A single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants

Abstract

The application relates to a single-pointer instrument automatic reading method for intelligent inspection of a nuclear power plant, comprising the following steps: determining a dial area view of an instrument image on a single-pointer instrument; detecting four key geometric feature points on the dial area view, wherein the four key geometric feature points comprise a pointer rotation center (O), a pointer tip (P), a range zero point (Z) defined based on a scale line outer endpoint centroid and a range full point (F); calculating an included angle theta of a vector OP and a vector OZ, and an included angle Phi of a vector OF and the vector OZ; and according to a preset instrument range [Range_min, Range_max], the application calculates a reading value according to a formula Value=Range_min+(theta / Phi)*(Range_max-Range_min). The instrument reading problem is converted into a detection and spatial relationship analysis problem of a few feature points with constant geometric relationships on a dial, so that the technical problem that the prior art excessively depends on scale line integrity and digit recognition accuracy is solved.

Description

Technical Field

[0001] This application belongs to the interdisciplinary field of computer vision and industrial automation, and specifically relates to an automatic reading method for single-pointer instruments used in intelligent inspection of nuclear power plants. Background Technology

[0002] In intelligent inspection systems for nuclear power plants, automatically reading data from a large number of single-pointer instruments (pressure, temperature, flow, etc.) is a core requirement. Current mainstream technical solutions mainly include: Traditional methods based on template matching and Hough transform require pre-establishing a standard template for each instrument, matching the template to position the dial, and then using the Hough transform to detect the pointer's linearity. This method is prone to failure in the complex environment of nuclear power plants (such as glare and contamination) and has poor generalization ability.

[0003] Semantic segmentation methods based on deep learning employ networks such as U-Net to simultaneously segment pointers and tick marks, and then calculate the angle. These methods rely on large amounts of labeled data, incur high computational costs, and experience significant performance degradation when tick marks are blurred or partially occluded.

[0004] End-to-end methods based on regression or keypoint detection, such as those described in patent application CN202210963854.4, directly regress readings or detect predefined keypoints. However, these methods typically require accurate identification of dial digits (OCR) to determine the measurement range and lack stability when digits are obscured or fonts are varied.

[0005] Analysis revealed a common dependency in existing technologies: the requirement for complete and clear scale line image features and / or digital recognition results as a necessary condition for calculating readings. This is a critical vulnerability in the reflective environment of nuclear power plant instrument panels. Summary of the Invention

[0006] In view of this, this application provides an automatic reading method for single-pointer instruments for intelligent inspection of nuclear power plants. By transforming the instrument reading problem into the problem of detecting and spatially resolving a few feature points with invariant geometric relationships on the dial, this method solves the technical problem of excessive reliance on the integrity of scale lines and the accuracy of digital recognition in existing technologies.

[0007] The first aspect of this application provides a method for automatic reading of a single-pointer instrument for intelligent inspection of nuclear power plants, which includes the following steps.

[0008] Step S1: Determine the dial area view of the instrument image on the single-pointer instrument.

[0009] Step S2: Detect four key geometric feature points on the dial area view. The four key geometric feature points include the pointer rotation center (O), the pointer tip (P), the zero point (Z) of the range defined based on the centroid of the outer endpoint of the scale line, and the full-range point (F).

[0010] Step S3: Calculate the angle θ between vectors OP and OZ, and the angle Φ between vectors OF and OZ.

[0011] Step S4: Calculate the reading according to the preset instrument range [Range_min, Range_max] using the formula Value=Range_min+(θ / Φ)*(Range_max-Range_min).

[0012] In one specific embodiment of this application, step S3 includes steps S31 to S33.

[0013] Step S31: Calculate vector OZ=V_z=O->Z, vector OF=V_f=O->F, and vector OP=V_p=O->P.

[0014] Step S32: Calculate the angle θ between vector OP and vector OZ = arccos((V_p·V_z) / (|V_p||V_z|)), and determine the quadrant according to the sign of the cross product to obtain θ in the range [0, 360°).

[0015] Step S33: Calculate the angle Φ between vector OF and vector OZ = arccos((V_f·V_z) / (|V_f||V_z|)).

[0016] In one specific embodiment of this application, step S4 includes steps S41 and S42.

[0017] Step S41: Automatically obtain the preset instrument range [Range_min,Range_max].

[0018] Step S42: Calculate the reading according to the preset instrument range [Range_min,Range_max] using the formula Value=Range_min+(θ / Φ)*(Range_max-Range_min).

[0019] In one specific embodiment of this application, step S1 includes steps S11 to S13.

[0020] Step S11: Obtain the instrument image.

[0021] Step S12: Preprocess the instrument image to eliminate the overshadowing of key geometric feature points by highlights.

[0022] Step S13: Locate the dial area based on the instrument image.

[0023] In one specific embodiment of this application, step S12 includes step S121.

[0024] Step S121: Perform illumination normalization and adaptive contrast enhancement on the instrument image, focusing on using a physical model-based reflective area detection and repair algorithm to eliminate the overshadowing of key point features by highlights.

[0025] In one specific embodiment of this application, step S13 includes steps S131 and S132.

[0026] Step S131: Use a lightweight object detection network or elliptic detection algorithm to locate the dial area.

[0027] Step S132: When the dial area is detected to be elliptical, the perspective transformation matrix is ​​calculated by fitting the ellipse parameters to correct the dial area into a circular area image.

[0028] In one specific embodiment of this application, step S2 includes step S21.

[0029] Step S21: A unified keypoint detection network (KP-Net) is used to directly process the dial area, and the normalized coordinates of the following four key geometric feature points are regressed. These four key geometric feature points include the pointer rotation center (O), the pointer tip (P), the zero point (Z) defined based on the centroid of the outer endpoints of the scale lines, and the full-scale point (F). The backbone of the keypoint detection network adopts a feature pyramid structure, and a pointer tip attention module is introduced into the prediction head. The pointer tip attention module enhances the pointer tip features in the feature map by analyzing the gradient direction consistency and highlight symmetry of the pointer region, suppressing interference features caused by dirt or reflections.

[0030] In one specific embodiment of this application, reference is made to Figure 3 After step S2, the method for automatic reading of single-pointer instrument for intelligent inspection of nuclear power plants also includes step S10.

[0031] Step S10: Establish a geometric constraint model and perform geometric verification and correction on key geometric feature points.

[0032] In one specific embodiment of this application, step S10 includes steps S101 to S103.

[0033] Step S101: Determine whether the deviation between the pointer rotation center (O) and the centroid (O') of the dial area view exceeds a threshold. If the deviation exceeds the threshold, replace the pointer rotation center (O) with the centroid (O') of the dial area. If the deviation does not exceed the threshold, keep the pointer rotation center (O) unchanged.

[0034] Step S102: Determine whether the deviation between the zero point (Z) and the full point (F) of the range, as well as the center of the circle fitted by the outer endpoints of other main scale lines and the pointer rotation center (O) exceeds the threshold. If the deviation exceeds the threshold, use the RANSAC algorithm to fit the best circle and correct the zero point (Z) or full point (F) of the range that deviates from the best circle so that it is projected onto the best circle.

[0035] Step S103: Verify whether the pointer tip (P) is within the annular region with the pointer rotation center (O) as the center and [R_min, R_max] as the radius.

[0036] In one specific embodiment of this application, after step S4, the method for automatic reading of single-pointer instrument for intelligent inspection of nuclear power plants further includes step S5.

[0037] Step S5: Output readings and confidence scores. The confidence score is a weighted average of three components: the average score of the KP-Net predicted keypoints, the conformity of the geometric validation, and the smoothness of the current reading compared to recent historical readings under process constraints. The formula for calculating the confidence score is: C = α * S_net + β * S_geo + γ * S_temp Where S_net, S_geo, and S_temp are the independent confidence scores for the three dimensions, and α, β, and γ are the corresponding configurable weight coefficients, satisfying α + β + γ = 1.

[0038] A second aspect of this application provides a computer device including a processor and a memory. The processor is used to execute a method for automatic reading of single-pointer instruments for intelligent inspection of nuclear power plants, as described in the first aspect of this application. The memory is used to store executable instructions of the processor.

[0039] A third aspect of this application provides a computer-readable storage medium having executable instructions stored thereon. When executed by a processor, the executable instructions implement a method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant, as described in the first aspect of this application.

[0040] The fourth aspect of this application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants according to the first aspect of this application.

[0041] The beneficial effects of the technical solution in this application are as follows: 1. Extremely high robustness: By detecting geometric feature points that are not sensitive to contamination and reflection (such as the outer endpoints of scale lines), instead of detecting fine-grained features that are easily disturbed (such as complete scale lines and numbers), the stability of the algorithm in harsh environments is fundamentally improved.

[0042] 2. High computational efficiency: It avoids complex full-scale segmentation or high-precision OCR processes. The algorithm is lightweight and suitable for deployment on mobile inspection equipment or edge computing units.

[0043] 3. Avoiding technical barriers: The core of this solution, "key point detection and geometric calculation without OCR", effectively avoids the mainstream technical routes that rely on "template matching + Hough transform" or "segmentation + OCR". Attached Figure Description

[0044] Figure 1 The diagram shown is a flowchart illustrating an automatic reading method for a single-pointer instrument used in intelligent inspection of a nuclear power plant, according to an embodiment of this application.

[0045] Figure 2 The diagram shows the included angles θ and Φ.

[0046] Figure 3 The diagram shown is a simplified flowchart of a single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants, provided by another embodiment of this application.

[0047] Figure 4 The diagram shows a flowchart of the process of using the KP-Net key point detection network to process the dial area.

[0048] Figure 5 The diagram shows the process of correcting the measurement from full-range F to calibration point F'. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] At least one embodiment of this application provides an automatic reading method for single-pointer instruments used in intelligent inspection of nuclear power plants. This method is particularly suitable for complex industrial environments with uneven lighting, glare, contamination, and irregular instrument installation angles. Figure 1As shown, the automatic reading method for single-pointer instruments used in intelligent inspection of nuclear power plants includes the following steps.

[0051] Step S1: Determine the dial area view of the instrument image on the single-pointer instrument.

[0052] Step S2: Detect four key geometric feature points on the dial area view. The four key geometric feature points include the pointer rotation center O, the pointer tip P, the zero point Z and the full point F of the range defined based on the centroid of the outer endpoint of the scale line.

[0053] It should be noted that the detection of the zero point Z and the full point F of the range does not rely on the optical character recognition of the dial numbers, but is determined by detecting the cluster of endpoints on the outer side of the scale lines and calculating their centroid.

[0054] The pointer rotation center O is also the pointer axis. The pointer tip P is also the geometric center of the pointer tip. The range zero point Z is also the range zero reference point, defined as the centroid of the set of endpoints of the outermost points of the range's starting scale line (no need to identify the number "0"). The range full-scale point F is also the range full-scale reference point, defined as the centroid of the set of endpoints of the range's ending scale line (no need to identify the number "full-scale value").

[0055] Step S3: Calculate the angle θ between vectors OP and OZ, and the angle Φ between vectors OF and OZ.

[0056] Step S4: Calculate the reading according to the preset instrument range [Range_min, Range_max] using the formula Value=Range_min+(θ / Φ)*(Range_max-Range_min).

[0057] The above embodiments demonstrate a method for achieving highly robust, adaptive single-pointer instrument readings in complex environments by detecting only a few key geometric feature points. This method eliminates the need for pre-stored instrument templates and OCR recognition, significantly improving its practicality and reliability in real-world nuclear power plant scenarios.

[0058] In at least one embodiment of this application, step S3 includes steps S31 to S33.

[0059] Step S31: Calculate vector OZ=V_z=O->Z, vector OF=V_f=O->F, and vector OP=V_p=O->P.

[0060] Step S32: Calculate the angle θ between vector OP and vector OZ = arccos((V_p·V_z) / (|V_p||V_z|)), and determine the quadrant according to the sign of the cross product to obtain θ in the range [0, 360°).

[0061] It should be noted that the angle between vector OP and vector OZ is the angle of the pointer relative to the zero point.

[0062] Step S33: Calculate the angle Φ between vector OF and vector OZ = arccos((V_f·V_z) / (|V_f||V_z|)).

[0063] It should be noted that the angle between vector OF and vector OZ is the effective angle of the measurement range. References for vectors OP, OZ, and OF, the included angle θ, and the included angle Φ are provided. Figure 2 As shown.

[0064] In at least one embodiment of this application, step S4 includes steps S41 and S42.

[0065] Step S41: Automatically obtain the preset instrument range [Range_min,Range_max].

[0066] For example, preset instrument ranges can be automatically obtained by linking to the nuclear power plant's equipment asset database, or they can be manually entered once during initialization, and then automatically retrieved afterward. It should be noted that the preset instrument ranges [Range_min, Range_max] represent the range of that instrument.

[0067] Step S42: Calculate the reading according to the preset instrument range [Range_min,Range_max] using the formula Value=Range_min+(θ / Φ)*(Range_max-Range_min).

[0068] In this embodiment of the application, the image OCR recognition process is completely avoided by automatically obtaining the preset instrument range [Range_min,Range_max].

[0069] In at least one embodiment of this application, step S1 includes steps S11 to S13.

[0070] Step S11: Obtain the instrument image.

[0071] Specifically, such as Figure 3 As shown, it can be used to input instrument images from nuclear power plants.

[0072] For example, instrument images can be captured using a fixed camera or a mobile robot.

[0073] Step S12: Preprocess the instrument image to eliminate the overshadowing of key geometric feature points by highlights.

[0074] Step S13: Locate the dial area based on the instrument image.

[0075] In at least one embodiment of this application, step S12 includes step S121.

[0076] Step S121: Perform illumination normalization and adaptive contrast enhancement on the instrument image, focusing on using a physical model-based reflective area detection and repair algorithm to eliminate the overshadowing of key point features by highlights.

[0077] In at least one embodiment of this application, step S13 includes steps S131 and S132.

[0078] Step S131: Use a lightweight object detection network or elliptic detection algorithm to locate the dial area.

[0079] For example, lightweight object detection networks such as YOLO-Fastest.

[0080] Step S132: When the dial area is detected to be elliptical, the perspective transformation matrix is ​​calculated by fitting the ellipse parameters to correct the dial area into a circular area image I_dial.

[0081] It should be noted that step S131 is the coarse positioning step of the dial area of ​​the instrument image, and step S132 is the perspective correction step.

[0082] In at least one embodiment of this application, step S2 includes step S21.

[0083] Step S21: The unified keypoint detection network KP-Net is used to directly process the dial area I_dial, and the normalized coordinates of four key geometric feature points are regressed. These four key geometric feature points include the pointer rotation center O, the pointer tip P, the zero point Z and the full-scale point F defined based on the centroid of the outer endpoints of the scale lines. The backbone network of KP-Net adopts a feature pyramid structure, and a "pointer tip attention module" is introduced into the prediction head. This module enhances the pointer tip features in the feature map by analyzing the gradient direction consistency and highlight symmetry of the pointer region, suppressing interference features caused by dirt or reflections.

[0084] refer to Figure 4 The specific data processing procedure of the Key Point Detection Network (KP-Net) is as follows.

[0085] KP-Net is an end-to-end multi-task neural network specifically designed for key point detection in instruments. Its data processing flow can be broken down into the following four stages.

[0086] Phase 1: Image Input and Multi-Scale Feature Extraction Input: The input received by the network is "input image (I_dial)", which is the image of the instrument panel from the frontal view after preprocessing and perspective correction, that is, the circular area image I_dial.

[0087] Feature Pyramid Construction: The input image is fed into a "feature extraction backbone network (lightweight FPN structure)". This backbone network first extracts feature maps of different depths, and then fuses them from top to bottom through a Feature Pyramid Network (FPN) structure, outputting a set of multi-scale feature maps. Figure 4 The labels are "C5", "C4", and "C3".

[0088] C5: The deepest feature map, containing the strongest semantic information (such as understanding the overall concepts of "meter" and "pointer"), but with the lowest spatial resolution.

[0089] C3: A shallower feature map that retains rich details and edge information (such as pointer tips and tick mark endpoints) and has the highest spatial resolution.

[0090] C4: Intermediate layer features between the two. This structure allows subsequent predictions to utilize both high-level semantics and low-level details, ensuring adaptability to instruments of different sizes and resolutions.

[0091] Phase 2: Parallel Processing of Multi-Task Prediction Heads The multi-scale features extracted from the backbone network are simultaneously fed into four parallel prediction branches, each responsible for learning different attributes of the key points.

[0092] Attention module branch (Pointer-Head) This is one of the core innovative modules of this network, also known as the pointer tip attention module. Instead of directly performing pixel-level segmentation, it focuses on salient features of the pointer region through a specific attention mechanism. Its internal operations include "gradient orientation focusing" and "highlight symmetry enhancement." This attention module analyzes the gradient orientation field of the feature map, enhancing linear features with consistent orientations (corresponding to the pointer edges), while simultaneously detecting symmetrical highlight patterns (common in cylindrical pointers), thereby accurately locking onto the contextual features of the "point P (point of the pointer)" in complex backgrounds.

[0093] Center-Head Prediction Branch The center point prediction branch is responsible for regressing the "pointer rotation center (O point)", and usually outputs a heatmap, where the peak position corresponds to the coordinates of point O. Its learning objective is to locate the physical rotation axis of the pointer.

[0094] Offset prediction branch (Offset-Head) This offset prediction branch predicts "sub-pixel-level offsets." Because the feature map is downsampled, coordinates regressed directly from the feature map contain quantization errors. This offset prediction branch learns the subtle offset of each keypoint from its approximate location to its true location to achieve sub-pixel-level localization accuracy, thereby improving the precision of the final coordinates.

[0095] Geometric verification auxiliary branch (Aux-Head): This geometric validation auxiliary branch is responsible for extracting "auxiliary features" for subsequent geometric validation. For example, it might output a preliminary map of the scale endpoints or features used to evaluate the geometric consistency of the point set. Figure 5 The subsequent spatial verification module shown provides preliminary, learnable clues rather than relying on fixed rules.

[0096] Phase 3: Feature Fusion and Coordinate Regression Feature fusion: The feature maps or preliminary prediction results output from the four prediction branches are fed into the "Feature Fusion and Coordinate Regression" module. This module is not a simple concatenation, but a learnable fusion network.

[0097] Fusion logic: Deeply fuse the pointer features focused by the Pointer-Head, the center position context provided by the Center-Head, the fine offset information of the Offset-Head, and the geometric auxiliary features of the Aux-Head. For example, the final localization of the pointer tip (P) will integrate its own attention features, its relative positional relationship with the predicted circle center (O), and the sub-pixel offset.

[0098] Phase 4: Unified Output Final Output: The fusion module outputs its normalized coordinates for four key points in a single pass through a lightweight regression layer. For example... Figure 4 As shown, the output format is: "Output: O, P, Z, F normalized coordinates".

[0099] Output meaning: The coordinates here are normalized values ​​relative to the input image I_dial (usually in the range [0,1]), representing the precise two-dimensional positions of the pointer rotation center (O), pointer tip (P), zero point (Z), and full-scale point (F). The coordinates of points Z and F are predictions obtained by the network directly learning the concept of the "centroid of the set of endpoints outside the scale lines," completely avoiding the OCR process.

[0100] In at least one embodiment of this application, reference is made to Figure 3 After step S2, the method for automatic reading of single-pointer instrument for intelligent inspection of nuclear power plants also includes step S10.

[0101] Step S10: Establish a geometric constraint model and perform geometric verification and correction on key geometric feature points.

[0102] It should be noted that step S10 is the step of verifying and adaptively correcting the spatial relationship of feature points.

[0103] In at least one embodiment of this application, step S10 includes steps S101 to S103.

[0104] Step S101: Determine whether the deviation between the pointer rotation center (O) and the centroid (O') of the dial area view exceeds a threshold. If the deviation exceeds the threshold, replace the pointer rotation center (O) with the centroid (O') of the dial area. If the deviation does not exceed the threshold, keep the pointer rotation center (O) unchanged.

[0105] Step S102: Determine whether the deviation between the zero point (Z) and the full point (F) of the range, as well as the center of the circle fitted by the outer endpoints of other main scale lines and the pointer rotation center (O) exceeds the threshold. If the deviation exceeds the threshold, use the RANSAC algorithm to fit the best circle and correct the zero point (Z) or full point (F) of the range that deviates from the best circle so that it is projected onto the best circle.

[0106] Step S103: Verify whether the pointer tip (P) is within the annular region with the pointer rotation center (O) as the center and [R_min, R_max] as the radius.

[0107] It should be noted that R_min and R_max are adaptively determined based on the radius of the dial area. Steps S101 to S103 are all geometric verification rules. Step S101 is a concentricity verification rule (e.g., Figure 5 Rule 1). Step S102 is the common arc verification rule (e.g.) Figure 5 Rule 2). Step S103 is the pointer validity verification rule.

[0108] In the above embodiments, by adopting a unified KP-Net and geometric verification rules, different models and ranges of instruments can be adaptively processed, thereby enabling the automatic reading method of single-pointer instruments for intelligent inspection of nuclear power plants to have strong adaptive capabilities, and without the need to establish image templates for each type of instrument.

[0109] In at least one embodiment of this application, reference is made to Figure 3 After step S4, the method for automatic reading of single-pointer instruments for intelligent inspection of nuclear power plants also includes step S5.

[0110] Step S5: Output the reading and confidence level. The confidence level is composed of three weighted components: the average score of the key points predicted by KP-Net, the compliance with geometric validation, and the smoothness of the current reading compared with recent historical readings under process constraints.

[0111] In this embodiment, while outputting the final reading Value, a comprehensive confidence level C (range [0, 1]) is also simultaneously output. This confidence level is not a simple judgment based on a single indicator, but rather a multi-dimensional evaluation system that integrates perceived reliability, physical rationality, and process continuity to quantify the credibility of the automatic reading result. Its calculation formula is as follows: C = α * S_net + β * S_geo + γ * S_temp Here, S_net, S_geo, and S_temp are the independent confidence scores for the three dimensions, and α, β, and γ are the corresponding configurable weight coefficients, satisfying α + β + γ = 1. The default recommended weights are α=0.2, β=0.5, and γ=0.3. This setting emphasizes the core importance of physical consistency and process smoothness, which aligns with the safety philosophy of high-reliability nuclear power applications. The calculation methods for each item are detailed below.

[0112] 1. KP-Net predicted keypoint average score (S_net): perceptual confidence This score represents the degree to which the Keypoint Detection Network (KP-Net) is confident in its own prediction results.

[0113] Data Source: During the forward propagation of the KP-Net network, in addition to outputting the normalized coordinates of the four keypoints (O, P, Z, F), its output layer synchronously generates a confidence score s_i (i ∈ {O, P, Z, F}) for each point, with a value range of [0, 1]. This score originates from the network's Softmax output or a specific confidence branch, reflecting the probability that the model correctly determines the location of the point based on the features of the input image.

[0114] Calculation method: S_net takes the arithmetic mean of the confidence scores of the four points. S_net = (s_O + s_P + s_Z + s_F) / 4 Technical implications: S_net directly evaluates the visual quality (such as sharpness and illumination) and discriminability of target features of the current input image I_dial. This value will decrease significantly when the image is severely blurred or contains strong interference.

[0115] 2. Geometric verification compliance (S_geo): Confidence level of physical plausibility. This fractional quantification network's preliminary prediction results agree well with the inherent physical and geometric constraints of the instrument.

[0116] Data source: Calculated based on the results of step S10 (feature point spatial relationship verification and correction). The core basis is the optimal circle C (center O_fit, radius R_fit) fitted by the RANSAC algorithm and the set of all correctly judged interior points (including the corrected Z, F' points and other tick mark endpoints).

[0117] Calculation method: Calculate the average fitting residual: Calculate the average radial absolute error E_avg from all interior points to the fitted circle C: E_avg = mean( abs( dist(P_i, O_fit) - R_fit ) ), where P_i is the coordinate of the i-th interior point.

[0118] Normalization to conformity score: Set a maximum permissible error threshold T_max based on the dial pixel radius adaptively or empirically (e.g., R_fit * 0.02, i.e., 2% of the radius). S_geo is calculated by the following formula: S_geo = max( 0, 1 - (E_avg / T_max) ). When all interior points are perfectly on the circle (E_avg = 0), S_geo = 1; when the average error reaches or exceeds the threshold (E_avg ≥ T_max), S_geo = 0.

[0119] Technical implications: S_geo evaluates whether the detection results conform to the basic physical fact that "the dial is circular and the key points are concyclic". Even if S_net is high, if S_geo is too low, it indicates that the network prediction is geometrically inconsistent and may have serious false positives.

[0120] 3. Historical reading smoothness (S_temp): Confidence level of process consistency This score uses the knowledge that changes in nuclear power plant instrument parameters are typically continuous and gradual to assess the timing appropriateness of the current reading.

[0121] Data source: The system maintains a recent historical reading queue of length N (e.g., N=5) {Value_t-1, Value_t-2, ..., Value_t-N}. The current reading is Value_t.

[0122] Calculation method: Generate short-term forecasts: Use a first-order linear extrapolation or simple moving average model to predict the current reading, Value_pred, based on historical data. For example, using first-order extrapolation: Value_pred = Value_t-1 + (Value_t-1 - Value_t-2).

[0123] Calculate the relative deviation: Calculate the relative absolute deviation D_rel between the current reading and the predicted value, and normalize it to the instrument range [Range_min, Range_max]: D_rel = abs(Value_t - Value_pred) / (Range_max - Range_min) Normalization to smoothness score: A maximum permissible instantaneous relative change threshold D_max is set based on process knowledge (e.g., 0.5% of the range, i.e., 0.005). S_temp is calculated by the following formula: S_temp = max(0, 1 - (D_rel / D_max)). When the reading perfectly matches the trend (D_rel = 0), S_temp = 1; when a drastic abnormal jump occurs (D_rel ≥ D_max), S_temp = 0.

[0124] Technical implications: S_temp introduces prior knowledge of process control. A reading that appears visually and geometrically reasonable, but deviates drastically from recent operating trends, may be an identification error or an early sign of a real equipment malfunction, and should be marked with low confidence.

[0125] 4. Weighted Fusion and Decision-Making Applications The three scores mentioned above provide relatively independent and credible evidence from different dimensions. After weighted fusion to obtain the overall confidence level C, the system will make intelligent decisions based on this: High reliability output: When the C value is higher than the preset action threshold (e.g., C_threshold = 0.85), the reading is considered highly reliable and the result is directly used for recording, monitoring and data analysis.

[0126] Low reliability handling: When the C value is below the threshold, the anomaly handling mechanism is triggered. The system can automatically mark the data as "requiring manual review" or issue a "re-collection request" instruction to the inspection robot.

[0127] The following example illustrates a method for automatic reading of a single-pointer instrument for intelligent inspection of a nuclear power plant, using the inspection image processing of a pressure gauge (range 0-10.0 MPa) in the main feedwater system of a nuclear power plant.

[0128] 1. Image input (corresponding to step S11): Obtain an original instrument image with reflections from the top light.

[0129] 2. Preprocessing and Positioning (corresponding to steps S12 and S13): Perform reflection repair and enhancement. Position the dial using a lightweight detector and correct it to obtain the dial area view I_dial.

[0130] 3. Keypoint Detection (corresponding to step S2): Input the dial area view I_dial into the trained keypoint detection network KP-Net. The keypoint detection network KP-Net outputs the normalized coordinates of four points, which are then calculated as follows: O: (320.5, 240.3); P: (320.1, 120.7) / / The pointer is pointing upwards; Z: (200.2, 240.8) / / Assuming the pointer rotation center (O) is on the left; F: (320.6, 100.5) / / Assuming the 10MPa point is slightly off-center at the top.

[0131] 4. Geometric Verification and Correction (corresponding to step S10): The centroid (O') of the dial area view is calculated to be (320.0, 240.0), which is close to the pointer rotation center (O), thus verification is successful. (Reference) Figure 5 It was found that the full-scale point (F) was not concyclic with other scale endpoints. The circle was fitted using RANSAC and corrected to the calibration point F':(319.9, 99.9).

[0132] 5. Reading Calculation (corresponding to steps S3, S4, and S5): The calculated θ (angle between vectors OP and OZ) is approximately 90.1 degrees. The calculated Φ (angle between vector OF' and OZ) is approximately 270.0 degrees. The system database indicates that the instrument's tag number range is [0, 10.0]. Final reading: Value = 0 + (90.1 / 270.0) * 10.0 ≈ 3.34 MPa.

[0133] Output: Output reading 3.34 MPa, overall confidence level C=0.95.

[0134] The calculation process for a comprehensive confidence level C=0.95: Current reading: Value_t = 3.34 MPa Recent historical readings: Assume the reading at the previous time step Value_{t-1} = 3.30 MPa, and the reading at the two time steps before that Value_{t-2} = 3.28 MPa (for trend prediction).

[0135] Keypoint detection scores: Assume that the confidence scores of the four keypoints (O, P, Z, F) output by KP-Net are 0.92, 0.95, 0.90, and 0.93, respectively.

[0136] Geometric verification results: Assuming that after RANSAC fitting, the average radial error of all interior points is E_avg = 0.4 pixels, and the error threshold T_max = 10 pixels is set according to the size of the dial image.

[0137] Process parameters: Instrument range = 10.0 MPa, and the maximum allowable instantaneous change in reading D_max = 0.05 MPa (i.e. 0.5% of the range) is set according to process knowledge.

[0138] Weighting coefficients: The default weights are α=0.2, β=0.5, and γ=0.3.

[0139] Step-by-step calculation process Step 1: Calculate the perceptual confidence score (S_net) S_net is the average of the four keypoint confidence scores output by KP-Net: S_net = (0.92 + 0.95 + 0.90 + 0.93) / 4 = 3.70 / 4 = 0.925 Step 2: Calculate the geometric fit (S_geo) S_geo is calculated based on the ratio of the average fitting error to the maximum allowable error: S_geo = max(0, 1 - (E_avg / T_max)) = max(0, 1 - (0.4 / 10)) = 1 - 0.04 = 0.96 Step 3: Calculate the time smoothness (S_temp) Predicted values ​​are generated using first-order linear extrapolation. Value_pred = Value_{t-1} + (Value_{t-1} - Value_{t-2}) = 3.30 + (3.30 - 3.28) = 3.32 MPa Calculate the relative deviation: D_rel = abs(Value_t - Value_pred) / Range = abs(3.34 -3.32) / 10.0 = 0.02 / 10.0 = 0.002 Calculate the smoothness score: S_temp = max(0, 1 - (D_rel / D_max)) = max(0, 1 -(0.002 / 0.05)) = 1 - 0.04 = 0.96 Step 4: Weighted fusion to calculate the overall confidence level (C) C = α * S_net + β * S_geo + γ * S_temp = 0.2*0.925 + 0.5*0.96 + 0.3*0.96 C = 0.185 + 0.48 + 0.288 = 0.953 ≈ 0.95 At least one embodiment of this application also provides a computer device including a processor and a memory. The processor is used to execute a single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants provided in any of the above embodiments of this application. The memory is used to store executable instructions of the processor, such as application programs. There can be one or more processors. The application programs stored in the memory can include one or more modules, each corresponding to a set of instructions. Furthermore, the processor is configured to execute instructions to perform the above-described single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants.

[0140] The computer device may also include a power supply component configured for power management, a wired or wireless network interface configured to connect the computer device to a network, and an input / output (I / O) interface. The computer device can operate on an operating system stored in memory, such as Windows Server. TM Mac OSX TM Unix TM Linux TM FreeBSD TM Or similar.

[0141] At least one embodiment of this application also provides a computer-readable storage medium storing executable instructions for a computer thereon. When executed by a processor, the executable instructions implement a method for automatic reading of a single-pointer instrument for intelligent inspection of a nuclear power plant, as provided in any of the above embodiments of this application.

[0142] A non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by the processor of the computer device, enables the computer device to perform the aforementioned method for automatically reading single-pointer instruments for intelligent inspection of nuclear power plants. This method for automatically reading single-pointer instruments for intelligent inspection of nuclear power plants is executed by an agent program.

[0143] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0144] At least one embodiment of this application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements a single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants provided in any of the above embodiments of this application.

[0145] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a computer program product. This computer program 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 steps of the single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants according to various embodiments of this application. The aforementioned storage medium includes various media capable of storing program verification codes, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0146] It should be noted that the combination of the technical features in the embodiments of this application is not limited to the combination methods described in the embodiments of this application or the combination methods described in specific embodiments. All technical features described in this application can be freely combined or combined in any way, unless they contradict each other.

[0147] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the term "comprising" only indicates that it includes the explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0148] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications or equivalent substitutions made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for automatic reading of single-pointer instruments for intelligent inspection in nuclear power plants, characterized in that, include: Step S1: Determine the dial area view of the instrument image on the single-pointer instrument; Step S2: Detect four key geometric feature points on the dial area view. The four key geometric feature points include the pointer rotation center (O), the pointer tip (P), the zero point (Z) of the range defined based on the centroid of the outer endpoint of the scale line, and the full-scale point (F). Step S3: Calculate the angle θ between vector OP and vector OZ, and the angle Φ between vector OF and vector OZ; Step S4: Calculate the reading according to the preset instrument range [Range_min, Range_max] using the formula Value=Range_min+(θ / Φ)*(Range_max-Range_min).

2. The method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant according to claim 1, characterized in that, Step S1 includes: Step S11: Acquire instrument image; Step S12: Preprocess the instrument image to eliminate the overshadowing of key geometric feature points by highlights; Step S13: Locate the dial area based on the instrument image.

3. The method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant according to claim 1, characterized in that, Step S12 includes: Step S121: Perform illumination normalization and adaptive contrast enhancement on the instrument image, focusing on using a physical model-based reflective area detection and repair algorithm to eliminate the overshadowing of key point features by highlights. And / or, step 13 includes: Step S131: Use a lightweight object detection network or elliptic detection algorithm to locate the dial area; Step S132: When the dial area is detected to be elliptical, the perspective transformation matrix is ​​calculated by fitting the ellipse parameters to correct the dial area into a circular area image.

4. The method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant according to claim 1, characterized in that, Step S2 includes: Step S21: A unified key point detection network (KP-Net) is used to directly process the dial area, and the normalized coordinates of the following four key geometric feature points are regressed. The four key geometric feature points include the pointer rotation center (O), the pointer tip (P), the zero point (Z) and the full point (F) of the range defined based on the centroid of the outer endpoint of the scale line. The backbone network of the key point detection network adopts a feature pyramid structure, and a pointer tip attention module is introduced into the prediction head. The pointer tip attention module strengthens the pointer tip features in the feature map by analyzing the gradient direction consistency and highlight symmetry of the pointer area, and suppresses the interference features caused by dirt or reflection.

5. The method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant according to claim 1, characterized in that, Following step S2, the following is also included: Step S10: Establish a geometric constraint model and perform geometric verification and correction on key geometric feature points.

6. The method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant according to claim 1, characterized in that, Step S10 includes: Step S101: Determine whether the deviation between the pointer rotation center (O) and the centroid (O') of the dial area view exceeds the threshold. If the deviation exceeds the threshold, replace the pointer rotation center (O) with the centroid (O') of the dial area. If the deviation does not exceed the threshold, keep the pointer rotation center (O) unchanged. Step S102: Determine whether the deviation between the zero point (Z) and the full point (F) of the range, as well as the center of the circle fitted by the outer endpoints of other main scale lines and the pointer rotation center (O) exceeds the threshold. If the deviation exceeds the threshold, use the RANSAC algorithm to fit the best circle and correct the zero point (Z) or full point (F) of the range that deviates from the best circle so that it is projected onto the best circle. Step S103: Verify whether the pointer tip (P) is within the annular region with the pointer rotation center (O) as the center and [R_min, R_max] as the radius.

7. A method for automatic reading of a single-pointer instrument for intelligent inspection in a nuclear power plant according to any one of claims 1 to 6, characterized in that, Following step S4, the following is also included: Step S5: Output the reading and confidence score. The confidence score is composed of three weighted components: the average score of the keypoints predicted by KP-Net, the conformity of the geometric validation, and the smoothness of the current reading compared to recent historical readings under process constraints. The formula for calculating the confidence score is: C = α * S_net + β * S_geo + γ * S_temp Where S_net, S_geo, and S_temp are the independent confidence scores for the three dimensions, and α, β, and γ are the corresponding configurable weight coefficients, satisfying α + β + γ = 1.

8. A computer device, characterized in that, include: A processor for executing a single-pointer instrument automatic reading method for intelligent inspection of a nuclear power plant, as described in any one of claims 1 to 7. as well as Memory for storing the executable instructions of the processor.

9. A computer-readable storage medium having executable instructions stored thereon, characterized in that, When the executable instructions are executed by the processor, they implement the single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants according to any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the single-pointer instrument automatic reading method for intelligent inspection of nuclear power plants according to any one of claims 1 to 7.

Images