampering and data silos. It provides a camera-assisted intelligent on-site calibration method for testing instruments, achieving accurate identification, real-time monitoring, environmental compensation, reliable evidence collection, and intelligent closed-loop management of the entire on-site calibration process. Compared to existing technologies that only provide single-point functions, this method significantly improves calibration accuracy, efficiency, and traceability, demonstrating significant technological advancement and economic value.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] A camera-assisted intelligent on-site calibration method for testing instruments includes the following sequential steps:

[0014] 1) Confirmation of the legality of the venue and personnel:

[0015] The coordinates of the calibration site are obtained through GPS and / or Beidou dual-mode positioning and compared with the coordinates of the pre-stored work orders; and the identity of the calibration personnel is uniquely verified by dual facial and voiceprint recognition.

[0016] 2) Multi-camera panoramic acquisition and calibration:

[0017] No fewer than three high-resolution cameras with preset positions are arranged around the instrument being calibrated. The internal and external parameters of the cameras are obtained by Zhang Zhengyou's calibration method and uploaded to the cloud. Each camera is started synchronously to generate a multi-view video stream with timestamps.

[0018] 3) Visual-data fusion calibration:

[0019] The convolutional neural network-OCR combined model is called to automatically identify the standard instrument reading and the value displayed by the instrument being calibrated in the video stream, and write them into the local cache according to the one-to-one correspondence; the real-time identified value and the standard calibration point are differentially calculated to obtain the first calibration error;

[0020] 3) Comprehensive judgment of abnormal actions and environment:

[0021] The human posture estimation algorithm is used to detect whether the action sequence of the calibration personnel deviates from the preset working conditions; temperature, humidity, light and vibration data are collected simultaneously, and environmental drift compensation is performed on the first calibration error to obtain the second calibration error;

[0022] 4) Blockchain-based evidence storage and AR prompts:

[0023] After performing SHA-3-256 calculation on the image hash, environmental data and second calibration error generated in steps 2)-4), the data is written into the consortium blockchain and digitally signed by elliptic curve P-256. If the error or abnormal score exceeds the threshold, the correction instruction is immediately displayed on the wearable AR terminal, and remote experts are allowed to annotate in the AR screen in real time via WebRTC.

[0024] 5) Intelligent report generation:

[0025] Based on the differential calculation results, the first frame screenshot of the camera evidence image, and the environmental drift compensation model, a structured calibration report with predictive maintenance suggestions is automatically generated, and on-chain transaction hashes are attached.

[0026] Preferably, step 2) specifically includes the following steps:

[0027] a) Arrange at least three 4K industrial cameras around the instrument being calibrated at three or more non-collinear points, so that their optical axes intersect to form a minimum triangular mesh;

[0028] b) Perform Zhang Zhengyou calibration on each camera sequentially using a 9×6 checkerboard calibration board to obtain the intrinsic parameter matrix K, distortion coefficients k1, k2, p1, p2 and extrinsic parameter matrix [R|t], and then calculate the SHA-3-256 hash of the parameter file and upload it to the cloud storage.

[0029] c) Synchronize the time base of each camera system using the IEEE-1588 precision clock to ensure that the difference between the trigger acquisition times is no more than 5ms, and write a nanosecond-level timestamp into the frame header;

[0030] d) Each camera starts recording simultaneously under the same trigger command, outputting a multi-view video stream with a unified timestamp, providing spatiotemporally consistent image data for subsequent vision-data fusion calibration.

[0031] Preferably, step 3) specifically includes the following steps:

[0032] a) At least three cameras synchronized with an IEEE-1588 clock capture timestamped I-frames at the same trigger time;

[0033] b) Perform contrast-limited adaptive histogram equalization, bilateral filtering denoising, and YOLO-based ROI auto-cropping sequentially on the I-frame;

[0034] c) The cropped image is fed into the ResNet-BiLSTM-CTC convolution-OCR combined network to obtain the standard instrument reading and the instrument under calibration and their confidence levels, respectively.

[0035] d) Write the two types of readings into the local cache one by one according to the timestamp, and fuse the results from multiple cameras through a confidence-weighted voting mechanism;

[0036] e) Perform differential calculation between the fused reading and the preset standard calibration point to generate the first calibration error and archive it with frame hash and confidence label.

[0037] Preferably, step 4) specifically includes the following steps:

[0038] a) Integrate the key I-frame hash, environmental baseline hash and second calibration error collected by multiple cameras into evidence storage metadata, calculate the SHA-3-256 digest and generate a digital signature using the ECDSA-P256 private key;

[0039] b) Adopting a “consortium blockchain + IPFS” layered strategy, the metadata and the IPFS content address of the original video and sensor files are packaged together into a transaction structure and written into the blockchain through BFT-DPoS consensus, with a block time of ≤5 seconds.

[0040] c) When the compensation error or the overall anomaly score is higher than the preset threshold of 0.1%FS or 0.4 points, a red correction instruction will immediately pop up on the wearable AR terminal, and remote experts will be allowed to draw vector annotations in the AR view in real time through the WebRTC channel. The on-site operator can continue the subsequent calibration process only after completing the correction according to the annotations.

[0041] Preferably, the AR terminal automatically overlays the real-time differential results of step C onto the screen area of ​​the instrument being calibrated based on SLAM technology.

[0042] Preferably, when the error exceeds the threshold, the system automatically sends a freeze command to the enterprise ERP to lock the current calibration work order and prevent the instrument from being put into use; the structured calibration report includes a calibration reliability index, which is obtained by weighting the abnormal action score, the environmental steady state score and the error residual score.

[0043] Furthermore, the present invention also provides an on-site intelligent calibration system for testing instruments used to implement the method, comprising:

[0044] a) Positioning and identity verification unit, including GPS / BeiDou module and face-voiceprint integrated recognition terminal;

[0045] b) A camera array unit consisting of at least three 4K cameras, with hardware synchronization and self-calibration capabilities;

[0046] c) Vision-data fusion engine, deployed on GPU edge gateway, is used to perform OCR recognition, differential calculation and pose analysis;

[0047] d) Environmental perception and compensation module, including temperature and humidity, light and vibration sensors and LSTM-based drift compensation submodule;

[0048] e) Blockchain client module, used to write image hashes, compensation data and calibration errors, and execute digital signatures;

[0049] f) The AR collaboration module supports on-site visual guidance and real-time annotation by remote experts;

[0050] g) Intelligent report generation and ERP interface module, used to output structured reports and interact with enterprise information systems.

[0051] As a preferred embodiment, the camera array unit supports H.265 encoding and RTSP streaming, and has low-light imaging capability in scenarios with a resolution of ≤10 lux; the vision-data fusion engine supports 60 FPS real-time inference with a latency of ≤200 ms; the environmental perception and compensation module has a temperature measurement accuracy of ±0.1℃ and a relative humidity measurement accuracy of ±1%RH; the blockchain client module adopts the BFT-DPoS consensus algorithm with an average block time of ≤5s; and the AR collaboration module is compatible with the OpenXR standard and can run on mainstream AR glasses and smartphones.

[0052] Furthermore, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor implements the steps of the method.

[0053] Furthermore, the present invention also provides a computer program product comprising program instructions that, when the instructions are run on an electronic device, cause the electronic device to perform the method.

[0054] The present invention, by adopting the above-described technical solution, has the following technical effects:

[0055] 1. Dual identification and site verification throughout the entire process: Through real-time GPS / BeiDou positioning and face-voiceprint dual-factor authentication, calibration personnel are bound to work orders and site coordinates one by one, completely eliminating cross-site operations and unlicensed personnel calibrating on behalf of others, significantly improving the compliance and traceability of on-site calibration activities.

[0056] 2. Automated acquisition of readings with zero transcription error: After unified calibration, the multi-view camera array synchronously acquires the display values ​​of the standard and the instrument being calibrated; the convolutional-OCR joint network completes digital analysis and mapping in milliseconds, realizing automatic differential calculation of instrument-standard readings "one-to-one", eliminating manual transcription and secondary input errors.

[0057] 3. Real-time warning of abnormal behavior by multimodal AI: Combining human posture estimation and voice context analysis, the system can identify illegal actions, irrelevant dialogues or unauthorized device access within <200ms, and push correction instructions to the AR terminal in real time, reducing the on-site violation rate by more than 70% and significantly ensuring the standardization of the calibration process.

[0058] 4. Real-time environmental drift compensation and uncertainty compression: Through continuous monitoring of multi-dimensional environmental factors such as temperature, humidity, vibration, and electromagnetic noise via LoRa-WAN micro-sensor network, and in conjunction with the LSTM drift model to correct calibration errors online, the uncertainty of high-precision instruments in complex field environments is reduced to the laboratory level (10^-6 order of magnitude), and the error convergence speed is increased by 2 times.

[0059] 5. Non-repudiable on-chain evidence collection and data integrity: Adopting a "consortium blockchain + IPFS sharding" layered storage strategy, image hashes, environmental data and calibration results are automatically written to the blockchain and encrypted with elliptic curve digital signatures; third parties can verify data integrity at any time, effectively preventing later tampering and enhancing legal validity and customer trust.

[0060] In summary, this invention has achieved systematic technological breakthroughs in multiple dimensions, including identity verification, reading accuracy, anomaly warning, environmental compensation, data evidence collection, remote collaboration, reporting efficiency, and security compliance. It comprehensively improves the accuracy, efficiency, and reliability of on-site calibration of testing instruments, and has significant technological advancement and industry promotion value. Attached Figure Description

[0061] Figure 1 This is a block diagram of the overall structure of an on-site auxiliary monitoring and calibration system for a testing instrument according to the present invention.

[0062] Figure 2 This is a flowchart of a camera-assisted on-site intelligent calibration method for testing instruments.

[0063] Figure 3 This is a functional structure diagram of the camera-assisted acquisition and image recognition module in the system of this invention.

[0064] Figure 4 This is a flowchart of the AI ​​abnormal behavior recognition module in this invention.

[0065] Figure 5 This is a schematic diagram of the model structure and data flow of the environmental factor drift compensation module in this invention.

[0066] Figure 6 This is a flowchart of the on-chain evidence storage module in this invention.

[0067] Figure 7 This is a flowchart illustrating the result binding process based on OCR and automatic analysis of calibration errors in this invention. Detailed Implementation

[0068] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0069] like Figure 1 , Figure 2 As shown, a camera-assisted intelligent on-site calibration method for testing instruments includes the following sequential steps:

[0070] 1) Verification of site and personnel legitimacy: The coordinates of the calibration site are obtained through GPS / BeiDou dual-mode positioning and compared with the coordinates of the pre-stored work order; and the identity of the calibration personnel is uniquely verified by dual facial and voiceprint recognition.

[0071] 2) Multi-camera panoramic acquisition and calibration: At least three high-resolution cameras with preset positions are arranged around the instrument being calibrated. The internal and external parameters of the cameras are obtained by Zhang Zhengyou's calibration method and uploaded to the cloud. Each camera is started synchronously to generate a multi-view video stream with timestamps.

[0072] 3) Visual-data fusion calibration: The convolutional neural network-OCR combined model is called to automatically identify the standard instrument reading and the value displayed by the instrument being calibrated in the video stream, and write them into the local cache according to the one-to-one correspondence; the real-time identified value and the standard calibration point are differentially calculated to obtain the first calibration error;

[0073] 3) Comprehensive judgment of abnormal actions and environment: The human posture estimation algorithm is used to detect whether the action sequence of the calibration personnel deviates from the preset working conditions; temperature, humidity, light and vibration data are collected simultaneously, and environmental drift compensation is performed on the first calibration error to obtain the second calibration error;

[0074] 4) Blockchain Evidence Storage and AR Prompts: The image hash, environmental data and second calibration error generated in steps 2)-4) are calculated using SHA-3-256 and written into the consortium blockchain, and digitally signed by elliptic curve P-256; if the error or abnormal score exceeds the threshold, the correction instruction is immediately displayed on the wearable AR terminal, and remote experts are allowed to annotate in the AR screen in real time via WebRTC.

[0075] 5) Intelligent report generation: Based on the differential calculation results, the first frame screenshot of the camera evidence image and the environmental drift compensation model, a structured calibration report with predictive maintenance suggestions is automatically generated and an on-chain transaction hash is attached.

[0076] The following description follows eleven major stages and thirty-two steps: "Preliminary Preparation—Identity and Site Rights Confirmation—Multi-Camera Calibration—Environmental Factor Baseline Establishment—Calibration, Data Acquisition and Error Calculation—AI Anomaly Detection—Real-time Environmental Drift Compensation—On-Chain Evidence Storage—AR Collaboration—Report Generation—Data Archiving and Maintenance".

[0077] Phase 1: Preliminary Preparations

[0078] 1.1 Hardware Layout Planning

[0079] The on-site supervisor first determines the installation positions of the camera array and environmental sensor nodes based on the model, installation height, and operating channel width of the instrument being calibrated (hereinafter referred to as "calibrated instrument"). In principle, at least three cameras should form a minimum triangular mesh to obtain complete complementary viewing angles. If the calibrated instrument has more than two display panels, additional cameras can be installed directly in front of the corresponding panels. All cameras and sensor nodes are fixed using industrial magnetic bases or expansion bolts, meeting the requirements for shock resistance and IP65 protection.

[0080] 1.2 Edge gateway power-on and network connectivity check

[0081] The GPU-edge-gateway (referred to as "edge gateway") is mounted in a wall-mounted cabinet within 5 meters of the instrument. The edge gateway connects to the enterprise's backbone network via Gigabit Ethernet and is configured with dual-band Wi-Fi 6APs to provide wireless backup links for cameras, AR glasses, and environmental nodes. Technicians connect a 220V industrial power supply and verify that the UPS module passes its self-test before proceeding to the next stage.

[0082] Phase Two: Identity and Site Ownership Confirmation

[0083] 2.1 Upload Work Order Geographic Coordinates

[0084] The calibration scheduling system automatically writes latitude and longitude coordinates (WGS-84) into the work order and pushes it to the PostgreSQL database of the edge gateway. The error threshold for comparing these coordinates with the real-time readings of the GPS / BeiDou module is set at 10m; exceeding the threshold will trigger a red warning and prevent calibration from starting.

[0085] 2.2 Issuance of Personnel Identity Templates

[0086] The enterprise digital identity platform calls the OAuth2.0 interface to send the facial feature vector, voiceprint feature vector, and employee number of the authorized personnel for this calibration to the local TEE of the edge gateway; at the same time, a one-time token is generated for the calibration personnel to log in by scanning a QR code on the AR glasses. This token expires after 10 minutes to prevent interception and replay.

[0087] 2.3 On-site check-in and confirmation of rights

[0088] Calibration personnel arrive on-site and use AR glasses to perform liveness detection and compare facial vectors. They then read aloud a five-digit verification code randomly generated by the system to complete voiceprint verification. The system signs and writes the two-factor authentication result along with real-time GPS coordinates into the chain. If any step fails, the calibration process automatically terminates and is recorded as an "illegal attempt" event.

[0089] Phase 3 Multi-camera Calibration

[0090] Step 3.1 External parameter measurement and coarse calibration

[0091] A laser rangefinder is used to obtain the distance and angle between the camera and the center point of the surface of the instrument being calibrated. The ranging data is input into the calibration assistant program, which automatically generates the initial extrinsic parameter matrix E0 for each camera. This matrix includes a rotation matrix R ∈ R3×3 and a translation vector t ∈ R3×1.

[0092] Step 3.2 Precise calibration of the chessboard grid

[0093] A 9×6 checkerboard calibration board was placed next to the instrument being calibrated. The system acquired 20 frames of still images and used Zhang Zhengyou's algorithm to obtain the intrinsic parameter matrix K and distortion coefficients (k1,k2,p1,p2). A calibration report file was generated and its hash value, denoted as Hcamcal, was calculated using SHA-3-256.

[0094] Step 3.3 Time Synchronization Mark

[0095] To ensure frame alignment across multiple cameras, IEEE 1588PTP (Precision Time Protocol) was enabled on the edge gateway to synchronize the camera system clock with the gateway's local crystal oscillator. Finally, the frame synchronization error was verified to be less than 5ms by testing the blinking of indicator lights.

[0096] Phase IV: Establishment of Environmental Factor Baselines

[0097] The goal of this phase is to establish a set of dynamic baselines for key environmental parameters such as temperature, humidity, vibration, electromagnetic noise, air pressure, and illumination within 15 minutes before calibration begins, so that the subsequent compensation model can perform high-precision corrections for instantaneous environmental drift. The implementation steps are carried out in six sub-processes: hardware initialization → sensor node self-calibration → wireless networking → baseline acquisition → statistical integration → model hot loading.

[0098] 4.1 Hardware Initialization

[0099] (1) Node types and distribution

[0100] Each system is equipped by default with 6 types of LoRa-WAN miniature sensor nodes: temperature-humidity integrated node TH-01, triaxial acceleration node V-03, broadband electromagnetic noise node EM-02, absolute pressure node P-01, illumination node LUX-01, and backup AI camera node VIS-B (for image association).

[0101] Node placement principles: TH-01 and P-01 are installed 20cm from the sensor panel of the instrument being calibrated; V-03 is attached to the base plate; EM-02 is placed on the top of the cabinet; and LUX-01 is suspended 1m above the horizontal center of the instrument. The straight-line distance from each node to the center of the instrument being calibrated does not exceed 1m to ensure that the environmental measurements are representative.

[0102] (2) Power-on self-test

[0103] After the node is powered on, the built-in BIST (Built-InSelf-Test) sequence in the MCUBoot-ROM is executed first to quickly scan the sensor chip, power supply stability and RTC clock drift.

[0104] If the self-test result identifier contains any bit errors (e.g., temperature ADC checksum mismatch), the node will issue a warning via a local LED flashing three times and refuse to join the LoRa network.

[0105] 4.2 Sensor Node Self-Calibration

[0106] (1) Temperature-humidity integrated node

[0107] The node internally stores the multi-point correction coefficient C from the previous traceability verification. T,i C RH,i During node self-testing, the level of the on-chip 100kΩ precision reference thermistor is read and compared with the calibration curve in the EEPROM; if the deviation exceeds ±0.05℃, second-order polynomial compensation is enabled.

[0108]

[0109] (2) Triaxial vibration node

[0110] The static bias estimation b is achieved through the internal MEMS accelerometer. x ,b y ,b z Calculate the magnitude of the gravity vector, if Automatically perform zero-bias reset and log the event.

[0111] (3) Electromagnetic noise nodes

[0112] It supports 9kHz–30MHz wideband FFT. During the self-calibration phase, it performs phase-locked scanning on a 100kHz fixed signal generator to verify that the passband gain error does not exceed 0.5dB. If the limit is exceeded, the node switches to the backup gain table.

[0113] (4) Air pressure and light nodes

[0114] The pressure node reads the comparison value from the factory-sealed vacuum chamber; the illumination node uses the reference response induced by the on-chip 550nm LED to perform a slope calibration to ensure that the linearity error is <1%.

[0115] After self-calibration is complete, all correction coefficients are written to the node's Shadow RAM, and a 32-bit CRC checksum is generated and reported along with it.

[0116] 4.3 LoRa-WAN Wireless Networking

[0117] (1) Join-Procedure

[0118] The node joins the network using OTAA (Over-The-Air-Activation) and sends a Join-Request along with the previous CRC; the edge gateway LoRa Server performs fast CRC verification.

[0119] After the server returns Join-Accept, the node automatically switches to the latest Adaptive Data Rate, ensuring 250bps in open workshops and 980bps in metal-intensive workshops.

[0120] (2) Time synchronization

[0121] LoRa-WAN Class B downlink Beacon frames are transmitted every 128 seconds. The nodes use temperature-compensated crystal oscillators to align with the BeaconCounter, achieving a time base consistency of ±20ppm, and providing a unified time axis for subsequent sliding window averaging.

[0122] 4.4 Baseline Data Acquisition

[0123] (1) Sampling frequency

[0124] TH-01: once every 2 seconds; V-03: once every 1 second; EM-02: once every 5 seconds; P-01 and LUX-01: once every 10 seconds. Each sample includes the measured value and the Node-TX-Timestamp (μs), which is accompanied by the gateway RX-Timestamp when it reaches the gateway via LoRa-WAN.

[0125] (2) Edge gateway caching

[0126] The gateway writes the data to a circular buffer and categorizes it according to Node ID, forming the following time series.

[0127] {T i (t k )},{RH i (t k )},{V RMS,i (t k )},{N PSD,i (t k )},{P i (t k )},{L i (t k )};

[0128] Where i is the node index, t k The sampling time.

[0129] (3) Outlier removal

[0130] Perform a three-point moving median filter on each sequence; if a single point deviates from the mean before and after by more than 3σ, it is marked as an instantaneous anomaly and the NodeID and deviation value are recorded in the log. Subsequently, the point is replaced by linear interpolation.

[0131] 4.5 Statistical Integration

[0132] (1) Calculation of sliding window mean and variance

[0133] Select a 10-minute window (approximately 300 TH-01 sampling points) and calculate the mean values ​​of each environmental quantity. With standard deviation σ T ,σ RH ,…。 If σ T >0.3℃ or σ RH If the value exceeds the limit (e.g., >2), it is considered an unstable environment. The system will issue a yellow warning on the AR glasses and extend the baseline acquisition time by 5 minutes.

[0134] (2) Baseline Vector Generation

[0135] When the standard deviation of all environmental quantities falls within the threshold, the edge gateway calculates the baseline vector.

[0136]

[0137] And store it in the local Influx-DB.

[0138] (3) Hash digest

[0139] Pack all the original sampled sequences into a Message Pack and calculate the SHA-3-256 hash to obtain H. env_raw Then for E baseThe hash H is calculated by packing the standard deviation array together. env_bas e. Both are written into the blockchain along with the camera's calibration hash.

[0140] 4.6 Model Hot Loading and Drift Threshold Setting

[0141] (1) Loading the LSTM model

[0142] Based on the current instrument model and range being calibrated, the gateway retrieves the latest version number v from the model repository. n LSTM drift model weight file Θ vn The weights are quantized in 16-bit INT8 format. After loading, the ECDSA-P256 signature of the file is compared with the SHA-3-256 checksum; only when both pass the checksum are the data entered into the GPU inference RAM.

[0143] (2) Temperature window threshold matching

[0144] in accordance with The temperature range is determined by referring to a table to obtain the temperature sensitivity coefficient K. T And add a linear gain before the model inference layer:

[0145]

[0146] (3) Quick drift valve

[0147] To avoid excessive calculation deviations during the LSTM cold start phase, a fast drift valve is implemented:

[0148] |δ env (t)|≤C·σ Δ,spec ;

[0149] Where C = 0.5–0.7, σ Δ,spec The maximum permissible standard deviation given by the instrument specifications. If the predicted value exceeds the threshold, it is replaced with the threshold value and marked "Estimation Uncertain".

[0150] Through the six sub-processes described above, this invention completes self-calibration, stability assessment, and baseline determination for multiple environmental parameters within 15 minutes, ensuring that the subsequent LSTM drift model can be built on reliable and highly consistent input data. The precise implementation of this stage directly determines the compensation accuracy, thus affecting the overall calibration uncertainty. Actual measurements show that even in the harsh environment of a metal heat treatment workshop (temperature fluctuation ±2℃, humidity fluctuation ±6%RH), this solution can still control the calibration residual after drift compensation within 110% of the laboratory-level uncertainty, fully verifying the effectiveness and robustness of this stage of the process.

[0151] Phase 5 Calibration Data Acquisition and Error Calculation

[0152] This stage is the core measurement process that officially begins after the environmental baseline is established, camera calibration is completed, and personnel identification is confirmed. It directly determines the metrological traceability validity of the calibration results. To enable those skilled in the art to replicate this process in different testing projects, this implementation method subdivides the fifth stage into twelve steps. The following description uses a 0.1% class pressure transmitter (range 0–1 MPa) as the instrument being calibrated and a 0.02% class pressure control / measurement module as the standard; however, except for the range and unit, all steps can be seamlessly replaced with other physical quantities such as temperature, flow rate, pH, and conductivity.

[0153] 5.1 Steady-state verification of the standard instrument

[0154] (1) Steady-state start command

[0155] The edge gateway sends a steady-state preparation command to the pressure control module via the SCPI instruction INIT:PROCSTABLE. The standard's internal closed-loop feedback PID control loop locks the output pressure to the current calibration point P. set .

[0156] (2) Steady-state criterion

[0157] The gateway reads the standard's digital interface return value P at 5Hz. std (t). If in consecutive n... s = Satisfy within 30 seconds

[0158]

[0159] Then it is determined that a steady state has been reached; where ∈ s =0.01%FS.

[0160] (3) Temperature Co-check

[0161] Simultaneously, it was verified that the temperature node acquisition value T(t) satisfies: |T(t)-T(t-30s)|<ε T Otherwise, the process will be delayed until the next step is reached and a message will be displayed indicating "waiting for steady state".

[0162] 5.2 Synchronous Triggering and Time Base Locking

[0163] (1) PTP clock fine alignment

[0164] The edge gateway acts as the PTPPGrandmaster, broadcasting nanosecond-level timestamps to all camera, microphone array, and standard interface log threads; the camera's internal PLL phase-locked loop reports SYNC_OK.

[0165] (2) Single-trigger design

[0166] Once the steady-state conditions are met, the gateway sends a TRIG multicast message. After receiving the TRIG, the camera array captures an I-frame at the same PTP time t0 and immediately sends back the frame sequence number and timestamp; if the time difference is >5ms, it is recorded as ASYNC_WARN and enters the reshoot branch (see 5.9).

[0167] 5.3 Adaptive Cropping of Image ROI

[0168] (1) From panoramic to local

[0169] Each camera transmits 4K raw frames to the GPU-edge-gateway; the gateway runs YOLOv8-Det to detect the candidate bounding boxes B between the calibrated instrument's display window and the standard's display window. dev B std .

[0170] (2) Parallax correction

[0171] The bounding box is projected onto the common reference plane based on the camera extrinsic parameter matrix R,t. If the projected bounding box is offset from the center of the previous frame by more than 10px, it is considered that the lens may be vibrating and needs to be automatically adjusted or prompted for manual reset.

[0172] 5.4 Convolution-OCR Readout Extraction

[0173] (1) Feature Pyramid Enhancement

[0174] The captured ROI image is enhanced using CLAHE (Contrast-Limited Adaptive Histogram Equalization) and then fed into CRNN-OCR. The lightweight PP-OCR-v4Head is used for character sequence recognition, returning the result string_dev and confidence score p_dev; similarly, string_std and p_std are obtained.

[0175] (2) Character post-processing

[0176] Execute regular expression rules ^ \d{1,3}(\.\d{1,4})? $ Ensures the value is valid; if intermittent character reflection causes the confidence level p<0.85, call the cosine similarity backup language model LM_corr for correction and prediction.

[0177] 5.5 Consistency Arbitration of Multi-Camera Results

[0178] (1) Voting-Confidence Mechanism

[0179] The readings from the three cameras are denoted as {P} dev ,j},{P std,j}, j=1,2,3. If the results of two stations are consistent and the confidence level of each is >0.9, then take their arithmetic mean; if the results of all three stations are inconsistent, take the mode; if there is still no mode, then select the single station with the highest confidence level and note it as SINGLE_RELY.

[0180] (2) Adaptive viewpoint weighting

[0181] The system maintains a camera reliability weight table w j If a camera is judged as low confidence or SINGLE_RELY 5 times consecutively, its weight w will be adjusted. j Decrease by 0.1, down to a minimum of 0.2; the weights are used for subsequent weighted averages.

[0182]

[0183] 5.6 Digital Interface Cross-Check

[0184] (1) Redundancy comparison of standard devices

[0185] In addition to image OCR, the gateway retrieves the internal A / D readings of the standard via RS-485.

[0186] P obtained by comparing OCR std and If the difference

[0187]

[0188] The system will issue an alarm and request a repeat recording to prevent misreading due to localized glare on the display screen.

[0189] 5.7 Calculation of First Difference Error

[0190] (1) Definition

[0191] Original difference: ΔP1=P dev -P std , Unit: Pa.

[0192] (2) Range normalization

[0193] To facilitate comparison of devices with different measuring ranges, the normalized error is calculated:

[0194]

[0195] 5.8 Repeated Measures and Statistical Filtering

[0196] (1) Setting the number of loops

[0197] For each calibration point, nr = 5 repeated acquisitions were performed, with an interval of 3 seconds; the acquisition sequence is denoted as follows.

[0198] (2) Deviation Removal Rules

[0199] Calculate the average With standard deviation σ δ If it exists

[0200]

[0201] This is then identified as an outlier and marked as OUTLIER_k in the log. Outliers will be removed before the final average is calculated.

[0202]

[0203] Where n r * The number of valid samples (≥3).

[0204] 5.9 Reshoot and Rewind Mechanism

[0205] (1) Frame asynchrony compensation

[0206] If ASYNC_WARN is detected in 5.2, the gateway immediately sends a second TRIG and records RETAKE_FLAG=1 in the metadata; only the latest I-frame that was successfully synchronized is saved.

[0207] (2) OCR failure rollback

[0208] If the OCR confidence of a camera is less than 0.7 for three consecutive times, the "angle suggestion algorithm" will be activated: the possible best shooting position will be calculated based on the extrinsic parameter matrix and the operator will be prompted to adjust it via AR arrows.

[0209] 5.10 Real-time visual feedback on-site

[0210] (1) AR data floating window

[0211] AR glasses HUD display P in real time std P dev With δ1 * The values ​​are differentiated by color (green: |δ| < 0.05% FS; yellow: 0.05–0.1%; red: > 0.1%). If red is detected, the system pauses and proceeds to step 5.11, allowing the operator to select "Repeat / Confirm Fault".

[0212] (2) Voice broadcast

[0213] The Text-to-Speech kernel announces "Calibration point X successful, error Yppm" after successful sampling, improving on-site operation efficiency.

[0214] 5.11 Calibration Point Sequence Management

[0215] (1) Automatic step pressure control

[0216] The standard pressure is automatically switched using P_set_next = P_set_prev + ΔP_step, with ΔP_step set to 20% FS according to the procedure. The gateway maintains the state machine STATE = {IDLE,STABILIZING,CAPTURING,COMPLETE} and updates it in real time on the ERP work order.

[0217] (2) Two-way return test

[0218] After ascending the stairs, the computer automatically descends the same number of stairs, and the return error is compared.

[0219]

[0220] If δ return If the deviation exceeds 0.02% of FS, a mechanical lag risk is indicated, and a "backward deviation" event is reported.

[0221] 5.12 Data Packetization and Temporary Storage

[0222] (1) Sampling packet format

[0223] The data is compressed using MessagePack and written to the edge gateway SSD, and a signature sig_step5 is generated.

[0224] (2) Temporary backup and confirmation

[0225] After each calibration point is completed, the SSD will push an encrypted copy to the nearest NAS in the data center via SFTP; only after the NAS returns an ACK will the gateway switch the state machine to the next calibration point.

[0226] Through a refined twelve-step process, this invention completes a highly reliable acquisition and differential calculation of a calibration point within a chain logic of standard steady-state—synchronous triggering—multi-source consistency—repetitive statistics—instant feedback—sequence management—packet signature. Compared to the traditional hand-copying-comparison process, the dual redundancy of OCR and digital bus, outlier detection, and bidirectional backhaul detection significantly reduce random errors and systematic biases, ensuring that the normalized error residual remains better than 0.05%FS under industrial conditions. Simultaneously, the synchronization frame mechanism ensures that image, audio, and environmental data form digital evidence in the same time domain, laying a data integrity foundation for subsequent on-chain evidence storage and automatic reporting.

[0227] Phase 6 AI Abnormal Behavior Detection

[0228] The core objective of this phase is to use a multimodal artificial intelligence model to identify and assess the compliance of operational behaviors, voice content, and equipment status during the on-site calibration of testing instruments in real time. This aims to prevent violations such as operator substitution, non-compliance with procedures, access to non-standard equipment, and cheating collusion, thereby enhancing the credibility and non-repudiation of the entire calibration process.

[0229] 6.1 Video Stream Input and Pose Skeleton Extraction

[0230] (1) Synchronous video stream capture

[0231] A multi-camera array (≥3 units) begins recording at each calibration action point (e.g., turning on the standard, reading the display, connecting the pipeline), with a frame rate of 30fps. The video stream data is uniformly transcoded into RGB image sequence format and clock-synchronized via PTP (Precision Time Protocol) to ensure subsequent multimodal alignment.

[0232] (2) Human pose estimation network processing

[0233] Each frame of image is input into the YOLOv8-Pose model, and the output is the two-dimensional coordinates of 17 key points of the human body:

[0234] P i (t)=(x i (t),y i (t)), i∈{1,2,...,17},

[0235] Where P i (t) represents the pixel coordinates of the i-th body key point (such as the right shoulder or left wrist) at time t.

[0236] (3) Construction of skeleton trajectory time series

[0237] The 17 key points are concatenated into a sequence of attitude trajectory vectors over consecutive time frames:

[0238]

[0239] And construct a 10-second sliding time window data segment. Used for behavior recognition.

[0240] 6.2 Behavioral Compliance Identification and Abnormal Action Detection

[0241] (1) Loading of Action Classification Network

[0242] The above sliding window pose sequence is fed into a lightweight action recognition network (such as ST-GCN or MoveNet-LSTM hybrid model) to output the current action label A(t) and confidence level p. A(t). Model categories include, but are not limited to:

[0243] Compliance procedures: Hold the device, press the button to confirm, align the device, and observe the display;

[0244] Abnormal behavior: prolonged absence from work, obstructing camera view with body, facing others, using non-standard tools, etc.

[0245] (2) Calculation of abnormal action score

[0246] Based on action labels and confidence levels, define an abnormal action scoring function:

[0247]

[0248] If S in a continuous time window (e.g., 5 seconds) pose If (t)>0.3 more than 3 times, the "abnormal action" event will be triggered.

[0249] (3) Multi-person operation detection

[0250] The system uses the number of inter-frame skeletons The system detects whether other personnel have entered the operating area or identified unauthorized personnel operating the calibrated instrument. Once detected, a pop-up window immediately appears in the AR glasses to indicate "unauthorized personnel intervention".

[0251] 6.3 Voice Monitoring and Semantic Anomaly Recognition

[0252] (1) Microphone array and sound source localization

[0253] A microphone array consisting of four high-sensitivity directional microphones was used for beamforming to pinpoint the direction of the main speech source and filter out background noise and irrelevant conversations. A 10-second audio segment was extracted. It is fed into the speech recognition system.

[0254] (2) Speech transcription and context classification

[0255] ASR (Automatic Speech Recognition) processing is performed using either Whisper or Conformer-RNN models, outputting the speech-to-text T. voice (t).

[0256] The transcribed text is input into a natural language classifier (based on BERT or ERNIE model) for context label determination:

[0257] Compliance statements: reading pressure value, operating instructions, and execution confirmation;

[0258] Abnormal statements include: discussing non-standard operations, private coordination, irrelevant chatter, and refusing supervision.

[0259] (3) Semantic anomaly scoring

[0260] Let the current context label be L voice (t), whose corresponding anomaly score is defined as:

[0261]

[0262] If the average value is within a 30-second window Mark "voice abnormality".

[0263] 6.4 Device Identification and Detection of Unauthorized Tool Access

[0264] (1) Object recognition and tag comparison

[0265] YOLOv8-Object-Detection was used to detect tools, connectors, cables, and other equipment in the field video frames; and the legitimate equipment library was imported from the ERP work orders.

[0266] For each frame recognition result D t ={tool i Perform a difference comparison with the set of legitimate devices:

[0267]

[0268] If an illegal device (such as a non-standard pipe fitting or a non-branded standard device) is detected, record an image screenshot and mark it as "tool abnormal".

[0269] (2) Equipment anomaly score calculation

[0270] Define device anomaly score:

[0271]

[0272] That is, the percentage of illegal tools, if 3 consecutive frames S device >0.2, triggering an alarm.

[0273] 6.5 Multimodal Integrated Anomaly Scoring and Alarm Mechanism

[0274] (1) Anomaly scoring fusion model

[0275] Construct a multimodal anomaly scoring function:

[0276] S abn (t)=α·S pose (t)+β·S speech (t)+γ·S device (t),

[0277] The recommended weights are: α = 0.4, β = 0.3, γ = 0.3, which can be adjusted according to the actual industry sensitivity.

[0278] (2) Window averaging and event triggering

[0279] Calculate the average anomaly score within a 15-second sliding window. If the following conditions are met:

[0280]

[0281] If this round of operations is deemed to have a violation risk, the following actions will be triggered:

[0282] a) A red box pops up in the AR glasses with the message "Operation error, please check again";

[0283] b) System voice prompts;

[0284] c) The data acquisition and error recording process will be automatically paused, requiring on-site personnel to re-execute the process;

[0285] Pack the anomaly-related video, audio clips, and device images into the chain and prepare to write the block.

[0286] 6.6 Log Generation and On-Chain Marking

[0287] (1) Log format

[0288] The system generates an AI anomaly analysis log for each calibration point.

[0289] (2) Hashing and Chaining Preparation

[0290] Package the logs to generate a hash H ai_log =SHA-3-256(log), which is subsequently written into the blockchain along with the image hash and calibration error value to ensure that the anomaly handling results are non-repudiable and have third-party auditing capabilities.

[0291] This phase combines three sensing methods—pose recognition, voice analysis, and device image recognition—and constructs a fusion scoring system to ensure comprehensive, real-time, and intelligent monitoring of operational behavior during instrument calibration. Its significant technical effects include:

[0292] a) Achieve millisecond-level response to events such as unauthorized personnel, incorrect procedures, and verbal cheating;

[0293] b) Enhance the compliance, credibility, and legal validity of the calibration process;

[0294] c) Automatically generate AI-based judgment logs to provide valid evidence for digital forensics and on-chain evidence storage.

[0295] The multimodal AI anomaly recognition model of this invention can be continuously iterated and updated to adapt to more industry standard operating procedures, and has good versatility and technical scalability.

[0296] Stage 7 Real-time Environmental Drift Compensation

[0297] This phase aims to address errors in testing instruments caused by fluctuations in factors such as temperature, humidity, air pressure, vibration, and electromagnetic interference in the field environment, resulting in zero-point shifts, nonlinear responses, or hysteresis. The goal is to dynamically correct the original error data, improving the objectivity and accuracy of the calibration data. This invention proposes a real-time environmental drift compensation method based on Long Short-Term Memory (LSTM) neural network modeling. It combines real-time acquired multi-dimensional environmental parameters, performs rapid inference on an edge computing gateway, and fuses the results with the current error data to output a compensation value. The implementation process unfolds in the following seven steps:

[0298] 7.1 Environmental Data Stream Acquisition and Timing Processing

[0299] 7.1.1: Multi-source synchronous acquisition

[0300] All environmental sensor nodes (temperature, humidity, air pressure, vibration, electromagnetic) deployed around the instrument being calibrated are connected to the edge gateway via LoRa or Wi-Fi, with the following sampling frequencies:

[0301] Temperature and humidity: every 2 seconds;

[0302] Air pressure and light intensity: every 5 seconds;

[0303] Triaxial vibration: RMS calculation every 1 second;

[0304] Electromagnetic noise: spectral average every 10 seconds.

[0305] 7.1.2: Standardized Definition of Environment Vector

[0306] Define a multidimensional environment state vector at each time point t:

[0307] E(t)=[T(t),RH(t),P(t),V(t),N(t)] T ,

[0308] in:

[0309] T(t): Temperature (°C);

[0310] RH(t): Relative humidity (%);

[0311] P(t): air pressure (hPa);

[0312] V(t): Triaxial composite vibration RMS (m / s) 2 );

[0313] N(t): Mean electromagnetic noise (dBμV).

[0314] All environmental variables need to be linearly normalized before being input into the LSTM:

[0315]

[0316] Where μ i ,σ i The empirical mean and standard deviation for each environmental dimension (can be set based on the sliding window statistics of the most recent 30 minutes or historical values ​​of the laboratory).

[0317] 7.2 LSTM Drift Model Architecture and Loading

[0318] 7.2.1: Model Structure Definition

[0319] This invention uses a two-layer stacked LSTM network structure to construct an environment drift prediction model:

[0320] Input layer dimension: 5 (number of environment variables);

[0321] First LSTM layer: Hidden layer size is 64;

[0322] The second LSTM layer has a hidden layer size of 32;

[0323] Fully connected layer: Output predicted drift δ env (The unit is the same as the quantity being calibrated, such as Pa or °C);

[0324] Time window: Construct a time series input using 60 seconds of historical data, with a step size of 1 second.

[0325] 7.2.2: Model Training and Hot Update Mechanism

[0326] During the offline phase, the model is trained under supervision using a laboratory-calibrated environment and error pairs, employing the MSE loss function. After online deployment, the edge gateway receives a new weight packet (.pth format) every quarter for hot replacement. Before updating the weights, integrity verification (SHA-256 hash comparison + ECDSA signature verification) must be performed.

[0327] 7.3 Construction and Input Preparation of Historical Time Windows

[0328] 7.3.1: Historical Sequence Caching

[0329] The edge gateway maintains a sequence of one-dimensional environment vectors sampled every second for the past 60 seconds:

[0330]

[0331] A 60×560×5 input tensor is constructed and fed into the LSTM model for inference.

[0332] 7.4 Drift Error Prediction and Output

[0333] 7.4.1: LSTM Inference Process

[0334] The input tensor E(t) is forward propagated through the model, and the predicted environment drift error at the current time step is output:

[0335]

[0336] in:

[0337] A trained LSTM model;

[0338] δ env (t): The unit is the same as the physical quantity being checked, such as pressure (Pa), temperature (°C), flow rate (L / min), etc.

[0339] 7.4.2: Error Stability Judgment

[0340] To avoid model cold start jitter, a maximum drift threshold δ is set. max ,like:

[0341] |δ env (t)|≤δ max =0.05%·FS,

[0342] If this value is exceeded, it is judged as an abnormal prediction, and the amplitude limit value is used instead, and DRIFT_CLIP_FLAG=1 is marked.

[0343] 7.5 Error Compensation and Real-time Update

[0344] 7.5.1: Definition of Original Error

[0345] The raw differential error (uncompensated) obtained during calibration data acquisition is as follows:

[0346] △ raw (t)=V dev (t)-V std (t).

[0347] 7.5.2: Calculation of error after compensation

[0348] The corrected difference error is obtained by compensating for the drift predicted by the LSTM model.

[0349] △ corr (t)=△ raw (t)-δ env (t),

[0350] The compensated error will be used as the official error value for subsequent on-chain evidence storage and report generation.

[0351] 7.6 Real-time Drift Status Indication and Alarm Mechanism

[0352] 7.6.1: Confidence Score Mechanism

[0353] Calculate the confidence score C(t) for current drift prediction, which is defined as follows:

[0354]

[0355] If C(t) > 0.9, it is displayed as a green status (model is reliable);

[0356] If 0.7 < C(t) ≤ 0.9, it is displayed as yellow (medium confidence);

[0357] If C(t) ≤ 0.7, it is displayed as red, and the "drift uncertainty alarm" is triggered.

[0358] 7.6.2: AR Real-time Display

[0359] In wearable AR glasses, the system superimposes the real-time graphical representation of the currently predicted drift value, confidence level, and compensated error curve on the operation interface, and provides voice prompts such as "environmental compensation started / drift is too high".

[0360] 7.7 Compensation Data Recording and Traceable Storage

[0361] 7.7.1: Recording of Original Values and Compensated Values

[0362] For each sampling point, the system records: the original error Δ raw (t); the environmental drift value δ env (t); the compensated error Δ corr (t); the drift confidence score C(t).

[0363] 7.7.2: Data Packaging and Signature

[0364] Pack the above data and the sampling timestamp into a JSON document, generate a SHA-3-256 hash digest, and perform ECDSA-P256 signature with the gateway private key, so as to prepare it as a part of the on-chain evidence storage data structure.

[0365] The present invention constructs an LSTM-based environmental drift error prediction model to realize real-time impact modeling and quantification of measurement accuracy caused by complex dynamic on-site environments, and accurately applies drift compensation to each calibration sampling point through an online inference mechanism, effectively improving the comparability and confidence of data. The system design has the following advantages:

[0366] a) Highly adaptable: The model can be transferred for training and is suitable for different physical quantities and devices;

[0367] b) High real-time performance: GPU inference latency <100ms;

[0368] c) Visual and intuitive: With real-time AR prompts, operators can proactively perceive drift risks;

[0369] d) Traceability: All compensation data is signed and uploaded to the blockchain to ensure the credibility of the entire calibration closed loop.

[0370] This stage is the core of ensuring the accuracy of the entire intelligent calibration system, and it is also one of the key innovations that distinguishes this invention from traditional methods.

[0371] Phase 8 On-chain Evidence Storage

[0372] The I-frame image file, standard value, instrument value, environment vector, and anomaly score of this calibration point are concatenated into a raw byte stream, and then a hash H is generated using SHA-3-256. pt Using the ECDSA-P256 private key in the edge gateway TPM chip, for H pt Perform digital signature σ TPM Then, a blockchain transaction is constructed:

[0373] Tx = <H pt ,σ TPM ,timestamp,nodeID〉,

[0374] And broadcast to the consortium blockchain. Main chain confirmation time ≤ 5s. Transaction index hash H tx This will be cited in the report.

[0375] Phase 9 AR Real-Time Collaboration

[0376] If on-site personnel require assistance, they can click the AR menu to initiate a WebRTC session and invite metrology experts to join. Experts can view a panoramic video stitched together from multiple cameras in their browser and draw lines and circles anywhere on the screen; all markings are projected back into the AR field of view in real time. The system automatically generates 3D arrows or text bubbles based on expert markings and error feedback. For example, "Rotate the adjustment knob to 1.2 bar" appears next to the knob; "Change the calibration point to 150℃" appears above the temperature control knob. This entire interactive dialogue is recorded via WebSocket and a signature hash is generated.

[0377] Phase 10 report generated

[0378] The edge gateway uses Matplotlib-CPU to plot the curve of ΔV2(t) over time, overlaid with the environment vector, and outputs a PNG file. The first frame image is cropped using a function and then included as an illustration in the report. The template engine is used to write the following fields into XML:

[0379] Equipment Information: Model, Serial Number, Work Order Number

[0380] Calibration parameters: List of reference points and corresponding ΔV2

[0381] Environmental baseline and instantaneous values

[0382] List and timestamps of AI anomalies

[0383] On-chain transaction hash H tx List

[0384] Calibrate the credibility index (CI).

[0385] in:

[0386]

[0387] Parameter k is the steepness coefficient, S abn,avg σ represents the average anomaly score. spec This represents the maximum permissible deviation specified in the instrument specifications.

[0388] The generated PDF report is signed with an X.509 digital certificate and pushed to the enterprise ERP. After receiving the confirmation, the ERP system sets the status to "Calibration Completed" and unlocks the on-site equipment.

[0389] Phase 11 Data Archiving and Maintenance

[0390] After the edge gateway completes its task, it copies the original I-frame, environmental data, OCR results, and on-chain transaction list to the RAID6-NAS, and then transfers them to the offline tape repository after being signed by the on-duty engineer. The system performs cluster analysis on the collected environment-error pairs quarterly and retrains the LSTM drift model. Once the new model is validated, it is upgraded using a rolling upgrade mechanism; the old version is automatically archived and retained for one year. In accordance with ISO / IEC 27040 requirements, critical information logs are retained for at least ten years. A blockchain transaction reconciliation is performed monthly, randomly selecting 5% of calibration points to compare the local I-frame hash with the on-chain hash to verify consistency. If discrepancies are found, a security investigation process must be initiated.

[0391] Test case

[0392] I. Test Subjects and Experimental Setup

[0393] The device being calibrated is a WZ-PT800 intelligent pressure transmitter with a range of 0–1 MPa and an accuracy class of 0.1% FS.

[0394] Standard instrument: Fluke 6270A high-precision pressure controller, range 0–1MPa, accuracy 0.02%FS.

[0395] Calibration point settings: 0%, 20%, 40%, 60%, 80%, 100%FS (6 points total), plus two cycles of ascent and descent, for a total of 12 points.

[0396] Site conditions: Steel structure workshop (non-air-conditioned) in summer, temperature fluctuation ±2.5℃, humidity fluctuation ±8%RH, intermittent fan and vibration interference.

[0397] The system deployment and configuration are as follows:

[0398] camera array Three 4K industrial cameras, 120° field of view, hardware synchronization Environmental sensing nodes Temperature and humidity, electromagnetic noise, triaxial vibration, air pressure, and light. AR devices Rokid Glass 2: Real-time feedback and expert collaboration AI Engine YOLOv8-Pose, Whisper, and a mini BERT (edge ​​reasoning) Blockchain Platform Hyperledger Fabric + IPFS (Consortium Blockchain Deployment)

[0399] II. Experimental Procedure

[0400] 1. Setting up the experimental control group

[0401] Group A (Control Group) Traditional manual calibration is used, with values ​​read manually and records kept on paper; there is no compensation or video recording. Group B (This invention) Full system enabled: camera OCR, AI behavior recognition, environmental compensation, and on-chain evidence storage.

[0402] Each group performs on-site calibration of 5 devices, comparing a total of 60 calibration points.

[0403] 2. Calibration Implementation Process

[0404] The control group's data was manually recorded by senior inspectors using standard forms. The invention group's data was automatically acquired via camera; the system recorded real-time images and errors, AI monitored behavioral compliance in real-time, and a drift model based on LSTM predicted error shifts in real-time. All Group B data was automatically signed and written to the consortium blockchain, connecting in real-time with the ERP system.

[0405] III. Key Experimental Data and Results

[0406] 1. Comparison of error statistics

[0407] 20%FS 0.2005 0.2002 -0.15％ 40%FS 0.4007 0.4001 -0.30％ 60%FS 0.5996 0.6000 +0.07％ 80%FS 0.7992 0.8001 +0.11％ 100%FS 1.0008 1.0002 -0.06％

[0408] Conclusion: The average error deviation of the present invention group after compensation is less than ±0.1%FS, while the traditional group has a maximum FS offset of 0.3%, indicating that the environmental drift compensation model effectively reduces the interference of temperature and humidity.

[0409] 2. Anomaly detection accuracy statistics

[0410]

[0411]

[0412] Conclusion: The AI ​​model has a high detection capability for key violations and can promptly block potential operational risks.

[0413] 3. Data credibility and auditability

[0414] Error source traceability × √(Image + Environment + Error Chain) Process compliance record × √(AI Behavior Log) Abnormal operation alarm mechanism × √(AR pop-up + sound prompt) Digital Certificate + Report Signature × √(Blockchain hash + signature) Report Verifiability × √(Includes IPFS file pointers)

[0415] IV. Results Analysis

[0416] This example demonstrates that the present invention achieves the following technical effects:

[0417] 1. Enhanced Data Trustworthiness: Through on-chain writing, calibration error data, images, and compensation values ​​are tamper-proof and traceable;

[0418] 2. Improved error control capability: In complex and fluctuating environments, the error control capability of this invention is about 200-300% better than that of traditional methods;

[0419] 3. Transparent Operation Process: The overall accuracy rate of abnormal operation identification exceeds 85%, significantly enhancing the ability to supervise processes;

[0420] 4. Automated delivery of results: Report generation time has been reduced from 30 minutes manually to 2 minutes automatically completed by the system;

[0421] 5. System robustness verification: Under ambient temperature fluctuations of ±2.5℃ and humidity of ±8%RH, the system operates stably and the compensation effect is obvious.

[0422] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A camera-assisted intelligent on-site calibration method for detection instruments, characterized in that, It includes the following sequential steps: 1) Confirmation of the legality of the venue and personnel: The coordinates of the calibration site are obtained through GPS and / or Beidou dual-mode positioning and compared with the coordinates of the pre-stored work orders; and the identity of the calibration personnel is uniquely verified by dual facial and voiceprint recognition. 2) Multi-camera panoramic acquisition and calibration: No fewer than three high-resolution cameras with preset positions are arranged around the instrument being calibrated. The internal and external parameters of the cameras are obtained by Zhang Zhengyou's calibration method and uploaded to the cloud. Each camera is started synchronously to generate a multi-view video stream with timestamps. 3) Visual-data fusion calibration: The convolutional neural network-OCR combined model is called to automatically identify the standard instrument reading and the value displayed by the instrument being calibrated in the video stream, and write them into the local cache according to a one-to-one correspondence; the real-time identified value and the standard calibration point are differentially calculated to obtain the first calibration error; 4) Comprehensive judgment of abnormal actions and environment: The human posture estimation algorithm is used to detect whether the action sequence of the calibration personnel deviates from the preset working conditions; temperature, humidity, light and vibration data are collected simultaneously, and environmental drift compensation is performed on the first calibration error to obtain the second calibration error; 5) Blockchain-based evidence storage and AR prompts: After performing SHA-3-256 calculation on the image hash, environmental data and second calibration error generated in steps 2)-4), the data is written into the consortium blockchain and digitally signed by elliptic curve P-256. If the error or abnormal score exceeds the threshold, the correction instruction is immediately displayed on the wearable AR terminal, and remote experts are allowed to annotate in the AR screen in real time via WebRTC. 6) Intelligent report generation: Based on the differential calculation results, the first frame screenshot of the camera evidence image, and the environmental drift compensation model, a structured calibration report with predictive maintenance suggestions is automatically generated, and on-chain transaction hashes are attached.

2. The method according to claim 1, characterized in that, Step 2) specifically includes the following steps: a) Arrange at least three 4K industrial cameras around the instrument being calibrated at three or more non-collinear points, so that their optical axes intersect to form a minimum triangular mesh; b) Perform Zhang Zhengyou calibration on each camera sequentially using a 9×6 checkerboard calibration board to obtain the intrinsic parameter matrix K, distortion coefficients (k1,k2,p1,p2) and extrinsic parameter matrix [R|t]. Calculate the SHA-3-256 hash of the parameter file and upload it to the cloud for storage. c) Synchronize the time base of each camera system using the IEEE-1588 precision clock to ensure that the difference between the trigger acquisition times is no more than 5ms, and write a nanosecond-level timestamp into the frame header; d) Each camera starts recording simultaneously under the same trigger command, outputting a multi-view video stream with a unified timestamp, providing spatiotemporally consistent image data for subsequent vision-data fusion calibration.

3. The method according to claim 1, characterized in that, Step 3) specifically includes the following steps: a) At least three cameras synchronized with an IEEE-1588 clock capture timestamped I-frames at the same trigger time; b) Perform contrast-limited adaptive histogram equalization, bilateral filtering denoising, and YOLO-based ROI auto-cropping sequentially on the I-frame; c) The cropped image is fed into the ResNet-BiLSTM-CTC convolution-OCR combined network to obtain the standard instrument reading and the instrument under calibration and their confidence levels, respectively. d) Write the two types of readings into the local cache one by one according to the timestamp, and fuse the results from multiple cameras through a confidence-weighted voting mechanism; e) Perform differential calculation between the fused reading and the preset standard calibration point to generate the first calibration error and archive it with frame hash and confidence label.

4. The method according to claim 1, characterized in that, Step 5) specifically includes the following steps: a) Integrate the key I-frame hash, environmental baseline hash and second calibration error collected by multiple cameras into evidence storage metadata, calculate the SHA-3-256 digest and generate a digital signature using the ECDSA-P256 private key; b) Adopting a "consortium blockchain + IPFS" layered strategy, the metadata and the IPFS content address of the original video and sensor files are packaged together into a transaction structure and written into the blockchain through BFT-DPoS consensus, with a block time of ≤5 seconds. c) When the compensation error or the overall anomaly score is higher than the preset threshold of 0.1%FS or 0.4 points, a red correction instruction will immediately pop up on the wearable AR terminal, and remote experts will be allowed to draw vector annotations in the AR view in real time through the WebRTC channel. The on-site operator can continue the subsequent calibration process only after completing the correction according to the annotations.

5. The method according to claim 1, characterized in that, when the error exceeds the threshold, the system automatically sends a freeze command to the enterprise ERP to lock the current calibration work order and prevent the instrument from being put into use; the structured calibration report includes a calibration reliability index, which is obtained by weighting the abnormal action score, the environmental steady-state score and the error residual score.

6. A field intelligent calibration system for a testing instrument used to implement the method according to any one of claims 1-5, comprising: Positioning and identity verification unit, including GPS / BeiDou module and face-voiceprint integrated recognition terminal; The camera array unit consists of at least three 4K cameras and has hardware synchronization and self-correction functions; The vision-data fusion engine, deployed on a GPU edge gateway, is used to perform OCR recognition, differential calculation, and pose analysis. The environmental perception and compensation module includes temperature and humidity sensors, light and vibration sensors, and an LSTM-based drift compensation submodule. The blockchain client module is used to write image hashes, compensation data, and calibration errors, and to execute digital signatures. The AR collaboration module supports on-site visual guidance and real-time annotation by remote experts. The intelligent report generation and ERP interface module is used to output structured reports and interact with enterprise information systems.

7. The system according to claim 6, characterized in that, The camera array unit supports H.265 encoding and RTSP streaming, and has low-light imaging capability in scenes with a resolution of ≤10 lux; the vision-data fusion engine supports 60 FPS real-time inference with a latency of ≤200 ms; the environmental perception and compensation module has a temperature measurement accuracy of ±0.1℃ and a relative humidity measurement accuracy of ±1%RH; the blockchain client module adopts the BFT-DPoS consensus algorithm with an average block time of ≤5s; the AR collaboration module is compatible with the OpenXR standard and can run on mainstream AR glasses and smartphones.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when executed by a processor, the computer program causes the processor to perform the steps of the method according to any one of claims 1-5.

9. A computer program product comprising program instructions that, when executed on an electronic device, cause the electronic device to perform the method of any one of claims 1-5.

Images