Construction method of high-wear-resistance composite floor, sensing detection method and related products

Abstract

The invention relates to the technical field of building energy conservation and green building material technologies and building engineering quality detection, in particular to a construction method, a sensing detection method and related products of a high-wear-resistance composite terrace, and the construction method comprises the steps of preparing a base layer, applying a wear-resistant middle layer, laying a decorative surface layer, curing and finish machining, and carrying out multi-mode quality acceptance inspection. The detection method comprises the steps of collecting a data set; generating a plurality of intermediate state attributes representing different physical dimensions; obtaining a current actual measurement comprehensive state representation representing the surface of the terrace; outputting a quality detection result of the terrace; the high-hardness wear-resistant layer made of the carborundum aggregate and the terrazzo surface layer containing the decorative aggregate are integrally and compositely laid; and through systematic fusion of sensing technologies of different physical dimensions, comprehensive information of chemical substance components of the surface layer, internal structure defects at the depth of several centimeters close to the surface and a dynamic friction coefficient can be obtained at the same time in one detection task, and three-dimensional reconnaissance of the quality of the terrace is realized.

Description

Technical Field

[0001] This invention relates to the fields of building energy conservation and green building materials technology and building engineering quality testing technology, specifically to a construction method, sensing and testing method and related products for a high wear-resistant composite floor, aiming to reduce building energy consumption and improve indoor environmental friendliness. Background Technology

[0002] Currently, there is a growing global focus on building energy conservation and environmental protection, and the development of low-energy, environmentally friendly green buildings has become a core industry trend. As an important component of the building envelope, the thermal performance, environmental friendliness of materials, and life-cycle cost of flooring have a significant impact on the overall energy efficiency and sustainability of buildings.

[0003] Traditional concrete or terrazzo floors have high thermal conductivity and insufficient insulation properties, especially in the ground floor or underground spaces. This can easily lead to heat loss in winter or cool air loss in summer, directly increasing the energy load on air conditioning and heating systems. Furthermore, some flooring adhesives may release harmful substances such as formaldehyde during production and use, posing a potential threat to indoor air quality and human health, and failing to meet the ecological requirements of modern green buildings.

[0004] In flooring engineering practice, terrazzo flooring is favored for its decorative appeal, but its traditional manufacturing process not only has limited wear resistance but also lacks special design for thermal insulation, making it difficult to meet the standards for energy-efficient buildings. While emery abrasion-resistant flooring possesses extremely high mechanical strength, effectively extending its service life and reducing resource consumption from renovations, its appearance is monotonous, and it also fails to fully consider thermal performance and energy-saving requirements.

[0005] Terrazzo flooring is a product made by mixing aggregates such as crushed stone, glass, and quartz with cementitious binders to form a concrete product, which is then ground and polished. Terrazzo made with cementitious binders is called inorganic terrazzo, while terrazzo made with epoxy binders is called epoxy terrazzo or organic terrazzo. Terrazzo is also classified according to its construction process into cast-in-place terrazzo and precast terrazzo flooring.

[0006] Traditional terrazzo flooring enjoys a huge market due to its unique advantages such as low cost, customizable colors and patterns, and convenient construction. However, the inherent material properties of traditional terrazzo flooring also limit its surface hardness and wear resistance. When exposed to high-frequency pedestrian and material traffic or heavy equipment, its surface is prone to scratches, wear, and even dust.

[0007] In contrast, emery flooring involves evenly spreading a dry-spread hardener composed of emery, silicate cement, and other admixtures onto the concrete surface during the initial setting stage of a newly poured concrete base. This is then mechanically smoothed and compacted to form a dense, high-strength, wear-resistant integral surface layer with the concrete base. The most prominent advantages of this type of flooring are its extremely high surface hardness and excellent wear resistance and impact resistance. However, its function-oriented nature also brings significant limitations: its appearance is usually relatively simple and rugged, with limited color choices, mainly presenting the original cement color or a limited number of industrial colors, lacking refined decorative effects, and failing to meet the spatial aesthetic pursuits of modern architecture.

[0008] The current technology of flooring presents a situation where it is difficult to achieve both "decorative performance" and "mechanical wear resistance" efficiently. Even if the combination of terrazzo flooring and emery flooring is completed, the current testing methods used to evaluate the final installation quality have significant limitations.

[0009] Firstly, the most common method is manual visual inspection and measurement with hand tools. Inspectors rely on their personal experience to observe the floor surface for cracks, sandblasting, color differences, stains, etc., and use simple tools such as straightedges and feeler gauges to measure flatness, or to make a preliminary judgment on the presence of hollow spots by tapping and listening to the sound. While this method is simple, its results are highly dependent on the inspector's experience and subjective judgment, lacking unified, quantifiable objective standards, and easily leading to disputes during the acceptance process. Furthermore, for large-area floor projects, manual inspection is not only inefficient and labor-intensive, but also difficult to ensure comprehensive coverage of inspection points, easily overlooking localized or hidden defects.

[0010] Secondly, some tests use single-function instruments for point-to-point measurement. For example, using a pendulum friction meter to measure slip resistance or a rebound hammer to test surface hardness only provides discrete, single-dimensional physical parameters. These methods cannot reveal deeper, comprehensive quality conditions of the flooring, such as: the uniformity of surface chemical composition and the presence of residual contaminants; whether different materials like terrazzo and corundum have achieved good microscopic integration; and the tightness of the bond between the surface layer and the base layer, and whether there is a potential risk of delamination.

[0011] Therefore, developing a new composite flooring process that combines the advantages of terrazzo and corundum, and simultaneously developing an advanced testing system that can scientifically and accurately verify and ensure the quality of this new process, has become an important technical issue that urgently needs to be addressed in this field. Summary of the Invention

[0012] To address the aforementioned problems, this invention provides a construction method, a sensing and detection method, and related products for a high-wear-resistant composite floor.

[0013] This invention is achieved through the following technical solution:

[0014] A construction method for a high-wear-resistant composite floor, comprising:

[0015] Preparation of the base layer: On the treated ground surface, lay and vibrate the concrete base layer;

[0016] Apply a wear-resistant intermediate layer: When the concrete base layer reaches the predetermined plastic state, a wear-resistant mortar layer containing diamond aggregate is evenly laid on it to form a high-hardness wear-resistant base.

[0017] Laying the decorative surface layer: Before the wear-resistant intermediate layer initially sets and solidifies, a terrazzo surface layer material containing decorative aggregate is laid on its surface to ensure a strong chemical and physical bond between the two layers;

[0018] Curing and finishing: The laid composite floor is cured and maintained in a standardized manner, and then its surface is subjected to multi-stage coarse grinding, fine grinding and chemical polishing until a composite finished surface is formed that exposes the decorative aggregate and part of the diamond aggregate.

[0019] Multimodal quality acceptance inspection: A fusion inspection method based on multimodal sensing is used to conduct the final quality acceptance of the finished surface.

[0020] A sensing and detection method for high wear-resistant composite flooring is provided for performing the multimodal quality acceptance testing described above. The sensing and detection method includes:

[0021] A set of interconnected datasets is collected in the target ground area. The datasets include at least: a hyperspectral dataset for characterizing the material composition of the area, an active ultrasonic dataset for characterizing the near-surface internal structure of the area, and a frictional force dataset for characterizing the dynamic frictional properties of the surface of the area.

[0022] Each type of data in the dataset undergoes preliminary processing to generate multiple intermediate state attributes representing different physical dimensions.

[0023] In a common spatial coordinate system, multiple intermediate state attributes are fused to obtain the current measured comprehensive state representation of the ground surface;

[0024] The comprehensive state representation is input into a pre-trained standard reference model for quantization and comparison, and the output is the quality inspection result of the floor.

[0025] Optionally, the method for obtaining the hyperspectral dataset and characterizing the material composition of the region includes:

[0026] A reflection intensity map of the target ground area is generated by pre-scanning, and based on the reflection intensity map, at least one acquisition parameter is dynamically adjusted during the formal acquisition to obtain a hyperspectral data cube. The acquisition parameter is selected from the light source irradiance intensity or the sensor integration time.

[0027] The hyperspectral data cube is input into a physical information neural network, which demixes the mixed pixels to extract a set of endmembers representing the spectral characteristics of pure substances and their respective abundances.

[0028] The endmembers are compared with reference spectra in a standard spectral library, and a weighted similarity measurement algorithm is used to determine the best-matching material component for each endmember, so as to generate intermediate state attributes that characterize the material components.

[0029] Optionally, methods for acquiring active ultrasonic datasets and characterizing the near-surface internal structure of the region include:

[0030] An ultrasonic transducer array is used, in which each element of the array sequentially transmits a pulse-compressed and encoded excitation signal, and all elements of the array synchronously receive the echo, thus acquiring a full matrix dataset covering all transmit-receive channel combinations.

[0031] At at least one calibration point in the target floor area, the round-trip propagation time of sound waves on a floor base layer of known thickness is measured using an ultrasonic transducer array, and the local sound velocity value of the current area is calculated.

[0032] The full matrix dataset is processed using the total focusing method, and the acoustic cross-sectional image of the near-surface internal structure of the target ground area is reconstructed using local sound velocity values;

[0033] The acoustic cross-sectional image is input into a pre-trained image segmentation model to identify and segment defect regions in the image, and the geometric parameters of the defect regions are quantified to generate intermediate state attributes characterizing the internal structure. The geometric parameters include depth, size, or orientation.

[0034] Optionally, the method for obtaining the friction force dataset and characterizing the dynamic frictional properties of the surface in this region includes:

[0035] A friction probe integrating a multi-axis force sensor and a vibration sensor is used to simultaneously collect a set of multi-dimensional force vector data and a set of vibration signal data during the process of moving relative to the ground surface with a preset slip ratio and constant normal pressure.

[0036] The dynamic friction coefficient time series is calculated based on the multidimensional force vector data, and the vibration signal data is subjected to time-frequency analysis to extract the vibration feature spectrum;

[0037] The dynamic friction coefficient time series and the vibration feature spectrum are fused into a composite feature vector, and the composite feature vector is input into a preset evaluation model to output the classification or quantitative evaluation results of the surface dynamic friction characteristics, so as to generate intermediate state attributes characterizing dynamic friction characteristics.

[0038] Optionally, the method for fusing the multiple intermediate state attributes into a comprehensive state representation includes:

[0039] The geometry of the target ground area is mapped into a spatial graph, wherein the nodes of the spatial graph correspond to predetermined cells within the area, the edges of the spatial graph represent the spatial adjacency relationship between the cells, and the initial feature vector of each node is composed of the multiple intermediate state attributes of its corresponding cell.

[0040] A graph neural network is used to perform multi-round iterative message passing and node update operations on the spatial graph. In each round of operation, each node updates its own feature representation by aggregating the features of its neighboring nodes.

[0041] The final set of node feature vectors output after processing by the graph neural network is defined as the comprehensive state representation.

[0042] Optionally, methods for performing quantitative comparison and outputting quality inspection results include:

[0043] A generative deep learning model trained on a comprehensive state representation of multiple standard compliant floor samples is used as the standard reference model.

[0044] The comprehensive state representation of the ground is input into the standard reference model, and the reconstructed comprehensive state representation generated by the standard reference model is obtained.

[0045] Calculate the reconstruction error between the comprehensive state representation and the reconstructed comprehensive state representation, and quantify the reconstruction error as an anomaly score characterizing the degree to which the floor to be accepted deviates from the quality standard;

[0046] The quality inspection result is generated by comparing the abnormal score with a preset judgment threshold.

[0047] Optionally, methods for generating quality inspection results include:

[0048] Extreme value theory is used to statistically model the set of abnormal scores obtained from multiple standard compliant floor samples, and a probability judgment model is established.

[0049] For each cell of the flooring, the anomaly score is calculated using a probability judgment model to determine whether it belongs to the standard compliant flooring. The confidence score represents the probability that the cell meets the quality standard.

[0050] Based on the confidence score, a quality inspection result is generated, and by analyzing the intermediate state attribute components that constitute the cell reconstruction error, the source of defects in non-compliant areas is output.

[0051] Furthermore, the method also includes a system calibration step, which includes:

[0052] The multimodal sensor assembly is instructed to measure a standard reference containing multiple regions of known physical properties, and the multimodal sensor assembly is calibrated based on the difference between the measured data and the known physical properties. The known physical property regions include at least a standard spectral reflectance region, a standard acoustic property region, or a standard friction coefficient region.

[0053] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the sensing and detection method for high-wear-resistant composite flooring as described above.

[0054] A computer program product includes a computer program / instructions that, when executed by a processor, implement the sensing and detection method for high-wear-resistant composite flooring as described above.

[0055] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0056] This invention, through a specific composite structural design, exhibits lower thermal conductivity compared to traditional flooring, effectively blocking heat conduction from the ground and providing excellent thermal insulation. This helps maintain stable indoor temperatures, significantly reducing building energy consumption for heating in winter and cooling in summer, aligning with the development requirements of low-energy green buildings.

[0057] This invention integrates a high-hardness, wear-resistant layer of corundum aggregate with a terrazzo surface layer containing decorative aggregate through a specific construction sequence, followed by deep grinding and polishing processes. Finally, a multimodal sensing and detection method is used for final quality acceptance, and objective and comprehensive data analysis is used to ensure that the finished flooring meets its high-standard performance design.

[0058] This invention utilizes a composite structure design, with a corundum wear-resistant layer as a solid base and terrazzo as an artistic surface layer. This design combines the superior mechanical properties of corundum—its ultra-high hardness, impact resistance, and wear resistance—with the rich colors, diverse patterns, and seamless aesthetic appeal of terrazzo. This results in a flooring product that meets both high-intensity usage requirements and decorative appeal.

[0059] This invention constructs a multimodal sensing system integrating a hyperspectral imager, an active ultrasonic detection array, and a dynamic friction probe. First, the raw data collected by each sensor is processed to generate their own independent intermediate state attributes. Then, these attributes are fused to obtain a comprehensive state representation. Finally, the comprehensive state representation is compared with a reference model to output a quality acceptance conclusion.

[0060] This invention systematically integrates three sensing technologies from different physical dimensions: hyperspectral imaging, active ultrasound, and dynamic frictional vibration. In a single inspection task, it can simultaneously acquire comprehensive information on the surface chemical composition, internal structural defects within centimeters of the surface, and dynamic friction coefficient, thus achieving a three-dimensional survey of floor quality.

[0061] This invention employs standardized sensor acquisition and mathematical model-based algorithms (such as PINN, TFM, GNN) for fully automated analysis, establishing quality assessment based on quantifiable data and models, eliminating interference from human factors, and ensuring the objectivity of the test results.

[0062] This invention analyzes the contribution of each error source in the final judgment model, which can identify the possible sources of defects that lead to non-compliance, providing clear guidance for subsequent responsibility identification and remediation work. Attached Figure Description

[0063] The accompanying drawings illustrate exemplary embodiments of the present invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, but do not constitute a limitation on the embodiments of the present invention.

[0064] Figure 1 This is a schematic flowchart of a sensing and detection method for a high wear-resistant composite flooring according to the present invention.

[0065] Figure 2 This is a flowchart illustrating the method for acquiring the hyperspectral dataset and characterizing the material composition of the region according to the present invention.

[0066] Figure 3 This is a flowchart illustrating the method for acquiring active ultrasonic datasets and characterizing the near-surface internal structure of the region according to the present invention.

[0067] Figure 4 This is a schematic flowchart of the method for obtaining the friction force dataset and characterizing the dynamic friction characteristics of the surface in the region according to the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0069] It should also be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0070] Where there is no conflict, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0071] This invention provides a construction method for a high-wear-resistant composite floor, comprising:

[0072] Preparation of the base course: On the treated (compacted, leveled and moisture-proofed) ground foundation (concrete floor slab or backfill layer of the building); lay the concrete base course according to the design thickness and use vibrating equipment (such as a plate vibrator or vibrating beam) to fully vibrate it.

[0073] Apply a wear-resistant intermediate layer: When the concrete base layer reaches the predetermined plastic state, a wear-resistant mortar layer containing diamond aggregate is evenly laid on it to form a high-hardness wear-resistant base; the predetermined plastic state means that the concrete has initially solidified and has sufficient load-bearing capacity to support the weight of workers and equipment, but its internal hydration reaction is still in progress, and the surface still has a certain "activity" and has not yet been fully hardened.

[0074] During this time window, a pre-prepared wear-resistant mortar layer containing a high proportion of corundum aggregate is evenly laid on the concrete base layer, and immediately compacted and leveled using a mechanical trowel. This forms a high-hardness, wear-resistant substrate, ensuring sufficient physical interlocking and chemical cross-linking between the wear-resistant mortar and the plastic concrete base layer, resulting in a strong bond.

[0075] Laying the decorative surface layer: Before the wear-resistant intermediate layer initially sets and solidifies, a terrazzo surface layer material containing decorative aggregate is laid on its surface to ensure a strong chemical and physical bond between the two layers; before initial setting and solidification means that the wear-resistant mortar layer has begun to lose fluidity, but has not yet generated sufficient strength, and the surface still retains the ability to chemically bond with new materials.

[0076] At this precise point in time, terrazzo surface material containing pre-selected decorative aggregates (such as colored glass fragments, marble, quartz, etc.) is laid on the surface of the wear-resistant intermediate layer. This core technology ensures a seamless, high-strength integrated bond between the two functional layers, effectively avoiding interface delamination or hollow areas that may occur in traditional layered construction.

[0077] Curing and finishing: The laid composite floor is cured and maintained in a standardized manner, and then its surface is subjected to multi-stage coarse grinding, fine grinding and chemical polishing until a composite finished surface is formed that exposes the decorative aggregate and part of the diamond aggregate.

[0078] For the completed composite flooring, standard procedures are followed for covering, watering, and other curing and maintenance to ensure the full development of its internal strength. After the curing period, a professional large-scale floor grinding machine is used with diamond grinding discs of varying grits from coarse to fine to perform "multi-stage coarse grinding and fine grinding" on the floor surface.

[0079] Multimodal quality acceptance inspection: A fusion inspection method based on multimodal sensing is used to conduct the final quality acceptance of the finished surface.

[0080] After the flooring construction is completed, a final quality acceptance is conducted using a sensory testing method for high-wear-resistant composite flooring. Traditional project acceptance methods are often controversial due to a lack of objective data. The testing method provided by this invention can generate a quantitative quality report. Based on the quantitative data and defect source analysis in the quality report, precise process optimization can be performed, improving the level of project quality control and acceptance efficiency.

[0081] Example 1

[0082] This embodiment discloses a highly automated and systematic detection method, which mainly consists of a perception layer and a decision layer.

[0083] The sensing layer is the foundation of the entire system, and its goal is to acquire comprehensive, accurate, and unavailable ground data that other technologies cannot obtain, including hyperspectral imaging modulus, active ultrasonic detection arrays, and dynamic friction probes.

[0084] Hyperspectral imaging modules can capture dozens or even hundreds of continuous, narrow-band electromagnetic spectrum segments from the visible to the near-infrared region. Each substance on the ground (water, oil, cleaning agents, paint) exhibits different reflectance and absorptivity under different wavelengths of light, forming a unique "spectral fingerprint." As the hyperspectral imaging module scans the ground line by line, acquiring the complete spectral curve data of each pixel, a three-dimensional data cube is formed. .

[0085] Active ultrasonic detector arrays utilize the principle that ultrasonic waves have different propagation speeds and attenuation characteristics in different media. This module is not a single probe, but an array of multiple (e.g., 8x8 or 16x16) piezoelectric transducers. Some transducers act as transmitters, emitting specially encoded focused ultrasonic pulses toward the ground; the remaining transducers act as receivers, receiving echo signals that penetrate the ground or are reflected back from different levels.

[0086] The dynamic friction probe simulates the critical state of a human foot about to slip but not quite, thus measuring the slip ratio during dynamic processes. The module consists of a small, motor-driven roller covered with a standard test material (such as neoprene). The roller rotates at a constant linear velocity, slightly below the walking speed, while a pressure sensor ensures it maintains constant pressure in contact with the ground. This results in minute slippage between the roller and the ground. A high-precision encoder measures the actual rotational speed of the roller, and a torque sensor measures its driving torque.

[0087] The decision-making level implements a sensing and detection method for a highly wear-resistant composite flooring, such as... Figure 1 As shown, it specifically includes:

[0088] The first step is to collect a set of correlated datasets in the target surface area. Correlatedness means that all collected data can be precisely mapped to specific spatial locations on the surface. The dataset must contain at least:

[0089] Hyperspectral datasets are used to characterize the material composition of a region. Hyperspectral technology can capture spectral information from hundreds of consecutive bands of light reflected from an object's surface, forming a unique "spectral fingerprint." By analyzing this fingerprint, specific chemical substances on the floor surface can be identified, such as the presence of oil stains, chemical residues, or the use of incorrect maintenance agents.

[0090] The active ultrasonic dataset used to characterize the near-surface internal structure of the area involves the detection system actively emitting ultrasonic pulses toward the floor and analyzing the internal condition by receiving the echoes. This is used to detect structural defects in the concrete that are not visible to the naked eye, such as hollow areas, delamination, and insufficient density.

[0091] The friction dataset is used to characterize the dynamic frictional properties of the surface in this area. It accurately identifies which areas have decreased friction due to material, wear, or contamination, which is related to the anti-slip performance and walking safety of the floor.

[0092] The second step involves preliminary processing of each type of data in the dataset to generate multiple intermediate state attributes representing different physical dimensions. After processing the hyperspectral data, a chemical map is generated that indicates the distribution of different chemical pollutants. After processing the ultrasonic data, a structural defect map is generated that indicates the location and size of internal voids. After processing the friction data, a slip risk map is generated.

[0093] The third step is to ensure that the chemical map is in a common spatial coordinate system. Location of points, structural defects on the map Location of the point, slip risk map The points are perfectly matched.

[0094] By fusing multiple intermediate state attributes, a comprehensive representation of the current state of the floor surface is obtained, which is a multi-layered, highly information-based digital twin of the floor, in which any coordinate point simultaneously contains multiple information such as chemical composition, internal structure, and friction characteristics.

[0095] The fourth step is to input the measured comprehensive state representation into the pre-trained standard reference model for quantitative comparison and output the quality inspection results of the floor.

[0096] Quantitative comparison involves using algorithms to accurately calculate the degree of difference between the measured representation and the standard model. The system will output quality inspection results based on the magnitude of the difference to determine whether the floor paving meets the standards.

[0097] The working principle of the method in this embodiment is to capture comprehensive physical, chemical and structural information of the ground through multimodal sensing technology, and to construct a comprehensive digital model of the actual state of the site through hierarchical intelligent processing and fusion.

[0098] First, a one-time, multi-dimensional data scan of the floor is performed, including hyperspectral, ultrasonic, and frictional data. Next, the algorithm module extracts these raw data into intermediate attribute maps of chemical composition, internal structure, and frictional characteristics. Then, the system merges these maps into a measured comprehensive state representation under a unified spatial coordinate system. Finally, by quantitatively comparing this measured comprehensive state representation with a reference model representing the "ideal standard," a final inspection conclusion on the quality of the floor paving is obtained.

[0099] Example 2

[0100] This embodiment details how to acquire and analyze hyperspectral data in the overall detection process to accurately identify the material composition of the ground surface. First, an adaptive scanning method ensures optimal quality of the acquired raw spectral data. Second, an artificial intelligence model incorporating physical laws is used to "purify" the spectral characteristics of individual substances from complex mixed signals. Finally, the purified features are precisely matched with a standard database to complete the final determination of the surface material and generate corresponding intermediate state attributes.

[0101] like Figure 2 As shown, the method for obtaining the hyperspectral dataset and characterizing the material composition of the region includes:

[0102] (1) In order to solve the problem of data quality degradation caused by uneven surface material, color and smoothness in actual testing, adaptive acquisition and data cube generation are carried out: the reflectance intensity map of the target surface area is generated by pre-scanning, and based on the reflectance intensity map, at least one acquisition parameter is dynamically adjusted in the formal acquisition to obtain a hyperspectral data cube. The acquisition parameter is selected from the light source irradiance intensity or the sensor integration time.

[0103] Light source irradiance refers to the brightness of the airborne active light source, which is increased when scanning darker areas and decreased when scanning highly reflective areas.

[0104] Sensor integration time is similar to camera exposure time, extending exposure in low light and shortening exposure in high light.

[0105] A high-quality "hyperspectral data cube" is obtained through adaptive acquisition, where two dimensions (X-axis and Y-axis) represent the spatial coordinates of the ground surface, while the third dimension (Z-axis) represents the spectral information consisting of hundreds of continuous bands collected at each coordinate point.

[0106] (2) Input the hyperspectral data cube into the physical information neural network, and the physical information neural network demixes from the mixed pixels a set of endmembers representing the spectral characteristics of pure substances and their respective abundances.

[0107] In practice, the tiny area covered by a single pixel of a sensor often contains more than one substance. In this case, the sensor receives a mixed spectral signal. The purpose of this step is to separate the original spectra of each pure substance from this mixed signal.

[0108] During training, Physical Information Neural Networks (PINNs) are not only required to learn the patterns in the data itself, but also to ensure that their outputs comply with known physical laws (such as the light reflection and absorption model in optics).

[0109] An endmember is a "spectral fingerprint" that is separated from the mixed spectrum by the model and represents a specific pure substance. Abundance refers to the area proportion or contribution of each endmember in the original mixed pixel. For example, a pixel can be unmixed into "95% epoxy resin" and "5% hydraulic oil".

[0110] (3) The endmembers are compared with the reference spectra in the standard spectral library, and a weighted similarity measurement algorithm is used to determine the most matching material component for each endmember in order to generate intermediate state attributes that characterize the material components.

[0111] This step compares each endmember spectrum extracted in the previous step with a large number of reference spectra in a standard spectral library. The standard spectral library is a pre-built database containing standard spectral fingerprints of hundreds or thousands of known substances (such as various building materials, coatings, oil stains, chemicals, etc.).

[0112] For each endmember, the algorithm finds one or more best-matching candidate substances from the standard library and provides a similarity score. This final identification result (e.g., in...) The point identifies substance A, in The point identification of substance B) forms a distribution map over the entire floor area. This map is the "intermediate state attribute characterizing the substance composition" described in Example 1.

[0113] This embodiment specifically includes the following steps:

[0114] Traditional methods typically use fixed lighting and camera parameters, which can easily lead to problems such as overexposure (signal saturation) or underexposure (low signal-to-noise ratio) in some areas when the reflective properties of the flooring material vary greatly (e.g., from high-gloss ceramic tiles to dark rubber flooring).

[0115] Before the formal line-by-line scan, a rapid "Z"-shaped or sparse grid movement is performed on the target ground area. Simultaneously, an auxiliary, low-resolution photodiode array or RGB camera quickly captures the light intensity of the entire area, generating a preliminary, low-precision reflectance map. .

[0116] For each pixel in the formal data acquisition The system will determine its performance based on its position in the system. The corresponding intensity value is used to adjust the driving current of the light source in real time and in a feedforward manner. and / or integration time of hyperspectral camera The adjustment strategy aims to adjust the signal strength received by the sensor. Maintain it within an ideal range, such as 70%-90% of the sensor's dynamic range.

[0117] A single pixel on a floor often contains multiple substances (such as the tile itself, water stains, and dust), i.e., a "mixed pixel". Our goal is to "decipher" the spectrum (endmembers) of each pure substance and their area proportion (abundance) from the mixed spectrum.

[0118] Suppose that for a mixed pixel, its observed spectral vector is ,in This represents the number of spectral bands. According to the linear mixture model, it can be expressed as: ,in, It is an endmember matrix, and each column of it... Representing the The spectrum of a pure substance; It is an abundance vector, and each element of it is a vector. Representing the The area percentage of a particular substance in that pixel; It is the total amount of pure matter in that pixel; It is a noise vector.

[0119] PINN model In pixel coordinates As input, it simultaneously outputs the predicted endmember matrix. abundance vector The training objective of the model is to minimize a compound loss function. : ,in These are the parameters of the neural network. , , It is the weight hyperparameter for each loss.

[0120] Data reconstruction loss Ensure that the spectrum reconstructed by the model is consistent with the observed spectrum. ,in, It is the total number of pixels used for training. It is the first The observed spectrum of each pixel, The model is for the first The endmember matrix and abundance vector predicted for each pixel. This represents the square of the L2 norm, which is the square of the Euclidean distance.

[0121] The Hapk model is introduced to constrain the shape of the endmember spectrum. The Hapk model describes the behavior of light scattering on granular surfaces. A simplified Hapk model can be used to represent reflectance. Expressed as single-scattering albedo A function of other physical parameters forces the network to predict each endmember. All of them should conform to this physical law.

[0122] Physical model constraint loss : ,in, Represents the Hapk reflection model function. The model is the first Endmember prediction, wavelength-dependent Varying single-scattering albedo, Representative and the The set of other physical parameters associated with each endmember (such as phase function parameters, surface roughness, etc.) can be predicted by the network or set to constant values. This loss term penalizes oddly shaped spectra that do not conform to the laws of optical physics, making the unmixing results more stable and realistic.

[0123] Abundance constraint loss to ensure that the abundance vector conforms to its physical meaning : ,in, It is a row vector consisting entirely of 1s, with the first term having a forced sum of abundance of 1. The first term sets all positive abundance values ​​to 0, keeping only the negative values. The second term penalizes any negative abundance value, forcing it to be non-negative.

[0124] By minimizing this composite loss function A neural network was trained that can robustly extract pure substance spectra that conform to physical laws from complex mixed spectra.

[0125] To obtain pure endmember spectra Then, it was compared with the standard spectral library. In The reference spectra of a known substance were matched.

[0126] This embodiment uses a weighted hybrid similarity measurement algorithm. ,in, It is an extracted endmember spectrum. It is a reference spectrum in the library. It is a weighting factor that can be adjusted according to specific application scenarios.

[0127] First item Standard spectral angular similarity is used to measure shape: .

[0128] Second item It is a normalized Euclidean distance similarity used to measure magnitude difference: The normalization design of the denominator ensures that its range also falls within the range of... The range is not sensitive to overall amplitude scaling.

[0129] Finally, for each end element The optimal matching reference spectral index is found by solving the following optimization problem. Ultimately, it will be linked to the index. The corresponding substance identity, pre-defined in the standard spectral library, serves as this endmember. The recognition results.

[0130] Example 3

[0131] This embodiment details how to acquire and analyze active ultrasonic data in the overall testing process. First, data acquisition is performed. Second, key physical parameters of the material under test are acquired in real time on-site. Then, using these parameters, a focusing algorithm is used to reconstruct an extremely clear internal cross-sectional image from the original data. Finally, the image is automatically interpreted to locate and quantify internal defects.

[0132] like Figure 3 As shown, methods for obtaining active ultrasonic datasets and characterizing the near-surface internal structure of the region include:

[0133] (1) An ultrasonic transducer array is used, in which each element in the array sequentially transmits a pulse-compressed and encoded excitation signal, and all elements in the array synchronously receive the echo, thus collecting a full matrix dataset covering all transmit-receive channel combinations.

[0134] The pulse compression coding excitation signal is a complex acoustic signal that has been specially encoded, has a longer duration, and higher energy. Through pulse compression processing technology, the signal energy can be refocused at the receiving end, greatly improving the signal's ability to penetrate the ground (especially high-attenuation materials such as concrete) and its anti-interference ability without sacrificing detection accuracy.

[0135] The operation of the full matrix dataset is as follows: the first element in the array transmits a signal, and all other elements (including the first element) are responsible for receiving the echo; then the second element transmits, and all elements receive again... and so on, until all elements have worked as transmitters at least once. Ultimately, the system will collect a huge dataset containing all "transmit-receive" channel combinations, recording information on all possible propagation paths of the sound waves in the medium.

[0136] (2) At at least one calibration point in the target floor area, use an ultrasonic transducer array to measure the round-trip propagation time of the sound wave on the floor base with a known thickness, and calculate the local sound velocity value of the current area.

[0137] This step addresses the uncertainty of sound velocity in ultrasonic testing. The accuracy of any ultrasonic imaging is highly dependent on the local sound velocity value of the sound wave in the tested medium, and this value varies in different batches of flooring materials under different curing conditions.

[0138] Before performing large-area scanning imaging, the system performs an "in-situ calibration" at one or more "calibration points" on the target floor. This involves using an ultrasonic transducer array to precisely measure the round-trip time required for sound waves to travel vertically downwards, reach the bottom of the floor base with a known thickness, and then reflect back to the array surface. Since the design thickness of the floor is known, the sound velocity can be easily calculated.

[0139] (3) The total focusing method is applied to process the full matrix dataset, and the acoustic cross-sectional image of the near-surface internal structure of the target ground area is reconstructed using local sound velocity values;

[0140] The total focusing method is an imaging algorithm whose basic principle is: for each pixel in the desired internal cross-sectional image, the total focusing method will backtrack to the huge "full matrix dataset" and accurately and coherently superimpose the signal energy of all possible sound wave propagation paths that could pass through the pixel.

[0141] Acoustic cross-sectional images will clearly show conditions such as hollowness, delamination, cracks, or uneven aggregate distribution.

[0142] (4) Input the acoustic cross-section image into a pre-trained image segmentation model to identify and segment the defect region in the image, and quantify the geometric parameters of the defect region to generate intermediate state attributes that characterize the internal structure. The geometric parameters include depth, size or orientation.

[0143] The generated acoustic cross-sectional image is input into a pre-trained image segmentation model (such as the U-Net network commonly used in medical image analysis in deep learning). After training on a large number of labeled acoustic images, the model can automatically identify feature patterns representing defects such as "hollow areas" and "cracks" in the image. The model automatically performs pixel-level classification of the image, "segmenting" the identified defective areas from the background to form precise contours.

[0144] Once the defect area is identified, the system can automatically calculate its key "geometric parameters," such as the depth of the defect, its lateral and longitudinal dimensions, and its direction of extension (orientation) in space, which ultimately constitute the "intermediate state attributes characterizing the internal structure" described in Embodiment 1.

[0145] This embodiment specifically includes the following steps:

[0146] Traditional methods use a short, high-voltage pulse with limited energy. To address the issue of strong attenuation of ultrasonic waves by flooring materials (especially concrete), this embodiment uses a linear frequency-modulated pulse as the excitation signal. .

[0147] While ensuring the peak power is not excessively high, a large amount of total energy is injected by extending the transmission time. At the receiving end, the received echo signal... With the original excitation signal Pulse compression processing can refocus the broadened energy into a sharp pulse, thereby achieving a very high signal-to-noise ratio gain (typically 20-30dB) without sacrificing resolution.

[0148] Linear ultrasonic transducer array containing The data acquisition process for each array element includes:

[0149] The first element transmits the coded signal, and all elements in the array... Each array element simultaneously acts as a receiver, recording... The echo time-domain signal is denoted as ,in .

[0150] The second array element is launched, all of them. Each array element receives and records the data again. echo signal .

[0151] ...

[0152] No. Each element is emitted, all The last reception of each array element was recorded. echo signal .

[0153] Total received These time-domain waveform signals constitute a three-dimensional full matrix dataset. (Dimension: launch channel) Receive channel ,time This dataset contains information on all possible propagation paths of sound waves in a medium.

[0154] The total focusing algorithm discretizes the near-surface region to be imaged into individual pixels. For each pixel, using the FMC dataset, the energy of all acoustic signals "passing" through that point is coherently superimposed. If there is indeed a defect at that point (such as voids or aggregate boundaries), the signals will be superimposed in phase, resulting in enhanced energy; if the point is a homogeneous medium, the signals will be superimposed out of phase, canceling each other out.

[0155] Reconstruct any point in the image pixel grayscale value It is calculated using the following formula: .

[0156] in, : Represents the coordinates in the image. The final grayscale value (intensity) of the pixel.

[0157] It is the total number of array elements in the ultrasonic transducer array.

[0158] and These are the index numbers of the transmitting and receiving array elements, ranging from 1 to... .

[0159] From the full matrix dataset Extracted from the launch array elements Transmitted, received by array elements The received time-domain signal. Typically, a Hilbert transform is applied to it before computation to obtain an analytic signal and its envelope.

[0160] It is the calculated theoretical flight time, that is, the time it takes for the sound wave to travel from the transmitting array element. Departure, arrival at the scattering point Then reflect back to the receiving array element Total time required.

[0161] and These are the launch array elements. and receiving array elements Coordinates on the array surface (assuming the array surface is...) ).

[0162] It is the speed at which ultrasonic waves propagate in the flooring material.

[0163] This means that after summing all signals using complex numbers, the absolute value (modulus) is taken to obtain the coherent superposition energy at that point.

[0164] By analyzing each pixel in the ROI Performing the above calculations yields an acoustic cross-sectional image focused on each pixel.

[0165] The U-Net network was trained using a batch of TFM images manually labeled by professional nondestructive testing personnel. Different regions in the images, such as "hollow", "crack", "dense aggregate" and "background material", were labeled to obtain a pre-learned deep learning image segmentation network. The network takes the acoustic cross-sectional image generated by TFM as input.

[0166] The trained image segmentation network can perform pixel-level classification on new TFM images and output a defect mask. On this mask, image processing algorithms (such as connected component analysis) are used to automatically calculate key geometric parameters such as the depth (z-coordinate), lateral dimension (span in the x-direction), area, and principal axis orientation of each defect region, which ultimately constitute the intermediate state attributes representing the internal structure as described in Example 1.

[0167] Example 4

[0168] This embodiment details how to specifically acquire and analyze friction-related data in the overall testing process to deeply characterize the dynamic frictional properties of the floor surface.

[0169] like Figure 4 As shown, the method for obtaining the friction force dataset and characterizing the dynamic frictional properties of the surface in this region includes:

[0170] (1) A friction probe integrating a multi-axis force sensor and a vibration sensor is used to simultaneously collect a set of multi-dimensional force vector data and a set of vibration signal data during the process of moving relative to the ground surface with a preset slip ratio and constant normal pressure.

[0171] Slip ratio refers to the small but constant difference between the sliding speed of the probe and the moving speed of the robot body.

[0172] Multi-axis force sensors can simultaneously measure forces in multiple dimensions (such as front-back, left-right, and up-down). Vibration sensors are used to collect and record subtle, high-frequency vibration signals generated when the probe rubs against the floor surface.

[0173] (2) The dynamic friction coefficient time series is calculated based on the multidimensional force vector data, and the vibration signal data is subjected to time-frequency analysis to extract the vibration feature spectrum.

[0174] Based on the collected multidimensional force vector data, the "dynamic friction coefficient time series" that varies with time and location can be calculated using physical formulas (e.g., the ratio of tangential resultant force to normal force), reflecting the continuous change in the magnitude of friction force on the floor surface.

[0175] The system performs time-frequency analysis on the collected vibration signal data to reveal the frequency components of the vibration signal at each moment, and then extracts the vibration characteristic spectrum. The vibration characteristics of vibrations that rub against a smooth surface are different from those that rub against a rough surface with tiny particles.

[0176] (3) The dynamic friction coefficient time series and the vibration feature spectrum are fused into a composite feature vector, and the composite feature vector is input into a preset evaluation model to output the classification or quantitative evaluation results of the surface dynamic friction characteristics, so as to generate intermediate state attributes characterizing dynamic friction characteristics.

[0177] This embodiment specifically includes the following steps:

[0178] The friction probe includes a contact plate mounted on a standard elastomer (such as neoprene), and an electromagnetic actuator is responsible for applying and maintaining a constant normal pressure. .

[0179] Multi-axis force sensors can simultaneously measure forces in three directions. and torque in three directions The vibration sensor picks up the high-frequency vibration signals generated during friction due to surface irregularities and stick-slip effects. .

[0180] Instantaneous dynamic friction coefficient Defined by the ratio of the resultant tangential frictional force to the normal force: ;in, This represents the force component (longitudinal friction force) measured along the robot's forward direction. This is the force component (lateral friction) that I measured perpendicular to the robot's direction of travel. The normal pressure perpendicular to the floor surface is maintained constant by the controller. Represents time.

[0181] Vibration signals were analyzed using short-time Fourier transform. Generate a speech graph. The calculation formula is: ,in, It is the signal at a certain point in time. and frequency point The complex value of the modulus squared It is the intensity of the graph. It is the original vibration time-domain signal. A window function (such as the Hanning window) is used to extract a small segment of a signal for analysis.

[0182] Extract vibration feature spectra (or feature vectors) from the generated graph. ,For example:

[0183] Spectral centroid: The "center" frequency of the energy spectrum, reflecting the "pitch" of the vibrating sound.

[0184] Spectral bandwidth: The width of the energy spectrum, reflecting the "richness" or "noise" of the sound.

[0185] Energy ratio in a specific frequency band: For example, the ratio of high-frequency band (>10kHz) energy to total energy can reflect the degree of microscopic roughness.

[0186] At every point in time This includes information in two dimensions: the coefficient of kinetic friction. and vibration eigenvectors They are merged into a composite feature vector. .

[0187] Finally, this composite feature vector is fed into a pre-trained machine learning classifier (e.g., Support Vector Machine (SVM) or Gradient Boosting Tree (GBT)). This classifier is trained to identify different “friction patterns” and outputs a final classification assessment of the surface’s dynamic friction properties.

[0188] Example output categories:

[0189] Category 1: Cleaning and Drying (Higher frequency, vibration energy concentrated in the mid-frequency range).

[0190] Category 2: Liquid water pollution Extremely low, vibrations are damped by the liquid, energy is weak and concentrated at low frequencies.

[0191] Category 3: Oily contamination It is low, but its viscosity produces unique low-frequency vibration modes that are different from those of water.

[0192] Category 4: Particulate dust pollution It may not be low, but the vibration signal contains a large number of random, high-frequency impact components, and the spectral bandwidth is very wide.

[0193] Example 5

[0194] This embodiment details how to deeply integrate multiple independent "intermediate state attributes" (such as chemical composition diagrams, internal defect diagrams, and friction characteristic diagrams) generated in the preceding steps during the overall testing process, in order to generate a final digital model that can comprehensively and holistically reflect the condition of the floor.

[0195] The method for fusing the multiple intermediate state attributes into a comprehensive state representation includes:

[0196] (1) The geometric structure of the target ground area is mapped into a spatial graph, wherein the nodes of the spatial graph correspond to the predetermined cells in the area, the edges of the spatial graph represent the spatial adjacency relationship between the cells, and the initial feature vector of each node is composed of the multiple intermediate state attributes of its corresponding cell.

[0197] (2) A graph neural network (GNN) is used to perform multiple rounds of iterative message passing and node update operations on the spatial graph. In each round of operation, each node updates its own feature representation by aggregating the features of its neighboring nodes.

[0198] Aggregation means "looking around" at all neighboring nodes that are directly connected to it by "edges" and "collecting" their feature vector information.

[0199] The updated information is aggregated from the neighbors and combined with its own feature vector from the previous round. This process is then performed through a neural network layer to generate a new, more informative feature vector for itself.

[0200] Through multiple iterations, information spreads throughout the graph network, so that the final feature representation of each node incorporates the contextual information of its surrounding environment.

[0201] (3) The final set of node feature vectors output by the graph neural network after processing is defined as the comprehensive state representation.

[0202] The comprehensive status representation is a structured, overall digital model that can reflect the spatial relationships between various floor conditions.

[0203] This embodiment specifically includes the following steps:

[0204] Node definition: Dividing the ground map into equally sized cells or irregular superpixels generated by image segmentation algorithms. Each cell / superpixel becomes a node in the map. One of the nodes .

[0205] For each node Initial feature vector It is composed of all intermediate state attributes on the ground area corresponding to that node. For example, in the location... The node at the location Its initial feature vector is: ,in, Chemical composition attribute vectors derived from hyperspectral analysis (such as probability distributions of various pollutants). Internal structural attribute vectors derived from ultrasonic analysis (such as defect category and depth value). Frictional property vectors derived from frictional analysis (such as friction coefficient values ​​and vibration characteristics). This represents the concatenation operation of vectors.

[0206] Definition of an edge: If two nodes (cells) are spatially adjacent (e.g., sharing an edge or a vertex), an edge is created between them. .

[0207] After constructing the graph, a graph convolutional neural network (GNN) is used to learn the complex relationships between nodes. GNNs achieve this by stacking multiple graph convolutional layers. For the... The update rule for the node features of a graph convolutional layer can be expressed as: .

[0208] in, It is the first All layers The feature matrix of each node. The feature dimension of each node is... In particular, This refers to the initial node feature matrix constructed in the previous step.

[0209] It is the first The weight matrix that the layer network needs to learn is trained using the backpropagation algorithm.

[0210] It is the adjacency matrix of the graph Add an identity matrix .

[0211] yes The degree matrix (a diagonal matrix), whose diagonal elements yes Matrix number The sum of the elements in the row.

[0212] It performs symmetric normalization on the adjacency matrix to prevent node features from becoming too large or too small during multi-level transmission.

[0213] It is a non-linear activation function, such as ReLU. ), used to increase the expressive power of the model.

[0214] The update process will repeat. Second-rate( (This refers to the total number of layers in the GNN). In each round of iterations, the final feature vector of each node... It not only includes its own initial information, but also incorporates... Information about all nodes within the jump neighbor range.

[0215] After processing by GNN, the final node feature matrix is ​​obtained. Each of its rows It is a node The final feature representation. For example, a node may initially only have two isolated attributes: "crack" and "high friction". But through GNN, it learns from its neighboring nodes that it is on a long, linear "crack zone", and the final feature representation will contain the pattern information of this "linear structure".

[0216] Example 6

[0217] This embodiment details how to perform "quantitative comparison" and output the final quality inspection results in the final stage of the overall testing process.

[0218] Methods for quantitative comparison and outputting quality inspection results include:

[0219] (1) A generative deep learning model trained on the comprehensive state representation of multiple standard compliant floor samples is used as the standard reference model.

[0220] (2) Input the comprehensive state representation of the ground into the standard reference model and obtain the reconstructed comprehensive state representation generated by the standard reference model.

[0221] The "comprehensive state representation" of the floor to be inspected (i.e., the digital model generated in Example 5 that represents its current actual condition) is input into the "standard reference model" trained in the previous step. Since this model only "recognizes" the pattern of compliant samples, it will try to reproduce (i.e. "reconstruct") the input representation using its own understanding of the "perfect standard", thereby generating a "reconstructed comprehensive state representation".

[0222] (3) Calculate the reconstruction error between the comprehensive state representation and the reconstructed comprehensive state representation, and quantify the reconstruction error into an anomaly score that characterizes the degree to which the floor to be accepted deviates from the quality standard.

[0223] If the test surface is compliant, its data pattern is consistent with the "perfect standard" learned during model training, and the model can easily reconstruct it with high quality, so the difference between the two representations will be very small.

[0224] If the test floor has defects, its data contains "abnormal" patterns that the model has never seen before. The model will have difficulty understanding and reproducing these patterns, resulting in poor reconstruction results and a large difference between the two representations.

[0225] This difference is precisely calculated using mathematical formulas (such as mean squared error) and then quantified.

[0226] (4) The abnormal score is compared with the preset judgment threshold to generate a quality detection result.

[0227] If the abnormal score is lower than or equal to the threshold, the system generates a "qualified" quality inspection result.

[0228] If the abnormal score is higher than the threshold, a "non-compliant" quality inspection result is generated.

[0229] This embodiment specifically includes the following steps:

[0230] Model selection: Variational Autoencoder (VAE) or Generative Adversarial Network (GAN). VAE will be used as an example here.

[0231] Training data: A large amount of data was collected from "golden sample" flooring that fully complies with design specifications and has excellent construction quality. This data underwent the same pre-processing to obtain a series of perfect comprehensive state representations.

[0232] Training process: A VAE consists of an encoder and a decoder. The encoder learns to compress the input data into a low-dimensional, statistically well-distributed latent space, while the decoder learns to recover (reconstruct) the data from this latent space as losslessly as possible. The goal of training is to make the reconstructed data as similar as possible to the original input data.

[0233] Final Model: After training, this VAE becomes the standard reference model.

[0234] When the measured comprehensive condition of a piece of flooring awaiting acceptance indicates When input into a trained VAE model, the model obtains a reconstructed comprehensive state representation. .

[0235] If the input is compliant: and They will be very close.

[0236] If the input is non-compliant: and The reconstruction error between them can be very large.

[0237] Calculate the error fraction using the mean squared error: ,in, It refers to the number of floor cells (nodes). It is the dimension of the feature vector of each node.

[0238] Based on experience or industry standards, a judgment threshold is pre-set. .

[0239] if If the paving is deemed satisfactory, then the area is deemed to be "qualified".

[0240] if If so, the area is deemed "unqualified in terms of paving".

[0241] The final inspection report can include an "anomaly score heatmap" that visually highlights which areas have the largest reconstruction errors, thereby accurately locating the specific location of construction defects.

[0242] Example 7

[0243] This embodiment further optimizes Embodiment Six, providing a better method for generating quality inspection results, including:

[0244] (1) The extreme value theory (EVT) is used to statistically model the abnormal score set obtained on multiple standard compliant floor samples and establish a probability judgment model.

[0245] (2) For the abnormal score of each cell of the floor, the confidence score of the cell belonging to the standard compliant floor is calculated using the probability judgment model. The confidence score represents the probability that the cell meets the quality standard.

[0246] The confidence score is a value between 0 and 1, representing the probability that the cell meets the quality standard. For example, a confidence score of 0.99 means that there is a 99% certainty that the cell's outlier score is within the "normal" range; while a confidence score of 0.15 means that there is an 85% probability that it is outlier (non-compliant).

[0247] (3) Generate quality inspection results based on the confidence score, and output the source of defects in non-compliant areas by analyzing the intermediate state attribute components that constitute the cell reconstruction error.

[0248] The system can generate a final conclusion based on a "confidence score." For example, a confidence threshold (such as 95%) can be set; a score above this threshold is considered "qualified," while a score below it is considered "unqualified."

[0249] Anomaly scores originate from reconstruction errors, which are composed of intermediate state attribute components representing different physical dimensions (i.e., hyperspectral, ultrasonic, and frictional characteristics). By analyzing which component has the largest error, the primary source of the defect can be inferred. For example, if the reconstruction error of a cell mainly comes from the ultrasonic attribute component, the system can make an inference.

[0250] This embodiment includes the following specific steps:

[0251] Collect all outlier scores generated during the training phase for "standard compliant flooring" to form a baseline distribution. Select a high quantile (e.g., the 95th quantile) as the initial threshold. For all exceeding the threshold The scores were fitted using a generalized Pareto distribution (GPD). The GPD is the standard distribution used in EVT for modeling data exceeding the threshold.

[0252] The cumulative distribution function of GPD is: ,in, This refers to an abnormal score. Exceeding the threshold The amount. (shape parameters) and (Scale parameters) are model parameters obtained through the fitting process.

[0253] A new anomaly score calculated for a cell of flooring awaiting inspection. Using the established probability judgment model, a ratio is calculated. The probability of more extreme scores appearing in the "standard compliance" distribution, i.e., the p-value. ,in, The score in the training samples exceeds Quantity, This represents the total number of training samples. The smaller the p-value, the lower the likelihood that the paving stone is compliant.

[0254] Define confidence score And set a confidence threshold, such as 95% (i.e., 0.95).

[0255] If a cell If so, its compliance conclusion is "qualified".

[0256] if If so, the conclusion is "unqualified".

[0257] The process of calculating cell reconstruction error, i.e. and The difference between the two vectors is that they are composed of different intermediate state attributes (chemical, structural, frictional), so the partial reconstruction error of each attribute can be calculated:

[0258]

[0259]

[0260]

[0261] If an invalid cell If the result is significantly greater than the other two, the system can automatically generate a defect source analysis conclusion in the final quality inspection report: "Area X is unqualified (confidence level 12%), the main reason may be internal structural defects (such as hollowness or insufficient density)."

[0262] Example 8

[0263] To ensure that the measurement accuracy and data consistency of the entire detection system, especially its core sensor components, have been verified and calibrated before performing formal detection tasks, the method also includes a system calibration step, which includes:

[0264] The multimodal sensor assembly is instructed to measure a standard reference containing multiple regions of known physical properties, and the multimodal sensor assembly is calibrated based on the difference between the measured data and the known physical properties. The known physical property regions include at least a standard spectral reflectance region, a standard acoustic property region, or a standard friction coefficient region.

[0265] First, the detection system equipped with multimodal sensors performs a complete measurement scan of the standard reference object. Second, the real-time measurement data collected by the sensors is compared with the pre-stored real physical characteristic values ​​of the reference object to identify and quantify the current measurement error of the system. Finally, based on this error, the system automatically generates and applies corresponding calibration parameters to correct the sensors or subsequent processing algorithms, thereby ensuring the accuracy and reliability of the data collected in the subsequent actual floor inspection tasks.

[0266] Example 9

[0267] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the sensing and detection method for high-wear-resistant composite flooring as described above.

[0268] Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instruction data structures, program modules, or other data. Computer storage media includes RAM, ROM, EPROM, EEPROM, flash memory or other solid-state storage technologies, CD-ROM, DVD or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The aforementioned system memories and mass storage devices can be collectively referred to as memory.

[0269] A computer program product includes a computer program / instructions that, when executed by a processor, implement the sensing and detection method for high-wear-resistant composite flooring as described above.

[0270] Computer program products include computer programs or instruction sets used to perform specific tasks or achieve specific functions. These programs or instructions are designed to be executed by a processor to implement a series of predefined steps or operations. The program product may be stored in various forms of computer storage media, such as memory, hard disks, solid-state drives, optical discs, or other forms of digital storage devices. It may exist in the form of compiled binary code or in the form of scripts or bytecode that can be executed by an interpreter. Through carefully designed algorithms and logical instructions, the program product enables the processor to process data in a specific order and manner, performing various functions such as data analysis, user interaction, and device control.

[0271] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0272] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0273] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above invention, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A construction method for a high wear-resistant composite floor, characterized in that, include: Preparation of the base layer: On the treated ground surface, lay and vibrate the concrete base layer; Apply a wear-resistant intermediate layer: When the concrete base layer reaches the predetermined plastic state, a wear-resistant mortar layer containing diamond aggregate is evenly laid on it to form a high-hardness wear-resistant base. Laying the decorative surface layer: Before the wear-resistant intermediate layer initially sets and solidifies, a terrazzo surface layer material containing decorative aggregate is laid on its surface to ensure a strong chemical and physical bond between the two layers; Curing and finishing: The laid composite floor is cured and maintained in a standardized manner, and then its surface is subjected to multi-stage coarse grinding, fine grinding and chemical polishing until a composite finished surface is formed that exposes the decorative aggregate and part of the diamond aggregate. Multimodal quality acceptance inspection: A fusion inspection method based on multimodal sensing is used to conduct the final quality acceptance of the finished surface.

2. A sensing and detection method for high wear-resistant composite flooring, characterized in that, The sensing and detection method is used to perform the multimodal quality acceptance testing as described in claim 1, comprising: A set of interconnected datasets is collected in the target ground area. The datasets include at least: a hyperspectral dataset for characterizing the material composition of the area, an active ultrasonic dataset for characterizing the near-surface internal structure of the area, and a frictional force dataset for characterizing the dynamic frictional properties of the surface of the area. Each type of data in the dataset undergoes preliminary processing to generate multiple intermediate state attributes representing different physical dimensions. In a common spatial coordinate system, multiple intermediate state attributes are fused to obtain the current measured comprehensive state representation of the ground surface; The measured comprehensive state representation is input into the pre-trained standard reference model for quantitative comparison, and the quality inspection results of the floor are output.

3. The sensing and detection method for high wear-resistant composite flooring according to claim 2, characterized in that, The method for obtaining the hyperspectral dataset and characterizing the material composition of the region includes: A reflection intensity map of the target ground area is generated by pre-scanning, and based on the reflection intensity map, at least one acquisition parameter is dynamically adjusted during the formal acquisition to obtain a hyperspectral data cube. The acquisition parameter is selected from the light source irradiance intensity or the sensor integration time. The hyperspectral data cube is input into a physical information neural network, which demixes the mixed pixels to extract a set of endmembers representing the spectral characteristics of pure substances and their respective abundances. The endmembers are compared with reference spectra in a standard spectral library, and a weighted similarity measurement algorithm is used to determine the best-matching material component for each endmember, so as to generate intermediate state attributes that characterize the material components.

4. The sensing and detection method for high wear-resistant composite flooring according to claim 2, characterized in that, Methods for acquiring active ultrasonic datasets and characterizing the near-surface internal structure of the region include: An ultrasonic transducer array is used, in which each element of the array sequentially transmits a pulse-compressed and encoded excitation signal, and all elements of the array synchronously receive the echo, thus acquiring a full matrix dataset covering all transmit-receive channel combinations. At at least one calibration point in the target floor area, the round-trip propagation time of sound waves on a floor base layer of known thickness is measured using an ultrasonic transducer array, and the local sound velocity value of the current area is calculated. The full matrix dataset is processed using the total focusing method, and the acoustic cross-sectional image of the near-surface internal structure of the target ground area is reconstructed using local sound velocity values; The acoustic cross-sectional image is input into a pre-trained image segmentation model to identify and segment defect regions in the image, and the geometric parameters of the defect regions are quantified to generate intermediate state attributes characterizing the internal structure. The geometric parameters include depth, size, or orientation.

5. The sensing and detection method for high wear-resistant composite flooring according to claim 2, characterized in that, The method for obtaining the friction force dataset and characterizing the dynamic frictional properties of the surface in this region includes: A friction probe integrating a multi-axis force sensor and a vibration sensor is used to simultaneously collect a set of multi-dimensional force vector data and a set of vibration signal data during the process of moving relative to the ground surface with a preset slip ratio and constant normal pressure. The dynamic friction coefficient time series is calculated based on the multidimensional force vector data, and the vibration signal data is subjected to time-frequency analysis to extract the vibration feature spectrum; The dynamic friction coefficient time series and the vibration feature spectrum are fused into a composite feature vector, and the composite feature vector is input into a preset evaluation model to output the classification or quantitative evaluation results of the surface dynamic friction characteristics, so as to generate intermediate state attributes characterizing dynamic friction characteristics.

6. The sensing and detection method for high wear-resistant composite flooring according to claim 2, characterized in that, The method for fusing the multiple intermediate state attributes into a comprehensive state representation includes: The geometry of the target ground area is mapped into a spatial graph, wherein the nodes of the spatial graph correspond to predetermined cells within the area, the edges of the spatial graph represent the spatial adjacency relationship between the cells, and the initial feature vector of each node is composed of the multiple intermediate state attributes of its corresponding cell. A graph neural network is used to perform multi-round iterative message passing and node update operations on the spatial graph. In each round of operation, each node updates its own feature representation by aggregating the features of its neighboring nodes. The final set of node feature vectors output after processing by the graph neural network is defined as the comprehensive state representation.

7. The sensing and detection method for high wear-resistant composite flooring according to claim 2, characterized in that, Methods for quantitative comparison and outputting quality inspection results include: A generative deep learning model trained on a comprehensive state representation of multiple standard compliant floor samples is used as the standard reference model. The comprehensive state representation of the ground is input into the standard reference model, and the reconstructed comprehensive state representation generated by the standard reference model is obtained. Calculate the reconstruction error between the comprehensive state representation and the reconstructed comprehensive state representation, and quantify the reconstruction error as an anomaly score characterizing the degree to which the floor to be accepted deviates from the quality standard; The quality inspection result is generated by comparing the abnormal score with a preset judgment threshold.

8. The sensing and detection method for high wear-resistant composite flooring according to claim 7, characterized in that, Methods for generating quality inspection results include: Extreme value theory is used to statistically model the set of abnormal scores obtained from multiple standard compliant floor samples, and a probability judgment model is established. For each cell of the flooring, the anomaly score is calculated using a probability judgment model to determine whether it belongs to the standard compliant flooring. The confidence score represents the probability that the cell meets the quality standard. Based on the confidence score, a quality inspection result is generated, and by analyzing the intermediate state attribute components that constitute the reconstruction error of the cell, the source of defects in the non-compliant area is output.

9. The sensing and detection method for high wear-resistant composite flooring according to claim 2, characterized in that, It also includes a system calibration step, which includes: The multimodal sensor assembly is instructed to measure a standard reference containing multiple regions of known physical properties, and the multimodal sensor assembly is calibrated based on the difference between the measured data and the known physical properties. The known physical property regions include at least a standard spectral reflectance region, a standard acoustic property region, or a standard friction coefficient region.

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 sensing and detection method for high wear-resistant composite flooring as described in any one of claims 2-9.

Images