Building structure detection data anomaly identification method and system

Abstract

The application provides a building structure detection data anomaly identification method and system, and relates to the technical field of building structure detection.The method comprises the following steps: adopting an adaptive data management mechanism to dynamically weight and fuse the original data stream of building structure detection and to adaptively time window regularize, embedding pre-acquired environmental parameters, component construction information and historical quality trajectories of similar components to obtain a working condition enhanced detection data set carrying the historical quality trajectories; performing working condition deviation compensation calculation and dynamic threshold fitting on the working condition enhanced detection data set to generate an adaptive dynamic reference interval and obtain a deviation mapping feature sequence; dynamically identifying a plurality of feature attractors from the deviation mapping feature sequence, taking the plurality of feature attractors as control vertices to construct a minimum external envelope polytope.The application can realize accurate anomaly identification of detection data, confirm abnormal causes and traceability paths, and effectively meet the actual needs of a plurality of building structure detection scenes.

Description

Technical Field

[0001] This invention relates to the field of building structure testing technology, and in particular to a method and system for identifying anomalies in building structure testing data. Background Technology

[0002] Routine monitoring of old buildings, construction quality control of new buildings, and operation and maintenance inspection of special structures such as industrial plants and bridges all require accurate identification of anomalies in inspection data to promptly identify structural hazards and mitigate safety risks, providing a scientific basis for structural repair, reinforcement, and management decisions. With the development of the Internet of Things and sensor technology, building structure inspection has entered a stage of real-time acquisition of multi-source data. Sensor networks can simultaneously acquire multi-dimensional inspection data such as concrete strength, rebar cover thickness, and structural flatness.

[0003] However, most current mainstream methods for identifying anomalies in inspection data suffer from the following technical shortcomings: they fail to fully integrate operational information and historical quality trajectories of similar components during the anomaly identification process, relying solely on single inspection data or fixed thresholds for judgment, making it difficult to adapt to complex and ever-changing actual inspection scenarios. This deficiency leads to low anomaly identification accuracy and a high likelihood of misjudgments and missed detections. For example, in the inspection of old residential buildings, normal data fluctuations caused by environmental parameters are often misjudged, and early-stage hidden structural damage may be missed due to the lack of integration with historical trajectories, delaying the handling of potential hazards. Furthermore, it results in a lack of reliability and traceability of anomaly results, making it difficult to determine the cause of the anomaly and the data tracing path, causing inconvenience for subsequent hazard investigation and maintenance. This is especially problematic in the inspection of large industrial plants, where it makes it difficult for maintenance personnel to accurately locate the root cause of the anomaly, increasing maintenance costs and workload. Summary of the Invention

[0004] This invention provides a method and system for identifying anomalies in building structure inspection data, which can achieve accurate anomaly identification in inspection data, confirm the cause of the anomaly and trace the source path, and effectively meet the actual needs of various building structure inspection scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for identifying anomalies in building structure inspection data, the method comprising: An adaptive data governance mechanism is used to dynamically weight and fuse the raw data stream of building structure inspection and perform adaptive time window normalization. Pre-acquired environmental parameters, component construction information and historical quality trajectories of similar components are embedded to obtain a condition-enhanced inspection dataset carrying historical quality trajectories. The working condition enhancement detection dataset is subjected to working condition deviation compensation calculation and dynamic threshold fitting to generate an adaptive dynamic benchmark interval and obtain a deviation mapping feature sequence. Multiple feature attractors are dynamically identified from the deviation mapping feature sequence. Using these multiple feature attractors as control vertices, a minimum externally oriented envelope polycell is constructed. The minimum externally oriented envelope polycell is subjected to an affine transformation and fitted to a standard hyperellipsoidal space to generate a geometrically constrained hypersurface. The correlation weighted orthogonal statistics of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface are calculated to obtain the geometrically enhanced feature sequence. The geometrically enhanced feature sequence is input into the pre-trained hybrid anomaly recognition model, and static threshold screening, temporal evolution trend analysis and multi-dimensional weight correlation inference are performed sequentially to obtain anomaly discrimination probability sequence and anomaly type label. The confidence level of the anomaly discrimination probability sequence and the anomaly type label are verified and logically mutually exclusive. After obtaining the verification result, associated operating condition parameters and data traceability path, the results are structured and encapsulated to obtain anomaly identification alarm instructions and send them to the management terminal.

[0006] Furthermore, an adaptive data governance mechanism is employed to dynamically quality-weightedly fuse and adaptively normalize the raw data stream of building structure inspection, embedding pre-acquired environmental parameters, component construction information, and historical quality trajectories of similar components to obtain a condition-enhanced inspection dataset carrying historical quality trajectories, including: The signal-to-noise ratio and timestamp sparsity of each data source in the original data stream are analyzed, and the data quality weight of each data source is dynamically calculated to obtain a multi-source data stream with quality weights. Based on the multi-source data stream with quality weights, the width of the regularized window is adaptively adjusted according to the timestamp distribution of each data source, and time window alignment is performed to obtain a time-synchronized multi-source data sequence. Based on the aforementioned data quality weights, the time-synchronized multi-source data sequences are weighted and fused to form an initial fused dataset; The collected environmental parameters, component construction information, and historical quality trajectories of similar components are used as embedded information and are temporally associated and bound with the initial fusion dataset to form a condition-enhanced detection dataset carrying historical quality trajectories.

[0007] Furthermore, the working condition enhancement detection dataset is subjected to working condition deviation compensation calculation and dynamic threshold fitting to generate an adaptive dynamic benchmark interval, resulting in a deviation mapping feature sequence, including: Environmental parameters and component construction information are extracted from the enhanced working condition detection dataset. The deviation compensation amount of the current working condition relative to the standard working condition is calculated to obtain the working condition compensation coefficient. The detection values ​​in the enhanced working condition detection dataset are corrected based on the working condition compensation coefficient to eliminate the influence of working condition deviation and obtain the working condition normalized detection sequence. The mean and standard deviation within each window are calculated using the sliding window method for the normalized test sequence under working conditions. The upper and lower boundaries of the adaptive dynamic benchmark interval are dynamically fitted to generate the adaptive dynamic benchmark interval. The deviation of each detection value in the normalized detection sequence under operating conditions relative to the adaptive dynamic reference interval is calculated to form a deviation mapping feature sequence.

[0008] Furthermore, multiple feature attractors are dynamically identified from the bias mapping feature sequence. These feature attractors are gradient mutation stationary points in the temporal-parameter joint space of the bias mapping feature sequence, with their coordinates precisely located at the intersection of a specific timestamp and the corresponding parameter dimension. Using these multiple feature attractors as control vertices, a minimum outer envelope polytope is constructed, including: The directional derivative magnitude of each point in the time-parameter joint space is calculated for the deviation mapping feature sequence. The local maxima of the directional derivative magnitude are located as candidate feature attractors. The precise coordinates of each candidate feature attractor are recorded. The precise coordinates are composed of a specific timestamp and the corresponding concrete strength deviation value, rebar cover thickness deviation value, rebar spacing deviation value, and flatness deviation value, thus obtaining an initial set of feature attractors. Non-maximum suppression is applied to the initial set of feature attractors to remove redundant stationary points whose linear distance in the feature space is less than a preset distance threshold, and feature attractors whose gradient direction angle between adjacent feature attractors is greater than a preset angle threshold, thus forming an optimized set of feature attractors. Using the precise coordinates of each feature attractor in the optimized feature attractor set as control vertices, a geometric method of constructing a minimum closed polytope around the control vertices is adopted. The outer vertices are connected sequentially to form an outer envelope surface, forming a minimum outer envelope polytope containing all control vertices, thus obtaining the minimum outer envelope polytope.

[0009] Furthermore, an affine transformation is performed on the minimum outer envelope multicell and fitted to the standard hyperellipsoid space to generate a geometrically constrained hypersurface. The correlation weighted orthogonal statistics of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the hypersurface normal deflection angle are calculated to obtain the geometrically enhanced feature sequence, including: Apply an affine transformation to the minimum externally enclosed polycell, stretching, rotating, and translating it to the standard hyperellipsoid space, so that the minimum externally enclosed polycell is inscribed in the standard hyperellipsoid, generating a geometrically constrained hypersurface; Project the coordinates of each point in the deviation mapping feature sequence onto the geometrically constrained hypersurface, calculate the weighted orthogonal statistic of the correlation between each projection point and the origin, and obtain the weighted orthogonal distance value of each point. Calculate the angle between the normal direction of each point in the deviation mapping feature sequence and the tangent plane of the hypersurface at this point, and obtain the hypersurface normal deflection angle at each point; The weighted orthogonal distance value and the hypersurface normal deflection angle are combined as a vector to form a geometrically enhanced feature sequence.

[0010] Furthermore, the geometrically enhanced feature sequence is input into a pre-trained hybrid anomaly recognition model. The pre-trained hybrid anomaly recognition model sequentially performs static threshold screening, temporal evolution trend analysis, and multi-dimensional weighted correlation inference to obtain anomaly discrimination probability sequences and anomaly type labels, including: The geometrically enhanced feature sequence is input into a pre-trained hybrid anomaly recognition model, and a static threshold screening is performed, including comparing each feature value in the geometrically enhanced feature sequence with a preset static threshold, marking feature points that exceed the threshold range as candidate anomalies, and obtaining a preliminary anomaly set. A time-series evolution trend analysis is performed on the candidate anomalies in the initial anomaly set, including calculating the rate of change of the feature values ​​of each candidate anomaly in the forward and backward neighborhoods on the time axis, and selecting candidate anomalies with consistent change rate signs and cumulative change exceeding the preset trend threshold to obtain a trend anomaly set. Multidimensional weighted correlation reasoning is performed on the feature points in the trend anomaly set, including calculating weighted anomaly scores based on preset weights of each parameter dimension, and fusing the correlation constraints between different dimensions to obtain anomaly discrimination probability sequences and anomaly type labels corresponding to each feature point.

[0011] Furthermore, the anomaly discrimination probability sequence and anomaly type label are subjected to confidence verification and logical mutual exclusion determination. After obtaining the verification-passing determination result, associated operating condition parameters, and data traceability path, they are structured and encapsulated to obtain anomaly identification alarm instructions, which are then sent to the management terminal, including: The confidence level of each anomaly discrimination probability value in the anomaly discrimination probability sequence is verified, and anomaly discrimination probabilities and their corresponding anomaly type labels that are lower than the preset confidence threshold are removed to obtain a high-confidence anomaly set. Logical mutual exclusion is performed on the labels of each anomaly type in the high confidence anomaly set. The system checks whether there are contradictory anomaly type labels at the same detection location and within the same time window. If they exist, the label with the highest priority is retained and the contradictory label is removed, thus obtaining the anomaly set that passes the mutual exclusion check. Extract the associated working condition parameters and data tracing paths corresponding to each anomaly point from the set of anomalies that have passed mutual exclusion verification. The associated working condition parameters include the environmental parameters and component construction information corresponding to the anomaly point. The data tracing paths include the original data stream source identifier, timestamp, and parameter dimension index. The verified judgment result, associated operating condition parameters, and data traceability path are structured and encapsulated according to the preset alarm format to form an anomaly identification alarm command, which is then sent to the control terminal.

[0012] Secondly, the building structure inspection data anomaly identification system includes: The data acquisition module is used to collect raw data streams and related data such as environmental parameters, component construction information, and historical quality trajectories of similar components for building structure inspection. The data governance module is used to perform dynamic quality weighted fusion and adaptive time window normalization on the raw data stream of building structure inspection using an adaptive data governance mechanism, and to embed environmental parameters, component construction information and historical quality trajectories of similar components to obtain a condition-enhanced inspection dataset carrying historical quality trajectories. The baseline interval generation module is used to perform working condition deviation compensation calculation and dynamic threshold fitting on the working condition enhancement detection dataset to generate an adaptive dynamic baseline interval and obtain a deviation mapping feature sequence. The geometric feature enhancement module is used to dynamically identify multiple feature attractors from the deviation mapping feature sequence, and construct a minimum outer envelope polycell using the multiple feature attractors as control vertices; perform an affine transformation on the minimum outer envelope polycell and fit it to the standard hyperellipsoid space to generate a geometrically constrained hypersurface; calculate the correlation weighted orthogonality statistic of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface to obtain the geometrically enhanced feature sequence; The anomaly detection module is used to input the geometrically enhanced feature sequence into the pre-trained hybrid anomaly detection model, and sequentially perform static threshold screening, temporal evolution trend analysis and multi-dimensional weight correlation inference to obtain anomaly discrimination probability sequence and anomaly type label. The alarm delivery module is used to perform confidence verification and logical mutual exclusion judgment on the anomaly discrimination probability sequence and anomaly type label. After obtaining the judgment result that passes the verification, the associated operating condition parameters and the data traceability path, it is structured and encapsulated to obtain the anomaly identification alarm instruction and send it to the management terminal.

[0013] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: An adaptive data governance mechanism is employed to dynamically weight and fuse multi-source raw data streams, and to regularize them over time windows. Environmental parameters, construction information, and historical quality trajectories are embedded, overcoming the problem of low anomaly identification accuracy caused by insufficient integration of working conditions and historical information in traditional methods. This achieves high-quality alignment and working condition enhancement of multi-source heterogeneous data. An adaptive baseline interval is generated through working condition deviation compensation and dynamic threshold fitting, obtaining a deviation mapping feature sequence. This overcomes the limitation of fixed thresholds in adapting to complex detection scenarios, achieving adaptive compensation for working condition fluctuations and improving the accuracy of initial anomaly screening. Feature attractors are dynamically identified from the deviation mapping feature sequence, a minimum outer envelope polycellular structure is constructed and fitted to a standard hyperellipsoidal space, generating a geometrically constrained hypersurface, and a correlation-weighted orthogonal statistic is calculated. The method incorporates hypersurface normal deflection angles to overcome the shortcomings of traditional methods in characterizing multidimensional data geometric structures and local mutation features, thereby enhancing the geometric distinguishability and expressive dimension of abnormal features. By inputting the geometrically enhanced feature sequence into the hybrid anomaly identification model, static threshold screening, time-series trend analysis, and multidimensional weighted inference are sequentially performed. This overcomes the problem of misjudgment and missed judgments easily generated by single-discrimination logic, thus achieving layer-by-layer accurate identification of anomalies from coarse screening to trend verification and then to multidimensional correlation. The anomaly discrimination results are subjected to confidence verification and logical mutual exclusion determination. After extracting associated operating condition parameters and data traceability paths, the results are structured, packaged, and distributed, overcoming the lack of reliability and traceability of anomaly results. This provides operation and maintenance personnel with verifiable and locatable alarm information, reducing the cost and workload of hazard investigation. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method for identifying anomalies in building structure inspection data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a building structure inspection data anomaly identification system provided in an embodiment of the present invention; Figure 3 This is a simulation diagram of the three-dimensional parameter space distribution of the deviation mapping feature sequence provided in the experimental example of the present invention; Figure 4 This is a statistical chart showing the distribution of the number of abnormality type labels detected in the experimental examples of this invention; Figure 5 This is a comparison chart of the performance indicators of different anomaly identification methods provided in the experimental examples of this invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, an embodiment of the present invention proposes a method for identifying anomalies in building structure inspection data, the method comprising the following steps: Step 1: Adaptive data governance mechanism is used to perform dynamic quality weighted fusion and adaptive time window normalization on the raw data stream of building structure inspection, and pre-acquired environmental parameters, component construction information and historical quality trajectories of similar components are embedded to obtain a working condition enhanced inspection dataset carrying historical quality trajectories. Step 2: Perform working condition deviation compensation calculation and dynamic threshold fitting on the working condition enhancement detection dataset to generate an adaptive dynamic benchmark interval and obtain the deviation mapping feature sequence. Step 3: Dynamically identify multiple feature attractors from the deviation mapping feature sequence, and construct a minimum outer envelope polycell using the multiple feature attractors as control vertices; perform an affine transformation on the minimum outer envelope polycell and fit it to the standard hyperellipsoid space to generate a geometrically constrained hypersurface; calculate the correlation weighted orthogonal statistics of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface to obtain the geometrically enhanced feature sequence; Step 4: Input the geometrically enhanced feature sequence into the pre-trained hybrid anomaly recognition model, and sequentially perform static threshold screening, temporal evolution trend analysis and multi-dimensional weight correlation inference to obtain anomaly discrimination probability sequence and anomaly type label; Step 5: Perform confidence verification and logical mutual exclusion judgment on the anomaly discrimination probability sequence and anomaly type label. After obtaining the judgment result that passes the verification, the associated working condition parameters and the data traceability path, perform structured encapsulation to obtain the anomaly identification alarm instruction and send it to the management terminal.

[0019] In this embodiment of the invention, by integrating multi-source data with environmental, construction, and historical information, data quality and operational adaptability are effectively improved. The use of operational deviation compensation and dynamic threshold fitting overcomes the limitations of fixed thresholds in complex scenarios, improving the accuracy of initial anomaly screening. By identifying feature attractors and constructing geometrically constrained hypersurfaces, the geometric feature expression and mutation identification capabilities of multidimensional detection data are enhanced. A hybrid anomaly identification model is used for static threshold screening, time-series trend analysis, and multidimensional weighted inference, achieving layer-by-layer anomaly discrimination from coarse screening to fine verification, effectively reducing false positives and false negatives. Confidence verification and logical mutual exclusion determination are performed on the discrimination results, and associated operational parameters and data traceability paths are encapsulated, ensuring the reliability and traceability of alarm information, facilitating rapid location of the root cause of anomalies by maintenance personnel, and effectively saving on hidden danger investigation costs.

[0020] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Analyze the signal-to-noise ratio (SNR) and timestamp sparsity of each data source in the original data stream, dynamically calculate the data quality weight of each data source, and obtain a multi-source data stream with quality weights. Specifically, this includes: calculating the SNR and timestamp distribution sparsity of each data source for the original data stream collected by multiple sensors deployed at the building structure inspection site; the SNR is determined by the ratio of the effective signal amplitude to the root mean square of the background noise, and the calculation formula is: In the formula, This represents the signal-to-noise ratio of the i-th data source. This represents the effective amplitude of the signal from the data source. The root mean square value of the background noise is used; timestamp sparsity is characterized by the number of valid data points per unit time, and the calculation formula is as follows: in, Let be the timestamp sparsity of the i-th data source. For the time period The number of valid data points in this data source. The observation time window length is specified. Each data source is weighted and scored based on signal-to-noise ratio and sparsity, and the data quality weight of each data source is dynamically calculated using the following formula: In the formula Let i be the data quality weight of the i-th data source. and These are preset adjustment coefficients for signal-to-noise ratio and sparsity, respectively, satisfying... M represents the total number of data sources; through the above calculations, multi-source stream data with quality weight labels is generated.

[0021] Step 1.2: Based on the quality-weighted multi-source data stream, adaptively adjust the regularization window width according to the timestamp distribution of each data source, and perform time window alignment to obtain a time-synchronized multi-source data sequence. Specifically, this includes: based on the aforementioned quality-weighted multi-source data stream, analyzing the distribution pattern of timestamps from each data source, and automatically calculating the optimal regularization window width required for each data source; specifically, firstly, statistically analyzing the time interval sequence between adjacent timestamps of each data source, and calculating the percentile distribution of these intervals. For data sources with generally smaller timestamp intervals, it indicates that their collection frequency is high and the data is dense. In this case, a narrower regularization window is used to retain the detailed changes in the original data; for data sources with generally larger timestamp intervals, it indicates that their collection frequency is low and the data is sparse. In this case, a wider regularization window is used to ensure that no effective information is lost during time alignment; according to the statistical characteristics of each data source, an adaptive regularization window width is determined respectively. Based on this, time window alignment is performed on each data stream according to each adaptive window width, unifying the time base of all data sources to the same time axis, thereby obtaining a time-synchronized multi-source data sequence.

[0022] Step 1.3 involves weighted fusion of time-synchronized multi-source data sequences based on the data quality weights to form an initial fused dataset. Specifically, this includes: weighted fusion of detection values ​​from different data sources at the same time in the time-synchronized multi-source data sequences, based on the calculated data quality weights of each data source; assuming there are K data sources providing valid detection values ​​at time t. The corresponding data quality weight is The value of the initial fused dataset at time t is given by the following formula: in, The weighted fusion value at time t. Let be the detection value of the i-th data source at time t. The data quality weight of this data source is the sum of the weights of all valid data sources in the denominator. Through time-by-time weighted fusion processing, the multi-source data sequences are merged into a unified initial fusion dataset.

[0023] Step 1.4 involves embedding the collected environmental parameters, component construction information, and historical quality trajectories of similar components as embedded information, and temporally associating and binding them with the initial fusion dataset to form a condition-enhanced detection dataset carrying historical quality trajectories. Specifically, this includes: embedding the pre-collected environmental parameters, component construction information, and historical quality trajectories of similar components retrieved from the database as embedded information, and temporally associating and binding them with the initial fusion dataset; specifically, for each time point in the initial fusion dataset, obtaining the ambient temperature and humidity values ​​at that moment, reading the corresponding construction procedure type from the construction management end, such as concrete pouring, vibration, curing, etc., and simultaneously retrieving historical detection value sequences belonging to the same category and within the same service life of the currently detected component from the historical database; further, appending this information to the corresponding time point record in the initial fusion dataset in a time-aligned manner to form a complete record containing the current detection value, environmental parameters, construction information, and historical quality trajectories; after arranging all the time point records in chronological order, the condition-enhanced detection dataset carrying historical quality trajectories is obtained.

[0024] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Extract environmental parameters and component construction information from the enhanced working condition detection dataset, calculate the deviation compensation amount of the current working condition relative to the standard working condition, and obtain the working condition compensation coefficient. Specifically, this includes: extracting environmental parameters and component construction information from the enhanced working condition detection dataset time-by-time. The environmental parameters specifically include the ambient temperature value at that moment, in degrees Celsius, and the ambient humidity value, in percentage. The component construction information specifically includes the type of construction process currently in progress, such as concrete pouring, vibration, curing, formwork removal, etc., and the construction strength level of the process, such as quantitative indicators like pouring speed and vibration frequency. This method internally stores a set of standard working condition conditions, where the standard temperature is set to 20 degrees Celsius, the standard humidity is set to 60%, and the standard construction process is set to the standard curing process. For each moment, first calculate the difference between the current temperature and the standard temperature, and record it as the temperature deviation difference; then calculate the difference between the current humidity and the standard humidity, and record it as the humidity deviation difference.

[0025] Based on this, according to the differences between the current construction procedure type and the standard construction procedure type, the procedure deviation factor is determined according to the pre-set procedure deviation mapping table: if the current procedure type is exactly the same as the standard procedure type, the procedure deviation factor is 0; if the current procedure type is concrete pouring and the standard procedure is curing, the procedure deviation factor is 0.3; if the current procedure type is vibration and the standard procedure is curing, the procedure deviation factor is 0.2; if the current procedure type is formwork removal and the standard procedure is curing, the procedure deviation factor is 0.1; for other unlisted procedure types, the deviation factor is calculated by linear interpolation: that is, based on the differences in process parameters between the procedure and the standard procedure. The similarity is proportionally distributed between 0 and 0.3; the process deviation mapping table and interpolation rules are obtained based on historical experimental data, which can quantify the average influence of different processes on concrete strength test values; finally, the temperature deviation difference, humidity deviation difference, and process deviation factor are multiplied by preset influence coefficients, including temperature influence coefficient, humidity influence coefficient, and process influence coefficient, and the sum of the three is 1. The three products are added together to obtain the comprehensive deviation compensation amount at that moment. This compensation amount is the working condition compensation coefficient; when the working condition compensation coefficient is positive, it indicates that the test value needs to be corrected upward, and when it is negative, it indicates that the test value needs to be corrected downward. The absolute value of the coefficient reflects the magnitude of the required correction.

[0026] Step 2.2: Correct the detected values ​​in the enhanced working condition detection dataset according to the working condition compensation coefficient to eliminate the influence of working condition deviation and obtain the normalized working condition detection sequence. Specifically, this includes: correcting the detected values ​​at each time point in the enhanced working condition detection dataset according to the calculated working condition compensation coefficient. First, determine the type of the detected value at the current time. For concrete strength detection values ​​in megapascals, correct them according to the linear correction method: that is, the corrected strength value = original strength value × (1 + working condition compensation coefficient). For example, if the working condition compensation coefficient is 0.05, it means that the strength is underestimated by 5% due to the low ambient temperature, then the corrected strength value is increased by 5%. For steel reinforcement protective layer thickness detection values ​​in millimeters, a similar method is used. Similar linear correction, but the correction coefficients may be different. For example, the thickness of the protective layer is more affected by humidity, so another set of coefficients can be preset. Iterate through each time point in the enhanced working condition test dataset, read the concrete strength value, the thickness of the steel reinforcement protective layer, the steel reinforcement spacing value, and the flatness deviation value at that point, and correct them with their respective working condition compensation coefficients (different influence coefficient calculation methods can be pre-bound for different test types). After correction, all test values ​​are unified to equivalent values ​​under standard working conditions, eliminating the systematic deviation caused by environmental fluctuations and differences in construction procedures, thus forming a normalized working condition test sequence. Each data point in this sequence represents the test value that should be measured under ideal standard working conditions.

[0027] Step 2.3 involves using the sliding window method to calculate the mean and standard deviation within each window of the normalized test sequence, dynamically fitting the upper and lower boundaries of the adaptive dynamic benchmark interval, and generating the adaptive dynamic benchmark interval. Specifically, this includes: processing the obtained normalized test sequence using the sliding window method to dynamically fit the adaptive benchmark interval; first, determining the length of the sliding window based on the sampling frequency of the test data. For data with a high sampling frequency, such as collecting data points once per second, the window length is set to 60 data points, corresponding to a one-minute time span; for data with a low sampling frequency, such as collecting data points once per hour, the window length is set to 24 data points, corresponding to a one-day time span; for data with sampling frequencies between these two, the window length is determined using a linear interpolation method, specifically: window length = sampling frequency × preset base time span, which can be set to 60 seconds or 1440 minutes.

[0028] After determining the window length, the window is placed at the beginning of the normalized detection sequence, covering the first data point to the Lth data point, where L is the window length. Within the window, the arithmetic mean of all detected values ​​is calculated, which is the sum of the values ​​of each data point within the window divided by the number of data points. Simultaneously, the standard deviation of all detected values ​​within the window is calculated, which is the square root of the sum of the squares of the differences between each data point value and the mean, divided by the number of data points. The arithmetic mean reflects the central tendency of the data within the window, while the standard deviation reflects the dispersion of the data. The upper boundary of the adaptive dynamic baseline interval is set to the arithmetic mean plus twice the standard deviation, and the lower boundary is set to the arithmetic mean minus twice the standard deviation. After completing the calculation for the first window, the window is slid backward by one data point, with a step size of 1. At this point, the window covers the second data point to the Lth plus one data point. At the new window position, the arithmetic mean, standard deviation, and upper and lower boundaries are repeatedly calculated. Each slide yields a baseline interval boundary value corresponding to a time point, which is typically the center or end position of the window.

[0029] The window is continuously slid in a step of 1 until its right endpoint reaches the last data point in the sequence. Once the window reaches the end of the sequence, the upper boundary values ​​of all time points are arranged chronologically to form the upper boundary sequence; similarly, the lower boundary values ​​of all time points are arranged chronologically to form the lower boundary sequence. The upper and lower boundary sequences together constitute an adaptive dynamic baseline interval covering the entire detection sequence. This interval automatically adjusts its width according to local fluctuations in the detection data: in areas of stable data, the standard deviation is small, resulting in a narrower interval; in areas of drastic data fluctuation, the standard deviation is large, and the interval automatically widens. This dynamic adjustment mechanism ensures that the baseline interval accurately reflects the normal fluctuation range of the detection data over different time periods, avoiding misjudgments caused by using fixed thresholds.

[0030] Step 2.4 calculates the deviation of each detection value in the normalized detection sequence relative to the adaptive dynamic reference interval, forming a deviation mapping feature sequence. Specifically, this includes: for each detection value at any given time, first determining whether the value falls between the upper and lower boundaries corresponding to that time. If the detection value is within the interval, the deviation is recorded as a small positive value close to zero, such as 0.001, to indicate that the point is within the normal fluctuation range but retains a weak signal for subsequent processing. If the detection value is higher than the upper boundary, the deviation is recorded as the difference between the detection value and the upper boundary (positive value), and the magnitude of this difference reflects the severity of exceeding the normal upper limit. If the detection value is lower than the lower boundary, the deviation is recorded as the difference between the detection value and the lower boundary (negative value), and its absolute value reflects the severity of falling below the normal lower limit. After point-by-point calculation, all deviations are arranged in the original time order to form a one-dimensional sequence, i.e., the deviation mapping feature sequence. Each element in this sequence quantitatively expresses the deviation direction and magnitude of the detection value at that time from the dynamic normal fluctuation range, providing a basic input for subsequent feature attractor identification and geometric enhancement.

[0031] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Calculate the directional derivative magnitude of each point in the time-parameter joint space of the deviation mapping feature sequence. Locate the local maxima of the directional derivative magnitude as candidate feature attractors. Record the precise coordinates of each candidate feature attractor. The precise coordinates are composed of a specific timestamp and the corresponding concrete strength deviation value, rebar cover thickness deviation value, rebar spacing deviation value, and flatness deviation value, thus obtaining an initial set of feature attractors. Specifically, for each point in the deviation mapping feature sequence, represent it as a five-dimensional space vector, specifically including the timestamp dimension, concrete strength deviation dimension, rebar cover thickness deviation dimension, rebar spacing deviation dimension, and flatness deviation dimension. Let the deviation mapping feature sequence have a total of The point, of which the first Each point is recorded as , , For timestamps, This represents the concrete strength deviation value. This represents the deviation value of the concrete cover thickness for reinforcing bars. This represents the deviation value of the rebar spacing. This represents the flatness deviation value.

[0032] Calculate the magnitude of the directional derivative at each point; this magnitude measures the degree of change of that point along each coordinate axis in five-dimensional space. For interior points (…), First, calculate the rate of change of each deviation dimension relative to time: rate of change of concrete strength deviation. ,in , ; Rate of variation in the thickness deviation of the reinforcing steel protective layer ,in ; Rate of change of rebar spacing deviation ,in ; Flatness deviation change rate ,in For the endpoints of the sequence ( or Using one-sided difference, for example hour, , The same applies to the other dimensions; therefore, the first... directional derivative magnitude at each point Defined as the square root of the sum of the squares of the four rates of change mentioned above, the calculation formula is as follows: ; in, , , , The first Approximate values ​​of the directional derivatives at each point in the dimensions of concrete strength deviation, rebar cover thickness deviation, rebar spacing deviation, and flatness deviation; the modulus of the directional derivatives at all points is calculated. Then, locate the local maximum point of the magnitude of the directional derivative, that is, satisfy... and (For endpoints, only one-sided comparisons are needed); mark these local maxima points as candidate feature attractors, and record the precise coordinates of each candidate feature attractor, with the coordinates represented by a timestamp. and the corresponding concrete strength deviation value Depth of concrete cover for reinforcing bars Rebar spacing deviation value Flatness deviation value Together they constitute; all candidate feature attractors form the initial feature attractor set. .

[0033] Step 3.2 involves performing non-maximum suppression on the initial feature attractor set, removing redundant stationary points whose linear distance in the feature space is less than a preset distance threshold, and retaining feature attractors whose gradient direction angle between adjacent feature attractors is greater than a preset angle threshold, thus forming an optimized feature attractor set. Specifically, this includes: adjusting the initial feature attractor set... Perform nonmaximum suppression to eliminate redundant stationary points; first for any two candidate feature attractors and ( ), calculate the straight-line distance between them in the feature space, which is the five-dimensional Euclidean distance, and the calculation formula is: ; Preset a distance threshold ,like Then it is believed and Since the two are redundant stationary points, the one with the larger directional guide modulus is retained, and the other is discarded. After the first round of distance filtering, the set of intermediate feature attractors is obtained. ;right The feature attractors in the array are arranged in ascending order of timestamp. For two adjacent feature attractors... and Calculate the angle between them along the gradient direction; the gradient direction is characterized by the gradient vector at that point, the th... The gradient vector at each point The components are the four directional derivatives calculated in step 3.1, namely: Adjacent two feature attractors and gradient direction angle Calculated using the vector dot product formula: in, The dot product of two gradient vectors. and These are the magnitudes of the two gradient vectors; a preset angle threshold is also provided. ,like Then retain both and ;like Then, the one with the smaller directional derivative magnitude is removed; after filtering by gradient direction angle, the remaining candidate feature attractors constitute the optimized feature attractor set. .

[0034] Step 3.3: Using the precise coordinates of each feature attractor in the optimized feature attractor set as control vertices, a geometric method of constructing a minimum closed polytope around the control vertices is adopted. The outer vertices are connected sequentially to form an outer envelope surface, resulting in a minimum outer envelope polytope containing all control vertices. Specifically, this involves: first, treating all points in the optimized feature attractor set as a discrete set of points in five-dimensional space; each point contains five coordinate components: timestamp, concrete strength deviation value, rebar cover thickness deviation value, rebar spacing deviation value, and flatness deviation value. These points represent the key locations with the most drastic changes in the deviation mapping feature sequence; based on this, a convex hull construction algorithm is used to determine the location containing all these points. The minimum convex multicell is obtained by: selecting a starting point on the boundary from the point set, usually choosing a point with an extreme value in a certain coordinate dimension (e.g., the smallest timestamp or the smallest concrete strength deviation) as the initial vertex; further, sequentially finding adjacent vertices that form a polar edge with the current vertex (i.e., all other points are located on the same side of this edge), gradually expanding the boundary; each time a new vertex is found, it is added to the boundary vertex set, and the search for the next polar edge vertex continues from this vertex as the starting point. When all polar edge vertices are found and finally return to the starting point, the boundary vertex sequence of the convex hull is obtained; the lines connecting these boundary vertices in five-dimensional space constitute the edges of the convex hull, and the hyperplane patches enclosed by these edges constitute the faces of the convex hull.

[0035] After obtaining the convex hull, it needs to be further processed into a minimal convex hull polytope. "Convex" means that the outer surface of the polytope is formed by sequentially connecting the outer vertices of the convex hull, creating a closed geometry without any depressions. Specifically, during construction, all boundary vertices of the convex hull are connected sequentially according to their adjacency on the convex hull surface, ensuring that each face is convex. For a five-dimensional convex hull, each face is part of a four-dimensional hyperplane. Connecting all outer vertices in a certain order (e.g., according to polar angle or topological order of the convex hull) forms a complete convex hull surface. The enclosed region is the minimum convex envelope polytope. To ensure that the polytope is the minimum, it is necessary to verify that there is no convex polytope that can contain all control vertices and has a smaller volume. Since the convex envelope itself is the minimum convex set containing the set of points, the convex envelope polytope constructed based on the convex envelope naturally has the minimum volume. In addition, the degeneracy case needs to be handled during the construction process: when all control vertices are coplanar or collinear, the dimension of the convex envelope will decrease. At this time, the minimum convex envelope polytope degenerates into a low-dimensional geometry, such as a hyperplane, line segment, or a single point, but it still needs to be constructed according to the same principle, that is, to contain all vertices and have the minimum convexity property.

[0036] The final output is this minimal outer envelope polytope, which will be used in subsequent steps for affine transformation and fitting to a standard hyperellipsoidal space to generate a geometrically constrained hypersurface. The entire construction process ensures that the polytope tightly encloses all feature attractors, thus providing precise boundary constraints for subsequent geometric feature enhancement.

[0037] In a preferred embodiment of the present invention, step 3 above may include: Step 3.4: Apply an affine transformation to the minimum outer envelope polycell, stretching, rotating, and translating it to the standard hyperellipsoid space, so that the minimum outer envelope polycell is inscribed in the standard hyperellipsoid, generating a geometrically constrained hypersurface. Specifically, the implementation process is divided into the following four sub-stages: Sub-stage 1: Calculate the geometric center of the minimum outer envelope polycell in five-dimensional space. Traverse the coordinates of all vertices of the polycell and calculate the arithmetic mean of the timestamp dimension, concrete strength deviation dimension, rebar cover thickness deviation dimension, rebar spacing deviation dimension, and flatness deviation dimension. The point formed by the five average values ​​is the geometric center. Use this geometric center as the reference point for translation. The purpose of translation is to make the center of the polycell coincide with the center of the subsequent standard hyperellipsoid. The center of the standard hyperellipsoid is set as the origin of the five-dimensional space, that is, the point where all coordinates in each dimension are zero.

[0038] In the second sub-stage, the semi-axis lengths of the minimum externally oriented polytope in each coordinate axis direction are analyzed. For each coordinate axis direction, the maximum and minimum values ​​of all vertices of the polytope in that direction are calculated, and the larger of the two absolute values ​​is taken as the semi-axis length in that direction. For example, in the concrete strength deviation dimension, the maximum and minimum values ​​of all vertices in that dimension are found. If the absolute value of the maximum value is greater than the absolute value of the minimum value, the semi-axis length is taken as the absolute value of the maximum value; otherwise, the absolute value of the minimum value is taken. The semi-axis lengths of each of the five dimensions are obtained in turn. The stretching coefficient required to transform the polytope into a unit hyperellipsoid is further calculated. For each dimension, the stretching coefficient is equal to the reciprocal of the semi-axis length of that dimension. The coordinates of each vertex of the polytope in each dimension are multiplied by the stretching coefficient of the corresponding dimension to achieve non-uniform stretching, so that the maximum span of the polytope in each dimension after stretching is from -1 to +1. After stretching, the polytope is exactly inscribed in a unit hyperellipsoid, which is defined as the sum of the squares of the coordinates in each dimension being equal to 1.

[0039] In the third stage, the deviation between the principal axis direction of the stretched polytope and the coordinate axis direction of the unit hyperellipsoid is determined, and a rotation matrix is ​​constructed accordingly. Specifically, the covariance matrix of the coordinates of each vertex of the stretched polytope is calculated, and eigenvectors are obtained through eigenvalue decomposition. These eigenvectors are the principal axis directions. The angles between the principal axis directions and the standard coordinate axis directions (i.e., the unit vectors of each dimension) are compared to obtain the rotation angles. An orthogonal rotation matrix is ​​constructed based on these angles, such that after multiplying the stretched polytope by the rotation matrix, its principal axis is completely aligned with the standard coordinate axis. During the rotation, the relative geometric relationship between the vertices of the polytope remains unchanged, and only its orientation is changed.

[0040] In the fourth sub-stage, translation, stretching, and rotation operations are sequentially applied to the original minimal outer envelope polycell. First, the polycell is translated so that its geometric center is moved to the origin. Then, the coordinates of each vertex are stretched by a stretching coefficient. Finally, the rotation is completed by multiplying the stretched coordinates by a rotation matrix. The combination of these three operations is the affine transformation. After the affine transformation, the original minimal outer envelope polycell is mapped to a convex polycell inscribed in a unit hyperellipsoid, and all vertices of this convex polycell are located on or inside the surface of the unit hyperellipsoid. At this point, the surface of the unit hyperellipsoid is the generated geometrically constrained hypersurface. The hypersurface is a smooth closed surface in 5-dimensional space, and its expression is: the sum of the squares of all coordinate components equals 1. All vertices of the polycell after the affine transformation satisfy this expression or are located inside the region enclosed by the expression. The geometrically constrained hypersurface serves as a reference for subsequent steps and is used to calculate the radial deviation and normal deflection angle of each point in the deviation mapping feature sequence relative to this surface.

[0041] Step 3.5: Project the coordinates of each point in the deviation mapping feature sequence onto the geometrically constrained hypersurface, calculate the weighted orthogonal statistic of the correlation between each projected point and the origin, and obtain the weighted orthogonal distance value of each point. Specifically, this includes: Let the first point in the deviation mapping feature sequence be... The points are ,in The value range is from 1 to the total number of points in the deviation mapping feature sequence. First, apply the same stretching, rotation, and translation to the point using the same affine transformation parameters as in step 3.4 to obtain the transformed point. Based on this, starting from the origin (i.e., the zero vector in five-dimensional space), connect the origin and... A ray is formed that intersects the geometrically constrained hypersurface (i.e., the standard hyperellipsoid) at a single point, denoted as the projection point. .

[0042] Projection point The coordinates satisfy the standard hyperellipsoid equation, which is expressed as follows: ; and and The directions are the same, meaning there exists a scalar. Make Further calculations Euclidean distance from the origin and projection point The distance from the origin is always 1; define the correlation weighted orthogonal statistic. for: ,in The standard deviation of the distances from all points in the bias-mapped feature sequence to the origin after an affine transformation; this statistic reflects The radial deviation relative to the geometrically constrained hypersurface is normalized using the global standard deviation to obtain a dimensionless weighted orthogonal distance value for each... Calculate the corresponding .

[0043] Step 3.6: Calculate the angle between the normal direction of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the tangent plane of the hypersurface at that point, to obtain the hypersurface normal deflection angle at each point. Specifically, this includes: for the transformed points obtained in Step 3.5... Its corresponding projection point Located on the standard hyperellipsoid, in At that point, the normal direction of the geometrically constrained hypersurface is from the origin to... The radial direction is denoted as the unit normal vector. (because ); at the same time, in A tangent plane can be constructed at a point on the hypersurface, and this tangent plane is perpendicular to the normal vector. A vertical five-dimensional hyperplane; connection and , to obtain vector ;vector With normal vector The angle between This is the normal deflection angle of the hypersurface; the calculation formula is: in, For dot product, for The length of the mold, ;because and directional relationship, The value range is 0 to ;when When it is exactly located on the hypersurface, Define at this time This included angle reflects whether the detection point deviates from the hypersurface along the normal or tangential direction: if... A value close to 0 indicates that the deviation is primarily radial; if... near This indicates that the deviation is primarily tangential; by calculating the corresponding value for each point... The hypersurface normal deflection angle at each point is obtained.

[0044] Step 3.7: Combine the weighted orthogonal distance value with the hypersurface normal deflection angle to form a geometrically enhanced feature sequence. Specifically, this includes: combining the obtained weighted orthogonal distance value... With the obtained hypersurface normal deflection angle Vector combinations are performed to form a geometrically enhanced feature sequence; specifically, for each point in the deviation mapping feature sequence... Construct a two-dimensional feature vector The first component of this two-dimensional eigenvector is the normalized radial deviation distance, and the second component is the normal deflection angle; [This is followed by a list of components and their relationships.] The corresponding two-dimensional feature vectors are arranged in the order of the original timestamps, forming a two-dimensional geometrically enhanced feature sequence. Each vector in this two-dimensional geometric enhancement feature sequence simultaneously contains the radial degree and directional characteristics of the detection point's deviation from the geometrically constrained hypersurface. Compared with the original deviation mapping feature sequence, it has a stronger geometric discrimination ability, providing highly discriminative input features for subsequent hybrid anomaly recognition models.

[0045] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Input the geometrically enhanced feature sequence into the pre-trained hybrid anomaly recognition model. The pre-trained hybrid anomaly recognition model performs static threshold screening, including comparing each feature value in the geometrically enhanced feature sequence with a preset static threshold, marking feature points exceeding the threshold range as candidate anomalies, and obtaining a preliminary anomaly set. Specifically, the pre-trained hybrid anomaly recognition model is based on a dual-branch heterogeneous integrated architecture design of a temporal convolutional network and a lightweight gradient booster. The temporal convolutional network branch is used to capture the local dependencies and long-term evolution patterns of the geometrically enhanced feature sequence in the time dimension, and the lightweight gradient booster branch is used to perform nonlinear classification on the two-dimensional vector (weighted orthogonal distance value and hypersurface normal deflection angle) of each feature point. The outputs of the two branches are weighted and combined through a learnable fusion layer to finally output the anomaly discrimination probability and anomaly type label. Compared with a single model, this hybrid architecture can simultaneously utilize the feature representation capabilities of deep learning and the robustness of tree models, and is especially suitable for real-world scenarios with strong noise interference and imbalanced positive and negative samples in building structure inspection data.

[0046] The training process of the model is as follows: First, a large number of geometrically enhanced feature sequence samples are collected from historical projects that have been completed and tested. Each sample contains two-dimensional feature vectors from multiple consecutive time points. Senior testing engineers annotate the detection results at each time point to determine whether they are abnormal and the type of abnormality, such as low concrete strength, insufficient rebar cover thickness, excessive rebar spacing, and excessive flatness. The annotated samples constitute the training dataset. Based on this, a sliding window method is used to divide the training dataset into multiple fixed-length subsequences, each subsequence serving as a training sample. For the temporal convolutional network branch, subsequences are used... Supervised learning is performed within the complete time window of the sequence, and weighted cross-entropy is used as the loss function to alleviate the class imbalance problem. For the lightweight gradient booster branch, the two-dimensional feature vector at each time point is trained as an independent sample, and log loss is used as the loss function. The parameters of the fusion layer of the two branches are jointly optimized through end-to-end backpropagation. During training, five-fold cross-validation is used to adjust hyperparameters, including the kernel size, number of layers, number of channels of the temporal convolutional network, as well as the maximum depth and learning rate of the gradient booster tree. The trained model achieves an anomaly detection accuracy of over 95% on the validation set, with a false positive rate controlled within 3%.

[0047] After the model training is completed, the geometrically enhanced feature sequence obtained in step 3.7 is input into this pre-trained hybrid anomaly recognition model. First, a static threshold screening is performed: the model internally sets a static threshold range for each feature component, where the normal range for the weighted orthogonal distance value is set between -2 and +2, and the normal range for the hypersurface normal deflection angle is set between 0 and 1.57 radians. For each feature point in the sequence, it is determined whether its weighted orthogonal distance value exceeds the preset threshold range, and whether its hypersurface normal deflection angle exceeds the preset threshold range. If either of the two components exceeds the corresponding normal range, the feature point is marked as a candidate anomaly. All candidate anomalies are collected to form a preliminary anomaly set. This preliminary screening process can quickly filter out points that significantly deviate from the normal fluctuation range, reducing the computational load of subsequent analysis. The static threshold screening, as a pre-filter of the model, together with the subsequent temporal evolution trend analysis and multi-dimensional weighted correlation inference, constitutes a three-layer progressive discrimination logic, effectively reducing computational overhead and improving the real-time performance of anomaly recognition.

[0048] Step 4.2 involves performing time-series evolution trend analysis on the candidate anomalies in the initial screening set. This includes calculating the rate of change of eigenvalues ​​in the forward and backward neighborhoods of each candidate anomaly on the time axis, and selecting candidate anomalies with consistent signs of change and cumulative changes exceeding a preset trend threshold to obtain a trend anomaly set. Specifically, this includes performing time-series evolution trend analysis on each candidate anomaly in the initial screening set. Assume there are a total of [number missing] anomalies in the initial screening set. The nth candidate outlier, for the nth one Candidate outliers ( First, determine its temporal position index in the original geometric enhancement feature sequence. Then, taking that point as the center, take forward... Take the nearest neighbor points and move backward. Each of the three neighboring points forms a forward neighborhood and a backward neighborhood. For example, the preset neighborhood radius. The forward neighborhood contains the index from arrive Feature points, backward neighborhood contains index from arrive Feature points; calculate the rate of change of weighted orthogonal distance values ​​in the forward neighborhood for each feature point. The calculation formula is: in, This represents the weighted orthogonal distance between the endpoints of the forward neighborhood. This is the weighted orthogonal distance value between the first endpoint of the forward neighborhood. and These are the corresponding timestamps; similarly, the rate of change of the weighted orthogonal distance values ​​within the backward neighborhood is calculated. The formula is as follows: in, This represents the weighted orthogonal distance between the endpoints of the backward neighborhood. This is the weighted orthogonal distance value between the first endpoint of the backward neighborhood; further, it is determined whether the signs of the forward and backward rates of change are consistent, i.e. If the signs are consistent, then calculate the cumulative change in the forward neighborhood. and cumulative change in the backward neighborhood The larger of the two values ​​is taken as the comprehensive cumulative change of the candidate outlier. ,Will With preset trend threshold If a comparison is made, If a candidate outlier is found to have a continuous evolutionary trend, it is retained; otherwise, it is removed from the initial screening set of outliers. After time-series evolution trend analysis, the retained candidate outliers constitute the trend outlier set.

[0049] Step 4.3 involves performing multi-dimensional weighted correlation reasoning on the feature points in the trend anomaly set. This includes calculating a weighted anomaly score based on preset weights for each parameter dimension and fusing the correlation constraints between different dimensions to obtain anomaly discrimination probability sequences and anomaly type labels corresponding to each feature point. Specifically, this includes performing multi-dimensional weighted correlation reasoning on each feature point in the trend anomaly set. Assume there are a total of [number missing] feature points in the trend anomaly set. For the nth feature point, for the nth feature point Feature points ( Its geometric enhancement feature vector is ,in The weighted orthogonal distance value. The normal deflection angle of the hypersurface; the model pre-sets weight coefficients for these two dimensions, let... The weights for the weighted orthogonal distance values, for example , For example, the weight of the hypersurface normal deflection angle. ,and Then the weighted anomaly score of this feature point. The calculation formula is: in, Take the absolute value to reflect the degree of deviation. Normalize the normal deflection angle to interval, because The value range is 0 to Simultaneously, the correlation constraints between dimensions were also considered: correlation factors were defined. Its value is based on and The combination of possibilities is determined; specifically, if and ,but This corresponds to a high probability of abnormal concrete strength; if and ,but This corresponds to an abnormal protective layer thickness; if but ,but This corresponds to the flatness deviation; other cases This corresponds to normal or slight fluctuations; the final anomaly detection probability value. The result is obtained by multiplying the weighted outlier score by the association factor and then mapping it using the Sigmoid function: in, The value ranges from 0 to 1, with a larger value indicating a higher probability of an anomaly; simultaneously, according to The value category is selected, and the corresponding exception type label is output. For example, low concrete strength, insufficient concrete cover thickness, excessive rebar spacing, or excessive flatness; the discrimination results of all feature points are arranged in chronological order to form an anomaly discrimination probability sequence. and exception type label sequence This is for use in subsequent confidence level verification.

[0050] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1 involves performing a confidence check on each anomaly probability value in the anomaly discrimination probability sequence, removing anomaly probabilities and their corresponding anomaly type labels that are below a preset confidence threshold to obtain a high-confidence anomaly set. Specifically, this includes: performing a confidence check on each anomaly probability value in the anomaly discrimination probability sequence; sequentially reading the anomaly discrimination probability value corresponding to each feature point from the anomaly discrimination probability sequence obtained in Step 4.3, pre-setting a confidence threshold, for example, 0.6; comparing each anomaly discrimination probability value with the confidence threshold: if the probability value is greater than or equal to the threshold, the discrimination result is considered to have high confidence and is retained; if the probability value is lower than the threshold, it is determined to be an unreliable result, and the probability value and its corresponding anomaly type label are removed; after the confidence check, all retained anomaly points and their discrimination results constitute a high-confidence anomaly set. This step effectively filters out low-confidence misjudgment results, improving the reliability of the final alarm.

[0051] Step 5.2 involves performing logical mutual exclusion checks on the anomaly type labels in the high-confidence anomaly set. This involves detecting whether contradictory anomaly type labels exist within the same detection location and time window. If such labels exist, the highest priority label is retained, and contradictory labels are removed, resulting in an anomaly set that passes the mutual exclusion check. Specifically, this includes performing logical mutual exclusion checks on the anomaly type labels in the high-confidence anomaly set. Specifically, the anomalies in the high-confidence anomaly set are grouped according to their detection location (e.g., the same component, the same floor, or the same test area) and time window (e.g., within ten consecutive minutes). For multiple anomalies within the same group, it is checked whether there are logical contradictions between their corresponding anomaly type labels. For example, labels indicating low concrete strength and high concrete strength cannot appear simultaneously in the same location and within the same time window; similarly, labels indicating insufficient concrete cover thickness and excessive concrete cover thickness are mutually exclusive. The model internally stores a mutually exclusive label mapping table, listing all label pairs that cannot appear simultaneously. For each set of anomalies, if a mutually exclusive label pair is detected, the label with higher priority is retained and the label with lower priority is removed according to a preset priority rule. The priority setting is based on engineering experience: for example, abnormal concrete strength has a higher priority than abnormal concrete cover thickness, and abnormal concrete cover thickness has a higher priority than abnormal flatness. After mutual exclusion judgment, all retained anomalies constitute a set of anomalies that have passed the mutual exclusion verification.

[0052] Step 5.3: Extract the associated working condition parameters and data tracing paths corresponding to each anomaly from the set of anomalies that have passed mutual exclusion verification. The associated working condition parameters include the environmental parameters and component construction information corresponding to the anomaly. The data tracing paths include the original data stream source identifier, timestamp, and parameter dimension index. Specifically, this includes: extracting the associated working condition parameters and data tracing paths corresponding to each anomaly from the set of anomalies that have passed mutual exclusion verification; for each anomaly in the set, tracing back its record in the original data processing flow, the associated working condition parameters include the environmental temperature, environmental humidity, and construction conditions at the time of the anomaly. The sequence type and construction strength level are already bound to the detection values ​​in the enhanced detection dataset generated in step 1.4. The data traceability path includes three parts: the first part is the original data stream source identifier, i.e., the sensor number or device number that generated the detection data; the second part is the timestamp, accurate to the millisecond, indicating the specific time of data collection; the third part is the parameter dimension index, indicating which dimension of concrete strength, rebar cover thickness, rebar spacing, or flatness the anomaly belongs to. This information is extracted one by one according to the order of the anomalies to form a traceability record corresponding to each anomaly.

[0053] Step 5.4 involves structurally encapsulating the verified judgment results, associated operating parameters, and data tracing paths according to a preset alarm format to form an anomaly identification alarm command. This command is then sent to the management terminal. Specifically, the preset alarm format is a standardized data packet structure containing an alarm header and an alarm body. The alarm header includes the alarm number, generation timestamp, and alarm level, categorized into high, medium, and low levels based on the anomaly discrimination probability value. The alarm body includes a precise location description of the anomaly (composed of a timestamp and parameter dimension index), anomaly type label, anomaly discrimination probability value, corresponding environmental parameters and construction information, and a complete data tracing path. All anomaly information is packaged into one or more alarm commands in chronological order. This anomaly identification alarm command is then sent to the management terminal via wired or wireless network for on-site management personnel or a remote monitoring center to view and process. Upon receiving the alarm command, the management terminal can parse and display the anomaly details, helping maintenance personnel quickly locate the root cause of the problem and take appropriate measures.

[0054] like Figure 2 As shown, embodiments of the present invention also provide a building structure inspection data anomaly identification system, including: The data acquisition module is used to collect raw data streams and related data such as environmental parameters, component construction information, and historical quality trajectories of similar components for building structure inspection. The data governance module is used to perform dynamic quality weighted fusion and adaptive time window normalization on the raw data stream of building structure inspection using an adaptive data governance mechanism, and to embed environmental parameters, component construction information and historical quality trajectories of similar components to obtain a condition-enhanced inspection dataset carrying historical quality trajectories. The baseline interval generation module is used to perform working condition deviation compensation calculation and dynamic threshold fitting on the working condition enhancement detection dataset to generate an adaptive dynamic baseline interval and obtain a deviation mapping feature sequence. The geometric feature enhancement module is used to dynamically identify multiple feature attractors from the deviation mapping feature sequence, and construct a minimum outer envelope polycell using the multiple feature attractors as control vertices; perform an affine transformation on the minimum outer envelope polycell and fit it to the standard hyperellipsoid space to generate a geometrically constrained hypersurface; calculate the correlation weighted orthogonality statistic of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface to obtain the geometrically enhanced feature sequence; The anomaly detection module is used to input the geometrically enhanced feature sequence into the pre-trained hybrid anomaly detection model, and sequentially perform static threshold screening, temporal evolution trend analysis and multi-dimensional weight correlation inference to obtain anomaly discrimination probability sequence and anomaly type label. The alarm delivery module is used to perform confidence verification and logical mutual exclusion judgment on the anomaly discrimination probability sequence and anomaly type label. After obtaining the judgment result that passes the verification, the associated operating condition parameters and the data traceability path, it is structured and encapsulated to obtain the anomaly identification alarm instruction and send it to the management terminal.

[0055] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0056] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0057] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0058] Methods and Systems for Identifying Anomalies in Building Structure Inspection Data (Example): I. Overview of Experimental Conditions This experimental example uses the structural inspection of an underground parking garage in a large commercial complex as the simulation object. Four core types of inspection data were collected: concrete strength, rebar cover thickness, rebar spacing, and structural flatness. Ambient temperature and humidity were also collected as auxiliary parameters. The data sources included six types of sensors: concrete strength sensor (signal-to-noise ratio 35dB), rebar cover thickness gauge (signal-to-noise ratio 42dB), rebar spacing detector (signal-to-noise ratio 38dB), structural flatness detector (signal-to-noise ratio 30dB), ambient temperature and humidity sensor (signal-to-noise ratio 28dB), and displacement sensor (signal-to-noise ratio 33dB). The data collection period was from January 2025 to June 2025, with a sampling interval of 30 minutes. The adaptive time window width of the dynamic quality-weighted fusion was automatically adjusted according to the timestamp sparsity of each data source, ranging from 30 minutes to 2 hours. The sliding window width of the adaptive dynamic baseline interval was 60 sampling points, and the upper and lower bounds of the baseline interval were the mean plus or minus 2 standard deviations. The hybrid anomaly detection model comprises three levels: static threshold initial screening, time-series evolution trend analysis, and multi-dimensional weighted correlation inference. The confidence threshold is 0.85, and the preset static threshold is 1.5 times the upper bound of the baseline interval.

[0059] II. Experimental Procedures and Results Step 1 (Adaptive Data Governance and Condition Enhancement): Analyze the signal-to-noise ratio (SNR) and timestamp sparsity of the six data sources, and dynamically calculate the data quality weight of each data source. The rebar cover thickness gauge has the highest SNR (42dB), obtaining the highest quality weight of 19.7%; the environmental temperature and humidity sensor has the lowest SNR (28dB), obtaining the lowest quality weight of 13.1%. Adaptively adjust the time window width according to the timestamp distribution of each data source, perform weighted fusion on the time-synchronized multi-source data sequences, embed environmental parameters, component construction information, and historical quality trajectories of similar components, forming a condition enhancement detection dataset carrying historical quality trajectories, totaling 2000 sampling points.

[0060] Step 2 (Work Condition Deviation Compensation and Dynamic Benchmark Interval Generation): Environmental parameters and component construction information are extracted from the enhanced work condition detection dataset to calculate the deviation compensation amount of the current work condition relative to the standard work condition. The work condition compensation coefficient ranges from 0.85 to 1.15. After correcting the detected values ​​based on the compensation coefficient, the mean and standard deviation are calculated using a sliding window method with 60 sampling points, and the upper and lower bounds of the adaptive dynamic benchmark interval are dynamically fitted. The deviation is calculated point by point for 2000 sampling points to form a deviation mapping feature sequence. At sampling points 200-210, 350-365, and 420-430 of the time series, the deviation values ​​significantly exceed the dynamic benchmark interval, and are marked as three abnormal areas.

[0061] Step 3 (Geometric Feature Enhancement and Anomaly Recognition): Identify 8 feature attractors from the deviation mapping feature sequence, and construct a minimum outer envelope polytope using these as control vertices (e.g., Figure 3 (As shown). An affine transformation is applied to the multicellular object to fit it to the standard hyperellipsoid space, generating a geometrically constrained hypersurface. The weighted orthogonal distance and normal deflection angle of each point relative to the hypersurface are calculated to form a geometrically enhanced feature sequence. The geometrically enhanced feature sequence is input into the hybrid anomaly recognition model, and static threshold screening (screening out 186 candidate anomalies), temporal evolution trend analysis (screening out 72 trend anomalies), and multidimensional weighted correlation inference (screening out 45 anomalies) are performed sequentially. After confidence verification (removing 7 points below the 0.85 threshold) and logical mutual exclusion judgment (removing 3 contradictory labels), 35 high-confidence anomalies are finally confirmed. The anomaly types are distributed as follows: 10 rebar cover anomalies, 8 concrete strength anomalies, 6 comprehensive multidimensional anomalies, 6 rebar spacing anomalies, and 5 flatness anomalies (such as...). Figure 4 Show).

[0062] Step 4 (Effect Comparison and Verification): Compare the method of this invention with the fixed threshold method, isolated forest algorithm, LSTM autoencoder method, and principal component analysis method (e.g., Figure 5 (As shown). The accuracy of the method of this invention is 96.8%, the recall is 94.3%, and the F1 score is 95.5% (as shown). Figure 5 As shown in the figure, both methods significantly outperform the comparison methods. The fixed threshold method has an accuracy of only 78.5% and a recall of 72.1%. The LSTM autoencoder method has an accuracy of 89.6% and a recall of 87.2%, which is the best among the comparison methods.

[0063] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying anomalies in building structure inspection data, characterized in that, The method includes: Step 1: Adaptive data governance mechanism is used to perform dynamic quality weighted fusion and adaptive time window normalization on the raw data stream of building structure inspection, and pre-acquired environmental parameters, component construction information and historical quality trajectories of similar components are embedded to obtain a working condition enhanced inspection dataset carrying historical quality trajectories. Step 2: Perform working condition deviation compensation calculation and dynamic threshold fitting on the working condition enhancement detection dataset to generate an adaptive dynamic benchmark interval and obtain the deviation mapping feature sequence. Step 3: Dynamically identify multiple feature attractors from the deviation mapping feature sequence, and construct a minimum outer envelope polycell using the multiple feature attractors as control vertices; perform an affine transformation on the minimum outer envelope polycell and fit it to the standard hyperellipsoid space to generate a geometrically constrained hypersurface; calculate the correlation weighted orthogonal statistics of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface to obtain the geometrically enhanced feature sequence; Step 4: Input the geometrically enhanced feature sequence into the pre-trained hybrid anomaly recognition model, and sequentially perform static threshold screening, temporal evolution trend analysis and multi-dimensional weight correlation inference to obtain anomaly discrimination probability sequence and anomaly type label; Step 5: Perform confidence verification and logical mutual exclusion judgment on the anomaly discrimination probability sequence and anomaly type label. After obtaining the judgment result that passes the verification, the associated working condition parameters and the data traceability path, perform structured encapsulation to obtain the anomaly identification alarm instruction and send it to the management terminal.

2. The method for identifying anomalies in building structure inspection data according to claim 1, characterized in that, An adaptive data governance mechanism is employed to dynamically weight and fuse the raw data stream of building structure inspection, and to adapt it to time window normalization. Pre-acquired environmental parameters, component construction information, and historical quality trajectories of similar components are embedded to obtain a condition-enhanced inspection dataset carrying historical quality trajectories, including: The signal-to-noise ratio and timestamp sparsity of each data source in the original data stream are analyzed, and the data quality weight of each data source is dynamically calculated to obtain a multi-source data stream with quality weights. Based on the multi-source data stream with quality weights, the width of the regularized window is adaptively adjusted according to the timestamp distribution of each data source, and time window alignment is performed to obtain a time-synchronized multi-source data sequence. Based on the aforementioned data quality weights, the time-synchronized multi-source data sequences are weighted and fused to form an initial fused dataset; The collected environmental parameters, component construction information, and historical quality trajectories of similar components are used as embedded information and are temporally associated and bound with the initial fusion dataset to form a condition-enhanced detection dataset carrying historical quality trajectories.

3. The method for identifying anomalies in building structure inspection data according to claim 2, characterized in that, The operating condition enhancement detection dataset is subjected to operating condition deviation compensation calculation and dynamic threshold fitting to generate an adaptive dynamic benchmark interval, resulting in a deviation mapping feature sequence, including: Environmental parameters and component construction information are extracted from the enhanced working condition detection dataset. The deviation compensation amount of the current working condition relative to the standard working condition is calculated to obtain the working condition compensation coefficient. The detection values ​​in the enhanced working condition detection dataset are corrected based on the working condition compensation coefficient to eliminate the influence of working condition deviation and obtain the working condition normalized detection sequence. The mean and standard deviation within each window are calculated using the sliding window method for the normalized test sequence under working conditions. The upper and lower boundaries of the adaptive dynamic benchmark interval are dynamically fitted to generate the adaptive dynamic benchmark interval. The deviation of each detection value in the normalized detection sequence under operating conditions relative to the adaptive dynamic reference interval is calculated to form a deviation mapping feature sequence.

4. The method for identifying anomalies in building structure inspection data according to claim 3, characterized in that, Multiple feature attractors are dynamically identified from the bias mapping feature sequence. These feature attractors are gradient abrupt change stationary points in the temporal-parameter joint space of the bias mapping feature sequence, with their coordinates precisely located at the intersection of a specific timestamp and the corresponding parameter dimension. Using these multiple feature attractors as control vertices, a minimum outer envelope polytope is constructed, including: The directional derivative magnitude of each point in the time-parameter joint space is calculated for the deviation mapping feature sequence. The local maxima of the directional derivative magnitude are located as candidate feature attractors. The precise coordinates of each candidate feature attractor are recorded. The precise coordinates are composed of a specific timestamp and the corresponding concrete strength deviation value, rebar cover thickness deviation value, rebar spacing deviation value, and flatness deviation value, thus obtaining an initial set of feature attractors. Non-maximum suppression is applied to the initial set of feature attractors to remove redundant stationary points whose linear distance in the feature space is less than a preset distance threshold, and feature attractors whose gradient direction angle between adjacent feature attractors is greater than a preset angle threshold, thus forming an optimized set of feature attractors. Using the precise coordinates of each feature attractor in the optimized feature attractor set as control vertices, a geometric method of constructing a minimum closed polytope around the control vertices is adopted. The outer vertices are connected sequentially to form an outer envelope surface, forming a minimum outer envelope polytope containing all control vertices, thus obtaining the minimum outer envelope polytope.

5. The method for identifying anomalies in building structure inspection data according to claim 4, characterized in that, Affine transformation is performed on the smallest externally oriented envelope multicell and fitted to the standard hyperellipsoid space to generate a geometrically constrained hypersurface. The weighted orthogonal statistic of the correlation of each point in the deviation mapping feature sequence with respect to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface are calculated to obtain the geometrically enhanced feature sequence, including: Apply an affine transformation to the minimum externally enclosed polycell, stretching, rotating, and translating it to the standard hyperellipsoid space, so that the minimum externally enclosed polycell is inscribed in the standard hyperellipsoid, generating a geometrically constrained hypersurface; Project the coordinates of each point in the deviation mapping feature sequence onto the geometrically constrained hypersurface, calculate the weighted orthogonal statistic of the correlation between each projection point and the origin, and obtain the weighted orthogonal distance value of each point. Calculate the angle between the normal direction of each point in the deviation mapping feature sequence and the tangent plane of the hypersurface at this point, and obtain the hypersurface normal deflection angle at each point; The weighted orthogonal distance value and the hypersurface normal deflection angle are combined as a vector to form a geometrically enhanced feature sequence.

6. The method for identifying anomalies in building structure inspection data according to claim 5, characterized in that, The geometrically enhanced feature sequence is input into a pre-trained hybrid anomaly recognition model. The pre-trained hybrid anomaly recognition model sequentially performs static threshold screening, temporal evolution trend analysis, and multi-dimensional weighted correlation inference to obtain anomaly discrimination probability sequences and anomaly type labels, including: The geometrically enhanced feature sequence is input into a pre-trained hybrid anomaly recognition model, and a static threshold screening is performed, including comparing each feature value in the geometrically enhanced feature sequence with a preset static threshold, marking feature points that exceed the threshold range as candidate anomalies, and obtaining a preliminary anomaly set. A time-series evolution trend analysis is performed on the candidate anomalies in the initial anomaly set, including calculating the rate of change of the feature values ​​of each candidate anomaly in the forward and backward neighborhoods on the time axis, and selecting candidate anomalies with consistent change rate signs and cumulative change exceeding the preset trend threshold to obtain a trend anomaly set. Multidimensional weighted correlation reasoning is performed on the feature points in the trend anomaly set, including calculating weighted anomaly scores based on preset weights of each parameter dimension, and fusing the correlation constraints between different dimensions to obtain anomaly discrimination probability sequences and anomaly type labels corresponding to each feature point.

7. The method for identifying anomalies in building structure inspection data according to claim 6, characterized in that, The anomaly discrimination probability sequence and anomaly type label are subjected to confidence verification and logical mutual exclusion determination. After obtaining the verification-passing determination result, associated operating condition parameters, and data traceability path, they are structured and encapsulated to obtain anomaly identification alarm instructions, which are then sent to the management terminal, including: The confidence level of each anomaly discrimination probability value in the anomaly discrimination probability sequence is verified, and anomaly discrimination probabilities and their corresponding anomaly type labels that are lower than the preset confidence threshold are removed to obtain a high-confidence anomaly set. Logical mutual exclusion is performed on the labels of each anomaly type in the high confidence anomaly set. The system checks whether there are contradictory anomaly type labels at the same detection location and within the same time window. If they exist, the label with the highest priority is retained and the contradictory label is removed, thus obtaining the anomaly set that passes the mutual exclusion check. Extract the associated working condition parameters and data tracing paths corresponding to each anomaly point from the set of anomalies that have passed mutual exclusion verification. The associated working condition parameters include the environmental parameters and component construction information corresponding to the anomaly point. The data tracing paths include the original data stream source identifier, timestamp, and parameter dimension index. The verified judgment result, associated operating condition parameters, and data traceability path are structured and encapsulated according to the preset alarm format to form an anomaly identification alarm command, which is then sent to the control terminal.

8. A system for identifying anomalies in building structure inspection data, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect raw data streams and related data such as environmental parameters, component construction information, and historical quality trajectories of similar components for building structure inspection. The data governance module is used to perform dynamic quality weighted fusion and adaptive time window normalization on the raw data stream of building structure inspection using an adaptive data governance mechanism, and to embed environmental parameters, component construction information and historical quality trajectories of similar components to obtain a condition-enhanced inspection dataset carrying historical quality trajectories. The baseline interval generation module is used to perform working condition deviation compensation calculation and dynamic threshold fitting on the working condition enhancement detection dataset to generate an adaptive dynamic baseline interval and obtain a deviation mapping feature sequence. The geometric feature enhancement module is used to dynamically identify multiple feature attractors from the deviation mapping feature sequence, and construct a minimum outer envelope polycell using the multiple feature attractors as control vertices; perform an affine transformation on the minimum outer envelope polycell and fit it to the standard hyperellipsoid space to generate a geometrically constrained hypersurface; calculate the correlation weighted orthogonality statistic of each point in the deviation mapping feature sequence relative to the geometrically constrained hypersurface and the normal deflection angle of the hypersurface to obtain the geometrically enhanced feature sequence; The anomaly detection module is used to input the geometrically enhanced feature sequence into the pre-trained hybrid anomaly detection model, and sequentially perform static threshold screening, temporal evolution trend analysis and multi-dimensional weight correlation inference to obtain anomaly discrimination probability sequence and anomaly type label. The alarm delivery module is used to perform confidence verification and logical mutual exclusion judgment on the anomaly discrimination probability sequence and anomaly type label. After obtaining the judgment result that passes the verification, the associated operating condition parameters and the data traceability path, it is structured and encapsulated to obtain the anomaly identification alarm instruction and send it to the management terminal.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images