A method and device for predicting stress of a fine stone concrete steel bar truss floor support plate

Abstract

The application discloses a fine stone concrete steel bar truss floor support plate stress prediction method and device, and relates to the field of building structure safety monitoring. Real-time collection of initial stress data, environmental parameters and load data of stress monitoring key points arranged at key stress positions of a floor support plate, construction of a to-be-predicted sample by using a sliding window method; adaptive normalization of the initial stress data in the to-be-predicted sample based on wavelet threshold denoising and a Gaussian mixture model dynamic segmentation, to obtain target stress data. The application pre-processes original monitoring data by wavelet threshold denoising, eliminates noise interference, and then uses a Gaussian mixture model to perform soft clustering segmentation on the time series data of each key point along the time axis, realizes adaptive local normalization, can effectively cope with the characteristics that the scale difference between multiple monitoring key points is significant and the time series data is non-stationary, and avoids data feature distortion or smoothing of local dynamic information caused by global normalization.

Description

Technical Field

[0001] This invention relates to the field of building structural safety monitoring, and more specifically, to a method and apparatus for predicting stress in fine aggregate concrete reinforced truss floor slabs. Background Technology

[0002] In the field of civil engineering, with the rapid development of construction technology towards industrialization and intelligence, fine aggregate concrete reinforced concrete truss floor slabs have been increasingly widely used in high-rise buildings and large-span structures due to their high overall stiffness, convenient construction, and excellent seismic performance. This type of composite floor slab combines the reinforced concrete truss with the fine aggregate concrete, bearing the self-weight of the wet concrete and construction loads during the construction phase, and participating in the overall load-bearing as part of the floor structure during the service phase. The stability and reliability of its load-bearing performance are directly related to the safety of the building structure.

[0003] Existing methods for predicting stress in bearing plates have several technical shortcomings. In the data preprocessing stage, global normalization tends to smooth out local abrupt changes, and fixed-time-window segmented normalization fails to capture irregular temporal variations in data distribution, both leading to data feature distortion and making it difficult to adapt to the significant scale differences in stress values ​​at multiple key points and the non-stationary nature of time-series data. In the feature extraction stage, single or fixed-scale convolutional kernels can only focus on specific time ranges and cannot simultaneously capture both short-term instantaneous impact and long-term continuous loading characteristics of stress data.

[0004] Meanwhile, existing stress prediction methods neglect the mechanical coupling relationship between different monitoring points and fail to consider the nonlinear modulation effect of environmental factors such as temperature, humidity, and load on stress evolution, making it impossible to model the dynamic spatial correlation between key points. Furthermore, most prediction models are purely data-driven, focusing solely on minimizing the error between predicted and actual values ​​during training. They lack consideration for fundamental laws of structural mechanics, and in situations where training data is not covered, they are prone to producing physically unreasonable and mechanically inconsistent predictions. This makes it difficult to guarantee the engineering applicability of the prediction results and fails to provide reliable support for the safety monitoring of floor decking structures. Summary of the Invention

[0005] In view of this, the present invention provides a method and apparatus for predicting stress in fine aggregate concrete reinforced truss floor slabs, which solves the problems of existing stress prediction methods, such as significant differences in scale of multiple key monitoring points, non-stationary time series, complex coupling of environmental factors, and lack of physical interpretability.

[0006] To achieve the above objectives, the following solution is proposed: A method for predicting stress in fine aggregate concrete reinforced truss floor slabs includes: Initial stress data, environmental parameters and load data of key stress monitoring points are collected in real time, and the sample to be predicted is constructed using the sliding window method. The target stress data is obtained by adaptively normalizing the initial stress data in the sample to be predicted based on wavelet threshold denoising and dynamic segmentation of Gaussian mixture model. The sample to be predicted is input into the pre-trained stress prediction model, which outputs the predicted stress values ​​of each monitoring key point at several future times.

[0007] Preferably, the process of adaptively normalizing the initial stress data in the sample to be predicted based on wavelet threshold denoising and Gaussian mixture model dynamic segmentation includes: The initial stress data is decomposed into multiple resolutions using discrete wavelet transform. The wavelet coefficients obtained from the decomposition are subjected to soft thresholding to suppress noise; The processed wavelet coefficients are reconstructed by inverse discrete wavelet transform to obtain denoised stress data; The denoised stress data is segmented into soft clusters along the time axis using a Gaussian mixture model. Based on the membership weight of each segment at each time point, the normalized parameters corresponding to each segment are weighted and fused to obtain the target stress data.

[0008] Preferably, the process of the stress prediction model predicting stress values ​​at several future times includes: Multi-scale convolutional feature extraction of target stress data is performed using parallel multi-scale one-dimensional convolutional layers; Based on the attention mechanism, the fusion weights are dynamically calculated according to the global information of the feature maps at each scale, and the feature maps at all scales are weighted and summed to generate the target stress features after multi-scale feature fusion. The environmental parameters and load data are normalized and feature-encoded, and an enhanced feature matrix is ​​generated by adaptive interaction between the nonlinear coupled gating unit and the target stress characteristics. Each monitoring key point is regarded as a graph node, and the enhanced features corresponding to each monitoring key point are used as the initial features of the node. The attention coefficient between nodes is dynamically calculated using a multi-head graph attention mechanism, and neighborhood information is aggregated through message passing to generate a spatiotemporal feature matrix. The decoder generates predicted stress values ​​for several future moments based on the spatiotemporal feature matrix.

[0009] Preferably, the process of multi-scale convolutional feature extraction of target stress data includes: The target stress data of each monitoring key point are input in parallel into several one-dimensional convolutional layers with different kernel sizes. Each convolutional layer extracts features at a specific time scale and introduces nonlinearity by modifying the linear unit activation function.

[0010] Preferably, the stress prediction model is trained using a hybrid loss function that integrates mean square error and physical constraints, and incorporates engineering index constraints on maximum tensile stress.

[0011] Preferably, after obtaining the predicted stress values ​​of each key monitoring point at several future times, the method further includes: Perform inverse normalization on the predicted stress values ​​and then display the results.

[0012] Preferably, it further includes: Determine whether the predicted stress values ​​at several future moments are greater than the allowable stress of the material; If the value is greater than the threshold, an early warning mechanism will be triggered, and a safety alarm will be issued.

[0013] Preferably, before inputting the sample to be predicted into the pre-trained stress prediction model, the method further includes: The samples to be predicted are labeled according to the construction stage or environmental conditions.

[0014] A stress prediction device for fine aggregate concrete reinforced truss floor slabs includes: The data acquisition module collects initial stress data, environmental parameters, and load data at key stress monitoring points in real time, and uses the sliding window method to construct the sample to be predicted. The data normalization module adaptively normalizes the initial stress data in the sample to be predicted by dynamically segmenting the data based on wavelet threshold denoising and Gaussian mixture model to obtain the target stress data. The stress prediction module takes the sample to be predicted and inputs it into a pre-trained stress prediction model, and outputs the predicted stress values ​​of each monitoring key point at several future times.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: 1. This invention preprocesses the original monitoring data by wavelet threshold denoising to remove noise interference. Then, it uses a Gaussian mixture model to perform soft clustering segmentation on the time series data of each key point along the time axis to achieve adaptive local normalization. This can effectively cope with the characteristics of significant scale differences between multiple monitoring key points and non-stationary time series data, and avoid data feature distortion or smoothing of local dynamic information caused by global normalization.

[0016] 2. The stress prediction model of this invention uses a parallel multi-scale convolutional feature extraction and attention fusion mechanism to extract short-term fluctuation and long-term trend features in the stress monitoring data in the time dimension by using multiple one-dimensional convolutional layers with different kernel sizes. The model also uses a lightweight attention mechanism to dynamically calculate the fusion weight based on the global information of the feature maps at each scale and performs a weighted summation of the feature maps at all scales. This allows the model to capture local and global features at different time scales simultaneously, avoiding the loss of short-term abrupt change information or the weakening of long-term trends caused by single or fixed-scale convolutional kernels.

[0017] 3. This invention proposes an environment-structure coupling feature enhancement and cross-keypoint graph attention spatiotemporal fusion. First, an enhanced feature matrix is ​​generated by adaptively interacting environmental variables with multi-scale fusion features through a nonlinear coupling gating unit, modeling the modulation effect of environmental factors on structural stress. Each monitored keypoint is treated as a graph node, and a multi-head graph attention mechanism is used to dynamically calculate the attention coefficients between nodes. Neighborhood information is aggregated through message passing to capture the dynamic spatial dependencies between keypoints as structural stiffness distribution and load transfer paths change. This invention effectively models the nonlinear modulation effect of external environmental factors on structural response and the dynamic spatial dependencies between keypoints, improving the prediction accuracy of the stress prediction model.

[0018] 4. This invention constructs a hybrid loss function that integrates mean square error, physical consistency terms, and engineering safety constraints. It introduces physical residuals based on the elasticity equilibrium equations as a penalty term, while also incorporating the maximum tensile stress engineering index constraint. The predicted stress is substituted into a simplified physical model to calculate the residuals, and penalties are applied to predictions exceeding the material's allowable stress limit. This loss function ensures that the prediction results not only conform to the data distribution but also satisfy the fundamental laws of structural mechanics, thus improving the accuracy of predicting key safety variables. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a method for predicting stress in fine aggregate concrete reinforced truss floor slabs, provided in an embodiment of the present invention; Figure 2 This is a comparison diagram of multi-scale convolutional feature extraction provided in an embodiment of the present invention; Figure 3 This is a comparison chart of the adaptive normalization effect provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of a stress prediction device for a fine aggregate concrete reinforced truss floor slab provided in an embodiment of the present invention. Detailed Implementation

[0021] 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.

[0022] First, combined Figure 1 The stress prediction method for fine aggregate concrete reinforced truss floor slabs provided in the embodiments of the present invention is introduced, such as... Figure 1 As shown, the method includes: Step S01: Collect initial stress data, environmental parameters and load data of key stress monitoring points in real time, and construct the sample to be predicted using the sliding window method.

[0023] Specifically, in the stress prediction model construction and training phase, the first step is to conduct systematic stress monitoring data acquisition. Several key stress monitoring points are deployed at critical stress locations in the floor decking. These key points are typically selected at locations with high steel reinforcement stress or abrupt structural geometric changes, such as mid-span, near supports, and truss joint areas. Each key monitoring point is equipped with a high-precision vibrating wire strain gauge or fiber optic grating sensor to collect the micro-strain of the steel reinforcement in real time and convert it into stress values. Simultaneously, to comprehensively capture external factors affecting stress evolution, temperature and humidity sensors, anemometers, and load sensors are deployed at each key monitoring point to record environmental parameters and load data such as ambient temperature, humidity, wind load, and construction live load. Data is continuously collected at a fixed sampling frequency (e.g., once per minute or once every ten seconds) to obtain long-term continuous time-series monitoring data.

[0024] After obtaining the raw monitoring data, the initial stress data, environmental parameters, and load data need to be preprocessed and labeled to generate training samples. A sliding window approach is used to construct the samples: a fixed-length time window is set as the input sequence length, and a prediction step size is defined. Initial stress data, environmental parameters, and load data at multiple consecutive times within the window are used as sample features, and stress monitoring values ​​at multiple future times after the window ends are used as the corresponding label (i.e., the true stress value). Each sliding window moves one time step to generate a sample, and all samples together constitute the training dataset.

[0025] Once the stress prediction model has been trained and its accuracy verified to meet requirements, it is applied to the stress prediction of actual fine-aggregate concrete reinforced truss floor slabs. In a real-time monitoring scenario, stress data at current and historical moments are continuously collected from each key monitoring point, while corresponding environmental parameters and load data such as ambient temperature, humidity, wind load, and construction live load are simultaneously collected. For each prediction moment, the historical stress sequence and environmental variable sequence within the most recent fixed-length time window are extracted to form a sample to be predicted.

[0026] Step S02: Adaptive normalization is performed on the initial stress data in the sample to be predicted based on wavelet threshold denoising and dynamic segmentation using a Gaussian mixture model.

[0027] Specifically, an adaptive normalization method based on wavelet threshold denoising and Gaussian mixture model dynamic segmentation is adopted. First, wavelet threshold denoising is used to remove noise from the monitoring data, providing a cleaner input for subsequent processing. Then, Gaussian mixture model is used to perform soft segmentation of the time series of each monitoring key point, and adaptive normalization is performed according to the statistical characteristics of each segment to eliminate scale differences and retain local time series features. The specific steps are as follows: (1) Wavelet thresholding for noise reduction For the initial stress data of each monitoring key point of each sample, multi-resolution decomposition is performed using discrete wavelet transform. The obtained wavelet coefficients are then subjected to soft thresholding to suppress noise. Finally, the denoised stress data is reconstructed using inverse discrete wavelet transform, as follows:

[0028] Where i represents the sample index, which is a positive integer and its value ranges from 1 to 2. , represents the total number of samples to be predicted. j represents the keypoint index, a positive integer with a value range of 1 to 2. Where J is the total number of monitored key points. t represents the time point index, which is a positive integer with a value range of... , where T is the time series length of each sample, i.e., the total number of time steps.

[0029] The enhanced stress value of the j-th monitoring key point of the i-th sample to be predicted at time t is the denoised stress data obtained after wavelet threshold denoising. This represents the initial stress data of the j-th monitoring key point of the i-th sample to be predicted at time t, which is directly acquired by the stress sensor; This represents the discrete wavelet transform function, which is used to decompose a one-dimensional time series signal into wavelet coefficients of multiple frequency sub-bands to achieve multi-resolution analysis and thus separate noise components in the signal. The soft threshold function is defined as follows: This is used to shrink wavelet coefficients, reducing the absolute value to a threshold value. Set the coefficient to zero, and set the absolute value to the threshold. The coefficient shrinks to zero to remove noise; The threshold parameter used in the soft thresholding function is calculated according to a general thresholding criterion and is expressed as follows: ,in This is an estimate of the noise standard deviation. This parameter represents the total number of time steps and controls the intensity of noise reduction. This represents the inverse discrete wavelet transform function, used to reconstruct the wavelet coefficients after soft thresholding into a time-domain signal, thus obtaining denoised stress data.

[0030] (2) Adaptive normalization based on Gaussian mixture model To avoid distorting the data distribution through global normalization, the denoised stress data at each monitoring key point is softly clustered along the time axis using a Gaussian mixture model. Each time point belongs to one of K Gaussian segments with a certain membership degree. During normalization, the normalization parameters corresponding to each segment are weighted and fused according to the membership degree weight of each time point to obtain the target stress data, achieving adaptive local normalization, as shown below:

[0031] in, represents the target stress data of the j-th monitoring key point of the i-th sample to be predicted at time t; K represents the total number of mixture components in the Gaussian mixture model, i.e. the number of Gaussian segments, which is dynamically determined by the Bayesian information criterion and is used to characterize the complexity of the time series data distribution. This represents the denoised stress data of all samples to be predicted within the k-th Gaussian segment for the j-th monitoring key point. The mean is a scalar, estimated together with the parameters of the Gaussian mixture model using the expectation-maximization algorithm; This represents the denoised stress data of all samples to be predicted within the k-th Gaussian segment for the j-th monitoring key point. The standard deviation is a scalar, estimated together with the parameters of the Gaussian mixture model using the expectation-maximization algorithm. This represents a very small positive integer, used to prevent the denominator from being zero and to ensure the numerical stability of normalized calculations. Examples of its values ​​are shown below. .

[0032] The membership weight of the j-th monitoring key point at time t belonging to the k-th Gaussian segment is a scalar, satisfying the following condition: The posterior probability, calculated from the Gaussian mixture model, is expressed as:

[0033] in, Let be the mixing coefficient of the k-th Gaussian segment of the j-th monitoring key point, which is a scalar and satisfies . , representing the prior weight of the segment in the time distribution, estimated by the expectation-maximization algorithm; The center point of the k-th Gaussian segment of the j-th monitoring key point on the time axis, i.e. the mean of the segment in the time dimension, is estimated by the expectation-maximization algorithm; Let be the covariance of the k-th Gaussian segment of the j-th monitoring key point on the time axis. Since time t is a one-dimensional scalar, In actual calculations, it degenerates into variance. This is used to characterize the span or range of the segment on the time axis, and is estimated using the expectation-maximization algorithm; Let k be the probability density function of a Gaussian distribution. k represents the Gaussian segment index, which is a positive integer ranging from 1 to K; k' represents the dummy variable of the Gaussian segment index, which is a positive integer ranging from 1 to K.

[0034] Step S03: Input the sample to be predicted into the pre-trained stress prediction model and output the predicted stress values ​​of each monitoring key point at several future times.

[0035] Specifically, the normalized samples to be predicted are sequentially input into the trained multi-scale one-dimensional convolutional layer, environment-structure coupling enhancement layer, graph attention layer, and decoder. The stress prediction model forward propagates to calculate the predicted stress values ​​of each monitoring key point at multiple future times (such as the next ten minutes or the next hour).

[0036] Multi-scale convolutional feature extraction is performed on the target stress data through parallel multi-scale one-dimensional convolutional layers; based on the attention mechanism, the fusion weights are dynamically calculated according to the global information of the feature maps at each scale, and the feature maps at all scales are weighted and summed to generate the target stress features after multi-scale feature fusion. The environment-structure coupling enhancement layer normalizes and encodes environmental parameters and load data, and adaptively interacts with the target stress features through nonlinear coupling gating units to generate an enhanced feature matrix. The graph attention layer treats each monitoring key point as a graph node, uses the enhanced features corresponding to each monitoring key point as the initial features of the node, dynamically calculates the attention coefficients between nodes using a multi-head graph attention mechanism, and aggregates neighborhood information through message passing to generate a spatiotemporal feature matrix. Based on the spatiotemporal feature matrix, the decoder generates predicted stress values ​​for several future time moments time by time.

[0037] The present invention provides a detailed description of the prediction process of the stress prediction model, as follows: (1) Multi-scale convolution feature extraction This invention employs parallel multi-scale one-dimensional convolutional layers to extract features from the normalized time-series data of each monitoring key point. By setting convolutional kernels of different sizes, it synchronously captures patterns in the data at different time scales. The specific steps are as follows: For the normalized stress time-series data of each monitoring key point, it is used as a one-dimensional input sequence and fed in parallel into M one-dimensional convolutional layers with different kernel sizes. Each convolutional layer extracts features at a specific time scale and introduces nonlinearity by modifying the linear unit activation function, expressed as:

[0038] in, This represents the feature map extracted from the j-th monitoring key point through the m-th scale convolutional layer, with the shape being... . Let be the time step of the output after a one-dimensional convolution operation. To ensure that the length of the input and output remains unchanged in the time dimension, the convolution operation employs a padding strategy, which is set as follows: . This represents the number of feature channels output by the m-th scale convolutional layer. This represents the normalized target stress data matrix for the j-th monitoring key point, with the shape as follows: The normalized target stress data sequence of all test samples at the j-th monitoring key point, output by step S02, the time-series enhancement and adaptive normalization step. Reorganization along the sample dimension or the sequence of a single sample. This represents the one-dimensional convolution operation corresponding to the m-th scale convolutional layer. This represents the bias vector of the m-th scale convolutional layer operation for the j-th monitoring keypoint. This vector is a trainable parameter with dimension O(m). The number of output feature channels is the same. This indicates a modified linear unit activation function, introducing nonlinear transformation capabilities into the neural network. This represents the weight matrix of the m-th scale convolutional layer operation for the j-th monitoring keypoint. This matrix consists of trainable parameters of shape [formula missing]. 1 represents the number of input channels; The kernel size for the m-th scale convolution operation controls the receptive field size of the kernel to capture features at different time scales; for example, it can be set to... , , , m represents the scale index, which is a positive integer ranging from 1 to M; M represents the total number of parallel convolution scales, determined by the network architecture design.

[0039] (2) Multi-scale feature fusion based on attention mechanism A weighted fusion strategy based on an attention mechanism is adopted to adaptively integrate multi-scale features to comprehensively represent the dynamic changes of stress data, avoiding feature dimensionality expansion and redundancy caused by simple splicing. A lightweight attention mechanism is employed, which dynamically calculates the fusion weights based on the global information of the feature maps at each scale and performs a weighted summation of the feature maps at all scales to generate the fused feature representation, as follows:

[0040] in, The multi-scale target stress characteristics of the j-th monitoring key point have the following shape: , , This represents the number of channels in the fused feature map, which is the sum of the number of channels in the feature maps at each scale. Let be the fusion weight of the m-th scale feature map of the j-th monitoring key point, which is a scalar.

[0041] The fusion weights are dynamically calculated using an attention mechanism and are used to measure the importance of features at that scale during fusion. The calculation method is expressed as follows:

[0042] in, This represents the Sigmoid activation function, which maps the linearly transformed weight scores to the (0,1) interval, making the fused weight values ​​easier to interpret and optimize. This is a global average pooling operation used to analyze the target stress feature map. In the time dimension Average it and compress it into one. A dimensional vector is used to capture global statistical information about the features at this scale. u is a learnable weight vector. is the transpose of the weight vector. c is a learnable bias scalar.

[0043] (3) Enhanced environment-structure coupling characteristics The stress variation of fine-aggregate concrete reinforced truss floor slabs is not only affected by historical stress states, but also closely related to external environmental factors and construction loads. To fully utilize this heterogeneous information, the environmental parameters and load data synchronously collected at each key monitoring point are first normalized and feature-encoded. Then, through adaptive interaction between nonlinear coupled gated units and multi-scale fusion features, an enhanced feature matrix is ​​generated to model the modulation effect of environmental factors on structural stress, expressed as:

[0044] in, Let be the enhanced feature matrix of the j-th monitoring key point, which is the feature matrix of the j-th monitoring key point after environment-structure coupling enhancement, with dimension . . Let denote the weight matrix of the first linear transformation, which is a trainable parameter with dimension . This is used to linearly transform environmental features to a feature space, the shape of which is the number of output feature channels multiplied by the total number of external features. This represents the environmental and load feature vector corresponding to the j-th monitoring key point. It has a dimension of L and is composed of L external monitoring variables (such as temperature, humidity, wind load, and construction live load) after normalization and embedding encoding. An example value is L=4. Let be the bias vector of the first linear transformation, and be trainable parameters with dimension . . This represents element-wise multiplication (Hadamard product), used to multiply gated signals with multi-scale fused features. This represents the hyperbolic tangent activation function, with an output range of . This allows environmental information to be incorporated into features in a nonlinear additive manner, resulting in an environmental bias vector. Let be the weight matrix of the second linear transformation, which are trainable parameters with dimension . , used to generate environment bias terms. Let be the bias vector of the second linear transformation, and be trainable parameters with dimension . . Let T' be an all-1 vector. This indicates that the environment bias vector is copied T' times, where T' represents the time step of the feature map, with an example value of T'=T=256.

[0045] (4) Spatiotemporal fusion of attention across key point graphs The stress evolution at key points of the fine-aggregate concrete reinforced truss floor slab exhibits spatial correlation, which varies with structural stiffness distribution and load transfer path. To capture this dynamic spatial dependence, each monitored key point is treated as a graph node, and its enhanced features are used as initial node features. A multi-head graph attention mechanism is employed to dynamically calculate the attention coefficients between nodes, and neighborhood information is aggregated through message passing to generate node features that integrate global spatial information, represented as follows:

[0046] in, This represents the spatiotemporal feature matrix of the j-th monitoring key point after graph attention fusion, with dimension 1. That is, its time step remains T', and its feature dimension is ; The feature dimension of the graph attention layer output is determined by the network structure design, and an example value is shown below. . This represents a vector concatenation operation, used to concatenate the outputs of H attention heads along the feature dimension. H represents the total number of attention heads, determined by the network architecture design; an example value is H=4. This represents the set of nodes adjacent to the j-th monitoring key point, predefined according to the structural topology (such as mechanical connection relationships). This represents the value transformation matrix of the h-th attention head, with dimension 1. , are trainable parameters. This represents the enhanced feature matrix of the p-th monitoring key point. This represents the normalized attention coefficient of the j-th node under the h-th attention head to its neighbor p, used to measure the importance of node p to node j.

[0047] Normalized attention coefficients satisfy The calculation method is expressed as follows:

[0048] Among them, LeakyReLU ( ) represents the modified linear unit activation function with leakage, and the negative slope is 0.2. exp( ) represents the natural exponential function. || ] indicates a vector concatenation operation. Let h represent the learnable weight vector of the h-th attention head, with dimension n. This is used to calculate attention scores. This represents the query transformation matrix for the h-th attention head, with dimension 1. , are trainable parameters. This represents the key transformation matrix of the h-th attention head, with dimension 1. , are trainable parameters. p is the index of the first neighbor node, q is the index of the second neighbor node, and all nodes in Ne(j) are traversed. This represents the enhanced feature matrix of the p-th monitoring key point.

[0049] (5) Multi-step stress prediction decoder Based on node features that integrate environmental factors and spatial correlations, stress values ​​at multiple future time points are predicted. Considering the temporal dependence of stress evolution and the continuous nature of the prediction target, a decoding network based on gated recurrent units and residual connections is adopted to generate prediction sequences time-by-time. An attention mechanism is introduced to dynamically select relevant information from historical features to improve long-term prediction accuracy, expressed as:

[0050] in, Let Δ represent the predicted stress value of the j-th monitoring key point of the i-th sample at the current time t, which is a scalar. Δ represents the prediction step size index, which ranges from 1, 2, ..., P, where P is the maximum prediction step size. This indicates that the gated loop unit in the decoder has the following hidden state: The input is the context vector from the previous time step. Features of integration with the current moment (i.e., take) At any moment of (dimensional vector). The input is the context vector from the previous time step, with dimension . . Let be the spatiotemporal feature matrix at the current moment. t represents the index of the current prediction start time, which must satisfy... , where T is the original time series data length, i.e. the total number of time steps.

[0051] Context vector This is obtained by weighted summation of the node features at all historical moments, and is represented as:

[0052] in, This represents the index of a historical moment, with values ​​ranging from 1, 2, ..., T', where T' is the time step of the feature map (set T'=T). The attention weights are determined by the current GRU hidden state. Characteristics of historical moments It is obtained through additive attention calculation. This represents the spatiotemporal feature matrix of the j-th monitoring key point at time τ. 3D eigenvectors.

[0053] The process of calculating attention weights is expressed as follows:

[0054] Where v is the first trainable parameter, with dimension . . The second trainable parameter has a dimension of . The third trainable parameter has a dimension of .

[0055] After constructing the stress prediction model, it needs to be trained end-to-end using the labeled training dataset. This embodiment of the invention describes the training process of the stress prediction model as follows: The collected and preprocessed sample data is divided into training, validation, and test sets, typically in chronological order to avoid future information leakage. Each training sample contains stress data from multiple monitoring key points within a historical time window, synchronized environmental parameters, load data, and corresponding future real stress values ​​as labels.

[0056] During training, samples are input into the stress prediction model, which then undergoes temporal enhancement and adaptive normalization, multi-scale convolutional feature extraction, environment-structure coupling feature enhancement, spatiotemporal fusion of cross-keypoint graph attention, and a decoder, ultimately outputting predicted stress values ​​for multiple future time points.

[0057] The stress prediction model calculates a hybrid loss function based on the difference between the output predicted stress value and the label's true stress value. This loss function is a weighted combination of a mean squared error term, a physical consistency term, and an engineering safety constraint term. Subsequently, the gradient of the loss function with respect to the trainable parameters of each layer is calculated using the backpropagation algorithm, and the model weights are updated using optimization algorithms such as the Adam optimizer or stochastic gradient descent to gradually reduce the loss function.

[0058] During training, each complete traversal of the training set is called a epoch. After each epoch, the model performance is evaluated on the validation set (e.g., calculating the root mean square error on the validation set) and the change in the loss value is monitored. When the performance metric on the validation set stops improving or begins to decline for several consecutive epochs, it indicates that the model may be overfitting, and training is stopped at this point. Furthermore, training will also terminate if the preset maximum number of epochs (e.g., 1000 epochs) is reached. Finally, the model parameters with the best performance on the validation set are retained as the trained stress prediction model for subsequent stress prediction tasks in practical engineering.

[0059] To ensure that the prediction results conform to both the data distribution and the fundamental laws of structural mechanics, a hybrid loss function is constructed that integrates mean square error and physical constraints. The physical constraints are based on the elastic equilibrium equations of fine-aggregate concrete reinforced truss floor slabs. The predicted stress is substituted into a simplified physical model, and the residuals are calculated as a penalty term. Simultaneously, the maximum tensile stress engineering index constraint is introduced to improve the prediction accuracy of key safety variables. The hybrid loss function is expressed as:

[0060]

[0061]

[0062]

[0063] Here, Loss represents the total loss function, which is used to optimize model parameters during training. This represents the mean square error term. Indicates a physical consistency term. This indicates engineering safety constraints. This represents the actual monitoring stress value of the j-th monitoring key point of the i-th sample at the Δ-th step after the current time t, which is directly acquired by the sensor. The weighting coefficients for the physical consistency term are hyperparameters. The weighting coefficients of the engineering safety constraints are hyperparameters. A represents a linear or nonlinear operator matrix established based on structural mechanics. Its dimension is related to the length of the prediction sequence and is used to map the predicted stress vector to the equivalent load. This matrix is ​​not a trainable parameter but is predetermined based on the mechanical model. Let P be the vector of predicted stress values ​​for the j-th monitoring key point of the i-th sample under test at P future times. This represents the external load vector, which is known based on actual working conditions or calculated from environmental characteristics, and has a dimension of P. It represents the square of the L2 norm, used to measure the magnitude of the residuals in the physical equations. This represents the maximum predicted stress value at all predicted times for all monitored key points, i.e., the maximum tensile stress of the reinforcing steel. This represents the upper limit of allowable stress for the material, determined based on the strength grade of the fine aggregate concrete and the reinforcing steel. It is a constant, and an example value is given below. . This indicates a positive function that incurs a penalty when the predicted maximum stress exceeds the allowable value, otherwise it is 0.

[0064] Furthermore, to facilitate the model's learning of stress patterns under different working conditions, the stress prediction method for fine-aggregate concrete reinforced truss floor slabs in this embodiment of the invention may also perform the following steps before inputting the sample to be predicted into the pre-trained stress prediction model in step S03: The samples to be predicted are labeled according to the construction stage or environmental conditions.

[0065] Specifically, samples are roughly classified and labeled according to the construction stage or environmental conditions. For example, the construction stage (such as concrete pouring period, curing period, normal use period) or environmental conditions (such as high temperature, low temperature, strong wind) at the time of sample collection can be labeled. Although this category information is not directly used as input for the stress prediction model, it can be used for subsequent stress prediction model evaluation or detailed analysis of specific scenarios.

[0066] To provide more timely guidance for taking appropriate measures on-site, after outputting the predicted stress values ​​of each monitoring key point at several future times in step S03, this embodiment of the invention can also perform the following steps: Perform inverse normalization on the predicted stress values ​​and then display the results; An early warning is issued when the predicted stress value exceeds the upper limit of the material's allowable stress.

[0067] Specifically, the system performs denormalization as needed to restore the stress value to a physical unit and displays it on the monitoring interface in real time. It can also trigger an early warning mechanism—if the predicted maximum tensile stress value exceeds the material's allowable stress limit, a safety alarm will be issued in a timely manner to guide the on-site implementation of appropriate measures.

[0068] Next, the feasibility of the provided method for predicting stress in fine aggregate concrete reinforced truss floor slabs will be verified in the embodiments of the present invention.

[0069] (1) Comparison of multi-scale convolution feature extraction Taking key monitoring points across the river as an example, this study compares the impact of different convolution kernel scale settings on the prediction of future stress values ​​to verify the effectiveness of the multi-scale parallel convolution and attention fusion mechanism. Figure 2 As shown, the horizontal axis represents the prediction step size (steps), and the vertical axis represents the stress value (unit: megapascals). The figure plots the actual stress (black dotted line) and the prediction results of three methods: the method with a single-scale convolution kernel size of 3 (dashed line), the method with a single-scale convolution kernel size of 7 (dashed line), and the multi-scale fusion method used in this invention (solid line). The actual stress curve shows small fluctuations within the prediction step size range, reflecting the complexity of stress evolution. The predicted stress value curve with a single-scale convolution kernel of 3 is generally low and slow to respond at local fluctuations; the predicted stress value with a single-scale convolution kernel of 7 is generally high and shows periodic deviations from the actual trend, indicating that fixed-scale convolution kernels are difficult to capture both short-term mutations and long-term dependencies simultaneously. In contrast, the multi-scale fusion prediction curve (solid line) provided by this invention is closest to the actual stress curve, with a high degree of agreement in both amplitude and fluctuation phase, indicating that parallel multi-scale convolution extracts features from different time scales and dynamically weights and fuses them through an attention mechanism, enabling the model to adaptively combine useful information.

[0070] (2) Comparison of adaptive normalization effects like Figure 3As shown, the stress time series data for three key monitoring points (mid-span, support, and node) are presented, with the horizontal axis representing time (in seconds) and the vertical axis representing stress value (in megapascals). Each subplot contains three curves: initial stress data (solid line), stress calculated back to the original scale after global normalization (red dashed line), and target stress data calculated back after adaptive normalization (blue dotted line). The initial stress data exhibits obvious non-stationary fluctuations, local shocks, and noise. Although the overall amplitude range of the curve after global normalization is compressed, the deviation from the original curve is large in local abrupt change areas because the global statistics smooth out local details. The target stress data curve calculated using the adaptive normalization method of this invention, while maintaining the overall trend, can more closely follow the local fluctuations of the original curve, proving that this invention effectively preserves the dynamic information in the time series through soft segmentation and weighted normalization on the time axis.

[0071] The stress prediction device for fine aggregate concrete reinforced truss floor slabs provided in the embodiments of the present invention will be described below. The stress prediction device for fine aggregate concrete reinforced truss floor slabs described below can be referred to in correspondence with the stress prediction method for fine aggregate concrete reinforced truss floor slabs described above.

[0072] First, combine Figure 4 This paper introduces a stress prediction device for fine aggregate concrete reinforced truss floor slabs, such as... Figure 4 As shown, the stress prediction device for fine aggregate concrete reinforced truss floor slabs may include: The data acquisition module 100 collects initial stress data, environmental parameters and load data of key stress monitoring points in real time, and uses the sliding window method to construct the sample to be predicted. The data normalization module 200 adaptively normalizes the initial stress data in the sample to be predicted by dynamically segmenting the data based on wavelet threshold denoising and Gaussian mixture model to obtain the target stress data. The stress prediction module 300 inputs the sample to be predicted into the pre-trained stress prediction model and outputs the predicted stress values ​​of each monitoring key point at several future times.

[0073] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0074] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0075] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting stress in fine aggregate concrete reinforced truss floor slabs, characterized in that, include: Initial stress data, environmental parameters and load data of key stress monitoring points are collected in real time, and the sample to be predicted is constructed using the sliding window method. The target stress data is obtained by adaptively normalizing the initial stress data in the sample to be predicted based on wavelet threshold denoising and dynamic segmentation of Gaussian mixture model. The sample to be predicted is input into the pre-trained stress prediction model, which outputs the predicted stress values ​​of each monitoring key point at several future times.

2. The method for predicting stress in fine aggregate concrete reinforced truss floor slabs according to claim 1, characterized in that, The process of adaptively normalizing the initial stress data in the sample to be predicted based on wavelet threshold denoising and Gaussian mixture model dynamic segmentation includes: The initial stress data is decomposed into multiple resolutions using discrete wavelet transform. The wavelet coefficients obtained from the decomposition are subjected to soft thresholding to suppress noise; The processed wavelet coefficients are reconstructed by inverse discrete wavelet transform to obtain denoised stress data; The denoised stress data is segmented into soft clusters along the time axis using a Gaussian mixture model. Based on the membership weight of each segment at each time point, the normalized parameters corresponding to each segment are weighted and fused to obtain the target stress data.

3. The method for predicting stress in fine-aggregate concrete reinforced truss floor slabs according to claim 1, characterized in that, The process by which a stress prediction model predicts stress values ​​at several future moments includes: Multi-scale convolutional feature extraction of target stress data is performed using parallel multi-scale one-dimensional convolutional layers; Based on the attention mechanism, the fusion weights are dynamically calculated according to the global information of the feature maps at each scale, and the feature maps at all scales are weighted and summed to generate the target stress features after multi-scale feature fusion. The environmental parameters and load data are normalized and feature-encoded, and an enhanced feature matrix is ​​generated by adaptive interaction between the nonlinear coupled gating unit and the target stress characteristics. Each monitoring key point is regarded as a graph node, and the enhanced features corresponding to each monitoring key point are used as the initial features of the node. The attention coefficient between nodes is dynamically calculated using a multi-head graph attention mechanism, and neighborhood information is aggregated through message passing to generate a spatiotemporal feature matrix. The decoder generates predicted stress values ​​for several future moments based on the spatiotemporal feature matrix.

4. The method for predicting stress in fine aggregate concrete reinforced truss floor slabs according to claim 3, characterized in that, The process of multi-scale convolutional feature extraction from target stress data includes: The target stress data of each monitoring key point are input in parallel into several one-dimensional convolutional layers with different kernel sizes. Each convolutional layer extracts features at a specific time scale and introduces nonlinearity by modifying the linear unit activation function.

5. The method for predicting stress in fine aggregate concrete reinforced truss floor slabs according to claim 1, characterized in that, The stress prediction model is trained using a hybrid loss function that integrates mean square error and physical constraints, and incorporates engineering index constraints on maximum tensile stress.

6. The method for predicting stress in fine-aggregate concrete reinforced truss floor slabs according to claim 1, characterized in that, After obtaining the predicted stress values ​​for each key monitoring point at several future times, the process also includes: Perform inverse normalization on the predicted stress values ​​and then display the results.

7. The method for predicting stress in fine aggregate concrete reinforced truss floor slabs according to claim 1, characterized in that, Also includes: Determine whether the predicted stress values ​​at several future moments are greater than the allowable stress of the material; If the value is greater than the threshold, an early warning mechanism will be triggered, and a safety alarm will be issued.

8. The method for predicting stress in fine-aggregate concrete reinforced truss floor slabs according to any one of claims 1-7, characterized in that, Before inputting the sample to be predicted into the pre-trained stress prediction model, the following steps are also included: The samples to be predicted are labeled according to the construction stage or environmental conditions.

9. A stress prediction device for fine aggregate concrete reinforced truss floor slabs, characterized in that, include: The data acquisition module collects initial stress data, environmental parameters, and load data at key stress monitoring points in real time, and uses the sliding window method to construct the sample to be predicted. The data normalization module adaptively normalizes the initial stress data in the sample to be predicted by dynamically segmenting the data based on wavelet threshold denoising and Gaussian mixture model to obtain the target stress data. The stress prediction module takes the sample to be predicted and inputs it into a pre-trained stress prediction model, and outputs the predicted stress values ​​of each monitoring key point at several future times.

Images