Intelligent monitoring system for shear structure of fabricated net rack

Abstract

The invention discloses an intelligent monitoring system for an assembled net rack shear structure, and relates to the technical field of net rack structure health monitoring. A fiber grating sensor, an ultrasonic array, an electromagnetic coil and an energy assembly are integrated through a sensor integration module and a modular support; optical fiber strain, ultrasonic signals and electromagnetic impedance data are collected in real time; after the feature fusion module extracts key features through preprocessing, a fusion feature vector is obtained through an adaptive manifold projection fusion method; the shear force value prediction module takes the fusion feature vector as input, and obtains a shear force value prediction model through training by using an LSTM (Long Short Term Memory) model; and the net rack judgment module is combined with a DBSCAN algorithm to divide neighbor sensor network groups, and is used for judging the structure defect type based on a local anomaly concentration ratio LAC and an overall cooperative change rate GCR. According to the system, intelligent monitoring of the fabricated net rack shear structure is realized, and the monitoring precision and the defect identification accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of grid structure health monitoring, in particular to an intelligent monitoring system for assembled grid shear structure. BACKGROUND

[0002] As a lightweight and large-span space structure, the assembled grid is widely used in large-scale venues, transportation hubs and other projects. The shear state of the assembled grid is directly related to the safety and durability of the structure, so it is of great significance to monitor the shear state of the assembled grid intelligently.

[0003] The existing monitoring technology for the shear structure of the assembled grid has many limitations: first, the deployment of sensors mostly uses single type sensors, and the relative position of the sensors is unstable, resulting in poor data consistency; second, the feature processing mostly uses simple splicing and linear fusion methods, without considering the physical heterogeneity of multi-modal data such as fiber strain, ultrasonic signal and electromagnetic impedance, making it difficult to dig deep correlations between features; third, the shear value prediction mostly relies on empirical formulas, which cannot capture the dynamic change law of the structure state; fourth, the defect judgment is mostly based on a single index, lacking comprehensive analysis of spatial distribution and overall collaborative change, and being prone to misjudgment of local defects and overall structure deterioration.

[0004] Therefore, there is an urgent need for an intelligent monitoring system that integrates multi-modal fusion, high-precision prediction and accurate discrimination to solve the shortcomings of the existing technology. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides an intelligent monitoring system for the shear structure of the assembled grid, which solves the problems of insufficient multi-sensor integration, poor feature fusion, low shear prediction and defect discrimination accuracy.

[0006] To achieve the above purpose, the present application realizes the following technical scheme: an intelligent monitoring system for the shear structure of the assembled grid, comprising:

[0007] A sensor integration module is installed on the assembled grid to collect real-time fiber strain data, ultrasonic signal and electromagnetic impedance data, and the sensor network specifically includes fiber grating sensors, ultrasonic arrays, electromagnetic coils and energy components.

[0008] A feature fusion module is used to preprocess the fiber strain data, ultrasonic signal and electromagnetic impedance data, extract key features, and fuse the key features through an adaptive manifold projection fusion method to obtain a fusion feature vector.

[0009] A shear value prediction module is used to obtain historical data of each sensor network and construct a shear value formula, take the fusion feature vector as the input of LSTM, take the shear value as the output of LSTM, and train to obtain a shear value prediction model for each sensor network.

[0010] The net frame judgment module obtains the real-time shear value of each sensor network, and automatically divides the near-neighbor sensor network group by using the DBSCAN algorithm, and judges the assembly type net frame according to the local anomaly concentration degree LAC and the overall collaborative change rate GCR.

[0011] As a further scheme of the present application, the optical fiber grating sensor, the ultrasonic array, the electromagnetic coil and the energy component are integrated into one body by using the modular support, and are fixed on the reserved mounting position of the prefabricated rod member through bolts.

[0012] As a further scheme of the present application, the key features extracted from the pretreated optical fiber strain data include the strain peak value, the strain rate and the strain average value; the key features extracted from the pretreated ultrasonic signal include the echo amplitude attenuation rate, the defect depth and the wave velocity change rate; and the key features extracted from the pretreated electromagnetic impedance data include the impedance real part change amount, the impedance phase angle and the impedance modulus change rate.

[0013] As a further scheme of the present application, the specific operation steps of fusing the key features by the adaptive manifold projection fusion method are as follows:

[0014] The key features of the optical fiber strain, the ultrasonic signal and the electromagnetic impedance are respectively subjected to Z-score standardization, wherein the optical fiber strain includes the strain peak value, the strain rate and the strain average value, the ultrasonic signal includes the echo amplitude attenuation rate, the defect depth and the wave velocity change rate, and the electromagnetic impedance includes the impedance real part change amount, the impedance phase angle and the impedance modulus change rate.

[0015] The standardized three subspaces ψ1, ψ2 and ψ3 are respectively used to construct local neighborhood graphs by using the adaptive k-neighborhood algorithm, and are subjected to dimension reduction by using the local linear embedding LLE, so as to reduce the three-dimensional single-mode features to d-dimensional low-dimensional manifold features of each mode And is an integer;

[0016] A cross-modal contrast loss function is constructed, and the specific formula is Wherein, represents the low-dimensional feature of the i-th sample in the optical fiber mode, represents the low-dimensional feature of the j-th sample in the ultrasonic mode, represents the low-dimensional feature of the r-th sample in the electromagnetic mode.

[0017] The contrast loss is optimized by gradient descent to obtain the projection matrices P1, P2 and P3 of each mode to the common space;

[0018] The low-dimensional manifold features of each mode are mapped to the common space by the projection matrices to obtain are directly spliced into the final fusion vector:

[0019] As a further aspect of the present invention, in the process of constructing a local neighborhood graph using the adaptive k-nearest neighbor algorithm, according to the formula ρ i =∑ j exp(-||x i -x j || 2 Calculate the local density of the sample points. The k-value range is [5,8] for regions with high density and [2,4] for regions with low density.

[0020] As a further aspect of the present invention, the shear force value is calculated according to the formula V=V0×(1-α×R)×(1-β×δ), where V is the actual shear force value under defective or abnormal conditions, V0 is the basic shear force under normal structural conditions, R is the crack reflection coefficient of the ultrasonic signal, δ is the relative change rate of electromagnetic impedance, α is the influence coefficient of microcracks on shear force, and β is the influence coefficient of sleeve loosening on shear force.

[0021] As a further aspect of the present invention, based on the relationship between shear strain and shear stress of the space frame material, and combined with the shear strain data measured by the fiber optic grating sensor, the foundation shear force V0 under normal structural conditions is calculated according to the formula V0=G×γ×A, where G is the material shear modulus, γ is the shear strain measured by the fiber optic grating, and A is the cross-sectional area of ​​the component.

[0022] As a further aspect of the present invention, the specific steps for training the sensor network shear force prediction model are as follows:

[0023] From the fused feature vector set Z, continuous time segments are extracted in chronological order: for the shear force value F(t) at time t, the fused features {Z(tT), Z(t-T+1),..., Z(t-1)} of the previous T times are used as the input sequence to construct sample pairs (input sequence, F(t)), which are divided into training set, validation set and test set in a ratio of 7:2:1, where T is the time step;

[0024] For the input fusion feature sequence, Z-score normalization is performed using the mean and standard deviation of the training set; for the shear force value, min-max normalization is performed using the maximum and minimum values ​​of the training set.

[0025] A regression network consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer is constructed. The specific structural parameters are as follows: input dimension = temporal length T × fusion feature dimension d; the number of LSTM layers is set to 2, the number of hidden units in the first layer is 32, the number of hidden units in the second layer is 16, and the activation function is tanh for both; the number of neurons in the fully connected layer is 8, and the activation function is ReLU; the output dimension is 1.

[0026] The specific parameters for model compilation are as follows: the optimizer is Adam, the learning rate is set to 0.001, and the loss function is the mean squared error (MSE).

[0027] The parameters are updated iteratively using the training set. In each round, the MSE of the training set and the MSE of the validation set are calculated, and the loss curve is plotted.

[0028] Calculate MSE, mean absolute error (MAE), and coefficient of determination (R²) on the test set, and evaluate the model.

[0029] According to formula F pred =F pred,norm ×(F max -F min )+F min The predicted value F of the output layer pred,norm Perform inverse normalization, where F pred For the predicted true shear force value, F max F min To train the maximum and minimum concentrated shear force values.

[0030] As a further aspect of the present invention, the local anomaly concentration degree LAC is calculated as follows: (number of shear force anomalies within the group / total number of networks within the group) × average deviation of anomalies, where the average deviation of anomalies is the average deviation rate between the abnormal shear force within the group and the benchmark value.

[0031] The overall coordinated rate of change (GCR) is calculated as follows: (Number of networks with consistent shear force change trends / Total number of networks) × Mean trend correlation coefficient. Consistent trend means that the shear force values ​​change in the same direction over time. The trend correlation coefficient is calculated using the Pearson coefficient.

[0032] As a further aspect of the present invention, the specific operation for determining the prefabricated space frame status based on LAC and GCR is as follows:

[0033] If LAC≥LACth and GCR<GCRmin, then the prefabricated space frame is judged to have local defects.

[0034] If LAC < LACth and GCR ≥ GCRmax, then the prefabricated space frame is judged to have an overall defect.

[0035] If LAC≥LACth and GCR≥GCRmax, then the prefabricated space frame is judged to have mixed defects.

[0036] Otherwise, the prefabricated space frame is judged to be without defects;

[0037] Wherein, LACth is the threshold for local anomaly concentration, and GCRmin and GCRmax are the upper and lower limits of the overall coordinated change rate.

[0038] This invention provides an intelligent monitoring system for prefabricated space frame shear structures, which has the following advantages compared with the prior art:

[0039] (1) This invention integrates multiple types of sensors through a modular bracket, realizing the integrated deployment of fiber optic gratings, ultrasonic arrays, electromagnetic coils and energy components, ensuring the relative position stability of the sensors. At the same time, it adopts a self-powered design, eliminating the need for an external power supply, thereby improving the integration and engineering applicability of the monitoring system.

[0040] (2) The present invention uses an adaptive manifold projection fusion method to fuse multimodal features, effectively eliminating the mode gap between fiber strain, ultrasonic signal and electromagnetic impedance, preserving local geometric features of data, and improving the representation ability of fused feature vectors.

[0041] (3) This invention combines the LSTM model to dynamically predict the shear force value and uses DBSCAN clustering and LAC and GCR indicators to comprehensively identify the defect type, realizing the intelligent process from multimodal data to accurate defect identification, overcoming the problem of traditional methods relying on a single indicator and having low discrimination accuracy. Attached Figure Description

[0042] Figure 1 This is the system principle block diagram of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] like Figure 1 This invention provides an intelligent monitoring system for prefabricated space frame shear structures, comprising:

[0045] The sensor integration module has several sensor networks installed on a prefabricated grid frame for real-time acquisition of fiber optic strain data, ultrasonic signals and electromagnetic impedance data. The sensor network specifically includes fiber optic grating sensors, ultrasonic arrays, electromagnetic coils and energy components.

[0046] Single-mode fiber optic grating sensors can be selected as fiber optic sensors, which have the characteristics of anti-electromagnetic interference and corrosion resistance, and can directly capture the minute deformation of the space frame members under load.

[0047] An 8-channel ring ultrasonic transducer array is selected for the ultrasonic array. The array is evenly distributed along the inner wall of the node sleeve and is bound to the transmission link of the fiber optic grating sensor through shielded wires. It can emit high-frequency sound waves and accurately detect defects such as voids and cracks in the node grouting layer by utilizing the changes in the amplitude and propagation time of the echo signal, thus making up for the inability of fiber optic gratings to directly monitor internal defects of materials.

[0048] The electromagnetic coil is a miniature electromagnetic coil, which covers 2 / 3 of the total length of the sleeve. It is integrated with the ultrasonic components through a miniature circuit board. When the electromagnetic coil is subjected to alternating current, its impedance change is linearly related to the sleeve preload, which can monitor the bolt loosening in real time.

[0049] The energy component combines an L-shaped cantilever piezoelectric vibrator with an electromagnetic coil. At the point of most significant vibration, the L-shaped piezoelectric vibrator is fixed, and a permanent magnet is arranged next to the vibrator. The permanent magnet and the miniature coil at the end of the vibrator form an electromagnetic induction component. The output of both is rectified and filtered before being connected to the power supply interface of the sensor circuit.

[0050] The piezoelectric vibrator generates voltage at common vibration frequencies of the grid structure (such as when people walk or equipment runs), and the electromagnetic coil generates electricity through the relative motion between the permanent magnet and the vibrator. The combined energy density of the two can meet the continuous power supply requirements of fiber optic gratings, ultrasonic components, and electromagnetic modules without the need for an external power supply.

[0051] The modular bracket, made of ABS material, integrates the fiber optic sensor (arranged along the bracket axis), ultrasonic array (annular groove at the front of the bracket), electromagnetic coil (groove in the middle of the bracket), and energy component (slot at the rear of the bracket) into one unit, which is then fixed to the reserved installation position of the prefabricated rod with bolts.

[0052] Modular brackets ensure that the relative positions of each sensor are fixed, such as the distance between the ultrasonic array and the sleeve always remains standard, avoiding measurement deviations caused by displacement during prefabrication or hoisting. At the same time, ABS material is lightweight, has a wide temperature range, good compatibility with prefabricated components such as steel and concrete, and will not cause electrochemical corrosion.

[0053] The feature fusion module preprocesses fiber strain data, ultrasonic signals, and electromagnetic impedance data, extracts key features, and fuses these key features using an adaptive manifold projection fusion method to obtain a fused feature vector.

[0054] Denoising of fiber optic strain data was achieved by using wavelet thresholding. A db4 wavelet basis was selected, and the number of decomposition layers was 5. This eliminated interference from construction vibration and temperature drift, while retaining the effective strain signal.

[0055] The db4 wavelet has good time-frequency localization characteristics, making it suitable for processing non-stationary strain signals. It can effectively separate noise (high-frequency random components) from effective signals (low-frequency or specific frequency band strain information). Setting a 5-level decomposition can gradually strip away noise in different frequency bands while retaining effective strain components related to the structural state.

[0056] Fiber optic signals may lose data due to temporary sensor failures, etc. If the missing values ​​are retained directly, the continuity of the time series will be destroyed. For the very small percentage of missing data, linear interpolation is used to smoothly fill in the missing data based on the effective data before and after the missing point, which can ensure the continuity of the time series without significantly distorting the original strain trend.

[0057] Ultrasonic signals exhibit continuous fluctuations in the time domain, making them difficult to distinguish directly in time domain analysis. However, defect echoes and interference echoes differ in the frequency domain. Therefore, short-time Fourier transform (STFT) is used to convert the time domain signal into a time-frequency domain signal.

[0058] Using STFT to process time-domain signals can preserve the time positioning of the signal and reflect the frequency components of different echoes.

[0059] Adaptive thresholding (based on the maximum inter-class variance method) is applied to the time-frequency domain signal to retain the weak signal of the defect reflection;

[0060] Electromagnetic impedance measurement is susceptible to external electromagnetic interference, especially 50Hz power frequency interference, which often comes from power equipment and power line radiation. This type of interference will be superimposed on the actual impedance signal, causing fluctuations in the measured value.

[0061] Using a Butterworth low-pass filter to remove power frequency interference ensures that the frequency components of the effective signal pass through without distortion, while steeply attenuating the noise after the cutoff frequency.

[0062] Changes in ambient temperature can cause changes in coil resistance and conductivity of the measured structure, which in turn can cause impedance drift and lead to monitoring errors. Normalization by dividing by the initial reference value can cancel out the impedance changes caused by temperature and retain only the impedance changes related to the preload.

[0063] Key features extracted from preprocessed fiber strain data include peak strain, strain rate, and mean strain.

[0064] When a structure is under stress, the peak strain usually appears in the stress concentration area. Its magnitude directly reflects the local load-bearing capacity of the structure. When the peak strain exceeds the yield strength of the material, it may indicate that the structure is about to fail.

[0065] Strain rate reflects the dynamic response characteristics of a structure; abnormal strain rate may correspond to sudden load changes or structural stiffness degradation.

[0066] The mean strain reflects the overall stress state of the structure, and its variation may be related to temperature, long-term load accumulation, or overall deformation.

[0067] For the preprocessed ultrasonic signal, the key features extracted include echo amplitude attenuation rate, defect depth, and wave velocity change rate:

[0068] When ultrasound encounters a defect during propagation, its energy is scattered or absorbed, causing the echo amplitude to attenuate. The attenuation rate is positively correlated with the size and density of the defect.

[0069] Defect depth can be used to quantify the defect level. For example, defects on the surface of a structure may expand more quickly, while deep defects in the core area have a greater impact on load-bearing capacity.

[0070] Internal damage to a material can alter its elastic modulus, leading to changes in the propagation speed of ultrasonic waves.

[0071] Key features extracted from the preprocessed electromagnetic impedance data include the change in the real part of the impedance, the impedance phase angle, and the rate of change of the impedance magnitude.

[0072] The real part of the impedance is mainly related to the resistance and eddy current loss of the structure under test. Changes in preload will alter the contact state of the structure, thereby affecting the current path and resistance value.

[0073] The impedance phase angle reflects the phase difference between current and voltage, and is related to the inductive characteristics of the structure under test. Changes in structural stiffness will alter the magnetic field distribution, thus affecting the phase angle.

[0074] The impedance modulus comprehensively reflects the changes in the real and imaginary parts, and is more sensitive to the overall state of the structure.

[0075] The specific steps for obtaining the fused feature vector using the adaptive manifold projection fusion method are as follows:

[0076] Z-score normalization was performed on nine features of fiber strain (peak strain, strain rate, mean strain), ultrasonic signal (echo amplitude attenuation rate, defect depth, wave velocity change rate), and electromagnetic impedance (change in real part of impedance, impedance phase angle, change in impedance modulus).

[0077] For the three standardized subspaces ψ1, ψ2, and ψ3, local neighborhood graphs are constructed using the adaptive k-nearest neighbor algorithm:

[0078] According to the formula ρ i =∑ j exp(-||x i -x j || 2 Calculate the local density of the sample points. The k-value range is [5,8] for regions with high density and [2,4] for regions with low density.

[0079] Neighborhood graphs with a fixed k value cannot adapt to the uneven distribution of data. Adaptive k-nearest neighbors can dynamically match the local data density, ensuring that the neighborhood graph truly reflects the local geometry of the manifold.

[0080] According to the formula The elements of the neighborhood graph weight matrix W, where x i x j All are samples, σ i For sample x i The local bandwidth can be obtained through ρ i Regulation;

[0081] For each neighborhood graph, dimensionality reduction is performed using Local Linear Embedding (LLE):

[0082] Its objective function is: min∑ i ||x i -∑ j W ij 'x j || 2 The constraint condition is: ∑ j W ij =1;

[0083] LLE preserves the local topology of data through linear reconstruction, avoiding the loss of geometric information caused by linear dimensionality reduction methods such as traditional PCA.

[0084] The 3D single-modal features are reduced to d dimensions, where d ∈ [1,3] and is an integer, to obtain the low-dimensional manifold features of each modality.

[0085] Construct a cross-modal contrastive loss function to align low-dimensional manifolds of different modalities:

[0086] For samples of the same structural state, such as fiber optic strain, ultrasonic, and electromagnetic data under loosened bolt conditions, the distance between them in the common space is forced to be minimized: By optimizing the contrastive loss using gradient descent, the projection matrices P1, P2, and P3 from each mode to the common space are obtained, where... This represents the low-dimensional feature of the i-th sample in the fiber mode. This represents the low-dimensional feature of the j-th sample in the ultrasonic modality. This represents the low-dimensional feature of the r-th sample in the electromagnetic mode;

[0087] The physical meanings of features from different modalities differ greatly, creating a modal gap. Direct concatenation can lead to weak feature associations. Cross-modal alignment forces multimodal features of the same state to aggregate in a common space through contrastive loss, thereby strengthening the semantic associations between multimodalities.

[0088] Low-dimensional manifold features of each mode By mapping to the common space using a projection matrix, we obtain... Directly concatenate them into the final fused vector:

[0089] For every N new sets of data, the adaptive k-value and weight matrix W of the neighborhood graph are recalculated, and the projection matrices P1, P2, and P3 are updated using incremental gradient descent. The specific formula is as follows: Among them, P i new For the updated projection matrix, P i old Let η be the original projection matrix, η be the learning rate, and ζ be the cross-modal contrastive loss.

[0090] During long-term monitoring, data distribution may drift due to environmental changes, and static models may gradually become ineffective. Dynamically adjusting the manifold structure and projection matrix ensures that the fused vector always reflects the true physical state.

[0091] The shear force prediction module acquires historical data from each sensor network and constructs a shear force formula. It uses the fused feature vector as the input to the LSTM and the shear force as the output of the LSTM to train each sensor network and obtain the corresponding shear force prediction model.

[0092] To construct a quantitative shear force formula for associated defects or anomalies, it is necessary to combine the influence law of space frame material and defects on the mechanical properties of the structure, as well as the characteristic parameters of multimodal monitoring data, to establish a mathematical relationship between directly measurable physical quantities and shear force values ​​under defective conditions. The core logic includes the following three parts: (1) basic shear force under normal structural conditions (theoretical value based on shear strain); (2) shear force attenuation caused by microcracks (ultrasonic signal characteristic quantization); (3) shear force loss caused by sleeve loosening (electromagnetic impedance characteristic quantization).

[0093] The shear force is calculated using the formula V = V0 × (1 - α × R) × (1 - β × δ), where V is the actual shear force under defective or abnormal conditions; V0 is the basic shear force under normal structural conditions, which can be calculated from the theoretical value of shear strain; R is the crack reflection coefficient of the ultrasonic signal, used to quantify the degree of microcracks; δ is the relative change rate of electromagnetic impedance, used to quantify the degree of sleeve loosening; α is the influence coefficient of microcracks on shear force, usually calibrated from historical data; and β is the influence coefficient of sleeve loosening on shear force, usually calibrated from historical data.

[0094] Based on the relationship between shear strain and shear stress of the space frame material, and combined with the shear strain data γ measured by the fiber optic grating sensor, the specific formula is: V0=τ0×A=G×γ×A, where τ0 is the shear stress under normal conditions; G is the material shear modulus; γ is the shear strain measured by the fiber optic grating; and A is the cross-sectional area of ​​the component.

[0095] (1-α×R) in the formula is the microcrack effect term. Microcracks will weaken the shear bearing capacity of the structure. It is quantified by the crack reflection coefficient R of the ultrasonic signal, which represents the ratio of the reflected wave amplitude to the incident wave amplitude at the crack. The value range is [0,1]. The deeper / longer the crack, the larger R is.

[0096] The (1-β×δ) in the formula is the sleeve loosening effect term. Sleeve loosening will reduce the node stiffness and lead to abnormal shear force distribution. It is quantified by the relative change rate δ of electromagnetic impedance, which represents the impedance value under normal conditions and the impedance change under loose conditions. The more serious the loosening, the larger δ is.

[0097] The specific steps for training each sensor network to obtain the corresponding shear force prediction model are as follows:

[0098] From the fusion feature vector set Z, extract continuous time segments in chronological order: for the shear force value F(t) at time t, use the fusion features {Z(tT), Z(t-T+1),..., Z(t-1)} of the previous T times as the input sequence to construct sample pairs (input sequence, F(t)), where T is the time step;

[0099] The dataset is divided into training, validation, and test sets in a 7:2:1 ratio, while maintaining temporal continuity during the division.

[0100] For the input fusion feature sequence, Z-score normalization is performed using the mean and standard deviation of the training set; for the shear force value, min-max normalization is performed using the maximum and minimum values ​​of the training set.

[0101] The activation function of LSTM has a limited output range. If the input features have large scale differences, the large-scale features will dominate the update during gradient descent, and the gradients of small-scale features will be submerged. Normalization can balance the contributions of each feature.

[0102] Construct a regression network consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer. The specific structural parameters are as follows:

[0103] Input dimension = temporal length T × fusion feature dimension d;

[0104] A 2-layer LSTM, with 32 hidden units in the first layer and 16 hidden units in the second layer, and the activation function is set to tanh for both layers;

[0105] The network depth is increased by using two LSTM layers, with the number of hidden units decreasing from 32 to 16, gradually compressing temporal information and avoiding overfitting;

[0106] The fully connected layer has 8 neurons and uses ReLU as the activation function.

[0107] ReLU activation can filter out invalid feature combinations and improve the fitting ability to the laws of shear force attenuation caused by microcracks and shear force abrupt change caused by sleeve loosening.

[0108] The output dimension is 1;

[0109] The specific parameters for model compilation are:

[0110] The optimizer is Adam, the learning rate is set to 0.001, and the update step size is adjusted by the first moment and the second moment to solve the common gradient vanishing / exploding problem in LSTM training.

[0111] The loss function chosen is the mean squared error (MSE), which is sensitive to large errors in shear force prediction.

[0112] The parameters are updated iteratively using the training set. The MSE of the training set and the MSE of the validation set are calculated in each round, and the loss curve is plotted (in order to intuitively judge the model status). If the MSE of the validation set is consistently higher than the MSE of the training set, it indicates overfitting and the number of hidden units of the LSTM needs to be reduced. If both are high, the number of training rounds needs to be increased or the learning rate needs to be increased.

[0113] The MSE, mean absolute error (MAE), and coefficient of determination (R²) were calculated on the test set, and the model was evaluated: MSE reflects the overall error, MAE reflects the average deviation, and R² measures the explanatory power of the features for shear force variation.

[0114] According to formula F pred =F pred,norm ×(F max -F min )+F min The predicted value F of the output layer pred,norm Perform inverse normalization to map back to the true shear force value, where F pred For the predicted true shear force value, F max F min To train the maximum and minimum concentrated shear force values.

[0115] The space frame judgment module inputs the real-time fused feature vector of each sensor network into the corresponding shear force prediction model to obtain the shear force value. Based on the center coordinates of each sensor network, it automatically divides the nearest sensor network group using the DBSCAN algorithm and judges the condition of the prefabricated space frame according to the local anomaly concentration (LAC) and the overall cooperative rate of change (GCR).

[0116] The three-dimensional coordinates (x, y, z) of all sensor networks are extracted from the space frame BIM model. The DBSCAN algorithm is used to cluster them. A distance threshold and a minimum number of clusters are set, and several nearest neighbor network groups are automatically output. A single network can belong to multiple groups.

[0117] The core function of sensor networks is to monitor specific parts of the grid structure, such as critical nodes and weak stress areas. The installation position itself is designed around the critical node. For example, modular brackets are fixed to the reserved positions of prefabricated rods, and the reserved positions usually correspond to the critical nodes of the grid structure.

[0118] At this point, the coordinates of the key nodes are highly correlated with the actual spatial location of the sensor network. Using the coordinates of the key nodes to represent the location of the sensor network can directly reflect the spatial attributes of the monitored object, avoiding the problem that simply calculating the physical location of the sensor may be disconnected from the monitored target.

[0119] The Local Anomaly Concentration (LAC) is calculated as follows: (Number of shear force anomalies within a group / Total number of networks within a group) × Average deviation of outliers. This is used to quantify the proportion of anomalous shear force values ​​within a neighboring group. The average deviation of outliers is the average deviation rate between the anomalous shear force within a group and the baseline value. The higher the LAC, the more concentrated the shear force anomalies are in the local area.

[0120] The overall coordinated change rate (GCR) is calculated as follows: (Number of networks with consistent shear force change trends / Total number of networks) × Mean trend correlation coefficient. Consistent trend means that the shear force value changes in the same direction over time. The trend correlation coefficient is calculated using the Pearson coefficient. The higher the GCR, the more coordinated the shear force change of the entire network structure.

[0121] The specific steps for determining the condition of a prefabricated space frame based on LAC and GCR are as follows:

[0122] If LAC≥LACth and GCR<GCRmin, it indicates that there is an abnormal area within a certain range, and the monitoring data of the abnormal area has a weak correlation with other areas, indicating that the abnormality is isolated and does not affect the overall force transmission path. Therefore, it is judged that the prefabricated space frame has a local defect.

[0123] If LAC < LACth and GCR ≥ GCRmax, it indicates that the local anomaly range is small. The monitoring data of this small number of anomaly areas show a strong correlation with the monitoring data of other areas, indicating that the anomaly originates from a systemic problem of the overall structure. The local anomaly is just a manifestation of the overall problem. Therefore, it is judged that the prefabricated space frame has an overall defect.

[0124] If LAC≥LACth and GCR≥GCRmax, it indicates the existence of significant local anomaly regions. These local anomalies are strongly correlated with the overall region, indicating that local defects occur on the basis of overall structural performance deterioration. The overall problem exacerbates the damage to local weak points, thus it is judged that the prefabricated space frame has mixed defects.

[0125] Otherwise, it indicates that the local anomaly range is extremely small and has no global correlation, suggesting that the anomaly may be measurement noise or accidental disturbance, and has not formed a substantial defect. Therefore, it is judged that the prefabricated space frame is defect-free.

[0126] Among them, LACth is the threshold for local anomaly concentration, and GCRmin and GCRmax are the upper and lower limits of the overall coordinated change rate, which need to be set according to the actual situation.

[0127] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0128] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An intelligent monitoring system for prefabricated space frame shear structures, characterized in that, include: The sensor integration module has several sensor networks installed on a prefabricated grid frame for real-time acquisition of fiber optic strain data, ultrasonic signals and electromagnetic impedance data. The sensor network specifically includes fiber optic grating sensors, ultrasonic arrays, electromagnetic coils and energy components. The feature fusion module preprocesses fiber strain data, ultrasonic signals, and electromagnetic impedance data, extracts key features, and fuses these key features using an adaptive manifold projection fusion method to obtain a fused feature vector. The shear force prediction module acquires historical data for each sensor network and constructs a shear force formula. It uses the fused feature vector as the input to the LSTM and the shear force value as the output of the LSTM to train a shear force prediction model for each sensor network. The space frame judgment module obtains the real-time shear force value of each sensor network and uses the DBSCAN algorithm to automatically divide the nearest sensor network groups. It judges the condition of the prefabricated space frame based on the local anomaly concentration (LAC) and the overall cooperative rate of change (GCR).

2. The intelligent monitoring system for prefabricated space frame shear structures according to claim 1, characterized in that, A modular bracket is used to integrate the fiber optic grating sensor, ultrasonic array, electromagnetic coil, and energy components into one unit, which is then fixed to the pre-installed mounting position on the prefabricated rod with bolts.

3. The intelligent monitoring system for prefabricated space frame shear structures according to claim 1, characterized in that, For the preprocessed fiber strain data, the key features extracted include peak strain, strain rate, and average strain; for the preprocessed ultrasonic signal, the key features extracted include echo amplitude attenuation rate, defect depth, and wave velocity change rate. The key features extracted from the preprocessed electromagnetic impedance data include the change in the real part of the impedance, the impedance phase angle, and the rate of change of the impedance magnitude.

4. The intelligent monitoring system for prefabricated space frame shear structures according to claim 1, characterized in that, The specific steps for fusing key features using the adaptive manifold projection fusion method are as follows: Z-score normalization was performed on the key characteristics of fiber strain, ultrasonic signal and electromagnetic impedance respectively. Among them, fiber strain includes peak strain, strain rate and strain mean; ultrasonic signal includes echo amplitude attenuation rate, defect depth and wave velocity change rate; electromagnetic impedance includes change of real part of impedance, impedance phase angle and change of impedance modulus. For the three standardized subspaces ψ1, ψ2, and ψ3, local neighborhood graphs are constructed using the adaptive k-nearest neighbor algorithm, and dimensionality reduction is performed using Local Linear Embedding (LLE) to reduce the 3D single-modal features to d dimensions, thus obtaining the low-dimensional manifold features of each modality. d∈[1,3] and is an integer; Construct a cross-modal contrastive loss function, the specific formula of which is as follows: in, This represents the low-dimensional feature of the i-th sample in the fiber mode. This represents the low-dimensional feature of the j-th sample in the ultrasonic modality. This represents the low-dimensional feature of the r-th sample in the electromagnetic mode; By optimizing the contrastive loss through gradient descent, the projection matrices P1, P2, and P3 of each mode onto the common space are obtained. Low-dimensional manifold features of each mode By mapping to the common space using a projection matrix, we obtain... Directly concatenate them into the final fused vector:

5. The intelligent monitoring system for prefabricated space frame shear structures according to claim 4, characterized in that, In the process of constructing a local neighborhood graph using the adaptive k-nearest neighbor algorithm, according to the formula ρ i =∑ j exp(-||x i -x j || 2 Calculate the local density of the sample points. The k-value range is [5,8] for regions with high density and [2,4] for regions with low density.

6. The intelligent monitoring system for prefabricated space frame shear structures according to claim 1, characterized in that, The shear force value is calculated using the formula V = V0 × (1 - α × R) × (1 - β × δ), where V is the actual shear force value under defective or abnormal conditions, V0 is the foundation shear force under normal structural conditions, R is the crack reflection coefficient of the ultrasonic signal, δ is the relative change rate of electromagnetic impedance, α is the influence coefficient of microcracks on shear force, and β is the influence coefficient of sleeve loosening on shear force.

7. The intelligent monitoring system for prefabricated space frame shear structure according to claim 6, characterized in that, Based on the relationship between shear strain and shear stress of the space frame material, and combined with the shear strain data measured by the fiber optic grating sensor, the foundation shear force V0 under normal structural conditions is calculated according to the formula V0=G×γ×A, where G is the material shear modulus, γ is the shear strain measured by the fiber optic grating, and A is the cross-sectional area of ​​the component.

8. The intelligent monitoring system for prefabricated space frame shear structure according to claim 1, characterized in that, The specific steps for training the sensor network shear force prediction model are as follows: From the fused feature vector set Z, continuous time segments are extracted in chronological order: for the shear force value F(t) at time t, the fused features {Z(tT), Z(t-T+1),..., Z(t-1)} of the previous T times are used as the input sequence to construct sample pairs (input sequence, F(t)), which are divided into training set, validation set and test set in a ratio of 7:2:1, where T is the time step; For the input fusion feature sequence, Z-score normalization is performed using the mean and standard deviation of the training set; for the shear force value, min-max normalization is performed using the maximum and minimum values ​​of the training set. A regression network consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer is constructed. The specific structural parameters are as follows: input dimension = temporal length T × fusion feature dimension d; the number of LSTM layers is set to 2, the number of hidden units in the first layer is 32, the number of hidden units in the second layer is 16, and the activation function is tanh for both; the number of neurons in the fully connected layer is 8, and the activation function is ReLU; the output dimension is 1. The specific parameters for model compilation are as follows: the optimizer is Adam, the learning rate is set to 0.001, and the loss function is the mean squared error (MSE). The parameters are updated iteratively using the training set. In each round, the MSE of the training set and the MSE of the validation set are calculated, and the loss curve is plotted. Calculate MSE, mean absolute error (MAE), and coefficient of determination (R²) on the test set, and evaluate the model. According to formula F pred =F pred,norm ×(F max -F min )+F min The predicted value F of the output layer pred,norm Perform inverse normalization, where F pred For the predicted true shear force value, F max F min To train the maximum and minimum concentrated shear force values.

9. The intelligent monitoring system for prefabricated space frame shear structure according to claim 1, characterized in that, Calculate the local anomaly concentration degree LAC = (number of shear force anomalies within the group / total number of networks within the group) × average deviation of outliers, where the average deviation of outliers is the average deviation rate of abnormal shear forces within the group from the baseline value; The overall coordinated rate of change (GCR) is calculated as follows: (Number of networks with consistent shear force change trends / Total number of networks) × Mean trend correlation coefficient. Consistent trend means that the shear force values ​​change in the same direction over time. The trend correlation coefficient is calculated using the Pearson coefficient.

10. The intelligent monitoring system for prefabricated space frame shear structures according to claim 1, characterized in that, The specific steps for determining the condition of a prefabricated space frame based on LAC and GCR are as follows: If LAC≥LACth and GCR<GCRmin, then the prefabricated space frame is judged to have local defects. If LAC < LACth and GCR ≥ GCRmax, then the prefabricated space frame is judged to have an overall defect. If LAC≥LACth and GCR≥GCRmax, then the prefabricated space frame is judged to have mixed defects. Otherwise, the prefabricated space frame is judged to be without defects; Wherein, LACth is the threshold for local anomaly concentration, and GCRmin and GCRmax are the upper and lower limits of the overall coordinated change rate.

Images