Pipe-jacking construction pipeline settlement abnormity detection method

Abstract

The invention discloses a pipe-jacking construction pipeline settlement abnormity detection method. The method comprises the following steps: collecting monitoring data of a multi-source sensor; self-adaptive sliding binning normalization: using a sliding window with the current time step as the center, extracting a data set of a corresponding sensor in the window, and normalizing the original sensor reading of the current time step based on the minimum value and the maximum value of the data set; constructing a pipe-jacking construction pipeline settlement anomaly detection model; and pipe-jacking construction pipeline settlement anomaly detection: after model training is completed, the pipe-jacking construction pipeline settlement anomaly detection process is applied to a real-time or offline monitoring scene. According to the method, a self-adaptive sliding binning normalization method is adopted, the normalization range is dynamically adjusted according to local data distribution, the problem that traditional binning is fuzzy in feature at a sudden change point is solved, and the model is more sensitive to response to tiny settlement change in an abnormal time period; through parallel modeling of a long-range dependence path and a local mutation path, multi-scale feature capture of a complex settlement behavior is realized.

Description

Technical Field

[0001] This invention relates to a method for detecting abnormal settlement of pipelines, and more particularly to a method for detecting abnormal settlement of pipelines constructed by pipe jacking. Background Technology

[0002] Pipe jacking, as an important method for constructing urban underground pipe networks, is widely used in complex geological conditions and high-density urban environments. However, during the pipe jacking process, the pipeline is highly susceptible to uneven settlement due to various factors such as construction disturbance, geological heterogeneity, changes in groundwater pressure, and fluctuations in construction parameters. If abnormal settlement is not detected and addressed in a timely manner, it may lead to joint leakage, pipe section misalignment, or even structural damage, resulting in serious safety accidents such as ground subsidence and settlement of surrounding buildings. Therefore, real-time and accurate anomaly detection of the settlement status of pipe jacking construction pipelines has significant engineering importance and safety assurance value.

[0003] Currently, multi-source sensor monitoring systems are commonly used on engineering sites for settlement monitoring, collecting data in real time through devices such as inclinometers, strain gauges, and earth pressure gauges. However, this sensor data is characterized by significant non-stationarity, high noise levels, irregular abrupt changes, and large dimensional differences, making traditional monitoring data processing and anomaly identification methods difficult to effectively adapt. Specifically, existing technologies have the following main shortcomings:

[0004] 1) Existing technologies mostly use fixed window or global normalization methods, which cannot cope with the non-stationary characteristics of settlement monitoring data, resulting in the weakening of abnormal signals after normalization and insufficient detection sensitivity.

[0005] 2) Conventional single-path neural networks (such as LSTM or CNN) have difficulty capturing both long-term trend changes and instantaneous mutation features at the same time, resulting in a weak ability to jointly model gradual settlement and sudden anomalies.

[0006] 3) Existing feature fusion methods typically use simple splicing or fixed weights, lacking an adaptive adjustment mechanism. They cannot flexibly balance the contributions of global and local features according to different working conditions, thus affecting detection accuracy.

[0007] 4) Conventional anomaly detection methods lack the distinction between the importance of time and sensor dimensions, and do not consider the uncertainty assessment of prediction results, which can easily lead to false alarms, false negatives, and the inability to quantify confidence. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to propose a method for detecting abnormal settlement of pipelines in pipe jacking construction. This method can dynamically adjust the normalization range according to the local data distribution, realize the multi-scale feature capture of complex settlement behavior, and improve the accuracy and robustness of anomaly identification.

[0009] Technical solution: This invention includes the following steps:

[0010] S1. Multi-source sensor monitoring and data acquisition;

[0011] S2, Adaptive sliding bin normalization;

[0012] S3. Construct a model for detecting abnormal settlement of pipelines during pipe jacking construction, specifically including:

[0013] S31. Construct a dual-path feature extraction module: By adopting a dual-path parallel structure, global temporal patterns and local mutation features are extracted separately to separate global trends and local jitter.

[0014] S32. Feature Fusion Gating Mechanism: By adopting a gating mechanism to adaptively fuse dual-path outputs, the weights of long-range dependent paths and local mutation paths are dynamically allocated using learnable gating vectors, so that the model focuses on the global trend in the stable segment and enhances the local response in the mutation segment.

[0015] S33, Enhanced attention to sedimentation sensitivity perception: By adopting a sensor time-dimension dual-dimensional attention mechanism, the importance is evaluated in both the time dimension and the sensor dimension, focusing on the abnormal signals of key sensors at critical moments;

[0016] S34. Calculate the loss function: By combining weighted cross-entropy and outlier margin loss, class weights are used to alleviate sample imbalance, while forcing outlier classes and normal classes to be significantly separated in the feature space.

[0017] S35. Anomaly Confidence Assessment Output: The confidence level is calculated based on the predicted probability distribution. The relative probability of the anomaly class and the predictive certainty are considered together. The reliability of anomaly detection is quantified by using a dual assessment mechanism.

[0018] S36. Model Training and Parameter Update: The settling-sensitive loss is minimized by using an adaptive moment estimation optimizer, the generalization ability is improved by combining weight decay, and gradient pruning is used to prevent gradient explosion and oscillation during training.

[0019] S37. Model Iterative Training and Parameter Update: The model iterative training process involves traversing the training dataset multiple times, using an adaptive moment estimator to minimize the settlement-sensitive loss function, and gradually updating the model parameters.

[0020] S4. Detection of abnormal settlement of pipelines during pipe jacking construction: After the model training is completed, the process of detecting abnormal settlement of pipelines during pipe jacking construction will be applied to real-time or offline monitoring scenarios.

[0021] S2 uses a sliding window centered on the current time step to extract the data set of the corresponding sensor within the window, and normalizes the original sensor readings of the current time step based on the minimum and maximum values ​​of the data set.

[0022] S31 specifically includes:

[0023] S311, Long-range dependency path feature extraction: The forward-gated recurrent unit is used to process the normalized value vector sequence to generate the hidden state that captures the forward temporal dependency. At the same time, the backward-gated recurrent unit is used to process the normalized value vector sequence to generate the hidden state that captures the backward temporal dependency. The forward and backward hidden states at the same time step are concatenated to form the output feature vector of the long-range dependency path.

[0024] S312. Local mutation path feature extraction: Using dilated causal convolution kernels, the normalized value vector sequence is convolved. The modified linear unit activation function is applied to the convolution result to generate the output feature vector of the local mutation path. The receptive field is expanded by dilated convolution while maintaining temporal causality.

[0025] Specifically, S32 includes:

[0026] S321. Calculate the feature fusion gating vector: concatenate the output feature vector of the long-range dependent path and the output feature vector of the local mutation path, perform a linear transformation on the concatenated features through the gating weight matrix and the gating bias vector, and apply the Sigmoid activation function to the linear transformation result to generate the fusion gating vector.

[0027] S322. Perform gated feature fusion: Use the fusion gate vector to weight the output feature vector of the long-range dependent path element by element, and at the same time, weight the output feature vector of the local mutation path element by element. Add the two weighted results to generate the final fusion feature vector.

[0028] Specifically, S33 includes:

[0029] S331. Calculate the attention weights in the time dimension: Apply a linear transformation to the fused feature vector at each time step using the time attention weight matrix and the time attention bias vector, and activate it using the hyperbolic tangent activation function. Calculate the dot product similarity between the activation result and the time attention query vector to obtain the unnormalized attention score. Apply an exponential function to the unnormalized scores of all time steps and normalize them to generate the time attention weights for each time step. Use the time attention weights to perform a weighted summation of the fused feature vectors at each time step to obtain the time context vector.

[0030] S332. Calculate sensor dimension attention weights: Apply a linear transformation to the feature slices of each sensor channel using the sensor attention weight matrix and the sensor attention bias vector, and activate them using the hyperbolic tangent activation function. Calculate the dot product similarity between the activation result and the sensor attention query vector to obtain the unnormalized attention score. Apply an exponential function to the unnormalized scores of all sensor channels and normalize them to generate the channel attention weights for each sensor. Use the channel attention weights to perform a weighted summation of the feature slices of each sensor to obtain the sensor context vector.

[0031] S333. Generate a two-dimensional attention context vector: Concatenate the temporal context vector and the sensor context vector to form a two-dimensional attention context vector.

[0032] S34 specifically includes:

[0033] S341. Calculate the class weight coefficient: Calculate the corresponding class weight coefficient based on the number of samples of each class in the training set;

[0034] S342. Calculate the weighted cross-entropy loss components: Use the class weight coefficient to weight the cross-entropy loss of each sample, and calculate the weighted cross-entropy loss components.

[0035] S343. Calculate the abnormal interval loss component: Calculate the feature center of the abnormal class in the embedding space and the feature center of the normal class in the embedding space, calculate the distance between the two feature centers, and generate the abnormal interval loss component when the distance is less than the set minimum class interval.

[0036] S344. Combined total loss function: The weighted cross-entropy loss component and the outlier interval loss component are added together to form the total loss function.

[0037] Specifically, S35 includes:

[0038] S351. Calculate the relative probability of the abnormal class: Divide the predicted probability of the abnormal class by the sum of the exponents of the predicted probabilities of all abnormal classes to calculate the relative probability component of the abnormal class.

[0039] S352. Calculate the entropy of the predicted probability distribution: Calculate the information entropy of the predicted probability distribution based on the probability distribution vector predicted by the model;

[0040] S353. Calculate the overall confidence score: Based on the results of the relative probability components of the anomaly class and the information entropy of the predicted probability distribution, generate an overall confidence score.

[0041] Specifically, S36 includes:

[0042] S361. Perform adaptive moment estimation optimizer parameter update: update the model parameter vector based on the bias correction vector of the first-order moment estimation and the bias correction vector of the second-order moment estimation, combined with the learning rate and weight decay coefficient.

[0043] S362. Implement gradient clipping constraints: Calculate the L2 norm of the gradient of the total loss function. If the norm exceeds the preset gradient clipping threshold If so, the gradient vector will be scaled proportionally to the threshold range.

[0044] In step S37, in each iteration, a batch of normalized sensor data sequences are input, and the outputs of dual-path feature extraction, feature fusion gating mechanism, and sedimentation-sensitive perception attention enhancement are calculated through forward propagation to obtain the predicted probability distribution; the total loss of weighted cross-entropy loss and abnormal interval loss is calculated, and the gradient is calculated through backpropagation.

[0045] In step S37, gradient clipping constraints are implemented, and the L2 norm of the gradient vector is checked to see if it exceeds a preset threshold. If it does, the gradient is scaled proportionally. Parameter updates are based on the rules of the adaptive moment estimator optimizer, combined with bias corrections from first-order and second-order moment estimations, to adjust the learning rate and apply weight decay.

[0046] Beneficial effects: The present invention has the following advantages:

[0047] (1) The adaptive sliding bin normalization method can dynamically adjust the normalization range according to the local data distribution, overcome the problem of feature ambiguity at the point of change in traditional equal width or equal frequency bins, and make the model more sensitive to small settlement changes during abnormal periods.

[0048] (2) Construct a dual-path feature extraction structure, and model the long-range dependent path and the local mutation path in parallel to effectively take into account the gradual trend and instantaneous mutation in the settlement process, and realize the multi-scale feature capture of complex settlement behavior.

[0049] (3) An adaptive feature fusion mechanism based on learnable gating vectors is adopted, which can dynamically adjust the weight distribution of global and local information according to the working conditions, so that the model can automatically focus on different feature sources in the stable and abnormal segments, thereby improving the accuracy and robustness of anomaly identification.

[0050] (4) By adopting a sensor time-dimensional attention mechanism and an anomaly confidence assessment module, it can focus on key risk signals in both time and sensor dimensions, and quantify the reliability of detection results through probability distribution entropy, thereby reducing the false alarm rate while improving the credibility of anomaly detection results. Attached Figure Description

[0051] Figure 1This is a comparison chart of the inclinometer data processing effects in an embodiment of the present invention;

[0052] Figure 2 This is a comparison chart of strain gauge data processing effects in embodiments of the present invention;

[0053] Figure 3 This is a comparison chart of the data processing effects of the earth pressure gauge 1 in this embodiment of the invention;

[0054] Figure 4 This is a comparison chart of the data processing effects of the earth pressure gauge 2 in this embodiment of the invention;

[0055] Figure 5 This is a diagram illustrating the dual-path feature extraction effect in an embodiment of the present invention.

[0056] Figure 6 This is a time-dimensional attention distribution diagram in an embodiment of the present invention;

[0057] Figure 7 This is a sensor dimension attention distribution diagram in an embodiment of the present invention;

[0058] Figure 8 This is a comparison chart of the confidence distribution of different methods in the embodiments of the present invention;

[0059] Figure 9 This is a performance comparison chart of different methods in pipeline settlement anomaly detection in the embodiments of the present invention. Detailed Implementation

[0060] The invention will now be further described with reference to the accompanying drawings.

[0061] Example 1

[0062] The method for detecting abnormal settlement of pipelines during pipe jacking construction in this embodiment includes the following steps:

[0063] S1, Multi-source sensor monitoring data acquisition

[0064] During pipe jacking construction, in order to comprehensively monitor the settlement status of the pipeline, a variety of sensors are deployed for real-time data acquisition, including key sensors such as inclinometers, strain gauges, and earth pressure gauges. These sensors are installed at key locations on the pipeline and continuously record data at a fixed sampling frequency. For example, inclinometers monitor changes in the pipeline's tilt angle, strain gauges measure the strain values ​​of the pipeline structure, and earth pressure gauges detect the pressure distribution of the surrounding soil.

[0065] The data acquisition process involved collecting historical sensor data from multiple pipe jacking construction projects, covering different geological conditions, construction stages, and working conditions to ensure data diversity and representativeness. The collected data was then labeled based on expert knowledge and on-site monitoring reports, categorizing the sensor data sequence at each time step into three classes: normal, warning, and abnormal. The normal class indicates that pipeline settlement is within a safe range with no significant risk; the warning class indicates that settlement is approaching a critical value or showing slight deviations, requiring enhanced monitoring; and the abnormal class indicates that settlement exceeds the safe threshold or shows drastic changes, indicating potential faults.

[0066] S2, Adaptive sliding bin normalization

[0067] Multi-source sensor monitoring data during pipe jacking construction, such as readings from inclinometers, strain gauges, and earth pressure gauges, are characterized by non-stationarity, significant local abrupt changes, and large differences in the dimensions of different sensors. Conventional processing methods use equal-width or equal-frequency bin normalization, but this is difficult to adapt to the abrupt changes in the distribution of settlement data near the abrupt change point, and the fixed boundary will blur the local anomaly features, resulting in insufficient detection capability for key changes.

[0068] This embodiment uses a sliding window centered on the current time step to extract the data set of the corresponding sensor within the window. Based on the minimum and maximum values ​​of this data set, the original sensor readings of the current time step are normalized to eliminate dimensional differences between different sensors, as shown below:

[0069]

[0070] In the formula, Indicates the first The first time step Normalized values ​​from each sensor are used to scale the raw data to... Intervals, eliminating dimensional differences; Indicates the first The first time step The raw readings of each sensor; Indicates the first Within the sliding window centered on the [time step] The data set from each sensor is used for local statistical calculations, dynamically adapting to the data distribution, and is represented as follows: ; This indicates the window length, which is an odd number. It controls the local statistical range and is set according to the sampling frequency and settlement response time. To represent extremely small positive numbers, prevent the denominator from being zero, and enhance numerical stability, it is preferred to set it to 0.0001; This represents the function that takes the minimum value of a set and is used to calculate the lower limit of data within a window. This represents the function to find the maximum value of a set, used to calculate the upper limit of data within a window; This indicates a function to round down, ensuring that the window index is an integer; Represents a set of positive integers, limiting the range of time step indices; Indicates the index of the current time step; Indicates the sensor channel index; This represents the index of the time step within the sliding window.

[0071] like Figures 1-4 As shown, this embodiment compares the effects of different sensor data processing methods. The experiment simulates the monitoring data of key sensors such as inclinometers, strain gauges, and earth pressure gauges in an actual construction environment over 200 time steps. Abnormal events are introduced near time steps 50, 120, and 180 to verify the advantages of the proposed adaptive sliding bin normalization method over conventional normalization and fixed-window sliding normalization. Four sensor types were configured in the experiment, and three different normalization methods were applied to each sensor. The ability of each method to retain abnormal signals was evaluated by comparing the processed data curves. Figures 1-4 As shown in the comparison charts of the four sensors, it can be observed that conventional normalization methods, due to the use of global maximum and minimum values, compress local abnormal signals, resulting in insignificant data changes during abnormal periods. While sliding window normalization can reflect local changes to some extent, it exhibits noticeable jumps at window boundaries and has limited enhancement of abnormal signals. In contrast, the adaptive normalization method of this invention, by dynamically adjusting the window size and combining it with local statistical characteristics, can significantly enhance abnormal signals during abnormal periods while maintaining the continuity of the data curve. Particularly during the period of anomaly occurrence, the curve of the method of this invention shows obvious peak changes, which is consistent with the abrupt changes in sensor readings under abnormal conditions in actual engineering. Experimental results show that adaptive sliding bin normalization can dynamically adjust the normalization parameters according to the local data distribution, effectively overcoming the limitations of conventional methods when processing non-stationary sensor data.

[0072] S3. Construct a model for detecting abnormal settlement of pipelines during pipe jacking construction, specifically including:

[0073] S31. Construct a dual-path feature extraction module

[0074] Settlement anomalies often manifest as a global trend deviation superimposed with local high-frequency fluctuations. Conventional single-path deep neural networks, such as fully connected networks or unidirectional recurrent neural networks, are unable to simultaneously capture long-range dependencies and instantaneous abrupt changes, resulting in insufficient ability to model both gradual settlement processes and sudden anomalies.

[0075] This embodiment employs a dual-path parallel structure to extract global temporal patterns and local abrupt change features separately, separating global trends from local fluctuations. This effectively addresses the joint modeling problem of long-range dependencies and transient abrupt changes, achieving more comprehensive feature capture of settlement anomalies. The specific steps are as follows:

[0076] S311, Long-range Dependency Path Feature Extraction

[0077] A forward-gated recurrent unit is used to process the normalized value vector sequence to generate hidden states that capture forward temporal dependencies. Simultaneously, a backward-gated recurrent unit is used to process the normalized value vector sequence to generate hidden states that capture backward temporal dependencies. Then, the forward and backward hidden states at the same time step are concatenated to form the output feature vector of the long-range dependency path, represented as:

[0078]

[0079]

[0080]

[0081] In the formula, Indicates the first The normalized vector of all sensors at each time step; Indicates the forward-gated loop unit in the first... The hidden state of a time step is used to capture forward temporal dependencies; This indicates that the backward-gated loop unit is in the first... The hidden state of a time step is used to capture backward temporal dependencies; Indicates the forward-gated loop unit in the first... The hidden state of a time step serves as the initial state of the current time step; This indicates that the backward-gated loop unit is in the first... The hidden state of a time step serves as the initial state of the current time step; The network weight parameters of the forward gated recurrent unit are trainable parameters, including the weights related to the update gate, reset gate, and candidate state. The network weight parameters of the backward gated recurrent unit are trainable parameters, including the weights related to the update gate, reset gate, and candidate state. This indicates that the long-range dependency path is at the 1st position. The output feature vector at each time step is composed of the concatenation of the forward and backward hidden states; This indicates a vector concatenation operation, used to combine forward and backward features; This represents the calculation function of the gated loop unit, which is used to extract temporal features.

[0082] S312, Local Mutation Path Feature Extraction

[0083] Using a dilated causal convolution kernel with a specific dilation rate, a normalized value vector sequence is convolved. Then, a modified linear unit activation function is applied to the convolution result to generate an output feature vector of local mutation paths. By expanding the receptive field and preserving temporal causality through dilated convolution, transient anomaly features are effectively captured, as shown below:

[0084]

[0085] In the formula, Indicates the local mutation path in the th order. The output feature vector at each time step represents the local mutation pattern; Indicates the first The input feature vector at each time step is sampled through dilated convolution; This indicates the number of convolutional kernels, controlling the complexity of feature extraction; This represents the inflation rate, which increases exponentially with the number of network layers to expand the receptive field and controls the sampling interval of the convolutional kernels. Indicates the first The weight matrix of each convolutional kernel is a trainable parameter used to extract specific local patterns. The bias vector representing the convolution operation is a trainable parameter that enhances the model's expressive power. This represents the modified linear unit activation function; Represents the index of an element within the convolution kernel, from arrive .

[0086] S32, Feature Fusion Gating Mechanism

[0087] In dual-path feature fusion, conventional processing methods use simple splicing or weighted summation, but it is difficult to dynamically adjust the contribution ratio of global and local information. Especially in scenarios where gradual and abrupt settlement coexist, fixed fusion strategies cannot adapt to the differentiated needs of different working conditions for long-range trends and instantaneous changes, resulting in poor feature fusion performance.

[0088] This embodiment employs a gating mechanism to adaptively fuse dual-path outputs. It dynamically allocates weights for long-range dependent paths and locally abrupt paths using learnable gating vectors, enabling the model to emphasize global trends in stable segments and enhance local responses in abrupt segments. This solves the problem of adaptive feature selection under conditions of coexisting gradual and abrupt changes. The specific steps are as follows:

[0089] S321. Calculate the feature fusion gating vector.

[0090] The output feature vectors of the long-range dependent path and the local mutation path are concatenated. The concatenated features are then linearly transformed using a gating weight matrix and a gating bias vector. A sigmoid activation function is applied to the linear transformation result to generate a fusion gating vector, thereby dynamically adjusting the fusion ratio of the two path features. This is expressed as:

[0091]

[0092] In the formula, Indicates the first The fusion gate vector at each time step takes values ​​in the range of . This is used to dynamically adjust the fusion ratio of dual-path features; This represents the Sigmoid activation function, which maps the input to... The interval is used to realize the nonlinear transformation of the gating mechanism; The gating weight matrix represents the trainable parameters used to learn the linear mapping relationship between dual-path features and gating vectors. The gating bias vector is a trainable parameter that enhances the expressive power of the gating mechanism.

[0093] S322, Execution Gating Feature Fusion

[0094] The output feature vectors of long-range dependent paths are weighted element-wise using a fusion gating vector, and the output feature vectors of locally abrupt paths are also weighted element-wise. The two weighted results are then added to generate the final fusion feature vector, thereby achieving an adaptive combination of the features from the two paths, as shown below:

[0095]

[0096] In the formula, Indicates the first The fused feature vector at each time step; This represents element-wise multiplication, which is used to assign weights to feature vectors. Indicates and A vector of all 1s with the same dimension.

[0097] S33, Enhanced sensitivity to settling and attention

[0098] Conventional attention mechanisms pay equal attention to all time steps and sensor channels, but sedimentation anomalies are often triggered by a few key sensors at specific moments. For example, a sudden change at a certain point can trigger a chain reaction, resulting in insufficient perception of weak anomaly features and difficulty in focusing on truly critical risk signals, thus affecting the accuracy of anomaly detection.

[0099] This embodiment employs a sensor-time dual-dimensional attention mechanism to evaluate importance in both the time and sensor dimensions, focusing on anomalous signals from key sensors at critical moments, thus significantly improving the ability to perceive weak anomalous features. The specific steps are as follows:

[0100] S331, Calculate the attention weights in the time dimension.

[0101] For each time step, the fused feature vector is linearly transformed using the temporal attention weight matrix and temporal attention bias vector, and then activated using the hyperbolic tangent activation function. The activation result is then multiplied by the temporal attention query vector to calculate the dot product similarity, yielding an unnormalized attention score. An exponential function is applied to and normalized the unnormalized scores for all time steps to generate temporal attention weights for each time step. These weights are then used to weight and sum the fused feature vectors for each time step to obtain the temporal context vector. This vector is then used to focus on the time steps important for anomaly detection, as shown below:

[0102]

[0103] In the formula, Indicates the first The temporal attention weights for each time step quantify the importance of that time step for anomaly detection; This represents the temporal attention query vector, used for similarity calculation with the transformed features; This represents the transpose of the temporal attention query vector, used to calculate similarity by performing a dot product with the transformed features; The temporal attention weight matrix is ​​a trainable parameter that performs a linear transformation on the input features. The temporal attention bias vector is a trainable parameter that enhances the expressive power of the attention mechanism. Indicates the first The fused feature vector at each time step; This represents the fused feature vector at the j-th time step; This represents the hyperbolic tangent activation function, which maps the input to... interval; This represents an exponential function used to calculate the normalized score of attention weights; The temporal context vector is obtained by weighted summation of features from each time step according to attention weights, and the calculation method is expressed as follows: ; This indicates the total number of time steps and determines the length of the time series.

[0104] S332, Calculate the attention weights for the sensor dimension

[0105] For each sensor channel's feature slice, a linear transformation is applied using the sensor attention weight matrix and sensor attention bias vector, followed by activation using the hyperbolic tangent activation function. The activation result is then compared with the sensor attention query vector using a dot product similarity calculation to obtain an unnormalized attention score. An exponential function is applied to and normalized the unnormalized scores of all sensor channels to generate channel attention weights for each sensor. These channel attention weights are then used to weight and sum the feature slices of each sensor to obtain the sensor context vector. This vector is then used to focus on sensors important for anomaly detection, as shown below:

[0106]

[0107] In the formula, Indicates the first The channel attention weights of each sensor are used to quantify the importance of that sensor for anomaly detection; The sensor context vector is obtained by weighting and summing the features of each sensor according to the attention weights. The calculation method is expressed as follows: ; This represents the sensor attention query vector, used for similarity calculation with sensor features; This represents the sensor attention weight matrix, which is a trainable parameter used to linearly transform sensor features; This represents the sensor attention bias vector, which is a trainable parameter that enhances the expressive power of the attention mechanism. Indicates all time steps Feature slices of each sensor are used to form feature vectors for the sensor channels; This represents the feature slice of the nth sensor at all time steps, forming the feature vector of the sensor channel; This indicates the total number of sensors and determines the number of sensor channels. This represents the sensor channel index, used to identify the first... One sensor; This represents the sensor channel index, used to identify the first... One sensor.

[0108] S333, Generate a two-dimensional attention context vector

[0109] The temporal context vector and the sensor context vector are concatenated to form a two-dimensional attention context vector, which integrates key information from both temporal and sensor dimensions, and is represented as follows:

[0110]

[0111] In the formula, This represents a two-dimensional attention context vector.

[0112] S34. Calculate the loss function

[0113] Settlement anomaly samples are scarce and class imbalanced, with normal samples far outnumbering warning and anomaly samples. Conventional cross-entropy loss is insufficient to optimize the decision boundary and lacks sensitivity to the identification of rare anomaly samples, which can easily lead to the model biasing towards the majority class and ignoring key anomaly signals.

[0114] This embodiment combines weighted cross-entropy and anomaly margin loss, using class weights to alleviate the sample imbalance problem, while forcing anomalous and normal classes to be significantly separated in the feature space, thereby improving the sensitivity to identify rare anomalous samples. The specific steps are as follows:

[0115] S341. Calculate the category weight coefficients.

[0116] Based on the number of samples in each class in the training set, calculate the corresponding class weight coefficient, which is used for weighted cross-entropy loss to give higher weights to rare classes and alleviate the sample imbalance problem. This is expressed as:

[0117]

[0118] In the formula, Indicates category The weighting coefficients are used for weighted cross-entropy loss; This represents the total number of samples in the training set, which is the sum of the number of samples in all categories. Indicates category The number of samples reflects the richness of samples in that category; This represents the category index, with a value of These correspond to categories such as normal, warning, and abnormal, respectively. This indicates the total number of categories, determining the number of categories for the classification task.

[0119] S342. Calculate the weighted cross-entropy loss components.

[0120] The cross-entropy loss of each sample is weighted using class weight coefficients. The weighted cross-entropy loss components are calculated to balance the contributions of different classes of samples to model training, ensuring that rare outliers are adequately learned. This is represented as:

[0121]

[0122] In the formula, This represents the weighted cross-entropy loss component, used to handle class imbalanced classification tasks; This represents the one-hot encoded class label vector, indicating when a sample belongs to class [class name missing]. hour Otherwise ; The model represents the categories The predicted probability is obtained by applying the Softmax function to the two-dimensional attention context vector. Calculated; Indicates category The weighting coefficients are used to balance the contributions of samples from different categories; This represents a logarithmic function, with the default base being the natural constant.

[0123] S343. Calculate the abnormal interval loss component.

[0124] Calculate the feature centers of the abnormal class and the normal class in the embedding space, and calculate the distance between the two feature centers. When this distance is less than the set minimum class interval, an abnormality interval loss component is generated, thereby forcing the abnormal class and the normal class to remain separated in the feature space, improving the separability of abnormal features, as expressed as:

[0125]

[0126] In the formula, This represents the abnormal interval loss component, used to enhance feature separation between abnormal and normal classes; This represents the interval loss weighting coefficient, which controls the contribution of interval loss to the total loss. It is preferably set to 0.2. This represents the maximum value function, ensuring that the interval loss is non-negative; This represents the desired minimum inter-class interval, and sets the minimum distance threshold between abnormal and normal classes in the feature space. The feature center of the anomaly class in the embedding space is obtained by averaging the feature vectors of all anomaly samples. The feature center of the normal class in the embedding space is obtained by averaging the feature vectors of all normal samples.

[0127] S344, Combined Total Loss Function

[0128] The weighted cross-entropy loss component is added to the outlier margin loss component to form the total loss function, which simultaneously optimizes classification accuracy and feature space structure, as expressed below:

[0129]

[0130] In the formula, This represents the total loss function, also known as the settlement-sensitive loss, which is used as the overall optimization objective for model training.

[0131] S35, Output of Anomaly Confidence Assessment

[0132] Conventional anomaly detection methods rely on a single threshold for decision-making, making it difficult to quantify the reliability of detection results. In settlement early warning scenarios, this can easily lead to false alarms and missed alarms, failing to provide a reliable reference for engineering decisions.

[0133] This embodiment calculates confidence based on the predicted probability distribution, comprehensively considers the relative probability of anomaly classes and prediction certainty, and uses a dual evaluation mechanism to quantify the reliability of anomaly detection, thus solving the trade-off problem between false positives and false negatives. The specific steps are as follows:

[0134] S351. Calculate the relative probability of anomaly classes.

[0135] The predicted probability of the anomalous class is divided by the sum of the exponents of the predicted probabilities of all abnormal classes to calculate the relative probability component of the anomalous class. By assessing the relative importance of the anomalous class in the risk categories, the dominance of the normal class is avoided. This is expressed as:

[0136]

[0137] In the formula, This represents the relative probability component of the anomaly class, used to assess the relative importance of the anomaly class within the risk category; This represents the predicted probability of an anomaly class, reflecting the model's confidence level in the anomaly state. Indicates category The predicted probability is calculated using the Softmax function; This represents the normal category label, used as an exclusion category in relative probability calculations; The label represents the anomaly category, which serves as the core category for relative probability calculation.

[0138] S352. Calculate the entropy of the predicted probability distribution.

[0139] The information entropy of the predicted probability distribution is calculated based on the probability distribution vector predicted by the model, quantifying the uncertainty of the model prediction results, and is expressed as:

[0140]

[0141] In the formula, Information entropy represents the predicted probability distribution and is used to quantify the uncertainty of the prediction results; This represents the predicted probability vector for all categories. ; Indicates category The natural logarithm of the predicted probability.

[0142] S353, Calculate the overall confidence level

[0143] Based on the results of the relative probability components of the anomaly class and the information entropy of the predicted probability distribution, a comprehensive confidence score is generated. This score, combined with the significance of the anomaly signal and the certainty of the prediction, is expressed as:

[0144]

[0145] In the formula, This represents the overall confidence score, with a value range of [value range missing]. This is used to quantify the reliability of anomaly detection results.

[0146] S36, Model Training and Parameter Update

[0147] During the training process of sedimentation anomaly samples, there are problems such as gradient instability and optimization oscillation. Conventional optimization methods are prone to getting stuck in local optima in class imbalance scenarios, and the sparsity of anomaly samples makes it difficult for the model to converge.

[0148] This embodiment minimizes the settling-sensitive loss by employing an adaptive moment estimation optimizer, enhances generalization ability by combining weight decay, and uses gradient pruning to prevent gradient explosion and oscillation during training, ensuring the stability of training on outlier samples. The specific steps are as follows:

[0149] S361. Perform adaptive moment estimator optimizer parameter update.

[0150] Based on the bias correction vectors estimated by the first and second moments, the model parameter vectors are updated by combining the learning rate and weight decay coefficient, and then the parameter update step size is adaptively adjusted, as expressed in:

[0151]

[0152] In the formula, Indicates the first The model parameter vector from the next iteration is used as the initial parameters for the next round of training. Indicates the first The model parameter vector for the next iteration contains all weights and biases; This represents the learning rate, which controls the step size of parameter updates; it is preferably set to 0.001. This represents the bias correction vector for the first-order moment estimate, used to correct the gradient mean. This represents the bias correction vector for the second-order moment estimate, used to correct the mean of the squared gradient. To represent the numerical stability constant and prevent the denominator from being zero, it is preferable to set it to... ; This represents the weight decay coefficient, which controls the strength of parameter regularization; it is preferably set to 0.2. This represents the training iteration index, indicating the current training round.

[0153] S362. Implement gradient clipping constraints.

[0154] Calculate the L2 norm of the gradient of the total loss function. This is used to measure the magnitude of the gradient. If the norm exceeds a preset gradient clipping threshold... The gradient vector is then scaled proportionally to a threshold range to prevent gradient explosion and violent oscillations during training. Setting a gradient clipping threshold is preferable. It is 0.5.

[0155] S37, Model Iterative Training and Parameter Update

[0156] The model iterative training process involves traversing the training dataset multiple times, using an adaptive moment estimator to minimize the settlement-sensitive loss function, and gradually updating the model parameters.

[0157] In each iteration, a batch of normalized sensor data sequences is input, and the outputs of dual-path feature extraction, feature fusion gating mechanism, and sedimentation-sensitive perception attention enhancement are calculated through forward propagation to obtain the predicted probability distribution. The total loss of weighted cross-entropy loss and abnormal interval loss is calculated, and the gradient is calculated through backpropagation.

[0158] To ensure training stability, gradient clipping constraints are implemented. The L2 norm of the gradient vector is checked to see if it exceeds a preset threshold; if so, the gradient is scaled proportionally. Parameter updates are based on the rules of the adaptive moment estimator optimizer, incorporating bias corrections from first and second-order moment estimations, adjusting the learning rate, and applying weight decay to prevent overfitting.

[0159] Iterative training continues until stopping conditions are met. These conditions include the total loss function no longer significantly decreasing on the validation set for multiple consecutive cycles (e.g., loss variation is less than a preset tolerance), reaching the maximum number of training iterations, or the validation set accuracy stabilizing. During training, loss and performance metrics are monitored to ensure effective model convergence and avoid overfitting. The optimized model parameters are then used for subsequent anomaly detection tasks.

[0160] To verify the effectiveness of the core module of this invention, the experiment configured 100 time steps of simulated data, including readings from six different types of sensors, and set abnormal periods at time steps 20 to 30 and 60 to 70. Figures 5-8The effects of dual-path feature extraction, temporal attention distribution, sensor attention distribution, and confidence assessment are presented separately. In the dual-path feature extraction effect diagram, the horizontal axis represents the time step, and the vertical axis represents the feature intensity. The global temporal feature curve smoothly changes, reflecting the long-term trend; the local abrupt change feature curve shows a significant peak during abnormal periods; and the fused feature curve combines the advantages of both, highlighting abnormal signals while maintaining trend continuity. The temporal dimension attention distribution diagram shows that the attention weight increases significantly during abnormal periods, indicating that the model can automatically focus on the key time steps where anomalies occur. The sensor dimension attention distribution diagram shows that the horizontal axis represents the sensor type, and the vertical axis represents the attention weight. It can be seen that earth pressure gauges and strain gauges are given higher weights, which is consistent with the understanding that these sensors are most sensitive to settlement anomalies in actual engineering. The confidence distribution comparison diagram uses a violin plot to show the distribution characteristics of the output confidence of each method. The confidence distribution of conventional methods is scattered and low, indicating that their judgment uncertainty is high; while the confidence distribution of the method of this invention is more concentrated and at a higher level, indicating that the model is more confident in the detection results. Experimental results show that the dual evaluation mechanism of the anomaly confidence assessment module in this invention considers both the relative probability of anomaly classes and the information entropy of the predicted probability distribution, thereby quantifying the reliability of anomaly detection results and providing more valuable reference for engineering decisions.

[0161] S4. Detection of abnormal settlement in pipe jacking construction

[0162] After the model is trained, the process of detecting abnormal settlement of pipes during pipe jacking construction will be applied to real-time or offline monitoring scenarios.

[0163] The newly acquired multi-source sensor data first undergoes adaptive sliding bin normalization to eliminate dimensional differences and adapt to local data distribution.

[0164] The normalized data sequence is input into the trained pipe jacking construction pipeline settlement anomaly detection model for forward propagation calculation;

[0165] The model captures global temporal patterns and local mutation features through a dual-path feature extraction module, and adaptively combines these features using a feature fusion gating mechanism.

[0166] The sedimentation-sensitive perception attention enhancement module assesses importance in both the time and sensor dimensions, focuses on key time steps and sensor channels, and generates a two-dimensional attention context vector.

[0167] Based on a two-dimensional attention context vector, the model outputs the predicted probability distribution for each time step, including the probabilities of normal, warning, and abnormal categories.

[0168] Calculate the anomaly confidence assessment, combine the relative probability of the anomaly class and the entropy of the predicted probability distribution to generate a comprehensive confidence score. If the predicted probability of the anomaly class exceeds a preset threshold or the comprehensive confidence score is high, it is determined that there is a settlement anomaly at that time step and an alarm is triggered.

[0169] The test results can be visualized or integrated into the monitoring system to provide engineers with real-time decision support and timely intervention measures to prevent safety accidents.

[0170] A comparative analysis of the performance of different methods in detecting settlement anomalies was conducted to evaluate the overall performance differences between the method of this invention and conventional detection methods in detecting pipeline settlement anomalies. Figure 9 As shown, the horizontal axis represents five detection methods (Support Vector Machine, Random Forest, Long Short-Term Memory Network, Convolutional Neural Network, and the method of this invention), and the vertical axis represents the performance index value (dimensionless proportion). Three sets of bar charts present the anomaly detection rate, early warning detection rate, and false alarm rate, respectively. The experimental configuration used the same construction dataset (containing sensor data of 2000 time steps), and the tests were repeated five times under five types of geological conditions, with the average value taken. The experimental results show that the method of this invention (rightmost bar group) is significantly higher than other methods in anomaly detection rate (red bars) and early warning detection rate (orange bars), while having the lowest false alarm rate (blue bars). The dual-path structure (long-range dependence + local mutation) of this invention accurately captures small settlement shifts. The bar height shows that its detection rate is about 15% higher than the suboptimal method. In addition, the feature fusion gating mechanism makes the early warning detection rate bar significantly higher than that of traditional methods, proving that it can identify early risks. The blue bar is the lowest and the error bar is the shortest, reflecting that the sensor's time-dimension attention effectively filters out interference signals. Furthermore, the length of the error bar indicates that the performance fluctuation range of this invention is minimal, verifying the training stability of gradient clipping and adaptive optimization.

Claims

1. A method for detecting abnormal settlement of pipelines during pipe jacking construction, characterized in that, Includes the following steps: S1. Multi-source sensor monitoring and data acquisition; S2, Adaptive sliding bin normalization; S3. Construct a model for detecting abnormal settlement of pipelines during pipe jacking construction, specifically including: S31. Construct a dual-path feature extraction module: By adopting a dual-path parallel structure, global temporal patterns and local mutation features are extracted separately to separate global trends and local jitter. S32. Feature Fusion Gating Mechanism: By adopting a gating mechanism to adaptively fuse dual-path outputs, the weights of long-range dependent paths and local mutation paths are dynamically allocated using learnable gating vectors, so that the model focuses on the global trend in the stable segment and enhances the local response in the mutation segment. S33, Enhanced attention to sedimentation sensitivity perception: By adopting a sensor time-dimension dual-dimensional attention mechanism, the importance is evaluated in both the time dimension and the sensor dimension, focusing on the abnormal signals of key sensors at critical moments; S34. Calculate the loss function: By combining weighted cross-entropy and outlier margin loss, class weights are used to alleviate sample imbalance, while forcing outlier classes and normal classes to be significantly separated in the feature space. S35. Anomaly Confidence Assessment Output: The confidence level is calculated based on the predicted probability distribution. The relative probability of the anomaly class and the predictive certainty are considered together. The reliability of anomaly detection is quantified by using a dual assessment mechanism. S36. Model Training and Parameter Update: The settling-sensitive loss is minimized by using an adaptive moment estimation optimizer, the generalization ability is improved by combining weight decay, and gradient pruning is used to prevent gradient explosion and oscillation during training. S37. Model Iterative Training and Parameter Update: The model iterative training process involves traversing the training dataset multiple times, using an adaptive moment estimator to minimize the settlement-sensitive loss function, and gradually updating the model parameters. S4. Detection of abnormal settlement of pipelines during pipe jacking construction.

2. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, Specifically, S2 involves using a sliding window centered on the current time step to extract the data set of the corresponding sensor within the window, and normalizing the original sensor readings of the current time step based on the minimum and maximum values ​​of the data set.

3. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, S31 specifically includes: S311, Long-range dependency path feature extraction: The forward-gated recurrent unit is used to process the normalized value vector sequence to generate the hidden state that captures the forward temporal dependency. At the same time, the backward-gated recurrent unit is used to process the normalized value vector sequence to generate the hidden state that captures the backward temporal dependency. The forward and backward hidden states at the same time step are concatenated to form the output feature vector of the long-range dependency path. S312. Local mutation path feature extraction: Using dilated causal convolution kernels, the normalized value vector sequence is convolved. The modified linear unit activation function is applied to the convolution result to generate the output feature vector of the local mutation path. The receptive field is expanded by dilated convolution while maintaining temporal causality.

4. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, Specifically, S32 includes: S321. Calculate the feature fusion gating vector: concatenate the output feature vector of the long-range dependent path and the output feature vector of the local mutation path, perform a linear transformation on the concatenated features through the gating weight matrix and the gating bias vector, and apply the Sigmoid activation function to the linear transformation result to generate the fusion gating vector. S322. Perform gated feature fusion: Use the fusion gate vector to weight the output feature vector of the long-range dependent path element by element, and at the same time, weight the output feature vector of the local mutation path element by element. Add the two weighted results to generate the final fusion feature vector.

5. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, Specifically, S33 includes: S331. Calculate the attention weights in the time dimension: Apply a linear transformation to the fused feature vector at each time step using the time attention weight matrix and the time attention bias vector, and activate it using the hyperbolic tangent activation function. Calculate the dot product similarity between the activation result and the time attention query vector to obtain the unnormalized attention score. Apply an exponential function to the unnormalized scores of all time steps and normalize them to generate the time attention weights for each time step. Use the time attention weights to perform a weighted summation of the fused feature vectors at each time step to obtain the time context vector. S332. Calculate sensor dimension attention weights: Apply a linear transformation to the feature slices of each sensor channel using the sensor attention weight matrix and the sensor attention bias vector, and activate them using the hyperbolic tangent activation function. Calculate the dot product similarity between the activation result and the sensor attention query vector to obtain the unnormalized attention score. Apply an exponential function to the unnormalized scores of all sensor channels and normalize them to generate the channel attention weights for each sensor. Use the channel attention weights to perform a weighted summation of the feature slices of each sensor to obtain the sensor context vector. S333. Generate a two-dimensional attention context vector: Concatenate the temporal context vector and the sensor context vector to form a two-dimensional attention context vector.

6. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, S34 specifically includes: S341. Calculate the class weight coefficient: Calculate the corresponding class weight coefficient based on the number of samples of each class in the training set; S342. Calculate the weighted cross-entropy loss components: Use the class weight coefficient to weight the cross-entropy loss of each sample, and calculate the weighted cross-entropy loss components. S343. Calculate the abnormal interval loss component: Calculate the feature center of the abnormal class in the embedding space and the feature center of the normal class in the embedding space, calculate the distance between the two feature centers, and generate the abnormal interval loss component when the distance is less than the set minimum class interval. S344. Combined total loss function: The weighted cross-entropy loss component and the outlier interval loss component are added together to form the total loss function.

7. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, Specifically, S35 includes: S351. Calculate the relative probability of the abnormal class: Divide the predicted probability of the abnormal class by the sum of the exponents of the predicted probabilities of all abnormal classes to calculate the relative probability component of the abnormal class. S352. Calculate the entropy of the predicted probability distribution: Calculate the information entropy of the predicted probability distribution based on the probability distribution vector predicted by the model; S353. Calculate the overall confidence score: Based on the results of the relative probability components of the anomaly class and the information entropy of the predicted probability distribution, generate an overall confidence score.

8. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, Specifically, S36 includes: S361. Perform adaptive moment estimation optimizer parameter update: update the model parameter vector based on the bias correction vector of the first-order moment estimation and the bias correction vector of the second-order moment estimation, combined with the learning rate and weight decay coefficient. S362. Implement gradient clipping constraints: Calculate the L2 norm of the gradient of the total loss function. If the norm exceeds the preset gradient clipping threshold If so, the gradient vector will be scaled proportionally to the threshold range.

9. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 1, characterized in that, In S37, in each iteration, a batch of normalized sensor data sequences are input, and the outputs of dual-path feature extraction, feature fusion gating mechanism, and sedimentation-sensitive perception attention enhancement are calculated through forward propagation to obtain the predicted probability distribution; the total loss of weighted cross-entropy loss and abnormal interval loss is calculated, and the gradient is calculated through backpropagation.

10. The method for detecting abnormal settlement of pipelines during pipe jacking construction according to claim 9, characterized in that, In step S37, a gradient clipping constraint is implemented, and it is checked whether the L2 norm of the gradient vector exceeds a preset threshold. If it does, the gradient is scaled proportionally.

Images