Dynamic monitoring method for construction process of floor support plate

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a method for dynamically monitoring the construction process of a floor support plate, which comprises the following steps of: acquiring stress data of the whole construction period of the floor support plate, and constructing a training data set with a construction stage category label; performing adaptive multi-scale morphological filtering processing on the stress data in the training data set; on the basis of a predefined discrete structural element scale set, processing the stress data by adopting two composite operations of performing closing operation after morphological opening operation and performing opening operation after the closing operation; constructing a construction monitoring model and finishing model training by using the training data set so as to dynamically monitor the construction process; wherein the construction monitoring model takes a deep convolutional neural network with adaptive deep convolution and cross-layer feature bridging as a trunk feature extractor, and determines the classification probability of a construction stage through an attention classification layer based on external feature guidance and a weighted loss function of mutation perception.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method for dynamic monitoring of the construction process of floor decking. Background Technology

[0002] During the construction of floor decking, activities such as vibration and loading induce complex stress responses in the decking. These stress signals contain crucial information reflecting construction quality and structural safety. Therefore, real-time monitoring and precise analysis of stress signals during construction, enabling effective identification of different construction stages, are essential for ensuring the quality of floor decking construction and mitigating structural safety risks.

[0003] Currently, technologies such as signal processing, feature engineering, and machine learning have been gradually introduced into the field of construction monitoring in an attempt to improve the level of intelligence in monitoring. However, existing monitoring methods still have many technical defects that need to be addressed in practical applications, making it difficult to meet the high-precision monitoring needs in complex construction environments.

[0004] In the signal preprocessing stage, conventional frequency domain filtering and wavelet threshold denoising methods are currently the mainstream signal denoising methods. However, when processing non-stationary stress signals generated during floor deck construction, these methods may over-smooth the local abrupt changes caused by construction activities such as vibration and loading, resulting in the loss of key engineering information.

[0005] In the feature extraction stage, conventional feature engineering methods or single-path autoencoders struggle to simultaneously capture both the macroscopic trend changes and microscopic fluctuation details of stress signals. The extracted features often suffer from limited discriminative power, which in turn affects classification performance.

[0006] In the deep feature extraction and pattern recognition stages, standard convolutional neural networks are often used for feature learning and stage recognition of stress signals. However, these networks use convolutional kernels of fixed size and lack effective cross-layer connection mechanisms, making it impossible to adapt to the multi-scale characteristics of construction stress signals, resulting in insufficient deep feature extraction.

[0007] More importantly, existing monitoring methods generally suffer from a disconnect between technical processes, treating signal processing, feature engineering, and pattern recognition as independent procedures without integrating prior knowledge of construction mechanics into the entire algorithm design process. This results in insufficient generalization ability and a high false alarm rate when facing complex on-site interference.

[0008] In summary, existing floor decking construction monitoring methods generally suffer from problems such as insufficient capture of key stress change characteristics, weak anti-interference ability, low stage identification accuracy, and poor generalization ability. These methods are insufficient to meet the actual construction process's demand for intelligent and adaptive monitoring, thus hindering the improvement of floor decking construction quality and structural safety assurance levels. Summary of the Invention

[0009] To address the aforementioned technical issues, this application provides a dynamic monitoring method for the construction process of floor decking. The method aims to achieve precise filtering and feature enhancement of stress signals throughout the entire construction cycle. By utilizing a neural network model embedded with prior engineering physics, it ultimately achieves automated, high-precision real-time identification and status tracking of key construction stages such as concrete pouring, vibration, curing, and loading. This provides reliable data support and decision-making basis for construction quality control and safety management.

[0010] The technical solution provided in this application is as follows: A method for dynamic monitoring of the construction process of floor decking includes: Collect stress data throughout the entire construction cycle of floor decking and construct a training dataset with construction stage category labels; Adaptive multi-scale morphological filtering is performed on the stress data in the training dataset, including: predefining a discrete set of structuring element scales; processing the stress data using two composite operations: morphological opening followed by closing and closing followed by opening; determining the dominant scale by analyzing the residual energy distribution of the stress data and the results of the two composite operations at each scale; calculating adaptive weights for the results of the two composite operations based on the dominant scale; and weighting and combining the results of the two composite operations according to the adaptive weights to obtain the filtered stress data. A construction monitoring model is constructed and trained using the training dataset to perform dynamic monitoring of the construction process. The construction monitoring model uses a deep convolutional neural network with adaptive deep convolution and cross-layer feature bridging as the backbone feature extractor, and determines the classification probability of the construction stage through an attention classification layer guided by external features and a weighted loss function for mutation awareness.

[0011] In one possible implementation, the collected stress data includes stress time-series signals, corresponding timestamps, and construction activity markers; The construction of the training dataset with construction stage category labels includes: Continuous stress data is cut into fixed-length segments or segments corresponding to the duration of construction activities. Based on the main construction activities within the time period of each data segment, each data segment is assigned a corresponding construction stage category label, including construction preparation stage, concrete pouring stage, curing and hardening stage, and loading test stage. The training dataset consists of data segments labeled with construction stage categories.

[0012] One possible implementation involves processing the stress data using two composite operations: morphological opening followed by closing, and closing followed by opening. An adaptive multi-scale morphological filtering framework is constructed, and the structuring element is used to perform opening-closing and closing-opening operations on the acquired raw stress data respectively. The two computational results are weighted and fused based on adaptive weights that satisfy complementary relationships to obtain filtered stress data that removes positive and negative impulse noise.

[0013] In one possible implementation, the method for determining the result of the opening / closing operation / the result of the closing / opening operation includes: Using a flat linear structural element under the dominant scale, opening-closing operations and closing-opening operations are performed on the original stress data to obtain the opening-closing operation results and the closing-opening operation results. The adaptive weights include adaptive weights corresponding to the opening and closing operation results and adaptive weights corresponding to the closing and opening operation results, both of which are determined based on the residual vector under the dominant scale. The residual vectors include the first residual vector between the original stress data and the opening and closing operation results under the dominant scale, and the second residual vector between the original stress data and the closing and opening operation results under the dominant scale.

[0014] In one possible implementation, the construction monitoring model further includes a dual-path feature extraction module, which is used to extract the global contour feature vector and local significant fluctuation feature vector of the filtered stress data, and fuse them to obtain the input of the adaptive depth convolution.

[0015] In one possible implementation, the dual-path feature extraction module includes a first path and a second path; The first path is used to adaptively segment the filtered stress data, calculate weighted features based on the linear fitting parameters of each segment, and concatenate them to obtain the global contour feature vector. The second path is used to extract overlapping subsequence fragments from the filtered stress data, encode the fragments using a pre-trained learnable sparse coding dictionary, calculate the mean and standard deviation of the coding coefficients, and integrate them through a fully connected layer to obtain the local significant fluctuation feature vector. The dual-path feature extraction module is used to concatenate the global contour feature vector with the local significant fluctuation feature vector, and then fuse them through a linear transformation layer to obtain an enhanced feature vector, which is used as the input of the adaptive depth convolution.

[0016] In one possible implementation, the improved deep convolutional neural network is configured with convolutional blocks, each of which includes adaptive deep convolution, batch normalization, activation function and cross-layer feature bridging; the adaptive deep convolution introduces a channel attention mechanism, which calculates channel attention weights through global average pooling and fully connected layers to modulate the contribution of each channel convolution. The cross-layer feature bridging adjusts the dimension of the shallow layer feature map through a one-dimensional 1×1 convolution with a stride of 2, and then adds it to the current layer feature map to achieve fusion.

[0017] In one possible implementation, the construction monitoring model processes the output feature map of the deep convolutional neural network through an attention classification layer guided by external features, fuses the statistical features of the filtered data, and outputs the classification probability through a fully connected layer and a Softmax function; the attention classification layer is configured with three types of transformation weight matrices: query, key, and value. The query transformation weight matrix is ​​used to map the external feature vector to the query vector; the external feature vector is determined based on the statistical feature vector calculated from the filtered stress data vector. The key transformation weight matrix and the value transformation weight matrix are used to map the feature vectors of each time position of the stress data into key vectors and value vectors, respectively; the attention calculation process is also associated with the dimensions of the query vector and the key vector, the weight matrix of the classification output layer, and the bias vector.

[0018] In one possible implementation, training the construction monitoring model further includes: An adaptive moment estimation optimization algorithm is used to iteratively update the model parameters, and the model training is completed using the total loss function; wherein, the total loss function is composed of the sum of the mutation-aware weighted cross-entropy loss and the attention regularization loss after being weighted by the regularization coefficient; The mutation-aware weighted cross-entropy loss is determined based on preset mutation weight coefficients, the average gradient magnitude of the filtered stress data vector, the construction stage category labels, and the elements of the predicted classification probability vector. The attention regularization loss is determined based on the attention weights corresponding to each time position of the stress data.

[0019] One possible implementation involves monitoring the construction process based on the construction monitoring model, including: The real-time filtered data is segmented using a sliding window method, input into the construction monitoring model to obtain the probability of each category, and the category with the highest probability is output as the identification result of the current construction stage. When an abnormal stress pattern or stage deviation is detected, an early warning signal is triggered. Compared with the prior art, the technical solution provided in this application has the following beneficial effects: This application employs an adaptive multi-scale morphological filtering method. By constructing multi-scale structuring elements and an adaptive weight selection mechanism, it resolves the contradiction in conventional filtering methods that smooth out key abrupt features during noise suppression. A dual-path feature extraction module is used, innovatively fusing global contour features approximated by adaptive segmentation and aggregation with local fluctuation features captured by a learnable sparse coding dictionary, overcoming the limitations of single-path feature extraction. An improved deep convolutional neural network is employed, innovatively introducing a channel attention mechanism to optimize deep convolution and designing a cross-layer feature bridging function to achieve adaptive capture and enhanced feature reuse across multiple scales. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for dynamic monitoring of the construction process of floor decking, provided in Embodiment 1 of this application.

[0021] Figure 2 This is a flowchart of a dynamic monitoring method for the construction process of floor decking provided in Embodiment 2 of this application.

[0022] Figure 3 This is a comparison chart showing the processing effects of different filtering methods provided in Embodiment 3 of this application on the stress monitoring data of floor decking construction.

[0023] Figure 4 This is a comparison chart of the performance indicators of different filtering methods provided in Embodiment 3 of this application.

[0024] Figure 5 This is a spatial distribution map of conventional manual features provided in Embodiment 4 of this application.

[0025] Figure 6 This is a PCA feature space distribution diagram provided in Embodiment 4 of this application.

[0026] Figure 7 This is a spatial distribution map of dual-path features provided in Embodiment 4 of this application.

[0027] Figure 8 This is a comparison chart showing the robustness of different methods provided in Embodiment 5 of this application under different noise levels.

[0028] Figure 9 A comparison chart showing the robustness of different methods provided in Embodiment 5 of this application in the case of missing data. Detailed Implementation

[0029] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0030] Example 1 See Figure 1 This is a flowchart of a dynamic monitoring method for the construction process of floor decking provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include: Step 101: Collect stress data throughout the entire construction cycle of the floor decking and construct a training dataset with construction stage category labels.

[0031] Step 102: Perform adaptive multi-scale morphological filtering on the stress data in the training dataset.

[0032] The aforementioned adaptive multi-scale morphological filtering process includes: predefining a discrete set of structuring element scales; processing the stress data using two composite operations: morphological opening followed by closing, and closing followed by opening. The dominant scale is determined by analyzing the residual energy distribution of the stress data and the results of the two composite operations at each scale. Adaptive weights are calculated based on these dominant scales, and the results are then weighted and combined according to these adaptive weights to obtain the filtered stress data.

[0033] Step 103: Construct a construction monitoring model and train it using the aforementioned training dataset to dynamically monitor the construction process. The construction monitoring model uses a deep convolutional neural network with adaptive deep convolution and cross-layer feature bridging as the backbone feature extractor, and determines the classification probability of each construction stage through an attention classification layer guided by external features and a weighted loss function for mutation awareness.

[0034] Compared with the prior art, the technical solution provided in this application has the following beneficial effects: This application employs an adaptive multi-scale morphological filtering method. By constructing multi-scale structural elements and an adaptive weight selection mechanism, it resolves the contradiction in conventional filtering methods that smooth out key construction mutation features while suppressing noise. An improved deep convolutional neural network is used, innovatively introducing a channel attention mechanism to optimize deep convolution and designing a cross-layer feature bridging function to achieve adaptive capture and enhanced feature reuse across multiple scales. An external feature-guided attention classification layer and a mutation-aware weighted loss function are employed, systematically integrating the physical prior knowledge (mutation characteristics) of stress signals into the neural network classification decision-making process for the first time. Ultimately, this achieves automated, high-precision real-time identification and status tracking of key construction stages such as concrete pouring, vibration, curing, and loading, thus providing reliable data support and decision-making basis for construction quality control and safety management.

[0035] Example 2 Embodiment 2 of this application is a further refinement or supplement to the technical solution of Embodiment 1, so as to fully disclose the technical solution of this application. See also Figure 2 This is a flowchart of a dynamic monitoring method for the construction process of floor decking provided in Embodiment 2 of this application. Figure 2 As shown, the specific implementation of the above method includes: Step 201: Collect stress data during the construction of the floor decking.

[0036] Data acquisition in this application is performed by pre-embedding or installing high-precision resistive or fiber optic stress sensors at multiple key stress-bearing locations in the floor deck. Optionally, these key stress-bearing locations include, but are not limited to, mid-span and near supports.

[0037] The sensor continuously collects stress data throughout the entire construction process at a set high-frequency sampling rate (such as 10Hz or 50Hz), thereby capturing the dynamic stress response at each stage, from rebar tying, concrete pouring, vibration, curing to subsequent loading tests. In addition to the stress time-series signal, the collected raw stress data also includes synchronously recorded timestamps and construction activity markers accurate to the minute provided by the on-site construction log.

[0038] Step 202: Cut the above stress data according to the preset cutting rules to obtain a series of data segments of fixed length or corresponding to the duration of construction activities.

[0039] The realization of dynamic monitoring of the floor decking construction process relies on a high-quality and accurately labeled training dataset. The core of constructing the training dataset lies in the segmentation and labeling of continuous stress signals. As one feasible approach, based on the construction log, the continuous stress data stream is cut into a series of data segments of fixed length or corresponding to the duration of construction activities. Each data segment is assigned a clear stage category label based on the main construction activities that occurred within its collection period.

[0040] In some embodiments, the annotation categories defined above specifically include: the construction preparation stage, the concrete pouring stage, the curing and hardening stage, and the loading test stage. During the construction preparation stage, load changes are small, and the signal is stable. During the concrete pouring stage, the signal shows a continuous increasing trend accompanied by short-term, sharp fluctuations caused by pouring and vibration. During the curing and hardening stage, the signal exhibits slow relaxation or stable changes. During the loading test stage, the signal shows significant, step-like abrupt changes corresponding to the loading procedure.

[0041] Ultimately, the constructed training dataset includes a large number of labeled stress data fragments and their corresponding category labels, which are used for subsequent model training and validation.

[0042] Furthermore, this application considers that the stress monitoring data during the construction of floor decking is a one-dimensional high-frequency sampled signal, which is subject to strong noise interference and contains local abrupt changes caused by construction activities. It also exhibits baseline drift. Conventional frequency domain filtering or wavelet thresholding denoising methods smooth out key local abrupt changes when suppressing noise, and have limited ability to handle non-stationary baseline drift.

[0043] To address the aforementioned technical problems, this application constructs multi-scale structural elements and designs an adaptive weight selection mechanism to achieve morphological filtering. This suppresses noise and baseline drift while preserving the characteristics of local stress abrupt changes generated at different construction stages to the greatest extent possible. The specific implementation process is as follows: Step 203: Construct an adaptive multi-scale morphological filtering framework to filter the stress data based on predefined multi-scale structural elements.

[0044] Specifically, this application employs a weighted combination of the results of morphological opening and closing operations as an adaptive multi-scale morphological filtering framework to process the original stress data. This framework effectively filters out positive and negative impulse noise and smooths the baseline. Its output is expressed as follows: Xfilt = ωoc . OC ( Xraw , SEs ) + ωco . CO ( Xraw , SEs ) In the formula, Xfilt This represents a filtered stress data vector that suppresses noise and baseline drift while preserving key local mutation features. ωoc Represents the result of opening and closing operations OC ( Xraw , SEsThe larger the value of the adaptive weight corresponding to the opening and closing operation, the greater the contribution of the opening and closing operation result to the final filtered output. OC ( , ) represents a composite operator that performs a closing operation immediately after a morphological opening operation. Xraw This represents the original stress monitoring data vector, which is directly acquired by the stress sensor.

[0045] SEs The scale is represented as s The structuring element is used to define the neighborhood template of the local shape of the probe signal in morphological operations. An example is a flat linear structuring element. s This represents the current scale parameter of the structuring element and controls the size of the neighborhood range for morphological operations. ωco Represents the result of the opening and closing operation. CO ( Xraw , SEs The larger the value of the adaptive weight corresponding to the closing-opening operation, the greater the contribution of the closing-opening operation result to the final filtered output. CO ( , ) represents a composite operator that performs an opening operation immediately after a morphological closing operation.

[0046] Based on the above formula, this application constructs an adaptive multi-scale morphological filtering framework, using structuring elements to perform opening-closing and closing-opening operations on the acquired raw stress data. The two operation results are then weighted and fused according to adaptive weights that satisfy complementary relationships to obtain a filtered stress data vector that removes positive and negative impulse noise.

[0047] To handle noise and features of varying widths, this application predefines a set of discrete scales. By analyzing the energy distribution of the residuals between the original signal and the morphological filtering results at multiple scales, the dominant filtering scale is dynamically determined. Based on the residual intensity of the dominant scale, adaptive weights are calculated for the results of the two composite operations. The calculation method is expressed as follows:

[0048]

[0049] In the formula, Indicates the dominant scale Below, the original stress monitoring data vector Xraw The result of opening and closing operations OC ( X raw SE sm ) The first residual vector between. Indicating at the dominant scale Below, the original stress monitoring data vector Xraw The result of opening and closing operations OC(X) raw SE sm ) The second residual vector between. m Represents the scale index, used to traverse a predefined set of multi-scale structuring element scales. S The value range is {1,2,...,} M} M Represents scale set S The total number of mesoscale values, with an example value of 5. This represents the index corresponding to the dominant scale, i.e., the importance score. When taking the maximum value, the corresponding m value. Indicates the dominant scale value. SE sm The scale value represents s m Flat linear structure elements. The L1 norm of a vector is represented. This represents a very small positive constant.

[0050] Indicates the first m Individual scale s m The average energy of the residuals from the closing-opening operation is expressed as:

[0051] In the formula, L Represents signal vector Xraw The length of the sample is the total number of sampling points. This represents the L2 norm of a vector.

[0052] The above importance score This value reflects the likelihood that the scale will be the dominant filtering scale; a larger value indicates that the scale is more important. It is expressed as:

[0053] In the formula, Indicates the first j Individual scale s j Below, the average energy of the residuals from opening and closing operations. Indicates the first j Individual scale s j Below, the average energy of the residuals from the closing-opening operation. exp(.) denotes the natural exponential function. This represents the sharpening factor, which controls the degree of concentration of importance score distribution. The larger the value, the more concentrated the score distribution is on a few scales. An example value is 0.5.

[0054] Noise is suppressed in the filtered stress data, but direct input into the neural network has the problems of high dimensionality and the feature representation has not been enhanced by human priors. Conventional feature engineering or single-path autoencoders are difficult to capture local detail patterns and global shape context information in the stress signal at the same time.

[0055] To address the aforementioned technical issues, this application constructs and trains a construction monitoring model to complete the subsequent dynamic monitoring of the construction process. The construction monitoring model employs a dual-path feature extraction module. One path approximates the global contour features of the signal through adaptive segmentation and aggregation, while the other path captures local significant fluctuation features through a learnable sparse coding dictionary. Finally, the two paths are fused to form an enhanced low-dimensional feature representation. The specific implementation process is as follows: Step 204: Adaptively segment the filtered stress data, calculate the weighted features based on the linear fitting parameters of each segment, and concatenate them to obtain the global contour feature vector.

[0056] Specifically, the filtered stress data is adaptively segmented, and a weighted feature is calculated based on the linear fitting parameters of each segment to characterize the overall shape and trend of the signal. This feature scalar is represented as:

[0057] In the formula, Indicates the first k Global contour feature scalar for each segment. k This represents a segmented index, with values ​​ranging from {1, 2, ...,}. Kg} Kg This represents the total number of segments obtained through adaptive segmentation. Indicates the first k The slope parameter obtained after performing least-squares linear fitting on segmented data represents the overall trend of change of the segmented data. Indicates the first k The intercept parameter is obtained by performing least-squares linear fitting on segmented data. Indicates all Kg Intercepts of each segment The arithmetic mean. Indicates all Kg Intercepts of each segment The standard deviation of . tanh(.) represents the hyperbolic tangent activation function, which maps the input to the interval (-1, 1). β This represents an adjustable scaling factor used to control the strength of the impact of the degree of trend change on the feature; an example value is 2.0.

[0058] All Kg The feature scalars of each segment are concatenated sequentially to obtain a dimension of... Kg The global contour feature vector is represented as:

[0059] In the formula, [.]T represents vector concatenation and transpose operations.

[0060] Step 205: Extract overlapping subsequence segments from the filtered stress data, encode the subsequence segments using a pre-trained learnable sparse coding dictionary, and calculate the mean and standard deviation of the coding coefficients to obtain the local significant fluctuation feature vector.

[0061] In this embodiment, the mean and standard deviation of the statistical coding coefficients are used as features to describe local fluctuations, and are ultimately integrated and represented as follows:

[0062] In the formula, This represents a local significant fluctuation feature vector with dimension . . This represents the corresponding bias vector, which is a trainable parameter. This represents the local feature vector of the p-th block, calculated as follows: Dimensions are 2N d .in, This represents the mean vector of the sparse coding coefficient vectors of all segments within the p-th block. Let represent the standard deviation vector of the sparse coding coefficient vectors of all segments within the p-th block. Let p represent the index of the block, with values ​​ranging from {1, 2, ...,}. K 1}. K 1 represents the total number of blocks into which the long sequence is evenly divided. This represents the weight matrix of the fully connected layer used for integration and dimensionality reduction, and consists of trainable parameters.

[0063] Step 206: Concatenate the global contour feature vector and the local significant fluctuation feature vector, and fuse them into a low-dimensional enhanced feature vector through a learnable linear transformation layer.

[0064] The aforementioned low-dimensional enhanced feature vector is represented as: Xfeat = Wf. [ Fglobal ; Flocal ]+ bf In the formula, Xfeat This indicates that the enhanced feature vector output after fusion will be used as the input to the subsequently improved deep neural network. WfThis represents the learnable feature fusion weight matrix, which consists of trainable parameters. bf This represents the learnable feature fusion bias vector, which is a trainable parameter. Fglobal ; Flocal ] indicates that the vector Fglobal and Flocal Perform splicing operations in the dimensional direction.

[0065] To address the local patterns and hierarchical abstract information contained in the enhanced feature vectors, conventional convolutional neural networks use fixed-size convolutional kernels, and the inter-layer feature transfer ignores the complementarity between features at different abstraction levels, making it difficult to adaptively capture multi-scale features and promote feature reuse.

[0066] To address the aforementioned technical problems, this application proposes constructing a deep convolutional neural network with adaptive depthwise convolution and cross-layer feature bridging as the backbone feature extractor. The output feature map of the previous layer... As input, features are extracted through adaptive depthwise convolution, and then added to features from shallower layers to achieve feature reuse and gradient flow enhancement. Furthermore, a channel attention mechanism optimizes the depthwise convolution, and a bridging function fuses cross-layer features to enhance information flow. The specific implementation process is as follows: Step 207: Use a deep convolutional neural network with adaptive deep convolution and cross-layer feature bridging as the backbone feature extractor, and determine the classification probability of the construction stage through an attention classification layer guided by external features and a weighted loss function with mutation awareness.

[0067] Specifically, the aforementioned deep convolutional neural network includes adaptive depthwise convolution, batch normalization, activation functions, and convolutional blocks that bridge cross-layer features. The output of the convolutional block is represented as follows:

[0068] In the formula, Indicates the first l The output feature maps of each convolutional block, and Z (0) =Xfeat . Indicates the first l The output feature map of layer 1 is used as the input of this convolutional block. Indicates the first l The output feature map of layer 1-2 is used to combine with the features of the current layer to achieve feature fusion. l Indicates the index of the convolution block. Conv 1 Dad (.) indicates an adaptive depthwise convolution operation. Indicates the first l Learnable kernel parameters for layer-adaptive depthwise convolution. Indicates the first lLearnable bias parameters for layer-adaptive depthwise convolution. Indicates the first l The layer-to-layer feature bridging function. BN() represents the batch normalization operation.

[0069] Furthermore, this application introduces a lightweight channel attention mechanism into adaptive depthwise convolution, modulating the contribution of each input channel in the convolution by calculating the attention weights of each input channel. Based on these contributions, the network can adaptively emphasize or suppress specific channels, and the calculation method of its output scalar value is expressed as follows:

[0070] In the formula, Conv 1 Dad ( .) c,t This indicates that the adaptive depthwise convolution output feature map is in the channel c Time and location t The scalar value at that location. Indicates the first l The channel attention weight vector of the layer is calculated as follows: , dimension .in, The channel descriptor vector obtained by global average pooling is represented as: , dimension . This represents a global average pooling operation along the channel dimension. Indicates the first l Feature map of layer -1 The number of channels. Indicates the first l In layer-adaptive depthwise convolution, the weight matrix of the first fully connected layer in the channel attention mechanism is a trainable parameter. Indicates the first l In layer-adaptive depthwise convolution, the weight matrix of the second fully connected layer in the channel attention mechanism is a trainable parameter. This represents the modified linear unit activation function. This represents the Sigmoid activation function.

[0071] Indicates the first l Layer channel attention weight vector The first in c Each element. Indicates the first l Layer depth convolution kernel in channels c , kernel offset τ The parameter value at that location. τ This represents the time offset index within the convolution kernel. Represents the input feature map In the passage c Time and location t + τ The scalar value at that location.

[0072] Furthermore, based on the aforementioned cross-layer feature bridging structure, this application employs a feature transformation function to downsample and adjust the number of channels of the shallow layer feature map through convolution operations, making it match the feature map of the current layer in both time and channel dimensions. Then, these features are summed to achieve cross-layer feature fusion. This function is expressed as:

[0073] In the formula, B ( l (.) indicates the first (.) l Layer-to-layer feature bridging function. This represents a one-dimensional 1×1 convolution operation with a stride of 2, used to adjust the time length and number of channels of the feature map.

[0074] The improved deep convolutional neural network output feature map is processed by an attention classification layer guided by external features, the statistical features of the filtered data are fused, and the classification probability is output through a fully connected layer and a Softmax function.

[0075] In this embodiment, it is considered that the feature maps extracted by deep convolutional neural networks contain multi-scale dynamic information of the construction process, but directly using them for classification has problems such as unoptimized mapping relationship between features and categories and loss function not incorporating physical priors of the construction process. Conventional classifiers and cross-entropy loss are difficult to utilize the abrupt change characteristics and attention focus of stress signals.

[0076] To address this technical problem, this application designs an attention classification layer with three types of transformation weight matrices (query, key, and value) guided by external features, and a weighted loss function for mutation awareness, thereby achieving classification probability output and loss calculation.

[0077] Specifically, this application employs an external feature-guided attention mechanism to aggregate key time-point information from the output feature map of a convolutional neural network and fuse it with external features extracted from filtered stress data. The classification probability is output through a fully connected layer and a Softmax function. P The calculation method is expressed as follows:

[0078] P = softmax(S) In the formula, S This represents the original classification score vector, with dimension 1. K It is used for Softmax transformation to obtain the predicted classification probability. Wo This represents the weight matrix of the classification output layer, with dimension 1. K×(dv+dg), are trainable parameters. This indicates a vector concatenation operation. WQ This represents the query transformation weight matrix, with dimensions [missing information]. dk×dg Used to transfer external features G The mapping is to a query vector, which are trainable parameters. WK This represents the key transformation weight matrix, with dimension 1. dk×D, Used to transform feature vectors zt The mapping is to a key vector, which are trainable parameters. WV The value transformation weight matrix has dimensions of . dv×D , used to transform feature vectors zt The mapping is to a value vector, which consists of trainable parameters. Among them, dk This indicates the dimension of the query and key vectors, used to scale the dot product; an example value is 64. dv This indicates the dimension of the value vector, with an example value of 128. D The sparse coding dictionary matrix is ​​obtained through pre-training and has dimensions of . d×Nd. d This indicates the length of the atoms in the dictionary, that is, the length of each subsequence segment used for encoding, with an example value of 64. Nd This represents the total number of atoms in the dictionary, with an example value of 128.

[0079] zt This represents the improved deep convolutional neural network. l Output feature map of the layer Z ( L ) in the t The feature vector at each time location has a dimension of . D ,and Z ( L The dimension of ) is D × T . This represents the improved deep convolutional neural network. l Output feature map of the layer Z ( L ) in the Feature vectors at each time location T Representation of feature map Z ( L The duration of time, i.e., the total number of time points. G Represents the filtered stress data vector Xfilt The extracted external feature vector has a dimension of . dg , used to introduce prior information about the construction process, denoted as G = ReLU(Wg) U+bg) .in, Wg This represents the weight matrix of the fully connected layer used for external feature extraction. bg This represents the corresponding bias vector.U Represents the filtered stress data vector Xfilt The calculated statistical feature vector has a dimension of du This includes the mean, standard deviation, skewness, kurtosis, and average gradient magnitude, expressed as... .in, express Xfilt The arithmetic mean, express Xfilt The standard deviation. express Xfilt The skewness is expressed as: .

[0080] express Xfilt The kurtosis is expressed as .

[0081] express Xfilt The average gradient magnitude, used to quantify the degree of signal abrupt change, is expressed as:

[0082] In the formula, express Xfilt The length of the sample is the total number of sampling points. express Xfilt The sampling point index, with a value range of {1} , 2 ,…,L} bo This represents the bias vector of the classification output layer. P This represents the predicted classification probability vector.

[0083] Step 208: The total loss function is constructed by using mutation-aware weighted cross-entropy loss and attention regularization loss, and the model parameters are iteratively updated using an adaptive moment estimation optimization algorithm.

[0084] The aforementioned mutation-aware weighted cross-entropy loss is determined based on preset mutation weight coefficients, the average gradient magnitude of the filtered stress data vector, the construction stage category labels, and elements of the predicted classification probability vector. The aforementioned attention regularization loss is determined based on the attention weights corresponding to each time position of the stress data.

[0085] Specifically, this application, considering the impact of stress mutations during construction on classification criticality, employs mutation-aware weighted cross-entropy loss. The loss weights are dynamically adjusted based on the average gradient magnitude of the filtered stress data to enhance the model's sensitivity to mutation features. The calculation method is expressed as follows:

[0086] In the formula, Lclass This represents the classification loss, which is a scalar. The smaller the value, the more accurate the classification prediction. η This represents the mutation weighting coefficient, used to control the strength of the impact of mutation degree on loss. An example value is 0.1. Represents the filtered stress data vector Xfilt The average gradient magnitude is used to quantify the degree of signal abrupt change. K This indicates the total number of categories. Yk The first element representing the actual construction stage category label vector. k Each element is one-hot encoded; if the true category is... k but Yk =1 if the value is equal to 1, otherwise 0. P represents the predicted classification probability vector. Pk Indicates belonging to the first k The probability of each category.

[0087] Furthermore, to ensure that the attention mechanism focuses on key construction time points and avoids weight dispersion, attention entropy regularization loss is adopted to minimize the information entropy of attention weights, thereby promoting the interpretability of model decisions. The calculation method is expressed as follows:

[0088] In the formula, This represents the attention regularization loss; it is a scalar value, and the smaller the value, the more focused the attention. T Indicates the duration of time. This indicates that the attention weights correspond to the time positions. t The weight value is expressed as:

[0089] The total loss function, based on the weighted cross-entropy loss and attention regularization loss described above, is expressed as follows: Ltotal = Lclass+γLattn In the formula, γ This is the regularization coefficient, used to balance the classification loss and the attention regularization loss. An example value is 0.01.

[0090] In this embodiment of the application, after the network structure and loss function are defined as described above, the model enters the iterative training and parameter update stage. The constructed training dataset is divided into a training set and a validation set for model learning and performance evaluation. The training process is based on forward propagation and backward propagation.

[0091] In each iteration, a small batch of data is sampled from the training set, including filtered stress data and its actual construction stage category labels. The data first passes through a dual-path feature extraction module to obtain enhanced feature vectors, then is input into an improved deep convolutional neural network module for deep feature abstraction, and finally the predicted probability is obtained through the classification result output layer. Based on the predicted probability and the actual label, combined with the attention weights guided by external features, the total loss function value is calculated. This total loss integrates the mutation-aware classification loss and the attention regularization loss.

[0092] Then, using the backpropagation algorithm, the gradient of the total loss with respect to all trainable parameters in the model (including the weights of the fusion layer in the dual-path feature extraction module, the parameters of the convolutional kernel and attention mechanism of the convolutional neural network, the weights of the classification layer, etc.) is calculated.

[0093] During training, an adaptive moment estimation optimization algorithm is employed to dynamically update all parameters based on the calculated gradients, minimizing the total loss function. The training process continues through multiple iterations, and the model performance is evaluated using an independent validation set after each iteration, monitoring changes in the validation set loss.

[0094] The stopping condition for model iterative training is determined based on a pre-set strategy. Usually, the training process terminates when the training reaches the preset maximum number of iterations, or when the loss on the validation set no longer decreases within several consecutive iterations (i.e., the early stopping strategy is triggered). At this time, the model parameters with the best performance on the validation set are saved as the final monitoring model after training is completed.

[0095] Once the model is trained, it can be deployed for actual dynamic monitoring during the construction process of floor decking. In actual construction scenarios, stress sensors continuously collect real-time stress data streams from the floor decking.

[0096] The monitoring system first performs adaptive multi-scale morphological filtering on the raw stress data stream acquired in real time, identical to the processing during the training phase, to suppress noise and baseline drift while preserving key stress abrupt change features, resulting in clean filtered stress data. Then, using a sliding window approach, the continuous filtered data is segmented into segments of the same length as the training data, and these segments are sequentially input into the pre-trained dynamic monitoring model.

[0097] The model performs a series of processing steps on the data segment. A dual-path feature extraction module obtains the fused features of its global contour and local fluctuations. An improved deep convolutional neural network module further mines deep abstract features; finally, an attention classification layer guided by external features outputs the probability that the construction process within this time period belongs to each preset category.

[0098] The system outputs the category with the highest probability as the construction stage identification result for the current monitoring time slice. By continuously processing the data stream and outputting the identification results in real time, the system can dynamically depict the evolution trajectory of the construction progress and realize automated and intelligent stage identification and status tracking of the entire process of floor decking construction (from concrete pouring to hardening and bearing).

[0099] In addition, when the construction monitoring model identifies abnormal stress patterns or stages that deviate significantly from the expected construction plan, it can trigger early warning signals, providing critical decision support for construction safety management.

[0100] Example 3 Embodiment 3 of this application compares the effects of different filtering methods on stress monitoring data processing in floor decking construction to evaluate the performance differences between the adaptive multi-scale morphological filtering method proposed in this application and conventional filtering techniques in terms of noise suppression, baseline drift correction, and key feature preservation.

[0101] The experiment compares four filtering techniques: moving average filtering, as the most basic linear smoothing technique, smooths the signal by calculating the mean through a sliding window; median filtering, as a representative of nonlinear filtering, replaces the current value with the neighborhood median to suppress impulse noise; wavelet thresholding, as an advanced method of frequency domain filtering, separates different frequency bands through wavelet transform and performs thresholding; and the adaptive multi-scale morphological filtering method proposed in this application.

[0102] In terms of experimental setup, stress monitoring data from the entire construction process of the floor decking was used, with a time span of 100 seconds, a sampling frequency of 10 Hz, and a total of 1000 data points. The data simulated the stress response characteristics of three typical stages: construction preparation, concrete pouring, and loading test. The concrete pouring stage simulated a continuous upward trend and short-term, drastic fluctuations caused by vibration, while the loading test stage simulated a step-like abrupt change. Gaussian white noise was added to the raw data to simulate random measurement errors, impulse noise to simulate occasional sensor anomalies, and slowly changing baseline drift to simulate systematic errors such as sensor temperature drift.

[0103] See Figure 3 , Figure 3The line graph clearly illustrates the filtering effects of different methods. The horizontal axis represents time in seconds; the vertical axis represents stress value in megapascals. The original monitoring data line shows obvious noise fluctuations and baseline drift, while the true stress signal line shows the ideal variation trend at each stage of construction. Although the moving average filter line smooths the noise, it lags significantly in abrupt change areas and over-smooths the vibration fluctuation characteristics; the median filter line effectively suppresses impulse noise, but its suppression effect on Gaussian noise is limited; the wavelet threshold denoising line is relatively smooth overall, but it loses some vibration fluctuation details during the concrete pouring stage; the method in this application, while maintaining overall smoothness, better preserves the abrupt change characteristics of the original signal in the pouring stage (marked by blue shading) and the loading test stage (marked by green shading).

[0104] In this embodiment, a bar chart is used to display the comparison of quantification performance. For example... Figure 4 As shown in the figure, the horizontal axis represents four filtering methods, and the vertical axis represents performance index values. The left bar represents the improvement in signal-to-noise ratio, expressed in decibels; a higher value indicates better noise suppression. The right bar represents the feature retention rate, expressed as a percentage; a higher value indicates more complete retention of key construction features. The figure shows that the method in this application significantly outperforms the comparative methods in both metrics, particularly in feature retention. This indicates that the method in this application, while suppressing noise, can better preserve key stress change features during construction, providing a higher-quality data foundation for subsequent stage identification.

[0105] Example 4 Embodiment 4 of this application analyzes the effectiveness of different feature extraction methods from the perspective of feature space distribution, and the results are as follows: Figures 5-7 As shown in the figure. The purpose of the experiment is to demonstrate the separability of features extracted by each method at different construction stages through visualization methods, and to evaluate the feature extraction module's ability to enhance inter-class differences.

[0106] The experiment compares three feature extraction strategies: conventional manual feature extraction directly calculates signal statistics; principal component analysis (PCA) auto-coding feature extraction obtains a compact representation through linear dimensionality reduction; and this application's dual-path feature extraction integrates global contour and local fluctuation information. The experiment uses simulated data from four typical construction stages: construction preparation, concrete pouring, vibration operation, and loading test. Six core feature dimensions are extracted from each sample: mean reflects the average signal level, variance reflects the degree of fluctuation, skewness reflects distribution symmetry, kurtosis reflects distribution sharpness, energy reflects signal strength, and abrupt change reflects the degree of change. These features constitute a six-dimensional feature space, visualized through a two-dimensional scatter plot. The scatter plot selects the mean and abrupt change features as the two most representative dimensions, with the horizontal axis representing the mean and the vertical axis representing the abrupt change. In the conventional manual feature extraction plot, the sample points from the four stages largely overlap, forming a mixed cloud. In the PCA auto-coding feature extraction plot, the point cloud begins to differentiate but still has blurred boundaries. In the method diagram of this application, the point clouds of the four stages form relatively independent clusters, indicating that the dual-path feature extraction of this application can effectively enhance the differences between classes and provide the classifier with more easily distinguishable feature representations.

[0107] Example 5 The robustness and stability of different methods in real-world applications were evaluated. The experiments aimed to test the performance retention capabilities of each method under noise interference and data loss conditions, as well as the consistency of results across multiple experiments. Four methods were compared: conventional statistical features combined with support vector machines, one-dimensional convolutional neural networks, long short-term memory networks, and the method proposed in this application. Two robustness test scenarios were designed for the experimental configuration. The first was a noise robustness test, in which different intensities of noise were added to the test data, increasing from 5% to 25%, simulating varying degrees of interference in actual monitoring. The second was a data loss robustness test, in which different proportions of data points were randomly deleted from the test data, increasing from 0% to 20%, simulating occasional sensor failure or data transmission loss. Each condition was repeated multiple times to evaluate stability.

[0108] like Figure 8 The diagram illustrates the comparison of noise robustness. The horizontal axis represents five noise levels, and the vertical axis represents classification accuracy as a percentage. The error bar represents the standard deviation across multiple experiments. As the noise level increases, the accuracy of all methods decreases, but the method described in this application decreases the slowest, maintaining high accuracy even at high noise levels. Furthermore, the shortest error bar indicates the most stable results. The method combining conventional statistical features with support vector machines is the most sensitive to noise, exhibiting the fastest performance degradation and the greatest fluctuation.

[0109] like Figure 9As shown in the diagram, in this embodiment, a bar chart is used to illustrate the comparison of robustness to missing data. The horizontal axis represents the five missing data scenarios, and the vertical axis represents the classification accuracy, in percentage. When the data is complete, the performance of each method is similar, but as the missing data rate increases, performance differences gradually become apparent. The method of this invention maintains the highest accuracy at different missing data rates, especially showing the most significant advantage at a 20% missing data rate, indicating that the method of this invention has a strong tolerance for partial data loss and is more reliable in practical applications.

[0110] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic monitoring of the construction process of floor decking, characterized in that, include: Collect stress data throughout the entire construction cycle of floor decking and construct a training dataset with construction stage category labels; Adaptive multi-scale morphological filtering is performed on the stress data in the training dataset, including: predefining a discrete set of structuring element scales; processing the stress data using two composite operations: morphological opening followed by closing and closing followed by opening; determining the dominant scale by analyzing the residual energy distribution of the stress data and the results of the two composite operations at each scale; calculating adaptive weights for the results of the two composite operations based on the dominant scale; and combining the results of the two composite operations according to the adaptive weights to obtain a filtered stress data vector. A construction monitoring model is constructed and trained using the training dataset to perform dynamic monitoring of the construction process. The construction monitoring model uses a deep convolutional neural network with adaptive deep convolution and cross-layer feature bridging as the backbone feature extractor, and determines the classification probability of the construction stage through an attention classification layer guided by external features and a weighted loss function for mutation awareness.

2. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, The collected stress data includes stress time-series signals, corresponding timestamps, and construction activity markers; The construction of the training dataset with construction stage category labels includes: Continuous stress data is cut into fixed-length segments or segments corresponding to the duration of construction activities. Based on the main construction activities within the time period of each data segment, each data segment is assigned a corresponding construction stage category label, including construction preparation stage, concrete pouring stage, curing and hardening stage, and loading test stage. The training dataset consists of data segments labeled with construction stage categories.

3. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, The stress data is processed using two composite operations: one performing a morphological opening operation followed by a closing operation, and the other performing a closing operation followed by an opening operation. An adaptive multi-scale morphological filtering framework is constructed, and the structuring element is used to perform opening-closing and closing-opening operations on the acquired raw stress data respectively. The two computational results are weighted and fused based on adaptive weights that satisfy complementary relationships to obtain a filtered stress data vector that removes positive and negative impulse noise.

4. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, The method for determining the result of the opening / closing operation / the result of the closing / opening operation includes: Using a flat linear structural element under the dominant scale, opening-closing operations and closing-opening operations are performed on the original stress data to obtain the opening-closing operation results and the closing-opening operation results. The adaptive weights include adaptive weights corresponding to the opening and closing operation results and adaptive weights corresponding to the closing and opening operation results, both of which are determined based on the residual vector under the dominant scale. The residual vectors include the first residual vector between the original stress data and the opening and closing operation results under the dominant scale, and the second residual vector between the original stress data and the closing and opening operation results under the dominant scale.

5. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, The construction monitoring model also includes a dual-path feature extraction module, which is used to extract the global contour feature vector and local significant fluctuation feature vector of the filtered stress data, and fuse them to obtain the input of the adaptive depth convolution.

6. The method for dynamic monitoring of the construction process of floor decking according to claim 5, characterized in that, The dual-path feature extraction module includes a first path and a second path; The first path is used to adaptively segment the filtered stress data, calculate weighted features based on the linear fitting parameters of each segment, and concatenate them to obtain the global contour feature vector. The second path is used to extract overlapping subsequence segments from the filtered stress data, encode the segments using a pre-trained learnable sparse coding dictionary, and calculate the mean and standard deviation of the coding coefficients to obtain the local significant fluctuation feature vector. The dual-path feature extraction module is used to concatenate the global contour feature vector with the local significant fluctuation feature vector, and then fuse them through a linear transformation layer to obtain an enhanced feature vector, which is used as the input of the adaptive depth convolution.

7. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, The deep convolutional neural network is configured with convolutional blocks, each of which contains adaptive depth convolution, batch normalization, activation function and cross-layer feature bridging; the adaptive depth convolution introduces a channel attention mechanism, which calculates channel attention weights through global average pooling and fully connected layers to modulate the contribution of each channel convolution. The cross-layer feature bridging adjusts the dimension of the shallow layer feature map through a one-dimensional 1×1 convolution with a stride of 2, and then adds it to the current layer feature map to achieve fusion.

8. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, The construction monitoring model processes the output feature map of the deep convolutional neural network through an attention classification layer guided by external features, fuses the statistical features of the filtered data, and outputs the classification probability through a fully connected layer and a Softmax function; the attention classification layer is configured with three types of transformation weight matrices: query, key, and value. The query transformation weight matrix is ​​used to map the external feature vector to the query vector; the external feature vector is determined based on the statistical feature vector calculated from the filtered stress data vector. The key transformation weight matrix and the value transformation weight matrix are used to map the feature vectors of each time position of the stress data into key vectors and value vectors, respectively; the attention calculation process is also associated with the dimensions of the query vector and the key vector, the weight matrix of the classification output layer, and the bias vector.

9. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, Training the construction monitoring model also includes: An adaptive moment estimation optimization algorithm is used to iteratively update the model parameters, and the model training is completed using the total loss function; wherein, the total loss function is composed of the sum of the mutation-aware weighted cross-entropy loss and the attention regularization loss after being weighted by the regularization coefficient; The mutation-aware weighted cross-entropy loss is determined based on preset mutation weight coefficients, the average gradient magnitude of the filtered stress data vector, the construction stage category labels, and the elements of the predicted classification probability vector. The attention regularization loss is determined based on the attention weights corresponding to each time position of the stress data.

10. The method for dynamic monitoring of the construction process of floor decking according to claim 1, characterized in that, Construction process monitoring is performed based on the aforementioned construction monitoring model, including: The real-time filtered data is segmented using a sliding window method, input into the construction monitoring model to obtain the probability of each category, and the category with the highest probability is output as the identification result of the current construction stage. An early warning signal is triggered when an abnormal stress pattern or stage deviation is detected.

Images