A building outer wall hollowing evaluation method and system

Abstract

The application belongs to the technical field of data processing, and specifically discloses a building outer wall hollow evaluation method and system, which comprises the following steps: first, obtaining a target building outer wall detection signal; after time window segmentation and normalization preprocessing, inputting the self-encoder model pre-trained with normal wall detection signals to obtain a reconstruction signal; calculating the residual error of the detection signal and the reconstruction signal; extracting the multi-dimensional features composed of residual energy, residual maximum absolute value, spike density and multi-scale feature scores from the residual error; substituting the multi-dimensional features into an abnormal score function to obtain an abnormal score value; comparing the abnormal score value with a threshold value based on the statistical characteristics of normal samples to determine whether the target outer wall region has hollow or potential hollow. The application can accurately capture subtle hollow signals, improve the objectivity and accuracy of detection, and meet the fine evaluation requirements of building outer wall hollow.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for assessing hollowness in building exterior walls. Background Technology

[0002] In the field of building maintenance, hollow areas in exterior walls, such as decreased adhesion between the finish and the substrate, early-stage cracks, or localized detachment, are common hidden dangers. If not detected in time, they can easily lead to safety problems such as wall detachment and water seepage. Current mainstream methods for detecting hollow exterior walls mainly rely on manual tapping and listening, conventional ultrasonic testing, or infrared thermography, but these methods have significant technical limitations.

[0003] Manual tapping relies on the experience of the inspectors, is highly subjective, has low accuracy in identifying minor hollow areas such as those with reduced bonding strength but not completely detached, and is difficult to cover the entire exterior wall area of ​​high-rise buildings. Although conventional ultrasonic or infrared detection achieves semi-automation, it is affected by environmental noise such as mechanical vibration, electromagnetic interference, and uneven wall materials. The characteristic signals of minor hollow areas are easily submerged in background noise, making it impossible to effectively distinguish noise interference from real hollow signals.

[0004] Existing technologies lack precise extraction and analysis methods for subtle hollow signals, making it difficult to establish a standardized anomaly judgment mechanism. This results in low detection efficiency and high rates of missed and false detections, failing to meet the needs of modern building exterior wall maintenance for early hazard identification. There is an urgent need for an efficient hollow assessment method that can enhance the sensitivity of subtle abnormal signals and reduce noise interference. Summary of the Invention

[0005] To address this issue, the present invention provides a method and system for assessing hollow areas in building exterior walls, thereby resolving the aforementioned technical problems.

[0006] This invention provides a method for assessing hollow areas in building exterior walls, comprising the following steps:

[0007] Acquire the detection signal of the exterior wall of the target building, and preprocess the detection signal, including time window segmentation and normalization processing;

[0008] The preprocessed detection signal is input into a pre-trained autoencoder model to reconstruct the detection signal and obtain a reconstructed signal. The autoencoder model is trained using normal wall detection signals as training samples.

[0009] Calculate the residual between the detected signal and the reconstructed signal;

[0010] Multi-dimensional features are extracted from the residuals, including residual energy, maximum absolute value of residuals, peak density, and multi-scale feature scores.

[0011] The multi-dimensional features are substituted into the anomaly scoring function to obtain an anomaly score value. The anomaly score value is then compared with a threshold obtained based on the statistical features of normal samples to determine whether there is hollowness or potential hollowness in the target exterior wall area.

[0012] Further, the step of substituting the multi-dimensional features into the anomaly scoring function to obtain the anomaly score includes: standardizing the difference between the mean of the residual energy and the residual energy of the normal samples, multiplying the result by a first weighting coefficient to obtain a first score; standardizing the difference between the maximum absolute value of the residual and the mean of the peak value of the normal sample residuals, multiplying the result by a second weighting coefficient to obtain a second score; multiplying the residual peak by a third weighting coefficient to obtain a third score; multiplying the feature scores under multiple scales by a fourth weighting coefficient to obtain a fourth score; and summing the first score, second score, third score, and fourth score to obtain the anomaly score.

[0013] Furthermore, by performing wavelet transform on the residuals, the ratio of the residual energy at each scale to the mean residual energy of the normal samples is calculated, and the weighted summation is used to obtain the multi-scale feature score.

[0014] Furthermore, the same exterior wall area is repeatedly inspected at different time periods, and the abnormal score values ​​of multiple inspections are compared. If the abnormal score value exceeds the threshold continuously within a preset time range, the area is determined to be a high-confidence potential hollow area.

[0015] Furthermore, cluster analysis is performed on spatially adjacent detection units. When the abnormal scores of adjacent units simultaneously exceed the threshold, the adjacent regions are jointly determined as hollow areas.

[0016] Furthermore, the autoencoder model is a one-dimensional convolutional autoencoder, and a temporal attention module is introduced in the latent representation layer to enhance sensitivity to minor anomalies.

[0017] In another aspect, this application also provides a building exterior wall hollowness assessment system, comprising:

[0018] The detection signal acquisition module is used to acquire the detection signal of the exterior wall of the target building and to preprocess the detection signal, the preprocessing including time window segmentation and normalization processing;

[0019] The reconstructed signal acquisition module is used to input the preprocessed detection signal into a pre-trained autoencoder model to reconstruct the detection signal and obtain a reconstructed signal. The autoencoder model is trained using normal wall detection signals as training samples.

[0020] The residual calculation module is used to calculate the residual between the detected signal and the reconstructed signal;

[0021] A multi-dimensional feature extraction module is used to extract multi-dimensional features from the residual, including residual energy, maximum absolute value of residual, peak density, and multi-scale feature score.

[0022] The hollow wall detection module is used to substitute the multi-dimensional features into the anomaly scoring function to obtain an anomaly score value, and compare the anomaly score value with a threshold obtained based on the statistical features of normal samples to determine whether there is hollow wall or potential hollow wall in the target exterior wall area.

[0023] In another aspect, this application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method for assessing hollowness in building exterior walls as described above.

[0024] In another aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor to implement a method for assessing hollowness in building exterior walls as described above.

[0025] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for assessing hollowness in building exterior walls.

[0026] This invention optimizes the quality of the detection signal through time window segmentation and normalization preprocessing, and combines a one-dimensional convolutional autoencoder with a time attention module to enhance the ability to capture weak signals of subtle hollow areas and reduce environmental noise interference. It extracts multi-dimensional features from the residuals and quantifies the degree of anomaly through an anomaly scoring function, setting thresholds based on the statistical characteristics of normal samples to improve the objectivity of the judgment. Combined with time stability verification and spatial clustering analysis, it reduces the false negative and false positive rates, accurately identifying potential hollow areas. Overall, it achieves automation, standardization, and refinement in the detection of hollow areas in building exterior walls, solving the problems of strong subjectivity and difficulty in identifying subtle hollow areas in traditional methods. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0029] Figure 1 A flowchart of a method for assessing hollowness in building exterior walls provided in an embodiment of the present invention.

[0030] Figure 2 A schematic diagram of the autoencoder model architecture is provided for embodiments of the present invention.

[0031] Figure 3 This is a schematic diagram of a building exterior wall hollowness assessment system provided in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] The technical solutions of this application will be described in detail below with reference to various embodiments.

[0035] like Figure 1 As shown, this embodiment of the invention discloses a method 100 for assessing hollow areas in building exterior walls, comprising the following steps:

[0036] S101, acquire the detection signal of the exterior wall of the target building, and preprocess the detection signal, the preprocessing including time window segmentation and normalization processing;

[0037] S102, the preprocessed detection signal is input into the pre-trained autoencoder model to reconstruct the detection signal and obtain the reconstructed signal, wherein the autoencoder model is trained using normal wall detection signals as training samples;

[0038] S103, calculate the residual between the detected signal and the reconstructed signal;

[0039] S104, extract multi-dimensional features from the residual, the multi-dimensional features including residual energy, maximum absolute value of residual, peak density and multi-scale feature score;

[0040] S105, substitute the multi-dimensional features into the anomaly scoring function to obtain an anomaly score value, and compare the anomaly score value with a threshold obtained based on the statistical features of normal samples to determine whether there is hollowness or potential hollowness in the target exterior wall area.

[0041] In some embodiments, for S101, exemplarily, the detection signals include two core types of signals: acoustic and image / thermal. Specifically, acoustic signal acquisition uses a dedicated microphone for building inspection, and the sampling rate needs to be set according to the equipment capabilities, preferably 16kHz or 44.1kHz. Image acquisition uses a standard camera with a resolution of 2-12MP; thermal image acquisition uses a thermal imager, and linear temperature calibration must be performed in advance to eliminate the influence of the equipment's own temperature error on the detection results. A resolution of 2-12MP can meet the detailed presentation of exterior wall panels or a 0.5m×0.5m detection grid, and the thermal image after linear temperature calibration can accurately reflect the temperature distribution differences on the wall surface.

[0042] Next, the detection units were divided according to a unified standard, and data was collected over multiple time periods. Specifically, the exterior wall of the target building was divided into several independent detection units based on the natural boundaries of the exterior wall panels or a 0.5m × 0.5m grid. This not only closely matches the actual structure of the exterior wall but also ensures the representativeness of the signal in each unit, avoiding the masking of local hollow signals due to excessively large units.

[0043] For each detection unit, at least three short-duration signals should be collected, each lasting 1-10 seconds. Selecting different time periods for collection, such as morning, noon, and evening, can eliminate occasional interference from environmental noise such as traffic noise and wind noise. Stable signal characteristics can then be screened out through subsequent time consistency analysis.

[0044] In this embodiment, the acquired signal is first synchronized and denoised, then time window segmentation and normalization are performed, and finally standardized model input samples are output.

[0045] Specifically, for acoustic signals and image / thermal imaging signals from the same detection unit, the acquisition time of the acoustic signal is used as a reference, and the image / thermal imaging is associated with the time window of the acoustic signal according to spatial blocks to achieve spatiotemporal alignment. For example, after the acoustic signal is divided into 1-second time windows, the image / thermal imaging blocks of the detection unit within the same time period are extracted to ensure that the temporal and spatial dimensions of different types of signals are consistent in subsequent analysis.

[0046] For acoustic signals only, bandpass filtering, such as 100Hz-8kHz, is used to remove noise. Low-frequency noise below 100Hz mainly comes from surrounding mechanical vibrations such as construction and vehicle movement, while high-frequency noise above 8kHz is equipment quantization noise. Bandpass filtering can preserve the effective acoustic frequency band related to hollowness and reduce the interference of irrelevant noise on subsequent processing. Image / thermal imaging signals do not require additional filtering, and their noise can be indirectly eliminated through subsequent normalization and model training.

[0047] In this embodiment, the core of time window segmentation is to divide a continuous long signal into subsequences with local features, which facilitates the extraction of subtle hollow features by the deep autoencoder.

[0048] Specifically, taking an acoustic signal as an example, it is divided into segments with a fixed window length of 1 second and an overlap rate of 50%. The 1-second window length can completely cover the short acoustic pulses that may be generated by the hollow sound, which are usually in the millisecond range. The 50% overlap rate can prevent the loss of signal features at the window boundaries and ensure the continuity of features between adjacent windows. For example, a 10-second acoustic signal, after being divided into segments with a 1-second window length and 50% overlap, can be divided into 19 time window subsequences. For example, the calculation formula is: number of subsequences = (total duration - window length) / overlap step + 1, overlap step = window length × 50% = 0.5s, that is, (10-1) / 0.5+1 = 19.

[0049] By combining the spatiotemporal alignment results, the image / thermal image blocks corresponding to the acoustic time window are divided into segments at the same time intervals. Each image / thermal image block is bound to an acoustic time window subsequence to form an acoustic-image / thermal image pairing sample.

[0050] Normalization can eliminate signal amplitude differences between different detection units and under different acquisition conditions, allowing the model to focus on feature patterns rather than absolute intensity. For example, a combination of zero-mean and unit variance normalization is used to process each time window sequence independently, and the calculation formula is as follows:

[0051] Zero-mean: Where x(t) is the signal value of the time window subsequence at time t, and N is the length of the time window subsequence. This is the mean of the sequence within that time window; the mean of the processed signal is 0.

[0052] Unit variance normalization: ,in This represents the standard deviation of the time window sequence, and the processed signal variance is 1. This can unify signals with different amplitude ranges, such as the differences in acoustic signal amplitudes from different detection units, to the same scale, i.e., a mean of 0 and a variance of 1, avoiding misjudging the importance of features due to signal amplitude magnitude.

[0053] During the normalization process, the original mean of each time window sequence is saved simultaneously. with standard deviation When calculating quantitative indicators of the reconstructed residuals, such as residual energy and maximum residual peak value, it is necessary to recover them in reverse from the original amplitude. This ensures that the residual index truly reflects the physical meaning of the signal, rather than a normalized relative value.

[0054] In some embodiments, for S102, such as Figure 2 As shown, the autoencoder model adopts a three-segment structure of encoder-attention module-decoder, focusing on feature compression and reconstruction of one-dimensional time-series signals, primarily acoustic, while also being compatible with image / thermal imaging signal adaptation processing.

[0055] Specifically, the encoder progressively compresses the preprocessed temporal window sub-sequences, such as the 1-second acoustic window and the corresponding image patch feature vectors, into low-dimensional latent representations through multiple convolutional layers and downsampling. For example, this includes:

[0056] Set up 3-6 1D convolutional layers, with each layer sequentially connected to a batch normalization layer, a ReLU activation function, and a downsampling layer. The choice of 3-6 layers can balance feature extraction capability and computational efficiency. Shallow layers (3-4 layers) are suitable for scenarios with simple signal features, such as a single concrete exterior wall, while deep layers (5-6 layers) can capture the subtle features of complex wall materials, such as composite structures of stone and insulation layers.

[0057] The convolution kernel length is set to 3-15, where the small kernel with a length of 3-5 is used to extract short-term details such as millisecond-level acoustic pulses generated by hollow drums, and the large kernel with a length of 10-15 is used to capture the periodic features of medium-term structures such as wall vibrations; downsampling is achieved by adjusting the step size or max pooling, with the step size set to 2. The signal length is halved after each downsampling layer to ensure that the feature dimension is gradually compressed.

[0058] The final output latent representation dimension is set to 16-128, which can be adjusted according to the length of the input time window. For example, a 1-second acoustic window with a sampling rate of 16kHz and 16,000 data points is downsampled to 1,000 dimensions after 4 layers of downsampling (each layer with a stride of 2), and then mapped to a 64-dimensional latent vector through a convolutional layer. This preserves key features while avoiding overfitting due to excessive dimensionality.

[0059] For the attention module, a temporal attention mechanism is introduced in the latent representation layer or the intermediate layer of the encoder, such as after the 2nd or 3rd convolutional layer, to highlight the intermittent acoustic signals of subtle hollows that recur in time.

[0060] Specifically, either the squeeze-and-excitation (SE) mechanism or the self-attention mechanism is employed. The SE mechanism redistributes the weights of feature channels through a squeeze-excitation operation, enhancing channel features related to hollowness, such as the 1-4kHz acoustic frequency band. The self-attention mechanism focuses on temporally recurring local features, such as the hollowness impact signal that occurs every 0.2 seconds, by calculating the correlation weights within features. The two mechanisms can be selected based on the signal type: the SE mechanism is preferred for acoustic signals, while the self-attention mechanism is preferred for image feature vectors.

[0061] For the decoder, the decoder adopts a structure symmetrical to the encoder. Through deconvolution and upsampling, the latent representation is restored to a reconstructed signal with the same length as the input window sequence.

[0062] Specifically, the number of layers is completely symmetrical to that of the encoder. For example, if the encoder has 4 layers, the decoder has 4 layers. Each layer consists of a deconvolution layer, a batch normalization layer, and a ReLU activation function, all sequentially connected in series, except for the last layer. The kernel length and stride of the deconvolution layer match the corresponding layer in the encoder to ensure that the signal dimension is gradually restored after upsampling. For example, if a certain layer in the encoder uses a "10-length convolution kernel and a stride of 2 for downsampling", the corresponding layer in the decoder will use a "10-length deconvolution kernel and a stride of 2 for upsampling".

[0063] The last layer of the decoder uses a linear activation function instead of ReLU to avoid truncating the signal amplitude, ensuring that the amplitude range of the reconstructed signal is consistent with the original preprocessed signal, and providing an accurate amplitude basis for subsequent residual calculation.

[0064] In one embodiment, the model is trained using only confirmed normal wall detection signals, and the model learns the characteristic patterns of normal signals through iterative optimization.

[0065] Specifically, multiple healthy buildings without hollow records were selected, covering different building ages and different wall materials, such as concrete, stone, and glass curtain walls. Following the detection unit division-signal acquisition-preprocessing process described in the aforementioned embodiment, at least 10,000 preprocessed time window sequences were obtained, including acoustic and image / thermal image paired samples, to ensure sample diversity.

[0066] For example, the normal sample set is divided into a training set (70%), a validation set (20%), and a test set (10%) in a ratio of 7:2:1. The training set is used for updating model parameters, the validation set is used to monitor overfitting, and the validation set reconstruction error is calculated after each round of training. The test set is used to finally evaluate the model's reconstruction accuracy of normal signals.

[0067] In this embodiment, the main loss function employs a hybrid loss of mean squared error (MSE) and frequency domain error. Specifically, time-domain MSE: calculates the time-domain amplitude difference between the input signal and the reconstructed signal, using the following formula:

[0068]

[0069] in, For input signal, To reconstruct the signal, N is the signal length, ensuring accurate reconstruction of the time-domain features;

[0070] Frequency domain error, representing the spectral amplitude error calculated by performing a short-time Fourier transform (STFT) on the input and reconstructed signals, is given by the following formula:

[0071]

[0072] in, , The STFT spectral amplitudes of the input and reconstructed signals are respectively used to enhance the reconstruction quality of the 1-4kHz hollow-drum correlation band.

[0073] Mixed loss weights, let , , , To represent weights, for example, , To achieve a balance between time-domain and frequency-domain reconstruction effects.

[0074] Optionally, L2 regularization can be added, with the weight decay coefficient set to 1e-5, or sparse constraints can be added, with L1 regularization applied to the latent vectors and the coefficient set to 1e-6, to encourage sparsity of model parameters, reduce the risk of overfitting, and are especially suitable for scenarios with a small sample size.

[0075] For example, the Adam optimizer is selected, and the learning rate is set to 1e-4 to 1e-3. In the early stage of training, 1e-3 is used to accelerate convergence, and in the later stage, it is reduced to 1e-4 for fine-tuning. The batch size is set to 32 to 256, with 256 used when the sample size is ≥10000 and 32 used when the sample size is <5000. The number of training epochs is set to 50 to 200. The training is stopped when the validation set reconstruction error does not decrease for 5 consecutive epochs.

[0076] In this embodiment, the preprocessed target building detection signal is input into the pre-trained autoencoder model, and the reconstructed signal is output through the feature compression-attention enhancement-signal restoration process.

[0077] Specifically, the preprocessing window sequence of the target building, such as the acoustic signal as a 1×16000 vector and the image patch as a 1×2048 feature vector, is organized according to the training format to ensure that the dimension matches the model input layer. For example, the encoder input needs to be a three-dimensional tensor of batch size × signal length × 1, and the batch size is set to 1 when inputting a single sample.

[0078] By inputting the acoustic time window and the corresponding image feature vector into different input branches of the encoder, the attention module will simultaneously adjust the weights of the potential features of the two signals to improve the reconstruction accuracy.

[0079] The model inference mode is activated, which disables the update function of the batch normalization layer. The preprocessed signal is compressed into a latent vector by the encoder. The attention module enhances the weights of weak features in the latent vector, such as acoustic features that resemble hollow sounds. The input signal is then fed into the decoder and gradually reconstructed through deconvolution and upsampling. Finally, the reconstructed signal with the same length as the input signal is output. .

[0080] In some embodiments, for S103, the temporal residual is the core of all subsequent feature extraction. Based on the preprocessed detection signal and the reconstructed signal output by the model, it is calculated according to the point-by-point difference rule to ensure that the residual can truly reflect the signal difference.

[0081] Specifically, since the preprocessing stage performs zero-mean and unit variance normalization on the detection signal, it is necessary to first use the saved original mean. with standard deviation The normalized detection signal and the reconstructed signal are then restored to their original amplitudes in a physically meaningful way, as shown in the following formula:

[0082] Recovery of the original amplitude of the detected signal:

[0083] Reconstruction of the original amplitude of the signal:

[0084] The reconstructed signal and the detection signal use the same set here. and Since both originate from the same time window sequence of the same detection unit, the amplitude reference is consistent, avoiding residual distortion caused by differences in normalization parameters.

[0085] Based on the point-by-point difference rule between the original detected signal and the original reconstructed signal, the time-domain residual sequence is calculated and denoted as r(t), with the following formula:

[0086]

[0087] t=1,2,…,N, where N is the length of the time window subsequence (e.g., 16000) and t is the time sampling point. After point-by-point calculation, a residual sequence with the same length as the original signal is obtained. This sequence contains anomalous information in the detected signal that was not accurately reconstructed by the model (based on normal signal training), i.e., the signal differences that may correspond to subtle hollows.

[0088] In some embodiments, for S103, before extracting multi-dimensional features, two types of core data are identified to provide a basis for feature calculation and anomaly detection:

[0089] Temporal residual sequence: The original temporal residual sequence calculated based on the above embodiments. N is the length of the time window sequence, which contains information about the difference between the detected signal and the normal signal, i.e., the model reconstruction.

[0090] Normal sample baseline statistics: From the normal sample library, i.e., the normal wall signal set used to train the autoencoder model, residual data of the same type of detection unit (i.e., the same wall material, the same detection equipment, and the same window length) are extracted, and key baseline indicators are statistically analyzed, including:

[0091] Mean residual energy Standard deviation ;

[0092] Maximum absolute value of residuals Standard deviation ;

[0093] The mean residual energy at each wavelet scale, where l is the number of wavelet decomposition levels;

[0094] These baseline statistics are used for subsequent feature standardization and outlier assessment to ensure the objectivity of feature extraction.

[0095] Specifically, for residual energy extraction, for example, residual energy reflects the overall energy intensity of the residual sequence. Higher energy indicates a greater overall deviation between the detected signal and the normal signal, serving as a fundamental time-domain characteristic for determining whether anomalies exist in the wall. For instance, ;

[0096] For the extraction of the maximum absolute value of residuals, for example, the maximum absolute value of residuals focuses on extreme outliers in the residual sequence, which correspond to short-term impact signals that may be generated by minor hollows, such as the sudden reflection of sound waves at the hollow interface, and can effectively identify local structural changes such as the initial stage of cracks.

[0097] Specifically, for each sampling point in the time-domain residual sequence r(t), its absolute value |r(t)| is calculated to eliminate the influence of the sign of the residual on the amplitude judgment; all |r(t)| are traversed, and the maximum value is extracted as the maximum absolute value of the residual, denoted as max|r|. max|r| is then compared with the baseline statistic of the normal samples. In comparison, if max|r|> +2 If the value is less than 2 standard deviations, then the point is marked as a potential outlier, providing a reference for subsequent peak density calculations.

[0098] For peak density extraction, for example, peak density measures the number of valid abnormal peaks in the residual sequence per unit time, which can effectively distinguish between isolated noise peaks (low density) and continuous / clustered peaks (high density) related to cavitation. Specifically, a peak determination threshold T is set based on the background noise level of the normal sample residuals. peak Multiply the mean of the absolute values ​​of the residuals of normal samples by 1.2-1.5 times to balance sensitivity and anti-interference, or use the mean of the absolute values ​​of the first 10% stationary segments of the target residual sequence multiplied by 1.5 times as a temporary threshold;

[0099] For the target residual sequence r(t), iterate through each sampling point, excluding the first and last points. If r(t)>r(t-1) and (r(t)>r(t+1), then the point is determined to be a local maximum point, ensuring that the selected peaks have local significance.

[0100] Count all local maxima where r(t) > T peak The number of points, denoted as C. peak ;

[0101] Using the duration of the time window subsequence (e.g., 1 second) as the time base, calculate the number of effective spikes per unit time, i.e., the spike density D. peak The formula is D peak =C peak / T win T win D is the duration of the time window subsequence (in seconds). For example, if there are 6 effective spikes within a 1-second time window, then D... peak =6 peaks / s; if the time window is 2s and the number of effective spikes is 10, then D peak =5 times / s, ensuring that the peak density of time windows of different durations is comparable.

[0102] For example, multi-scale feature score extraction can capture the abnormal features of residuals at different time scales through wavelet transform, reflecting whether the anomaly has cross-scale stability. That is, real hollows usually show abnormal behavior at multiple scales, while noise is mostly single-scale random fluctuation.

[0103] Specifically, Daubechies wavelets (such as db4 and db6) are selected to perform 3-5 layers of wavelet decomposition on the time-domain residual sequence r(t). The number of layers is adjusted according to the time window length. A 1-second acoustic window can have 4 layers. After decomposition, detail coefficients d1, d2, d3, and d4 are obtained at each scale, which correspond to the scale from high frequency to low frequency. d1 is the highest frequency and d4 is the lowest frequency. The overall trend is obtained together with the approximation coefficient a4.

[0104] Detail factor for each scale (l=1,2,3,4), calculate its energy. The formula is:

[0105]

[0106] in For detail coefficients The length of the wavelet decomposition is approximately 1 / 2^l of the original sequence, where the length of each detail coefficient after wavelet decomposition is approximately 1 / 2^l.

[0107] Based on the baseline statistic of the normal sample That is, the mean residual energy of normal samples at scale l, and the outlier ratio at each scale is calculated. =, if A value >1 indicates that the residual energy at this scale is higher than the normal level, indicating an anomaly.

[0108] Anomaly ratio across all scales Perform weighted summation to obtain multi-scale feature scores. The weighting should be combined with the scales sensitive to voids (e.g., high-frequency scales d1 and d2 are more sensitive to minor voids, so their weights can be set to 0.3, while the weights of low-frequency scales d3 and d4 can be set to 0.2).

[0109] The four extracted features include residual energy. , Maximum absolute value of residuals max|r|, Peak density D peak Multi-scale feature scores z-score standardization is performed based on the baseline statistics of normal samples to ensure that all features are on the same order of magnitude, providing standardized input for the subsequent anomaly scoring function.

[0110] In some embodiments, for S105, the core of the anomaly scoring function is to fuse four types of features—residual energy, maximum absolute value of residuals, peak density, and multi-scale feature scores—according to weights to quantify the overall anomaly degree of the residuals.

[0111] This includes standardizing the difference between the mean residual energy and the residual energy of normal samples, multiplying the result by a first weighting coefficient; standardizing the difference between the maximum absolute value of the residual and the mean peak value of the normal sample residuals, multiplying the result by a second weighting coefficient; multiplying the residual peak by a third weighting coefficient; multiplying the feature score under multiple scales by a fourth weighting coefficient; and summing the above four items with weights to obtain the anomaly score. This can be simplified to the following formula:

[0112]

[0113] The weights can be set through validation set learning or engineering experience. The core principle is to assign higher weights to features that are sensitive to voids. For example, w1=0.2 residual energy reflects the overall difference and has a moderate weight; w2=0.3 the maximum absolute value of the residual captures local peaks and is sensitive to subtle voids and has a high weight; w3=0.3 peak density distinguishes noise from real voids and has a high weight; w4=0.2 multi-scale feature score verifies cross-scale stability and has a moderate weight; and it satisfies w1+w2+w3+w=1.

[0114] In this embodiment, the threshold is obtained based on the statistical characteristics of normal samples, and the median absolute deviation (MAD) is used instead of the standard deviation to reduce the interference of abnormal samples, such as a small number of noisy samples in the normal sample library, on the threshold.

[0115] Compare the abnormal score S of the detection unit in the target exterior wall area with the set threshold T;

[0116] If S≤T: it is determined that the detection unit has no voids, indicating that its residual abnormality is within the normal range and there are no obvious void characteristics;

[0117] If S>T: The detection unit is determined to be a suspected hollow area, and further verification is needed to eliminate accidental noise interference.

[0118] Optionally, the same exterior wall area can be repeatedly inspected at different time periods, and the abnormal score values ​​of multiple inspections can be compared. If the abnormal score value exceeds the threshold continuously within a preset time range, the area is determined to be a high-confidence potential hollow area.

[0119] If the scores of all repeated tests are >T: it is judged as a high-confidence potential hollow area, indicating that the abnormal features have time stability and eliminating random noise interference;

[0120] If only one score is greater than T, and all others are less than or equal to T, it is judged as a low-confidence suspected area, which may be due to temporary noise during the detection, such as wind or surrounding vibration, and needs to be re-verified later.

[0121] Optionally, cluster analysis is performed on spatially adjacent detection units. When the abnormal score values ​​of adjacent units simultaneously exceed the threshold, the adjacent regions are jointly determined as hollow areas.

[0122] If the abnormal scores of adjacent units are all >T: these adjacent areas are jointly determined as hollow areas, because hollow areas usually have a certain spatial range and will not exist only in a single isolated unit;

[0123] If only a single isolated unit scores >T, while surrounding units are all normal: it is judged as an isolated suspected point, which may be due to signal interference caused by local surface noise such as wall stains. It is not judged as hollow at this time, and manual verification is required.

[0124] Figure 3 A building exterior wall hollowness assessment system 300 is shown. Embodiments of this system are... Figure 1 Corresponding to the illustrated method embodiments, this system can be specifically applied to various electronic devices. Specifically, it includes:

[0125] The detection signal acquisition module 301 is used to acquire the detection signal of the exterior wall of the target building and preprocess the detection signal, the preprocessing including time window segmentation and normalization processing;

[0126] The reconstructed signal acquisition module 302 is used to input the preprocessed detection signal into a pre-trained autoencoder model to reconstruct the detection signal and obtain a reconstructed signal, wherein the autoencoder model is trained using normal wall detection signals as training samples.

[0127] The residual calculation module 303 is used to calculate the residual between the detected signal and the reconstructed signal;

[0128] The multi-dimensional feature extraction module 304 is used to extract multi-dimensional features from the residual, the multi-dimensional features including residual energy, maximum absolute value of residual, peak density and multi-scale feature score;

[0129] The hollow wall detection module 305 is used to substitute the multi-dimensional features into the abnormal scoring function to obtain an abnormal score value, and compare the abnormal score value with a threshold obtained based on the statistical features of normal samples to determine whether there is hollow wall or potential hollow wall in the target exterior wall area.

[0130] Based on the same inventive concept, this application also provides an electronic device. The method corresponding to the electronic device can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. The electronic device provided in this application includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the methods and / or technical solutions of the foregoing embodiments of this application.

[0131] Figure 4 The diagram illustrates the structure of an apparatus suitable for implementing the methods and / or technical solutions in the embodiments of this application. The apparatus 400 includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 402 or a program loaded from a storage portion 408 into a random access memory (RAM) 403. The RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0132] The following components are connected to I / O interface 405: input section 406 including keyboard, mouse, touch screen, microphone, infrared sensor, etc.; output section 407 including cathode ray tube (CRT), liquid crystal display (LCD), LED display, OLED display, etc., and speakers, etc.; storage section 408 including one or more computer-readable media such as hard disk, optical disk, magnetic disk, semiconductor memory, etc.; and communication section 409 including network interface card such as LAN (local area network) card, modem, etc. Communication section 409 performs communication processing via a network such as the Internet.

[0133] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by the central processing unit (CPU) 401, it performs the functions defined in the methods of this application.

[0134] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0135] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application.

[0136] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for evaluating hollowing of a building exterior wall, characterized by, The method comprises the following steps: obtaining a target building outer wall detection signal, and preprocessing the detection signal, wherein the preprocessing comprises time window segmentation and normalization processing, specifically, time window segmentation is performed with a fixed window length, a preset overlap rate is set to form time window subsequences, and zero mean and unit variance normalization is adopted to independently process each time window subsequence, and the original mean and standard deviation of each time window subsequence are saved synchronously; inputting the preprocessed detection signal into a pre-trained autoencoder model to reconstruct the detection signal, and obtaining a reconstructed signal, wherein the autoencoder model takes normal wall detection signals as training samples, a time domain mean square error and a frequency domain error are mixed as a main loss function, the frequency domain error is a spectral amplitude error calculated by performing short-time Fourier transform on the input and reconstructed signals, and a time attention module is introduced in a latent representation layer, the time attention module adopts an SE mechanism or a self-attention mechanism to enhance the sensitivity to small abnormalities; calculating the residual error between the detection signal and the reconstructed signal; extracting multi-dimensional features from the residual error, wherein the multi-dimensional features include residual energy, residual maximum absolute value, spike density and multi-scale feature score; the multi-scale feature score is obtained by calculating the ratio of residual energy at each scale to the mean value of normal sample residual energy, and then weighting and integrating the ratio; inputting the multi-dimensional features into an anomaly scoring function to obtain an anomaly score value, comparing the anomaly score value with a threshold value obtained based on normal sample statistical characteristics to determine whether there is a hollow or potential hollow in the target outer wall area; and repeatedly detecting the same outer wall area at different time periods, comparing the anomaly score values obtained through multiple detections, and if the anomaly score value continuously exceeds the threshold value within a preset time range, determining that the outer wall area is a high-confidence potential hollow; and performing cluster analysis on spatially adjacent detection units, and when the anomaly score values of adjacent units simultaneously exceed the threshold value, jointly determining that the adjacent area is a hollow area; wherein the inputting the multi-dimensional features into the anomaly scoring function to obtain the anomaly score value comprises: standardizing the difference between the residual energy and the mean value of normal sample residual energy, multiplying the result by a first weight coefficient to obtain a first score; standardizing the difference between the residual maximum absolute value and the mean value of normal sample residual peak value, multiplying the result by a second weight coefficient to obtain a second score; multiplying the spike density by a third weight coefficient to obtain a third score; multiplying the multi-scale feature score by a fourth weight coefficient to obtain a fourth score; and summing the first score, the second score, the third score and the fourth score to obtain the anomaly score value.

2. A building exterior wall hollowing evaluation system characterized by, The method comprises the following steps: a detection signal acquisition module is configured to obtain a target building outer wall detection signal, and preprocess the detection signal, wherein the preprocessing comprises time window segmentation and normalization processing, specifically, time window segmentation is performed with a fixed window length, a preset overlap rate is set to form time window subsequences, and zero mean and unit variance normalization is adopted to independently process each time window subsequence, and the original mean and standard deviation of each time window subsequence are saved synchronously; The reconstruction signal acquisition module is configured to input the preprocessed detection signal into a pre-trained autoencoder model to reconstruct the detection signal and obtain a reconstruction signal, wherein the autoencoder model takes normal wall detection signals as training samples, a main loss function adopts a time domain mean square error and a frequency domain error hybrid loss, the frequency domain error is a spectral amplitude error calculated by performing a short-time Fourier transform on the input and the reconstruction signal, and a time attention module is introduced in a latent representation layer, and the time attention module adopts an SE mechanism or a self-attention mechanism to enhance the sensitivity to micro abnormalities; The residual error calculation module is configured to calculate a residual error between the detection signal and the reconstruction signal; The multi-dimensional feature extraction module is configured to extract multi-dimensional features from the residual error, including residual error energy, residual maximum absolute value, spike density and multi-scale feature score; wherein the multi-scale feature score is obtained by calculating the ratio of residual error energy at each scale to the average residual error energy of normal samples and then weighting and summarizing the ratio; The hollow judgment module is configured to input the multi-dimensional features into an anomaly scoring function to obtain an anomaly score value, compare the anomaly score value with a threshold value obtained based on the statistical characteristics of normal samples to determine whether the target external wall region has a hollow or potential hollow, repeatedly detect the same external wall region at different time periods, compare the anomaly score values obtained through multiple detections, and if the anomaly score value continuously exceeds the threshold value within a preset time range, determine that the external wall region is a high-confidence potential hollow; and perform cluster analysis on spatially adjacent detection units, and when the anomaly score values of adjacent units simultaneously exceed the threshold value, jointly determine that the adjacent region is a hollow region. The method comprises the following steps:

3. An electronic device, the electronic device comprising: obtaining a reconstruction signal by inputting a preprocessed detection signal into a pre-trained autoencoder model, wherein the autoencoder model takes normal wall detection signals as training samples, a main loss function adopts a time domain mean square error and a frequency domain error hybrid loss, the frequency domain error is a spectral amplitude error calculated by performing a short-time Fourier transform on the input and the reconstruction signal, and a time attention module is introduced in a latent representation layer, and the time attention module adopts an SE mechanism or a self-attention mechanism to enhance the sensitivity to micro abnormalities; calculating a residual error between the detection signal and the reconstruction signal; extracting multi-dimensional features from the residual error, including residual error energy, residual maximum absolute value, spike density and multi-scale feature score; wherein the multi-scale feature score is obtained by calculating the ratio of residual error energy at each scale to the average residual error energy of normal samples and then weighting and summarizing the ratio; inputting the multi-dimensional features into an anomaly scoring function to obtain an anomaly score value, comparing the anomaly score value with a threshold value obtained based on the statistical characteristics of normal samples to determine whether the target external wall region has a hollow or potential hollow, repeatedly detecting the same external wall region at different time periods, comparing the anomaly score values obtained through multiple detections, and if the anomaly score value continuously exceeds the threshold value within a preset time range, determining that the external wall region is a high-confidence potential hollow; and performing cluster analysis on spatially adjacent detection units, and when the anomaly score values of adjacent units simultaneously exceed the threshold value, jointly determining that the adjacent region is a hollow region.

4. A computer readable medium having stored thereon computer program instructions, characterized in that, The computer program instructions can be executed by the processor to implement the method of claim 1.

5. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the method of claim 1. The computer program is executed by the processor to implement the method of claim 1.

Images