Industrial meter reliable reading method based on visual perception and large model collaborative reasoning

Abstract

The application discloses an industrial meter trusted reading method based on visual perception and large model collaborative reasoning, comprising: processing the to-be-read image by using an industrial meter visual reading system to obtain a visual perception information package; integrating non-visual environment information and prior knowledge to construct cognitive context information; converting the visual perception information package and the cognitive context information into multi-modal prompt words and inputting the multi-modal prompt words into a large language model; driving the large language model to call a special function tool set to perform reasoning and verification, outputting a reasoning reading and a correction confidence; generating a final reading result through a dynamic decision strategy, and supporting performance evaluation and feedback optimization; the method can also be used for industrial equipment component reading and a method executed by using an intelligent agent architecture. Through deep coupling of visual perception and cognitive reasoning, the application effectively solves the misreading problem of industrial meters under complex working conditions, significantly improves the reliability and explainability of the reading result, and provides technical support for industrial intelligent operation and maintenance.

Description

Technical Field

[0001] This application relates to the fields of computer vision and artificial intelligence, and more specifically, to a reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning. Background Technology

[0002] In industrial inspections, equipment maintenance, and process monitoring, pressure gauges, temperature gauges, level gauges, and other industrial meters play a crucial role in indicating status. Existing automatic reading solutions for industrial meters typically rely on visual methods such as target detection, image segmentation, optical character recognition, image correction, and geometric calculations to extract visual elements such as the dial area, pointer position, and scale information, and output corresponding preliminary readings.

[0003] However, in actual industrial settings, image acquisition is easily affected by factors such as abnormal lighting, occlusion, and dirt. Preliminary readings obtained through purely visual methods may be inaccurate, making it difficult to judge their rationality and physical accuracy based solely on visual results. Furthermore, existing methods typically lack the unified utilization of non-visual information such as imaging location markers, image acquisition timestamps, environmental conditions, instrument range, units, normal operating range, historical trend data, and logical association rules. This makes it difficult to perform logical consistency checks, multi-source information conflict analysis, and reliability management on the reading results.

[0004] Similar issues exist in tasks involving the status identification of industrial equipment components such as switches, indicator lights, and valves. Relying solely on visual status judgments obtained from a single frame image makes it difficult to further determine whether the results conform to industrial scenario knowledge, operational procedures, and logical relationships.

[0005] In addition, after actual deployment, industrial meter vision recognition systems still suffer from the lack of online evaluation, difficult sample collection, and continuous optimization mechanisms, which hinders the system's continuous evolution in long-tail and challenging scenarios.

[0006] Therefore, the existing technology has defects and urgently needs improvement. Summary of the Invention

[0007] In view of the above problems, the purpose of this invention is to provide a reliable reading method for industrial meters based on visual perception and collaborative reasoning of large language models, so as to improve the reliability and interpretability of industrial meter reading results in complex industrial scenarios; at the same time, it provides solutions for industrial equipment component status reading, industrial meter reading system performance evaluation and evolution, and reliable reading of industrial meters based on intelligent agent architecture.

[0008] The first aspect of this invention provides a reliable reading method for industrial meters based on visual perception and large-model collaborative reasoning, characterized in that it includes: Acquire the image of the meter to be read; The images of the meters to be read are analyzed to extract visual perception information, which includes the visual perception information package of the industrial meters, the preliminary reading, and the original confidence level. Integrate non-visual environmental information and prior knowledge related to the meter to be read as cognitive context information; convert the visual perception information package and cognitive context information into multimodal prompt words containing task instructions in the industrial scenario through semantic encoding and structured mapping; The multimodal prompt words are input into the preset industrial meter reliable reading model. Based on the task instructions, cognitive reasoning and tool enhancement verification based on industrial domain knowledge are performed, and the inferred reading and corrected confidence are output. Based on the analysis of preliminary readings, original confidence levels, inferred readings, and corrected confidence levels, a dynamic decision-making strategy is implemented to generate reliable industrial meter reading data. An analysis report is generated based on the reading data from trusted industrial meters, and the trusted industrial meter reading data and analysis report are sent to a preset terminal.

[0009] In this solution, the image of the meter to be read is analyzed to extract visual perception information, including: Based on a pre-built industrial meter visual recognition system, the positioning information of the industrial meter body and dial area is obtained through a deep learning object detection model; The precise geometric features of the scale point sequence and pointer are obtained by using a preset precise geometric feature extraction method. Obtain scale value information through an optical character recognition model; A corrected dial image is obtained using an image correction algorithm; Preliminary readings and initial confidence levels are generated using purely visual methods; The visual perception information is obtained by integrating the positioning information of the industrial meter body and dial area, the scale point sequence and the precise geometric features of the pointer, the scale value information, the dial image, the preliminary reading and the original confidence level.

[0010] In this solution, cognitive reasoning and tool enhancement verification based on industry domain knowledge are performed according to task instructions, and the inference readings and corrected confidence scores are output, including: Perform a visual element logical consistency check on the multimodal cue words to obtain logical consistency check data; Information conflict analysis was performed on the visual perception information package and cognitive context information to obtain conflict analysis and environmental impact assessment data. Based on logical consistency check data and conflict analysis and environmental impact assessment data, the preliminary readings are verified and revised to generate inferred readings and revised confidence levels.

[0011] In this solution, visual element logical consistency checks are performed on multimodal cue words to obtain logical consistency check data, including: Based on the geometric features and element topology relationships in the industrial meter template knowledge base, a visual element logical consistency check is performed to verify the spatial geometric consistency between the pointer geometric features and the scale distribution pattern in the visual perception information package, and to verify the semantic correspondence between the identified scale values ​​and scale points. Based on the spatial geometric consistency verification results and semantic correspondence, logical consistency check data is generated.

[0012] In this scheme, information conflict analysis is performed on the visual perception information package and cognitive context information to obtain conflict analysis and environmental impact assessment data, including: By comparing visual perception information packets with cognitive context information, the differences between the two are determined; Acquire environmental metadata integrated into cognitive context information as environmental factors; Based on environmental factors, a predefined environmental interference analysis tool is invoked to assess the potential interference of environmental factors on the extraction of elements in the visual perception information package, and to identify the key visual features that cause interference. Based on differences, potential disturbances, and key visual features, conflict analysis and environmental impact assessment data are generated.

[0013] In this scheme, based on logical consistency check data and conflict analysis and environmental impact assessment data, preliminary readings are verified and revised to generate inferred readings and revised confidence levels, including: By combining logical consistency check data with conflict analysis and environmental assessment data, the rationality and physical accuracy of the initial readings are thoroughly verified. If inconsistencies, interference, and potential errors are detected, the preset industrial meter reliable reading model, based on its reasoning ability and tool calls, obtains the model's inference reading and its corrected confidence level through preset correction and calculation instructions.

[0014] This solution analyzes preliminary readings, initial confidence levels, inferred readings, and revised confidence levels, executes a dynamic decision-making strategy, and generates reliable industrial meter reading data, including: If the deviation between the initial reading and the inferred reading is within the preset allowable range, and both the original confidence level and the modified confidence level are higher than the high confidence threshold, then the initial reading is output and labeled as a high-confidence direct reading. If the deviation between the initial reading and the inferred reading is within the preset allowable range, and the original confidence level is lower than the high confidence threshold, then the inferred reading is taken as the final result and labeled with a model correction confirmation. At the same time, an explanation of the inference process of the preset industrial meter reliable reading model is provided. If the deviation between the initial reading and the inferred reading exceeds the preset allowable range, or if the preset industrial meter reliable reading model detects an unsolvable serious conflict during the inference process, a final decision containing detailed conflict analysis, cause inference, and potential risk warning will be output, and a label indicating that manual review is required will be added.

[0015] This solution generates an analysis report based on trusted industrial meter reading data, including: The analysis report is at least one of the following: a single meter reading analysis report and an industrial equipment operating status correlation analysis report; Obtain the final confirmed readings, confidence levels, and labels from trusted industrial meter readings; Intermediate data from visual element logical consistency checks, conflict analysis, and environmental impact assessment processes are used as inference data for model correction. Based on the evaluation results of the environmental interference analysis tool, a qualitative analysis is conducted on the specific visual interference factors that lead to questionable readings, and the visual interference analysis results are obtained. The deviation between the initial reading and the inferred reading is calculated. If the deviation exceeds the preset allowable range, the final decision data containing detailed conflict analysis, cause inference and potential risk warning is generated. Combined with the preset standard operating procedures and knowledge base, preliminary handling suggestions and contingency plans are generated. The risk assessment and handling suggestions are summarized based on the risk level assessment results. The final confirmed readings and confidence levels, along with the labels, inference data from model corrections, visual interference analysis results, risk assessments, and mitigation recommendations, are integrated to obtain a single meter reading analysis report. A list of equipment status is obtained by listing the current operating status of all identified industrial equipment within the inspection range; Based on the reading results from multiple devices, multi-device correlation analysis is performed on the detected anomalies to infer the potential causes of the anomalies, and an anomaly summary and logical analysis report is generated by combining the detected anomalies. The deviation between the initial reading and the inferred reading is calculated. If the deviation exceeds the preset allowable range, the final decision data containing detailed conflict analysis, cause inference and potential risk warning is generated. Combined with the preset standard operating procedures and knowledge base, preliminary handling suggestions and contingency plans are generated. The risk assessment and handling suggestions are summarized based on the risk level assessment results. Summarize and optimize key algorithms, model parameters, data augmentation strategies, and incremental training sample selection criteria for the system, and provide feedback suggestions for system optimization. By integrating the equipment status list, anomaly summary and logical analysis report, risk assessment and handling suggestions, and system optimization feedback suggestions, an industrial equipment operation status correlation analysis report is obtained.

[0016] This plan also includes: Based on preset conditions, a sample evaluation is conducted on the readings generated by the industrial meter vision recognition system and its correction process, and an online evaluation report of the system is generated. Based on the online evaluation report, the system automatically selects image samples that are difficult to identify because the deviation between the initial reading and the inferred reading exceeds the preset allowable range, or because the preset industrial meter reliable reading model detects an unsolvable serious conflict during the inference process, and the image samples that have been manually verified or whose corrected confidence level corresponding to the reading after correction by the preset industrial meter reliable reading model is higher than the preset high confidence threshold. Difficult example samples are added to the training set to update and optimize the industrial meter vision recognition system.

[0017] A second aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a program for a reliable industrial meter reading method based on visual perception and large-model collaborative reasoning, wherein when the program for the reliable industrial meter reading method based on visual perception and large-model collaborative reasoning is executed by a processor, it implements the steps of the reliable industrial meter reading method based on visual perception and large-model collaborative reasoning as described above.

[0018] This invention discloses a reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning. The method includes: processing the image to be read using an industrial meter visual reading system to obtain a visual perception information package; integrating non-visual environmental information and prior knowledge to construct cognitive context information; converting the visual perception information package and cognitive context information into multimodal prompt words and inputting them into a large language model; driving the model to call a dedicated functional toolset to perform reasoning and verification, outputting the inferred reading and corrected confidence level; generating the final reading result through a dynamic decision-making strategy, and supporting performance evaluation and feedback optimization. This method can also be used for reading industrial equipment components and executing methods using an intelligent agent architecture. By deeply coupling visual perception and cognitive reasoning, this invention effectively solves the problem of misreading industrial meters under complex operating conditions, significantly improving the reliability and interpretability of the reading results, and providing technical support for intelligent industrial operation and maintenance. Attached Figure Description

[0019] Figure 1 The flowchart of the reliable reading method for industrial meters based on visual perception and large model collaborative reasoning of the present invention is shown. Figure 2 A flowchart of an inference reading and corrected confidence acquisition method according to the present invention is shown; Figure 3 This invention illustrates a flowchart of a process for generating reliable industrial meter reading data by executing a dynamic decision-making strategy. Figure 4 The flowchart illustrates a method for performance evaluation and adaptive evolution of an industrial meter reading system according to the present invention. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] Figure 1 The flowchart of the reliable reading method for industrial meters based on visual perception and large model collaborative reasoning of the present invention is shown.

[0023] like Figure 1 As shown, this invention discloses a reliable reading method for industrial meters based on visual perception and large-model collaborative reasoning, including: S101, Obtain the image of the meter to be read; S102, Analyze the image of the meter to be read and extract visual perception information; the visual perception information includes the visual perception information package of the industrial meter, the preliminary reading and the original confidence level. S103, integrate non-visual environmental information and prior knowledge related to the meter to be read as cognitive context information; S104, through semantic encoding and structured mapping, transforms visual perception information packets and cognitive context information into multimodal prompt words containing task instructions for industrial scenarios; S105: Input the multimodal prompt words into the preset industrial meter reliable reading model, perform cognitive reasoning and tool enhancement verification based on industrial domain knowledge according to the task instructions, and output the reasoned reading and corrected confidence level. S106: Analyze the preliminary readings, original confidence levels, inferred readings, and corrected confidence levels, execute dynamic decision-making strategies, and generate reliable industrial meter reading data. S107: Generate an analysis report based on the trusted industrial meter reading data, and send the trusted industrial meter reading data and analysis report to a preset terminal.

[0024] According to an embodiment of the present invention, a pre-built industrial meter reading system is used to process the image of the meter to be read, extract and generate structured visual perception information; the visual perception information includes the industrial meter visual perception information package, as well as the preliminary reading and original confidence level generated by a pure vision method.

[0025] The visual perception information package for industrial meters includes at least: dial area positioning information, pointer geometry (e.g., coordinates of pointer rotation center and pointer endpoint), all identifiable scale point sequences, identified scale value information, and calibrated dial image.

[0026] It integrates cognitive context information, non-visual environmental information and prior knowledge related to the meter, including at least: imaging location identifier, image acquisition timestamp, current environmental conditions (e.g., weather conditions, artificial light source angle), and process parameters of the meter to be read (e.g., instrument range, unit, normal operating range, historical trend data) and logical association rules provided by the industrial meter template knowledge base.

[0027] Visual perception information and cognitive context information are converted into multimodal prompt words containing clear industrial scenario task instructions through semantic encoding and structured mapping.

[0028] It should be noted that the pre-built industrial meter reading system uses industry-standard visual inspection tools such as Halcon / OpenCV. This system is a dedicated image processing system developed based on machine vision algorithms, integrating a set of algorithm modules for object detection, geometric feature extraction, optical character recognition (OCR), and numerical calculation. Built upon industry-standard vision development libraries (such as OpenCV, Halcon, or deep learning inference frameworks), this system is used to complete low-level image feature extraction and preliminary reading tasks before large-scale model intervention.

[0029] It contains the following modules: The dial positioning module uses a pre-trained deep learning object detection model (such as the YOLO series) to infer the location information of the industrial meter body and dial area from the input image of the meter to be read.

[0030] Feature extraction module: Geometric extraction: Using image processing algorithms (such as Hough transform, edge detection combined with least squares fitting), the precise geometric features of the pointer on the image of the meter to be read (including the coordinates of the rotation center and the coordinates of the pointer endpoints) and the sequence distribution of the scale points are extracted.

[0031] Character recognition: The OCR model is called to recognize key character information on the dial (such as range value, unit, scale value). OCR technology is a commonly used technology and will not be described here.

[0032] Image preprocessing module: Apply image correction algorithms (such as perspective transformation and distortion correction) to perform geometric correction on the cropped dial area and generate a front view that is easy to read.

[0033] Preliminary reading generation module: Based on pure vision-based geometric calculation logic (such as calculating the angle between the pointer and the baseline, combined with the scale distribution model), it generates preliminary reading values ​​and outputs the "raw confidence level" of the result (usually calculated based on the IOU of the detection box or the feature matching degree).

[0034] The semantic encoding and structured mapping involve the following specific steps: First, data preprocessing and template mapping are performed, reading visual perception information packets (including dial positioning information, pointer geometric features, etc.) and cognitive context information (including instrument range, historical trends, and environmental parameters). The system then transforms this structured data into natural language descriptions based on predefined JSON Schema or string templates.

[0035] For example, this instrument is a pressure gauge with an initial reading of 0.5 MPa. However, based on historical trend analysis, this value usually fluctuates between 0.75 and 0.9 MPa. Furthermore, the current ambient light is dim, which may cause visual recognition interference due to insufficient light.

[0036] Then, semantic vector enhancement is performed. For non-visual environmental information (such as complex environmental metadata), the system can call the word embedding model to transform it into a high-dimensional semantic vector to capture the potential impact of environmental factors on the readings.

[0037] Next, the system performs instruction assembly and construction, assembling the instructions according to the logic of "system instructions + context + visual data + task requirements," where: System instructions: Define model roles (e.g., "Industrial Operations and Maintenance Expert").

[0038] Contextual background: Inject prior knowledge (such as "the normal operating pressure range of this device is 1 to 10").

[0039] Visual data: Fill in the natural language description content obtained from data preprocessing and template mapping.

[0040] Task requirements: Clear instructions (e.g., "Please use the tool to calculate the accurate reading and check the logical consistency").

[0041] The final generated multimodal prompt is a complete string containing text descriptions, numerical data, and task instructions.

[0042] Multimodal cue words contain at least one of the following information: Instrument type identification information: used to indicate the model or type of the instrument to be read.

[0043] Historical reading trend information: used to preset the reliable reading model of industrial meters to evaluate the rationality and continuity of the current reading.

[0044] Task priority or security level: Used to preset the reliable reading model of industrial meters and adjust the inference depth and level of caution.

[0045] Evaluation requirements for the performance of visual reading systems: to instruct a preset industrial meter reliable reading model to conduct performance spot checks on the visual reading system.

[0046] The multimodal prompt words are input into the preset industrial meter trusted reading model. The preset industrial meter trusted reading model performs cognitive reasoning and tool enhancement verification based on industrial domain knowledge according to the task instructions, and outputs the inferred reading and the corrected confidence level.

[0047] Specifically, the generated analysis report and trusted industrial meter reading data are sent to the preset terminal by calling the interface.

[0048] The preset terminals specifically include: user terminals and downstream system terminals.

[0049] According to an embodiment of the present invention, analyzing the image of the meter to be read and extracting visual perception information specifically includes: Based on a pre-built industrial meter visual recognition system, the positioning information of the industrial meter body and dial area is obtained through a deep learning object detection model.

[0050] The precise geometric features of the scale point sequence and pointer are obtained by using a preset precise geometric feature extraction method.

[0051] The scale value information is obtained through an optical character recognition model.

[0052] A corrected dial image is obtained using an image correction algorithm.

[0053] Preliminary readings and initial confidence levels are generated using a purely visual method.

[0054] The visual perception information is obtained by integrating the positioning information of the industrial meter body and dial area, the scale point sequence and the precise geometric features of the pointer, the scale value information, the dial image, the preliminary reading and the original confidence level.

[0055] Among them, the preset industrial meter trusted reading model is a multimodal large language model that can integrate text (such as structured prompt words) and visual (image of the meter to be read) information for deep reasoning and call a predefined toolset for enhanced verification.

[0056] The model is trained and optimized based on a general large model architecture (such as LLaMA, ChatGLM), and constructed by combining an industrial domain knowledge base, following these steps: Multimodal training: Using labeled data to learn text and image features, train the instrument to understand the semantic mapping between visual elements of industrial meters (such as pointer angle and scale value) and text commands.

[0057] Tool Integration: Built-in tool call interface can automatically trigger the call of tools in a predefined set of dedicated tools according to reasoning needs, ensuring that the results conform to physical laws and business logic.

[0058] Credibility assessment: Combining preliminary visual readings, tool verification results, and contextual information, outputs corrected credible readings and corrected confidence levels, and generates interpretable reasoning logic.

[0059] It should be noted that the dial area positioning information can be obtained by using a deep learning object detection model to acquire the positioning information of the industrial meter body and the dial area.

[0060] The precise geometric features of the pointer and the sequence of all identifiable tick marks can be obtained through deep learning segmentation models, traditional image processing algorithms, or a combination of both.

[0061] The identified scale value information can be obtained through scale semantic recognition using an optical character recognition (OCR) model; OCR technology is a commonly used technology and will not be elaborated here.

[0062] The corrected dial image can be obtained through image pre-correction using an image correction algorithm; the image correction algorithm is a commonly used existing technology and will not be described here.

[0063] The initial readings and original confidence levels can be generated using a purely visual method. This method employs a three-stage process commonly used in the industry: dial positioning, pointer analysis, and reading mapping. This technique is widely adopted and will not be elaborated upon here.

[0064] The deep learning object detection model's core algorithm is based on mature industry object detection model frameworks (such as YOLO series, RetinaNet, CenterNet, etc.) and is trained using industrial scenario-specific datasets. The network structure is lightweighted to enhance the recognition capability of industrial meters. The model development and deployment are based on general deep learning frameworks (such as PyTorch, TensorFlow).

[0065] Deep learning segmentation models use semantic segmentation networks (such as U-Net and Mask R-CNN) to directly output the pixel-level mask and precise location and contour of pointer regions through end-to-end learning, which is suitable for high-resolution or complex background images.

[0066] Traditional image processing algorithms, based on Canny edge detection, Hough line detection, etc., locate pointer edges and tick lines through steps such as filtering, edge extraction, and line fitting. They are suitable for clear, structured scenes with simple backgrounds.

[0067] How to combine the two: Use alone: ​​Deep learning segmentation models are used when image quality is high and interference is low; traditional image processing algorithms are used when real-time requirements are high or resources are limited.

[0068] The combined strategy is to use a deep learning segmentation model to locate the pointer region and a traditional image processing algorithm to refine the scale lines.

[0069] According to embodiments of the present invention, during the process of performing cognitive reasoning and tool enhancement verification based on industry domain knowledge according to task instructions, the predefined dedicated toolset invoked includes at least one of the following: Visual perception toolset: This tool is used to perform specific visual perception tasks on raw image data according to instructions from an industrial meter reliable reading model. Specific implementation methods include: Industrial meter vision reading tool: This is a pre-built industrial meter reading system that is called as a tool here; it processes the images of the meters to be read, extracts and generates a structured industrial meter vision perception information package.

[0070] Target detection tool: Implemented using lightweight models (such as YOLOv8 / v10), it locates specific industrial equipment bodies or their components (such as dials, switch bodies, indicator lights, valve handles) in the original image and outputs their location information.

[0071] Feature extraction and segmentation tools: Deep learning networks (such as semantic segmentation networks, U-Net, Mask R-CNN) are used to perform pixel-level classification of the original image to accurately locate the outlines and regions of key components such as pointers, scales, and valve handles; then, traditional image processing algorithms (such as Canny edge detection and Hough transform) are combined to refine the extraction of local geometric features (such as corners and edges) to generate high-precision feature mask information.

[0072] Image correction and enhancement tools: First, traditional algorithms (such as OpenCV's perspective transformation) are used to perform geometric correction on the image to eliminate distortion caused by the shooting angle; then, deep learning models (such as CNN-based denoising networks) or equivalent methods (such as adaptive histogram equalization) are integrated to enhance the image, effectively suppressing noise and improving the contrast of key areas, thereby outputting clear, distortion-free, high-quality images.

[0073] Geometric Quantization and Transformation Toolset: Used to perform geometric modeling and spatial numerical calculations based on visual perception results, including but not limited to: The scale geometry analysis tool is based on geometric modeling and mathematical optimization algorithms (such as least squares method, Hough transform, and circle fitting algorithm). By performing spatial analysis on the scale point sequence on the image, the tool uses optimization algorithms such as least squares method to fit and calculate key geometric parameters such as the rotation center and radius of the dial, and establishes the mapping relationship between the scale point distribution and the angle and distance.

[0074] Pointer pose analysis tool: Based on geometric calculation and keypoint detection algorithms (such as vector angle calculation, YOLOv8-pose keypoint regression, and Hough line transform); firstly, the pointer endpoints and rotation center are accurately located through a keypoint detection model or edge analysis, and then vector geometric relationships (such as...) are used to analyze the pointer pose. The deflection angle or position of the pointer relative to the zero position on the dial is calculated, and the physical quantity mapping is finally completed by combining the range parameters. y is the deflection angle (or position) of the pointer relative to the zero position of the dial, x is the horizontal coordinate of the pointer tip in the image coordinate system, y is the vertical coordinate of the pointer tip in the image coordinate system, x0 is the horizontal coordinate of the dial rotation center in the image coordinate system, and y0 is the vertical coordinate of the dial rotation center in the image coordinate system.

[0075] Coordinate transformation and projection tools: Based on geometric transformation algorithms (such as formal transformation, affine transformation, or perspective transformation and homography matrix); the tool calculates the spatial mapping relationship between the coordinate system of the original image and the target coordinate system (such as a standard front view template), and rotates, scales and translates the extracted pointer tip pixel coordinates and tick point coordinates, thereby accurately projecting the feature points into a unified standard coordinate system.

[0076] Reading calculation and interpolation toolset: Based on the geometric analysis results, spatial feature quantities are mapped to physical readings. The mapping process includes, but is not limited to: Numerical interpolation tool: Based on interpolation algorithms in numerical analysis (such as linear interpolation, polynomial interpolation or spline interpolation); the tool establishes a mathematical mapping relationship between identified adjacent standard scale points based on the angular position or arc distance of the pointer tip on the dial surface, thereby calculating continuous and accurate physical reading values.

[0077] Angle reading tool: Based on trigonometric function calculation and geometric mapping algorithms (such as polar coordinate transformation and least squares fitting); the tool directly converts the angle value into an accurate physical reading value by rotating the pointer relative to the reference zero point and combining it with the pre-stored "angle-value" calibration curve or linear mapping relationship.

[0078] Data alignment and consistency verification toolset: Based on multi-source constraints, multi-dimensional data matching and rationality verification are performed. The verification process includes, but is not limited to: Template matching and residual calculation tool: Based on a geometrically constrained robust matching model (such as RANSAC combined with homography matrix estimation and ICP iterative nearest point algorithm); the tool solves for optimal (such as rigid body / affine) transformation parameters by establishing the spatial correspondence between the standard scale points of the template and the observed scale points; the matching process strictly constrains the local topological consistency of the point set (such as neighborhood distance wallpaper or angular relationship), and finally calculates the distance between the observed point and the transformed template point as the point-by-point residual to quantitatively evaluate the alignment accuracy between visual inspection and the template.

[0079] Perspective correction quality assessment tool: Based on geometric consistency verification and error analysis models (such as reprojection error calculation, linear structure retention assessment or key point residual analysis), it evaluates the accuracy of known key points (such as scale marks, dial boundaries) in the original image by transforming them with the theoretical positions in the standard template coordinate system after correction matrix transformation.

[0080] Mechanical limit verification tool: Based on rule constraints and statistical verification models (such as physical boundary threshold checks, multi-pointer logic verification, and confidence weighting algorithms); first, the mechanical limit parameters of the instrument (e.g., the 0~100MPa range of the pressure gauge) are obtained, and the readings are directly compared to see if they exceed the mechanical limits; if multiple pointers or breakage anomalies are detected (e.g., identifying more than two valid pointers or angle mutations exceeding the historical fluctuation threshold), an error correction mechanism is triggered, and the readings are corrected based on the consistency of the geometric historical reading trend with adjacent scales (e.g., using a sliding window midpoint filter); the corrected confidence is calculated by weighting the residual distribution, boundary compliance, and logical self-consistency.

[0081] Anomaly Analysis and Assisted Judgment Toolset: Used to perform interference assessment and logical conflict determination based on environmental factors and prior logic. The determination process includes, but is not limited to: Environmental interference analysis tool: Based on image quality degradation modeling and environmental perception assessment algorithms (such as contrast adaptive histogram analysis, frequency domain noise analysis, and illumination non-uniformity quantification); by extracting key image quality indicators (such as local contrast standard deviation, high-frequency energy density, and color deviation index), and combining them with a preset environmental interference threshold library (such as contrast > 0.85 under strong illumination, and high-frequency components < 15% under water mist interference), the tool automatically quantifies the impact of interference on visual element extraction; the evaluation results are expressed as an interference coefficient of 0 to 1.

[0082] Logical conflict detection tool: Implemented based on rule-based reasoning and geometric constraint verification models (such as first-order logic rule engine, topological relationship consistency check or SAT-based satisfiability solution); By predefining the core logic rules of the pointer-scale system (such as "the pointer angle must be within the minimum / maximum scale range" and "the occluded area should not cover the valid reading range"), combined with the detected geometric parameters (such as pointer endpoint coordinates, scale line distribution, occluded area mask), it verifies whether the dial state conforms to physical logic; If a conflict is found (such as the pointer pointing to a prohibited area but the system still outputs a valid reading), an anomaly flag is triggered and error correction suggestions are generated, such as interpolation based on the most recent historical readings or switching to an alternative detection mode.

[0083] Figure 2 A flowchart of a method for obtaining inference readings and corrected confidence levels according to the present invention is shown.

[0084] like Figure 2 As shown in the embodiment of the present invention, cognitive reasoning and tool enhancement verification based on industrial domain knowledge are performed according to task instructions, and the reasoning reading and corrected confidence level are output, including: S201, Perform a visual element logical consistency check on the multimodal prompt words to obtain logical consistency check data; S202, perform information conflict analysis on the visual perception information package and cognitive context information to obtain conflict analysis and environmental impact assessment data; S203, based on logical consistency check data and conflict analysis and environmental impact assessment data, performs reading verification and correction decisions on the preliminary readings, and generates inferred readings and corrected confidence levels.

[0085] According to an embodiment of the present invention, a visual element logical consistency check is performed on multimodal cue words to obtain logical consistency check data, including: Based on the geometric features and element topological relationships in the industrial meter template knowledge base, a visual element logical consistency check is performed to verify the spatial geometric consistency between the pointer geometric features and the scale distribution pattern in the visual perception information package, and to verify the semantic correspondence between the identified scale values ​​and scale points.

[0086] Based on the spatial geometric consistency verification results and semantic correspondence, logical consistency check data is generated.

[0087] It should be noted that the industrial meter template knowledge base is constructed in the following way: Data collection and annotation: Collect sample images of various industrial meters and annotate geometric elements and topological relationships such as pointers, scales, and dials (e.g., the adjacent / intersecting relationship between pointers and scales).

[0088] Feature extraction and modeling: Extract the geometric parameters (position, angle, spacing) and topological structure of the features, establish parameterized templates, and define spatial constraints between features (such as uniform distribution of scale and pointer rotation range).

[0089] Structured storage: Templates are categorized by type and stored as a queryable database or file, supporting rapid matching of meter form and topology rules during inference.

[0090] The process verifies the spatial geometric consistency between the pointer geometric features and the scale distribution pattern in the visual perception information package. It invokes the scale geometry analysis tool and pointer posture analysis tool from a predefined dedicated toolset. The scale geometry analysis tool calculates key geometric parameters such as the fitted rotation center and radius based on the scale point sequence in the visual perception information package, and constructs the mapping relationship between scale points and angles and distances. The pointer posture analysis tool accurately determines the pointer tip position based on the pointer geometric features (e.g., bounding box, centerline, endpoints) and the fitted rotation center in the visual perception information package, and calculates its angle or position in the dial coordinate system. The results from both tools are combined to obtain the corresponding geometric parameters. These parameters are then compared with the content in the industrial meter template knowledge base to verify spatial geometric consistency and semantic correspondence.

[0091] According to an embodiment of the present invention, information conflict analysis is performed on visual perception information packets and cognitive context information to obtain conflict analysis and environmental impact assessment data, including: The differences between the visual perception information package and the cognitive context information are determined by comparing the two.

[0092] Obtain environmental metadata integrated into cognitive context information as environmental factors.

[0093] Based on environmental factors, a predefined environmental interference analysis tool is invoked to assess the potential interference of environmental factors on the extraction of elements in the visual perception information package, and to identify the key visual features that cause the interference.

[0094] Based on differences, potential disturbances, and key visual features, conflict analysis and environmental impact assessment data are generated.

[0095] It should be noted that by calling the environmental interference analysis tool in the predefined dedicated working set, the environmental metadata (such as strong light, shadow, water mist) in the visual perception information package is evaluated to assess its potential interference with image quality and visual element extraction, and the corresponding interference level assessment is provided to identify the key visual features that cause interference.

[0096] The difference between the visual perception information package and the cognitive context information is determined by the comparison and conflict analysis module built into the preset industrial meter reliable reading model. The specific process is as follows: Cross-modal alignment and feature comparison: The model first maps visual features (such as pointer position and scale value) and cognitive context (such as historical readings and environmental parameters) to a unified semantic space through a cross-modal encoder, and compares the inconsistencies between the two in terms of numerical values, trends, or logic.

[0097] The rule engine is combined with machine learning: Based on predefined industry rules (such as measurement range and reading change threshold) and abnormal patterns in the training data, the model automatically identifies data conflicts (such as visual readings that exceed historical range or do not match temperature parameters).

[0098] Environmental interference assessment: By calling the environmental interference analysis tools in the predefined dedicated working set, the environmental metadata (such as strong light, shadow, water mist) in the visual perception information package is evaluated to assess its potential interference to image quality and visual element extraction, and the corresponding interference level assessment is provided to identify the key visual features that cause interference.

[0099] According to an embodiment of the present invention, based on logical consistency check data and conflict analysis and environmental impact assessment data, a reading verification and correction decision is made on the preliminary readings to generate inferred readings and corrected confidence levels, including: By combining logical consistency check data with conflict analysis and environmental assessment data, the rationality and physical accuracy of the initial readings are thoroughly verified.

[0100] If inconsistencies, interference, and potential errors are detected, the preset industrial meter reliable reading model, based on its reasoning ability and tool calls, obtains the model's inference reading and its corrected confidence level through preset correction and calculation instructions.

[0101] It should be noted that a deep verification of the rationality and physical accuracy of the initial reading is performed. Specifically, the initial reading and its original confidence level are input, along with logical consistency check data and conflict analysis and environmental impact assessment data. Then, the mechanical limit verification tool in the predefined dedicated toolset is called to count the number of pointers and readings in the visual perception information package. This verifies whether the pointer reading is within the mechanical limit range of the meter (i.e., the maximum physical boundary of the pointer's movement) and whether there are errors such as multiple pointers or pointer breakage, and corrects these errors. All indicators are then summarized to calculate the reading and correct the confidence level.

[0102] If inconsistencies, interference, or potential errors are detected, corrections will begin using the following specific methods: Recalculation: When the confidence level of the visual reading is low, the angle reading tool or the numerical interpolation tool in the predefined dedicated toolset is invoked. The angle reading tool can calculate the position of the pointer in the scale-value mapping relationship based on the pointer angle in the visual perception information packet and return an accurate physical reading value. The numerical interpolation tool can calculate the position of the pointer tip in the scale mapping relationship based on the pointer tip position in the visual perception information packet and perform linear or nonlinear interpolation between adjacent standard scale points to calculate an accurate physical reading value.

[0103] Logic Correction: When a “pointer abnormality” or “scale abnormality” is detected, the logic conflict detection tool in the predefined dedicated toolset is invoked to analyze the relative position of the pointer and scale in the visual perception information packet, determine whether there is a logic conflict such as “pointer zeroing abnormality” or “dial partial occlusion causing reading failure”, and provide error correction suggestions; based on the error correction suggestions, the reading is corrected to a value that verifies the physical laws.

[0104] Figure 3 The flowchart illustrates a process of generating reliable industrial meter reading data by executing a dynamic decision-making strategy according to the present invention.

[0105] like Figure 3 As shown, according to an embodiment of the present invention, analysis is performed based on preliminary readings, initial confidence levels, inferred readings, and revised confidence levels to execute a dynamic decision-making strategy and generate reliable industrial meter reading data, including: S301, If ​​the deviation between the initial reading and the inferred reading is within the preset allowable range, and both the original confidence level and the modified confidence level are higher than the high confidence threshold, then output the initial reading and label it as a high-confidence direct reading. S302, if the deviation between the initial reading and the inferred reading is within the preset allowable range and the original confidence level is lower than the high confidence threshold, then the inferred reading is used as the final result and labeled with the model correction confirmation, while providing an explanation of the inference process of the preset industrial meter reliable reading model. S303 If the deviation between the initial reading and the inferred reading exceeds the preset allowable range, or if the preset industrial meter reliable reading model detects an unsolvable serious conflict during the inference process, a final decision containing detailed conflict analysis, cause inference and potential risk warning will be output, and a label indicating that manual review is required will be added.

[0106] It should be noted that the preset allowable range is set based on the instrument's maximum permissible error (MAE). For example, for a pointer-type pressure gauge with an accuracy class of ±1.5%, the allowable range is usually set to ±2% of the range (including measurement uncertainty) to ensure that the deviation is within the metrological compliance range.

[0107] The high confidence threshold is determined by the optimal F1 score of the visual model on the validation set based on the original confidence threshold (e.g., 0.85); the modified confidence threshold (e.g., 0.9) is dynamically adjusted in conjunction with industry safety standards (e.g., IEC 61511 error tolerance requirements for SIL systems).

[0108] Both the preset tolerance range and the high confidence threshold need to be verified through historical data backtracking (such as statistical analysis of 90 days of on-site data) and expert rule verification to ensure that they meet the reliability requirements of industrial scenarios.

[0109] Provides detailed explanations of the reasoning process for a pre-defined reliable reading model for industrial meters, including but not limited to: Comparison of geometric parameters before and after correction, analysis of the specific types and impacts of logical conflicts or environmental interference, and the mathematical basis of the correction strategy.

[0110] The pre-defined industrial meter reliable reading model detects specific instances of irresolvable serious conflicts during the inference process, including but not limited to: Geometric conflicts, environmental interference, and logical contradictions.

[0111] According to an embodiment of the present invention: generating an analysis report based on trusted industrial meter reading data includes: The analysis report is at least one of the following: a single meter reading analysis report and an industrial equipment operating status correlation analysis report.

[0112] Obtain the final confirmed readings, confidence levels, and labels from trusted industrial meter readings.

[0113] Intermediate data from visual element logical consistency checks, conflict analysis, and environmental impact assessment processes are acquired and used as inference data for model correction.

[0114] Based on the evaluation results of the environmental interference analysis tool, a qualitative analysis was conducted on the specific visual interference factors that led to questionable readings, and the visual interference analysis results were obtained.

[0115] The deviation between the initial reading and the inferred reading is calculated. If the deviation exceeds the preset allowable range, the final decision data containing detailed conflict analysis, cause inference and potential risk warnings is generated. Combined with the preset standard operating procedures and knowledge base, preliminary handling suggestions and contingency plans are generated. The results are summarized in conjunction with the risk level assessment to obtain risk assessment and handling suggestions.

[0116] The final confirmed readings and confidence levels, along with the labels, inference data from model corrections, visual interference analysis results, risk assessments, and mitigation recommendations, are integrated to obtain a single meter reading analysis report.

[0117] A list of equipment statuses is obtained by displaying the current operating status of all identified industrial equipment within the inspection area.

[0118] Based on the reading results from multiple devices, multi-device correlation analysis is performed on the detected anomalies to infer the potential causes of the anomalies. Combined with the detected anomalies, an anomaly summary and logical analysis report is generated.

[0119] The deviation between the initial reading and the inferred reading is calculated. If the deviation exceeds the preset allowable range, the final decision data containing detailed conflict analysis, cause inference and potential risk warnings is generated. Combined with the preset standard operating procedures and knowledge base, preliminary handling suggestions and contingency plans are generated. The results are summarized in conjunction with the risk level assessment to obtain risk assessment and handling suggestions.

[0120] Summarize and optimize key algorithms, model parameters, data augmentation strategies, and incremental training sample selection criteria for the system, and provide feedback suggestions for system optimization.

[0121] By integrating the equipment status list, anomaly summary and logical analysis report, risk assessment and handling suggestions, and system optimization feedback suggestions, an industrial equipment operation status correlation analysis report is obtained.

[0122] It should be noted that the preset standard operating procedure (SOP) is a standardized operating guide for handling equipment anomalies in industrial scenarios. It clarifies the anomaly classification, handling steps, division of responsibilities and safety specifications (such as the pressure vessel anomaly handling SOP, which specifies the pressure relief operation procedure triggered by the overpressure threshold).

[0123] The knowledge base is a structured collection of industrial knowledge, including equipment parameter databases, historical failure case databases, industry safety standards (such as ISO 13849 safety specifications for machinery) and expert experience rules.

[0124] The summary of system optimization feedback suggestions needs to be achieved through a four-step closed-loop analysis method: Performance attribution analysis: Based on the error classification matrix and learning curve, we locate algorithm bottlenecks and focus on identifying high-impact problems.

[0125] Parameter sensitivity verification: The impact of key parameters was quantified through ablation experiments.

[0126] Data-algorithm co-evaluation: When selecting incremental samples, prioritize covering high-frequency error patterns and verify the effectiveness of data augmentation strategies.

[0127] Engineering constraint calibration: Combining deployment environment limitations, the number of model parameters is compressed to within the threshold to ensure that the optimization solution can be implemented.

[0128] List all identified industrial equipment within the inspection area, including meters and other industrial equipment logically related to the meters.

[0129] List the current operating status of all identified industrial equipment within the inspection range, including: for example, instrument readings, switch status, indicator light colors, and valve handle opening.

[0130] Detected anomalies include over-range readings and status mismatches.

[0131] Inferring potential causes of anomalies includes: inferring the underlying root cause or cascading effects of the anomaly (e.g., a low reading on pressure gauge Y due to valve X not being fully open).

[0132] Figure 4 The flowchart illustrates a method for performance evaluation and adaptive evolution of an industrial meter reading system according to the present invention.

[0133] like Figure 4 As shown in the embodiments of the present invention, the reliable reading method for industrial meters based on visual perception and large model collaborative reasoning also includes a performance evaluation and evolution method for a visual recognition system for the status of industrial equipment components.

[0134] S401, based on preset conditions, sample and evaluate the readings generated by the industrial meter vision recognition system and its correction process, and generate an online evaluation report for the system; S402, based on the online evaluation report, automatically selects image samples that have a deviation between the preliminary reading and the inferred reading that exceeds the preset allowable range or that the preset industrial meter reliable reading model detects an unsolvable serious conflict during the inference process, and that have been manually verified or whose corrected confidence level corresponding to the reading after correction by the preset industrial meter reliable reading model is higher than the preset high confidence threshold as difficult case samples. S403 adds difficult example samples to the training set to update and optimize the industrial meter vision recognition system.

[0135] It should be noted that the preset conditions are user-defined conditions, which may be set as a specific time period or the number of tasks to be executed, and can be set by those skilled in the art according to actual needs.

[0136] Based on different conditions set by the user, the reliable industrial meter reading data generated by the industrial meter reading system is sampled using different sampling methods to extract relevant image samples and multimodal prompt words.

[0137] Each user-defined condition corresponds to a specific sampling method, and the relationship between the sampling method and the user-defined conditions is set by those skilled in the art based on actual needs. Different sampling methods include: full sampling, random sampling, and risk-based sampling.

[0138] After sampling, the multimodal prompt words of the sampled data are input into the preset industrial meter reliable reading model. Based on the task instructions, cognitive reasoning and tool enhancement verification based on industrial domain knowledge are performed, and the inferred reading and corrected confidence level are output.

[0139] Consistency assessment and verification are performed based on sampled readings, original confidence levels, inferred readings, and corrected confidence levels to generate an online assessment report for the industrial meter reading system.

[0140] Difficult example samples are added to the training set as training samples for industrial meter reading systems or as the basis for optimizing core algorithms.

[0141] The online assessment report includes the following: Core data comparison: Displays sampled readings, preliminary readings, inferred readings and corresponding confidence levels, analyzes bias and consistency, and labels high confidence / correction / conflict.

[0142] Abnormal Event Log: Summarizes anomalies such as exceeding tolerance limits and logical conflicts, with conflict analysis, environmental interference assessment and risk level attached.

[0143] System performance statistics: Statistics on key indicators such as recognition accuracy, response latency, and confidence distribution are presented by device / scenario.

[0144] Optimization suggestions: Based on abnormal patterns and performance bottlenecks, we propose algorithm parameter tuning, data augmentation directions, and model iteration requirements.

[0145] The steps to generate an online assessment report are as follows: Data alignment and aggregation: Real-time acquisition of readings and confidence levels, aligned by time / device dimension.

[0146] Consistency assessment: Calculate the reading deviation and determine the confidence label based on the confidence threshold.

[0147] Anomaly detection and classification: Identify conflicts through machine learning and quantify interference by associating environmental metadata.

[0148] Indicator statistics and visualization: generating charts based on data.

[0149] Report integration and delivery: Output results and generate online evaluation reports.

[0150] According to embodiments of the present invention, this method can also be used for industrial equipment component status identification based on visual perception and large model collaborative reasoning, and the specific steps are as follows: Acquire images of the industrial equipment components to be identified.

[0151] An industrial equipment component status visual recognition system is used to analyze images of industrial equipment components to be identified and extract visual perception information. The visual perception information includes the industrial equipment component visual perception information package, preliminary status judgment results, and original confidence level.

[0152] Integrate non-visual environmental information and prior knowledge related to industrial equipment components as cognitive context information; convert visual perception information packages and cognitive context information into multimodal prompt words containing task instructions for industrial scenarios through semantic encoding and structured mapping.

[0153] The multimodal prompts are input into the preset trusted identification model of industrial equipment components. Based on the task instructions, cognitive reasoning and tool enhancement verification based on industrial domain knowledge are performed, and the reasoning status and corrected confidence are output.

[0154] Based on the initial state, original confidence level, inference state, and revised confidence level, an analysis is performed to execute a dynamic decision-making strategy and generate reliable industrial equipment component identification data.

[0155] An analysis report is generated based on the identification data of industrial equipment components, and the trusted industrial equipment component identification data and analysis report are sent to a preset terminal.

[0156] It should be noted that the industrial equipment component status visual recognition system is a commonly used visual inspection tool in the industry, such as Halcon / OpenCV.

[0157] It is a dedicated image processing system developed based on machine vision algorithms, integrating a set of algorithm modules for object detection, geometric feature extraction, optical character recognition (OCR), and numerical calculation. The system is built on industry-standard vision development libraries (such as OpenCV, Halcon, or deep learning inference frameworks) and is used to complete low-level image feature extraction and preliminary reading tasks before large models are involved.

[0158] It contains the following modules: Component localization module: Based on a pre-trained deep learning object detection model (such as the YOLO series), it performs inference on the input image of the industrial equipment to be identified, and accurately locates the position and bounding box of various components.

[0159] Feature extraction module: Geometric extraction: Using image processing algorithms (such as edge detection, contour analysis, template matching) to extract the geometric features (such as edge contours) of components.

[0160] Character recognition: Retrieves key character information (such as serial number and specifications) from industrial equipment components in the OCR model. OCR technology is a commonly used technology and will not be elaborated here.

[0161] Image preprocessing module: Apply image correction algorithms (such as perspective transformation and distortion correction) to perform geometric correction on the image and generate a front view that is easy to read.

[0162] Preliminary reading generation module: Based on pure vision-based geometric calculation logic (such as calculating the size of the component), it generates preliminary state assessment results and outputs the "raw confidence level" of the results.

[0163] The preset industrial meter trusted reading model is a multimodal large language model that can integrate textual (such as structured prompt words) and visual (image of the meter to be read) information for deep reasoning and call a predefined toolset for enhanced verification.

[0164] The model is trained and optimized based on a general large model architecture (such as LLaMA, ChatGLM), and constructed by combining an industrial domain knowledge base, following these steps: Multimodal training: Using labeled data to learn text and image features, train the instrument to understand the semantic mapping between visual elements of industrial meters (such as outlines and light colors) and text commands.

[0165] Tool Integration: Built-in tool call interface can automatically trigger the call of tools in a predefined set of dedicated tools according to reasoning needs, ensuring that the results conform to physical laws and business logic.

[0166] Credibility assessment: Combining preliminary visual readings, tool verification results, and contextual information, outputs corrected credible readings and corrected confidence levels, and generates interpretable reasoning logic.

[0167] The visual perception package for industrial equipment components includes at least: the positioning information of the equipment component itself, and the current visual characteristics of the equipment component (such as the open / closed position of a switch, the on / off state or color of an indicator light, and the rotation angle of a valve handle).

[0168] Among them, the positioning information of the equipment components can be obtained by using a deep learning object detection model to acquire the positioning information of the industrial equipment components.

[0169] The current visual features of the equipment components can be obtained from the images of the industrial equipment components to be identified using feature extraction and segmentation tools.

[0170] The system integrates cognitive context information of industrial equipment components, including non-visual environmental information and prior knowledge related to the components. This information includes at least: imaging location identifier, image acquisition timestamp, current environmental conditions (e.g., weather conditions, artificial light source angle), and the normal operating status range, standard operating procedures (SOPs), and logical association rules of the equipment component.

[0171] In the process of cognitive reasoning and tool enhancement verification, the specific steps of the device component state logic consistency check are as follows: based on the device component template information and state rules in the industrial scenario knowledge base, verify whether the extracted current visual features of the device component match the expected state (e.g., on / off, bright / dark, specific color or angle); based on the verified device state, generate device component state logic consistency check data.

[0172] According to embodiments of the present invention, this method can also be used to implement cognitive reasoning and tool enhancement verification through an intelligent agent architecture, with the specific steps as follows: Integrate non-visual environmental information and prior knowledge related to the meter to be read as cognitive context information; convert the cognitive context information into multimodal prompt words containing task instructions in the industrial scenario through semantic encoding and structured mapping.

[0173] The multimodal prompt words are input into the preset industrial meter reliable reading model. The model calls the industrial meter visual reading tools in the predefined dedicated toolset according to the task instructions; the image of the meter to be read is processed to extract and generate a structured industrial meter visual perception information package, preliminary reading results and original confidence level.

[0174] The model also performs cognitive reasoning and tool enhancement verification based on industry domain knowledge according to task instructions, and outputs inference readings and corrected confidence scores.

[0175] Based on the analysis of preliminary readings, initial confidence levels, inferred readings, and revised confidence levels, a dynamic decision-making strategy is implemented to generate reliable industrial meter reading data.

[0176] An analysis report is generated based on the reading data from trusted industrial meters, and the trusted industrial meter reading data and analysis report are sent to a preset terminal.

[0177] A second aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a program for a reliable industrial meter reading method based on visual perception and large model collaborative reasoning, wherein when the program for the reliable industrial meter reading method based on visual perception and large model collaborative reasoning is executed by a processor, it implements the steps of the reliable industrial meter reading method based on visual perception and large model collaborative reasoning as described in any of the embodiments above.

[0178] All information (including but not limited to user equipment information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals (including but not limited to signals transmitted between user terminals and other devices) involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the "meter image to be read" and "non-visual environmental information" involved in this disclosure were obtained with full authorization.

[0179] This invention discloses a reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning. The method includes: processing the image to be read using an industrial meter visual reading system to obtain a visual perception information package; integrating non-visual environmental information and prior knowledge to construct cognitive context information; converting the visual perception information package and cognitive context information into multimodal prompt words and inputting them into a large language model; driving the model to call a dedicated functional toolset to perform reasoning and verification, outputting the inferred reading and corrected confidence level; generating the final reading result through a dynamic decision-making strategy, and supporting performance evaluation and feedback optimization. This method can also be used for reading industrial equipment components and executing methods using an intelligent agent architecture. By deeply coupling visual perception and cognitive reasoning, this invention effectively solves the problem of misreading industrial meters under complex operating conditions, significantly improving the reliability and interpretability of the reading results, and providing technical support for intelligent industrial operation and maintenance.

[0180] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0181] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0182] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0183] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0184] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

Claims

1. A reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning, characterized in that, include: Acquire the image of the meter to be read; The image of the meter to be read is analyzed to extract visual perception information; The visual perception information includes industrial meter visual perception information packages, preliminary readings, and initial confidence levels. Integrate non-visual environmental information and prior knowledge related to the meter to be read as cognitive context information; convert the visual perception information package and the cognitive context information into multimodal prompt words containing task instructions in the industrial scenario through semantic encoding and structured mapping; The multimodal prompt words are input into a preset industrial meter reliable reading model. Based on the task instructions, cognitive reasoning and tool enhancement verification based on industrial domain knowledge are performed, and the reasoned reading and corrected confidence level are output. Based on the preliminary readings, the original confidence levels, the inferred readings, and the revised confidence levels, an analysis is performed to execute a dynamic decision-making strategy and generate reliable industrial meter reading data. An analysis report is generated based on the trusted industrial meter reading data, and the trusted industrial meter reading data and the analysis report are sent to a preset terminal.

2. The reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning as described in claim 1, characterized in that, The step of analyzing the image of the meter to be read and extracting visual perception information includes: Based on a pre-built industrial meter visual recognition system, the positioning information of the industrial meter body and dial area is obtained through a deep learning object detection model; The precise geometric features of the scale point sequence and pointer are obtained by using a preset precise geometric feature extraction method. Obtain scale value information through an optical character recognition model; A corrected dial image is obtained using an image correction algorithm; Preliminary readings and initial confidence levels are generated using purely visual methods; The positioning information of the industrial meter body and dial area, the scale point sequence and the precise geometric features of the pointer, the scale value information, the dial image and the preliminary reading and the original confidence level are integrated to obtain visual perception information.

3. The reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning as described in claim 1, characterized in that, The process of performing cognitive reasoning and tool enhancement verification based on industry domain knowledge according to task instructions, and outputting inference readings and corrected confidence scores, includes: The visual element logical consistency check is performed on the multimodal prompt words to obtain logical consistency check data; Information conflict analysis is performed on the visual perception information package and the cognitive context information to obtain conflict analysis and environmental impact assessment data. Based on the logical consistency check data and the conflict analysis and environmental impact assessment data, the preliminary readings are verified and revised to generate inferred readings and revised confidence levels.

4. The reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning as described in claim 3, characterized in that, The step of performing a visual element logical consistency check on the multimodal cue words to obtain logical consistency check data includes: Based on the geometric features and element topology relationships in the industrial meter template knowledge base, a visual element logical consistency check is performed to verify the spatial geometric consistency between the pointer geometric features and the scale distribution pattern in the visual perception information package, and to verify the semantic correspondence between the identified scale values ​​and scale points. Based on the spatial geometric consistency verification results and the semantic correspondence, logical consistency check data is generated.

5. The reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning as described in claim 3, characterized in that, The step of performing information conflict analysis on the visual perception information package and the cognitive context information to obtain conflict analysis and environmental impact assessment data includes: The differences between the visual perception information package and the cognitive context information are determined by comparing the two. The environmental metadata integrated into the cognitive context information is obtained as environmental factors; Based on the aforementioned environmental factors, a predefined environmental interference analysis tool is invoked to assess the potential interference of environmental factors on the extraction of elements in the visual perception information package, and to identify the key visual features that cause the interference. Based on the aforementioned differences, potential interferences, and key visual features, conflict analysis and environmental impact assessment data are generated.

6. The reliable reading method for industrial meters based on visual perception and large-model collaborative reasoning as described in claim 3, characterized in that, The process of verifying and correcting the preliminary readings based on the logical consistency check data and the conflict analysis and environmental impact assessment data, generating inferred readings and corrected confidence levels, includes: By combining the logical consistency check data and the conflict analysis and environmental assessment data, the rationality and physical accuracy of the initial readings are thoroughly verified. If inconsistencies, interference, or potential errors are detected, the preset industrial meter reliable reading model, based on its reasoning ability and tool calls, obtains the model's inference reading and its corrected confidence level through preset correction and calculation instructions.

7. The reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning as described in claim 1, characterized in that, The process of analyzing the preliminary reading, the original confidence level, the inferred reading, and the revised confidence level, executing a dynamic decision-making strategy, and generating reliable industrial meter reading data includes: If the deviation between the preliminary reading and the inferred reading is within a preset allowable range, and both the original confidence level and the revised confidence level are higher than the high confidence threshold, then the preliminary reading is output and labeled as a high-confidence direct reading. If the deviation between the initial reading and the inferred reading is within a preset allowable range, and the original confidence level is lower than the high confidence threshold, then the inferred reading is taken as the final result and labeled with a model correction confirmation tag. At the same time, an explanation of the inference process of the preset industrial meter reliable reading model is provided. If the deviation between the preliminary reading and the inferred reading exceeds the preset allowable range, or if the preset industrial meter reliable reading model detects an unsolvable serious conflict during the inference process, a final decision containing detailed conflict analysis, cause inference, and potential risk warning will be output, and a label indicating that manual review is required will be added.

8. The reliable reading method for industrial meters based on visual perception and large-scale model collaborative reasoning according to claim 1, characterized in that, The step of generating an analysis report based on the trusted industrial meter reading data includes: The analysis report is at least one of a single meter reading analysis report and an industrial equipment operation status correlation analysis report; Obtain the final confirmed reading result, confidence level, and tag from the trusted industrial meter reading data; Intermediate data from the visual element logical consistency check, conflict analysis, and environmental impact assessment processes are obtained and used as inference data for model correction. Based on the evaluation results of the environmental interference analysis tool, a qualitative analysis is conducted on the specific visual interference factors that lead to questionable readings, and the visual interference analysis results are obtained. The deviation between the initial reading and the inferred reading is calculated. If the deviation exceeds the preset allowable range, the final decision data containing detailed conflict analysis, cause inference and potential risk warning is generated. Combined with the preset standard operating procedures and knowledge base, preliminary handling suggestions and contingency plans are generated. The risk assessment and handling suggestions are summarized based on the risk level assessment results. The final confirmed reading results and confidence levels, along with the labels, the inference data from the model correction, the visual interference analysis results, and the risk assessment and handling recommendations, are integrated to obtain a single meter reading analysis report. A list of equipment status is obtained by listing the current operating status of all identified industrial equipment within the inspection range; Based on the reading results from multiple devices, multi-device correlation analysis is performed on the detected anomalies to infer the potential causes of the anomalies, and an anomaly summary and logical analysis report is generated by combining the detected anomalies. The deviation between the initial reading and the inferred reading is calculated. If the deviation exceeds the preset allowable range, the final decision data containing detailed conflict analysis, cause inference and potential risk warning is generated. Combined with the preset standard operating procedures and knowledge base, preliminary handling suggestions and contingency plans are generated. The risk assessment and handling suggestions are summarized based on the risk level assessment results. Summarize and optimize key algorithms, model parameters, data augmentation strategies, and incremental training sample selection criteria for the system, and provide feedback suggestions for system optimization. The equipment status list, anomaly summary and logical analysis report, risk assessment and handling suggestions, and system optimization feedback suggestions are integrated to obtain the industrial equipment operation status correlation analysis report.

9. The reliable reading method for industrial meters based on visual perception and large-model collaborative reasoning as described in claim 2, characterized in that, Also includes: Based on preset conditions, the readings generated by the industrial meter vision recognition system and its correction process are sampled and evaluated to generate an online evaluation report for the system. Based on the online evaluation report, image samples that are automatically selected as difficult examples are those where the deviation between the preliminary reading and the inferred reading exceeds the preset allowable range, or where the preset industrial meter reliable reading model detects an unsolvable serious conflict during the inference process, and which have been manually verified or whose corrected confidence level corresponding to the reading after correction by the preset industrial meter reliable reading model is higher than the preset high confidence threshold. The difficult examples are added to the training set to update and optimize the industrial meter visual recognition system.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a program for a reliable industrial meter reading method based on visual perception and large-model collaborative reasoning. When the program is executed by a processor, it implements the steps of the reliable industrial meter reading method based on visual perception and large-model collaborative reasoning as described in any one of claims 1 to 9.

Images