Artificial intelligence-based prestress tensioning process anomaly identification method

Abstract

The application relates to an artificial intelligence-based prestress tensioning process anomaly identification method, and specifically as follows: four steel strands are symmetrically arranged in the ring direction of a bridge main tower for synchronous tensioning, sensors are installed on the tensioning equipment of each steel strand, sample data for training is synchronously collected and constructed, and the sample is labeled; an anomaly identification model is constructed, sample data is input into the model, and the anomaly identification result is obtained through the composite embedding and space-time position coding of tensioning time sequence data, a multi-stage dynamic feature modulation and anomaly sensitive feature mining module, and a multi-stage space-time anomaly detector, the total loss function of the model is calculated, the anomaly identification model is trained; the sensor network is deployed to collect related data of the target tensioning process to be identified, the related data is input into the trained anomaly identification model, and an anomaly classification result is output. The application can accurately identify prestress tensioning multi-class anomalies and improve the detection efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology in engineering, and in particular to an artificial intelligence-based method for identifying anomalies in the prestressing tensioning process. Background Technology

[0002] In cable-stayed bridge construction, the prestressing tensioning process of the stay cables is a core step determining the structural safety and long-term service performance of the bridge. As the main load-bearing component of a cable-stayed bridge, the tensioning accuracy of the stay cables directly affects the main girder alignment control, the stress balance of the main tower, and the overall structural stability. If abnormalities occur during the tensioning process, such as over-tensioning, under-tensioning, uneven stress on the steel strands, or synchronous deviations, it may not only cause permanent damage to the stay cables but also lead to structural hazards such as main tower misalignment and main girder deflection. In severe cases, it may even threaten the bridge's construction safety and subsequent operational lifespan. Traditional tensioning process monitoring mainly relies on the "dual control principle," namely controlling hydraulic pressure and elongation, and issuing alarms for abnormalities through preset thresholds. However, the actual engineering environment is complex and variable, and abnormal patterns often have hidden, coupled, and phased characteristics. How to intelligently discover abnormal signs and achieve a shift from "passive alarm" to "active diagnosis" has become an urgent need in the field of intelligent engineering monitoring.

[0003] Existing technologies objectively suffer from the following shortcomings: Firstly, they employ a uniform feature processing method for all time points, neglecting the fundamental differences between normal patterns and abnormal characteristics at different stages of tensioning (such as loading and holding). This leads to normal fluctuations during stage transitions being easily misjudged as anomalies. Secondly, existing technologies largely rely on single-point threshold alarms or simple statistical rules, lacking in-depth modeling of physical mechanisms such as the pressure-displacement theoretical relationship and geometric symmetry constraints between steel strands. This results in poor interpretability and high false negative rates for complex patterns such as friction anomalies and synchronization anomalies. Thirdly, most existing time-series or image models struggle to effectively and uniformly handle temporal evolution, spatial layout, and multi-stage transitions, often processing spatiotemporal dimensions in isolation and failing to capture spatiotemporal coupling anomalies occurring during collaborative operations between steel strands. Fourthly, standard classification loss functions only pursue overall accuracy, failing to explicitly optimize the separation between different anomaly categories at the feature learning level, leading to confusion and misjudgment when faced with anomaly types exhibiting similar features.

[0004] Therefore, this invention proposes an artificial intelligence-based method for identifying anomalies in the prestressing tensioning process to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention develops an artificial intelligence-based method for identifying anomalies in the prestressing tensioning process. This invention can accurately identify multiple types of anomalies in prestressing tensioning, improving detection efficiency and accuracy, and ensuring the safety and quality of bridge construction.

[0006] The technical solution of this invention to solve the technical problem is an artificial intelligence-based method for anomaly identification in the prestressing tensioning process, comprising the following steps:

[0007] S1. Four steel strands are symmetrically arranged in the circumferential direction of the main tower of the bridge and simultaneously tensioned. Each tensioning device of the steel strand is equipped with a sensor network to collect data synchronously at a fixed frequency. This data is used to construct training samples and to label the samples with abnormal patterns. At the same time, historical normal data is analyzed to obtain prior knowledge parameters.

[0008] S2. Construct an anomaly recognition model. Input sample data into the model. After composite embedding and spatiotemporal location encoding of tension time series data, multi-stage dynamic feature modulation and anomaly sensitive feature mining module, and multi-stage spatiotemporal anomaly detector, obtain the anomaly recognition result. Calculate the total loss function of the model and train the anomaly recognition model.

[0009] S3. Deploy a sensor network to collect relevant data during the tensioning process of the target to be identified, call the physical feature reconstruction coefficients from the training phase, input the collected data into the trained anomaly identification model, and output the anomaly classification results after analysis and processing.

[0010] S1 is as follows:

[0011] S1.1 The sample construction process is as follows:

[0012] Complete tensioning process records are extracted from historical engineering data. Complete tensioning cycles are extracted at fixed time lengths to form four-dimensional tensor samples of fixed duration. The four-dimensional tensors are the corresponding batch dimension, steel strand dimension, time dimension, and feature dimension.

[0013] Among them, the batch dimension represents an independent sample, the steel bundle dimension corresponds to four symmetrically arranged steel bundles, the time dimension represents a fixed time length, and the feature dimension represents the feature value obtained from the data collected by the sensor network.

[0014] The sensor network includes hydraulic sensors, displacement sensors, temperature sensors, and concrete strain gauges arranged below the anchorage, collecting data from the start of tensioning until the end of the holding phase.

[0015] S1.2 The sample labeling process is as follows:

[0016] The annotation method, which is based on the fusion of expert experience and multi-source information, first evaluates the samples manually based on the tensioning process curve, construction log and subsequent test report, and then uses the preset engineering threshold rules to screen automatic alarm records to assist in annotation. Each sample is assigned a unique category label.

[0017] The label categories correspond to the abnormal patterns to be identified, including normal, friction abnormality, synchronization abnormality, and equipment mechanism abnormality.

[0018] S2 is as follows:

[0019] S2.1 Input the sample data into the anomaly identification model, perform composite embedding and spatiotemporal position encoding of tension time series data, specifically perform feature semantic embedding calculation, absolute time position encoding calculation and steel strand circumferential position encoding calculation, and then add and fuse the calculation results to obtain the enhanced embedding tensor;

[0020] S2.2 Input the enhanced embedding tensor into the multi-stage dynamic feature modulation and anomaly-sensitive feature mining module. By applying stage-aware dynamic gating to the enhanced embedding tensor and simultaneously fusing the statistical relationship features across steel strands, the modulation feature tensor is obtained.

[0021] S2.3 Construct a physical consistency feature reconstruction tensor based on prior knowledge parameters. The multi-stage spatiotemporal anomaly detector performs deep spatiotemporal modeling based on the physical consistency feature reconstruction tensor and the modulation feature tensor, and outputs the anomaly category probability distribution.

[0022] S2.4 Calculate the total loss function including cross-entropy loss and outlier pattern comparison loss;

[0023] S2.5. Train the model using backpropagation based on the total loss function until the iteration stopping condition is met.

[0024] S2.1 is as follows:

[0025] Construct a composite embedded module to perform the following calculations:

[0026] (1) The feature semantic embedding calculation adopts a stage-aware feature transformation mechanism. Based on the tensor stage label, learnable weights are dynamically selected. Each feature is weighted and projected, and then nonlinear transformation and bias addition are performed to adjust the dimension to obtain the feature semantic embedding tensor.

[0027] (2) The absolute time position coding calculation specifically adopts the decoupled two-dimensional coding of the learnable parameter matrix to directly encode the order of each time step and obtain the absolute time position coding tensor;

[0028] (3) The calculation of the circumferential position coding of the steel strand adopts the decoupled two-dimensional coding of sine and cosine functions. Based on the symmetrical geometric relationship of the circumferential angle position coding of the steel strand, the circumferential position coding tensor of the steel strand is obtained.

[0029] The enhanced embedding tensor is obtained by superimposing the above three types of tensors.

[0030] S2.2 is as follows:

[0031] S2.2.1 Perform stage-aware gating weight tensor calculation. First, map the stage label to the stage code, and then interact with the current augmented embedding tensor. Generate a stage-gating weight tensor that is isomorphic to the augmented embedding tensor through linear transformation and Sigmoid activation function.

[0032] Among them, the stage coding tensor is obtained by mapping the stage marker feature tensor through a learnable embedding layer. The stage marker feature tensor is obtained from the input sample data and includes four tensioning stages: initialization, 10% loading, 100% loading, and holding load.

[0033] S2.2.2 Perform cross-strand interaction feature tensor calculation. First, calculate the mean, standard deviation and range along the strand dimension. Then, broadcast these statistics along the strand dimension and concatenate them with the enhanced embedding tensor of each strand. Finally, compress the concatenated high-dimensional features back to the original hidden dimension through a linear transformation layer to form the cross-strand interaction feature tensor.

[0034] S2.2.3 Multiply the stage-aware gating weights element-wise with the input embedded features, and combine them with the cross-strand interaction features to obtain the modulation feature tensor.

[0035] The construction process of the S2.3.1 Physical Consistency Feature Reconstruction Tensor is as follows:

[0036] (1) Calculate the normalized residual between the actual elongation and the theoretical expected elongation of each steel strand at each time step to obtain the theoretical-actual displacement residual characteristics;

[0037] (2) Based on the circumferential symmetry of the steel strands, the pressure difference between each steel strand and the symmetrical steel strands is calculated to obtain the symmetrical pressure equilibrium residual characteristics;

[0038] (3) The two types of residual features are spliced ​​together and projected onto the same dimension as the hidden layer of the model through a lightweight linear layer to obtain the physical consistency feature reconstruction tensor.

[0039] The S2.3.2 multi-stage spatiotemporal anomaly detector includes a time dynamic extraction module, a spatial relationship Transformer module, and a stage-aware classification head module, as detailed below:

[0040] (1) In the time dynamic extraction module, the modulation feature tensor and the physical consistency feature reconstruction tensor are first fused by adding elements to obtain the fused feature tensor; then the fused feature tensor is input into the stage-gated temporal convolutional network to obtain the output tensor of the stage-gated temporal convolutional network.

[0041] (2) In the spatial relationship Transformer module, the output tensor of the stage-gated temporal convolutional network is regarded as a set of sequences in the steel strand dimension. The global dependency relationship between all steel strands is modeled by the multi-head self-attention mechanism. Specifically, the state tensor is obtained through the multi-head self-attention mechanism and residual connection. The state tensor is then output as a high-level spatiotemporal feature tensor through the feedforward network and residual connection.

[0042] (3) In the stage-aware classification head module, adaptive max pooling is first applied to the high-level spatiotemporal feature tensor along the time dimension to extract the most critical signs of each steel strand in the entire tensioning process, and the output tensor of adaptive max pooling is obtained; then, based on the output tensor of adaptive max pooling, the tensor reconstructed by physical consistency features and the stage coding tensor, the probability distribution of abnormal categories is output through the stage-conditional fully connected layer.

[0043] S2.4 is as follows:

[0044] The total loss function includes cross-entropy loss and outlier pattern contrast loss;

[0045] The cross-entropy loss is calculated based on the negative log probability of the true label of each sample.

[0046] For any sample in a batch, for all other samples with the same true label, calculate the global feature vector between them, then calculate the similarity between the global feature vectors of the function through cosine similarity, and then calculate the abnormal pattern contrast loss.

[0047] The global feature vector of each sample is obtained by average pooling of the high-level spatiotemporal feature tensor in the time dimension and the steel strand dimension.

[0048] S2.5 is as follows:

[0049] First, all trainable parameters in the multi-stage spatiotemporal anomaly detector are initialized. Then, the training data is input into the model. After forward propagation, the probability distribution of anomaly categories is obtained. The total loss function of the model is then calculated. Backpropagation is performed based on the total loss to minimize the total loss function until the iteration stopping condition is met and the training ends.

[0050] 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:

[0051] This invention discloses an anomaly identification method for prestressed tensioning processes based on artificial intelligence. Through a stage-aware dynamic feature modulation mechanism, the model can adaptively adjust its focus on different physical features according to the tensioning process, solving the problem of key anomaly signals being overwhelmed by traditional methods that apply the same processing strategy to all time points. By fusing physical consistency features with deep learning features, and explicitly embedding engineering prior knowledge such as friction models and geometric symmetry into the model, the specific identification capability for friction anomalies and synchronization anomalies is enhanced. A decoupled spatiotemporal coding scheme and a cascaded spatiotemporal attention architecture are employed, using learnable parameters to encode the temporal sequence and trigonometric functions to encode circumferential symmetry relationships, respectively. Then, through gated temporal convolution and spatial Transformer collaborative modeling, the limitation of standard sequence models in simultaneously capturing spatiotemporal coupling relationships is overcome. An anomaly pattern comparison loss function is used to actively bring similar anomalies closer together and push away dissimilar samples in the feature space, improving the model's discriminative power in distinguishing similar anomaly patterns under complex working conditions. Attached Figure Description

[0052] 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.

[0053] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0054] Figure 2 The impact of hidden layer dimension hyperparameters on anomaly detection accuracy.

[0055] Figure 3 The effect of hidden layer dimension hyperparameters on the F1 score of anomaly detection.

[0056] Figure 4 A line graph comparing the phase-aware feature modulation mechanism of this method with the fixed feature weight strategy of the standard method.

[0057] Figure 5 This is a line graph comparing the anomaly detection performance of the method of the present invention at different tensioning stages. Detailed Implementation

[0058] 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.

[0059] Example 1

[0060] like Figure 1 As shown, an artificial intelligence-based method for anomaly identification in the prestressing tensioning process includes the following steps:

[0061] S1. Four steel strands are symmetrically arranged in the circumferential direction of the main tower of the bridge and simultaneously tensioned. Each tensioning device of the steel strand is equipped with a sensor network to collect data synchronously at a fixed frequency. This data is used to construct training samples and to label the samples with abnormal patterns. At the same time, historical normal data is analyzed to obtain prior knowledge parameters.

[0062] S2. Construct an anomaly recognition model. Input sample data into the model. After composite embedding and spatiotemporal location encoding of tension time series data, multi-stage dynamic feature modulation and anomaly sensitive feature mining module, and multi-stage spatiotemporal anomaly detector, obtain the anomaly recognition result. Calculate the total loss function of the model and train the anomaly recognition model.

[0063] S3. Deploy a sensor network to collect relevant data during the tensioning process of the target to be identified, call the physical feature reconstruction coefficients from the training phase, input the collected data into the trained anomaly identification model, and output the anomaly classification results after analysis and processing.

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

[0065] S1.1 At a specific circumferential position on the main tower of the bridge, four steel strands are symmetrically arranged for synchronous tensioning. Each steel strand tensioning device is equipped with a high-precision sensor network to continuously collect multi-dimensional physical quantities throughout the tensioning process. This network includes hydraulic sensors, displacement sensors, temperature sensors, and concrete strain gauges arranged below the anchorage. All sensors are synchronously sampled at a fixed frequency of once per second, from the start of tensioning until the end of the holding phase, ensuring complete coverage of the entire tensioning process and forming a continuous time-series data stream.

[0066] To construct a dataset suitable for training artificial intelligence models, a large number of complete tensioning process records were extracted from historical engineering data. For each record, a complete tensioning cycle was extracted with a fixed time length (corresponding to 600 seconds). If the original record length exceeded 600 seconds, the key stage was extracted; if it was insufficient, zero-padding or truncation was performed according to defined rules to form a sample with a fixed time length.

[0067] Each sample is constructed as a four-dimensional tensor with the shape (1,4,600,12), corresponding to the batch dimension (single sample), the steel bundle dimension (4 symmetrical steel bundles), the time dimension (600 consecutive time steps), and the feature dimension (12 physical features), respectively. The specific dimensions are explained in Table 1, and the 12 feature attributes of each time point and each steel bundle are shown in Table 2.

[0068] Table 1. Detailed Explanation of Four-Dimensional Tensor Samples in Each Dimension

[0069]

[0070] Table 2. 12 feature fields in the feature dimension

[0071]

[0072] S1.2 Data annotation based on an annotation method that integrates expert experience and multi-source information;

[0073] First, each historical tensioning sample is comprehensively evaluated manually based on the tensioning process curve, construction log, and subsequent inspection reports (such as grouting quality of ducts and bridge alignment measurement results) to determine whether the entire process is normal or abnormal. Second, automatic alarm records are screened and labeled in conjunction with preset engineering threshold rules (such as the allowable deviation range of dual control indicators of pressure and elongation, and the maximum pressure difference limit between symmetrical steel strands). Finally, each sample is assigned a unique category label, and all labels constitute the monitoring signal of the model.

[0074] There are four categories of labels, which correspond perfectly to the anomaly patterns to be identified: "Normal", indicating that all parameters in the tensioning process meet the design and specification requirements; "Frictional anomaly", which is characterized by a significant deviation of the pressure-elongation relationship from the theoretical curve, suggesting abnormal duct frictional resistance; "Synchronization anomaly", which is characterized by multiple steel strands in symmetrical positions progressing asynchronously in terms of pressure or elongation, with differences exceeding the allowable range; and "Equipment structure anomaly", which refers to anomalies caused by tensioning equipment failure (such as oil pump depressurization, anchor slippage) or local structural problems (such as abnormal concrete strain).

[0075] In addition, it is necessary to analyze and obtain the prior knowledge parameters required for subsequent steps from historical normal data in advance, including the pressure-displacement conversion coefficient, target pressure offset coefficient, and elongation residual normalization factor at each stage, which are used to construct physical consistency features during the model building stage.

[0076] In a specific implementation, S2 is as follows:

[0077] S2.1 Composite Embedding and Spatiotemporal Location Coding of Tensioning Time Series Data

[0078] This invention constructs a composite embedding module that integrates feature semantic embedding, absolute time position encoding, and steel strand circumferential position encoding to uniformly process features with different physical dimensions. Specifically, the original prestressed tensioning data is subjected to feature semantic embedding calculation, absolute time position encoding calculation, and steel strand circumferential position encoding calculation, and the three are added together to obtain an enhanced embedding tensor that integrates feature semantics, temporal order, and spatial position information, expressed as:

[0079]

[0080] In the formula, Represents the augmented embedding tensor with dimension . It integrates feature semantics, temporal order, and spatial location information; B represents the batch size, which is the number of samples used for batch input during the training of the artificial intelligence model; B=16. This represents the total number of steel strands, and its value is determined by the actual steel strand layout scheme of the tensioning project. =8; This represents the total number of time steps contained in the sample sequence, and its value is determined by the data acquisition frequency and the duration of the tensioning process. 256; This represents the hidden layer dimension and is an adjustable hyperparameter used to control the expressive power of the model. =256;

[0081] The feature semantic embedding tensor has a dimension of . It is obtained by transforming the original prestressed tension data through stage-aware feature transformation, and is used to project the original features with different physical meanings onto a unified semantic space.

[0082] This represents an absolute time-location encoding tensor with dimension . This is used to encode the absolute order information for each time step. This tensor is broadcast in the strand dimension and batch dimension and is added to the features of all samples and strands.

[0083] This represents the circumferential position encoding tensor of the steel strand, with dimension . This is used to encode the geometric position of each steel bundle in the circumferential space. This tensor is broadcast in the time dimension and the batch dimension and is added to the features of all time steps and samples.

[0084] S2.1.1 Calculate the feature semantic embedding tensor. Since the original prestressed tensioning data contains multidimensional features with different physical meanings and dimensions, to fully explore the intrinsic correlations of features and their interaction with the tensioning stages, a stage-aware feature transformation mechanism is adopted. Learnable weights are dynamically selected based on the tensioning stage labels, and each feature is weighted and projected. Then, after nonlinear transformation and bias addition, the dimensions are adjusted to obtain the feature semantic embedding tensor, represented as:

[0085]

[0086] In the formula, This represents the original prestressing tension data, with dimensions of [dimension number missing]. Data was collected by the tensioning monitoring system; This indicates the dimension of the original prestressing tensioning data; the default value is 12. This indicates the first [number] in the original prestressing tensioning data. Each feature is a slice across all batches, all strands, and all time steps, and is a three-dimensional tensor with dimension . ; Represents a dynamic weight tensor with dimension . The feature weighting used to achieve stage-awareness is calculated as follows: ; For the purpose of the phase and characteristics The embedded weight vector is specifically designed for the stage. and characteristics The embedding weight vector used for modeling has a dimension of , are trainable parameters; This is an indicator function that outputs 1 when the internal condition is true, and 0 otherwise. Indicates the index of the feature dimension, with a value range of 1. , used to traverse the twelve primitive features; The stage marker feature tensor has a dimension of . The range of values ​​for each element is . These represent the four tensioning stages: initialization, 10% loading, 100% loading, and holding load, respectively, based on the original prestressing tensioning data. Obtain from; Indicates the index of the tensioning stage, with a value range of 100. These correspond to initialization, 10% loading, 100% loading, and load holding, respectively. This represents the element-wise multiplication operator; This represents the projection bias vector, with dimension . , are trainable parameters that are biased after projection to enhance the model's expressive power; This indicates a dimension adjustment operation used to ensure that the final shape of the output tensor is... .

[0087] In practice, the dimension adjustment operation adjusts the shape of the input tensor to... ,because The result dimension of the item is Add the projection bias vector after broadcasting Since the shape remains unchanged, the dimension adjustment operation is usually an identity operation. However, to ensure that the output shape is clear, the reshaping function is still used to ensure that the dimensions are correct. For example, in element-wise multiplication, the three-dimensional tensor is first... Expanding dimensions, and then with four-dimensional tensors Perform element-wise multiplication.

[0088] S1.1.2 Calculate the absolute time position coding tensor and the circumferential position coding tensor of the steel strands. To solve the problem that the standard position coding cannot adapt to the two-dimensional spatiotemporal grid, a decoupled two-dimensional position coding scheme is adopted. Specifically, the absolute time position coding tensor uses a learnable parameter matrix to directly encode the order of each time step. The circumferential position coding tensor of the steel strands is generated using sine and cosine functions based on the angular position of the steel strands on the circumference to encode the symmetrical geometric relationship between multiple steel strands.

[0089] Define the absolute time-position encoding tensor ,in, The position encoding parameter matrix is ​​a trainable parameter matrix with dimension 1. ;

[0090] It should be noted that the dimension of the position encoding parameter matrix is ​​[dimension number missing]. Converted via broadcast The effective shape can be used to obtain the absolute time-position encoding tensor. Dynamic weight tensor The specific value depends on the feature index. and current stage markers That is, dynamic weight tensor The value is determined by two factors: feature index. and current stage markers Each feature has an independent learnable weight at each stage, thereby achieving stage-aware feature transformation.

[0091] S1.1.3, The calculation method for the elements in the circumferential position encoding tensor of the steel strand is defined as follows:

[0092]

[0093]

[0094] In the formula, Indicates the first The angle corresponding to the steel strand is calculated as follows: ; This indicates the tensor of the circumferential position encoding of the steel strand in the batch index. Steel strand index Time step index Hidden Dimension Index The value at that point is calculated using the sine function; This indicates the tensor of the circumferential position encoding of the steel strand in the batch index. Steel strand index Time step index Hidden Dimension Index The value at that point is calculated using the cosine function; Indicates batch index; Indicates the steel strand index, ; Indicates the time step index; Indicates the hidden dimension index. ,default It is an even number; Indicates the first The angle of the steel bundle in the circumferential direction, expressed in radians, is calculated as follows: This ensures that all steel strands are evenly distributed around the circumference, thereby encoding circumferential symmetry. This represents the wavelength adjustment factor, a pre-set constant used to adjust the wavelength of the sine wave in different dimensions, thereby generating position codes of different frequencies. =10000.

[0095] It should be noted that the circumferential position encoding tensor of the steel strand The value of the element depends only on the steel strand index. , with batch index and time step index Since it is irrelevant, a valid shape can be generated by broadcasting. Circumferential position encoding tensor of steel strand .

[0096] S2.2 Multi-stage dynamic feature modulation and anomaly-sensitive feature mining

[0097] A stage-aware feature modulator and a cross-strand interactive feature mining module are adopted. By applying stage-aware dynamic gating to the embedded features, the key features of different stages are enhanced and the secondary features are suppressed. At the same time, the statistical relationship features across the strands are integrated to directly construct features sensitive to patterns such as synchronization anomalies, thereby improving the model's ability to represent complex anomalies.

[0098] Specifically, the stage-aware gating weights are multiplied element-wise with the input embedded features to achieve dynamic scaling and modulation of the features. Simultaneously, the modulation feature tensor is obtained by combining cross-strand interaction features. This tensor integrates the stage-specific features of a single strand with the correlation features between multiple strands, and is represented as follows:

[0099]

[0100] In the formula, Represents the modulation feature tensor, with dimension . The characteristics of a single steel strand modulated during the fusion stage and the interaction characteristics across steel strands; The phase-aware gating weight tensor has a dimension of . Its element values ​​are between 0 and 1, and are used to dynamically scale the importance of each feature channel in the enhanced embedding tensor. This represents the element-wise multiplication operator; Denotes the cross-strand interaction feature tensor, with dimension . This is obtained by mining the statistical relationships between steel strands, and is used to enhance the model's ability to perceive abnormal patterns among multiple steel strands.

[0101] (1) Perform stage-aware gating weight tensor calculation. Dynamic gating weights are generated by utilizing stage label information. Specifically, the stage label is first mapped to the stage code, and then interacts with the current embedded features. A gating weight tensor with the same shape as the embedded features is generated through linear transformation and Sigmoid activation function. This determines the degree to which the original features should be enhanced or suppressed at each time point, each steel strand, and each feature channel. It is expressed as:

[0102]

[0103] In the formula, This represents the Sigmoid activation function, which maps the input to the range of 0 to 1, forming effective gating coefficients. This represents the first linear transformation layer, for the augmented embedding tensor. The transformation is performed, and its output dimension is ; This represents the second linear transformation layer, encoding the stage tensor. The transformation is performed, and its output dimension is ; This represents element-wise multiplication. The stage encoding tensor is represented by the stage flag feature tensor. It is obtained through a learnable embedding layer mapping, with a dimension of ; The dimension representing the stage encoding is an adjustable hyperparameter. =64; This represents the threshold adjustment gating bias term, which is a trainable parameter used to adjust the threshold for gating calculation.

[0104] In the specific implementation process, the phase coding tensor From the stage marker feature tensor The stage-specific feature tensor is obtained through learnable embedding layer mapping. Each element is a discrete stage value Through a learnable embedding layer, each stage value is mapped to... A vector of dimension, where the learnable embedding layer uses a learnable embedding matrix, where each row corresponds to a stage, for For each element in the array, find the embedding vector of the corresponding row to obtain the stage coding tensor. , dimension .

[0105] It should be noted that the linear transformation layer is the same as the fully connected layer in a neural network.

[0106] (2) Perform cross-strand interaction feature tensor calculation. To model the relationship between steel strands, calculate global statistics along the steel strand dimension and compare and fuse them with the local features of each steel strand. Specifically, first calculate the mean, standard deviation and range along the steel strand dimension, then broadcast these statistics along the steel strand dimension and concatenate them with the original embedded features of each steel strand. Finally, compress the concatenated high-dimensional features back to the original hidden dimension through a linear transformation layer. The resulting interaction feature is the cross-strand interaction feature tensor. ;

[0107] The definition of global statistics includes: the mean statistic tensor. is the mean statistic tensor calculated along the steel strand dimension, with dimension . This is used to characterize the average state of all steel strands, and the calculation method is expressed as follows: Standard deviation statistic tensor , is the standard deviation statistic tensor calculated along the steel strand dimension, with dimension . It is used to characterize the degree of dispersion between steel strands, and the calculation method is expressed as follows: Range statistic tensor , is the range statistic tensor calculated along the steel strand dimension, with dimension . This is used to characterize the maximum difference between steel strands, and the calculation method is expressed as follows: ;in, This indicates the operation of calculating the mean along the first dimension, which corresponds to the strand index dimension; This indicates the operation of calculating the standard deviation along the first dimension, which corresponds to the steel strand index dimension; This indicates the operation of calculating the maximum value along the first dimension, which corresponds to the strand index dimension; This indicates the operation of calculating the minimum value along the first dimension, which corresponds to the steel strand index dimension.

[0108] The cross-strand interaction feature tensor is calculated through feature synthesis and expressed as:

[0109]

[0110] In the formula, This indicates a broadcast copy operation performed in the steel strand dimension, causing the shape of the statistical tensor to change from... Become In order to enhance the embedding tensor To splice; This indicates a splicing operation, specifically splicing along the last dimension; This represents the third linear transformation layer, used to project the concatenated high-dimensional features back to the feature dimensions of the hidden layer. .

[0111] S2.3, Artificial intelligence model realizes anomaly identification in the prestressing tensioning process

[0112] The artificial intelligence model used is a multi-stage spatiotemporal anomaly detector. By employing a physical consistency feature reconstruction module and constructing a cascaded spatiotemporal attention mechanism, it simultaneously decouples and enhances the sensitivity to temporal dynamic anomalies, spatial correlation anomalies, and stage heterogeneity anomalies.

[0113] S2.3.1 Before using a multi-stage spatiotemporal anomaly detector, in order to inject prior physical knowledge in the field of prestressing tension and construct advanced features sensitive to specific anomalies, a physical consistency feature reconstruction tensor is constructed based on the original prestressing tension data and the stage marker feature tensor, thereby quantifying the deviation between actual observations and ideal physical models or engineering constraints.

[0114] (1) Definition The theoretical-actual displacement residual feature is used to sensitively capture elongation deviations caused by abnormal friction. The theoretical-actual displacement residual feature calculates the normalized residual between the actual elongation of each steel strand at each time step and the theoretically expected elongation based on the current pressure and the target pressure for the stage, expressed as:

[0115]

[0116] In the formula, This represents the real-time elongation characteristics of the original prestressing tension data, corresponding to the original prestressing tension data. Index 1 is displacement. Indicates the first The sample, the first steel bundle, first Real-time elongation at each time step; This represents the real-time hydraulic pressure characteristics of the original prestressing tension data, corresponding to the original prestressing tension data. Index 0 is pressure. Indicates the first The sample, the first steel bundle, first Real-time hydraulic pressure at each time step; This represents the target pressure characteristics of the original prestressing tension data, corresponding to the original prestressing tension data. Index 4 is target_pressure. Indicates the first The sample, the first steel bundle, first The pressure of achieving targets at each time step; Indicates the first The sample, the first steel bundle, first The theoretical-actual displacement residual at each time step directly quantifies the deviation between the actual elongation and the theoretical elongation calculated based on the current oil pressure and the target pressure of the stage, and is highly sensitive to frictional anomalies. The pressure-displacement conversion coefficient related to the stage represents the theoretical elongation that can be generated by a unit oil pressure under the current stage. This coefficient is obtained by performing linear regression analysis on historical normal data in stages. The target pressure bias coefficient, which is related to the stage, is used to correct the influence of the target pressure on the theoretical elongation. For example, in the holding stage, the target pressure remains constant, and its contribution to the change in elongation is different from that in the loading stage. This coefficient is obtained by performing linear regression analysis on historical normal data in stages. The normalization factor for the elongation residuals related to the stage is the standard deviation of the theoretical-to-actual displacement residuals of all normal samples in the training set corresponding to the stage. It is used to standardize the residuals so that they are comparable across different stages.

[0117] It should be noted that the theoretical-actual displacement residual characterizes whether the pressure-displacement relationship deviates from the expected physical model. A positive value indicates that the actual elongation is greater than the theoretical value, which may mean that the channel friction is less than expected or that there is slippage. A negative value indicates that the actual elongation is insufficient, which may mean that the friction is too large or the channel is blocked.

[0118] (2) Definition The symmetrical pressure equilibrium residual feature is used to directly enhance the perception of synchronous anomalies. This feature not only considers the pressure difference in the original prestressing tension data but also introduces the geometric symmetry relationship implicit in the circumferential position encoding. It calculates the pressure difference between each steel strand and its theoretically symmetrical strand, expressed as:

[0119]

[0120] In the formula, Indicates the first The sample, the first steel bundle, first Real-time hydraulic pressure at each time step; Indicates the relationship with the first The set of indices of adjacent steel bundles in the circumferential direction, such as one bundle on each side, is used to calculate the local average pressure as a reference. Represents the index of adjacent steel strands, and iterates through the set of steel strand indices. Elements in; Indicates the first The sample, the first steel bundle, first The symmetrical pressure equilibrium residuals at each time step amplify the asymmetric anomaly signal by combining the penalty term of absolute pressure difference and angle deviation, and are highly sensitive to synchronization anomalies and spatial asymmetry. Indicates the first The angle of the theoretically symmetrical steel strand; Indicates the first Index of theoretically symmetrical steel strands.

[0121] In the specific implementation process, since the steel strands are uniformly and symmetrically arranged in the circumferential direction, for the first... Steel strands, their theoretically symmetrical steel strand index It can be determined through circumferential geometric relationships, for example, assuming... If the number is even and the arrangement is symmetrical, then The corresponding angle That is, they are 180 degrees apart.

[0122] It should be noted that, The item introduces a penalty for symmetrical angle deviation. The larger the deviation between the actual symmetrical steel bundle and the theoretical symmetrical position, the larger the absolute value of this value, thereby amplifying the asymmetrical abnormal signal. The symmetrical pressure balance residual characterizes the symmetry and balance of pressure distribution among multiple steel bundles. The larger the value, the worse the pressure balance between the steel bundle and its symmetrical steel bundle, or the larger the deviation between the actual symmetrical relationship and the theoretical geometric symmetrical position.

[0123] (3) The reconstructed physical features and Concatenate along the new feature dimension and project it through a lightweight linear layer to the same dimension as the model's hidden layers. The physical consistency feature reconstruction tensor is obtained. , dimension ;

[0124] In the specific implementation process, the reconstructed physical characteristics and The two scalar features are respectively expanded to have the shape of The tensor is then concatenated along the last dimension to obtain a shape of... The tensor is then projected through a linear layer onto... dimension.

[0125] S2.3.2 The core of the multi-stage spatiotemporal anomaly detector consists of three cascaded sub-modules used to process the modulation feature tensor. Reconstructing the tensor with physical consistency features Deep spatiotemporal modeling is performed, as follows:

[0126] (1) The first sub-module of the multi-stage spatiotemporal anomaly detector is the time dynamic extraction module:

[0127] modulated feature tensor Reconstructing the tensor with physical consistency features The fusion is performed by adding elements one by one to obtain the fused feature tensor. , dimension The modulation characteristics and physical prior characteristics of the fusion stage;

[0128] Then, the fused feature tensor The input is fed into a stage-gated temporal convolutional network, which does not use fixed convolutional kernels but instead utilizes stage encoding tensors. Dynamically generating convolution weights allows the temporal feature extraction process to adapt to the dynamic characteristics of different tensioning stages, as expressed as:

[0129]

[0130] In the formula, The output tensor of the stage-gated temporal convolutional network has a dimension of . This characterizes the evolution and dynamic characteristics of each steel strand over time. This represents a stage-gated temporal convolutional network, the core of which is a one-dimensional convolutional layer, but its kernel weights are not fixed, but are determined by the stage encoding tensor. By dynamically generating a small linear layer of a neural network, the convolution operation can flexibly adjust its filtering characteristics according to the current tensioning stage. This represents the modified linear unit activation function; This indicates that the convolution kernel weights are encoded by the stage tensor. Dynamically generated one-dimensional convolution.

[0131] In the specific implementation process This project implements dynamic modulation of convolution kernel parameters based on stage information. Specifically, it first modulates the kernel parameters of dimension 1 to 2. Stage encoding tensor An averaging aggregation operation is performed at each time step to obtain the stage context vector for each sample-stem strand. Then, a specific one-dimensional convolutional kernel weight matrix corresponding to each sample-stem strand is generated through a fully connected layer. Finally, the dynamically generated convolutional kernel is used to fuse the feature tensors. One-dimensional convolution operation is performed on each steel strand and each feature channel.

[0132] (2) The second sub-module of the multi-stage spatiotemporal anomaly detector is the spatial relationship Transformer module:

[0133] The output tensor of the stage-gated time convolutional network In the strand dimension, it is viewed as a sequence, that is, each time step has a set of strand features. A multi-head self-attention mechanism is used to model the global dependencies between all strands and capture spatial patterns such as synchronization anomalies, represented as:

[0134]

[0135]

[0136] In the formula, This represents the output tensor after multi-head self-attention and residual connections, with dimension . This characterizes the mutual constraints and synergistic state of the steel strand group in the spatial dimension; This represents the output tensor after passing through the feedforward network and residual connections, with dimension . It represents the advanced spatiotemporal features after deep nonlinear transformation; The representation layer normalization operation normalizes the feature vector of each sample at each time step - the position of the steel strand; This represents a multi-head self-attention mechanism, which allows the model to focus on information from different locations in different representation subspaces. This represents a feedforward network, consisting of two linear layers and an activation function, specifically a modified linear unit activation function.

[0137] (3) The third sub-module of the multi-stage spatiotemporal anomaly detector is the stage-aware classification head module:

[0138] First, the output tensor after passing through the feedforward network and residual connection... Applying adaptive max pooling along the time dimension, the most critical signs for each steel strand during the entire tensioning process are extracted, yielding the output tensor of the adaptive max pooling. , dimension Extract the most important features of each steel strand over the entire time dimension;

[0139] Then, based on the output tensor of adaptive max pooling, the tensor reconstructed from physical consistency features, and the stage coding tensor, the anomaly class probability distribution is output through a stage-conditional fully connected layer, as follows:

[0140]

[0141] In the formula, This represents the probability distribution of the anomaly categories predicted by the model, with dimension 1. It consists of probability estimates for four categories; This represents the Softmax function, which maps the output of the fully connected layer to probability distributions for each category, such that the sum of the probabilities is 1. This represents a linear classification layer with an output dimension of . ; The total number of categories includes four categories: normal, friction abnormality, synchronization abnormality, and equipment structure abnormality. This indicates a max pooling operation. Term representation reconstructs the tensor from physical consistency features Perform max pooling operation over time, with an output dimension of... ; This indicates an averaging operation. Item representation pairs stage encoding tensor Averaging operations are performed along the time and tendon dimensions to obtain the stage context vector for each sample, with dimensions of [missing information]. Then broadcast in the steel strand dimension as The dimensions are used for splicing.

[0142] It should be noted that stage conditionalization refers to the modulation of the parameters or inputs of the classification layer by stage information. For stage-conditional fully connected layers, the input is concatenated with stage codes, so that the parameters of the fully connected layer make classification decisions based on both feature content and stage context during learning.

[0143] S2.4 Loss function calculation based on anomaly pattern comparison

[0144] The composite supervised loss function used consists of two parts: cross-entropy loss and anomalous pattern contrast loss, thereby improving the model's robustness in complex, ambiguous, and anomalous scenarios. Specifically,

[0145] Define the total loss function The calculation method is expressed as ;

[0146] in, This is a balancing hyperparameter of the loss function, used to adjust the strength of the contrastive learning term; an example value is 0.3. This represents the cross-entropy loss, which is the negative log probability of the true label calculated for each sample in the batch; The abnormal pattern contrast loss represents the ability to narrow the distance between samples of the same class in the feature space and widen the distance between samples of different classes, thereby enhancing the model's ability to distinguish abnormal patterns.

[0147] In the specific implementation process, the loss of abnormal patterns is compared. The feature tensor output by the spatial relation Transformer module is applied, specifically, the output tensor that has passed through the feedforward network and residual connections is first processed... Average pooling is performed along both the time and steel strand dimensions to obtain the global feature vector for each sample. The first... The global feature vector of each sample is Dimension Then, within the batch, the abnormal pattern contrast loss is calculated and expressed as:

[0148]

[0149] In the formula, This represents the cosine similarity calculation function, used to measure the similarity between two feature vectors; Indicates the batch and the first The set of indices of all other samples that have the same true label as a given sample is denoted as ; Indicates that, except for the first in the batch The set of indices of all other samples besides the sample itself, denoted as ; Indicates that the index belongs to the set. The sample index, i.e., the positive sample index; This represents a logarithmic function, with the default base being the natural constant. It is a natural exponential function; Indicates the first The global feature vector of each sample; Indicates the first The global feature vector of each sample; Indicates the first The global feature vector of each sample; Indicates belonging to the index set The sample index, i.e., the negative sample index; This represents the smoothness control coefficient, used to adjust the smoothness of the similarity distribution. =0.07.

[0150] S2.5 Model Training and Iterative Updates

[0151] Model training and iterative updates are a cyclical optimization process that aims to learn model parameters in a data-driven manner so that they can accurately identify anomalies in the prestressing tensioning process.

[0152] At the start of training, all trainable parameters of the constructed multi-stage spatiotemporal anomaly detector are initialized. In each iteration (i.e., a training step), a batch of data and its corresponding real labels are randomly sampled from the training dataset and input into the model. The model performs forward propagation, sequentially going through composite embedding and spatiotemporal location encoding, multi-stage dynamic feature modulation and anomaly-sensitive feature mining, and the core computation of the artificial intelligence model. Finally, it outputs the predicted probability distribution of each sample belonging to four categories. Then, based on the model prediction and the real labels, a composite supervised loss function is calculated as the total loss function. This total loss is obtained by weighted summation of cross-entropy loss and anomaly pattern contrast loss, which simultaneously measures classification accuracy and the clustering of similar samples and the separation of dissimilar samples in the feature space.

[0153] In the specific implementation process, backpropagation is performed based on the calculated total loss, and the gradient descent algorithm is used to update the model parameters. Specifically, the Adam optimizer is used to calculate the gradient of the loss function with respect to each model parameter. Then, the parameters are iteratively updated according to the optimizer rules and the preset learning rate to minimize the loss function.

[0154] Training will not continue indefinitely; explicit stopping conditions must be set to prevent overfitting. Specifically, the dataset is divided into training, validation, and test sets. After each complete training cycle (i.e., one training period), the validation set is used to evaluate the current model's performance, calculating metrics such as loss and classification accuracy on the validation set.

[0155] The stopping condition for iterative updates is based on an early stopping strategy for validation set performance: when the validation set loss no longer decreases over multiple consecutive training epochs, or the validation set accuracy no longer improves, it is considered that the model performance has reached saturation. Continuing training may lead to overfitting of the training set. At this point, the training iteration is stopped, and the model parameters are rolled back to the snapshot corresponding to the best validation set performance, which is taken as the final trained model.

[0156] In a specific implementation, S3 is as follows:

[0157] After training and optimization, the multi-stage spatiotemporal anomaly detector can be put into practical application to achieve anomaly identification in the prestressing tensioning process.

[0158] In practical applications, for the tensioning process to be identified (which can be real-time tensioning construction or historical data to be evaluated), the data processing flow is first followed exactly as in the training phase: the original time-series data is truncated or filled according to the same rules to form a fixed-length time window of 600 seconds, and the same 12 features are calculated to construct a four-dimensional input tensor of shape (1,4,600,12); at the same time, the physical feature reconstruction coefficients obtained in advance from historical normal data are loaded; then, the input tensor is fed into the trained model, which will automatically execute the complete processing flow, including feature semantic embedding and spatiotemporal location encoding, stage-aware feature modulation, physical consistency feature reconstruction, and cascaded spatiotemporal attention modeling. Finally, the stage-aware classification head module of the model will output a 4-dimensional probability vector, representing the confidence probability of the input tensioning process being judged as "normal", "friction abnormal", "synchronization abnormal", and "equipment structure abnormal".

[0159] The results of anomaly identification are analyzed based on the confidence probability distribution. Specifically, the category with the highest probability value is selected as the final identification result of the model. For example, if the probability value corresponding to "friction anomaly" is the highest, the system determines that there may be a problem of abnormal duct friction in this tensioning process. This builds a human-machine collaborative intelligent monitoring and diagnosis system to improve the quality control and safety level of prestressed tensioning construction.

[0160] Example 2

[0161] like Figure 2 and Figure 3 As shown, this paper analyzes the impact of hidden layer dimension hyperparameters on model performance to determine the optimal configuration and analyze the robustness of the method. The hidden layer dimension controls the expressive power of the model. This experiment compares the performance of the proposed method, the standard transformer model, and the Long Short-Term Memory (LSTM) time series model under different hidden layer dimensions, with multiple dimensionality settings ranging from small to large. Evaluation metrics include anomaly detection accuracy (…). Figure 2 ) and F1 score ( Figure 3 Precision represents the proportion of correctly identified samples, while the F1 score combines precision and recall; a higher value indicates that the model is both accurate and comprehensive in identifying anomalies. Figure 2 and Figure 3 The results showed that setting the hidden layer dimension to 256 resulted in higher anomaly detection accuracy and F1 score.

[0162] Example 3

[0163] like Figure 4As shown, the multi-stage dynamic feature modulation effect was verified, validating the ability of this method to dynamically adjust feature importance during multi-stage tensioning and revealing the working principle of the stage-aware feature modulation mechanism. The experiment simulated four key stages of the entire prestressing tensioning process: initialization stage, 10% loading stage, 100% loading stage, and holding stage, spanning 600 seconds and covering the complete tensioning cycle. The experiment compared the stage-aware feature modulation mechanism of this method with the fixed feature weight strategy of the standard method. Based on... Figure 4 The curves showing the feature gating weights over time reveal that this method can dynamically adjust the focus on the three core features according to the tensioning process. The weight of the real-time elongation feature rises rapidly during the 10% loading stage and peaks at the 100% loading stage, perfectly coinciding with the sensitive period of friction anomalies. The weight of the symmetrical pressure difference feature significantly increases during the 100% loading stage, accurately corresponding to the critical window for synchronous anomaly detection. The weight of the pressure change rate feature peaks during stage transitions, effectively capturing transient processes that may lead to equipment anomalies. In contrast, the fixed weight strategy of the standard method cannot adapt to the changing importance of features at different stages, resulting in insufficient sensitivity to anomaly signals in critical time windows. The time axis is in seconds, reflecting the actual timescale of the tensioning process; the gating weight axis is a unitless normalized value, representing the relative importance of feature channels.

[0164] Example 4

[0165] like Figure 5 As shown, the anomaly detection performance of the method of the present invention at different tensioning stages was evaluated to verify the effectiveness of its stage perception capability. The tensioning process includes four stages: initialization, 10% loading, 100% loading, and holding load. The normal operation mode and anomaly characterization at different stages are significantly different. Figure 5 The horizontal axis represents the false positive rate, indicating the proportion of normal samples misclassified as abnormal; the vertical axis represents the true positive rate, indicating the proportion of abnormal samples correctly identified; both are dimensionless proportional values. The closer the area under the curve is to 1, the better the model performance. Figure 5 shows the curves of the method of this invention in four stages, with the curve in the 100% loading stage being closest to the upper left corner and having the largest area under the curve, indicating the best performance; other stages also maintain a high area under the curve, proving that the stage-aware feature modulator can dynamically adjust the importance of features according to stage information, thereby improving the detection accuracy in key stages.

[0166] 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 artificial intelligence-based method for anomaly identification in the prestressing tensioning process, characterized in that, Includes the following steps: S1. Four steel strands are symmetrically arranged in the circumferential direction of the main tower of the bridge and simultaneously tensioned. Each tensioning device of the steel strand is equipped with a sensor network to collect data synchronously at a fixed frequency. This data is used to construct training samples and to label the samples with abnormal patterns. At the same time, historical normal data is analyzed to obtain prior knowledge parameters. S2. Construct an anomaly recognition model. Input sample data into the model. After composite embedding and spatiotemporal location encoding of tension time series data, multi-stage dynamic feature modulation and anomaly sensitive feature mining module, and multi-stage spatiotemporal anomaly detector, obtain the anomaly recognition result. Calculate the total loss function of the model and train the anomaly recognition model. S2 is as follows: S2.1 Input the sample data into the anomaly identification model, perform composite embedding and spatiotemporal position encoding of tension time series data, specifically perform feature semantic embedding calculation, absolute time position encoding calculation and steel strand circumferential position encoding calculation, and then add and fuse the calculation results to obtain the enhanced embedding tensor; S2.2 Input the enhanced embedding tensor into the multi-stage dynamic feature modulation and anomaly-sensitive feature mining module. By applying stage-aware dynamic gating to the enhanced embedding tensor and simultaneously fusing the statistical relationship features across steel strands, the modulation feature tensor is obtained. S2.2 is as follows: S2.2.1 Perform stage-aware gating weight tensor calculation. First, map the stage label to the stage code, and then interact with the current augmented embedding tensor. Generate a stage-gating weight tensor that is isomorphic to the augmented embedding tensor through linear transformation and Sigmoid activation function. Among them, the stage coding tensor is obtained by mapping the stage marker feature tensor through a learnable embedding layer. The stage marker feature tensor is obtained from the input sample data and includes four tensioning stages: initialization, 10% loading, 100% loading, and holding load. S2.2.2 Perform cross-strand interaction feature tensor calculation. First, calculate the mean, standard deviation and range along the strand dimension. Then, broadcast these statistics along the strand dimension and concatenate them with the enhanced embedding tensor of each strand. Finally, compress the concatenated high-dimensional features back to the original hidden dimension through a linear transformation layer to form the cross-strand interaction feature tensor. S2.2.3 Multiply the stage-aware gating weights element-wise with the input embedded features, and combine them with the cross-strand interaction features to obtain the modulation feature tensor. S2.3 Construct a physical consistency feature reconstruction tensor based on prior knowledge parameters. The multi-stage spatiotemporal anomaly detector performs deep spatiotemporal modeling based on the physical consistency feature reconstruction tensor and the modulation feature tensor, and outputs the anomaly category probability distribution. The construction process of the S2.3.1 Physical Consistency Feature Reconstruction Tensor is as follows: (1) Calculate the normalized residual between the actual elongation and the theoretical expected elongation of each steel strand at each time step to obtain the theoretical-actual displacement residual characteristics; (2) Based on the circumferential symmetry of the steel strands, the pressure difference between each steel strand and the symmetrical steel strands is calculated to obtain the symmetrical pressure equilibrium residual characteristics; (3) The two types of residual features are spliced ​​together and projected onto the same dimension as the hidden layer of the model through a lightweight linear layer to obtain the physical consistency feature reconstruction tensor; S2.4 Calculate the total loss function including cross-entropy loss and outlier pattern comparison loss; S2.

5. Train the model using backpropagation based on the total loss function until the iteration stopping condition is met; S3. Deploy a sensor network to collect relevant data during the tensioning process of the target to be identified, call the physical feature reconstruction coefficients from the training phase, input the collected data into the trained anomaly identification model, and output the anomaly classification results after analysis and processing.

2. The artificial intelligence-based anomaly identification method for prestressed tensioning process according to claim 1, characterized in that, The sample construction process is as follows: Complete tensioning process records are extracted from historical engineering data. Complete tensioning cycles are extracted at fixed time lengths to form four-dimensional tensor samples of fixed duration. The four-dimensional tensors are the corresponding batch dimension, steel strand dimension, time dimension, and feature dimension. Among them, the batch dimension represents an independent sample, the steel bundle dimension corresponds to four symmetrically arranged steel bundles, the time dimension represents a fixed time length, and the feature dimension represents the feature value obtained from the data collected by the sensor network. The sensor network includes hydraulic sensors, displacement sensors, temperature sensors, and concrete strain gauges arranged below the anchorage, collecting data from the start of tensioning until the end of the holding phase.

3. The artificial intelligence-based anomaly identification method for prestressed tensioning process according to claim 1, characterized in that, The sample annotation process is as follows: The annotation method, which is based on the fusion of expert experience and multi-source information, first evaluates the samples manually based on the tensioning process curve, construction log and subsequent test report, and then uses the preset engineering threshold rules to screen automatic alarm records to assist in annotation. Each sample is assigned a unique category label. The label categories correspond to the abnormal patterns to be identified, including normal, friction abnormality, synchronization abnormality, and equipment mechanism abnormality.

4. The artificial intelligence-based anomaly identification method for prestressed tensioning process according to claim 1, characterized in that, S2.1 is as follows: Construct a composite embedded module to perform the following calculations: (1) The feature semantic embedding calculation adopts a stage-aware feature transformation mechanism. Based on the tensor stage label, learnable weights are dynamically selected. Each feature is weighted and projected, and then nonlinear transformation and bias addition are performed to adjust the dimension to obtain the feature semantic embedding tensor. (2) The absolute time position coding calculation specifically adopts the decoupled two-dimensional coding of the learnable parameter matrix to directly encode the order of each time step and obtain the absolute time position coding tensor; (3) The calculation of the circumferential position coding of the steel strand adopts the decoupled two-dimensional coding of sine and cosine functions. Based on the symmetrical geometric relationship of the circumferential angle position coding of the steel strand, the circumferential position coding tensor of the steel strand is obtained. The enhanced embedding tensor is obtained by superimposing the above three types of tensors.

5. The artificial intelligence-based anomaly identification method for prestressed tensioning process according to claim 4, characterized in that, The S2.3.2 multi-stage spatiotemporal anomaly detector includes a time dynamic extraction module, a spatial relationship Transformer module, and a stage-aware classification head module, as detailed below: (1) In the time dynamic extraction module, the modulation feature tensor and the physical consistency feature reconstruction tensor are first fused by adding elements to obtain the fused feature tensor; then the fused feature tensor is input into the stage-gated temporal convolutional network to obtain the output tensor of the stage-gated temporal convolutional network. (2) In the spatial relationship Transformer module, the output tensor of the stage-gated temporal convolutional network is regarded as a set of sequences in the steel strand dimension. The global dependency relationship between all steel strands is modeled by the multi-head self-attention mechanism. Specifically, the state tensor is obtained through the multi-head self-attention mechanism and residual connection. The state tensor is then output as a high-level spatiotemporal feature tensor through the feedforward network and residual connection. (3) In the stage-aware classification head module, adaptive max pooling is first applied to the high-level spatiotemporal feature tensor along the time dimension to extract the most critical signs of each steel strand in the entire tensioning process, and the output tensor of adaptive max pooling is obtained; then, based on the output tensor of adaptive max pooling, the tensor reconstructed by physical consistency features and the stage coding tensor, the probability distribution of abnormal categories is output through the stage-conditional fully connected layer.

6. The artificial intelligence-based anomaly identification method for prestressed tensioning process according to claim 5, characterized in that, S2.4 is as follows: The total loss function includes cross-entropy loss and outlier pattern contrast loss; The cross-entropy loss is calculated based on the negative log probability of the true label of each sample. For any sample in a batch, for all other samples with the same true label, calculate the global feature vector between them, then calculate the similarity between the global feature vectors of the function through cosine similarity, and then calculate the abnormal pattern contrast loss. The global feature vector of each sample is obtained by average pooling of the high-level spatiotemporal feature tensor in the time dimension and the steel strand dimension.

7. The artificial intelligence-based anomaly identification method for prestressed tensioning process according to claim 6, characterized in that, S2.5 is as follows: First, all trainable parameters in the multi-stage spatiotemporal anomaly detector are initialized. Then, the training data is input into the model. After forward propagation, the probability distribution of anomaly categories is obtained. The total loss function of the model is then calculated. Backpropagation is performed based on the total loss to minimize the total loss function until the iteration stopping condition is met and the training ends.

Images