Multi-source data screening method

Abstract

The invention discloses a multi-source data screening method. Comprising the following steps: S1, acquiring virtual ideal data associated with a target object, and acquiring real observation data for the target object acquired by an acquisition terminal in real time; s2, calculating a first quality parameter representing the quality of the endogenous information based on the real observation data; s3, comparing the real observation data with the virtual ideal data, and calculating a second quality parameter representing the difference between the real observation data and the virtual ideal data; s4, generating a deviation field signal based on the first quality parameter and the second quality parameter; s5, based on the deviation field signal, generating and outputting a control instruction to reduce deviation; according to the method, endogenous and relative quality parameters are linked and fused, and a step flow from accurate evaluation to closed-loop guidance is constructed; the accuracy and efficiency of data acquisition are improved through a visual AR instruction, and long-term reliable operation is ensured through a double closed-loop adjustment mechanism of external user guidance and internal algorithm self-adaption.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and more specifically, relates to a method for filtering multi-source data. Background Technology

[0002] In the current wave of global digital transformation, industries are increasingly reliant on field data, and data-driven decision-making has become a core engine for improving the efficiency of industrial production, engineering management, and public safety. Especially in applications requiring precise perception of the physical world, extracting high-quality, high-value information from massive, multi-source field data has become a fundamental and critical technological requirement. Therefore, improving data quality and collection efficiency from the source of data acquisition to ensure that the optimal information enters the decision-making system is becoming a focus of continuous attention and development in this technological field.

[0003] To achieve the above goals, existing technologies typically employ two mainstream data quality assessment and screening strategies. The first is assessment based on intrinsic data attributes, focusing on the inherent quality of the data itself, such as image sharpness, signal-to-noise ratio, or information entropy. The second is assessment based on external benchmark comparison, which quantifies the degree of deviation between field-collected data and a pre-defined, idealized digital model or standard sample. However, applying these techniques to complex field environments still faces some unresolved technical challenges:

[0004] In practical applications, the two technical approaches described above often operate independently. While assessment methods based on endogenous attributes can determine the quality of the data itself (e.g., whether a photograph is clear), they cannot ascertain whether the data content meets the fundamental requirements of the task (e.g., whether the key parts of the equipment were photographed correctly). Conversely, methods based on external benchmarks, although defining ideal targets, may become inaccurate or even misleading when there are inherent differences or dynamic changes between the real-world environment and the ideal model. This information isolation between assessment dimensions makes it difficult for the system to form a comprehensive, accurate, and environmentally adaptable quality judgment.

[0005] Existing technologies, after completing quality assessments, typically output a quantified quality score or a simple "pass / fail" judgment. While this form of feedback informs operators whether the current data is "good" or "bad," it fails to provide specific, actionable guidance on "how to improve." For example, when the system indicates an unsuitable acquisition angle, operators still need to repeatedly try and find the correct angle. This lack of clear corrective instructions means that the data acquisition process remains highly dependent on the operator's personal experience, making it difficult to achieve standardized, efficient, closed-loop quality control.

[0006] To address the aforementioned challenges, the key to this invention lies in breaking down the barriers between different quality assessment dimensions and establishing a technological bridge that can transform abstract quality assessment results into concrete guiding behaviors. Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide a multi-source data filtering method. This method generates a unified quality parameter through a fusion model dynamically modulated by relative differences, and constructs a deviation field that includes potential, gradient, and time dynamics based on this parameter. Finally, the deviation field is used to drive a control system that integrates external user guidance and internal model adaptation.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-source data filtering method, comprising the following steps:

[0009] S1. Obtain virtual ideal data associated with the target object, as well as real observation data of the target object collected in real time by the acquisition terminal;

[0010] S2. Based on real observation data, calculate the first quality parameter that characterizes the quality of endogenous information;

[0011] S3. Compare the real observation data with the virtual ideal data, and calculate the second quality parameter that characterizes the difference between the real observation data and the virtual ideal data;

[0012] S4. Based on the first and second mass parameters, generate a deviation field signal that characterizes the multi-dimensional deviation between the real observation data and the ideal acquisition state.

[0013] S5. Based on the deviation field signal, generate and output control commands to guide the acquisition terminal to adjust its acquisition behavior in order to reduce deviation.

[0014] Furthermore, the steps in S2 for calculating the first mass parameter based on real observation data are as follows:

[0015] Calculate the visual information entropy of the real observation data; and compare the real-time spatial pose of the acquisition terminal with the preset digital model to calculate the geometric-semantic consistency score; the first quality parameter is determined based on the visual information entropy and the geometric-semantic consistency score.

[0016] Furthermore, in S3, the actual observed data is compared with the virtual ideal data, and the steps for calculating the second mass parameter are as follows:

[0017] Real observation data and virtual ideal data are respectively input into a pre-trained deep neural network model to extract their respective high-dimensional feature vectors; the projection distance between the high-dimensional feature vectors in the feature space is calculated, and the projection distance is used as the second quality parameter.

[0018] Furthermore, the specific steps for calculating the fusion weight of the first and second quality parameters are as follows:

[0019] The first quality parameter is adjusted by weighting the ratio of visual information entropy and geometric-semantic information, and the weight is determined by a Gaussian decay function with feature-space projection distance as the independent variable.

[0020] The first quality parameter and the second quality parameter are weighted and fused to obtain the final unified data quality parameter.

[0021] Furthermore, the deviation field signal in S4 is generated from the real-virtual deviation field; the deviation field signal is a multidimensional vector field, which includes a scalar potential field, a vector gradient field, and a deviation field time gradient scalar.

[0022] Furthermore, the steps for generating the deviation field signal are as follows:

[0023] A scalar potential field representing the magnitude of the deviation at each location is generated by unifying data quality parameters; based on the scalar potential field, a vector gradient field is calculated to indicate the direction of correction.

[0024] The steps to generate a scalar potential field include:

[0025] Based on the pre-defined semantic saliency map, the deviations at different positions are weighted and calculated.

[0026] Based on the change of the scalar potential field in continuous time frames, the deviation field time gradient scalar is calculated to characterize the rate of deviation change.

[0027] Furthermore, the control commands generated in S5 are augmented reality visual guidance signals, which are superimposed on the display interface of the acquisition terminal.

[0028] Furthermore, S5 includes dynamically adjusting the weights or thresholds used to calculate the first or second mass parameter based on the deviation field signal.

[0029] Furthermore, the control commands include external control commands for guiding the acquisition terminal and internal control commands for adjusting internal algorithm parameters;

[0030] The external control commands are augmented reality visual signals, which include macroscopic pose guidance elements generated by global geometric deviation based on the deviation field; and microscopic view guidance elements generated by modified gradient field based on the deviation field.

[0031] The internal control command updates the value of the integral deviation accumulator based on the deviation field signal.

[0032] Furthermore, the visual properties of augmented reality visual signals are dynamically adjusted based on the bias field time gradient scalar;

[0033] Internal control commands for adjusting internal algorithm parameters are generated only when the value of the integral deviation accumulator exceeds a preset adaptive trigger threshold.

[0034] The technical effects and advantages of this invention are as follows:

[0035] This invention utilizes a second quality parameter to dynamically adjust the fusion weights of each component within the first quality parameter, enabling real-time self-optimization based on the degree of conformity between reality and virtuality. When the physical reality and the digital model are highly consistent, geometric constraints are trusted and emphasized. However, when there are significant differences between the two, the reliance on unreliable models is intelligently reduced, and the information content of the image itself is trusted more. This fundamentally solves the problem of misjudgment that may occur in static evaluation models in complex and changing environments, ensuring that the final unified data quality parameters can more accurately and reliably reflect the true value of the currently collected data.

[0036] This invention further elevates the quality parameter into a multi-dimensional deviation field containing scalar potential, vector gradient, and time gradient; it successfully transforms an abstract quality problem into a control problem with clear physical meaning and operability guidance; the vector gradient directly provides precise instructions to the acquisition terminal, while the time gradient dynamically reflects the trend of deviation changes and is used to trigger emergency warnings; by transforming the deviation field signal into intuitive augmented reality visual guidance elements, it provides unprecedented real-time, closed-loop guidance for on-site operators;

[0037] This invention constructs a higher-dimensional internal adaptive closed loop by introducing an integral deviation accumulator and an adaptive triggering mechanism. It can go beyond the instantaneous feedback of a single acquisition and instead perform long-term monitoring of persistent deviations that are difficult to correct through user behavior. When the accumulated deviation exceeds a preset threshold, it can autonomously trigger the adjustment of internal algorithm parameters to adapt to the current special and possibly unforeseen environment. The dual closed-loop architecture of externally guiding users and internally optimizing itself can co-evolve with the environment and users, has high robustness and autonomous learning capabilities, and ensures that it can maintain long-term stability and optimal performance under various complex, non-ideal and even extreme working conditions.

[0038] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the steps provided by the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0041] like Figure 1 As shown in the figure, an embodiment of the present invention provides a multi-source data filtering method, which includes the following steps:

[0042] S1. Obtain virtual ideal data associated with the target object, as well as real observation data of the target object collected in real time by the acquisition terminal;

[0043] In this embodiment, the key parameters of the virtual ideal data are defined as follows:

[0044] The digital model serves as a geometric and semantic benchmark, and is a BIM or CAD model file containing complete information about the target object. The digital model contains: a precise 3D mesh or geometric representation of the constructed entity, defining the shape, size, and spatial relationships of all components; semantic labels for each 3D component (in this embodiment: "pump-A38", "valve-V5", "weld-S12"); and basic material classification information.

[0045] The calibration image set serves as a benchmark for appearance and lighting, consisting of high-resolution, realistic images taken from multiple different viewpoints around the target object. Acquiring the calibration image set requires meeting the following conditions: high image quality, clear focus, and normal exposure; comprehensive coverage of shooting viewpoints, allowing observation of all key parts of the target object without blind spots; and environmental conditions (such as weather and lighting) during shooting should be as close as possible to ideal observation conditions.

[0046] For each image in the calibration image set, the camera's intrinsic and extrinsic parameters and pose must be precisely calculated. These parameters include: an intrinsic parameter matrix, which describes the camera's internal optical characteristics (focal length, principal point); and an extrinsic parameter (pose), a six-degree-of-freedom parameter that describes the camera's precise position and orientation (rotation) in the world coordinate system during shooting.

[0047] The intrinsic and extrinsic parameters are calculated in batches by running the motion reconstruction algorithm on the calibration image set.

[0048] The ideal observation path is a "script" generated from virtual data. It is a pre-planned virtual camera motion trajectory consisting of a series of discrete six-degree-of-freedom pose points. The virtual camera motion trajectory represents the best route and the best observation angle for experts to conduct equipment inspections or data acquisition. The virtual camera motion trajectory can be visualized and set in the digital model.

[0049] The parameter symbols of the neural radiation field model are: It is a deep neural network whose core function is to learn a mapping from spatial coordinates to pixel values.

[0050] The calculation logic is as follows: Neural radiation field model The input is a five-dimensional vector containing the coordinates of a point in three-dimensional space. and a two-dimensional observation direction Its output is a four-dimensional vector containing the color values ​​of points in space. and volume density Volume density It represents the probability that a point in space contains matter and blocks light;

[0051] The parameter symbol for the virtual ideal data packet is: It is a structured data set that corresponds one-to-one with each pose point in the ideal observation path; virtual ideal data packet. It must contain at least:

[0052] Ideal rendering image It is a high-fidelity color image;

[0053] Ideal semantic graph It is the ideal rendered image A pixel-aligned semantic segmentation map, where the color or ID value of each pixel corresponds to a semantic label from a digital model;

[0054] Ideal depth map It is the ideal rendered image A pixel-aligned depth map, where the value of each pixel represents the depth of a spatial point from the virtual camera;

[0055] Ideal camera pose It generates virtual ideal data packets. The pose of the six-DOF camera used;

[0056] The computation process in this embodiment is divided into two main stages: the offline training stage and the data generation stage.

[0057] Phase 1: Model Training Phase (Offline)

[0058] Data preprocessing and alignment: Run the SfM algorithm on the calibration image set to calculate the accurate camera intrinsic and extrinsic parameters and pose for each image, and generate a sparse 3D point cloud.

[0059] The coordinate system of the digital model is rigidly aligned with the 3D point cloud output by the SfM algorithm. This alignment is accomplished by manually selecting at least three pairs of points with the same name or by using a point cloud registration algorithm. This ensures that the geometric truth and the appearance truth are in the same spatial reference system.

[0060] Guided neural radiation field model Training:

[0061] Initialize the neural radiation field model ; Enter the training loop, and in each iteration, randomly select an image and its corresponding camera pose from the calibration image set;

[0062] A batch of pixels are randomly sampled from the image, and a corresponding three-dimensional spatial ray is generated for each pixel based on the camera pose and intrinsic parameters.

[0063] Layered sampling is performed along the 3D spatial light rays to generate 3D spatial points; the position and light direction of each 3D spatial point are used as inputs and fed into the neural radiation field model. Query the color value of a spatial point and volume density ;

[0064] Using volumetric rendering technology, the color values ​​of all sampled points on the light source in three-dimensional space are... and volume density The integral is used to calculate the predicted color value of the pixel corresponding to the light ray in three-dimensional space;

[0065] Calculate the loss function for the three parts:

[0066] Calculate the difference between the predicted color value and the true color value of the corresponding pixel in the calibrated image;

[0067] For each sampling point on the 3D ray, query its position in the aligned digital model; if the spatial point is located inside the surface of the digital model, then its predicted volume density... The predicted volume density should be relatively high; if it is located in free space outside the digital model, then its predicted volume density is relatively high. The value should be low; based on this principle, a loss is calculated to penalize density distributions that do not conform to the authoritative geometric model.

[0068] For each sampling point on the light ray, its predicted color should be consistent with the typical color of its semantic region (obtained statistically from the calibration image set); based on this principle, a loss is calculated to penalize rendering results where the color does not match the semantics.

[0069] The three losses are weighted and summed to obtain the total loss, and the neural radiation field model is updated using the backpropagation algorithm. The weights;

[0070] Repeat the operation until the neural radiation field model is established. Convergence occurs, and the total loss no longer decreases significantly.

[0071] Phase Two: Data Generation Phase (Offline or On-Demand)

[0072] Load the trained neural radiation field model And the aligned digital model;

[0073] Traversing the ideal path:

[0074] Ideal camera pose for traversing every six-DOF pose point along the ideal observation path ;

[0075] For the current ideal camera pose Set up a virtual camera;

[0076] Composite data rendering:

[0077] For each pixel within the virtual camera's view frustum, a ray is generated, and a trained neural radiation field model is used. Volume rendering technology calculates the color of each pixel and ultimately synthesizes a complete and ideal rendered image. ;

[0078] For each pixel within the virtual camera's view frustum, a ray is generated, and its intersection with the digital model is determined through ray projection. The semantic label of the intersection point is the semantic value of the pixel, and finally, a complete ideal semantic map is synthesized. ;

[0079] In the projection of light rays, the depth value of the intersection point is the depth value of that pixel, and finally, a complete and ideal depth map is synthesized. ;

[0080] Data packaging and storage:

[0081] Ideal camera pose for the current pose And generate ideal rendered images Ideal semantic graph Ideal depth map The parameter symbol for the common encapsulation into a virtual ideal data packet is: ;

[0082] The parameter symbol of the virtual ideal data packet is... Store in the database and use the ideal camera pose Alternatively, the path sequence number can be used as an index for quick querying and retrieval during the real-time acquisition phase;

[0083] S2. Based on real observation data, calculate the first quality parameter that characterizes the quality of endogenous information;

[0084] The steps for calculating the first mass parameter based on real observation data in S2 are as follows:

[0085] Calculate the visual information entropy of the real observation data; and compare the real-time spatial pose of the acquisition terminal with the preset digital model to calculate the geometric-semantic consistency score; the first quality parameter is determined based on the visual information entropy and the geometric-semantic consistency score.

[0086] By calculating the pixel grayscale distribution and texture complexity of image data, a visual information entropy is obtained to quantify the richness and clarity of information. The six-degree-of-freedom spatial pose is acquired through the inertial measurement unit and augmented reality engine built into the acquisition terminal, and then spatially aligned in real time with the building information model or digital twin model on the server. The current acquisition perspective is used to calculate whether the key semantic parts of the target object can be effectively observed, thus obtaining a geometric-semantic consistency score. The visual information entropy and the geometric-semantic consistency score together constitute the first quality parameter, which can not only evaluate the pixel-level quality of the image, but also judge the effectiveness of the acquisition behavior from the business logic level. By combining the content information and spatial intent, the intrinsic value of the data is accurately and multidimensionally quantified.

[0087] S3. Compare the real observation data with the virtual ideal data, and calculate the second quality parameter that characterizes the difference between the real observation data and the virtual ideal data;

[0088] In S3, the steps for comparing real observation data with virtual ideal data and calculating the second mass parameter are as follows:

[0089] Real observation data and virtual ideal data are respectively input into a pre-trained deep neural network model to extract their respective high-dimensional feature vectors; the projection distance between the high-dimensional feature vectors in the feature space is calculated, and the projection distance is used as the second quality parameter;

[0090] Based on the digital model of the target object and the ideal observation path, a series of virtual perfect images or video frames are generated using technologies such as neural rendering as virtual ideal data. When the real observation data of the acquisition terminal is received, the real image and the virtual image corresponding in time or space are simultaneously sent to a powerful visual feature extraction model (CLIP or DINOv2 in this embodiment). The visual feature extraction model can transform the image into a deep semantic feature vector that is not sensitive to illumination and small angle changes. By calculating the cosine similarity or Euclidean distance between these two vectors, a feature space projection distance that can accurately quantify the difference between the real data and the standard answer in terms of content and semantics is obtained, which is the second quality parameter. This makes the screening and guidance standards no longer dependent on human-set static rules that may have biases.

[0091] The specific steps for calculating the fusion weight of the first and second quality parameters are as follows:

[0092] The first quality parameter is adjusted by weighting the ratio of visual information entropy and geometric-semantic information, and the weight is determined by a Gaussian decay function with feature-space projection distance as the independent variable.

[0093] The first quality parameter and the second quality parameter are weighted and fused to obtain the final unified data quality parameter;

[0094] In this embodiment, the key parameters are defined as follows:

[0095] The parameter symbol for visual information entropy is: It is used to quantify the richness of information, texture complexity and clarity contained in real observation images; its physical meaning comes from Shannon entropy in information theory. A high entropy value represents rich image details, accurate focus and no obvious overexposure or underexposure.

[0096] The computational logic involves: converting the input color real-world observed image into an 8-bit grayscale image; counting the number of pixels appearing at each of the 256 grayscale levels from 0 to 255 in the grayscale image, and calculating the probability of each grayscale level; according to the definition of Shannon information entropy, multiplying the probability of each grayscale level by its base-2 logarithm, summing the results for all grayscale levels, and taking the negative value to obtain the original value of the image's visual information entropy; to facilitate subsequent fusion, the original value is normalized, linearly mapping it to the [0,1] interval; the normalization method is to divide the calculated original value of visual information entropy by the theoretical maximum entropy value at the image bit depth (in this embodiment, for an 8-bit grayscale image, the theoretical maximum entropy value is 8);

[0097] The parameter notation for the geometric-semantic consistency score is: It is used to quantify the degree of matching between the spatial pose (position and attitude) of the acquisition terminal when acquiring real observation data and its preset ideal pose for effectively observing key semantic parts of the target object; its value range is [0,1], and the closer the value is to 1, the more the acquisition position and angle meet the preset requirements;

[0098] The calculation logic is as follows: The inertial measurement unit and visual-inertial odometry integrated into the acquisition terminal are used to obtain the real-time six-degree-of-freedom spatial pose in the world coordinate system. From the digital model on the server, the coordinates of the three-dimensional bounding box of the key semantic part to be observed on the target object and its ideal observation direction vector are obtained. The three-dimensional bounding box is projected onto the current two-dimensional view plane of the acquisition terminal, and the intersection-union ratio (IU / I) of the projected two-dimensional bounding box with the center region of the view plane is calculated to obtain a spatial position score between [0,1]. The cosine value of the angle between the current shooting direction vector of the acquisition terminal and the ideal observation direction vector is calculated, and it is mapped to the [0,1] interval through a linear transformation ((cosine value + 1) / 2) to obtain a direction attitude score. The spatial position score and the direction attitude score are weighted and averaged, with the weights preset according to task requirements (in this embodiment, each accounts for 50%), to obtain the final geometric-semantic consistency score. .

[0099] The symbol for the feature-spatial projection distance is: , is used to quantify the degree of difference between real observation data and virtual ideal data at the level of deep semantic features. The value range is [0,1]. The closer the value is to 0, the more similar the two are in terms of content, structure and semantics.

[0100] The calculation logic is as follows: The real observed image and its corresponding virtual ideal image in time or space are input into a deep neural network model. The normalized high-dimensional feature vectors output by the deep neural network model for the two images are obtained. The cosine similarity between these two high-dimensional feature vectors is calculated, with a value range of [-1, 1]. To convert this into a parameter representing distance or dissimilarity, with a value range of [0, 1], the following linear transformation is used: Subtract the calculated cosine similarity from 1, and then divide the result by 2 to obtain the final feature-space projection distance. ;

[0101] The process for calculating uniform data quality parameters is as follows:

[0102] Input includes real-time observation images from the camera of the acquisition terminal; real-time spatial pose of the acquisition terminal synchronized with the images; and virtual ideal images corresponding to the pose and associated digital model information obtained from the server.

[0103] Parallel computing steps:

[0104] Visual information entropy was calculated based on real observed images. ;

[0105] Based on real-time spatial pose and digital model information, the geometric-semantic consistency score is calculated. ;

[0106] The feature-spatial projection distance is calculated based on real observed images and virtual ideal images. ;

[0107] Linkage weight adjustment steps (logical judgment and processing):

[0108] Based on feature-spatial projection distance Dynamically determine the entropy used for fusing visual information. Geometric-semantic consistency score The weighting coefficients; the judgment logic is: if the feature-spatial projection distance... A low value (below 0.1 in this example) indicates that the digital model highly matches physical reality. In this case, the geometric constraints based on the digital model should be trusted, and therefore a geometric-semantic consistency score is assigned. Higher weights; conversely, if the feature-spatial projection distance... A high value (above 0.5 in this embodiment) indicates a significant difference between the two. In this case, the reliance on unreliable geometric constraints should be reduced, and more attention should be paid to the information content of the image itself. Therefore, visual information entropy is assigned. Higher weight;

[0109] Parameter fusion steps:

[0110] Using dynamically determined weights, the visual information entropy is... Geometric-semantic consistency score By performing a weighted summation, an adaptive endogenous quality parameter is obtained. ;

[0111] Endogenous quality parameters The parameter characterizing the relative quality (derived from the feature-spatial projection distance) In this embodiment, the feature-spatial projection distance is subtracted from 1. A second fusion is performed to obtain the final unified data quality parameters. ; represents a scalar value in the interval [0,1].

[0112] Used for fusing visual information entropy Geometric-semantic consistency score The weighting coefficients, i.e., the weights of endogenous mass sub-units ( and The value is not a preset fixed value, but is determined in real time through the following calculation method:

[0113] Geometric-semantic consistency score weight For a feature-space projection distance The Gaussian decay function, and its linguistic computational model, is: weights. Characteristic spatial projection distance equal to the power of the natural constant e, with a negative exponent. Squared by twice the preset sensitivity parameter The square of;

[0114] Sensitivity parameters The rate of weight decay was controlled, and its value needs to be calibrated through offline experiments. The calibration method is as follows: prepare a validation dataset containing various known model-reality matching degrees (from perfect match to severe mismatch), and find an optimal sensitivity parameter through grid search. The value makes the sensitivity parameter Under these conditions, the final output quality parameters have the highest correlation with the annotation results of human experts;

[0115] To ensure that the sum of the two weights is 1, the weight of visual information entropy... Determined to be 1 minus .

[0116] Through this mechanism, when the feature-spatial projection distance When the value approaches 0 (reality and virtuality are highly consistent), the weight... Approaching 1, mainly depends on geometric constraints; when the feature-space projection distance When the weight increases, It will decay rapidly in a non-linear manner, and the decision-making basis will smoothly transition to an entropy value that relies more on the data itself.

[0117] S4. Based on the first and second mass parameters, generate a deviation field signal that characterizes the multi-dimensional deviation between the real observation data and the ideal acquisition state.

[0118] The deviation field signal in S4 is generated from the real-virtual deviation field; the deviation field signal is a multi-dimensional vector field, which includes a deviation potential scalar field, a correction gradient two-dimensional vector field, and a deviation field time gradient scalar.

[0119] The steps for generating the deviation field signal are as follows:

[0120] A scalar potential field representing the magnitude of the deviation at each location is generated by unifying data quality parameters; based on the scalar potential field, a vector gradient field is calculated to indicate the direction of correction.

[0121] The steps to generate a scalar potential field include:

[0122] Based on the pre-defined semantic saliency map, the deviations at different positions are weighted and calculated.

[0123] Based on the change of the scalar potential field in continuous time frames, the deviation field time gradient scalar is calculated to characterize the rate of deviation change.

[0124] In this embodiment, the key parameters used to construct and describe the reality-virtual deviation field are defined as follows:

[0125] The parameter sign of localized visual information entropy is: Used to quantize real-world observed images in specified grid cells The information richness within the range is normalized to [0,1].

[0126] The sign of the parameter for localized feature-spatial projection distance is: Used to quantize the difference between real observed images and virtual ideal images in a specified grid cell. Deep semantic differences within, Represented as the x-coordinate of the grid weight. Represented as the ordinate of the grid weight;

[0127] The computational logic is as follows: The complete real-world observed image and the virtual ideal image are input into a visual feature extraction model based on the ViT architecture (DINOv2 is used in this embodiment); in the intermediate layer of the visual feature extraction model, the feature vector sequence corresponding to each image patch is extracted; based on the grid cells... Based on the spatial correspondence of the tiles, the real and virtual feature vectors corresponding to the units are extracted; the cosine similarity between these two local feature vectors is calculated and converted into a distance metric with a value range of [0,1], thus obtaining the localized feature-spatial projection distance. .

[0128] The parameter sign of the global geometric deviation vector is: , is a six-dimensional vector used to fully describe the overall deviation between the current real-time spatial pose of the acquisition terminal and the ideal spatial pose, including translational deviation and rotational deviation;

[0129] The calculation logic is as follows: the translational deviation component is obtained by calculating the Euclidean distance vector between the current pose and the ideal pose's three-dimensional coordinates. The rotational deviation component is obtained by calculating the angle difference or product between the quaternions representing the two poses, and is represented as a three-dimensional rotation vector. The two three-dimensional vectors together constitute a six-dimensional global geometric deviation vector. ;

[0130] The parameter notation of the semantic saliency map is , is a two-dimensional matrix of the same size as the actual observed image, used to identify the importance of each pixel or grid cell in the image to the current observation task. Its value range is [0,1], and the closer the value is to 1, the more important the region is.

[0131] The calculation logic is as follows: Based on the current pose of the acquisition terminal, the 3D mesh marked as key semantic parts in the digital model is subjected to frustum projection to generate a binary mask image, in which the pixel value of the key part projection area is 1 and the background is 0. In order to make the importance transition smoothly, a Gaussian blur filter with a preset kernel size is applied to the binary mask image to generate the final semantic saliency map. ;

[0132] The parameter sign of the deviation potential is: It is a scalar that is assigned to each grid cell. , is used to quantize the total deviation caused by multiple deviation sources and weighted by semantic importance on the unit. Its value range is [0,1]. The larger the value, the greater the gap between the acquisition quality of the local area and the ideal state.

[0133] The parameter sign of the corrected gradient vector is... It is a two-dimensional vector assigned to each grid cell. Its direction points to the direction in which the deviation potential of the cell decreases the fastest in its neighborhood, and its magnitude represents the rate of decrease, providing users with intuitive guidance on how to adjust the viewing angle on the image plane.

[0134] The parameter sign of the time gradient scalar of the deviation field is: , is a scalar used to quantify the rate of change of the total energy of the entire bias potential field over time. A positive value indicates that the acquisition quality is deteriorating, a negative value indicates that it is improving, and the magnitude of the absolute value indicates the degree of change.

[0135] The calculation process for generating the reality-virtual deviation field is as follows:

[0136] Input the current time frame The acquired parameters include localized visual information entropy. Localized features - spatial projection distance Global geometric deviation vector and pre-generated semantic saliency graphs It is also necessary to start from the previous time frame Total bias potential of the cache;

[0137] The field of view of the real observed image is logically divided into an N×M grid, and each grid cell... Initialize an empty deviation data structure;

[0138] Deviation potential Calculation steps:

[0139] Traverse all grid cells Within each unit, the local content bias is calculated by subtracting the localized visual information entropy bias from the information entropy bias (i.e., 1 minus the localized visual information entropy). ) and semantic content bias (i.e., localization feature-spatial projection distance) We obtain the result by weighted summation;

[0140] Multiply the local content deviation by the semantic saliency map corresponding to the unit. This amplifies the deviation falling in the critical region, while suppressing the deviation in the background region; the deviation potential of the output unit... ;

[0141] Correcting the gradient vector Calculation steps:

[0142] For a complete two-dimensional scalar deviation potential Apply a numerical gradient operator,

[0143] The calculation logic is as follows: For each unit Correcting the gradient vector The x-axis component is calculated by the right neighbor. With the left neighbor The difference between them is obtained; the y-axis component is obtained by calculating the difference between the upper neighbors. With the next neighbor The difference between them is obtained. To obtain the correction direction, the calculated gradient vector is inverted to obtain the corrected gradient vector pointing in the direction of potential energy decrease. ;

[0144] deviation field time gradient scalar Calculation steps:

[0145] Calculate the current time frame Total deviation potential The calculation method is the deviation potential for all grid cells. Perform summation;

[0146] Read the cached previous time frame Total deviation potential ;

[0147] Calculate the difference between the two and divide by the time interval between the two time frames. The time gradient scalar of the bias field is obtained. ;

[0148] Final Output: The final output is a structured deviation field signal, which consists of three components: an N×M deviation potential. N×M corrected gradient, two-dimensional corrected gradient vector Global bias field time gradient scalar ;

[0149] S5. Based on the deviation field signal, generate and output control commands to guide the acquisition terminal to adjust its acquisition behavior in order to reduce deviation.

[0150] The control commands generated in S5 are augmented reality visual guidance signals, which are superimposed on the display interface of the acquisition terminal.

[0151] The generated real-to-virtual deviation field is rendered in real time as an intuitive AR visual overlay. In this embodiment, heatmaps of different colors are used on the screen of the acquisition terminal to visualize the overall deviation magnitude of each region; simultaneously, dynamically changing arrows or path indicators are used to visualize measures to reduce the geometric deviation component. The required direction and distance of movement are determined by the deviation field, which is recalculated in real time as the user moves the device according to the AR guidance. The visual guidance is also dynamically updated until the deviation is reduced to below the threshold and the guidance signal disappears, indicating that the current acquisition quality meets the standard.

[0152] The control commands include external control commands for guiding the acquisition terminal and internal control commands for adjusting internal algorithm parameters;

[0153] The external control commands are augmented reality visual signals, which include macroscopic pose guidance elements generated by global geometric deviation based on the deviation field; and microscopic view guidance elements generated by modified gradient field based on the deviation field.

[0154] The internal control command updates the value of the integral deviation accumulator based on the deviation field signal;

[0155] The visual properties of augmented reality visual signals are dynamically adjusted based on the time gradient scalar of the bias field;

[0156] An internal control command for adjusting the internal algorithm parameters is generated only when the value of the integral deviation accumulator exceeds the preset adaptive trigger threshold.

[0157] S5 and beyond also include dynamically adjusting the weights or thresholds used to calculate the first or second mass parameter based on the deviation field signal;

[0158] It not only guides users externally, but also makes adaptive adjustments internally; specifically, it continuously monitors the average amplitude of the global deviation field.

[0159] Example of linkage control: Detecting geometric deviation components in a certain task scenario. Generally high, automatically reducing geometric-semantic consistency scores. The weights in the calculation of the first quality parameter are also increased, along with the weights of the visual information entropy.

[0160] Graceful degradation strategy trigger: When an interruption in the real observation data stream from the camera is detected, causing the first quality parameter to remain invalid, the deviation field module will be unable to calculate. At this time, the control logic will trigger degradation, automatically switching to pure geometry-guided mode, providing pose guidance only based on IMU data and virtual ideal data from the acquisition terminal, and issuing an alarm to the user for the loss of visual signals;

[0161] In this embodiment, the key parameters used to achieve control are defined as follows:

[0162] The parameter symbols of the deviation field signal packet are as follows: It is a structured data packet containing all the calculated deviation field information, specifically including: deviation potential. Correcting the gradient vector and global geometric deviation vector ; and the global bias field time gradient scalar ;

[0163] The parameter sign of the integral deviation accumulator value is: , is an internal state variable used to smooth and accumulate the total deviation over a period of time to reflect persistent acquisition quality issues. Its physical meaning is similar to a low-pass filter, which filters out instantaneous noise and retains the long-term trend of the deviation.

[0164] The calculation logic is as follows: at each time step The value of the integral deviation accumulator Update the value of the integral deviation accumulator in the following way: Equal to the preset smoothing factor Multiply by the total deviation potential of the current frame Plus (1 minus the smoothing factor) Multiply by its previous time step value ;

[0165] Smoothing factor The range of its value is (0,1), which determines the memory length of the accumulator and the smoothing factor. The smaller the value, the longer the memory of historical biases, the more stable the system, but the slower the response; this is the smoothing factor. The optimal balance between stability and response speed can be determined through offline simulation experiments. In this embodiment, the value is set to 0.05.

[0166] The control logic threshold set is a pre-defined set of scalar thresholds used to trigger different control behaviors:

[0167] AR activation threshold Used to determine whether AR guidance needs to be displayed, when the total deviation potential... If the value is below this, the data acquisition quality is considered to be up to standard, and no guidance will be displayed; the value range is (0, 0.3], with an optimal value of 0.1.

[0168] Emergency warning threshold Used to determine if the rate of deviation deterioration is too fast, when the global deviation field time gradient scalar When this value is exceeded, an AR emergency alert is triggered; the value range is (0.4, 0.8], with an optimal value of 0.6.

[0169] Adaptive trigger threshold Used to determine if the deviation is too large, when the value of the integral deviation accumulator... Exceeding the adaptive trigger threshold When this occurs, the internal parameters are adaptively adjusted; the value range is [0.6, 0.9), with an optimal value of 0.75.

[0170] Adaptive Reset Threshold Used to determine whether the deviation has been eliminated, when the value of the integral deviation accumulator... Fall back to adaptive reset threshold The adaptive state is exited when the following conditions are met; the value range is [0.2, 0.5), with an optimal value of 0.3.

[0171] All thresholds were jointly calibrated through user experience testing and performance evaluation in multiple simulated scenarios. The goal of the calibration was to ensure that the guiding signal appears when necessary to avoid information overload, while ensuring that the adaptive mechanism is activated only when necessary to maintain system stability.

[0172] The parameter symbols of the AR guidance signal packet are: The output of the external control loop is a structured data packet containing all the information needed to drive the acquisition terminal to render all AR visual elements.

[0173] The parameter notation of the intrinsic parameter weight vector is: , is the output of the internal control loop, which is a vector containing the weights of the sub-parameters of the upstream algorithm that should be adjusted;

[0174] The complete calculation process for achieving internal and external dual closed-loop control is as follows:

[0175] The input is the deviation field signal packet of the current time frame. And the value of the integral deviation accumulator of the previous time frame in the internal cache. and internal parameter weight vector ;

[0176] External control loop: AR guidance signal generation:

[0177] From the deviation field signal packet Extracting the total deviation potential (Through the deviation potential) (summated to obtain)

[0178] Determine the total deviation potential Is it greater than the AR activation threshold? If not, generate an empty AR guidance signal packet. If so, then jump to the internal control loop; if so, continue.

[0179] Initialize a non-empty AR boot signal packet. Please fill in the following data:

[0180] Macroscopic guidance layer: This layer stores the deviation field signal packet of the current time frame. Global geometric deviation vector in Directly store AR guidance signal packet The rendering engine will then render a large arrow in 3D space to indicate how the user should translate and rotate.

[0181] Microscopic guidance layer: This layer stores the deviation field signal packet of the current time frame. Corrected gradient vector Store AR guidance signal package The rendering engine will then overlay a dynamic flow or particle effect onto the two-dimensional plane of the screen, indicating finely adjustable directions.

[0182] Status alert layer: This layer stores the deviation field signal packet of the current time frame. Deviation potential in Store AR guidance signal package A heatmap used to render the background;

[0183] From the deviation field signal packet of the current time frame Extract the global deviation field time gradient scalar. Determine whether it exceeds the emergency warning threshold. If so, then in the AR guidance signal packet The emergency status flag is set to true; the rendering engine will switch the color of all AR elements to the warning color (set to red in this embodiment) and add a flashing effect based on this flag.

[0184] Internal control loop: Adaptive adjustment of model parameters:

[0185] Using the current total deviation potential and smoothing factor Update the integral deviation accumulator to obtain its value. ;

[0186] If the current state is normal, and the value of the integral deviation accumulator is... Greater than the adaptive trigger threshold If so, the state will switch to an adaptive state;

[0187] If the current state is adaptive, and the value of the integral deviation accumulator is... Less than the adaptive reset threshold If so, the state will be switched back to the normal state;

[0188] If the state switches to or remains in the adaptive state, the weight adjustment logic is executed; the weight adjustment logic analyzes the resulting value of the integral deviation accumulator. The primary sources of persistently high deviations (in this embodiment, by comparing the contributions of global geometric deviation and average local content deviation) generate a new internal parameter weight vector. In this embodiment, if geometric deviation is determined to be the main cause, the new weight vector will reduce the weight of the geometric-semantic consistency score while increasing the weight of the visual information entropy.

[0189] If the system is in a normal state, no adjustments are made, and the current internal parameter weight vector is set. Equal to the internal parameter weight vector of the previous time step .

[0190] Final output: Two parallel signals: an AR guidance signal packet sent to the rendering engine of the acquisition terminal. ; Send back upstream data internal parameter weight vector .

[0191] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source data filtering method, characterized in that, The method comprises the following steps: S1, obtaining virtual ideal data associated with a target object, and collecting real observation data of the target object collected by a terminal in real time; S2, calculating a first quality parameter representing the quality of endogenous information based on the real observation data; S3, comparing the real observation data with the virtual ideal data to calculate a second quality parameter representing the difference between the real observation data and the virtual ideal data; S4, generating a deviation field signal representing the multi-dimensional deviation between the real observation data and the ideal collection state based on the first quality parameter and the second quality parameter; S5, generating and outputting a control instruction for guiding the collection terminal to adjust the collection behavior based on the deviation field signal to reduce the deviation.

2. The multi-source data screening method of claim 1, wherein: The step of calculating the first quality parameter based on the real observation data in S2 is as follows: Calculate the visual information entropy of the real observation data; and compare the real-time spatial pose of the collection terminal with the preset digital model to calculate a geometric-semantic consistency score; the first quality parameter is determined based on the visual information entropy and the geometric-semantic consistency score.

3. The multi-source data screening method of claim 2, wherein: The step of comparing the real observation data with the virtual ideal data to calculate the second quality parameter in S3 is as follows: Input the real observation data and the virtual ideal data into a pre-trained deep neural network model respectively to extract high-dimensional feature vectors thereof; calculate the projection distance between the high-dimensional feature vectors in the feature space, and take the projection distance as the second quality parameter.

4. The multi-source data screening method of claim 3, wherein: The specific steps of calculating the fusion weight of the first quality parameter and the second quality parameter are as follows: The first quality parameter adjusts the proportion of the visual information entropy and the geometric-semantic through the weight, and the weight is determined by a Gaussian decay function with the feature-space projection distance as the independent variable; The first quality parameter and the second quality parameter are weighted and fused to obtain the final unified data quality parameter.

5. The method of claim 4, wherein: The deviation field signal in S4 is generated by a real-virtual deviation field; the deviation field signal is a multi-dimensional vector field, and the deviation field signal includes a scalar potential field, a vector gradient field, and a deviation field time gradient scalar.

6. The multi-source data filtering method of claim 5, wherein: The generation steps of the deviation field signal are as follows: Generate a scalar potential field representing the deviation size at each position through the unified data quality parameter; Calculate a vector gradient field for indicating the correction direction based on the scalar potential field; The steps of generating the scalar potential field include: Weighted calculation of the deviation at different positions according to a preset semantic saliency map; Calculate a deviation field time gradient scalar for representing the deviation change rate based on the change of the scalar potential field in consecutive time frames.

7. The multi-source data filtering method of claim 6, wherein: The control instruction generated in S5 is an augmented reality visual guidance signal, which is superimposed and displayed on the display interface of the collection terminal.

8. The multi-source data filtering method of claim 7, wherein: After S5, it further includes dynamically adjusting the weight or threshold used when calculating the first quality parameter or the second quality parameter based on the deviation field signal.

9. The multi-source data filtering method of claim 8, wherein: The control instruction includes an external control instruction for guiding the collection terminal and an internal control instruction for adjusting the internal algorithm parameters; The external control instruction is an augmented reality visual signal, which includes a macro pose guidance element generated based on the global geometric deviation of the deviation field; And a micro view angle guidance element generated based on the correction gradient field of the deviation field; The internal control instructions update the value of the integral deviation accumulator based on the deviation field signal.

10. The multi-source data screening method of claim 9, wherein: The visual properties of the augmented reality visual signal are dynamically adjusted according to the deviation field time gradient scalar; The internal control instructions for adjusting the internal algorithm parameters are generated only when the value of the integral deviation accumulator exceeds a preset adaptive triggering threshold.

Images