Construction condition intelligent classification method for cable-stayed bridge cantilever hanging basket

Abstract

The application relates to engineering intelligent management and monitoring, in particular to a construction working condition intelligent classification method for a cable-stayed bridge cantilever hanging basket, which specifically comprises the following steps: first, synchronously collecting double-channel vibration data of the front end of the hanging basket and a pouring beam section through a vibration acceleration sensor, and performing data set division after labeling; performing collaborative normalization processing on the data based on mutual information, and then adopting an adaptive multi-scale wavelet packet method to screen an optimal scale parameter set and generate an optimal multi-scale feature matrix; constructing a classification neural network containing a double-channel interactive attention module, modeling a collaborative or antagonistic relationship between the channels to obtain an enhanced feature matrix, simultaneously constructing environment disturbance agent features and fusing classification; calculating a total loss function, training and evaluating the model, and then applying the model to actual working condition classification, outputting working condition probability distribution, and taking a working condition corresponding to the maximum probability as the final result. The application provides a core judgment basis for digital and intelligent monitoring of cable-stayed bridge cantilever construction.

Description

Technical Field

[0001] This invention relates to the field of intelligent management and monitoring technology for engineering projects, and in particular to an intelligent classification method for construction conditions of cantilever formwork for cable-stayed bridges. Background Technology

[0002] In the construction of large cable-stayed bridges, the use of formwork for cantilever segment construction is a crucial step. This process involves multiple typical and continuous working conditions, including formwork movement, formwork installation, rebar tying, concrete pouring, prestressing tensioning, and system conversion. Under each working condition, the formwork-beam structure is in a different mechanical state, and its dynamic response characteristics vary significantly. Traditionally, monitoring of these construction stages relies mainly on manual inspections, construction logs, and video surveillance. This is not only labor-intensive but also suffers from subjective judgments, poor real-time performance, and difficulty in quantification, making it impossible to detect abnormal conditions in a timely manner and potentially creating safety hazards. With the development of structural health monitoring technology, it has become possible to collect structural response data during construction using vibration sensors. However, most existing technologies focus on the post-construction operation phase or only monitor single physical quantities during construction, lacking a systematic method for intelligent identification and classification of complex and dynamically changing construction conditions. In the few related studies, traditional signal processing techniques are typically used to extract fixed features and combined with conventional machine learning models for classification.

[0003] Existing technologies objectively suffer from the following shortcomings: Conventional normalization techniques typically scale the mean and standard deviation of data independently for each channel, neglecting the inherent mechanical synergy between the formwork and the vibration signals of the already cast beam segment during the cantilever construction of cable-stayed bridges. This results in the weakening or loss of crucial cross-channel phase and energy synchronization features for case classification during preprocessing. Traditional methods often employ fixed-scale wavelet transforms or predefined frequency band filtering analysis, failing to adaptively adjust based on the spectral characteristics of specific construction vibration data. The extracted features may contain a large amount of irrelevant frequency band information or omit certain subtle frequency band features specific to certain cases, leading to low feature efficiency. The discriminative power needs improvement; conventional classification models typically splice or process dual-channel features simply, lacking a dedicated structure for modeling the complex interaction mechanisms of dual-source signals under specific construction conditions, thus limiting the model's ability to extract deep coupling patterns from the data; standard classification task training usually only uses pure statistical loss functions such as cross-entropy, failing to incorporate the domain prior knowledge that cable-stayed bridge cantilever construction conditions occur in a fixed order and that adjacent conditions should have a certain continuity in features, which may cause the feature representations learned by the model to violate the physical logic of the construction process, affecting its stability and interpretability in complex real-world scenarios.

[0004] Therefore, this invention proposes an intelligent classification method for construction conditions of cantilever hanging baskets for cable-stayed bridges to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes an intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges. This invention can provide core judgment criteria for the digital and intelligent monitoring of cantilever construction of cable-stayed bridges.

[0006] The technical solution of this invention to solve the technical problem is an intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges, comprising the following steps:

[0007] S1. During the construction phase of the cantilever formwork of the cable-stayed bridge, vibration acceleration sensors installed at the front end of the formwork and the cast beam segment are used to synchronously collect dual-channel vibration monitoring data at a fixed sampling frequency. The collected data are labeled with the working condition category, and the labeled data are divided into training set, validation set and test set according to the proportion.

[0008] S2. The collected data is normalized using a mutual information-based collaborative normalization method. The normalization parameters are adjusted by quantifying the statistical dependencies between channels to generate collaboratively normalized data.

[0009] S3. An adaptive multi-scale wavelet packet feature extraction method is adopted. The optimal set of scale parameters is selected from the co-normalized data through optimization criteria. Based on this, the vibration energy characteristics of each channel at each scale are calculated to form the optimal multi-scale feature matrix.

[0010] S4. Construct a neural network model for construction condition classification. Model the collaborative or adversarial relationship between the two channels in the multi-scale feature matrix at different feature scales through a dual-channel interactive attention module to obtain an enhanced feature matrix. Then, use the obtained interactive relationship to construct an environmental disturbance proxy feature that reflects potential environmental disturbances. Merge the enhanced feature matrix and the proxy feature to classify the construction condition and obtain the predicted probability distribution of the construction condition.

[0011] S5. Calculate the total loss function of the model. Based on the standard classification loss, introduce the working condition smoothness constraint loss. Apply structured constraints to the feature space based on the enhanced feature matrix, and use the weight matrix that reflects the adjacency relationship of working condition categories to modulate the feature distance loss to obtain the working condition smoothness constraint loss term.

[0012] S6. Use the training set and validation set to train the construction condition classification neural network model to obtain the trained model, and then use the test set to evaluate the trained model.

[0013] S7. Apply the trained model to the intelligent classification task of actual cable-stayed bridge cantilever formwork construction, predict its construction conditions, output the probability distribution of each preset construction condition category, and take the condition category corresponding to the highest probability as the final judgment output.

[0014] S1 is as follows:

[0015] The two key locations for deploying vibration acceleration sensors are the main longitudinal beam or front transverse beam at the front end of the hanging basket, and the root of the most recent beam segment that has been poured and has reached the required strength. Data acquisition covers the entire cycle of hanging basket construction, including hanging basket movement, formwork installation and adjustment, rebar binding, concrete pouring, prestressing tensioning and system conversion.

[0016] Based on the construction site's supervision logs, construction records, and video surveillance data, the working conditions are categorized. Operators use precise timestamps to map each vibration data segment, lasting from several seconds to tens of seconds, to the specific construction activity stage at which it occurred, thus assigning a discrete working condition category label to each data sample. The working condition category labels include: hanging basket movement, formwork installation and adjustment, rebar tying, concrete pouring, prestressing tensioning, and system conversion.

[0017] The labeled dual-channel vibration data samples and their corresponding working condition labels are divided into training set, validation set and test set according to the proportion.

[0018] S2 is as follows:

[0019] S2.1. Obtain the joint probability distribution and marginal probability distribution from the discrete sampled data using the Gaussian kernel density estimation method, and calculate the mutual information between the data from the two vibration monitoring channels.

[0020] S2.2. Based on mutual information, the dual-channel vibration data is processed by co-normalization. By introducing mutual information as an adjustment factor, the traditional Z-score normalization result is scaled to retain the co-normalization characteristics. The exponentially decaying noise suppression term is used to attenuate abnormal noise points that deviate too far from the mean, so as to obtain the co-normalized data of each channel.

[0021] S3 is as follows:

[0022] S3.1. Set a set of candidate scale parameters from the co-normalized data of the two channels, and define a criterion function based on information entropy. By traversing the candidate scale combinations and minimizing the criterion, an optimal set of scales is automatically determined, and the optimal scale parameter set is obtained.

[0023] S3.2. Using the optimal scale parameter set, perform wavelet packet transform on the normalized data of each channel and extract the vibration energy features at each scale to form the optimal multi-scale feature matrix.

[0024] S3.2 is as follows:

[0025] For each scale in the optimal scale parameter set, the wavelet packet coefficients of each channel at that scale are calculated, and the feature values ​​are extracted by calculating the maximum absolute value of the wavelet packet coefficients. The maximum logarithmic wavelet packet coefficients of each channel at each optimal scale are obtained and used as elements to form the optimal multi-scale feature matrix.

[0026] S4 is as follows:

[0027] S4.1 Based on the optimal multi-scale feature matrix, the correlation strength between a certain scale feature of one channel and all scale features of another channel is measured by a learnable interaction weight matrix, so as to obtain the enhanced feature values ​​of each channel at each scale, which are used as elements to form the enhanced feature matrix.

[0028] S4.2 Utilize the inter-row differences and intra-row statistical properties of the enhanced feature matrix to construct an environmental disturbance proxy feature vector that characterizes potential time-varying disturbances in the construction environment;

[0029] S4.3 Flatten the enhanced feature matrix into a vector and concatenate it with the environmental disturbance proxy feature vector. Input the vector into a fully connected network with two hidden layers and output the predicted probability corresponding to various construction conditions through the Softmax function.

[0030] S4.2 is as follows:

[0031] The trainable parameters of the model are set during the calculation of surrogate features for environmental disturbances, including the weighting coefficients of each surrogate feature to each scale, the scaling factor of the contribution of channel 2 to channel 1 in each surrogate feature, and the contribution coefficient of the overall statistical term.

[0032] S5 is detailed below:

[0033] S5.1 Calculate the working condition smoothness constraint loss based on all samples in the current training mini-batch. By imposing constraints in the feature space, the model is encouraged to learn feature representations that conform to the continuity of the construction sequence.

[0034] Specifically, for any two samples, if their actual working condition categories are adjacent in the construction sequence, the model should make their distance in the feature space relatively small; if the categories are neither the same nor adjacent, they should be encouraged to have a larger feature distance.

[0035] S5.2. The total loss function of the model is obtained by summing the standard cross-entropy classification loss and the operating condition smoothness constraint loss with weights.

[0036] S6 is detailed below:

[0037] Using the training set, all trainable parameters in the neural network model are automatically adjusted through an iterative optimization algorithm. A gradient-based optimization algorithm is used to calculate the gradient of the total loss function with respect to all trainable parameters, and the parameters are updated accordingly to minimize the total loss function.

[0038] The training process is monitored synchronously on the validation set, and the model's performance metrics are evaluated using the validation set.

[0039] Training continues until the preset training termination conditions are met, resulting in a trained model. The trained model is then tested using a test set.

[0040] S7 is detailed below:

[0041] Real-time collected fixed-length dual-channel vibration data is input into the trained model, which outputs the probability distribution of each construction condition category. The condition category with the highest probability is taken as the result, and the hanging basket construction stage is identified in real time. The results are stored and visualized.

[0042] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0043] This invention discloses an intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges. It employs a cooperative normalization method based on mutual information, dynamically adjusting the normalization parameters using the statistical dependence of dual-channel vibration signals. This method uniquely preserves the cooperative vibration characteristics between channels, which are closely related to construction mechanics coupling, while eliminating dimensions and suppressing noise. This overcomes the limitation of traditional independent-channel normalization, which leads to the loss of cooperative information. Furthermore, it utilizes an adaptive multi-scale wavelet packet feature extraction mechanism. Through an optimization criterion based on information entropy, it dynamically selects the set of feature scales most sensitive to the construction conditions, automatically focusing on the discriminative frequency band from data-driven high-dimensional vibration signals, overcoming the limitations of fixed-scale methods. The decomposition may contain redundancy or loss of key details; a classification neural network model that integrates channel interactive attention and environmental disturbance surrogate features is adopted. This model not only models the dual-channel cross-scale vibration coupling mode through a learnable interactive weight matrix, but also constructs surrogate features to characterize non-operating environment disturbances, thereby enhancing the model's discriminative power and robustness; a loss function that combines operating condition smoothness constraints is adopted. By introducing structured constraints based on the adjacency relationship of construction sequence into the feature space, the model is guided to learn feature representations that conform to the physical continuity of the construction process, improving the model's generalization ability and physical interpretability, and surpassing the conventional training goal of only focusing on classification accuracy. Attached Figure Description

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0045] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0046] Figure 2 This is a schematic diagram of the operation process of S3 in this invention.

[0047] Figure 3 This is a time-frequency diagram of the hanging basket's walking operation.

[0048] Figure 4 This is a time-frequency diagram of the template installation and adjustment conditions.

[0049] Figure 5 This is a time-frequency diagram of the rebar tying operation.

[0050] Figure 6 A time-frequency diagram of concrete pouring conditions.

[0051] Figure 7 This is a time-frequency diagram of the prestressed tensioning condition.

[0052] Figure 8 This is a time-frequency diagram of the system's transition operating conditions.

[0053] Figure 9 This is a schematic diagram showing the characteristic distribution of vibration energy under different construction conditions. Detailed Implementation

[0054] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0055] Example 1

[0056] like Figure 1 As shown, an intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges includes the following steps:

[0057] S1. During the construction phase of the cantilever formwork of the cable-stayed bridge, vibration acceleration sensors installed at the front end of the formwork and the cast beam segment are used to synchronously collect dual-channel vibration monitoring data at a fixed sampling frequency. The collected data are labeled with the working condition category, and the labeled data are divided into training set, validation set and test set according to the proportion.

[0058] S2. The collected data is normalized using a mutual information-based collaborative normalization method. The normalization parameters are adjusted by quantifying the statistical dependencies between channels to generate collaboratively normalized data.

[0059] S3. An adaptive multi-scale wavelet packet feature extraction method is adopted. The optimal set of scale parameters is selected from the co-normalized data through optimization criteria. Based on this, the vibration energy characteristics of each channel at each scale are calculated to form the optimal multi-scale feature matrix.

[0060] S4. Construct a neural network model for construction condition classification. Model the collaborative or adversarial relationship between the two channels in the multi-scale feature matrix at different feature scales through a dual-channel interactive attention module to obtain an enhanced feature matrix. Then, use the obtained interactive relationship to construct an environmental disturbance proxy feature that reflects potential environmental disturbances. Merge the enhanced feature matrix and the proxy feature to classify the construction condition and obtain the predicted probability distribution of the construction condition.

[0061] S5. Calculate the total loss function of the model. Based on the standard classification loss, introduce the working condition smoothness constraint loss. Apply structured constraints to the feature space based on the enhanced feature matrix, and use the weight matrix that reflects the adjacency relationship of working condition categories to modulate the feature distance loss to obtain the working condition smoothness constraint loss term.

[0062] S6. Use the training set and validation set to train the construction condition classification neural network model to obtain the trained model, and then use the test set to evaluate the trained model.

[0063] S7. Apply the trained model to the intelligent classification task of actual cable-stayed bridge cantilever formwork construction, predict its construction conditions, output the probability distribution of each preset construction condition category, and take the condition category corresponding to the highest probability as the final judgment output.

[0064] In a specific implementation, S1 is as follows:

[0065] A training dataset for classifying construction conditions was constructed. Specifically, vibration monitoring data was collected synchronously at two key locations during the actual construction of the cable-stayed bridge cantilever formwork: the front end of the formwork and the poured beam segment. Based on detailed construction logs, accurate condition category labels were assigned to each collected data segment.

[0066] In the cantilever construction segment of a cable-stayed bridge, a vibration acceleration sensor is installed on the main longitudinal beam or front cross beam at the front end of the formwork, and another vibration acceleration sensor is installed at the root of the nearest beam segment that has been poured and has reached the required strength, thus forming a dual-channel synchronous monitoring system.

[0067] Data acquisition covers the entire cycle of hanging basket construction, including multiple typical construction stages such as hanging basket movement, formwork installation and adjustment, rebar binding, concrete pouring, prestressing tensioning and system conversion; during acquisition, a fixed high sampling frequency is used for continuous recording to ensure that the full-frequency information of structural vibrations excited by construction activities can be captured.

[0068] After obtaining the raw vibration time-series data, data annotation was performed. The annotation work was strictly based on the construction site supervision logs, construction records, and video monitoring data. Operators used precise timestamps to map each vibration data segment, lasting from several seconds to tens of seconds, to the specific construction activity stage in which it occurred, thus assigning a discrete working condition category label to each data sample. The labeled categories included: 1-Hanging basket movement, 2-Formwork installation and adjustment, 3-Reinforcement binding, 4-Concrete pouring, 5-Prestressing tensioning, and 6-System conversion, for a total of 6 categories.

[0069] The labeled dual-channel vibration data samples and their corresponding working condition labels are divided into training set, validation set and test set according to a certain ratio to form a complete dataset for model training, optimization and performance evaluation.

[0070] like Figures 3 to 8 As shown, the example presented uses the raw vibration monitoring data sequence collected from the front end of the hanging basket. A time-frequency diagram is generated using short-time Fourier transform to visually represent the differences in the distribution of vibration signal energy in the time and frequency dimensions under six different operating conditions. Figures 3 to 8 In the diagram, the horizontal axis represents time (seconds), the vertical axis represents frequency (Hz), and the colors from dark to light (corresponding to a color spectrum from black to yellow and white) represent energy intensity from low to high; for example... Figure 3 As shown, the energy during the hanging basket's walking operation is concentrated in the low-frequency region (around 20Hz) and has periodic impacts; such as Figure 4 As shown, energy accumulation and modulation phenomena occur during template installation and adjustment in the mid-frequency range (approximately 40Hz); as Figure 5 As shown, the rebar binding exhibits a broadband random energy distribution; as Figure 6 As shown, the concrete pouring exhibits strong, sustained low-frequency (approximately 15 Hz) energy; as Figure 7 As shown, prestressing tension exhibits significant energy spikes in the high-frequency range (approximately 60Hz and 90Hz); for example... Figure 8 As shown, the system conversion exhibits a complex pattern of multi-band mixing.

[0071] In a specific implementation, S2 is as follows:

[0072] Vibration monitoring data during the construction of cable-stayed bridges using cantilever formwork is collected from the front end of the formwork and the already poured beam segment. This data is characterized by dual-channel characteristics, high dimensionality, non-stationarity, and strong noise interference. Conventional normalization techniques independently calculate the mean and standard deviation of each channel for scaling, neglecting the inherent cooperative patterns between the two-channel vibration signals, such as phase synchronization. This easily leads to the loss of cooperative features closely related to the construction conditions in the normalized data. This invention employs a cooperative normalization method based on mutual information. By quantifying the statistical dependencies between channels to adjust the normalization parameters, it effectively preserves the cooperative vibration features crucial for classifying the construction conditions while eliminating dimensional differences and suppressing noise. The specific steps are as follows:

[0073] S2.1, Inter-channel mutual information calculation

[0074] The joint probability distribution and marginal probability distribution are obtained from the discrete sampled data using the Gaussian kernel density estimation method. The mutual information between the data from the two vibration monitoring channels is calculated to quantify their statistical dependence and serve as the basis for adjusting the co-normalization process, expressed as:

[0075]

[0076] In the formula, This represents the mutual information between the vibration data of channel 1 and channel 2, measured in nanots. It is approximately calculated based on discrete sampled data and is used to quantify the statistical dependence and degree of coordination between the two channels. A larger value indicates a higher degree of coordination. This represents the sequence of raw vibration monitoring data collected from the front end of the hanging basket, with dimensions of [dimension not specified]. , express One of the specific values; This represents the sequence of raw vibration monitoring data collected from the already poured beam segment, with dimensions of [dimension not specified]. , express One of the specific values; express and The joint probability distribution is obtained through kernel density estimation. In specific implementation, a bivariate Gaussian kernel function is used to... and Smooth the sample points and estimate their joint probability density; express The marginal probability distribution is obtained through kernel density estimation. In specific implementation, a univariate Gaussian kernel function is used to... Smooth the sample points and estimate their marginal probability density; express The marginal probability distribution is obtained through kernel density estimation. In specific implementation, a univariate Gaussian kernel function is used to... Smooth the sample points and estimate their marginal probability density; This represents a logarithmic function, with the default base being the natural constant.

[0077] S2.2, Cooperative Normalization and Noise Suppression Execution

[0078] Based on mutual information, the original dual-channel vibration data is co-normalized. Mutual information is introduced as a scaling factor to scale the traditional Z-score normalization result to preserve co-normalization features. An exponentially decaying noise suppression term is used to attenuate abnormal noise points that deviate too far from the mean. This is expressed as:

[0079]

[0080] In the formula, Indicates the first The vibration monitoring channel is in the first The co-normalized data at each moment is the first The vibration monitoring channel is in the first The data obtained after co-normalization and noise suppression at each time step is a scalar, the first... Coordinated normalized data from each vibration monitoring channel The dimension is still ; This represents the channel index, with a value of or ,in Corresponding to the front channel of the hanging basket, Corresponding to the already poured beam segment passage; Indicates the first The first channel in the Raw vibration monitoring data at any given time; Indicates the first The arithmetic mean of the raw data from each channel is used to eliminate DC offset in the data. The calculation method is as follows: ; Indicates the first The standard deviation of the raw data for each channel is used to scale the data to a uniform, comparable scale. The calculation method is as follows: ; Represents the discrete-time index, representing the first... Each sampling time, with a value range of [value]. arrive ; The signal length represents the total number of data points in a single sample, typically taking the value of [value to be filled in]. ; The coefficient represents the coordinated adjustment coefficient, a preset positive real hyperparameter used for control. The effect on normalized scaling intensity, typically taken as: ; This represents the attenuation coefficient, a preset positive real-valued hyperparameter used to adjust the sensitivity of the noise suppression term as the data points deviate from the mean. A typical value is [value missing]. ; This represents the natural exponential function.

[0081] It should be noted that the noise suppression term Abnormal noise is suppressed through an exponential decay mechanism, when Deviation When the value is large (which may be noise), the value of the noise suppression term decreases, thereby reducing the impact of this term on the normalization result. Conversely, data points close to the mean are less affected, thus enhancing the robustness of the normalized data and reducing noise interference while preserving collaborative features.

[0082] In specific implementation methods, such as Figure 3 As shown, S3 is as follows:

[0083] Because vibration data contains rich frequency components, different construction conditions such as concrete pouring and formwork adjustment will excite vibration modes in different frequency bands of the structure. Conventional wavelet transform or wavelet packet decomposition usually uses a predefined fixed scale or number of decomposition levels, which cannot adaptively focus on the frequency bands most effective in distinguishing the construction conditions according to the characteristics of the input data. This results in the extracted features possibly containing redundant information or losing key details. This invention adopts an adaptive multi-scale wavelet packet feature extraction method, which dynamically selects the most discriminative set of scales through optimization criteria, and calculates vibration energy features based on this set. This achieves a compact and effective representation of the sensitive frequency band information of the construction conditions. The specific steps are as follows:

[0084] S3.1 Adaptive Selection of Optimal Wavelet Packet Scale Set

[0085] A set of candidate scale parameters is defined, and a criterion function based on information entropy is defined. By traversing the combinations of candidate scales and minimizing the criterion, a set of optimal scales is automatically determined. The frequency bands corresponding to these scales can most effectively characterize the vibration mode distribution of the current data sample. Specifically,

[0086] Let the set of candidate scale parameters be... For any one containing subsets of different scales Calculate its total criterion value Then the optimal scale parameter set The selection method is expressed as follows:

[0087]

[0088] In the formula, The optimal scale parameter set is the set of optimal scale parameters selected through optimization. It is a set of feature scales most sensitive to construction conditions, selected through optimization, so that the wavelet packet energy features extracted at these scales can distinguish different conditions to the greatest extent and reduce redundant information. Represents the optimal scale parameter set The number of mesoscale, i.e. ;

[0089] This represents the operation of selecting parameters that minimize the objective function value. The term indicates that the number of items does not exceed a certain threshold across all scales. scale subset In the process, we search for the criterion function. The smallest subset is the set of optimal scaling parameters. ; Indicates the first The candidate scale parameter is greater than The larger the scale value, the higher the extension of the wavelet basis function, and the lower the frequency components of the analyzed signal. A typical range of values ​​is... ; Indicates a containing Subsets of different scales; Indicates the number of scales in the selected scale subset; This represents the maximum number of candidate scales, used to control the range of scale search; a typical value is [value to be filled in]. ; Represents the criterion function. Representation scale subset The total criterion value, which is the sum of the feature entropies corresponding to all scales within the subset, is calculated as follows: ; Indicating at the candidate scale The feature value sequence extracted after wavelet packet transform of the co-normalized data of all channels represents all channels at a specific candidate scale. The set of eigenvalues ​​calculated above is used to evaluate the discriminative power of this scale; express Information entropy, the smaller the entropy value, the more concentrated the feature distribution of the data at that scale and the stronger its discriminative power for operating conditions. It is calculated as follows: .

[0090] For each channel In candidate scale Feature value extraction ,but That is, the set of these maximum absolute eigenvalues ​​for all channels.

[0091] S3.2 Calculation of wavelet packet energy features based on the optimal scale parameter set

[0092] Wavelet packet transform is performed on the normalized data of each channel using the optimal scale parameter set, and vibration energy features at each scale are extracted to form an optimal multi-scale feature matrix. Specifically,

[0093] For the optimal set of scale parameters The first in Individual scale Calculate the first The wavelet packet coefficients of each channel at this scale The feature values ​​are extracted by calculating the maximum absolute value of the wavelet packet coefficients. , is represented as:

[0094]

[0095] In the formula, Indicates the first Each channel at the optimal scale The wavelet packet coefficients with the largest absolute value on the matrix are the optimal multi-scale feature matrix. The element in the i-th row and j-th column represents the element in the i-th row and j-th column. Each channel at the optimal scale The maximum absolute value of the wavelet packet coefficients, i.e., the maximum vibrational energy that the channel can reach at this characteristic scale, is used... To determine the sampling of the basis functions; This indicates that it belongs to the set of optimal scale parameters. The Each scale parameter; This indicates all translation indices. The corresponding value is the maximum value; Indicates the first Each channel at the optimal scale The above corresponds to the first The wavelet packet coefficients with translation indices, whose absolute values ​​characterize the signal component intensity near a specific scale and time point, are specifically obtained by calculating the inner product of the signal and the wavelet packet basis functions, expressed as follows: ; Indicates the first The first channel in the The data obtained from the "cooperative normalization and noise suppression" process at each sampling point is the input for this step. This indicates that it corresponds to the optimal scale. Translation index The discrete wavelet packet basis functions in the th The values ​​at each sampling point are generated by the mother wavelet function; This indicates a shift index, used to control the position of the basis function on the time axis; This represents a discrete time point index distinct from t, with a value range of 1. arrive ; This indicates the sampling frequency of the vibration monitoring data, expressed in Hz, with a typical value of [value missing]. Hz.

[0096] In a specific implementation, S4 is as follows:

[0097] Conventional classification networks directly input feature vectors into fully connected layers for processing, failing to fully exploit the unique interaction patterns of the dual-channel vibrations between the formwork front end and the already cast beam segment under construction mechanics coupling. For example, under certain construction conditions, the vibration energy of both channels may increase synchronously, while under other conditions, they may exhibit a trade-off. This invention constructs a classification neural network model that integrates channel interaction attention and environmental disturbance proxy features. First, through a dual-channel interaction attention module, it models the cooperative or antagonistic relationship between the two channels at different feature scales in a multi-scale feature matrix. Furthermore, it utilizes this interaction relationship to construct an environmental disturbance proxy feature vector reflecting potential environmental disturbances. The enhanced features and proxy features are then fused for construction condition classification. The specific steps are as follows:

[0098] S4.1, Dual-channel interactive attention module

[0099] Based on the optimal multi-scale feature matrix, a learnable interaction weight matrix is ​​used to measure the correlation strength between a certain scale feature of the first channel and all scale features of the second channel, and the original features are enhanced accordingly to capture the coupling mode of dual-channel vibration, as expressed as:

[0100]

[0101] In the formula, Indicates the first The first channel in the The enhanced eigenvalues ​​at each scale are scalars, representing the enhanced feature matrix. The element in the i-th row and j-th column, the enhanced feature matrix The dimension is This characterizes the first [unit / item] under specific construction conditions. The first channel in the Vibration modes at all optimal scales are considered, along with coupling relationships with another channel at all optimal scales, thus providing a more comprehensive reflection of the mechanical state of the structure. This represents the target scale index, corresponding to the enhanced feature matrix. The column index, with a value range of 100. arrive ; Indicates the first Each channel in scale The maximum absolute value of the wavelet packet coefficients; This indicates that it belongs to the set of optimal scale parameters. The Each scale parameter; The interaction enhancement coefficient controls the extent to which the interaction information enhances the original features; it is a trainable parameter. This represents a function indicating the instantaneous relationship between channels. The term is used to calculate the first term. The interaction strength between the original features of the two channels at the source scale is expressed as: ; This represents the hyperbolic tangent function, used to normalize the intensity of the instantaneous relationship between channels to a bounded range. Positive values ​​indicate an enhancing effect, while negative values ​​indicate an inhibiting effect. The relation sensitivity coefficient is a trainable parameter that determines how sensitive the model is to differences between channels. This represents the channel contribution balance coefficient, which is a trainable parameter that allows the model to learn the relative importance of the two channels in their interaction. Represents the cross-scale interactive propagation function. The term is used to measure the source scale. Interactions on the scale are propagated to the target scale based on interaction weights. The calculation method is expressed as follows: ; The propagation sharpness coefficient is a trainable parameter that controls the sharpness of the distribution of cross-scale interaction weights. Represents the interaction weight matrix No. Line number Column elements, interaction weight matrix These are trainable parameters, the interaction weight matrix. The dimension is It is used to encode the correlation strength between different scales.

[0102] S4.2 Construction of Environmental Disturbance Proxy Features

[0103] By leveraging the inter-row differences and intra-row statistical properties of the enhanced feature matrix, an environmental disturbance proxy feature vector characterizing potential time-varying disturbances in the construction environment is constructed to improve the model's robustness to vibration changes caused by non-operating conditions. This vector is expressed as:

[0104]

[0105] In the formula, Represents the surrogate feature vector of environmental disturbance The Each component is a scalar, representing the environmental disturbance proxy feature vector. Used to encode potential, non-operational-condition-related environmental disturbance information, environmental disturbance proxy feature vector Dimensions This is a preset value; a typical value is... ; The index represents the surrogate feature component, and its value range is... arrive ; Indicates the first The agent feature is related to the first The weighting coefficients at each scale are trainable parameters that allow the model to automatically learn the importance of features at different scales in characterizing different types of environmental disturbances. Indicates the first The contribution scaling factor of channel 2 relative to channel 1 in each proxy feature is a trainable parameter that helps the model capture different modes of response of the two channels to environmental disturbances. It represents the contribution coefficient of the overall statistical term, is a trainable parameter, and allows the model to balance the proportion of local variance and global statistics in the surrogate features; Represents the enhanced feature matrix The arithmetic mean of all elements in the set, i.e. .

[0106] S4.3 Feature Fusion and Operating Condition Classification

[0107] The enhanced feature matrix is ​​flattened into a vector and concatenated with the environmental perturbation proxy feature vector. This concatenation is then fed into a fully connected network with two hidden layers. The Softmax function outputs the predicted probabilities corresponding to various construction conditions, as follows:

[0108]

[0109]

[0110] In the formula, This indicates that the model predicts the sample belongs to the first... The predicted probability of a construction condition is a scalar, which is a predicted probability vector. The p-th element, and satisfying ; The total number of categories of construction conditions is determined based on the actual construction stage and monitoring requirements of the specific project. In one implementation method, it includes: 1-hanging formwork movement, 2-formwork installation and adjustment, 3-reinforcement binding, 4-concrete pouring, 5-prestressing tensioning, and 6-system conversion, for a total of 6 categories. The index representing the construction condition category has a value range of 100. arrive ; Indicates difference from The index of construction condition categories, with a value range of [value range missing]. arrive ; Indicates the first The score for each type of working condition is used to represent the model's determination of whether a sample belongs to the first type based on its input features. The strength of the original evidence for a given type of work condition; a higher value indicates a greater likelihood that it belongs to that category. This represents the number of neurons in the first hidden layer, typically taking the value of [value to be filled in]. ; This represents the number of neurons in the second hidden layer, typically taking the value of [value missing]. ; This represents the index of a neuron in the second hidden layer, with a value range of 1. arrive ; This represents the index of a neuron in the first hidden layer, with a value range of 1. arrive ; This represents the index of the input layer feature element, with a value range of 1. arrive ; Represents the output layer weight matrix Middle connection The second hidden layer neuron and the first Elements of each output node, output layer weight matrix The dimension is , are trainable parameters; Represents the weight matrix of the second hidden layer Middle connection The first hidden layer neuron and the first Elements of the second hidden layer neurons, the weight matrix of the second hidden layer The dimension is , are trainable parameters; Represents the weight matrix of the first hidden layer Middle connection The input feature and the first Elements of the first hidden layer neurons, the first hidden layer weight matrix The dimension is , are trainable parameters; Indicates the output layer number The bias terms of each node are trainable parameters; Indicates the second hidden layer The bias terms of each neuron are trainable parameters; Indicates the first hidden layer. The bias terms of each neuron are trainable parameters; This represents a modified linear unit activation function, enabling neural networks to learn and represent complex nonlinear relationships; This represents flattening and enhancing the feature vector, specifically by using a dimension of... Enhanced feature matrix The vector is obtained by concatenating the rows in order. This indicates that the feature vector is enhanced by flattening. Environmental disturbance proxy feature vector The concatenated eigenvectors of the first generation Each element.

[0111] In a specific implementation, S5 is as follows:

[0112] Conventional classification tasks typically employ the standard cross-entropy loss function, focusing only on the difference between the model's predicted probability and the sample's true label. This fails to adequately consider the natural transition and state continuity of the cantilever formwork construction conditions in cable-stayed bridges over time. During construction, adjacent or similar conditions should exhibit a certain degree of similarity in the vibration feature space. However, the standard loss function lacks guidance on this characteristic, potentially leading to the model learning feature representations that violate the physical continuity of the construction process, thus affecting the model's generalization ability and physical interpretability. This invention introduces a condition smoothness constraint loss based on the standard classification loss. This loss, based on an enhanced feature matrix, applies structured constraints to the feature space and modulates the feature distance loss using a weight matrix reflecting the adjacency relationship of condition categories. This encourages the model to learn feature representations that conform to the continuity of the construction process. The specific steps are as follows:

[0113] S5.1 Calculation of Smoothness Constraint Loss under Operating Conditions

[0114] The smoothness constraint loss is calculated based on all samples in the current training mini-batch. This loss term encourages the model to learn feature representations that conform to the continuity of the construction sequence by imposing constraints on the feature space. Specifically,

[0115] For any two samples, if their true working condition categories are adjacent in the construction sequence, the model should make their distance in the feature space smaller.

[0116] If the categories are neither the same nor adjacent, they should be encouraged to have a greater feature distance;

[0117] In practice, this goal is achieved by introducing a weight matrix based on the adjacency relationship of real working condition categories to perform a weighted summation of the feature distances between sample pairs, as shown below:

[0118]

[0119] In the formula, The loss represents the smoothness constraint of the working conditions. It is a scalar, and the smaller its value, the more the feature representation learned by the model conforms to the prior knowledge of smooth transition of the construction working conditions. This indicates the number of samples in the training mini-batch, used to control the range of samples covered in a single loss calculation; a typical value is 32. This represents the index of a sample within a mini-batch, used to iterate through samples in the mini-batch; its value range is... arrive ; Indicates difference from The index of samples within a small batch, used to compare with the first... Each sample is paired to calculate the pairwise loss, with values ​​ranging from [value range missing]. arrive And satisfy ; Indicates the first in a small batch The true working condition category label of each sample is between Total number of construction condition categories Integers between; Indicates the first in a small batch The actual working condition category label for each sample; This represents the symmetric modulation weights based on the adjacency relationship of operating condition categories, which are preset. Dimensional modulation weight matrix No. Line number The elements of the column are used to adjust the contribution of the feature distance to the total loss function based on the adjacency relationship of the sample to the real working condition category; Indicates the first The flattened and enhanced feature vector of each sample is obtained by flattening the enhanced feature matrix of the sample. Flattened in row order, the dimension is... ; Indicates the first Flattening and enhancing feature vectors of each sample; The feature space distance function is used to calculate the cosine distance between two flattened enhanced feature vectors, measuring the degree of dissimilarity between sample pairs in the feature space learned by the model. The calculation method is expressed as follows: ; This represents the Euclidean norm, also known as the L2 norm.

[0120] Modulation weight matrix No. Line number The elements in the column are set according to the construction process sequence; for example, if the category... With category If they are adjacent in the construction sequence, then Positive values ​​are used to penalize excessively large feature distances. If the categories are the same, zero values ​​are used to avoid interfering with intra-class compactness. If the categories are neither the same nor adjacent, negative values ​​are used to encourage feature separation.

[0121] S5.2 Calculation of Total Loss Function

[0122] The total loss function for model training is obtained by weighted summation of the standard cross-entropy classification loss and the working condition smoothness constraint loss. This total loss function is optimized to simultaneously ensure classification accuracy and the physical rationality of the feature representation, and is expressed as:

[0123]

[0124] In the formula, The total loss function for model training is a scalar and represents the ultimate goal of optimizing model parameters. This represents the standard cross-entropy classification loss, used to constrain and calculate the difference between the model's predicted probability distribution and the true label distribution of the samples; The weighting coefficient represents the loss due to the smoothness constraint of the operating conditions. It is used to balance the relative importance of the classification accuracy loss and the feature smoothness prior loss in the total loss function. A typical value is 0.3.

[0125] In a specific implementation, S6 is as follows:

[0126] After completing the construction of the training dataset, defining the neural network model structure and the total loss function, the model training phase begins. The goal of this phase is to use the training dataset and an iterative optimization algorithm to automatically adjust all trainable parameters in the neural network model, enabling the model to accurately identify construction conditions from the input dual-channel vibration data.

[0127] The training process begins with initializing the model parameters, followed by iterative forward and backward propagation on the training set in mini-batches. During forward propagation, each batch of sample data undergoes co-normalization and noise suppression, adaptive multi-scale wavelet packet feature extraction, and is then input into the constructed classification neural network model, ultimately yielding the predicted probability distribution for each sample corresponding to each work condition category. Simultaneously, based on the true labels of that batch and the enhanced feature matrix generated internally by the model, the standard cross-entropy classification loss and work condition smoothness constraint loss are calculated in parallel and weighted according to preset weights to obtain the total loss.

[0128] During backpropagation, a gradient-based optimization algorithm (such as the Adam optimizer) is used to calculate the gradient of the total loss with respect to all trainable parameters, and the parameters are updated accordingly to minimize the total loss function.

[0129] The entire training process is monitored synchronously on the validation set. After a certain number of training iterations or cycles, the model's classification accuracy and other performance metrics are evaluated on the validation set. Training continues until the model's performance metrics on the validation set no longer improve or even begin to decline, or until the preset maximum number of training iterations is reached. At this point, the early stopping mechanism is triggered, the iteration is stopped, and a snapshot of the model parameters with the best performance on the validation set is saved as the final trained construction condition classification model.

[0130] In a specific implementation, S7 is as follows:

[0131] Once the neural network model for construction condition classification is trained, it can be deployed for intelligent classification of construction conditions in actual cable-stayed bridge cantilever formwork construction.

[0132] In actual construction monitoring scenarios, vibration sensors deployed at the front end of the formwork and the poured beam segment continuously collect raw vibration signals. First, the system rigorously performs collaborative normalization and noise suppression processing based on mutual information on a fixed length of dual-channel vibration data collected in real time, following the steps described in S2, to eliminate dimensions, suppress noise, and retain collaborative features. Then, the processed data is fed into the adaptive multi-scale wavelet packet feature extraction module described in S3. This module uses the optimal set of scale parameters learned from the training data to calculate and output the optimal multi-scale feature matrix. This feature matrix is ​​then input into a pre-trained neural network model that includes a dual-channel interactive attention module and environmental disturbance proxy features. The model automatically performs feature enhancement, proxy feature generation, and fusion calculations, ultimately providing the probability distribution of the current data sample belonging to various preset construction condition categories through the Softmax output layer. The intelligent classification result uses the condition category corresponding to the highest probability as the final judgment output, thereby identifying the current construction stage of the formwork in real time, such as "concrete pouring" or "prestressing tensioning." The classification results can be displayed intuitively on the monitoring interface, or stored in the database for construction process traceability and safety analysis, providing a core basis for the digital and intelligent monitoring of cantilever construction of cable-stayed bridges.

[0133] Example 2

[0134] like Figure 9 As shown, box plots are used to compare the distribution of vibration energy characteristic values ​​under different construction conditions, verifying that the features extracted by the method of this invention have better intra-class consistency and inter-class distinguishability under different conditions. The experiment covers six construction conditions, including formwork movement, formwork installation, rebar tying, concrete pouring, prestressing tensioning, and system conversion. Figure 9The figure displays the median, quartile range, and outlier distribution of eigenvalues ​​for each working condition. The shorter bins and fewer outliers for each condition indicate a compact feature distribution and high intra-class consistency. This demonstrates that the present invention, through adaptive multi-scale wavelet packet feature extraction, can dynamically select the frequency band most sensitive to the working condition, and the calculated vibration energy features can effectively characterize the differences in vibration modes across different construction activities. The vertical axis in the figure represents dimensionless vibration energy eigenvalues, reflecting the maximum vibration energy of the signal at different optimal scales.

[0135] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. An intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges, characterized in that, Includes the following steps: S1. During the construction phase of the cantilever formwork of the cable-stayed bridge, vibration acceleration sensors installed at the front end of the formwork and the cast beam segment are used to synchronously collect dual-channel vibration monitoring data at a fixed sampling frequency. The collected data are labeled with the working condition category, and the labeled data are divided into training set, validation set and test set according to the proportion. S2. The collected data is processed using a collaborative normalization method based on mutual information. The normalization parameters are adjusted by quantifying the statistical dependencies between channels to generate collaboratively normalized data. S3. An adaptive multi-scale wavelet packet feature extraction method is adopted. An optimization criterion is designed based on the criterion function of information entropy. Then, the optimal scale parameter set is selected from the co-normalized data through the optimization criterion. On this basis, the vibration energy characteristics of each channel at each scale are calculated to form the optimal multi-scale feature matrix. S4. Construct a neural network model for construction condition classification. Model the collaborative or adversarial relationship between the two channels in the multi-scale feature matrix at different feature scales through a dual-channel interactive attention module to obtain an enhanced feature matrix. Then, use the obtained interactive relationship to construct an environmental disturbance proxy feature that reflects potential environmental disturbances. Merge the enhanced feature matrix and the proxy feature to classify the construction condition and obtain the predicted probability distribution of the construction condition. S4 is as follows: S4.1 Based on the optimal multi-scale feature matrix, the correlation strength between a certain scale feature of one channel and all scale features of another channel is measured by a learnable interaction weight matrix, so as to obtain the enhanced feature values ​​of each channel at each scale, which are used as elements to form the enhanced feature matrix. S4.2 Utilize the inter-row differences and intra-row statistical properties of the enhanced feature matrix to construct an environmental disturbance proxy feature vector that characterizes potential time-varying disturbances in the construction environment; S4.3 Flatten the enhanced feature matrix into a vector and concatenate it with the environmental disturbance proxy feature vector. Input the vector into a fully connected network with two hidden layers and output the predicted probability corresponding to various construction conditions through the Softmax function. S5. Calculate the total loss function of the model. Based on the standard classification loss, introduce the working condition smoothness constraint loss. Apply structured constraints to the feature space based on the enhanced feature matrix to encourage the model to learn feature representations that conform to the continuity of construction sequence. Modulate the feature distance loss using the weight matrix that reflects the adjacency relationship of working condition categories to obtain the working condition smoothness constraint loss term. Specifically, for any two samples, if their actual working condition categories are adjacent in the construction sequence, the model reduces their distance in the feature space; if the categories are neither the same nor adjacent, it encourages them to increase their distance in the feature space. S6. Use the training set and validation set to train the construction condition classification neural network model to obtain the trained model, and then use the test set to evaluate the trained model. S7. Apply the trained model to the intelligent classification task of actual cable-stayed bridge cantilever formwork construction, predict its construction conditions, output the probability distribution of each preset construction condition category, and take the condition category corresponding to the highest probability as the final judgment output.

2. The intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges according to claim 1, characterized in that, S2 is as follows: S2.

1. Obtain the joint probability distribution and marginal probability distribution from the discrete sampled data using the Gaussian kernel density estimation method, and calculate the mutual information between the data from the two vibration monitoring channels. S2.

2. Based on mutual information, the dual-channel vibration data is processed by co-normalization. By introducing mutual information as an adjustment factor, the traditional Z-score normalization result is scaled to retain the co-normalization characteristics. The exponentially decaying noise suppression term is used to attenuate abnormal noise points that deviate far from the mean, so as to obtain the co-normalized data of each channel.

3. The intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges according to claim 1, characterized in that, S3 is as follows: S3.

1. Set a set of candidate scale parameters from the co-normalized data of the two channels, and define a criterion function based on information entropy. By traversing the candidate scale combinations and minimizing the criterion, an optimal set of scales is automatically determined, and the optimal scale parameter set is obtained. S3.

2. Using the optimal scale parameter set, perform wavelet packet transform on the normalized data of each channel and extract the vibration energy features at each scale to form the optimal multi-scale feature matrix.

4. The intelligent classification method for construction conditions of cantilever formwork for cable-stayed bridges according to claim 3, characterized in that, S3.2 is as follows: For each scale in the optimal scale parameter set, the wavelet packet coefficients of each channel at that scale are calculated, and the feature values ​​are extracted by calculating the maximum absolute value of the wavelet packet coefficients. The maximum logarithmic wavelet packet coefficients of each channel at each optimal scale are obtained and used as elements to form the optimal multi-scale feature matrix.

5. The intelligent classification method for construction conditions of cantilever formwork for cable-stayed bridges according to claim 1, characterized in that, S4.2 is as follows: The trainable parameters of the model are set during the calculation of surrogate features for environmental disturbances, including the weighting coefficients of each surrogate feature to each scale, the scaling factor of the contribution of channel 2 to channel 1 in each surrogate feature, and the contribution coefficient of the overall statistical term.

6. The intelligent classification method for construction conditions of cantilever formwork in cable-stayed bridges according to claim 1, characterized in that, S5 is detailed below: S5.1 Calculate the working condition smoothness constraint loss based on all samples in the current training mini-batch. By imposing constraints in the feature space, the model is encouraged to learn feature representations that conform to the continuity of the construction sequence. S5.

2. The total loss function of the model is obtained by summing the standard cross-entropy classification loss and the operating condition smoothness constraint loss with weights.

7. The intelligent classification method for construction conditions of cantilever formwork for cable-stayed bridges according to claim 1, characterized in that, S6 Specifically as follows: Using the training set, all trainable parameters in the neural network model are automatically adjusted through an iterative optimization algorithm. A gradient-based optimization algorithm is used to calculate the gradient of the total loss function with respect to all trainable parameters, and the parameters are updated accordingly to minimize the total loss function. The training process is monitored synchronously on the validation set, and the model's performance metrics are evaluated using the validation set. Training continues until the preset training termination conditions are met, resulting in a trained model. The trained model is then tested using a test set.

8. The intelligent classification method for construction conditions of cantilever formwork for cable-stayed bridges according to claim 1, characterized in that, S1 is as follows: The two key locations for deploying vibration acceleration sensors are the main longitudinal beam or front transverse beam at the front end of the hanging basket, and the root of the most recent beam segment that has been poured and has reached the required strength. Data acquisition covers the entire cycle of hanging basket construction, including hanging basket movement, formwork installation and adjustment, rebar binding, concrete pouring, prestressing tensioning and system conversion. Based on the construction site's supervision logs, construction records, and video surveillance data, the working conditions are categorized. Operators use precise timestamps to map each vibration data segment, lasting from several seconds to tens of seconds, to the specific construction activity stage at which it occurred, thus assigning a discrete working condition category label to each data sample. The working condition category labels include: hanging basket movement, formwork installation and adjustment, rebar tying, concrete pouring, prestressing tensioning, and system conversion. The labeled dual-channel vibration data samples and their corresponding working condition labels are divided into training set, validation set and test set according to the proportion.

9. The intelligent classification method for construction conditions of cantilever formwork for cable-stayed bridges according to claim 1, characterized in that, S7 is detailed below: Real-time collected fixed-length dual-channel vibration data is input into the trained model, which outputs the probability distribution of each construction condition category. The condition category with the highest probability is taken as the result, and the hanging basket construction stage is identified in real time. The results are stored and visualized.

Images