Method for evaluating health level of gravity dam in operation combining temperature field and stress field

Abstract

The present application relates to the technical field of engineering health assessment, in particular to a gravity dam operation health grade evaluation method combined with temperature field and stress field, specifically as follows: a gravity dam thermal monitoring sample is constructed, the grid space one-to-one correspondence of the original temperature field and stress field matrix is realized, the real health grade is labeled and the training, verification and test sets are divided; the coupling enhancement feature map is generated through grid space unification, numerical standardization, gradient direction deviation and local coupling imbalance calculation; a typical operation mode library is constructed and matched, the mode maximum response map and response variance map are generated, and the network input feature map is spliced; a double branch network containing a bidirectional gate fusion mechanism is used to process the features, and the gravity dam health grade prediction result is output; the network is trained with sequence relationship weighted loss, and the trained network is deployed to complete the gravity dam operation health grade evaluation. The present application can solve the problem of inaccurate evaluation of single physical field, and efficiently evaluate the operation health state of gravity dam.

Description

Technical Field

[0001] This invention relates to the field of engineering health assessment technology, and in particular to a method for assessing the operational health level of gravity dams by combining temperature field and stress field. Background Technology

[0002] During the operation of a gravity dam, a natural thermoelastic coupling exists between the temperature field and stress field of the dam body. Under normal conditions, temperature changes trigger corresponding stress responses through thermal expansion and contraction, and the temperature gradient and stress gradient typically maintain a certain degree of consistency in spatial direction. However, when structural anomalies occur within the dam body, such as the propagation of microcracks, local stiffness degradation, the initiation of seepage channels, or changes in heat transfer paths, this thermo-mechanical coupling relationship deviates. This manifests as the temperature gradient direction no longer aligning with the stress gradient direction, or a significant mismatch between the intensity of temperature change and the intensity of stress response. This type of thermo-mechanical decoupling often occurs in the early stages of damage evolution, when the absolute amplitude of the temperature or stress values ​​may not yet have undergone significant abrupt changes and remains within the normal fluctuation range.

[0003] Existing technologies objectively suffer from the following shortcomings: Firstly, existing methods often treat temperature and stress fields as independent data or simply concatenate them into the network, lacking a quantitative expression of their coupling consistency / decoupling deviation along the gradient direction, making it difficult to effectively capture subtle thermal decoupling anomalies caused by early damage. Secondly, existing methods lack historical knowledge accumulation and comparison of typical local thermal coupling patterns of dams under different health states, only analyzing the absolute characteristics of the current sample, easily misjudging normal operating condition fluctuations such as seasonal temperature differences and water level changes as structural anomalies. Thirdly, existing identification networks often adopt a single-branch structure, directly mixing physical field information with historical reference information, leading to mutual interference between the two types of heterogeneous features during convolution or strong amplitude signals drowning out weak reference signals, making it difficult to achieve synergistic utilization. Fourthly, existing technologies often use standard cross-entropy as the loss function, ignoring the explicit order relationship between health levels, failing to impose additional penalties commensurate with the consequences of "directly misjudging normal states as severe anomalies," which is unacceptable in engineering. Therefore, this invention proposes a method for assessing the operational health level of gravity dams by combining temperature and stress fields to solve the above problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method for assessing the operational health status of gravity dams by combining temperature and stress fields. This invention can solve the problem of inaccurate assessment using a single physical field and efficiently assess the operational health status of gravity dams.

[0005] The technical solution of this invention is a method for assessing the operational health level of gravity dams by combining temperature and stress fields. The specific steps are as follows: S1. Collect data to construct a thermal monitoring sample of gravity dam. Make a one-to-one correspondence between the original temperature field matrix and the original stress field matrix in the sample in the grid space, and assign a real health level label to each sample. Then divide the dataset composed of the samples into training set, validation set and test set. S2. Enhance the samples by generating a coupled enhancement feature map through grid space uniformity, numerical standardization, gradient direction deviation calculation, and local coupling imbalance calculation. S3. Construct a typical operating mode library, perform local matching between the shape of the coupling enhancement feature map corresponding to each sample and the typical operating mode library, generate the maximum response map and the variance map of the mode, and then concatenate them with the coupling enhancement feature map to generate the network input feature map. S4. Construct a dual-branch gated fusion network, including a physically coupled branch, a pattern matching branch, and a bidirectional gated fusion network. Input the network input feature map into the two branches respectively, and then exchange information through the bidirectional gated fusion mechanism to output the gravity dam health level prediction result. S5. The dual-branch gated fusion network is trained using order relation weighted loss to obtain the trained dual-branch gated fusion network. S6. Deploy the training dual-branch gating fusion network in the gravity dam safety monitoring method to assess the operational health registration of the gravity dam.

[0006] The S1 operation is as follows: A thermal monitoring sample for gravity dams is constructed by collecting data from internal temperature sensors, strain gauges or stress inversion results, finite element calculation results, monitoring interpolation results, and manual inspection confirmation results. Spatial position calibration was performed on the original temperature field matrix and original stress field matrix in the thermal monitoring samples of the gravity dam. Using the unified cross-sectional coordinate system of the dam body as the reference, the original temperature field matrix and original stress field matrix were resampled to a unified spatial grid. A health level label is assigned to each gravity dam thermal monitoring sample, and the health level is divided into normal state, attention state, warning state and abnormal state; All gravity dam thermal monitoring samples were divided into training set, validation set and test set.

[0007] The specific operations for grid space uniformity and numerical normalization in S2 are as follows: Using the original temperature field matrix and original stress field matrix from the gravity dam thermal monitoring sample as input, spatial grid consistency and numerical standardization are performed on the two. First, check whether the spatial grids of the original temperature field matrix and the original stress field matrix are consistent. If the two have different sizes or inconsistent spatial coordinates, they are resampled to the same spatial grid. The original temperature field matrix is ​​standardized based on the statistical parameters of the training set, specifically by using the mean and standard deviation of the temperature fields in the training set to obtain the standardized temperature field matrix; the original stress field matrix is ​​also standardized based on the statistical parameters of the training set, specifically by using the mean and standard deviation of the stress fields in the training set to obtain the standardized stress field matrix.

[0008] The specific steps for calculating the gradient direction deviation of S2 are as follows: By calculating the gradient direction deviation response matrix, the Sobel operator is used to calculate the gradient components of the normalized temperature field matrix in the horizontal and vertical directions respectively, to obtain the temperature horizontal gradient matrix and temperature vertical gradient matrix. Then, the gradient components of the normalized stress field matrix in the horizontal and vertical directions are calculated in the same way to obtain the stress horizontal gradient matrix and stress vertical gradient matrix. The temperature gradient magnitude matrix is ​​calculated based on the horizontal and vertical temperature gradient matrices. The stress gradient magnitude matrix is ​​calculated based on the horizontal and vertical stress gradient matrices. The gradient direction deviation response matrix is ​​calculated based on the temperature gradient direction and the stress gradient direction. The temperature gradient magnitude and stress gradient magnitude are compared with the set flat region judgment threshold, and the flat region is suppressed based on the comparison results.

[0009] The specific calculation process of the coupling enhancement feature map in S2 is as follows: The temperature gradient magnitude matrix and stress gradient magnitude matrix are locally smoothed to obtain smoothed temperature gradient magnitude matrix and smoothed stress gradient magnitude matrix. The gradient magnitude imbalance matrix is ​​calculated based on the smoothed temperature gradient magnitude matrix and smoothed stress gradient magnitude matrix. Then, the coupling imbalance index matrix is ​​calculated by weighted summation based on the gradient direction deviation response matrix and the gradient magnitude imbalance matrix. The coupling imbalance index matrix is ​​then truncated and normalized to obtain the normalized coupling imbalance index matrix. The normalized temperature field matrix, normalized stress field matrix, gradient direction deviation response matrix and coupling imbalance index matrix are concatenated along the channel direction to obtain the coupling enhancement feature map, which contains four channels.

[0010] The specific operations in S3 are as follows: Based on the location of the dam structure, key areas are determined. Within each health level, local thermal coupling sample blocks are extracted from the coupling enhancement feature map using a sliding window, with each grid location within the key area as the center. All local thermal coupling sample blocks extracted under the same health level are flattened into vectors and K-means clustering is performed to obtain the prototype pattern corresponding to the health level. The mean is removed and the norm is normalized for each prototype pattern to obtain the normalized prototype pattern, which constitutes a typical operation mode library. Calculate the normalized cross-correlation response value between the current local thermal coupling sample block and each prototype mode in the typical operating mode library to obtain the mode response value. Traverse all health levels and all prototype modes, and count the maximum value of all mode response values ​​at each spatial location to obtain the mode maximum response map. Calculate the variance of all mode response values ​​at each spatial location to obtain the mode response variance map. Finally, concatenate the coupling enhancement feature map, the mode maximum response map, and the mode response variance map along the channel direction to obtain the network input feature map.

[0011] The physical coupling branch operations in S4 are as follows: The network input feature map has six channels, containing two types of heterogeneous information: the first four channels belong to physical coupling features, and the last two channels belong to pattern matching features. The network input feature map is split by channel to obtain the physical coupling input feature map. This physical coupling feature map is then input into the first convolutional layer of the physical coupling branch to obtain the initial physical coupling feature map. The initial physical coupling feature map is then input into the second convolutional layer of the physical coupling branch to obtain the middle physical coupling feature map. The coupling imbalance index matrix in the coupling enhancement feature map is then adjusted to the spatial size of the middle physical coupling feature map to obtain the scale-adapted coupling imbalance index matrix. A thermal decoupling gated weight map is generated based on the scale-adapted coupling imbalance index matrix. This thermal decoupling gated weight map is applied to the middle physical coupling feature map to obtain the gated enhanced physical coupling feature map. Finally, the gated enhanced physical coupling feature map is input into the third convolutional layer of the physical coupling branch to obtain the high-level physical coupling feature map.

[0012] The specific pattern matching branch operations in S4 are as follows: The network input feature map is split by channel to obtain the pattern matching input feature map. The pattern matching input feature map is then fed into the first convolutional layer of the pattern matching branch to obtain the initial pattern matching feature map. The initial pattern matching feature map is then fed into the second convolutional layer of the pattern matching branch to obtain the intermediate pattern matching feature map. Channel attention is calculated on the intermediate pattern matching feature map to obtain the pattern channel attention weights. The pattern channel attention weights are then multiplied by the intermediate pattern matching feature map channel by channel to obtain the channel-enhanced pattern matching feature map. The channel-enhanced pattern matching feature map is then fed into the third convolutional layer of the pattern matching branch to obtain the high-level pattern matching feature map.

[0013] The specific operation of the bidirectional gated fusion network in S4 is as follows: Global average pooling is performed on the physical coupling high-level feature map to obtain a physical coupling global vector. Simultaneously, global average pooling is performed on the pattern matching high-level feature map to obtain a pattern matching global vector. Physical branch adjustment weights are generated based on the pattern matching global vector, and pattern branch adjustment weights are also generated based on the physical coupling global vector. Based on the physical coupling high-level feature map and the pattern branch adjustment weights, as well as the pattern matching high-level feature map and the physical branch adjustment weights, interactively enhanced physical coupling feature maps and interactively enhanced pattern matching feature maps are obtained through residual enhancement. These interactively enhanced physical coupling feature maps and interactively enhanced pattern matching feature maps are concatenated along the channel direction to obtain a concatenated fused feature map. This concatenated fused feature map is then input into a 1×1 convolutional layer to obtain a fused high-level feature map. Global average pooling is then performed on the fused high-level feature map to obtain a classification feature vector. Finally, the classification feature vector is input into a fully connected classification layer to obtain a classification logical value vector. Softmax normalization is performed on the classification logical value vector to obtain a health level prediction probability vector. The health level with the highest probability is selected as the initial health level recognition result output by the network.

[0014] The specific operations in S5 are as follows: Calculate the health level prediction probability vector for each sample in the training set. Calculate the class balance weight based on the number of samples for each health level in the training set. The fewer the samples for a health level, the greater the corresponding class balance weight. Calculate the order relation weighted loss, which includes the class balance cross-entropy term and the order relation penalty term. The Adam optimizer is used to update all trainable parameters in the thermally decoupled sensitive dual-branch gated fusion network. After each training round, the model performance is evaluated using the validation set to obtain the trained dual-branch gated fusion network, which is then tested using the test set.

[0015] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a method for assessing the operational health level of gravity dams by combining temperature and stress fields. Addressing early-stage thermo-mechanical decoupling phenomena, this invention does not directly input the original field values. Instead, it extracts the gradient direction deviation response and coupling imbalance index of the temperature and stress fields, constructing an enhanced feature map that amplifies early anomalous signals such as microcracks and stiffness degradation. Furthermore, this invention proposes constructing a library of typical operational modes containing standardized temperature and stress fields and their coupling anomalies. By matching the local thermo-mechanical coupling patterns of the current sample with prototype modes under historical health levels, maximum response maps and variance maps of the modes are generated. This invention provides a historical reference for judging morphological deviations in the network. It also proposes a dual-branch gating fusion network that combines physical coupling and pattern matching. The physically coupled branch focuses on abnormal regions through a thermal decoupling gating mechanism, while the pattern matching branch strengthens key historical features through channel attention. Finally, bidirectional gating fusion is used to enhance the interaction of heterogeneous information, avoiding feature interference. Furthermore, an order relation weighted loss function is introduced during model training, incorporating the severity differences between health levels into the penalty term. This allows the model to impose a greater penalty on "serious misjudgments across levels," solving the problem that ordinary cross-entropy loss cannot distinguish the severity of errors. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a flowchart illustrating the method of the present invention.

[0018] Figure 2 This is a schematic diagram of the operation process for step S1.

[0019] Figure 3 This is a schematic diagram of the operation process for step S2.

[0020] Figure 4 This is a schematic diagram of the operation process for generating the pattern matching feature map in step S3.

[0021] Figure 5 This is a schematic diagram of the operation process for extracting physical coupling branch features in step S4.

[0022] Figure 6 This is a schematic diagram of the operation flow for feature extraction of pattern matching branches in step S4.

[0023] Figure 7 This is a comparison chart showing the changes in the probability of predicting abnormal states between the method of this invention and the conventional stress field method. Detailed Implementation

[0024] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0025] Example 1 like Figure 1 As shown, a method for assessing the operational health level of a gravity dam by combining temperature and stress fields is described, with the following specific steps: S1. Collect data to construct a thermal monitoring sample of gravity dam. Make a one-to-one correspondence between the original temperature field matrix and the original stress field matrix in the sample in the grid space, and assign a real health level label to each sample. Then divide the dataset composed of the samples into training set, validation set and test set. S2. Enhance the samples by generating a coupled enhancement feature map through grid space uniformity, numerical standardization, gradient direction deviation calculation, and local coupling imbalance calculation. S3. Construct a typical operating mode library, perform local matching between the shape of the coupling enhancement feature map corresponding to each sample and the typical operating mode library, generate the maximum response map and the variance map of the mode, and then concatenate them with the coupling enhancement feature map to generate the network input feature map. S4. Construct a dual-branch gated fusion network, including a physically coupled branch, a pattern matching branch, and a bidirectional gated fusion network. Input the network input feature map into the two branches respectively, and then exchange information through the bidirectional gated fusion mechanism to output the gravity dam health level prediction result. S5. The dual-branch gated fusion network is trained using order relation weighted loss to obtain the trained dual-branch gated fusion network. S6. Deploy the training dual-branch gating fusion network in the gravity dam safety monitoring method to assess the operational health registration of the gravity dam.

[0026] In a specific implementation, step S1 is performed as follows: Under normal thermoelastic response conditions, temperature changes typically induce corresponding stress changes, and the temperature gradient and stress gradient exhibit a certain degree of consistency in spatial direction. However, when microcracks, local stiffness degradation, seepage channels, heat transfer anomalies, or stress release anomalies exist in the dam body, the directions of the temperature gradient and stress gradient may deviate significantly. Simply inputting the temperature and stress fields as ordinary two-channel data into the network is insufficient to represent such thermo-mechanical decoupling anomalies. Therefore, this invention organizes each gravity dam thermal monitoring sample into a unified format and configures a health level label, enabling subsequent feature enhancement, pattern matching, and network training to be performed based on the same data standard. Figure 2 As shown, the specific steps of S1 are as follows: 1) This invention first constructs a thermal monitoring sample of a gravity dam. The sample sources may include internal temperature sensors, strain gauges or stress inversion results, finite element calculation results, monitoring interpolation results, and manual inspection confirmation results. Each thermal monitoring sample of a gravity dam includes an original temperature field matrix and an original stress field matrix. The original temperature field matrix is ​​used to characterize the temperature distribution on the dam cross section or a specified plane of the dam, and the original stress field matrix is ​​used to characterize the stress response distribution at the same spatial location. The original temperature field matrix and the original stress field matrix have a one-to-one correspondence on the spatial grid, that is, the same row and column position corresponds to the same physical position inside the dam.

[0027] Define the original temperature field matrix in the nth gravity dam thermal monitoring sample as: The original stress field matrix in the nth gravity dam thermal monitoring sample is defined as follows: Where n represents the sample index, which ranges from 1 to N, and N represents the total number of gravity dam thermal monitoring samples; and The dimensions are all H×W, where H represents the number of rows of the spatial grid and W represents the number of columns of the spatial grid.

[0028] In practical implementation, the spatial mesh can be derived from the finite element section mesh of the gravity dam, or from a regular mesh obtained by spatial interpolation of the sensor monitoring points; for example, if the typical cross-section of the dam is divided into 64 rows and 64 columns, then H=64 and W=64; if the finite element output mesh is an irregular mesh, it can be interpolated to a regular mesh based on the center coordinates or node coordinates of the finite element elements.

[0029] 2) Perform spatial position calibration on the original temperature field matrix and the original stress field matrix so that they correspond to the same physical position of the dam body at the same grid position.

[0030] In practical implementation, if the original temperature field matrix is ​​derived from temperature sensor interpolation results and the original stress field matrix is ​​derived from strain inversion or finite element calculation results, the two may have different spatial resolutions or boundary ranges. In this case, using the unified cross-sectional coordinate system of the dam body as a reference, both the original temperature field matrix and the original stress field matrix are resampled to a unified spatial grid of size H×W.

[0031] In practical implementation, resampling can be performed using bilinear interpolation, finite element shape function interpolation, or nearest neighbor interpolation. For missing measurement locations near the dam boundary, the mean of adjacent effective grids can be used to fill the gaps; for continuously missing measurement areas, linear interpolation of adjacent grids at the same elevation or interpolation of adjacent finite element elements can be used to fill the gaps.

[0032] 3) Assign a health level label to each gravity dam thermal monitoring sample. The health level label is denoted as follows: , This represents the true health level corresponding to the nth gravity dam thermal monitoring sample.

[0033] In one implementation, the number of health levels is defined as C, where C represents the total number of health levels to be identified, for example, C=4. When C=4, health level 0 represents a normal state, health level 1 represents a state of concern, health level 2 represents a state of alert, and health level 3 represents an abnormal state.

[0034] Among them, the normal state indicates that the monitoring indicators are consistent with the historical stable operating state; the attention state indicates that there is a slight thermal response deviation but it does not reach the alarm standard; the warning state indicates that the thermal coupling relationship is obviously abnormal and requires key inspection; the abnormal state indicates that there may be microcrack expansion, local seepage channels, stiffness degradation or other structural abnormalities, which require engineering treatment.

[0035] 4) All gravity dam thermal monitoring samples were divided into training, validation, and test sets. The training data in the training set was used to statistically standardize parameters, construct a typical operating mode library, and train the recognition network; the validation data in the validation set was used to select model parameters and adjust hyperparameters; and the test data in the test set was used to evaluate the final recognition performance.

[0036] In a specific implementation, step S2 is performed as follows: While the original temperature field matrix and the original stress field matrix can express the thermal distribution and mechanical response respectively, their direct concatenation still lacks an expression of whether temperature changes cause consistent stress changes. For early damage to gravity dams, anomalies are often not abrupt changes in temperature or stress values ​​themselves, but rather deviations in the coupling relationship between the temperature gradient and the stress gradient. Therefore, this invention uses the original temperature field matrix and the original stress field matrix as input, and generates a coupling enhancement feature map through spatial grid unification, numerical standardization, gradient direction deviation calculation, and local coupling imbalance calculation. Specifically, it is a coupling enhancement 4-channel feature map, which retains the original thermal field information while enhancing the thermal decoupling anomaly, such as... Figure 3 As shown, the specific steps of S2 are as follows: (1) Spatial grid uniformity and numerical standardization The original temperature field matrix and the original stress field matrix have different dimensions. Temperature is usually expressed in degrees Celsius, while stress is usually expressed in megapascals or kilopascals. If the gradient is directly calculated and jointly compared, the difference in numerical scale will affect the determination of the coupling relationship. This invention first performs spatial mesh unification and numerical normalization on the two matrices to obtain the normalized temperature field matrix and the normalized stress field matrix. The specific steps are as follows: 1) For the current gravity dam thermal monitoring sample to be processed, the original temperature field matrix is ​​defined as T, and the original stress field matrix is ​​defined as S, where the dimensions of both T and S are H×W. This represents the temperature value at the i-th row and j-th column. This represents the stress value at the i-th row and j-th column, where i represents the spatial row index and j represents the spatial column index. The value of i ranges from 1 to H, and the value of j ranges from 1 to W.

[0037] 2) Check whether the spatial grids of the original temperature field matrix and the original stress field matrix are consistent. If the two have different sizes or inconsistent spatial coordinates, resample them to the same spatial grid. After resampling, the i-th row and j-th column correspond to the same physical location of the dam body in the original temperature field matrix and the original stress field matrix.

[0038] In practical implementation, a finite element cross-sectional regular mesh can be used as a unified spatial mesh. For the original temperature field matrix, a regular matrix can be generated based on the temperature sensor coordinates using inverse distance weighted interpolation, radial basis function interpolation, or finite element node interpolation. For the original stress field matrix, it can be interpolated to the same regular matrix based on finite element stress results or strain inversion results. If there is no direct observation at a certain mesh location, the average value of adjacent valid meshes is used for filling. If the missing measurement location is consecutively more than 3 meshes, the interpolation results of adjacent finite element elements can be used first to fill the missing value to maintain spatial continuity.

[0039] In one embodiment, for example, the finite element cross-section mesh of a gravity dam is 80 rows and 80 columns, while the temperature sensor only covers the central region of 60 rows and 60 columns. Using the finite element mesh coordinates as a unified spatial mesh reference, the temperature sensor monitoring values ​​are interpolated using inverse distance weighted interpolation to generate an 80×80 regular matrix. The stress field is directly calculated using the 80×80 mesh result obtained from the finite element calculation. For missing temperature field locations near the boundary due to insufficient sensor coverage, if the consecutive missing values ​​do not exceed 3 meshes, the average value of adjacent valid meshes is used to fill the gap; if the consecutive missing values ​​exceed 3 meshes, the temperature interpolation result of adjacent finite element nodes at the same elevation is preferentially used to fill the gap, so that the final original temperature field matrix and the original stress field matrix are both 80×80, and the same row and column coordinates correspond to the same physical location on the dam body.

[0040] 3) Standardize the original temperature field matrix based on the statistical parameters of the training set to obtain the standardized temperature field matrix. , This represents the dimensionally unified temperature field matrix, with dimensions H×W, used for subsequent gradient calculations and network input.

[0041] In practical implementation, the mean and standard deviation of the temperature field in the training set are used for standardization, expressed as: ; in, The normalized temperature value at the i-th row and j-th column is represented by the normalized temperature field matrix. The element in the i-th row and j-th column; This represents the mean of temperature values ​​in the training set, used to characterize the overall center of temperature in the training set; ε represents the standard deviation of temperature values ​​in the training set, used to characterize the degree of temperature dispersion in the training set; ε represents a minimum constant to prevent the denominator from being zero, and can be taken as... .

[0042] 4) Standardize the original stress field matrix based on the statistical parameters of the training set to obtain the standardized stress field matrix. , This represents the dimensionally unified stress field matrix, with dimensions H×W, used for subsequent gradient calculations and network input.

[0043] In practical implementation, the mean and standard deviation of the training concentrated stress field are used for standardization, expressed as: ; in, The normalized stress value at the i-th row and j-th column is represented by the normalized stress field matrix. The element in the i-th row and j-th column; This represents the mean value of the concentrated stress in the training set, used to characterize the overall center of stress in the training set; The standard deviation of the stress values ​​in the training set represents the degree of stress dispersion in the training set; ε represents a constant to prevent the denominator from being zero, and can be taken as... .

[0044] It should be noted that the training set statistical parameters , , and It only calculates on the training set and directly reuses the statistical parameters obtained in the training stage in the verification, testing and online recognition stages, avoiding the introduction of future data or statistical bias of the current sample in the recognition stage.

[0045] (2) Under normal thermoelastic response, temperature field changes in a gravity dam typically induce corresponding stress field changes along the spatial direction. If the temperature gradient direction and stress gradient direction in a certain region are significantly inconsistent, it indicates that thermo-mechanical decoupling may occur in that region. This invention expresses this type of anomaly by calculating the gradient direction deviation response matrix. The specific steps are as follows: 1) The Sobel operator is used to calculate the gradient components of the normalized temperature field matrix in the horizontal and vertical directions, respectively, to obtain the horizontal temperature gradient matrix. and temperature vertical gradient matrix , This represents the gradient variation of the standardized temperature field matrix along the horizontal direction. This represents the gradient change of the standardized temperature field matrix along the vertical direction, and both have dimensions H×W.

[0046] In practical implementation, the standardized temperature field matrix... Convolution is performed using a horizontal Sobel kernel to obtain the temperature horizontal gradient matrix. For the standardized temperature field matrix Convolution is performed using a Sobel kernel in the vertical direction to obtain the temperature vertical gradient matrix. The convolution boundary can be filled by boundary copying to keep the output matrix size H×W.

[0047] Then, the gradient components of the normalized stress field matrix in the horizontal and vertical directions are calculated in the same way to obtain the stress horizontal gradient matrix. and stress vertical gradient matrix , This represents the gradient variation of the standardized stress field matrix along the horizontal direction. The gradient of the normalized stress field matrix along the vertical direction is represented by both H×W.

[0048] 3) Calculate the temperature gradient magnitude matrix based on the horizontal and vertical temperature gradient matrices. The stress gradient magnitude matrix is ​​calculated based on the stress horizontal gradient matrix and the stress vertical gradient matrix. ,in, This indicates the intensity of temperature change at each spatial location. This represents the intensity of stress change at each spatial location, and both have dimensions of H×W.

[0049] In practical implementation, the temperature gradient magnitude and stress gradient magnitude at the i-th row and j-th column can be expressed as: ; ; in, The value of the temperature gradient at the i-th row and j-th column is represented by the temperature gradient magnitude matrix. The element in the i-th row and j-th column; The stress gradient magnitude at the i-th row and j-th column is represented by the stress gradient magnitude matrix. The element in the i-th row and j-th column; The temperature horizontal gradient component at the i-th row and j-th column is represented by the temperature horizontal gradient matrix. The element in the i-th row and j-th column; The vertical temperature gradient component at the i-th row and j-th column is represented by the vertical temperature gradient matrix. The element in the i-th row and j-th column; The stress level gradient component at the i-th row and j-th column is represented by the stress level gradient matrix. The element in the i-th row and j-th column; The stress vertical gradient component at the i-th row and j-th column is represented by the stress vertical gradient matrix. The element in the i-th row and j-th column.

[0050] 4) Calculate the gradient direction deviation response matrix based on the temperature gradient direction and the stress gradient direction. , This indicates the degree of inconsistency between the gradient directions of the standardized temperature field matrix and the standardized stress field matrix at various spatial locations. Its size is H×W, and the larger the value, the more obvious the deviation between the temperature gradient direction and the stress gradient direction.

[0051] In practical implementation, the gradient direction deviation from the response at the i-th row and j-th column can be expressed as: ; in, The gradient direction deviation response at the i-th row and j-th column is represented by the gradient direction deviation response matrix. The element in the i-th row and j-th column; molecule The denominator represents the absolute value of the two-dimensional cross product of the temperature gradient vector and the stress gradient vector, used to characterize the degree of vertical deviation between the two gradient directions; This is used to normalize the gradient magnitude, preventing excessively large temperature or stress gradient magnitudes from causing deviations in response distortion; ε represents a minimal constant to prevent the denominator from being zero, and can be taken as... .

[0052] It should be noted that when the temperature gradient direction is nearly parallel to the stress gradient direction, Approaching 0; when the temperature gradient direction is nearly orthogonal to the stress gradient direction, The value is close to 1. Compared to directly calculating the inverse triangular angle, using the normalized cross product response reduces computational complexity and is sufficient to express the degree of directional deviation.

[0053] 5) Suppress flat regions to avoid false directional deviations caused by noise when the local gradient is close to 0.

[0054] In the specific implementation, let the threshold for determining flat regions be . , The parameter representing the region where the gradient magnitude is too small can be taken as follows: to ;when and At that time, directly ordered This indicates that there are no significant spatial changes in the temperature and stress fields at this location, and it is not considered a region of thermal decoupling anomaly.

[0055] In one embodiment, for example, at a certain grid location , , , If the temperature gradient is mainly along the horizontal direction, the stress gradient is also mainly along the horizontal direction, and the normalized cross product is relatively small. A value close to 0 indicates a consistent direction; if at another grid position... , , , If the temperature gradient is along the horizontal direction and the stress gradient is along the vertical direction, the normalized cross product is close to 1. A larger value indicates the possible existence of local thermal decoupling.

[0056] (3) While the gradient direction deviation response matrix alone can express directional inconsistency, it is still necessary to further enhance the expression for local coupling imbalances where "the temperature gradient is significant while the stress gradient is abnormally weak" or "the stress gradient is significant while the temperature gradient is abnormally weak". This invention constructs a coupling imbalance index matrix, enabling both directional deviation and amplitude imbalance to participate in the expression of anomalies. The specific steps are as follows: 1) For the temperature gradient magnitude matrix and stress gradient magnitude matrix Local smoothing is performed to obtain the smoothed temperature gradient magnitude matrix. and smooth stress gradient magnitude matrix ,in, Used to represent the steady-state intensity of temperature changes within a local neighborhood. Used to represent the stable strength of stress changes within a local neighborhood.

[0057] In practical implementation, a 5×5 mean filter or a Gaussian filter is used. and Smoothing is performed. The filter window size can be set according to the spatial grid resolution. For example, when H=64 and W=64, the filter window can be 5×5; when the spatial grid is denser, it can be 7×7. The purpose of local smoothing is to reduce the impact of individual grid noise on the judgment of coupling imbalance, so that subsequent exponents pay more attention to anomalies in continuous spatial regions.

[0058] 2) Calculate the gradient magnitude imbalance matrix based on the smoothed temperature gradient magnitude matrix and the smoothed stress gradient magnitude matrix. ,in, This indicates the degree of mismatch between the intensity of temperature change and the intensity of stress change, with dimensions of H×W.

[0059] In practical implementation, the gradient magnitude imbalance value at the i-th row and j-th column can be expressed as: ; in, The gradient magnitude imbalance value at the i-th row and j-th column is represented by the gradient magnitude imbalance matrix. The element in the i-th row and j-th column; This represents the magnitude of the smoothed temperature gradient at the i-th row and j-th column; The value represents the magnitude of the smoothed stress gradient at the i-th row and j-th column; ε represents a minimum constant to prevent the denominator from being zero, which can be taken as... When the temperature gradient is strong and the stress gradient is weak, or when the stress gradient is strong and the temperature gradient is weak, Increase.

[0060] 3) Based on the gradient direction deviation response matrix and gradient magnitude imbalance matrix Calculate the coupling imbalance index matrix B, where B represents the intensity of thermo-mechanical decoupling anomaly after comprehensively considering directional deviation and amplitude imbalance, and has a size of H×W.

[0061] In practical implementation, a weighted summation method is used to obtain the coupling imbalance index matrix, which is expressed as: ; in, This represents the coupling imbalance index at the i-th row and j-th column, and is an element in the i-th row and j-th column of the coupling imbalance index matrix B; This represents the direction deviation weight, used to control the contribution of the gradient direction deviation response to the coupling imbalance exponent; This represents the magnitude imbalance weight, used to control the contribution of gradient magnitude imbalance to the coupling imbalance exponent; and All are non-negative numbers and satisfy the following conditions: In a convenient implementation method, the following approach is preferable. , This indicates that priority is given to decoupling between the temperature gradient direction and the stress gradient direction, while also taking into account the magnitude mismatch.

[0062] Then, the coupling imbalance index matrix B is truncated and normalized to obtain the normalized coupling imbalance index matrix, still denoted as B. The truncation and normalization are used to avoid a few abnormally high values ​​dominating network training.

[0063] In practical implementation, the 1st and 99th percentiles of the coupling imbalance index in the training set can be statistically analyzed and denoted as follows: and ,in, The low-value cutoff boundary represents the coupling imbalance index of the training set. This represents the high-value cutoff boundary for the coupling imbalance index of the training set. For the current sample, it will be less than... The value is truncated to , will be greater than The value is truncated to Then it is linearly scaled to the range [0,1].

[0064] 4) Standardize the temperature field matrix Standardized stress field matrix Gradient direction deviates from response matrix and coupling imbalance index matrix By splicing along the channel direction, a coupling-enhanced 4-channel feature map is obtained. , This represents a feature map of the current gravity dam thermal monitoring sample after thermal coupling anomalous enhancement, with dimensions of 4×H×W.

[0065] In practical implementation, coupling enhancement is used for 4-channel feature maps. The first channel is the normalized temperature field matrix. The second channel is the normalized stress field matrix. The third channel represents the gradient direction deviation from the response matrix. The fourth channel is the coupling imbalance index matrix B.

[0066] In one embodiment, for example, if both the original temperature field matrix and the original stress field matrix are 64×64, then the coupling-enhanced 4-channel feature map... The dimensions are 4×64×64; the third channel can display the location where the temperature gradient and stress gradient directions are inconsistent, and the fourth channel can display the thermal decoupling intensity under the combined effect of directional deviation and amplitude imbalance.

[0067] It should be noted that this step does not simply input the original temperature field matrix and the original stress field matrix as two image channels into the network. Instead, it transforms the directional consistency and amplitude matching in the thermoelastic response into learnable features. This amplifies the early thermo-mechanical decoupling signals caused by microcracks, local stiffness degradation, heat transfer anomalies, and seepage channels before network training, reducing the subsequent network's dependence on a large number of anomalous samples.

[0068] In a specific implementation, step S3 is performed as follows: The thermal response of gravity dams is significantly affected by seasonal temperature differences, water level changes, solar radiation conditions, and operating conditions. Some health status changes do not manifest as obvious absolute amplitude anomalies in individual samples, but rather as local morphological deviations from historical typical operating patterns. Relying solely on the characteristics of the current sample itself can easily misjudge normal seasonal changes as anomalies, or misjudge early anomalies as normal fluctuations. This invention utilizes historical operating data with health level labels in the training set to construct a typical operating pattern library, and performs local matching between the current sample and the typical operating pattern library to generate a maximum response map and a variance map of the model response. These are then input into the network as historical reference features. The specific steps are as follows: (1) Under different health levels, the coupling patterns of the local temperature field and stress field of a gravity dam will exhibit different spatial patterns. For example, under normal conditions, the dam heel region may exhibit a stable thermal following pattern, while under abnormal conditions, a decoupling pattern may appear where the temperature gradient is concentrated but the stress gradient is weakened. This invention utilizes historical operating data labeled with health levels in the training set to extract typical local thermal coupling patterns under different health levels from the coupling enhancement 4-channel feature map, and constructs a typical operating pattern library. This pattern library provides a benchmark for subsequent pattern matching and historical reference. The specific steps are as follows: 1) Perform step S2 on each gravity dam thermal monitoring sample in the training set to obtain the coupled enhanced 4-channel feature map corresponding to that gravity dam thermal monitoring sample. .

[0069] In practical implementation, the coupled enhanced 4-channel feature map corresponding to the nth gravity dam thermal monitoring sample in the training set can be denoted as... , This represents the 4-channel feature map obtained after processing the nth gravity dam thermal monitoring sample in step S2, with dimensions of 4×H×W.

[0070] 2) Determine the key areas based on the location of the dam structure. Key areas may include the dam heel area, dam toe area, middle dam area, area near the upstream face, area near the downstream face, area near the gallery, and areas with existing cracks or seepage of concern.

[0071] In practical implementation, key regions can be determined using finite element mesh numbering, the geometric coordinate range of the dam body, or manually calibrated area masks. For example, the dam heel region can be composed of meshes within a certain range of the bottom upstream side, the dam toe region can be composed of meshes within a certain range of the bottom downstream side, and the central region of the dam body can be composed of meshes within a certain range near the center of the cross-section. If no key regions are preset, the entire dam cross-section can be used as candidate regions, but sampling should be prioritized in key regions to reduce interference from irrelevant background patterns on the pattern library.

[0072] In one embodiment, as an example, the dam cross-section grid is 64×64. The dam heel region is defined as the set of grids covered by row indices 56 to 64 and column indices 1 to 16, the dam toe region is defined as the set of grids covered by row indices 56 to 64 and column indices 49 to 64, and the central dam region is defined as the set of grids covered by row indices 28 to 36 and column indices 28 to 36. The key region mask is set to 1 in each designated region and 0 in all other locations. Subsequent extraction of local thermally coupled sample blocks and pattern clustering are performed within the range defined by the key region mask to reduce interference from non-key background regions on typical morphologies in the pattern library.

[0073] 3) Within each health level, extract local thermal coupling sample blocks from the coupling enhancement 4-channel feature map.

[0074] In the specific implementation, a sliding window is used to extract local thermally coupled sample blocks, centered on each grid location within the key region. The local window radius is defined as r, where r represents the number of grids extending upwards, downwards, leftwards, and rightwards from the center position; a value of 3 is suitable, resulting in a local window size of 7×7. The size of each local thermally coupled sample block is 4×(2r+1)×(2r+1). When the center position is close to the boundary, the coupling enhancement 4-channel feature map can be first copied and filled at the boundary before extracting the local thermally coupled sample blocks, ensuring that each local thermally coupled sample block has a consistent size.

[0075] In one embodiment, as an example, let the local window radius r=3 and the size of the local thermal coupling sample block be 4×7×7; for the grid position (30,30) in the key area, extract a 7×7 window centered at (30,30), and extract all the values ​​of the four channels in the window from the coupling enhancement 4-channel feature map to form a 4×7×7 local thermal coupling sample block; when the center position (2,30) is close to the boundary, first copy and fill the first row of the coupling enhancement 4-channel feature map upwards by 3 rows, and then extract a 7×7 window centered at (2,30) to ensure that the size of the extracted sample block is still 4×7×7.

[0076] 4) Flatten all the local thermally coupled sample blocks extracted under the same health level into vectors, and perform K-means clustering to obtain the prototype pattern corresponding to the health level.

[0077] In the specific implementation, the health level index is defined as g, with a value ranging from 0 to C-1; the number of prototype patterns under each health level is defined as K, where K represents the number of typical local thermal coupling patterns retained under that health level, and can be 3 to 8, for example, K=5. For health level g, all local thermal coupling sample blocks belonging to health level g are flattened into vectors and then K-means clustering is performed to obtain K cluster centers. Each cluster center is then restored to a tensor of size 4×(2r+1)×(2r+1), which serves as the prototype pattern under health level g.

[0078] In practical implementation, the k-th prototype pattern under the g-th health level is denoted as... Where k represents the prototype pattern index, and its value ranges from 1 to K; This represents the localized thermo-coupled spatial morphology that repeatedly appears in historical samples of health level g, with a size of 4×(2r+1)×(2r+1).

[0079] In one embodiment, as an example, assume a health level g=0, a number of prototype patterns K=5, and a local window radius r=3. The training set contains 5000 4×7×7 local thermal coupling sample blocks belonging to health level 0. Each sample block is flattened into a vector of length 4×7×7=196. K-means clustering is performed on these 5000 vectors to obtain 5 cluster center vectors. Each cluster center vector is then restored to a tensor of size 4×7×7, resulting in 5 prototype patterns under health level 0. These 5 prototype patterns represent the local thermal coupling morphology under different typical environmental conditions under health level 0, such as the normal morphology under high water levels in winter and the normal morphology under low water levels in summer.

[0080] Then, mean removal and norm normalization are performed on each prototype pattern to obtain the normalized prototype pattern, which is still denoted as . .

[0081] In the specific implementation, for each prototype pattern First, subtract the mean of all elements in the prototype mode, then divide by the sum of the Frobenius norm and the minimum constant of the prototype mode. This makes subsequent mode matching focus more on local spatial morphology, rather than being affected by the overall amplitude of temperature or stress.

[0082] In one embodiment, for example, the total number of health levels C=4, the number of prototype modes under each health level K=5, and the local window radius r=3. Then, the typical operating mode library includes a total of 20 prototype modes, each with a size of 4×7×7. Among them, the prototype modes under health level 0 mainly express the normal thermal coupling mode, and the prototype modes under health level 3 mainly express the abnormal thermal decoupling mode.

[0083] It should be noted that the typical operating mode library does not simply store the mean temperature or mean stress for each health level, but rather stores a four-channel local spatial model, including the combined morphology of the normalized temperature field, normalized stress field, gradient direction deviation response, and coupling imbalance index. Therefore, the typical operating mode library can express a historical reference of the "local structural morphology of the thermodynamic field," not just an amplitude reference.

[0084] (2) To quantitatively compare the local thermal coupling morphology of the current sample with historical typical operating modes, it is necessary to perform position-by-position pattern matching on the current sample based on the typical operating mode library, generating a maximum response map and a pattern response variance map. The maximum response map reflects the highest similarity between the current position and any historical typical mode, while the pattern response variance map reflects the degree of dispersion of the current position's response to different modes. Both serve as historical reference features, enabling the recognition network to perceive the degree of deviation between the current morphology and historical typical morphology, such as... Figure 4 As shown, the specific steps for generating pattern matching feature maps are as follows: 1) Perform step S2 on the current gravity dam thermal monitoring sample to be identified to obtain the coupled enhanced 4-channel feature map of the current sample. .

[0085] 2) Enhance the 4-channel feature map with the coupling of the current sample. Using each spatial location as the center, extract the current local thermal coupling sample block. , This represents a local thermally coupled sample block extracted with the i-th row and j-th column as the center, with a size of 4×(2r+1)×(2r+1).

[0086] In practical implementation, before extracting the current local thermal coupling sample block, it is possible to... Perform boundary copying and padding to ensure that the entire window can also be extracted at the boundary locations. Further, for... Perform the same mean removal and norm normalization as the prototype pattern to make it match the typical running pattern library on the same scale.

[0087] In one embodiment, as an example, suppose the local window radius r = 3, for the coupling enhancement 4-channel feature map Each boundary is copied outwards to fill 3 rows or 3 columns; for the spatial location (1,1), its local thermally coupled sample block The center is located at the top left corner of the original feature map, and the window covers rows -2 to 4 and columns -2 to 4 after filling, resulting in a sample block of size 4×7×7. Perform mean removal and norm normalization: first subtract The mean of the 196 elements in the dataset, divided by... The sum of the Frobenius norm and ε makes It is on the same matching scale as the prototype pattern.

[0088] 3) Calculate the current local thermal coupling sample block Each prototype pattern in the typical runtime pattern library The normalized cross-correlation response values ​​between them are used to obtain the model response values. .

[0089] In the specific implementation, the current local thermally coupled sample block and each prototype pattern Perform normalized cross-correlation calculation, expressed as follows: ; in, This represents the normalized cross-correlation response value between the current local thermally coupled sample block at the i-th row and j-th column and the k-th prototype mode of the g-th health level. This represents the sum of element-wise products of the current local thermally coupled sample block and the prototype mode at all four channels and window positions. This represents the Frobenius norm of the current local thermally coupled sample block; The Frobenius norm of the prototype mode is represented; ε represents a minimal constant to prevent the denominator from being zero, which can be taken as... .

[0090] It should be noted that the larger the normalized cross-correlation response value, the more similar the local thermo-coupling pattern at the current location is to the corresponding prototype mode; the smaller the normalized cross-correlation response value, the more significant the difference between the local thermo-coupling pattern at the current location and the corresponding prototype mode.

[0091] 4) Traverse all health levels and all prototype patterns to obtain C×K pattern response maps; then, at each spatial location, calculate the maximum value of all pattern response values ​​to obtain the pattern maximum response map. , This represents the highest similarity between the current sample and the typical operating mode library at each spatial location, with a size of H×W.

[0092] In the actual implementation, for the i-th row and j-th column, all... Compare and take the maximum value as .when A higher value indicates that the current position can find a similar local pattern in the historical typical operation pattern library; when... A lower value indicates that the local shape at the current location deviates from the historical typical pattern, which may be an abnormal or rare operating state.

[0093] 5) Calculate the variance of all model response values ​​at each spatial location to obtain the model response variance plot. , This represents the degree of dispersion of the current location's response to different health levels and different prototype modes, with a size of H×W.

[0094] In the actual implementation, for the i-th row and j-th column, all... As a set of numerical calculations, the variance is obtained , This represents all pattern response values ​​at row i and column j. The variance is the variance of the model response variance plot. The element in the i-th row and j-th column. When A higher value indicates that the current location responds highly to some modes and poorly to others, demonstrating strong discriminative power; when lower and A lower value indicates that the current location is dissimilar to all historical patterns, which may correspond to novel anomalies or local noise.

[0095] 6) Couple and enhance the 4-channel feature map Maximum response graph of the mode Pattern response variance plot By splicing along the channel direction, the network input feature map is obtained. , This represents a 6-channel feature map used as input to the subsequent recognition network, with dimensions of 6×H×W.

[0096] In practical implementation, the network input feature map The first channel is the normalized temperature field matrix. The second channel is the normalized stress field matrix. The third channel represents the gradient direction deviation from the response matrix. The fourth channel is the coupling imbalance exponential matrix B, and the fifth channel is the mode maximum response map. The 6th channel is the mode response variance plot. .

[0097] In one embodiment, for example, if H=64 and W=64, then the network input feature map The dimensions are 6×64×64; if the total number of health levels C=4 and the number of prototype patterns under each health level K=5, then a total of 20 pattern response maps are calculated during the pattern matching process, and then 1 pattern maximum response map and 1 pattern response variance map are obtained from the 20 pattern response maps.

[0098] It should be noted that this step provides historical references for the current sample through a typical operating mode library, enabling the network not only to see the thermal decoupling anomalies in the current sample, but also to determine whether the local thermal coupling pattern at the current location deviates from the typical patterns in the historical normal, watch, warning, or abnormal states. This can reduce misjudgments caused by seasonal temperature differences and water level changes, and improve the sensitivity to early changes in health level.

[0099] In a specific implementation, step S4 is performed as follows: The dual-branch gated fusion network consists of three parts: a physically coupled branch, a pattern matching branch, and a bidirectional gated fusion network. The physically coupled branch takes the physically coupled feature map as input and extracts high-level features of thermal coupling anomalies through convolutional layers and a thermal decoupling gating mechanism. The pattern matching branch takes the pattern matching feature map as input and extracts high-level features of historical pattern deviations through dilated convolution and a channel attention mechanism. After the two branches independently extract features, they interact and fuse information through the bidirectional gated fusion network, ultimately outputting the health level prediction result. This structure decouples the extraction processes of physically coupled features and pattern matching features, avoiding mutual interference between the two types of heterogeneous information during convolution.

[0100] Network input feature map It contains two types of heterogeneous information: the first four channels belong to physical coupling features, expressing temperature field, stress field, and thermo-decoupling anomalies; the latter two channels belong to pattern matching features, expressing the similarity and discriminability between the current sample and historical typical operating modes. If the six channels are directly input into a regular convolutional network, the physical coupling features and pattern matching features may interfere with each other. This invention constructs a thermo-decoupling-sensitive dual-branch gated fusion network, in which the physical coupling branch specifically processes physical coupling features, and the pattern matching branch specifically processes pattern matching features. Information interaction is then achieved through a bidirectional gated fusion mechanism, ultimately outputting the gravity dam health level prediction result. The specific steps are as follows: (1) The physical coupling branch specifically handles the four physical coupling channels: the normalized temperature field matrix, the normalized stress field matrix, the gradient direction deviation response matrix, and the coupling imbalance index matrix. This branch consists of three convolutional layers and one thermal decoupling gating module. The thermal decoupling gating module uses the coupling imbalance index matrix to generate spatial gating weights, which enhance the mid-layer features position by position, so that the network focuses on the thermal decoupling region in subsequent convolutions, such as... Figure 5 As shown, the specific steps for extracting features from physically coupled branches are as follows: 1) Input feature map into the network Decomposed by channel, the physically coupled input feature map is obtained. , Represents the network input feature map The first four channels, with dimensions of 4×H×W, include a normalized temperature field matrix. Standardized stress field matrix Gradient direction deviates from response matrix And the coupling imbalance index matrix B.

[0101] 2) Physically couple the input feature map The first convolutional layer of the input physical coupling branch is used to obtain the initial feature map of physical coupling. ,in, The shallow spatial characteristics of the physically coupled branch are represented by a dimension of 16×H×W.

[0102] In the specific implementation, the first convolutional layer can use a 3×3 convolution with 16 output channels, a stride of 1, padding of 1, followed by batch normalization and ReLU activation function; this layer is used to extract the local spatial texture of temperature field, stress field and thermal decoupling channel.

[0103] 3) Physically couple the initial feature map The second convolutional layer of the input physical coupling branch is used to obtain the feature map of the physical coupling middle layer. , This represents the mid-layer features of the physically coupled branches after spatial downsampling, with a size of [size missing]. .

[0104] In practice, the second convolutional layer can be a 3×3 convolution with 32 output channels, a stride of 2, padding of 1, followed by batch normalization and ReLU activation function. Indicates the number of spatial rows after downsampling. This indicates the number of spatial columns after downsampling. It can be set to H / 2 or W / 2. If H or W is not divisible by 2, boundary padding can be used to make the output size an integer.

[0105] 4) Adjust the coupling imbalance index matrix B to the physical coupling layer feature map. The spatial dimensions are used to obtain the scale-fit coupling imbalance index matrix. , This represents a coupling imbalance guide map with a spatial size consistent with the physical coupling layer feature map, and the size is... .

[0106] In the specific implementation, bilinear interpolation is used to scale B from H×W to [the desired value]. Since B is used to indicate the strength of thermal decoupling, its spatial position must be aligned with the feature map of the physical coupling layer.

[0107] 5) Based on the scale-adaptive coupling imbalance index matrix Generate a thermally decoupled gating weight graph , This represents a weighted graph in the physically coupled branch used to enhance the response in the thermally decoupled region, with a size of [size missing]. .

[0108] In the specific implementation, Input a 1×1 convolutional layer, map the number of channels from 1 to 32, and then pass it through a Sigmoid activation function to obtain... The Sigmoid activation function restricts the weights to between 0 and 1, making the gating process stable and controllable.

[0109] 6) Decouple the thermal gating weighted graph Effect on the feature map of the physical coupling layer The gating enhanced physical coupling feature map is obtained. , This represents the physical coupling characteristics enhanced by the thermal decoupling region, with a size of [missing information]. .

[0110] In practical implementation, thermal decoupling gating weight graphs will be used. Feature maps in physical coupling Channel-by-channel, position-by-position weighted enhancement is performed, and the calculation method is expressed as follows: ; in, This represents element-wise multiplication; Indicates and A matrix of all ones of the same size; This represents the gating enhancement coefficient, used to control the influence of the coupling imbalance index matrix on the enhancement of physical coupling features. It can be taken from 0.5 to 1.5, for example, 1.

[0111] It should be noted that the following is adopted: As an enhancing factor, rather than directly using By replacing the original features, the response of the thermal decoupling region can be improved while retaining the original physical coupling features, and the weak anomaly region can be completely suppressed.

[0112] 7) Enhance the physical coupling feature map of the gating system. The third convolutional layer of the input physical coupling branch is used to obtain the high-level feature map of physical coupling. ,in, This represents the high-level feature of the final output of the physically coupled branch, with a size of .

[0113] In practice, the third convolutional layer can be a 3×3 convolution with 64 output channels, a stride of 1, padding of 1, followed by batch normalization and ReLU activation function.

[0114] It should be noted that the physical coupling branch generates a thermally decoupling gated weight map through the coupling imbalance index matrix B, so that the network focuses on the location of abnormal coupling relationship between temperature field and stress field when extracting features, rather than relying solely on ordinary convolution to automatically learn abnormal regions, thereby enhancing the local response corresponding to microcracks, local stiffness degradation and seepage channels.

[0115] (2) The pattern matching branch specifically processes the two historical pattern matching channels: the maximum response map and the pattern response variance map. This branch consists of three convolutional layers and one channel attention module. The first convolutional layer uses dilated convolution to increase the receptive field. The channel attention module adaptively weights each channel, strengthening the pattern matching channels that contribute significantly to health level differentiation and suppressing irrelevant or noisy pattern responses, such as... Figure 6 As shown, the specific steps for feature extraction from pattern matching branches are as follows: 1) Input feature map into the network Split by channel to obtain the pattern matching input feature map. , Represents the network input feature map The last two channels, with dimensions of 2×H×W, include the mode maximum response map. Pattern response variance plot .

[0116] 2) Input pattern matching into the feature map The first convolutional layer of the input pattern matching branch is used to obtain the initial feature map for pattern matching. , The shallow spatial features of the pattern matching branch are represented by a size of 16×H×W.

[0117] In practical implementation, the first convolutional layer can use a 5×5 dilated convolution with a dilation coefficient of 2, 16 output channels, and a stride of 1, followed by batch normalization and ReLU activation function. The 5×5 dilated convolution can expand the receptive field with fewer parameters, making it easier to capture the regional distribution trends in the mode maximum response map and mode response variance map.

[0118] 3) Initialize the feature map for pattern matching. The second convolutional layer of the input pattern matching branch is used to obtain the feature map of the middle layer of pattern matching. , This represents the mid-level features of the pattern matching branch after spatial downsampling, with a size of [size missing]. .

[0119] In practice, the second convolutional layer can be a 3×3 convolution with 32 output channels, a stride of 2, and padding of 1. It is followed by batch normalization and ReLU activation function to make the feature map in the pattern matching layer have the same spatial size as the feature map in the physical coupling layer.

[0120] 4) For the feature map in the pattern matching layer Perform channel attention calculation to obtain the pattern channel attention weights. ,in, The importance weight of each channel in the pattern matching branch is represented by a dimension of 32.

[0121] In the specific implementation, first... Global average pooling is performed on each channel to obtain a channel statistical vector of length 32. This channel statistical vector is then sequentially fed into a dimensionality-reduced fully connected layer, a ReLU activation function, an up-dimensional fully connected layer, and a Sigmoid activation function to obtain the mode channel attention weights. The dimensionality reduction ratio can be set to 4, which means first reducing the 32 dimensions to 8 dimensions, and then restoring them to 32 dimensions.

[0122] In one embodiment, as an example, suppose the feature map in pattern matching is... The dimensions are 32×32×32. The channel attention calculation process is as follows: First, global average pooling is performed on each channel to obtain a channel statistical vector of length 32; this vector is then input into a fully connected layer with an output dimension of 8 (32 / 4=8) for dimensionality reduction, followed by a ReLU activation function; finally, a fully connected layer with an output dimension of 32 is input to restore the dimension, and the channel attention weights are obtained by a Sigmoid activation function. The weight vector has a length of 32, and each element takes a value between 0 and 1, controlling... The degree of enhancement or suppression of the corresponding channel.

[0123] 5) Adjust the attention weights of the pattern channels Feature maps in pattern matching Channel-by-channel multiplication yields the channel-enhanced pattern matching feature map. , This represents the pattern matching feature after channel selection, with a size of [size missing]. .

[0124] In the specific implementation, Expand to After scaling, it is compared with the feature map in the pattern matching middle layer. Element-wise multiplication can improve the response of the pattern-matching channel that contributes more to the differentiation of health levels, while reducing the influence of irrelevant or noisy mode responses.

[0125] 6) Enhance the channel pattern matching feature map The third convolutional layer of the input pattern matching branch is used to obtain the high-level feature map of pattern matching. , This represents the high-level features of the final output of the pattern matching branch, with a size of [size missing]. .

[0126] In practice, the third convolutional layer can be a 3×3 convolution with 64 output channels, a stride of 1, padding of 1, followed by batch normalization and ReLU activation function.

[0127] It should be noted that the pattern matching branch is specifically designed to handle historical pattern reference information, enabling the network to learn the "deviation relationship between the current local morphology and the historical typical operating mode". This branch is set up separately from the physical coupling branch, which can prevent the historical pattern matching information from being overwhelmed by the strong amplitude information of the temperature field and stress field.

[0128] (3) Dual-branch bidirectional gating fusion and health level output 1) Physically coupled high-level feature maps Perform global average pooling to obtain the physically coupled global vector. , This represents a global response summary of the physically coupled branch, with a dimension of 64.

[0129] Simultaneously, high-level feature maps for pattern matching Perform global average pooling to obtain the global vector for pattern matching. ,in, This represents the global response summary of the pattern matching branch, with a dimension of 64.

[0130] 2) Match global vectors based on patterns Generate physical branch adjustment weights , This represents the adjustment weight of the physical coupling branch channel generated by pattern matching information, with a dimension of 64.

[0131] In practical implementation, to enhance the relevant channels of physically coupled branches when historical pattern matching results indicate that certain regions or channels are more discriminative, the global pattern matching vector is used. Input a fully connected layer and a sigmoid activation function to obtain the physical branch adjustment weights. Then, adjust the weights of the physical branches. Expand to The dimensions (copying and expanding the vector in spatial dimensions to 64×) × The size), used for physically coupling high-level feature maps Multiplication or addition enhances the effect.

[0132] At the same time, based on the physical coupling global vector Generative mode branch adjustment weights , This represents the adjustment weight of the pattern matching branch channel generated by the physical coupling information, with a dimension of 64.

[0133] In practical implementation, to enhance the channels in the pattern matching branch related to the abnormal morphological reference when a significant thermal decoupling anomaly is detected in the physically coupled branch, the physical coupling global vector is strengthened. Inputting a fully connected layer and a sigmoid activation function yields the mode branch adjustment weights. Then, adjust the weights of the pattern branches. Expand to The dimensions (copying and expanding the vector in spatial dimensions to 64×) × The size), used for pattern matching high-level feature maps. Multiplication or addition enhances the effect.

[0134] 3) High-level feature maps based on physical coupling and mode branch adjustment weight and pattern matching high-level feature maps Adjusting weights for physical branches Interactive enhanced physical coupling feature maps were obtained through residual enhancement. and interaction enhancement pattern matching feature map .

[0135] In practical implementation, the original feature map is superimposed on the element-wise multiplication enhancement to preserve the original information. The calculation method is expressed as follows: ; ; in, The shape is 64× × A matrix in which all elements are 1; The interactive enhanced physical coupling feature map represents the enhanced high-level features of physical coupling after being guided by pattern matching information. The interactive enhanced pattern matching feature map represents the enhanced high-level features of pattern matching after being guided by physically coupled information; the physically coupled global vector... Adjusting weights for physical branches Before computation, all data are expanded to the spatial dimensions of the corresponding feature map (i.e., 64×). × ); This indicates element-wise multiplication.

[0136] 4) Enhance the physical coupling feature map with interaction Feature maps matching interaction enhancement patterns By stitching along the channel direction, a stitched and fused feature map is obtained. , This represents the intermediate feature after feature fusion of two branches, with a size of [size missing]. .

[0137] Then, the feature maps are spliced ​​and fused. Input a 1×1 convolutional layer to obtain a fused high-level feature map. , This represents the high-level feature after the completion of dual-branch information fusion, with a size of [size missing]. .

[0138] In practical implementation, a 1×1 convolutional layer is used to reassemble the channel information of the physical coupling branch and the pattern matching branch. The number of output channels can be set to 128, followed by batch normalization and ReLU activation function.

[0139] Then, merge the high-level feature maps. Perform global average pooling to obtain the classification feature vector v, where v represents the global health status of the current gravity dam thermal monitoring sample with a dimension of 128.

[0140] 5) Input the classification feature vector v into the fully connected classification layer to obtain the classification logical value vector z, where z represents the unnormalized score of the health level classification and has dimension C; when C=4, z contains 4 elements, corresponding to health level 0, health level 1, health level 2 and health level 3 respectively.

[0141] Then, Softmax normalization is performed on the classification logical value vector z to obtain the health level prediction probability vector p, where p represents the probability that the current gravity dam thermal monitoring sample belongs to each health level, with dimension C; the health level with the highest probability is used as the preliminary health level identification result output by the network.

[0142] In a specific implementation, step S5 is performed as follows: The health levels of a gravity dam exhibit a clear order relationship, with health levels 0 to 3 representing a gradual deterioration in condition. Ordinary cross-entropy loss only focuses on whether the classification is correct, failing to distinguish the severity of misclassifying a health level 0 as health level 1 versus misclassifying it as health level 3. This invention employs an order-weighted loss for training, thus penalizing erroneous predictions that deviate significantly from the true health level. The specific steps are as follows: 1) Perform steps S2 and S3 sequentially on each gravity dam thermal monitoring sample in the training set to obtain the corresponding network input feature map. Then, the network is input with feature maps. Input the thermally decoupled sensitive dual-branch gated fusion network constructed in step S4 to obtain the health level prediction probability vector.

[0143] 2) Definition Let represent the probability that the nth training sample is predicted to be of health level c, where n represents the training sample index, c represents the health level index, and c ranges from 0 to C-1; (Definition) This represents the true health level label of the nth training sample. For sparse integer labels.

[0144] 3) Calculate the class balance weights based on the number of samples for each health level in the training set; define... This represents the number of samples corresponding to health level c in the training set, defined as follows: This represents the largest sample size across all health levels. definition This represents the category balance weight corresponding to health level c. In the specific implementation, it can be set as follows: ,in, This represents the category balance weight for health level c; This represents the number of samples corresponding to the health level with the largest number of samples in the training set. ε represents the number of samples corresponding to health level c in the training set; ε represents a minimal constant to prevent the denominator from being zero, which can be taken as... .

[0145] It should be noted that the fewer the number of samples a health level has, the greater the corresponding category balance weight, thus receiving more attention during training.

[0146] 4) Calculate the order relation weighted loss Loss, which represents the total loss value within a training batch and is used to guide the network parameter update.

[0147] In its implementation, the loss function consists of a class-balanced cross-entropy term and an order relation penalty term, and its calculation method is as follows: ; in, This indicates the number of samples in a training batch. Indicates the true health level Corresponding category balance weights; Let represent the probability that the nth training sample is predicted to be the true health level; ε represents a minimal constant to prevent logarithmic overflow, which can be taken as εn. ; The order relation penalty coefficient is used to control the proportion of the health level distance penalty in the total loss, and can be taken as 0.1 to 0.5; This represents the difference between the predicted health level (c) and the actual health level. The hierarchical distance between them.

[0148] It should be noted that the first term of the loss function is the class-balancing cross-entropy, used to handle the problem of imbalanced sample sizes across different health levels; the second term is the order-based penalty term, used to penalize misclassifications at long distances. For example, if the model assigns a higher probability to health level 3 when the true health level is 0, then |3-0|=3, resulting in a larger penalty; if the model assigns a higher probability to health level 1, then |1-0|=1, resulting in a relatively smaller penalty. This design conforms to the engineering principle of the gradual deterioration of the health state of gravity dams.

[0149] 5) The Adam optimizer is used to update all trainable parameters in the thermally decoupled sensitive dual-branch gated fusion network. The trainable parameters include the convolution parameters of the physically coupled branch, the thermally decoupled gate parameters, the convolution parameters of the pattern matching branch, the channel attention parameters, the bidirectional gated fusion parameters, and the fully connected classification layer parameters.

[0150] In one implementation, for example, the initial learning rate can be chosen as follows: Batch size A value of 32 can be chosen, and the weight decay coefficient can be chosen as follows: The number of training rounds can be 200. If the validation set loss does not decrease for 10 consecutive training rounds, the learning rate can be reduced to 0.5 of the original value. During training, the model parameters corresponding to the highest macro-average F1 score of the validation set are retained as the final recognition model parameters.

[0151] 6) After each training round, evaluate the model performance using the validation set. Evaluation metrics may include overall accuracy, F1 score for each health level, and macro-average F1 score. The macro-average F1 score is obtained by taking the arithmetic mean of the F1 scores for health levels 0, 1, 2, and 3, which can avoid the evaluation results being biased towards normal states due to a large number of normal samples.

[0152] In a specific implementation, step S6 is performed as follows: After training is completed, this invention can be deployed in gravity dam safety monitoring methods to perform online health level identification of newly acquired temperature and stress fields. The specific steps are as follows: 1) Collect the original temperature field matrix and original stress field matrix at the current running moment using the same data format as in the training phase, and denot them as follows: and ,in, This represents the original temperature field matrix at the current running moment. This represents the original stress field matrix at the current operating moment, and both dimensions are adjusted to H×W.

[0153] In the specific implementation, if there is only discrete monitoring point data at the current running time, the matrix is ​​generated using the same spatial interpolation method as in the training phase; if there are a few missing sensor measurements at the current running time, the missing measurement filling method is used as in the training phase to ensure that the input format is consistent.

[0154] 2) The original temperature field matrix at the current operating moment and the original stress field matrix By performing step S2, the coupled enhanced 4-channel feature map at the current runtime is obtained. .

[0155] It should be noted that the standardized parameters used in the online recognition phase still use those obtained during the training phase. , , and This ensures that online samples and training samples are on the same numerical scale by not recalculating standardized parameters using the current samples.

[0156] 3) Using the typical running mode library built during the training phase, enhance the coupling of the 4-channel feature map at the current running moment. Perform the operation in step S3 to obtain the pattern maximum response graph. Pattern response variance plot .

[0157] It should be noted that the typical operating mode library is not rebuilt during the online recognition phase, but is directly called from the typical operating mode library obtained during the training phase to ensure that the historical reference standard is fixed. If new manually confirmed samples are accumulated in the future, the typical operating mode library can be updated during the model retraining phase.

[0158] 4) Enhance the coupling of the current runtime 4-channel feature map Maximum response graph of the mode Pattern response variance plot By splicing along the channel direction, the network input feature map at the current running time is obtained. The dimensions are 6×H×W.

[0159] 5) Input the network feature map at the current running time. Input the trained thermally decoupled sensitive dual-branch gated fusion network to obtain the health level prediction probability vector p; then, take the health level with the highest prediction probability as the health level identification result at the current running time, and take the highest prediction probability as the classification confidence.

[0160] In practical implementation, if the health level identification result is health level 2 or health level 3, and the classification confidence is greater than the preset alarm threshold, a warning or alarm will be triggered (the alarm threshold can be 0.7); if the health level is identified as health level 2 or health level 3 for 2 or 3 consecutive running times, a trend alarm can also be triggered to reduce false alarms caused by occasional noise.

[0161] Example 2 like Figure 7 As shown in the figure, the changes in the probability of predicting abnormal states using the method of this invention and the conventional stress field method are compared using a line graph simulation of a gravity dam operating online for 120 days. Figure 7 The horizontal axis represents the running time (days), and the vertical axis represents the predicted probability of abnormal states (dimensionless), with values ​​ranging from 0 to 1. Figure 7The conventional stress field method refers to a traditional monitoring and analysis approach that uses only the stress field matrix as input, combined with the moving average and threshold judgment, without using temperature field information or modeling the thermo-mechanical coupling relationship. The background blocks, in light orange and light red, mark the gradual change phase (days 60 to 80) and the abnormal operation phase (after day 80), respectively, reflecting the gradual deterioration of the actual health status. The prediction probability of the method in this invention remains at a low level with minimal fluctuation during the first 60 days of normal operation, steadily increases during the gradual change phase, and stabilizes at a high level after entering the abnormal phase. The overall trend closely matches the state changes with low noise. This invention effectively filters out normal fluctuations caused by seasonal temperature and operating condition changes through anomaly enhancement of the coupling gradient between the temperature field and stress field, as well as historical pattern matching. It can promptly improve the early warning probability when thermo-mechanical decoupling occurs in the early stages, while maintaining high stability and significantly reducing the risk of false alarms and missed alarms.

[0162] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for evaluating the operating health level of a gravity dam by combining temperature field and stress field, characterized in that, The specific steps are as follows: S1. Collect data to construct a thermal monitoring sample of gravity dam. Make a one-to-one correspondence between the original temperature field matrix and the original stress field matrix in the sample in the grid space, and assign a real health level label to each sample. Then divide the dataset composed of the samples into training set, validation set and test set. S2. Enhance the samples by generating a coupled enhancement feature map through grid space uniformity, numerical standardization, gradient direction deviation calculation, and local coupling imbalance calculation. The specific operations for grid space uniformity and numerical standardization are as follows: Using the original temperature field matrix and original stress field matrix from the gravity dam thermal monitoring sample as input, spatial grid consistency and numerical standardization are performed on the two. First, check whether the spatial grids of the original temperature field matrix and the original stress field matrix are consistent. If the two have different sizes or inconsistent spatial coordinates, they are resampled to the same spatial grid. The original temperature field matrix is ​​standardized based on the statistical parameters of the training set, specifically by using the mean and standard deviation of the temperature fields in the training set to obtain the standardized temperature field matrix; the original stress field matrix is ​​also standardized based on the statistical parameters of the training set, specifically by using the mean and standard deviation of the stress fields in the training set to obtain the standardized stress field matrix. The specific steps for calculating gradient direction deviation are as follows: By calculating the gradient direction deviation response matrix, the Sobel operator is used to calculate the gradient components of the normalized temperature field matrix in the horizontal and vertical directions respectively, to obtain the temperature horizontal gradient matrix and temperature vertical gradient matrix. Then, the gradient components of the normalized stress field matrix in the horizontal and vertical directions are calculated in the same way to obtain the stress horizontal gradient matrix and stress vertical gradient matrix. Calculate the temperature gradient magnitude matrix based on the horizontal and vertical temperature gradient matrices, calculate the stress gradient magnitude matrix based on the horizontal and vertical stress gradient matrices, and calculate the gradient direction deviation response matrix based on the temperature gradient direction and the stress gradient direction. The magnitudes of the temperature gradient and stress gradient are compared with the set threshold for determining flat regions, and flat regions are suppressed based on the comparison results. The calculation process of the coupling enhancement feature map is as follows: The temperature gradient magnitude matrix and stress gradient magnitude matrix are locally smoothed to obtain smoothed temperature gradient magnitude matrix and smoothed stress gradient magnitude matrix. The gradient magnitude imbalance matrix is ​​calculated based on the smoothed temperature gradient magnitude matrix and smoothed stress gradient magnitude matrix. Then, the coupling imbalance index matrix is ​​calculated by weighted summation based on the gradient direction deviation response matrix and the gradient magnitude imbalance matrix. The coupling imbalance index matrix is ​​then truncated and normalized to obtain the normalized coupling imbalance index matrix. The normalized temperature field matrix, normalized stress field matrix, gradient direction deviation response matrix and coupling imbalance index matrix are concatenated along the channel direction to obtain the coupling enhancement feature map, which contains four channels. S3. Construct a typical operating mode library, perform local matching between the shape of the coupling enhancement feature map corresponding to each sample and the typical operating mode library, generate the maximum response map and the variance map of the mode, and then concatenate them with the coupling enhancement feature map to generate the network input feature map. The specific operations in S3 are as follows: Based on the location of the dam structure, key areas are determined. Within each health level, local thermal coupling sample blocks are extracted from the coupling enhancement feature map using a sliding window, with each grid location within the key area as the center. All local thermal coupling sample blocks extracted under the same health level are flattened into vectors and K-means clustering is performed to obtain the prototype pattern corresponding to the health level. The mean is removed and the norm is normalized for each prototype pattern to obtain the normalized prototype pattern, which constitutes a typical operation mode library. Calculate the normalized cross-correlation response value between the current local thermal coupling sample block and each prototype mode in the typical operating mode library to obtain the mode response value. Traverse all health levels and all prototype modes, and count the maximum value of all mode response values ​​at each spatial location to obtain the mode maximum response map. Calculate the variance of all mode response values ​​at each spatial location to obtain the mode response variance map. Finally, concatenate the coupling enhancement feature map, the mode maximum response map, and the mode response variance map along the channel direction to obtain the network input feature map. S4. Construct a dual-branch gated fusion network, including a physically coupled branch, a pattern matching branch, and a bidirectional gated fusion network. Input the network input feature map into the two branches respectively, and then exchange information through the bidirectional gated fusion mechanism to output the gravity dam health level prediction result. S5. The dual-branch gated fusion network is trained using order relation weighted loss to obtain the trained dual-branch gated fusion network. S6. Deploy the training dual-branch gating fusion network in the gravity dam safety monitoring method to assess the operational health registration of the gravity dam.

2. The method for assessing the operational health level of gravity dams by combining temperature and stress fields as described in claim 1, characterized in that, The S1 operation is as follows: A thermal monitoring sample for gravity dams is constructed by collecting data from internal temperature sensors, strain gauges or stress inversion results, finite element calculation results, monitoring interpolation results, and manual inspection confirmation results. Spatial position calibration was performed on the original temperature field matrix and original stress field matrix in the thermal monitoring samples of the gravity dam. Using the unified cross-sectional coordinate system of the dam body as the reference, the original temperature field matrix and original stress field matrix were resampled to a unified spatial grid. A health level label is assigned to each gravity dam thermal monitoring sample, and the health level is divided into normal state, attention state, warning state and abnormal state; All gravity dam thermal monitoring samples were divided into training set, validation set and test set.

3. The method for assessing the operational health level of gravity dams by combining temperature and stress fields according to claim 1, characterized in that, The physical coupling branch operations in S4 are as follows: The network input feature map has six channels, containing two types of heterogeneous information: the first four channels belong to physical coupling features, and the last two channels belong to pattern matching features. The network input feature map is split by channel to obtain the physical coupling input feature map. This physical coupling feature map is then input into the first convolutional layer of the physical coupling branch to obtain the initial physical coupling feature map. The initial physical coupling feature map is then input into the second convolutional layer of the physical coupling branch to obtain the middle physical coupling feature map. The coupling imbalance index matrix in the coupling enhancement feature map is then adjusted to the spatial size of the middle physical coupling feature map to obtain the scale-adapted coupling imbalance index matrix. A thermal decoupling gated weight map is generated based on the scale-adapted coupling imbalance index matrix. This thermal decoupling gated weight map is applied to the middle physical coupling feature map to obtain the gated enhanced physical coupling feature map. Finally, the gated enhanced physical coupling feature map is input into the third convolutional layer of the physical coupling branch to obtain the high-level physical coupling feature map.

4. The method for assessing the operational health level of gravity dams by combining temperature and stress fields according to claim 3, characterized in that, The specific pattern matching branch operations in S4 are as follows: The network input feature map is split by channel to obtain the pattern matching input feature map. The pattern matching input feature map is then fed into the first convolutional layer of the pattern matching branch to obtain the initial pattern matching feature map. The initial pattern matching feature map is then fed into the second convolutional layer of the pattern matching branch to obtain the intermediate pattern matching feature map. Channel attention is calculated on the intermediate pattern matching feature map to obtain the pattern channel attention weights. The pattern channel attention weights are then multiplied by the intermediate pattern matching feature map channel by channel to obtain the channel-enhanced pattern matching feature map. The channel-enhanced pattern matching feature map is then fed into the third convolutional layer of the pattern matching branch to obtain the high-level pattern matching feature map.

5. The method for assessing the operational health level of gravity dams by combining temperature and stress fields according to claim 4, characterized in that, The specific operation of the bidirectional gated fusion network in S4 is as follows: Global average pooling is performed on the physical coupling high-level feature map to obtain a physical coupling global vector. Simultaneously, global average pooling is performed on the pattern matching high-level feature map to obtain a pattern matching global vector. Physical branch adjustment weights are generated based on the pattern matching global vector, and pattern branch adjustment weights are also generated based on the physical coupling global vector. Based on the physical coupling high-level feature map and the pattern branch adjustment weights, as well as the pattern matching high-level feature map and the physical branch adjustment weights, interactively enhanced physical coupling feature maps and interactively enhanced pattern matching feature maps are obtained through residual enhancement. These interactively enhanced physical coupling feature maps and interactively enhanced pattern matching feature maps are concatenated along the channel direction to obtain a concatenated fused feature map. This concatenated fused feature map is then input into a 1×1 convolutional layer to obtain a fused high-level feature map. Global average pooling is then performed on the fused high-level feature map to obtain a classification feature vector. Finally, the classification feature vector is input into a fully connected classification layer to obtain a classification logical value vector. Softmax normalization is performed on the classification logical value vector to obtain a health level prediction probability vector. The health level with the highest probability is selected as the initial health level recognition result output by the network.

6. The method for assessing the operational health level of gravity dams by combining temperature and stress fields according to claim 1, characterized in that, The specific operations in S5 are as follows: Calculate the health level prediction probability vector for each sample in the training set. Calculate the class balance weight based on the number of samples for each health level in the training set. The fewer the samples for a health level, the greater the corresponding class balance weight. Calculate the order relation weighted loss, which includes the class balance cross-entropy term and the order relation penalty term. The Adam optimizer is used to update all trainable parameters in the thermally decoupled sensitive dual-branch gated fusion network. After each training round, the model performance is evaluated using the validation set to obtain the trained dual-branch gated fusion network, which is then tested using the test set.

Images