Artificial intelligence-based dynamic evaluation method for construction quality of gravity dam concrete

Abstract

The present application relates to a kind of artificial intelligence-based gravity dam concrete construction quality dynamic evaluation method, belong to water conservancy engineering intelligent evaluation technical field.It includes the following steps: acquisition multi-source construction original data and label quality grade, based on the physical relationship of concrete material, eliminate physical inconsistency record, execute interlayer coupling noise suppression to thermal process attribute;Interlayer combination index, temperature difference cumulative damage factor and relative penetration defect index 3 types of derived characteristics are constructed, local linear embedding is executed according to mix proportion zoning, and dimension reduction is carried out by soft zoning mapping;Pouring sequence attention classification network is constructed, sequence formed by target layer and previous 5 layers is input, one-dimensional dilated convolution, bidirectional gated recurrent unit and attention pooling are fused to extract time sequence characteristics, and the model is trained using class weighted sequential constraint loss.This application can realize the fine dynamic evaluation of gravity dam concrete construction quality, reduce the risk of serious misjudgment across grade.

Description

Technical Field

[0001] This invention belongs to the field of intelligent assessment technology for water conservancy projects, specifically relating to a dynamic assessment method for the construction quality of gravity dam concrete based on artificial intelligence. Background Technology

[0002] In the layered pouring construction of large-scale hydraulic engineering projects such as gravity dams, the final quality of concrete depends not only on the materials and processes of the current layer, but also on temporal factors such as the bonding state with previous layers and cumulative damage from temperature differences. However, traditional quality assessments primarily rely on manual sampling and independent indicator evaluation, making it difficult to integrate multi-source monitoring data layer by layer in a real-time and systematic manner and capture inter-layer correlations. This leads to delayed detection of potential quality issues, highly subjective assessment results, and difficulties in tracing problems. Therefore, there is an urgent need in engineering to develop a dynamic quality assessment method that can combine the characteristics of the construction sequence, fully utilize multi-source on-site data, and achieve layer-by-layer refinement and traceability, in order to improve the overall controllability and safety of dam construction quality.

[0003] Existing methods suffer from the following drawbacks: Conventional methods typically assume outliers are numerically large or small, easily leading to the accidental deletion of construction records showing genuine quality anomalies while retaining physically invalid combinations, resulting in low reliability of training data; ordinary moving averages or low-pass filtering treat all changes as noise, simultaneously suppressing random fluctuations and real temperature abrupt changes caused by cold joints, resulting in the loss of key temporal information of interlayer bonding surfaces, affecting the judgment of quality degradation caused by intermittent anomalies; directly using original attributes to train the model cannot fully express nonlinear engineering mechanisms such as "interlayer bonding quality" and "cumulative temperature damage." Furthermore, applying uniform dimensionality reduction to different mix proportions can easily lead to the overlapping of local manifold structures in low and high water-cement ratio regions, causing blurred category boundaries; ordinary cross-entropy loss treats quality levels as disordered categories, failing to reflect that the cost of misjudging "qualified" and "basically qualified" is far less than the engineering risk of misjudging "severely unqualified" as "excellent." At the same time, conventional feedforward networks only consider single records, ignoring the interlayer recursive relationship where lower-layer states affect upper-layer bonding and upper-layer detection reveals lower-layer defects. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] This invention provides a dynamic evaluation method for the construction quality of gravity dam concrete based on artificial intelligence, comprising the following steps:

[0006] S1. Collect the original data of gravity dam concrete construction and mark the quality grade of each construction record;

[0007] S2. Based on the physical relationship of concrete materials, identify and remove construction records with physical inconsistencies, perform interlayer coupling noise suppression on thermal process-related attributes, and obtain a physically consistent construction feature vector.

[0008] S3. Based on the physical consistency construction feature vector, construct three types of derived features: interlayer bonding index, temperature difference cumulative damage factor and relative permeability defect index. Perform partitioned local linear embedding for different mix proportion conditions, and use soft partitioning mapping to obtain dimensionality-reduced construction feature vectors.

[0009] S4. Construct a casting sequence attention classification network. Take the fixed-length casting sequence consisting of the target casting layer and its preceding casting layers as input. Use one-dimensional dilated convolution to extract local combination patterns, use bidirectional gated recurrent units to capture the inter-layer correlation, and use attention pooling to highlight key casting layers. Finally, output the quality level probability.

[0010] S5. The pouring sequence attention classification network is trained using category-weighted order constraint loss to obtain the trained pouring sequence attention classification network.

[0011] S6. After processing the newly collected gravity dam concrete construction records through steps S2 and S3, input them into the trained pouring sequence attention classification network to obtain the gravity dam concrete construction quality assessment results.

[0012] Furthermore, the original data for the gravity dam concrete construction includes pouring temperature, ambient temperature, water-cement ratio, interval, vibration time, slump, air content, fly ash content, unit cementitious material dosage, compressive strength, elastic modulus, ultrasonic wave velocity, curing time, permeability coefficient, and defect depth; the quality grade labels include excellent, qualified, basically qualified, unqualified, and seriously unqualified; each construction record is simultaneously associated with the dam section number, pouring layer number, construction timestamp, and quality grade label.

[0013] Furthermore, a stable empirical correlation exists between the compressive strength, elastic modulus, ultrasonic velocity, permeability coefficient, and defect depth of gravity dam concrete. When the elastic modulus increases, the ultrasonic velocity usually increases synchronously; when the defect depth increases, the permeability coefficient usually increases accordingly. If the compressive strength is high but the elastic modulus and ultrasonic velocity are poor, the construction record often indicates abnormal detection, labeling errors, or sample mismatch. In step S2, this invention checks multiple material-related constraints for each original construction record. When the same construction record violates two or more constraints simultaneously, it is determined to be a physically inconsistent construction record and is discarded.

[0014] From the feature vector of the original construction records The compressive strength, elastic modulus, ultrasonic wave velocity, permeability coefficient, and defect depth are extracted. Based on the empirical relationship between elastic modulus and compressive strength, a strength consistency judgment result for the current construction record is generated. Based on the coupling relationship between permeability coefficient and defect depth, a permeability-defect consistency judgment result for the current construction record is generated. Based on the empirical relationship between elastic modulus and ultrasonic wave velocity, a wave velocity consistency judgment result for the current construction record is generated. The strength consistency judgment result, the permeability-defect consistency judgment result, and the wave velocity consistency judgment result are accumulated to obtain the physical contradiction count value of the current construction record.

[0015] Furthermore, during the layered casting process of a gravity dam, the three attributes of casting temperature, ambient temperature, and intermittent period exhibit significant interlayer transfer characteristics. If a simple moving average is directly applied, the actual abrupt changes before and after cold joint formation will be masked along with random measurement noise, leading to the loss of crucial information about the interlayer interface. In step S2, this invention adaptively filters the thermal process-related attributes according to the principles of the same dam section, similar intermittent periods, adjacent casting layers, and time-ordered arrangement. This suppresses random fluctuations while preserving the actual abrupt changes caused by cold joints.

[0016] For the retained construction records, sets of attributes to be filtered are established for the first dimension (pouring temperature), the second dimension (ambient temperature), and the fourth dimension (interval period). Construction records within each dam section are sorted in ascending order by construction timestamp, and each record is assigned a time rank within a layer. The later the time rank within a layer, the later the construction record appears in the current dam section. For the current construction record, neighboring construction records are searched within the same dam section to construct a set of neighboring construction records. For each attribute to be filtered in the current construction record, a neighborhood weighted mean is calculated: a time weight is generated based on the construction time difference between the neighboring construction records and the current construction record, and a working condition weight is generated based on the interval difference. The time weight and the working condition weight are multiplied to obtain a comprehensive neighborhood weight. Weights are calculated separately for the pouring temperature, ambient temperature, and interval period of the neighboring construction records. The mean is calculated; if no neighboring construction record that meets the conditions is found in the current construction record, the original value of the current construction record is directly retained without smooth replacement; the self-retention coefficient of the current construction record is calculated based on the time sequence within the layer and the interval of the current construction record; the pouring temperature, ambient temperature and interval are weighted and fused item by item according to the method of multiplying the original value of the current construction record by the self-retention coefficient and the neighboring weighted mean by the remaining weight, to obtain the filtered attribute value; in order to correct the slow drift of thermal process attributes in the same dam section during long-term acquisition, local weighted linear regression is performed on the pouring temperature and ambient temperature respectively with the construction timestamp as the independent variable in the same dam section to estimate the slow drift trend; then the drift trend component is subtracted from the current measurement value and the local mean is superimposed to obtain the final corrected value; finally, a physically consistent construction feature vector is obtained. , Indicates the first The feature vector of the construction record after physical consistency correction and inter-layer coupling filtering.

[0017] Furthermore, in step S3, based on the physically consistent construction feature vector, three derived features are constructed: interlayer bonding index, temperature difference cumulative damage factor, and relative permeability defect index.

[0018] For each physically consistent construction feature vector The interval period, vibration time, slump, air content and fly ash content were extracted. The interval period was used to reflect the construction interval between adjacent pouring layers, the vibration time was used to reflect the degree of compaction, the slump was used to reflect the fluidity of the mixture, and the air content and fly ash content were used to reflect the stability and interface adaptability of the mixture.

[0019] Calculate the interlayer bonding index , Indicates the first The interlayer bonding index of each construction record is a dimensionless characteristic; a larger value indicates better bonding quality between adjacent pouring layers. First, the foundation bonding score is obtained by dividing the vibration time by the sum of the interval and 0.1 days, where 0.1 days is a stability constant to prevent the denominator from being zero. The foundation bonding score reflects the basic principle that more thorough vibration and shorter intervals result in better interlayer bonding. Within the current dam section, the median and dispersion intervals of slump, air content, and fly ash content are statistically analyzed. If the slump of the current construction record falls within the middle range of the normal distribution of the current dam section, the compatibility weight is increased; if both the air content and fly ash content of the current construction record deviate from the normal distribution of the dam section, the compatibility weight is decreased. The foundation bonding score is multiplied by the compatibility weight, and the result is normalized within the dam section to obtain the interlayer bonding index. ;

[0020] Calculate the cumulative damage factor due to temperature difference , Indicates the first The cumulative temperature difference damage factor of each construction record is a dimensionless feature used to reflect the cumulative impact of the temperature difference of previous pouring layers on the current construction record. First, all historical pouring layers prior to the current construction record are identified by arranging them according to construction time within the same dam section. For each historical pouring layer, the temperature difference between the pouring temperature and the ambient temperature is calculated, and this temperature difference is used as the contribution of single-layer thermal stress. Based on the time interval between the historical pouring layer and the current construction record, the contribution of single-layer thermal stress is exponentially decayed. The shorter the time interval, the smaller the decay, indicating a greater influence of the previous layer on the current layer; the longer the time interval, the stronger the decay, indicating a gradual weakening of the influence of the previous layer. Finally, the decayed thermal stress contributions of all historical pouring layers are summed and normalized according to the statistical scale of the current dam section to obtain the cumulative temperature difference damage factor. ;

[0021] Calculate the relative penetration defect index , Indicates the first The relative permeability defect index of each construction record is a dimensionless feature used to comprehensively characterize the coupled risk of defect depth and permeability coefficient: The process involves calculating the average defect depth of all construction records within the current pouring layer or adjacent layers of the current dam section; dividing the defect depth of the current construction record by this average value to obtain the relative defect depth; performing logarithmic compression on the permeability coefficient of the current construction record to reduce the impact of maxima on model training; and multiplying the relative defect depth by the logarithmically compressed permeability coefficient to obtain the relative permeability defect index. ;

[0022] The three derived features are concatenated with the 15-dimensional physically consistent construction feature vector to obtain the extended construction feature vector. .

[0023] Furthermore, in step S3, different mix proportion strategies may be adopted for different dam sections or different construction stages. If dimensionality reduction is uniformly applied to all construction records directly, it is easy to mix local structures under different working conditions, resulting in blurred category boundaries. This invention, based on extended construction feature vectors, divides construction records into multiple partitions according to mix proportion-related attributes, and then independently performs local linear embedding within each partition. Dimensionally reduced construction feature vectors are obtained through soft partitioning mapping.

[0024] From extended construction feature vectors Two mix proportion-related attributes, water-cement ratio and fly ash content, were extracted to establish the zoning criteria. The water-cement ratio reflects the consistency and strength development of the slurry, while the fly ash content reflects the impact of admixtures on workability and later performance. Based on the water-cement ratio and fly ash content, all construction records were divided into three mix proportion zones. Within each mix proportion zone, the extended construction feature vector was standardized to ensure that the 18 dimensions had comparable numerical scales. For each construction record within each mix proportion zone, the 12 nearest neighbor construction records were searched using Euclidean distance. Construct a local neighborhood; for each construction record, solve for the local reconstruction weights: represent the current construction record as a weighted linear combination of its neighboring construction records, and require that the sum of all reconstruction weights be 1, with the goal of minimizing the reconstruction error; based on the local reconstruction weights of all construction records, perform local linear embedding within each mix proportion partition to obtain the corresponding 8-dimensional low-dimensional embedding coordinates; calculate the soft partition weights based on the distance from the current construction record to the center of each mix proportion partition, and weight and fuse the mapping results of each mix proportion partition according to the soft partition weights to obtain the dimensionality-reduced construction feature vector. Stack the dimensionality-reduced construction feature vectors of all construction records row by row to obtain the dimensionality-reduced construction feature matrix. .

[0025] Furthermore, in step S4, this invention uses one-dimensional dilated convolution. With a kernel length of only 3 and an dilation rate of 2, the receptive field of each temporal position covers the current layer, the previous two layers, and the previous four layers (a total of three positions). This effectively expands the temporal range without increasing the number of parameters or computational cost, avoiding the computational burden caused by stacking multiple convolutions or using large convolution kernels. Thus, it can expand the temporal receptive field with lower computational cost, improving the expressive power of local construction patterns. One-dimensional dilated convolution is used to extract local combination patterns.

[0026] The evaluation unit is constructed with the target pouring layer as the center: a reduced-dimensional construction feature matrix is ​​obtained. The OK That is, the first Dimensionally reduced construction feature vectors for each construction record; for each dimensionality-reduced construction feature vector... Within the same dam section, the top 5 valid construction records are retrieved in reverse order of construction time, and together with the current construction record, a pouring sequence of length 6 is formed.

[0027] Using a casting sequence of length 6 as network input, one-dimensional dilated convolutional encoding is performed on the casting sequence to obtain local convolutional encoding feature vectors. Local convolution encoding feature vectors Used to characterize the In the evaluation unit, the first The local combination pattern of each temporal position and its preceding layer is used, the convolution direction is unfolded along the construction time axis, the convolution kernel length is 3, the dilation rate is 2, and the number of output channels is 32.

[0028] Furthermore, in step S4, convolutional coding mainly reflects the combination relationship within a local time window. To further express the inter-layer dependency over a longer time span, this invention, based on the features of convolutional coding, uses bidirectional gated recurrent units to model the information of the pouring sequence in the forward and backward directions. Then, attention pooling is used to select the pouring layers that are more critical to the final quality assessment. The specific steps are as follows:

[0029] The local convolutional encoding feature vector is input into a bidirectional gated recurrent unit (BRN) network. The forward-gated BRN updates the hidden state from front to back in time, while the backward-gated BRN updates the hidden state from back to front in time. Both the forward and backward hidden states have a dimension of 64. For each temporal position, the forward and backward hidden states are concatenated along the channel dimension to obtain the bidirectional hidden state vector. For each bidirectional hidden state vector Attention scores are calculated using a fully connected transform layer and a hyperbolic tangent activation function. The attention scores for all valid temporal positions within the same evaluation unit are then normalized to obtain the attention weights. The context feature vector is obtained by weighting and summing all bidirectional hidden state vectors according to the attention weights. ; Context feature vector The data is fed into the output layer, and after fully connected mapping and Softmax normalization, a quality level probability vector is obtained. quality level probability vector The level with the highest probability, as the first The predicted quality level of each evaluation unit. Indicates the first The quality level probability vector of each evaluation unit.

[0030] Further, in step S5, the category-weighted order constraint loss is calculated:

[0031] For the Each evaluation unit has a true quality level denoted as [blank]. The quality level probability vector output by the network is denoted as... , its first Each component is denoted as , Indicates the first The evaluation unit belongs to the first The predicted probability of each quality level;

[0032] For each evaluation unit, calculate the negative logarithm of the probability of the true quality level: take the predicted probability corresponding to the true quality level of the current evaluation unit. Take the negative logarithm of it as the basic classification loss. For each evaluation unit, calculate the order deviation term. The predicted expected level is calculated based on the quality level probability vector. The five level numbers are multiplied by their corresponding predicted probabilities and summed to obtain the predicted center level of the current evaluation unit. The absolute difference between the predicted center level and the actual quality level is used as the order bias term. For each evaluation unit, calculate the level interval penalty item. The process involves iterating through all non-true quality levels, calculating the level difference between the current non-true quality level and the true quality level to obtain the level interval penalty term; calculating the class weights based on the number of evaluation units for each quality level in the training set; summing the basic classification loss, order bias term, and level interval penalty term in a weighted manner, and then multiplying by the loss weight corresponding to the true quality level to which the current evaluation unit belongs to obtain the total loss for the current evaluation unit; and averaging the total loss of all evaluation units to obtain the order constraint loss used in this round of training.

[0033] Furthermore, all evaluation units are divided according to dam sections to form training, validation, and test sets; the AdamW optimizer is used to update network parameters, with an initial learning rate of 0.0005 and a weight decay coefficient of [missing value]. The batch size can be 32; the learning rate is scheduled according to the training rounds. After every 10 training rounds, the current learning rate is multiplied by 0.7 to gradually decrease the learning rate; after each training round, the weighted ordinal accuracy is calculated using the validation set; training stops when the weighted ordinal accuracy on the validation set no longer improves for 15 consecutive training rounds.

[0034] The advantages of this invention are:

[0035] This invention does not rely on numerical distribution, but rather on the inherent physical constraints of concrete materials to detect and eliminate records of "physical inconsistencies." Only when the same record violates two or more empirical relationships is it considered an anomaly, thus retaining true quality anomalies and deleting only logically contradictory samples. For the layered casting characteristics of gravity dams, this invention employs neighborhood-weighted filtering based on "same dam section, similar intervals, and adjacent layer sequences." By calculating the self-holding coefficient, it suppresses random noise while accurately retaining true temperature abrupt changes such as cold joints, avoiding the loss of key interface information due to smoothing. This invention does not rely on black-box learning, but instead constructs three derived features: interlayer bonding index, temperature difference cumulative damage factor, and relative permeability defect index. Simultaneously, it divides the mix proportion into zones based on the water-cement ratio and fly ash content, independently performing local linear embedding within each zone, and solving the problem of category boundary ambiguity caused by cross-condition cascading through soft partition weight fusion. This invention constructs a fixed-length casting sequence composed of the target layer and preceding layers, using dilated convolution and bidirectional gated recurrent units to capture interlayer temporal dependencies, and automatically focusing on key anomaly layers through attention pooling. During training, a ranking-order constraint loss is used to impose a greater penalty on misjudgments across rankings, and the larger the ranking difference, the larger the probability interval is required. Attached Figure Description

[0036] 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.

[0037] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0038] Figure 2 This is a graph showing the variation of pouring temperature with the construction layer number.

[0039] Figure 3 This is a graph showing the change in ambient temperature as a function of the construction layer number.

[0040] Figure 4This is a diagram showing the results of dividing construction records into three water-cement ratio zones: low, medium, and high, using k-means clustering. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1

[0043] In this embodiment, as Figure 1 As shown, this invention provides a dynamic evaluation method for the construction quality of gravity dam concrete based on artificial intelligence, the specific steps of which include:

[0044] S1. Data Collection and Quality Grade Labeling for Gravity Dam Concrete Construction

[0045] During the layered concrete pouring construction of the gravity dam, 15-dimensional original attribute data corresponding to each construction record are collected simultaneously through various means, including temperature sensors embedded in the dam body, environmental meteorological stations, production control systems of the mixing plant, on-site vibration operation recorders, slump cone detectors, air content measuring instruments, material proportioning and feeding records, compressive strength tests of standard curing test blocks, elastic modulus testers, ultrasonic detectors, on-site permeability coefficient tests, and borehole or ground-penetrating radar defect detection.

[0046] Specifically, the pouring temperature is read immediately after pouring by a temperature sensor embedded inside the concrete; the ambient temperature is recorded hourly by a meteorological station near the dam site and aligned with the construction timestamp; material ratio information such as water-cement ratio, fly ash content, and unit cementitious material usage is exported from the mixing plant's production control system for each batch or layer of construction records; the interval is calculated by the difference in timestamps between two adjacent layers; the vibration time is automatically recorded by the on-site vibration operation recorder for each point; the slump and air content are tested on-site at the outlet and the surface of the container and manually entered; the compressive strength and modulus of elasticity are obtained by pressure tests and modulus of elasticity tests on concrete test blocks poured at the same time and cured according to standards at specified ages; the ultrasonic wave velocity is tested along the dam section using a non-metallic ultrasonic detector after the concrete reaches a certain strength; the permeability coefficient is obtained through on-site borehole water injection tests or indoor permeability tests; the defect depth is detected by ultrasonic imaging or borehole endoscopy to determine the extension depth of internal non-compact areas, cracks, and other defects.

[0047] After collecting the aforementioned 15-dimensional attributes, each construction record is simultaneously associated with four key identifiers: dam section number, used to distinguish the location of different dam sections of the gravity dam; pouring layer sequence number, used to record the construction sequence of that layer within its respective dam section; construction timestamp, accurate to the hour or minute, used for subsequent time-series analysis; and quality grade label, which is manually assigned by the on-site supervising engineer, quality inspection unit, and third-party laboratory based on the comprehensive evaluation standards in the record, combined with multiple indicators such as compressive strength, elastic modulus, ultrasonic wave velocity, permeability coefficient, defect depth, and interlayer bonding quality, after comprehensive judgment.

[0048] The quality grade labels are divided into five categories: Label 1 represents "Excellent", indicating that all indicators are better than the specifications and there are no defects; Label 2 represents "Pass", indicating that all indicators meet the design and specification requirements, and slight deviations are allowed but do not affect structural safety; Label 3 represents "Basic Pass", indicating that some indicators are slightly below the standard, but after design review, they can meet the functional requirements and need to be monitored more closely; Label 4 represents "Fail", indicating that there are obvious quality defects and repairs or reinforcement are required; Label 5 represents "Seriously Fail", indicating that there are serious structural defects and rework or scrapping is required.

[0049] All original construction records are stored in the database in chronological order of construction time, forming a feature vector of the original construction record and its corresponding dam section number, pouring layer number, construction timestamp, and quality grade label.

[0050] The feature vector of the original construction record is denoted as , ,in, Indicates the first The original construction record feature vector corresponding to each construction record has a dimension of 15; This indicates the total number of original construction records;

[0051] Of the 15 attributes, the first dimension is the pouring temperature (unit: The second dimension is ambient temperature (unit: The third dimension is the water-cement ratio (dimensionless), and the fourth dimension is the interval (unit: ...). The fifth dimension is the vibration time (unit: ...). The 6th dimension is the collapse degree (unit: The 7th dimension is the gas content (unit: gas content). The 8th dimension is the fly ash content (unit: The 9th dimension represents the amount of cementitious material used per unit (unit: ...). The 10th dimension is compressive strength (unit: The 11th dimension is the elastic modulus (unit: ). The 12th dimension is the ultrasonic wave velocity (unit: The 13th dimension is the maintenance duration (unit: The 14th dimension is the permeability coefficient (unit: ). The 15th dimension is the defect depth (unit: Each attribute unit retains the definition from the original detection record.

[0052] Each construction record is simultaneously associated with the dam section number, pouring layer sequence number, construction timestamp, and quality grade label. The dam section number identifies the dam section to which the current construction record belongs; the pouring layer sequence number identifies the pouring layer order corresponding to the current construction record; the construction timestamp identifies the actual construction time of the current construction record; and the quality grade label is denoted as follows: , The values ​​are 1, 2, 3, 4, and 5, which correspond to excellent, qualified, basically qualified, unqualified, and seriously unqualified, respectively.

[0053] S2. Physical consistency correction and interlayer coupling noise removal of gravity dam concrete construction data.

[0054] Gravity dam concrete is poured in layers, and there are obvious physical connections and temporal transmission relationships between different pouring layers. If contradictory combinations such as "high elastic modulus but abnormally low ultrasonic wave velocity", "high permeability coefficient but extremely small defect depth", and "high compressive strength but significantly insufficient material stiffness" appear in the original construction records, then such construction records are usually not outliers in the ordinary statistical sense, but rather physical inconsistencies in construction records caused by abnormal detection, mismatched records, or input errors.

[0055] Conventional cleaning based solely on numerical distribution can easily lead to the accidental deletion of genuine abnormal construction records or the retention of construction records that are not physically valid, thus affecting the training of subsequent classification models.

[0056] This invention first identifies and eliminates construction records with physical inconsistencies based on the physical relationships of concrete materials, and then performs interlayer coupling noise suppression on thermal process-related attributes to obtain a feature vector of physically consistent construction. The specific steps are as follows:

[0057] S201, Detection of Physical Inconsistency Construction Records Based on Material Correlation Constraints

[0058] There is a stable empirical correlation between the compressive strength, elastic modulus, ultrasonic velocity, permeability coefficient, and defect depth of gravity dam concrete. When the elastic modulus increases, the ultrasonic velocity usually increases synchronously; when the defect depth increases, the permeability coefficient usually increases accordingly. If the compressive strength is high but the elastic modulus and ultrasonic velocity are poor, the construction record often indicates abnormal testing, labeling errors, or sample mismatch.

[0059] This invention checks multiple material-related constraints for each original construction record. When the same construction record violates two or more constraints simultaneously, it is determined to be a physically inconsistent construction record and is removed. The specific steps are as follows:

[0060] 1) From the feature vector of the original construction record The compressive strength, elastic modulus, ultrasonic wave velocity, permeability coefficient, and defect depth were extracted from it.

[0061] Compressive strength is used to characterize the overall load-bearing capacity of concrete, elastic modulus is used to characterize the stiffness of the material, ultrasonic wave velocity is used to characterize the internal density, permeability coefficient is used to characterize the impermeability, and defect depth is used to characterize the extent to which internal defects extend into deeper layers.

[0062] 2) Based on the empirical relationship between elastic modulus and compressive strength, generate the strength consistency judgment result of the current construction record.

[0063] In practice, 30 GPa is used as the reference elastic modulus and 30 MPa as the reference compressive strength. The elastic modulus recorded in the current construction is converted into the corresponding reference compressive strength. The conversion principle is that "the higher the elastic modulus, the higher the lower limit of the allowable compressive strength". Then, the ratio of the measured compressive strength to the reference compressive strength is compared. When the ratio of the measured compressive strength to the reference compressive strength is less than 0.6, a strength conflict is recorded.

[0064] In one embodiment, for example, if the elastic modulus of the current construction record is 36 gigapascals, then the corresponding reference compressive strength can be obtained as approximately 36 megapascals based on a linear proportional relationship.

[0065] 3) Based on the coupling relationship between the permeability coefficient and the defect depth, generate the permeability-defect consistency judgment result of the current construction record.

[0066] In practice, the permeability coefficient and the defect depth are checked together. If the permeability coefficient is large and the defect depth is also large, it indicates that internal channels and deep damage exist at the same time, which is mutually corroborating. If the combination of the two significantly exceeds the upper limit of engineering experience, it is mostly a data anomaly.

[0067] As an easy-to-implement approach, the product of the penetration coefficient and the defect depth can be used as the joint risk factor. When the joint risk factor is greater than... One penetration-defect contradiction is recorded at centimeters per second per meter.

[0068] In one embodiment, for example, if the permeability coefficient of the current construction record is If the speed is centimeters per second and the defect depth is 0.7 meters, then the combined risk is... centimeters per second·meter, exceeding Centimeters per second per meter, indicating a contradiction between penetration and defects in the current construction record.

[0069] 4) Based on the empirical relationship between elastic modulus and ultrasonic wave velocity, generate the wave velocity consistency judgment result of the current construction record.

[0070] In practice, the reference ultrasonic wave velocity is first estimated based on the elastic modulus, and then the measured ultrasonic wave velocity is compared with the reference ultrasonic wave velocity. In order to avoid the instability caused by directly subtracting high-order values, the natural logarithm of the two can be taken first, and then the absolute value of the difference can be compared. When the logarithmic deviation is greater than 0.15, the wave velocity contradiction is recorded as 1.

[0071] In one implementation, the method of estimating the reference ultrasonic wave velocity based on the elastic modulus is expressed as follows: ,in, This represents the reference ultrasonic wave velocity estimated based on the elastic modulus, in meters per second. The value represents the elastic modulus of the current construction record, in gigapascals; 30 represents the reference elastic modulus, in gigapascals; 5000 represents the reference ultrasonic wave velocity, in meters per second. This formula indicates that the larger the elastic modulus, the higher the reference ultrasonic wave velocity.

[0072] In one embodiment, for example, if the elastic modulus of the current construction record is 36 gigapascals, then the reference ultrasonic wave velocity is approximately 5477 meters per second; if the measured ultrasonic wave velocity is only 3950 meters per second, then the logarithmic deviation between the two exceeds 0.15, and it is determined that there is a wave velocity contradiction in the current construction record.

[0073] 5) The results of the strength consistency assessment, the permeability-defect consistency assessment, and the wave velocity consistency assessment are summed to obtain the physical contradiction count value of the current construction record. Specifically,

[0074] When the physical contradiction count is greater than or equal to 2, the current construction record is judged as a physically inconsistent construction record and removed from the sample set; when the physical contradiction count is less than 2, the current construction record is retained for the next step of processing.

[0075] It should be noted that abnormalities in dam construction quality may themselves be genuine anomalies. Relying solely on statistical outlier rules can easily lead to accidental deletion. This invention does not delete construction records based on whether a single attribute is too large or too small. Instead, it judges the credibility of construction records based on whether multiple attributes simultaneously satisfy material physical relationships. By using multiple physical relationships for joint judgment, it can more accurately distinguish between "genuine anomalies" and "logical contradictions," thereby improving the credibility of training data and reducing misleading information caused by sample mismatch, instrument malfunction, and data entry errors.

[0076] S202, Thermal process property coupling filtering based on interlayer similar working conditions

[0077] During the layered pouring of gravity dams, the three attributes of pouring temperature, ambient temperature and interval period have obvious interlayer transfer characteristics. If ordinary moving average is used directly, the real abrupt change before and after the formation of cold joints will be smoothed out along with random measurement noise, resulting in the loss of key information of the interlayer interface.

[0078] This invention, following the principles of "same dam section, similar intermittent periods, adjacent pouring layers, and sorted by time," adaptively filters the thermal process-related attributes to suppress random fluctuations while preserving the true abrupt changes caused by cold joints. The specific steps are as follows:

[0079] 1) For the construction records retained in step S201, establish sets of attributes to be filtered for the first dimension of pouring temperature, the second dimension of ambient temperature, and the fourth dimension of intermittent period;

[0080] The construction records within each dam section are arranged in ascending order by construction timestamp, and each construction record is assigned a time priority within the layer; the later the time priority within the layer, the later the construction record appears in the current dam section.

[0081] 2) For the current construction record, search for neighboring construction records within the same dam section and construct a set of neighboring construction records. A neighboring construction record must simultaneously meet three conditions:

[0082] First, it belongs to the same dam section as the current construction record;

[0083] Second, the difference between the interval period and the current construction record is less than 0.5 days;

[0084] Third, the difference in the sequence number of the pouring layers shall not exceed 2 layers.

[0085] It should be noted that the purpose of constructing the neighborhood construction record set in this way is to smooth out construction records with similar construction conditions and inter-layer relationships as much as possible, and to avoid distortion caused by mixing across dam sections, working conditions and distant layers.

[0086] 3) For each attribute to be filtered in the current construction record, calculate the neighborhood weighted mean.

[0087] In the specific implementation, a time weight is first generated based on the construction time difference between the neighboring construction records and the current construction record, and then a working condition weight is generated based on the interval difference. The time weight and the working condition weight are then multiplied to obtain the comprehensive neighbor weight. Finally, the weighted average of the pouring temperature, ambient temperature and interval of the neighboring construction records are calculated respectively. If no neighboring construction record that meets the conditions is found for the current construction record, the original value of the current construction record is directly retained without smooth replacement.

[0088] In one embodiment, as an example, if the current construction record is located at the beginning of the dam section, and there are no other construction records in the same dam section with an interval difference of less than 0.5 days and a pouring layer number difference of no more than 2 layers, then no neighboring construction record that meets the conditions can be found. For example, if the current construction record is the first layer in the dam section with an interval of 0.3 days, but there are no other construction records in the dam section (or the intervals of subsequent records are all greater than 0.8 days), then the set of neighboring construction records is empty. In this case, no smooth replacement is performed, and the pouring temperature, ambient temperature, and interval of the current construction record retain their original values.

[0089] 4) Calculate the self-retention coefficient of the current construction record based on the time sequence within the layer and the interval between current construction records;

[0090] The larger the self-retention coefficient, the more the original value of the current construction record is preserved;

[0091] The smaller the self-maintaining coefficient, the more dependent it is on the neighborhood weighted mean.

[0092] In an easy-to-implement approach, the self-sustaining coefficient can be limited to between 0.5 and 0.95; when the current construction record is in the later stage of the dam section and the interval is relatively large, the self-sustaining coefficient is close to 0.95; when the current construction record is in the early stage of the dam section and the interval is normal, the self-sustaining coefficient is close to 0.5.

[0093] In practical implementation, the time sequence within each layer can be normalized to the 0-1 interval first, and then the intervals can be normalized to the 0-1 interval according to their distribution within the current dam section. Then, the self-retention coefficient can be calculated according to the principle of "preserving more of the original values ​​in later construction records and preserving more of the original values ​​in construction records with larger intervals." Specifically, the self-retention coefficient... The calculation consists of three steps:

[0094] a) Intra-layer time sequence normalization: Assuming the current dam section has a total of There are [number] construction records. The current construction record's time sequence within the floor is [number]. ( As the earliest, If it is the latest), then the normalized time order is... value range ;

[0095] b) Intermission Period Normalization: Within the current dam section, the intermission period of all construction records is statistically analyzed to obtain the minimum value. and maximum value The current construction record shows an interval of [period]. Then the normalization interval (like but );

[0096] c) Calculation of self-holding coefficient: The product form is used, and is limited to... Interval: when and When both are close to 1, When both are close to 0.95, It is close to 0.5.

[0097] In one embodiment, for example, a certain dam section has 10 construction records, and the current record is the 8th one. ),but The minimum interval between dam sections is 0.2 days, and the maximum is 1.5 days. The current interval is 1.2 days. ;product Self-holding coefficient .

[0098] 5) The pouring temperature, ambient temperature, and interval are weighted and fused one by one according to the method of "multiplying the original value of the current construction record by the self-holding coefficient and multiplying the neighborhood weighted mean by the remaining weight" to obtain the filtered attribute value;

[0099] After filtering, only the first, second, and fourth dimensions of the attribute are replaced, while the other dimensions retain their original values ​​as processed in step S201.

[0100] 6) To correct the slow drift of thermal process attributes within the same dam section during long-term data acquisition, locally weighted linear regressions are performed on the pouring temperature and ambient temperature, using the construction timestamp as the independent variable, to estimate the gradual drift trend. The drift trend component is then subtracted from the current measured value, and the local mean is superimposed to obtain the final corrected value. The locally weighted linear regression can use a cubic weighting function, and the regression window width can be 10% to 20% of the number of construction records for the current dam section.

[0101] It should be noted that if the total number of construction records for the current dam section is small, the window width can be fixed to no less than 5 construction records to avoid local regression instability.

[0102] In one embodiment, as an example, suppose the current dam section has a total of The window width is determined by the number of construction records. The regression window width threshold Furthermore, the lower limit of 5 conditions ensures regression stability, among which... This indicates the operation of finding the maximum value, for example:

[0103] like ,Pick ,but strip, Indicates the rounding operation; if ,Pick ,but However, since it is less than 5, we take... Article; if ,Pick ,but strip.

[0104] After step S2, the physically consistent construction feature vector is obtained. , Indicates the first The feature vector of the construction record after physical consistency correction and inter-layer coupling filtering has a dimension of 15; the total number of remaining construction records is denoted as... , This represents the total number of construction records after removing those that are physically inconsistent.

[0105] In one embodiment, a comparison is performed before and after coupling filtering of thermal process properties based on inter-layer similar working conditions (for a certain dam section), such as... Figure 2 As shown, the pouring temperature (°C) varies with the construction layer number, such as... Figure 3 As shown, the changes in ambient temperature (°C) are illustrated, with each graph including the original value curve (circles) and the filtered curve (squares). The horizontal axis represents the construction layer number (reflecting chronological order), and the vertical axis represents temperature (°C). The data demonstrate that the proposed inter-layer weighted fusion filtering method effectively suppresses random measurement noise while preserving the true temperature abrupt changes caused by cold joints.

[0106] S3. Derived feature construction based on layered construction mechanism and adaptive dimensionality reduction of mix proportion zoning.

[0107] The final quality of gravity dam concrete does not depend on a single attribute at a certain moment, but is simultaneously affected by factors such as interlayer bonding state, cumulative damage from temperature difference, and permeability-defect coupling. If the classification model is trained directly using 15 original attributes, it is difficult to fully express the nonlinear relationship between construction technology and final quality.

[0108] This invention first constructs three types of derived features: interlayer bonding index, temperature difference cumulative damage factor, and relative permeability defect index. Then, it performs partitioned local linear embedding for different mix proportions and uses soft partitioning mapping to obtain a unified 8-dimensional reduced construction feature vector. The specific steps are as follows:

[0109] S301, interlayer bonding index, temperature difference cumulative damage factor and relative permeability defect index construction

[0110] Based on the physically consistent construction feature vector, this invention constructs three derived features. The first derived feature describes the interlayer bonding quality, the second derived feature describes the cumulative thermal damage caused by temperature differences, and the third derived feature describes the combined risk of defect depth and permeability coefficient. The specific steps are as follows:

[0111] 1) For each physically consistent construction feature vector Extraction interval, vibration time, slump, air content, and fly ash content, among which,

[0112] Interval period is used to reflect the construction interval between adjacent pouring layers, vibration time is used to reflect the degree of compaction, slump is used to reflect the fluidity of the mixture, and air content and fly ash content are used to reflect the stability and interface adaptability of the mixture.

[0113] 2) Calculate the interlayer bonding index , Indicates the first The interlayer bonding index of a construction record is a dimensionless characteristic; the larger the value, the better the bonding quality between adjacent pouring layers.

[0114] In practical implementation, the foundation bonding score is first obtained by dividing the vibration time by "interval period plus 0.1 days," where 0.1 days is a stability constant to prevent the denominator from being zero. The foundation bonding score reflects the basic principle that "the more thorough the vibration and the shorter the interval, the better the interlayer bonding." Then, within the current dam section, the median and dispersion range of slump, air content, and fly ash content are statistically analyzed. If the slump recorded in the current construction record falls within the middle range of the normal distribution of the current dam section, the compatibility weight is increased; if both the air content and fly ash content recorded in the current construction record deviate from the normal distribution of the dam section, the compatibility weight is decreased. Finally, the foundation bonding score is multiplied by the compatibility weight, and the result is normalized within the dam section to obtain the interlayer bonding index. .

[0115] In one embodiment, for example, if the vibration time of the current construction record is 45 seconds and the interval is 0.4 days, then the foundation bonding score is 45 divided by 0.5, which is approximately 90. If the slump of the current construction record is in the normal range, but the air content is slightly high and the fly ash content is close to the median of the dam section, then the compatibility weight can be taken as 0.85. After product and normalization within the dam section, the interlayer bonding index of the current construction record is obtained.

[0116] It should be noted that, for the sake of simplicity in engineering implementation, all time quantities here are processed as numerical values, and no unit conversion is performed in actual calculations; they are only used for relative comparison of feature construction.

[0117] 3) Calculate the cumulative damage factor due to temperature difference , Indicates the first The cumulative damage factor of temperature difference in a construction record is a dimensionless feature used to reflect the cumulative impact of the temperature difference of the previous pouring layer on the current construction record.

[0118] In practice, the process begins by sorting all historical pouring layers within the same dam section according to their construction time, identifying all layers prior to the current construction record. For each historical pouring layer, the difference between the pouring temperature and the ambient temperature is calculated, and this temperature difference is used as the contribution of thermal stress to the single layer. Then, based on the time interval between the historical pouring layer and the current construction record, the contribution of thermal stress to the single layer is exponentially decayed. The shorter the time interval, the smaller the decay, indicating a greater influence of the preceding layer on the current layer; the longer the time interval, the stronger the decay, indicating a gradual weakening of the influence of the preceding layer. Finally, the decayed thermal stress contributions of all historical pouring layers are summed and normalized according to the statistical scale of the current dam section to obtain the cumulative temperature difference damage factor. .

[0119] In one implementation, the time decay time constant can be set to 7 days, which can take into account the engineering principle that the heat of hydration is significant in the early stage and the influence gradually weakens in the later stage.

[0120] 4) Calculate the relative penetration defect index , Indicates the first The relative permeability defect index of each construction record is a dimensionless feature used to comprehensively characterize the coupled risk of defect depth and permeability coefficient.

[0121] In practice, the mean defect depth of all construction records within the current pouring layer or adjacent layers of the current dam section is first calculated. Then, the defect depth of the current construction record is divided by this mean to obtain the relative defect depth. Next, logarithmic compression is performed on the permeability coefficient of the current construction record to reduce the impact of maxima on model training. Finally, the relative defect depth is multiplied by the logarithmically compressed permeability coefficient to obtain the relative permeability defect index. .

[0122] It should be noted that if the number of construction records for the current pouring layer is too small, resulting in an unstable average value, the average defect depth of the construction records of the three adjacent layers within the current dam section can be used instead.

[0123] 5) Combine the three derived features ( , and The extended construction feature vector is obtained by concatenating the 15-dimensional physical consistent construction feature vector with the 15-dimensional physical consistent construction feature vector. , Indicates the first The extended construction feature vector of each construction record has 18 dimensions. The first 15 dimensions are the physical consistency construction features, and the last 3 dimensions are the interlayer bonding index, the temperature difference cumulative damage factor, and the relative permeability defect index, respectively.

[0124] It should be noted that the construction quality of the dam body is affected by "interlayer bonding", "temperature difference accumulation" and "permeability-defect coupling". These relationships are difficult to be stably learned by the classification model directly from the original attributes. This invention can explicitly write the key engineering mechanism into the feature space by pre-constructing derived features, thereby improving the separability of subsequent dimensionality reduction and classification, and reducing the model's dependence on a large sample size.

[0125] S302, Local linear embedding dimensionality reduction with adaptive zoning of mix proportions

[0126] Different mix proportion strategies may be adopted for different dam sections or different construction stages. If the dimensionality is reduced uniformly on all construction records, it is easy to mix up the local structures under different working conditions, resulting in blurred category boundaries.

[0127] This invention first divides the construction records into multiple partitions based on mix proportion-related attributes, then independently performs local linear embedding within each partition, and finally obtains a unified 8-dimensional reduced construction feature vector through soft partitioning mapping. The specific steps are as follows:

[0128] 1) From the extended construction feature vector Two mix proportion-related attributes, water-cement ratio and fly ash content (the third dimension and the eighth dimension), were extracted to establish zoning criteria.

[0129] The water-cement ratio is used to reflect the consistency and strength development of the slurry, while the fly ash content is used to reflect the impact of the admixture on workability and later performance.

[0130] 2) Based on the water-cement ratio and fly ash content, all construction records were divided into 3 mix proportion zones.

[0131] In one implementation, the water-cement ratio and fly ash content can be normalized to the range of 0 to 1 first, and then... The mean clustering method divides the mix into three partitions, which correspond to the low water-cement ratio, medium water-cement ratio, and high water-cement ratio regions, respectively. It should be noted that the partitioning is used to ensure that local neighborhoods are only established between construction records with similar working conditions, which is more in line with the comparison logic in the actual construction process.

[0132] 3) Within each mix proportion zone, the extended construction feature vector is standardized within the zone to make the 18 dimensions have comparable numerical scales;

[0133] Standardization can be performed by setting the mean to 0 and the standard deviation to 1 within each partition. If the variance of a certain dimension is too small within the current partition, a very small constant is added to the denominator of the standard deviation to avoid instability in normalization.

[0134] 4) For each construction record within each mix ratio zone, search for the 12 nearest neighbor construction records using Euclidean distance to construct a local neighborhood. Here, 12 represents the number of neighborhood construction records in the local linear embedding, which is used to describe the linear reconstruction relationship of the construction records on the local manifold.

[0135] In practical implementation, if the number of construction records within the current mix design zone is small, the number of neighboring records is set to a maximum reasonable value that "does not exceed the number of construction records in the current mix design zone minus 1". That is, if the total number of construction records within the current mix design zone is small... If the number is small, the number of neighborhoods should be taken as... The minimum value in the range is used to ensure that each construction record has at least one distinct neighboring record. For example, if there are only 8 construction records in a certain mix design zone, the number of neighbors should be taken as... The minimum value in the search is 7, meaning that 7 nearest neighbor construction records are searched. If the partition cannot be constructed, it needs to be merged into a neighboring partition or global dimensionality reduction should be used.

[0136] 5) For each construction record, calculate the local reconstruction weights. Specifically,

[0137] The current construction record is represented as a weighted linear combination of its neighboring construction records, and the sum of all reconstruction weights is required to be 1; the goal of solving the problem is to minimize the reconstruction error.

[0138] If the neighborhood covariance matrix is ​​close to singular, then add [something] to the diagonal. Regularization terms of magnitude are added to ensure the stability of the solution.

[0139] 6) Based on the local reconstruction weights of all construction records, perform local linear embedding within each mix proportion zone to obtain the corresponding 8-dimensional low-dimensional embedding coordinates.

[0140] In practical implementation, since standard local linear embedding typically outputs low-dimensional coordinates directly instead of an explicit mapping matrix, to facilitate subsequent unified mapping of new construction records, this invention, after obtaining the 8-dimensional low-dimensional embedding coordinates of each matching score zone, then solves for the linear mapping matrix from "18-dimensional input to 8-dimensional output" using least-squares fitting. Specifically,

[0141] a) Constructing the data matrix: For the first There are several coordination ratio partitions, assuming that each partition contains... Each construction record is used to generate an 18-dimensional extended construction feature vector for each record. As row vectors, stacked to form the input matrix (size The corresponding 8-dimensional local linear embedding low-dimensional embedding coordinates are used as row vectors and stacked to form the output matrix. (size );

[0142] b) Least squares solution: Finding the linear mapping matrix , making That is, minimize The standard solution is ;like If it approaches singularity, add a regularization term: Among them, the regularization parameter Pick , for identity matrix for The transpose of .

[0143] 7) Calculate the soft partition weights based on the distances from the current construction records to the centers of each mix proportion partition. Then, weight and fuse the mapping results of each mix proportion partition according to the soft partition weights to obtain the final dimensionality-reduced construction feature vector. , Indicates the first The reduced-dimensional construction feature vector of each construction record, with a dimension of 8, is used for subsequent quality level classification.

[0144] In practical implementation, the current construction record is up to the [number]th [number]. The distance between the centers of each mix ratio zone is denoted as . , Characterize the current construction record and the first The degree of proximity of the working condition centers of each mix ratio zone; the soft zone weight is denoted as... , This indicates the current construction record for the [number]th [item]. The degree of adoption of each combination ratio partition mapping result is calculated as follows: ,in, This represents a minimal constant to prevent the denominator from being zero; for example, it can be taken as... .

[0145] Furthermore, the final dimensionality-reduced construction feature vector can be obtained by "weighted sum of the projection results of each mix proportion zone". That is, the current extended construction feature vector is projected using three mix proportion zone mapping matrices respectively, and then fused according to the soft zone weight.

[0146] In one embodiment, for example, if the distances from the current construction record to the centers of the three mix design partitions are 0.10, 0.60, and 0.90 respectively, then the corresponding soft partition weights are approximately 0.78, 0.13, and 0.09. This indicates that the current construction record mainly uses the mapping result of the first mix design partition, while retaining a small amount of mapping information from the other mix design partitions to reduce the discontinuity problem caused by hard partition boundaries.

[0147] 8) Stack the dimensionality-reduced construction feature vectors of all construction records row by row to obtain the dimensionality-reduced construction feature matrix. , This represents the feature matrix composed of all the reduced-dimensional construction feature vectors, with the following number of rows: The number of columns is 8; the corresponding quality level label vector is still denoted as , This represents the set of quality grade labels for all construction records.

[0148] It should be noted that the material behavior of gravity dam concrete is significantly affected by the mix proportion. The engineering meaning of the same property change under different working conditions is not exactly the same. This invention does not simply compress all construction records into a low-dimensional space. Instead, it first maintains the local geometric relationship between construction records with similar working conditions, and then achieves a smooth transition across working conditions through soft partitioning mapping, thereby improving the continuity and stability of the dimensionality reduction results and reducing the mapping jump caused by the boundary of the mix proportion partition.

[0149] In one embodiment, such as Figure 4 As shown, with the water-cement ratio (dimensionless) as the x-axis and the fly ash content (unit: %) as the y-axis, the results of dividing construction records into three water-cement ratio zones (low, medium, and high) using k-means clustering are presented. Different colored points represent different zones, and each zone presents a clear cluster boundary on the two-dimensional plane. This technique first performs zone processing based on mix proportion-related attributes, ensuring that subsequent local linear embedding only establishes neighborhood relationships between construction records with similar working conditions, avoiding dimensionality reduction distortion caused by cross-working-condition mixing, and demonstrating the engineering rationality of adaptive dimensionality reduction based on mix proportions.

[0150] S4. Construct a pouring sequence attention classification network oriented towards inter-layer transitive relationships.

[0151] The quality of gravity dam concrete exhibits a clear interlayer progression characteristic. The temperature difference, defect state, and curing state of the lower layer concrete can affect the interlayer bonding quality of the newly poured upper layer concrete. Conversely, the detection information collected during the construction of the upper layer may also reveal potential defects in the lower layer. Conventional feedforward classification networks only look at a single construction record and find it difficult to capture this interlayer sequential dependency.

[0152] This invention takes a fixed-length casting sequence consisting of the target casting layer and its preceding casting layers as input, uses one-dimensional dilated convolution to extract local combination patterns, then uses bidirectional gated recurrent units to capture inter-layer correlations, and highlights key casting layers through attention pooling, finally outputting the quality level probability. The specific steps are as follows:

[0153] S401. Construction of Fixed-Length Casting Sequence Based on Target Casting Layer and Local Convolutional Encoding

[0154] This invention utilizes one-dimensional dilated convolution. With a kernel length of only 3 and an dilation rate of 2, the receptive field of each temporal location covers the current layer, the two preceding layers, and the four preceding layers (a total of three locations). This effectively expands the temporal range without increasing the number of parameters or computational cost, avoiding the computational burden of stacking multiple convolutions or using large convolution kernels. Therefore, it can expand the temporal receptive field with lower computational cost and improve the expressive power of local construction patterns. The specific steps are as follows:

[0155] 1) Construct evaluation units centered on the target pouring layer; specifically,

[0156] Take the reduced-dimensional construction feature matrix The OK That is, the first The reduced-dimensional construction feature vector of each construction record;

[0157] For each dimension-reduced construction feature vector Within the same dam section, the top 5 valid construction records are retrieved in reverse order of construction time (or in forward order of construction time), and together with the current construction record, form a pouring sequence of length 6.

[0158] In practical implementation, when constructing the pouring sequence, the target pouring layer (the current construction record) is used as the center. Within the same dam section, the top 5 valid construction records are retrieved in reverse order of construction time (from the current time backward). If there are fewer than 5 historical construction records, zero-filling is performed on the left side of the pouring sequence, and valid position markers are generated simultaneously to ensure that subsequent loop layers and attention layers only perform valid calculations on the actual pouring layer positions. For example, if the construction timestamp order within a certain dam section is: record A (earliest), B, C, D, E, F (current target), then forward retrieval yields E, D, C, B, A, which, together with F, form the sequence. (Time from morning to night); If there are fewer than 5 historical records, such as target B, then only A is in the first 5 records, and the remaining 4 positions are filled with zeros, resulting in the sequence. and generate valid location markers. The quality grade label of the target pouring layer serves as the label for this sequence. .

[0159] The total number of evaluation units is denoted as , This represents the total number of casting sequences used for classification training or inference, the nth. The pouring sequence corresponding to each evaluation unit is denoted as . ,in, Indicates the first In the evaluation unit, the first The reduced-dimensional construction feature vectors corresponding to each time-series location have a dimension of 8. Indicates the temporal position in the pouring sequence;

[0160] No. The true quality level label of each evaluation unit is denoted as , It is equal to the quality grade label of the target pouring layer corresponding to the last position of the pouring sequence.

[0161] In one embodiment, as an example, the construction records for a certain dam section, arranged chronologically from earliest to latest, are: Record A (Quality Grade 2), B (Grade 3), C (Grade 1), D (Grade 4), E (Grade 2), and F (Current Target Layer, Grade 3). Taking a sequence length of 6, with F as the target layer, retrieving 5 records forward yields A, B, C, D, and E, which, together with F, form the sequence. The label of the sequence Take the level of the target layer F, that is If there are fewer than 5 historical records, for example, if the target layer is B (level 3) and there is only A (level 2) preceding it, then the sequence is: (0 indicates zero padding), tag .

[0162] 2) Using a casting sequence of length 6 as network input, perform one-dimensional dilated convolutional encoding on the casting sequence to obtain local convolutional encoding feature vectors. Local convolution encoding feature vectors Used to characterize the In the evaluation unit, the first The local combination pattern of each time position (cast layer) and its preceding layer (interval of 2 and 4 layers), and the convolution direction is unfolded along the construction time axis, the convolution kernel length is 3, the expansion rate is 2, and the number of output channels is 32.

[0163] Based on this, the current temporal position can see not only the information of the adjacent pouring layer during convolution, but also the information of the interlayer, which is beneficial for capturing the non-local effects caused by the lengthening of the interval period.

[0164] In practical implementation, for the current temporal position, the features of this layer, the features of the previous two layers, and the features of the previous four layers are fed into the same convolutional kernel for linear combination, then a bias term is added, and the result is processed by a modified linear activation function to obtain the local convolutional encoded feature vector. , Indicates the first In the evaluation unit, the first One-dimensional convolutional encoded feature vectors at each temporal location, with a dimension of 32, are used to characterize the local combination pattern between the current pouring layer and the previous pouring layer.

[0165] In the specific implementation, zero-filling input is used for locations at the boundary of the casting sequence where the receptive field is insufficient to construct a complete receptive field. For zero-filling locations, effective location markers are used to mask them in subsequent bidirectional gated recurrent units and attention pooling to avoid the filling value interfering with the real features of the casting layer.

[0166] In one embodiment, for example, if there are only two valid historical pouring layers before a target pouring layer, then the first three positions of the pouring sequence are zero-filled, and the last three positions are real pouring layers; after convolutional encoding, only the positions of the real pouring layers participate in the subsequent attention normalization.

[0167] It should be noted that the quality impact between gravity dam layers does not necessarily occur only between two adjacent layers. The curing status and temperature difference accumulation of earlier poured layers may also affect the current poured layer. This invention uses one-dimensional dilated convolution, with a convolution kernel length of only 3 and an dilation rate of 2, to enable the receptive field of each temporal position to cover the current layer, the two layers before it, and the four layers before it (a total of 3 positions). This effectively expands the temporal range without increasing the number of parameters and the amount of computation, avoiding the computational burden caused by stacking multiple layers of convolution or using large convolution kernels. Thus, it can expand the temporal receptive field with a lower amount of computation and improve the expressive ability of local construction patterns.

[0168] S402, Bidirectional Gated Recurrent Unit Modeling and Attention Pooling Classification

[0169] Convolutional coding primarily reflects the combinational relationships within a local time window. To further express inter-layer dependencies over longer time spans, this invention, based on the features of convolutional coding, employs bidirectional gated recurrent units to model the forward and backward directional information of the casting sequence. Then, attention pooling is used to select the casting layers that are more critical to the final quality assessment. The specific steps are as follows:

[0170] 1) Encode feature sequences using local convolution. Input a bidirectional gated recurrent unit network, where the forward gated recurrent unit updates the hidden state from front to back in time, and the backward gated recurrent unit updates the hidden state from back to front in time. The dimensions of both the forward and backward hidden states are 64.

[0171] 2) For each temporal position, concatenate the forward hidden state and the backward hidden state along the channel dimension to obtain a bidirectional hidden state vector. , Indicates the first In the evaluation unit, the first The bidirectional hidden state vector at each temporal location has a dimension of 128 and contains interlayer transfer information obtained from the convergence of the current pouring layer from the front and rear directions.

[0172] Furthermore, for each bidirectional hidden state vector Attention scores are calculated using a fully connected transform layer and a hyperbolic tangent activation function. Then, the attention scores for all valid temporal positions within the same evaluation unit are normalized to obtain the attention weights. , Indicates the first In the evaluation unit, the first The attention weight of each temporal position is a dimensionless coefficient; the sum of the attention weights of all valid temporal positions within the same evaluation unit is 1; zero-filled positions are assigned a minimum score before normalization and do not participate in the effective allocation.

[0173] 3) Perform a weighted summation of all bidirectional hidden state vectors according to the attention weights to obtain the context feature vector. , Indicates the first The context feature vector of each evaluation unit, with a dimension of 128, is used to represent the temporal characteristics of the overall interlayer quality of the current target pouring layer.

[0174] In one embodiment, for example, if the 5th and 6th time-series positions in an evaluation unit simultaneously exhibit a large intermittent period, high temperature difference accumulation, and high relative permeability defect index, then the attention weight of these two time-series positions will be significantly higher than that of the other time-series positions, thereby making the network pay more attention to the key anomaly layer most relevant to the current target pouring layer.

[0175] 4) Transfer the context feature vector The data is fed into the output layer, and after fully connected mapping and Softmax normalization, a quality level probability vector is obtained. , Indicates the first Each evaluation unit has a quality level probability vector with a dimension of 5. The five components correspond to the predicted probabilities of five levels: excellent, qualified, basically qualified, unqualified, and seriously unqualified.

[0176] Then, the quality level probability vector The level with the highest probability, as the first The predicted quality level of each evaluation unit; if used for real-time construction early warning, it can also output the sum of the probabilities of "unqualified" and "seriously unqualified" as a risk warning indicator.

[0177] It should be noted that different pouring layers have different importance to the final quality judgment. Simply averaging all time-series features will weaken the role of key anomaly layers. Through attention pooling, key pouring layers such as "intermittent anomaly layer", "temperature difference anomaly layer", and "significant defect layer" can be automatically assigned higher weights, thereby improving the sensitivity to a small number of key anomaly layers and improving the accuracy and interpretability of the final quality grade classification.

[0178] S5. Training of a classification model based on hierarchical order constraints

[0179] The quality grades of gravity dam concrete naturally have an order, and the risk of misclassifying "seriously substandard" as "excellent" is significantly higher than the risk of misclassifying "acceptable" as "basically acceptable." While using only ordinary cross-entropy loss can learn to distinguish categories, it cannot reflect the order of quality grades and the differences in error costs.

[0180] This invention employs class-weighted order constraint loss for training, and combines adaptive learning rate scheduling and early stopping strategies to improve model training stability and engineering applicability. The specific steps are as follows:

[0181] S501, Calculation of Category-Weighted Order Constraint Loss

[0182] 1) For the first Each evaluation unit has a true quality level denoted as [blank]. The quality level probability vector output by the network is denoted as... , its first Each component is denoted as , Indicates the first The evaluation unit belongs to the first The predicted probability of each quality level.

[0183] For each evaluation unit, first calculate the negative logarithm of the probability of the true quality level. Specifically, take the predicted probability corresponding to the true quality level of the current evaluation unit. Take the negative logarithm of it as the basic classification loss. .

[0184] 2) For each evaluation unit, further calculate the order deviation term. , specifically

[0185] The predicted expected level is calculated based on the quality level probability vector. This involves multiplying each of the five level numbers by its corresponding predicted probability and summing the results to obtain the predicted center level of the current evaluation unit. ,Right now Then, the absolute difference between the predicted center level and the actual quality level is used as the order deviation term. ,Right now The greater the order deviation, the further the overall level position predicted by the model deviates from the true level, and the greater the penalty needs to be imposed.

[0186] 3) For each evaluation unit, further calculate the level interval penalty item. , specifically

[0187] Iterate through all non-real quality levels and calculate the level difference between the current non-real quality level and the real quality level. When the grade difference is 1, the probability of the true quality grade must be at least 0.05 higher than the probability of the false quality grade; when the grade difference is 2, it must be at least 0.10 higher; when the grade difference is 3, it must be at least 0.15 higher; when the grade difference is 4, it must be at least 0.20 higher; if the corresponding minimum probability interval is not reached, the difference will be recorded as a grade interval penalty.

[0188] In practical implementation, the grade interval penalty term is used to ensure a sufficient gap between the probability of the true quality grade and the probability of each non-true grade, and the larger the grade difference, the larger the required gap. Specifically,

[0189] a) Traverse all non-real quality levels: Let the real quality level be... (Value) For each other level Calculate grade difference ;

[0190] b) Determine the minimum interval requirement: If ,Require ;like ,Require ;like ,Require ;like ,Require ;

[0191] c) Calculate the penalty value: If the actual difference is... Less than the required interval If the penalty is zero, then a penalty is applied; otherwise, the penalty is zero, and the penalty value is [value missing]. ;

[0192] Summing all non-true ratings yields the rating interval penalty for that rating unit. .

[0193] Based on this, the model not only needs to increase the probability of the true quality level, but also needs to create a sufficient gap between it and the incorrect quality level. The further the error is, the larger the required probability gap should be.

[0194] 4) Calculate the category weights based on the number of evaluation units for each quality level in the training set. Specifically, the first... The number of evaluation units for each quality level is denoted as: The maximum number of evaluation units across all quality levels is denoted as Then the first The loss weight for each quality level is denoted as: The calculation method is expressed as , Indicates the first The loss amplification factor for each quality level during training is such that the smaller the sample size of each quality level, the greater its loss weight.

[0195] 5) Add the classification basic loss, order bias term and level interval penalty term in a weighted manner, and then multiply by the loss weight corresponding to the true quality level to which the current evaluation unit belongs to obtain the total loss of the current evaluation unit; average the total loss of all evaluation units to obtain the order constraint loss used in this round of training (i.e. the loss value of this round of training).

[0196] In one implementation, the weight coefficient of the order bias term can be 0.3, and the weight coefficient of the level interval penalty term can be 0.2; based on this, the basic classification ability can be maintained, while the serious misclassification across multiple levels can be reduced.

[0197] S502, Adaptive Learning Rate Scheduling, Validation Selection and Early Stopping

[0198] 1) Divide all evaluation units into dam sections to form training sets, validation sets, and test sets;

[0199] Evaluation units within the same dam section are only included in one data subset to avoid information leakage caused by the same dam section appearing simultaneously during the training and verification phases.

[0200] 2) Update network parameters using the AdamW optimizer. The initial learning rate can be set to 0.0005, and the weight decay coefficient can be set to... The batch size can be 32; AdamW takes into account both adaptive gradient update and weight regularization, making it suitable for hybrid network structures containing convolutional layers, recurrent layers and attention layers in this method.

[0201] 3) The learning rate is scheduled according to the training rounds. Every 10 training rounds, the current learning rate is multiplied by 0.7 to gradually reduce the learning rate. Based on this, the learning rate can quickly approach the optimal region in the early stage of training and reduce parameter oscillations in the later stage of training to improve convergence stability.

[0202] 4) After each training round, calculate the weighted ordinal accuracy using the validation set;

[0203] Since quality grades have an inherent sequential attribute, engineering practices typically allow for small fluctuations between adjacent grades, but do not permit serious misjudgments across multiple grades. Therefore, the weighted ordinal accuracy is defined as the proportion of evaluation units whose absolute difference between the predicted quality grade and the actual quality grade does not exceed 1 out of all verification evaluation units.

[0204] 5) When the weighted ordinal accuracy on the validation set no longer improves for 15 consecutive training rounds, stop training and save the model parameters corresponding to the optimal performance on the validation set as the final model for the classification of gravity dam concrete construction quality assessment.

[0205] S6. Dynamic Evaluation of Concrete Construction Quality of Gravity Dams

[0206] After completing the classification model training based on hierarchical order constraints in step S5, the saved optimal model parameters are loaded into the evaluation system deployed at the construction site or cloud server.

[0207] For one or more newly collected gravity dam concrete construction records, the 15-dimensional original construction record feature vector, along with the associated dam segment number, pouring layer number, and construction timestamp, are first obtained according to the same attribute definitions and units as in step S1. This new construction record then proceeds to step S2, where a physical inconsistency construction record detection based on material association constraints is performed: checking whether its compressive strength, elastic modulus, ultrasonic wave velocity, permeability coefficient, and defect depth simultaneously violate two or more empirical constraints (such as strength consistency, permeability-defect consistency, and wave velocity consistency). If a physical inconsistency is determined, the system automatically marks the construction record as "data abnormal" and prompts for re-detection or re-entry, excluding it from subsequent quality assessment. If the detection passes, a further thermal process attribute coupling filter based on inter-layer similarity conditions is performed, searching for neighboring construction records with similar intermittent periods and pouring layer numbers within the same dam segment. Weighted fusion and drift correction are applied to pouring temperature, ambient temperature, and intermittent period to obtain a physically consistent construction feature vector.

[0208] Then, proceed to step S3, where three derived features are constructed based on the physically consistent construction feature vector: interlayer bonding index, temperature difference cumulative damage factor, and relative permeability defect index, forming an 18-dimensional extended construction feature vector. Then, based on the water-cement ratio and fly ash content, the construction record is divided into the corresponding mix proportion partition, and standardization and local linear embedding dimensionality reduction are performed within the partition. The soft partition weight is obtained by calculating the distance from the current construction record to the center of each mix proportion partition, and then weighted and fused to obtain a unified 8-dimensional reduced construction feature vector.

[0209] Then, proceed to step S4. Using the new construction record as the target pouring layer, retrieve the top 5 valid historical construction records in reverse order of construction time within the same dam section (fill with zeros if insufficient), construct a fixed-length pouring sequence of length 6, and input this sequence into the pre-trained pouring sequence attention classification network: the sequence first undergoes one-dimensional dilated convolutional encoding to extract the local combination patterns of this layer with the previous 2 and 4 layers; then, through bidirectional gated recurrent units, inter-layer information is transmitted in both forward and backward directions to obtain the bidirectional hidden state of each temporal position; then, through attention pooling layers, all valid temporal positions are weighted and summed to highlight the key pouring layers most relevant to the current construction quality (such as the intermittent period abnormal layer, temperature difference abnormal layer, or defect significant layer), to obtain the context feature vector; finally, through fully connected mapping and Softmax normalization, output the probability values ​​of 5 quality levels.

[0210] The system selects the grade with the highest probability as the predicted quality grade for the current construction record, and can also output the sum of the probabilities of "unqualified" and "severely unqualified" as a risk warning indicator. For gravity dams under continuous construction, the system automatically repeats the above evaluation process after each new pouring layer is completed and test data is obtained, updating the predicted quality grade results of each layer within the current dam section in real time, and displaying them to on-site management personnel in the form of time series curves, dam section cloud maps, or reports. If the predicted grade is "unqualified" or "severely unqualified," the system immediately issues a warning signal, prompting on-site verification or remedial measures. Through this dynamic evaluation mechanism, layer-by-layer, real-time, and traceable intelligent evaluation of the concrete construction quality of gravity dams is achieved.

[0211] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic evaluation method for the construction quality of gravity dam concrete based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect the original data of gravity dam concrete construction and mark the quality grade of each construction record; S2. Based on the physical relationship of concrete materials, identify and remove construction records with physical inconsistencies, perform interlayer coupling noise suppression on thermal process-related attributes, and obtain a physically consistent construction feature vector. S3. Based on the physical consistency construction feature vector, construct three types of derived features: interlayer bonding index, temperature difference cumulative damage factor and relative permeability defect index. Perform partitioned local linear embedding for different mix proportion conditions, and use soft partitioning mapping to obtain dimensionality-reduced construction feature vectors. The specific process is as follows: For each physically consistent construction feature vector Extraction interval, vibration time, slump, air content and fly ash content; Calculate the interlayer bonding index , Indicates the first The interlayer bonding index for each construction record is calculated as follows: First, the foundation bonding score is obtained by dividing the vibration time by the sum of the interval and 0.1 days. Within the current dam section, the median and dispersion intervals of slump, air content, and fly ash content are statistically analyzed. If the slump of the current construction record falls within the middle range of the normal distribution of the current dam section, the compatibility weight is increased. If both the air content and fly ash content of the current construction record deviate from the normal distribution of the dam section, the compatibility weight is decreased. The foundation bonding score is multiplied by the compatibility weight, and the result is normalized within the dam section to obtain the interlayer bonding index. ; Calculate the cumulative damage factor due to temperature difference , Indicates the first The cumulative damage factor of temperature difference in a construction record: First, sort the construction time in the same dam section to determine all historical pouring layers before the current construction record; for each historical pouring layer, calculate the temperature difference between the pouring temperature and the ambient temperature of that layer, and use the temperature difference as the contribution of single-layer thermal stress. Based on the time interval between the historical pouring layer and the current construction record, the contribution of single-layer thermal stress is exponentially decayed; The attenuated thermal stress contributions of all historical pouring layers are summed and normalized according to the statistical scale of the current dam section to obtain the cumulative thermal damage factor. ; Calculate the relative penetration defect index , Indicates the first The relative permeability defect index of a construction record is calculated as follows: The mean defect depth of all construction records within the current pouring layer or adjacent layers of the current dam section is calculated; the relative defect depth is obtained by dividing the current construction record's defect depth by this mean value; logarithmic compression is applied to the permeability coefficient of the current construction record; the relative defect depth is then multiplied by the logarithmically compressed permeability coefficient to obtain the relative permeability defect index. ; The three derived features are concatenated with the 15-dimensional physically consistent construction feature vector to obtain the extended construction feature vector. ; Based on the extended construction feature vector, the construction records are divided into multiple partitions according to the mix proportion related attributes. Then, local linear embedding is performed independently within each partition, and the dimensionality-reduced construction feature vector is obtained through soft partitioning mapping. From extended construction feature vectors Two mix proportion-related attributes, water-cement ratio and fly ash content, are extracted to establish the zoning criteria. Based on these attributes, all construction records are divided into three mix proportion zones. Within each mix proportion zone, the extended construction feature vector is standardized. For each construction record within each mix proportion zone, 12 nearest neighbor records are searched using Euclidean distance to construct a local neighborhood. For each construction record, the local reconstruction weight is calculated: the current construction record is represented as a weighted linear combination of its neighboring construction records, with the sum of all reconstruction weights set to 1, aiming to minimize the reconstruction error. Based on the local reconstruction weights of all construction records, local linear embedding is performed within each mix proportion zone to obtain the corresponding 8-dimensional low-dimensional embedding coordinates. The soft partition weights are calculated based on the distance from the current construction record to the center of each mix proportion zone. The mapping results of each mix proportion zone are then weighted and fused according to the soft partition weights to obtain the dimensionality-reduced construction feature vector. Stack the dimensionality-reduced construction feature vectors of all construction records row by row to obtain the dimensionality-reduced construction feature matrix. ; S4. Construct a casting sequence attention classification network. Take the fixed-length casting sequence consisting of the target casting layer and its preceding casting layers as input. Use one-dimensional dilated convolution to extract local combination patterns, use bidirectional gated recurrent units to capture the inter-layer correlation, and use attention pooling to highlight key casting layers. Finally, output the quality level probability. S5. The pouring sequence attention classification network is trained using category-weighted order constraint loss to obtain the trained pouring sequence attention classification network. S6. After processing the newly collected gravity dam concrete construction records through steps S2 and S3, input them into the trained pouring sequence attention classification network to obtain the gravity dam concrete construction quality assessment results.

2. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 1, characterized in that, In step S1, the original data for the gravity dam concrete construction includes pouring temperature, ambient temperature, water-cement ratio, interval, vibration time, slump, air content, fly ash content, unit cementitious material dosage, compressive strength, elastic modulus, ultrasonic wave velocity, curing time, permeability coefficient, and defect depth; the quality grade labels include excellent, qualified, basically qualified, unqualified, and seriously unqualified; each construction record is simultaneously associated with the dam section number, pouring layer number, construction timestamp, and quality grade label.

3. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 2, characterized in that, In step S2, multiple material-related constraints are checked for each original construction record. When the same construction record violates two or more constraints simultaneously, it is determined to be a physically inconsistent construction record and is removed. From the feature vector of the original construction records The compressive strength, elastic modulus, ultrasonic wave velocity, permeability coefficient, and defect depth are extracted. Based on the empirical relationship between elastic modulus and compressive strength, a strength consistency judgment result for the current construction record is generated. Based on the coupling relationship between permeability coefficient and defect depth, a permeability-defect consistency judgment result for the current construction record is generated. Based on the empirical relationship between elastic modulus and ultrasonic wave velocity, a wave velocity consistency judgment result for the current construction record is generated. The strength consistency judgment result, the permeability-defect consistency judgment result, and the wave velocity consistency judgment result are accumulated to obtain the physical contradiction count value of the current construction record.

4. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 3, characterized in that, In step S2, adaptive filtering is applied to the thermal process-related attributes according to the principles of the same dam section, similar intermittent periods, adjacent pouring layers, and time order, so as to suppress random fluctuations while retaining the real abrupt changes caused by cold joints. For the retained construction records, sets of attributes to be filtered are established for the first dimension (pouring temperature), the second dimension (ambient temperature), and the fourth dimension (interval period). Construction records within each dam section are sorted in ascending order by construction timestamp, and each record is assigned a time sequence within its layer. For the current construction record, neighboring construction records are searched within the same dam section to construct a set of neighboring construction records. For each attribute to be filtered in the current construction record, a neighborhood weighted mean is calculated: a time weight is generated based on the construction time difference between the neighboring construction records and the current construction record, and a working condition weight is generated based on the interval period difference. The time weight and the working condition weight are multiplied to obtain a comprehensive neighborhood weight, which is then applied to the pouring temperature, ambient temperature, and interval period of the neighboring construction records. The weighted mean is calculated for each period. If no neighboring construction record that meets the conditions is found in the current construction record, the original value of the current construction record is directly retained without smooth replacement. The self-retention coefficient of the current construction record is calculated based on the time sequence within the layer and the interval of the current construction record. The pouring temperature, ambient temperature and interval are weighted and fused item by item by multiplying the original value of the current construction record by the self-retention coefficient and the neighboring weighted mean by the remaining weight to obtain the filtered attribute value. In order to correct the slow drift of thermal process attributes in the same dam section during long-term collection, local weighted linear regression is performed on the pouring temperature and ambient temperature with the construction timestamp as the independent variable in the same dam section to estimate the slow drift trend. Next, subtract the drift trend component from the current measurement value and superimpose the local mean to obtain the final corrected value; finally, a physically consistent construction feature vector is obtained. , Indicates the first The feature vector of the construction record after physical consistency correction and inter-layer coupling filtering.

5. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 1, characterized in that, In step S4, one-dimensional dilated convolution is used to extract local combination patterns: The evaluation unit is constructed with the target pouring layer as the center: a dimension-reduced construction feature matrix is ​​obtained. The OK That is, the first Dimensionally reduced construction feature vectors for each construction record; for each dimensionality-reduced construction feature vector... Within the same dam section, the top 5 valid construction records are retrieved in reverse order of construction time, and together with the current construction record, a pouring sequence of length 6 is formed. Using a casting sequence of length 6 as network input, one-dimensional dilated convolutional encoding is performed on the casting sequence to obtain local convolutional encoding feature vectors. Local convolution encoding feature vectors Used to characterize the In the evaluation unit, the first The local combination pattern of each temporal position and its preceding layer is used, the convolution direction is unfolded along the construction time axis, the convolution kernel length is 3, the dilation rate is 2, and the number of output channels is 32.

6. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 5, characterized in that, In step S4, a bidirectional gated recurrent unit is used to capture the correlation between layers, and attention pooling is used to highlight key pouring layers, ultimately outputting the quality level probability: The local convolutional encoding feature vector is input into a bidirectional gated recurrent unit (BRN) network. The forward-gated BRN updates the hidden state from front to back in time, while the backward-gated BRN updates the hidden state from back to front in time. Both the forward and backward hidden states have a dimension of 64. For each temporal position, the forward and backward hidden states are concatenated along the channel dimension to obtain the bidirectional hidden state vector. For each bidirectional hidden state vector Attention scores are calculated using a fully connected transform layer and a hyperbolic tangent activation function. The attention scores for all valid temporal positions within the same evaluation unit are then normalized to obtain the attention weights. The context feature vector is obtained by weighting and summing all bidirectional hidden state vectors according to the attention weights. ; Context feature vector The data is fed into the output layer, and after fully connected mapping and Softmax normalization, a quality level probability vector is obtained. quality level probability vector The level with the highest probability, as the first The predicted quality level of each evaluation unit. Indicates the first The quality level probability vector of each evaluation unit.

7. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 1, characterized in that, In step S5, the category-weighted order constraint loss is calculated: For the Each evaluation unit has a true quality level denoted as [blank]. The quality level probability vector output by the network is denoted as... , its first Each component is denoted as , Indicates the first The evaluation unit belongs to the first The predicted probability of each quality level; For each evaluation unit, calculate the negative logarithm of the probability of the true quality level: take the predicted probability corresponding to the true quality level of the current evaluation unit. Take the negative logarithm of it as the basic classification loss. ; For each evaluation unit, calculate the order deviation term. The predicted expected level is calculated based on the quality level probability vector. The five level numbers are multiplied by their corresponding predicted probabilities and summed to obtain the predicted center level of the current evaluation unit. The absolute difference between the predicted center level and the actual quality level is used as the order bias term. ; For each evaluation unit, calculate the level interval penalty item. : Iterate through all non-true quality levels in sequence, calculate the level difference between the current non-true quality level and the true quality level to obtain the level interval penalty term; calculate the class weight based on the number of evaluation units for each quality level in the training set; add the classification basic loss, order bias term and level interval penalty term in a weighted manner, and then multiply by the loss weight corresponding to the true quality level to which the current evaluation unit belongs to obtain the total loss of the current evaluation unit. The order constraint loss used in this round of training is obtained by averaging the total loss of all evaluation units.

8. The method for dynamic evaluation of the construction quality of gravity dam concrete based on artificial intelligence according to claim 7, characterized in that, All evaluation units are divided according to dam sections to form training, validation, and test sets; the AdamW optimizer is used to update network parameters, with an initial learning rate of 0.0005 and a weight decay coefficient of [missing value]. The batch size can be 32; the learning rate is scheduled according to the training rounds. After every 10 training rounds, the current learning rate is multiplied by 0.7 to gradually decrease the learning rate; after each training round, the weighted ordinal accuracy is calculated using the validation set; training stops when the weighted ordinal accuracy on the validation set no longer improves for 15 consecutive training rounds.

Images