A method for identifying the health status of sluice structures on rock foundations

Abstract

The present invention relates to the field of artificial intelligence and data processing technology, and specifically to a method for identifying the health status of a sluice structure on a rock foundation. The method comprises the following steps: strain sensors are deployed at key locations on the surface of the sluice to collect strain data to form a training set, and the health status is annotated; the strain data in the training set is normalized using a fractional normalization strategy within a sliding time window; a health status identification model for the sluice structure on a rock foundation is constructed, and the normalized strain data is input into the model to identify the health status of the sluice structure on the rock foundation; the model is trained until preset conditions are met, at which point the training is terminated to obtain a trained model; and the newly collected strain data is input into the trained model to identify the health status of the sluice structure on the rock foundation. By capturing and analyzing weak change signals of the sluice structure, the present invention can promptly identify safety hazards existing in the sluice structure, thereby ensuring safe and stable operation of the sluice.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data processing technology, and in particular to a method for identifying the health status of a sluice structure on a rock foundation. Background Art

[0002] With the widespread application of sluice projects in water conservancy systems, their structural safety has become an increasing focus of attention. During long-term operation, rock-based sluices are affected by complex external environmental changes and continuous loads, and are prone to various damage modes such as crack expansion, foundation settlement, and fatigue failure.

[0003] For ordinary concrete sluice gates, the initial signs of structural damage typically manifest as localized distortion or dynamic variations in the strain field. These weak abnormal signals are easily masked by external noise, making early damage identification difficult. For sluice gates with stress absorption features, specifically those with a stress-absorbing layer between the base plate and the piers and those with cavities reserved within the piers filled with a tough material, the special stress-absorbing devices in these two structures can effectively alleviate some stress concentration issues and enhance the structural toughness and durability. However, these devices also complicate the stress-strain relationship. The dynamic response characteristics of the stress-absorbing devices are coupled with the mechanical behavior of the main sluice structure, resulting in intertwined and superimposed signals, further exacerbating the difficulty of extracting characteristic structural damage signals. Furthermore, performance degradation and aging of the stress-absorbing devices themselves can also pose new potential risks, creating new challenges for the real-time and accurate identification of the health status of sluice gate structures.

[0004] The existing methods for identifying the health status of sluice structures have the following problems: 1) The strain data collected from the sluice structure on the rock foundation are often accompanied by temperature drift, load fluctuation and external disturbance in actual operation, showing obvious non-stationary and multimodal distribution characteristics. Existing technologies mostly use static normalization methods such as Z-score, assuming that the data obeys a single Gaussian distribution, and cannot dynamically adjust the scaling range according to the time-varying data characteristics, resulting in small damage signals being masked by the normalization process, seriously affecting the early identification capability; 2) Conventional methods such as fast Fourier transform and wavelet transform require preset fixed frequency bands or basis functions in time-frequency analysis, making it difficult to respond effectively to unknown or diverse damage modes, and unable to achieve adaptive focusing in the frequency domain. In addition, the time domain positioning accuracy is limited, and the extracted features deviate from the actual damage performance, which seriously affects the accuracy of the damage detection. The generalization ability and applicability of the model are compromised; 3) The general attention mechanism in current deep learning mostly automatically generates feature channel weights based on the model training process, and lacks consideration of structural physical information such as sensor layout positions and key stress areas. In rock-based sluice structures, the degree of influence of different parts on overall safety varies significantly. The attention mechanism that ignores physical priors tends to overestimate the characteristics of non-critical areas, interfering with the accuracy of damage identification results; 4) The identity mapping path in the deep residual network is insufficiently responsive when processing micro-strain fluctuations, resulting in the weakening or even loss of early damage information during network propagation. At the same time, conventional single-scale feature fusion methods cannot take into account the characteristic expressions of various damage types such as trend changes and high-frequency disturbances, making the model's recognition performance unstable under complex working conditions and lacking broad adaptability.

[0005] Therefore, the present invention proposes a health status identification method for a sluice structure on a rock foundation to solve the above problems. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention develops a method for identifying the health status of a sluice structure on a rock foundation. The present invention captures and analyzes the weak change signals of the sluice structure through an optimized sluice structure health status identification method, promptly discovers the safety hazards of the sluice structure, and ensures the safe and stable operation of the sluice.

[0007] The technical solution to the technical problem of the present invention is a method for identifying the health status of a sluice structure on a rock foundation, comprising the following steps:

[0008] S1. Strain sensors are deployed at key locations on the sluice gate surface to collect strain data in real time. Strain data segments with representative time periods are selected as samples to form a training set. The health status of each sample is labeled based on actual test results and expert manual evaluation opinions.

[0009] S2, using the fractional normalization strategy within the sliding time window to normalize the strain data in the training set to obtain normalized strain data;

[0010] S3. Construct a health status identification model for the sluice structure on the rock foundation. The model includes a feature extraction layer, an attention mechanism layer, a gated differential residual structure layer, a multi-scale damage feature fusion layer, a prediction layer, and a confidence calibration layer. The normalized strain data is input into the model to identify the health status of the sluice structure on the rock foundation.

[0011] S4. Train the health status recognition model for the sluice structure on the rock foundation. After the input strain data is forward propagated, the time-series-aware focal loss function is calculated to obtain the error between the predicted result and the actual health status label. The error is then backpropagated to update the trainable parameters in the model. The model is iterated multiple times until the preset conditions are met, at which point the model training is terminated to obtain a trained model.

[0012] S5. Input the newly collected strain data into the trained model to identify the health status of the sluice structure on the rock foundation, and obtain the health status determination result of the sluice structure on the rock foundation at the current time.

[0013] S1 is as follows:

[0014] The stress concentration areas of the sluice structure are selected as key locations, including the piers, bottom plates, side walls and connecting transition areas;

[0015] The classification labels for health status include normal, minor injury, and major injury.

[0016] S2 is as follows:

[0017] The strain data in the training set are locally processed using a quantile normalization strategy within a sliding time window. The low and high quantiles of each strain sensor in the current time window are calculated, a dynamic range factor is constructed, and scaling and offset conversion operations are performed on the strain data in the training set to output the normalized strain data.

[0018] The operations in the feature extraction layer are as follows:

[0019] First, the normalized strain values ​​are weighted and integrated at each time point using the mean fusion method of multiple strain sensors. Then, adaptive time-frequency modeling is performed using a learnable kernel function to capture frequency bands and time domain patterns that are highly correlated with damage characteristics. The specific steps are as follows:

[0020] (1) Mean fusion of multiple strain sensors: By averaging the normalized strain values ​​of all strain sensors at a specified time point, a fused response curve is generated. The mean vector of the normalized strain values ​​of all strain sensors is expressed as , is the time point index, To distinguish from Another time point index of , and , Indicates the length of the time window;

[0021] (2) Learnable time-frequency feature extraction:

[0022] A time-frequency feature extraction method based on a learnable kernel function is adopted. By constructing multiple kernel functions with parameterized structures, weighted calculation is performed on the fused strain sequence in the time domain. The frequency domain modulation is performed in combination with a complex exponential function of a specified frequency to generate a multidimensional damage-sensitive feature vector. The generated multidimensional damage-sensitive feature is expressed as follows:

[0023] ,

[0024] in, express Multi-dimensional damage-sensitive characteristics of the moment, express Dimensions, ; express The moment damage-sensitive features, express The moment damage-sensitive features, express The moment damage-sensitive features;

[0025] The calculation formula of damage sensitivity characteristics is as follows:

[0026] ,

[0027] in, express Moment A parameterized kernel function; is a natural constant; represents an imaginary unit; Indicates the Preset center frequencies; for The differential symbol of .

[0028] The operations in the attention mechanism layer are as follows:

[0029] A weighted attention mechanism driven by physical priors is adopted to dynamically weight the strain energy density of damage-sensitive features. First, the mean strain energy density is calculated based on the mean square value of all damage-sensitive features over the entire time series. Then, the mean strain energy density is used as the attention weight. The attention weight corresponding to each damage-sensitive feature is expressed as , Indicates the The attention weight of the damage-sensitive feature, Indicates the The attention weight of the damage-sensitive feature, Indicates the The attention weight of the damage-sensitive feature, Indicates the The attention weight of each damage-sensitive feature;

[0030] The attention weight Applied to the feature vector, each damage-sensitive feature is Full time series characteristics The corresponding attention weight Perform channel-by-channel multiplication operations, perform weighted suppression or enhancement processing on the features of each channel, and finally generate a weighted feature matrix .

[0031] The operations in the gated differential residual structure layer are as follows:

[0032] Construct a gated differential residual structure, adopt a learnable gating mechanism for the output path of each residual block, and dynamically allocate information flow between the residual path and the differential path according to the input feature state. When the input is a steady-state feature, the model prioritizes the standard residual channel. When the input is a mutation feature, the differential channel enhances the short-term dynamic response. The gated differential residual block outputs the gated differential feature matrix .

[0033] The operations in the multi-scale damage feature fusion layer are as follows:

[0034] The gated differential feature matrix is ​​fused using a mechanically aware feature pyramid fusion strategy. Perform multi-scale time pooling separately to construct feature representations of multiple time scales , Indicates the Then, by calculating the strain energy density change rate of the corresponding feature of each scale, as the basis for measuring its damage evolution rate, the generation of dynamic weighting coefficients is guided; finally, all scale features are fused through the dynamic weighting coefficient to form a unified feature representation, and the multi-scale fusion feature is obtained. .

[0035] The operations in the prediction layer are as follows:

[0036] Use the fully connected layer to fuse multi-scale features Perform linear transformation and output prediction scores for each category , Indicates the The prediction scores of the categories, is a positive integer;

[0037] Then the prediction score Convert it into an output with a probabilistic explanation, use the Softmax normalized exponential function to normalize all category scores, perform an exponential transformation on the score of each category, and normalize it with the scores of other categories to generate a normalized probability distribution vector , Indicates that the model predicts The probability of a category.

[0038] The operations in the confidence calibration layer are as follows:

[0039] A damage propagation consistency calibration mechanism is adopted. By analyzing the strain energy density variation trend of each category at multiple scales under model prediction and comparing it with the strain energy gradient model constructed based on physical laws, the damage consistency factor is obtained. The damage consistency factor is then used as the weight to renormalize the original classification probability to generate the confidence-calibrated category prediction probability. The calculation formula is as follows:

[0040] ,

[0041] in, Indicates the first The predicted probability of each category; is the total number of categories of health status; The model predicts The probability of each category; The model predicts The probability of each category; is a positive integer; For the The damage consistency factor of each category; For the The damage consistency factor of each category.

[0042] S4 is as follows:

[0043] S4.1. Data forward propagation:

[0044] The labeled training dataset is input into the health status recognition model of the sluice structure on the rock foundation to execute the forward calculation process. It is transformed step by step through the feature extraction layer, attention mechanism layer, gated differential residual structure layer and multi-scale damage feature fusion layer. The fused features are fully connected through the prediction layer to obtain the scores of each category, and then converted into the initial probability distribution through the Softmax function. Finally, the confidence calibration layer uses the damage consistency factor to renormalize the probability under physical constraints and output the calibrated health status prediction probability.

[0045] S4.2, temporal-aware focal loss function:

[0046] A time-aware focal loss function is used. Based on the standard focal loss, a continuous label representing the damage severity is combined. The penalty intensity of each sample is adjusted through nonlinear mapping. When the damage severity approaches the threshold, the loss function automatically increases the misclassification penalty, encouraging the model to focus on these mixed samples, making the model better at identifying transition states. The calculation formula is as follows:

[0047] ,

[0048] in, It is a temporal-aware focal loss function; is a continuous value for injury severity; For the One-hot encoding of categories; It is a logarithmic function, with base 10 by default; is the dynamic modulation index, Nonlinear changes;

[0049] S4.3. Error back propagation:

[0050] A time-aware focus loss function calculates the error between the model's predictions and the true labels. The error signal propagates back layer by layer from the output layer to the input layer. The chain rule is used to calculate the gradient of each trainable parameter with respect to the loss function. Based on this gradient information, a gradient descent optimization algorithm is used to update the trainable parameters in the model, gradually reducing the loss and optimizing the model's performance.

[0051] S4.4, Iteration stop condition judgment:

[0052] The model uses the mini-batch gradient descent method for multiple rounds of iterative training. After each round of training, an independent validation dataset is used to evaluate the current model performance. The loss function value and classification accuracy index on the validation dataset are monitored. When the preset conditions are met, the training is terminated to obtain a trained model. The preset conditions are as follows:

[0053] 1) The validation set loss does not decrease significantly for Z consecutive rounds, where Z represents the preset number of rounds;

[0054] 2) The accuracy of the validation set reaches the preset threshold and remains stable;

[0055] 3) The preset maximum number of training rounds is reached.

[0056] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0057] Aiming at the characteristics of non-stationarity, strong interference and multimodal distribution of strain signals during the operation of sluice structures on rock foundations, the present invention adopts a quantile normalization method based on a sliding time window, and performs scaling transformation on the original signal by dynamically calculating the local low quantile and high quantile, which can effectively enhance the model's ability to retain and identify tiny abnormal signals and improve the detection effect of initial structural damage; the time-frequency feature extraction mechanism constructed with a learnable parameter kernel function proposed in the present invention automatically adjusts the time domain focusing ability and frequency domain modulation frequency during the training process, which can solve the problem that traditional time-frequency analysis methods cannot adapt to complex or unknown damage patterns, and realize the detection of key response patterns. Adaptive extraction; the present invention combines structural mechanics knowledge, designs an attention mechanism based on the average value of the strain energy density corresponding to the feature, dynamically adjusts the model's attention to various time-frequency feature channels, and suppresses redundant noise signals by enhancing high-energy feature responses, thereby significantly enhancing key damage features; the present invention adopts a gated differential residual module and a multi-scale feature fusion mechanism in the network structure, dynamically adjusts the ratio of the residual path to the differential path according to the input state, enhances the response capability to transient mutation signals, and dynamically weights and fuses features at different time granularities through a scale-aware mechanism, thereby improving the compatibility and discrimination of multiple types of damage.

[0058] In summary, the present invention provides an optimized method for identifying the health status of a sluice structure. By capturing and analyzing the weak change signals of the sluice structure, potential safety hazards of the sluice structure can be discovered in a timely manner, thereby ensuring the safe and stable operation of the sluice. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0060] Figure 1 Schematic diagram of the method of the present invention.

[0061] Figure 2 This is the surface strain detection data diagram of the acting sluice.

[0062] Figure 3This figure shows the enhancement effect of the normalization method in the present invention and the conventional Z-score normalization method on weak damage signals.

[0063] Figure 4 This is a comparison chart of the recognition accuracy of the gated differential residual structure in the present invention and the existing method in the damage evolution process.

[0064] Figure 5 This is a comparison chart of the early damage monitoring capabilities of the gated differential residual structure in the present invention and existing methods.

[0065] Figure 6 This is a comparison chart of the accuracy of the confidence calibration mechanism. DETAILED DESCRIPTION

[0066] In order to clearly illustrate the technical features of this solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0067] Example 1

[0068] The specific steps of a method for identifying the health status of a sluice structure on a rock foundation are as follows:

[0069] S1. Strain sensors are deployed at key locations on the sluice gate surface to collect strain data in real time. Strain data segments with representative time periods are selected as samples to form a training set. The health status of each sample is labeled based on actual test results and expert manual evaluation opinions.

[0070] S2, using the fractional normalization strategy within the sliding time window to normalize the strain data in the training set to obtain normalized strain data;

[0071] S3. Construct a health status identification model for the sluice structure on the rock foundation. The model includes a feature extraction layer, an attention mechanism layer, a gated differential residual structure layer, a multi-scale damage feature fusion layer, a prediction layer, and a confidence calibration layer. The normalized strain data is input into the model to identify the health status of the sluice structure on the rock foundation.

[0072] S4. Train the health status recognition model for the sluice structure on the rock foundation. After the input strain data is forward propagated, the time-series-aware focal loss function is calculated to obtain the error between the predicted result and the actual health status label. The error is then backpropagated to update the trainable parameters in the model. The model is iterated multiple times until the preset conditions are met, at which point the model training is terminated to obtain a trained model.

[0073] S5. Input the newly collected strain data into the trained model to identify the health status of the sluice structure on the rock foundation, and obtain the health status determination result of the sluice structure on the rock foundation at the current time.

[0074] The specific implementation is as follows, S1 is as follows:

[0075] The stress concentration areas of the sluice structure are selected as key locations. Key locations include the piers, bottom plate, side walls, and connection transition areas. Strain sensors are deployed at key locations. The strain sensors record the strain changes of the sluice structure in real time at a fixed frequency during operation to obtain strain data. Deploying strain sensors at the above key locations can accurately capture the stress response and local deformation information of the structure under different working conditions. The strain data collection process is as follows:

[0076] Lay on the surface of the sluice structure strain sensors with a sampling frequency Continuously collect data, all strain sensors in the time window The data in constitute the strain data matrix, which is expressed as:

[0077] ,

[0078] in, express The strain data matrix at the moment contains all strain sensors in the time window Data within; represents the time window length, , represents the empirical constant, Indicates the sampling frequency, set , 30 sampling points cover 1-2 gate operations, which can ensure that the time window includes the complete load cycle; Indicates the first strain sensor at time point The strain value at ; Indicates the first strain sensor at time point The strain value at ; Indicates the The strain sensor at time The strain value at ; Indicates the The strain sensor at time The strain value at ;

[0079] Any strain sensor is represented as , , any time is expressed as , , strain data matrix The strain value of any strain sensor at any time is expressed as ;

[0080] Strain data segments with representative time periods are selected as samples and form a training set. Then, based on the actual test results and expert manual evaluation opinions, the health status of each sample is annotated. The annotation content includes classification labels and injury severity indicators in the form of continuous values. , , 0 represents a completely normal state, 1 represents a complete failure of the sluice structure, and the classification labels include normal, slightly damaged, and severely damaged; the damage severity index in the form of a continuous value Structural inspection personnel conduct a comprehensive assessment based on factors such as crack width, settlement amplitude, historical evolution trends, etc. to form a quantitative damage level.

[0081] The specific implementation is as follows, S2 is as follows:

[0082] The quantile normalization strategy within the sliding time window is used to locally process the strain data in the training set. The low quantile and high quantile of each strain sensor in the current time window are calculated, and the dynamic range factor is constructed. The strain data in the training set are scaled and offset transformed, and the normalized strain data matrix is ​​output, which is expressed as:

[0083] ,

[0084] in, express The strain data matrix after time normalization eliminates the environmental noise and non-stationary effects in the strain data; Indicates the first strain sensor at time point Normalized strain value at ; Indicates the first strain sensor at time point Normalized strain value at ; Indicates the The strain sensor at time Normalized strain value at ; Indicates the The strain sensor at time Normalized strain value at ;

[0085] Normalized strain data matrix The normalized strain value of any strain sensor at any time is expressed as , the calculation formula is as follows:

[0086] ,

[0087] in, Indicates the Strain sensors in the time window 0.1 quantile within ; represents the dynamic range factor, , No. Strain sensors in the time window The dynamic range factor captures the non-Gaussian distribution characteristics of strain data through the quantile dynamics, solving the strain data interference problem caused by large environmental noise; Indicates a constant to prevent division by zero, set .

[0088] The specific implementation is as follows. The operations in the feature extraction layer are as follows:

[0089] Damage to sluice gates, such as cracks and foundation settlement, manifests itself in strain data as energy mutations in specific frequency bands. Conventional fast Fourier transforms ignore time-domain positioning information, while wavelet transforms require preset basis functions. Fixed-basis-function wavelets are difficult to match unknown damage patterns, and manual feature engineering has poor generalization. These methods cannot effectively combine time-domain and frequency-domain information, making it difficult to adapt to unknown damage patterns, resulting in inaccurate extraction of damage features. Therefore, considering the limitations of information from a single sensor, the present invention first uses a mean fusion method of multiple strain sensors to perform weighted integration of normalized strain values ​​at each time point. Then, a learnable kernel function is used for adaptive time-frequency modeling to capture frequency bands and time-domain patterns that are highly correlated with damage features, thereby improving adaptability and discrimination capabilities for unknown damage patterns. The specific steps are as follows:

[0090] (1) Mean fusion of multiple strain sensors:

[0091] By averaging the normalized strain values ​​of all strain sensors at a specified time point, a fused response curve is generated. This curve is used as the input basis for further feature extraction. By fusion of the mean values ​​of multiple strain sensors, the strain sensors can maintain spatial coverage information while reducing redundancy and improving modeling stability. The calculation formula is as follows:

[0092] ,

[0093] in, Indicates at a point in time The mean vector of the normalized strain values ​​of all strain sensors, with dimension ;

[0094] (2) Learnable time-frequency feature extraction:

[0095] A time-frequency feature extraction method based on a learnable kernel function is adopted. By constructing multiple kernel functions with parameterized structures, weighted calculations are performed on the fused strain series in the time domain, and frequency domain modulation is performed in combination with a complex exponential function of a specified frequency to generate a multidimensional damage-sensitive feature vector. The parameters of each kernel function can be automatically adjusted during the training process, enabling it to adaptively focus on different time positions and frequency intervals, thereby automatically capturing response change characteristics that are highly correlated with potential structural damage. The generated multidimensional damage-sensitive feature is expressed as follows:

[0096] ,

[0097] in, express Multi-dimensional damage-sensitive characteristics of the moment, express Dimensions; express The moment damage-sensitive features; express The first damage-sensitive feature at time t; express The Kth damage-sensitive feature at time t;

[0098] The calculation formula of damage sensitivity characteristics is as follows:

[0099] ,

[0100] in, express Moment A parameterized kernel function, , Indicates the A parameter that controls the width of the time domain, Indicates the A parameter that controls the oscillation frequency; is a natural constant; represents an imaginary unit; Indicates the A preset center frequency can be set to 、 、 .

[0101] It should be noted that the parameterized kernel function The training parameters are trained by gradient descent method to achieve adaptive optimization of parameters, and then adaptively determine the damage frequency band, such as the parameters that control the oscillation frequency We learned that 0.5–5 Hz corresponds to structural resonance, solving the problem of matching unknown damage patterns.

[0102] The specific implementation is as follows. The operations in the attention mechanism layer are as follows:

[0103] The contribution of strains in different parts of the sluice to the overall health status varies greatly. For example, the strain of the pier is more critical than that of the wing wall, and the severity of damage is nonlinearly related to the characteristics. The conventional attention mechanism treats all features equally and ignores the structural mechanics prior, which easily leads to over-response of secondary features and fails to effectively focus on high-energy features that are more sensitive to changes in health status, affecting the accuracy of structural damage location and identification. Therefore, in order to enhance the model's responsiveness to key parts of the structure, the present invention adopts a weighted attention mechanism driven by physical priors, and dynamically weights the strain energy density of damage-sensitive features. First, the mean strain energy density of all damage-sensitive features over the entire time series is calculated to measure the physical contribution of the feature to structural damage. Then, the mean strain energy density is used as the attention weight, and the attention weight corresponding to each damage-sensitive feature is expressed as , Indicates the The attention weight of each damage-sensitive feature can guide the attention mechanism to enhance the response of the significant feature channel, enabling the model to automatically focus on high-energy features that are more sensitive to changes in health status and suppress the interference of low-value features, thereby improving the accuracy of structural damage location and identification;

[0104] The calculation formula of attention weight is as follows:

[0105] ,

[0106] in, Indicates the damage-sensitive features The full time series characteristics of Indicates the damage-sensitive features The full time series characteristics of ; represents the calculation of the mean strain energy density function, , Indicates the length of time; represents the attention weight matrix; Indicates the The attention weight of each damage-sensitive feature;

[0107] The attention weight Applied to the feature vector, each damage-sensitive feature is Full time series characteristics The corresponding attention weight Perform channel-by-channel multiplication operations, perform weighted suppression or enhancement processing on the features of each channel, and finally generate a weighted feature matrix .

[0108] It should be noted that strain energy density is positively correlated with damage. The calculation method of the mean strain energy density of the feature conforms to the principles of structural mechanics, which can guide the model to focus on high-energy features and reduce the impact of interference from secondary features.

[0109] The specific implementation is as follows. The operations in the gated differential residual structure layer are as follows:

[0110] Health state classification requires a deep nonlinear mapping that captures strain characteristics, but the identity mapping of conventional residual neural networks fails in the early stages of damage. The standard residual block propagates features through identity mapping and nonlinear activation functions, is insensitive to weak signals, and gradients easily disappear in shallow layers, resulting in small damage being difficult to effectively propagate in the early stages. Therefore, in order to address the problem that small damage is difficult to propagate through standard residual blocks in the early stages, the present invention constructs a gated differential residual structure, adopts a learnable gating mechanism for the output path of each residual block, and dynamically allocates information flow between the residual path and the differential path according to the input feature state. When the input is a steady-state feature, the model prioritizes the standard residual channel. When the input is a mutation feature, the differential channel enhances the short-term dynamic response, and the gated differential residual block outputs a gated differential feature matrix. , in order to improve the model's sensitivity to weak anomalies, solve the problem of limited information propagation in the early stage of damage, and maintain the training stability and gradient validity of the deep network.

[0111] The calculation formula in the gated differential residual structure layer is as follows:

[0112] ,

[0113] in, represents the gated differential feature matrix; Represents the gate vector, which can control the fusion ratio of the residual and differential paths. , The range of is [0,1], represents the Sigmoid activation function, represents the weight matrix of the gating vector, represents the bias term of the gate vector; represents the feature activation value of the residual path, , represents the rectified linear unit activation function, represents the weight matrix of the residual path, represents the bias term of the residual path; Represents the time differential of the residual path, which can amplify the trend change caused by damage and capture the instantaneous gradient mutation. , represents the time differential step size, Represents the residual path in The feature activation value at time , Represents the residual path in Feature activation value at the moment; Represents the Hadamard product symbol.

[0114] It should be noted that in the gated differential feature matrix During the calculation of By adaptively selecting the signal enhancement method, steady-state features take the residual path, and transient features take the differential path, the problem of insensitivity to weak signals can be solved.

[0115] The specific implementation is as follows. The operations in the multi-scale damage feature fusion layer are as follows:

[0116] The damage of sluice structure varies at different time scales. For example, foundation settlement shows a low-frequency trend change, crack expansion is a medium-frequency mutation, and material fatigue is a high-frequency oscillation. It is difficult for a single-scale feature to fully characterize the health state. Conventional feature fusion such as splicing or average pooling ignores the physical correlation between scales, and fixed weights cannot adapt to the contribution differences of different damage modes. It is impossible to effectively combine feature information of different scales, resulting in insufficient recognition of complex damage patterns. Therefore, since different damage types show significant differences at the time scale, in order to achieve a more comprehensive description of the structural state, the present invention adopts a mechanically aware feature pyramid fusion strategy to transform the gated differential feature matrix into Multi-scale time pooling is performed separately to construct features of multiple time scales. Then, the strain energy density change rate of the corresponding feature of each scale is calculated as the basis for measuring its damage evolution rate, guiding the generation of dynamic weighting coefficients. Finally, all scale features are fused through the dynamic weight coefficient to form a unified feature representation, and the multi-scale fusion feature is obtained. , realizing the adaptive combination of diverse damage representations such as low-frequency trends and high-frequency disturbances, and enhancing the universality and discrimination of the model under different damage modes.

[0117] The calculation formula in the multi-scale damage feature fusion layer is as follows:

[0118] ,

[0119] ,

[0120] in, Represents multi-scale fusion features; Indicates the number of scales of the feature pyramid, set ; represents the symbol of partial derivative; Indicates the Scale features, different scales correspond to different time resolutions, usually by dividing the gated differential feature matrix Apply multi-scale time pooling, which can be average pooling or maximum pooling, and the pooling window size varies with the scale increase and grow; Indicates the Characteristics of scale The mean strain energy density; Indicates the Characteristics of scale The time rate of change of the upper strain energy density, , Indicates the Characteristics of scale exist The mean strain energy density at time , Indicates the Characteristics of scale exist The mean strain energy density at time t; Represents the weight matrix of the feature pyramid; Indicates the Dynamic weight coefficient of scale feature.

[0121] It should be noted that in the calculation process of the dynamic weight coefficient, the dynamic weight coefficient is positively correlated with the damage evolution rate, which solves the problem of missing scale correlation. For example, the minute-level weight is automatically increased during the crack propagation period.

[0122] The specific implementation is as follows. The operations in the prediction layer are as follows:

[0123] Use the fully connected layer to fuse multi-scale features Perform linear transformation and output the prediction scores for each category. The calculation formula is as follows:

[0124] ,

[0125] in, Represents the weight matrix of the fully connected layer; represents the bias term of the fully connected layer; Indicates the The prediction scores of the categories, is a positive integer;

[0126] Then the prediction score Convert it into an output with a probabilistic explanation, use the Softmax normalized exponential function to normalize all category scores, perform exponential transformation on the score of each category, and normalize it with the scores of other categories to generate a normalized probability distribution vector. The calculation formula is as follows:

[0127] ,

[0128] in, Indicates that the model predicts The probability of each category; is the total number of categories of health status; Indicates the The prediction scores of the categories.

[0129] It should be noted that in the Softmax normalized exponential function, Prediction scores for each category To calculate the probability of all categories, Prediction scores for each category Specifically refers to the current The scores of each category.

[0130] The specific implementation is as follows. The operations in the confidence calibration layer are as follows:

[0131] The confidence level of sluice health status prediction is crucial for operation and maintenance decisions. However, deep models tend to output overconfident probabilities in fuzzy areas of category boundaries, and environmental noise can interfere with prediction reliability. Conventional calibration methods such as temperature scaling ignore the physical characteristics of structural damage evolution, cannot adapt to the dynamic scenarios of sluice monitoring, and cannot effectively calibrate the confidence level of model output, resulting in low confidence levels in prediction results under actual monitoring environments. Therefore, in order to improve the credibility of model output under actual monitoring environments, the present invention adopts a damage propagation consistency calibration mechanism. By analyzing the strain energy density variation trend of each category at multiple scales under model prediction, and comparing it with the strain energy gradient model constructed based on physical laws, a damage consistency factor is obtained. The damage consistency factor is then used as a weight to renormalize the original classification probability to generate a confidence-calibrated category prediction probability. This can effectively suppress the overconfident probability output generated by the model under damage boundaries or environmental disturbances. The calculation formula is as follows:

[0132] ,

[0133] in, Indicates the first The predicted probability of each category; The model predicts The probability of each category; The model predicts The probability of each category; is a positive integer; For the The damage consistency factor of each category; For the The damage consistency factor of each category;

[0134] The damage consistency factor is calculated as follows:

[0135] ,

[0136] in, For the The damage consistency factor of each category, is the calibration sensitivity coefficient, and ,set up , represents the L2 norm, For the The time derivative of the injury severity vector of each category represents the changing trend of injury severity at different time scales;

[0137] Each component of the time derivative of the damage severity vector is approximated by numerical difference. For the characteristic representation of each scale, the mean strain energy density at two consecutive time points is selected. The difference between the two is calculated and divided by the time interval to obtain the rate of change of the strain energy density at that scale, that is, each component of the time derivative of the damage severity vector is characterized. The calculation formula is as follows:

[0138] ,

[0139] in, For the The injury severity vectors for each category are derived from multi-scale features. ; represents transpose; represents the theoretical strain energy gradient, constrained by the physical model, , represents the mechanical parameter matrix, Indicates the The material degradation coefficient of each category, Represents the mechanical parameter bias term.

[0140] It should be noted that the damage consistency factor is used to quantify the degree of matching between the model prediction and the physical law. For example, when the multi-scale damage evolution trend is highly consistent with the theoretical strain energy gradient, The term approaches 0, When it approaches 1, the original confidence is maintained. When the environmental noise or model misjudgment causes the two to deviate significantly, Increase, Exponential decay forces the confidence of the category to decrease.

[0141] The specific implementation is as follows, S4 is as follows:

[0142] S4.1. Data forward propagation:

[0143] The labeled training dataset is input into the health status recognition model of the sluice structure on the rock foundation to execute the forward calculation process. It is transformed step by step through the feature extraction layer, attention mechanism layer, gated differential residual structure layer and multi-scale damage feature fusion layer. The fused features are fully connected through the prediction layer to obtain the scores of each category, and then converted into the initial probability distribution through the Softmax function. Finally, the confidence calibration layer uses the damage consistency factor to renormalize the probability under physical constraints and output the calibrated health status prediction probability.

[0144] S4.2, temporal-aware focal loss function:

[0145] The health status of the sluice gate undergoes a gradual deterioration process, but the samples of minor damage are scarce and the boundary between them and the normal state is fuzzy. The standard cross entropy loss is sensitive to class imbalance, and FocalLoss tends to ignore the temporal continuity of damage evolution, cannot effectively deal with the class imbalance problem, and has insufficient recognition ability for samples of minor damage. To solve the problem of scarcity and fuzzy boundaries of samples of the minor damage category in the health status of the sluice gate, the present invention adopts a time-aware focal loss function. On the basis of the standard focal loss, combined with a continuous label representing the severity of the damage, the penalty intensity of each sample is adjusted through nonlinear mapping. When the damage degree approaches the threshold, the loss function automatically increases the misclassification penalty, encouraging the model to focus on these samples, making the model better at identifying transition states, and avoiding the decline in recognition ability caused by excessive separation between categories. The calculation formula is as follows:

[0146] ,

[0147] in, It is a temporal-aware focal loss function; is a continuous value for injury severity; For the One-hot encoding of categories; It is a logarithmic function, with base 10 by default; is the dynamic modulation index, Nonlinear changes;

[0148] The dynamic modulation index is calculated as follows:

[0149] ,

[0150] in, For the balance parameter of loss modulation, set ; is the hyperbolic tangent function; To control the shape parameter of the dynamic modulation index change, set ; is the damage threshold, set ;

[0151] It should be noted that in order to solve the problem of fuzzy sample feature boundaries, the dynamic modulation index At the damage threshold In its calculation process, the hyperbolic tangent function is used to ensure a smooth transition of the penalty.

[0152] S4.3. Error back propagation:

[0153] A time-aware focus loss function calculates the error between the model's predictions and the true labels. The error signal propagates back layer by layer from the output layer to the input layer. The chain rule is used to calculate the gradient of each trainable parameter with respect to the loss function. Based on this gradient information, a gradient descent optimization algorithm is used to update the trainable parameters in the model, gradually reducing the loss and optimizing the model's performance.

[0154] S4.4, Iteration stop condition judgment:

[0155] The model uses the mini-batch gradient descent method for multiple rounds of iterative training. After each round of training, an independent validation dataset is used to evaluate the current model performance. The loss function value and classification accuracy index on the validation dataset are monitored. When the preset conditions are met, the training is terminated to obtain a trained model. The preset conditions are as follows:

[0156] 1) The validation set loss does not decrease significantly for Z consecutive rounds, where Z represents the preset number of rounds;

[0157] 2) The accuracy of the validation set reaches the preset threshold and remains stable;

[0158] 3) The preset maximum number of training rounds is reached.

[0159] Example 2

[0160] like Figure 2 and Figure 3 As shown in the figure, in order to verify the effectiveness of the quantile normalization method in extracting weak damage signals, the processing results of the original strain signal, conventional Z-score normalization and the quantile normalization of the present invention are compared to simulate the enhancement ability of different methods on damage characteristics in a real engineering environment. Figure 2The original strain signal and the real damage signal are shown in the figure. The real damage signal includes three typical damage modes, namely, short-term high-incidence micro cracks, continuous low-frequency foundation settlement and randomly distributed material fatigue. These signals are all submerged by the environmental noise. Figure 3 The experimental results shown in the figure indicate that, although conventional Z-score normalization can suppress some noise, its enhancement effect on transient damage such as tiny cracks is limited. The quantile normalization of the present invention significantly amplifies the three types of damage features, especially completely extracting the tiny crack signal from the noise background while maintaining the continuity of foundation settlement and the distribution characteristics of material fatigue. This proves that quantile normalization can accurately capture non-Gaussian distribution characteristics through the dynamic range factor and can solve the problem of feature distortion caused by environmental noise interference.

[0161] Example 3

[0162] like Figure 4 As shown in the figure, in order to analyze the ability of the gated differential residual structure to capture the damage evolution process, the complete evolution process of the sluice structure from normal state to severe damage within 300 days is compared with the prediction effect of the present invention and the mainstream time series models (LSTM long short-term memory network model, SVM support vector machine model, and random forest model). Figure 4 The black baseline in the middle represents the actual damage development trajectory, and the other curves represent the prediction results of different methods, and are marked with the mean absolute error index (MAE). Figure 4 The experimental results shown indicate that the prediction curve (blue) of the present invention closely follows the actual damage evolution trajectory, demonstrating precise trend capture capabilities, especially in the damage acceleration stage (approximately 80 days) and the critical point of structural instability (approximately 150 days). However, the LSTM long short-term memory network model (green), SVM support vector machine model (red), and random forest model (purple) exhibit significant response lag and deviation amplification. The experimental results validate the role of the gated differential residual structure in the present invention. By dynamically selecting the residual path and the differential path, the differential channel response is automatically enhanced during the damage acceleration period, addressing the problem of conventional residual networks being insensitive to subtle trend changes.

[0163] like Figure 5 As shown in the figure, we selected the case data where the damage degree had a small change in the early stage and a sudden change occurred in the late stage. Figure 5The black solid line in the middle represents the actual damage level, which evaluates the detection sensitivity of the gated differential residual structure in the entire damage evolution cycle and compares it with the traditional conventional residual network, SVM traditional support vector machine and FFT fast Fourier spectrum analysis method; in the dual vertical axis chart, the actual damage level includes four stages: latent period, stable development, accelerated expansion, and severe deterioration. The colored curves are the detection probabilities of different methods. In the latent period, the detection probability of the technology of the present invention is the first to increase, and signs of damage are discovered more than 60 days earlier than other methods. In the accelerated expansion period of damage, the slope of the curve of the technology of the present invention is the largest, proving that the differential path effectively amplifies instantaneous gradient mutations. In addition, under the interference of random fluctuations, the detection probability curve of the technology of the present invention is closest to the actual damage trend, indicating that the gating mechanism can adaptively select the signal enhancement method.

[0164] Example 4

[0165] like Figure 6 As shown, in order to verify the improvement of the prediction reliability of the damage propagation consistency calibration mechanism, the uncalibrated model and the conventional temperature scaling method are compared. The horizontal axis of the reliability diagram is the prediction confidence, the vertical axis is the actual accuracy, and the three curves represent the calibration effects of different methods. The experimental results show that the technology of the present invention deviates from the ideal line in the high confidence interval (>0.8) before calibration, and is close to the ideal line after calibration, proving that physical constraints effectively correct misjudgments. In addition, in the confidence interval of 0.4-0.6, the calibration curve of the technology of the present invention is above the conventional method, indicating that the damage consistency factor improves the discrimination accuracy of easily confused samples, indicating that the calibration mechanism makes the prediction probability conform to the law of structural damage evolution by comparing the strain energy gradient with the theoretical model, thereby avoiding the blindness of pure data-driven.

[0166] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for identifying the health status of a sluice structure on a rock foundation, characterized by: The following steps are involved: S1. Strain sensors are deployed at key locations on the sluice gate surface to collect strain data in real time. Strain data segments with representative time periods are selected as samples to form a training set. The health status of each sample is labeled based on actual test results and expert manual evaluation opinions. S2, using the fractional normalization strategy within the sliding time window to normalize the strain data in the training set to obtain normalized strain data; S3. Construct a health status identification model for the sluice structure on the rock foundation. The model includes a feature extraction layer, an attention mechanism layer, a gated differential residual structure layer, a multi-scale damage feature fusion layer, a prediction layer, and a confidence calibration layer. The normalized strain data is input into the model to identify the health status of the sluice structure on the rock foundation. The operations in the feature extraction layer are as follows: First, the normalized strain values ​​are weighted and integrated at each time point using the mean fusion method of multiple strain sensors. Then, adaptive time-frequency modeling is performed using a learnable kernel function to capture frequency bands and time domain patterns that are highly correlated with damage characteristics. The specific steps are as follows: (1) Mean fusion of multiple strain sensors: By averaging the normalized strain values ​​of all strain sensors at a specified time point, a fused response curve is generated. The mean vector of the normalized strain values ​​of all strain sensors is expressed as , is the time point index, To distinguish Another time point index of , and , Indicates the length of the time window; (2) Learnable time-frequency feature extraction: A time-frequency feature extraction method based on a learnable kernel function is adopted. By constructing multiple kernel functions with parameterized structures, weighted calculation is performed on the fused strain sequence in the time domain. The frequency domain modulation is performed in combination with a complex exponential function of a specified frequency to generate a multidimensional damage-sensitive feature vector. The generated multidimensional damage-sensitive feature is expressed as follows: ， in, express Multi-dimensional damage-sensitive characteristics of the moment, express Dimensions, ; express The moment damage-sensitive features, express The moment damage-sensitive features, express The moment damage-sensitive features; The calculation formula of damage sensitivity characteristics is as follows: ， in, express Moment A parameterized kernel function; is a natural constant; represents an imaginary unit; Indicates the Preset center frequencies; for Differentiation symbol of ; S4. Train the health status recognition model for the sluice structure on the rock foundation. After the input strain data is forward propagated, the time-series-aware focal loss function is calculated to obtain the error between the predicted result and the actual health status label. The error is then backpropagated to update the trainable parameters in the model. The model is iterated multiple times until the preset conditions are met, at which point the model training is terminated to obtain a trained model. S5. Input the newly collected strain data into the trained model to identify the health status of the sluice structure on the rock foundation, and obtain the health status determination result of the sluice structure on the rock foundation at the current time.

2. The method for identifying the health status of a sluice structure on a rock foundation according to claim 1, wherein: S1 is as follows: The stress concentration areas of the sluice structure are selected as key locations, including the piers, bottom plates, side walls and connecting transition areas; The classification labels for health status include normal, minor injury, and major injury.

3. The method for identifying the health status of a sluice structure on a rock foundation according to claim 2, wherein: S2 is as follows: The strain data in the training set are locally processed using a quantile normalization strategy within a sliding time window. The low and high quantiles of each strain sensor in the current time window are calculated, a dynamic range factor is constructed, and scaling and offset conversion operations are performed on the strain data in the training set to output the normalized strain data.

4. According to the method for identifying the health status of a sluice structure on a rock foundation according to claim 3, the operations in the attention mechanism layer are as follows: A weighted attention mechanism driven by physical priors is adopted to dynamically weight the strain energy density of damage-sensitive features. First, the mean strain energy density is calculated based on the mean square value of all damage-sensitive features over the entire time series. Then, the mean strain energy density is used as the attention weight. The attention weight corresponding to each damage-sensitive feature is expressed as , Indicates the The attention weight of the damage-sensitive feature, Indicates the The attention weight of the damage-sensitive feature, Indicates the The attention weight of the damage-sensitive feature, Indicates the The attention weight of each damage-sensitive feature; The attention weight Applied to the feature vector, each damage-sensitive feature is Full time series characteristics The corresponding attention weight Perform channel-by-channel multiplication operations, perform weighted suppression or enhancement processing on the features of each channel, and finally generate a weighted feature matrix .

5. The method for identifying the health status of a sluice structure on a rock foundation according to claim 4, wherein: The operations in the gated differential residual structure layer are as follows: Construct a gated differential residual structure, adopt a learnable gating mechanism for the output path of each residual block, and dynamically allocate information flow between the residual path and the differential path according to the input feature state. When the input is a steady-state feature, the model prioritizes the standard residual channel. When the input is a mutation feature, the differential channel enhances the short-term dynamic response. The gated differential residual block outputs the gated differential feature matrix .

6. The method for identifying the health status of a sluice structure on a rock foundation according to claim 5, wherein: The operations in the multi-scale damage feature fusion layer are as follows: The gated differential feature matrix is ​​fused using a mechanically aware feature pyramid fusion strategy. Perform multi-scale time pooling separately to construct feature representations of multiple time scales , Indicates the Characteristics of scale; Then, by calculating the strain energy density change rate of the corresponding feature of each scale, as the basis for measuring its damage evolution rate, the generation of dynamic weighting coefficients is guided; finally, all scale features are fused through the dynamic weighting coefficients to form a unified feature representation, and the multi-scale fusion feature is obtained. .

7. The method for identifying the health status of a sluice structure on a rock foundation according to claim 6, wherein: The operations in the prediction layer are as follows: Use the fully connected layer to fuse multi-scale features Perform linear transformation and output prediction scores for each category , Indicates the The prediction scores of the categories, is a positive integer; Then the prediction score Convert it into an output with a probabilistic explanation, use the Softmax normalized exponential function to normalize all category scores, perform an exponential transformation on the score of each category, and normalize it with the scores of other categories to generate a normalized probability distribution vector , Indicates that the model predicts The probability of a category.

8. The method for identifying the health status of a sluice structure on a rock foundation according to claim 7, wherein: The operations in the confidence calibration layer are as follows: A damage propagation consistency calibration mechanism is adopted. By analyzing the strain energy density variation trend of each category at multiple scales under model prediction and comparing it with the strain energy gradient model constructed based on physical laws, the damage consistency factor is obtained. The damage consistency factor is then used as the weight to renormalize the original classification probability to generate the confidence-calibrated category prediction probability. The calculation formula is as follows: ， in, Indicates the first The predicted probability of each category; is the total number of categories of health status; The model predicts The probability of each category; The model predicts The probability of each category; is a positive integer; For the The damage consistency factor of each category; For the The damage consistency factor of each category.

9. The method for identifying the health status of a sluice structure on a rock foundation according to claim 8, wherein: S4 is as follows: S4.

1. Data forward propagation: The labeled training dataset is input into the health status recognition model of the sluice structure on the rock foundation to execute the forward calculation process. It is transformed step by step through the feature extraction layer, attention mechanism layer, gated differential residual structure layer and multi-scale damage feature fusion layer. The fused features are fully connected through the prediction layer to obtain the scores of each category, and then converted into the initial probability distribution through the Softmax function. Finally, the confidence calibration layer uses the damage consistency factor to renormalize the probability under physical constraints and output the calibrated health status prediction probability. S4.2, temporal-aware focal loss function: A time-aware focal loss function is used. Based on the standard focal loss, a continuous label representing the damage severity is combined. The penalty intensity of each sample is adjusted through nonlinear mapping. When the damage severity approaches the threshold, the loss function automatically increases the misclassification penalty, encouraging the model to focus on these mixed samples, making the model better at identifying transition states. The calculation formula is as follows: ， in, It is a temporal-aware focal loss function; is a continuous value for injury severity; For the One-hot encoding of categories; It is a logarithmic function, with base 10 by default; is the dynamic modulation index, Nonlinear changes; S4.

3. Error back propagation: A time-aware focus loss function calculates the error between the model's predictions and the true labels. The error signal propagates back layer by layer from the output layer to the input layer. The chain rule is used to calculate the gradient of each trainable parameter with respect to the loss function. Based on this gradient information, a gradient descent optimization algorithm is used to update the trainable parameters in the model, gradually reducing the loss and optimizing the model's performance. S4.4, Iteration stop condition judgment: The model uses the mini-batch gradient descent method for multiple rounds of iterative training. After each round of training, an independent validation dataset is used to evaluate the current model performance. The loss function value and classification accuracy index on the validation dataset are monitored. When the preset conditions are met, the training is terminated to obtain a trained model. The preset conditions are as follows: 1) The validation set loss does not decrease significantly for Z consecutive rounds, where Z represents the preset number of rounds; 2) The accuracy of the validation set reaches the preset threshold and remains stable; 3) The preset maximum number of training rounds is reached.

Images