A prefabricated beam intelligent maintenance method and system suitable for complex environmental conditions

Abstract

The present application relates to the technical field of precast beam intelligent maintenance, in particular to a precast beam intelligent maintenance method and system suitable for complex environmental conditions. The method comprises multi-source data acquisition, collecting environmental parameters and beam body state parameters related to precast beam maintenance and performing data fusion and preliminary correction; data preprocessing, cleaning, standardizing, feature engineering processing and data set division of the collected original data; neural network model construction, building a triple hybrid intelligent model based on improved BP neural network, fuzzy control and attention mechanism; model training and optimization, using a phased training strategy combined with an improved optimization algorithm to train the model, and preventing model overfitting through multiple regularization techniques; the present application optimizes the fuzzy rule base through genetic algorithm, improves the reasoning accuracy and efficiency by combining T-S fuzzy reasoning, and significantly improves the reliability of maintenance decision by combining expert experience and data-driven advantages.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance technology for precast beams, and in particular to an intelligent maintenance method and system for precast beams suitable for complex environmental conditions. Background Technology

[0002] In bridge construction, the curing of precast beams is a crucial step in ensuring their quality and performance. However, actual construction environments are often highly complex, with temperatures fluctuating wildly from -10℃ to 40℃ and humidity ranging dramatically from 30% to 95%. Traditional precast beam curing methods, such as manual periodic watering or simple timed spraying, have many drawbacks. In complex environments, manual curing struggles to accurately monitor environmental changes and adjust curing measures in a timely manner, leading to poor curing results. Timed spraying, on the other hand, cannot adjust curing strategies in real time based on the actual condition of the beam and environmental parameters. In extremely low or high temperatures or abnormal humidity conditions, over- or under-curing may occur. For example, in hot and dry environments, timed spraying may not meet the beam's moisture requirements, causing rapid evaporation of moisture from the beam surface and resulting in shrinkage cracks; in low-temperature and high-humidity environments, excessive spraying can cause the beam surface temperature to drop too low, affecting the strength development of the concrete. According to relevant statistics, due to the limitations of traditional maintenance methods in complex environments, precast beams are prone to defects such as cracks and insufficient strength, which seriously affect the quality and durability of bridge projects and increase the cost of later maintenance. Summary of the Invention

[0003] To address the problem that traditional precast beam maintenance methods in complex environments cannot accurately adjust maintenance strategies in real time, this invention proposes an intelligent maintenance method and system for precast beams suitable for complex environmental conditions. Through data acquisition, preprocessing, neural network model construction and training, and combined with intelligent decision-making algorithms, the maintenance strategy is dynamically optimized.

[0004] In a first aspect, the present invention provides an intelligent maintenance method for precast beams suitable for complex environmental conditions, which adopts the following technical solution: A smart curing method for precast beams suitable for complex environmental conditions includes: Step 1: Multi-source data acquisition, collecting environmental parameters related to the curing of precast beams and the beam's own state parameters, and performing data fusion and preliminary correction; Step 2, data preprocessing, involves cleaning, standardizing, performing feature engineering on the collected raw data, and partitioning the dataset; Step 3: Neural network model construction, building a triple hybrid intelligent model based on improved BP neural network, fuzzy control and attention mechanism; Step 4, Model Training and Optimization: A phased training strategy combined with an improved optimization algorithm is used to train the model, and multiple regularization techniques are used to prevent overfitting. Step 5: Model data processing and maintenance strategy generation. Real-time preprocessed data is input into the trained model to obtain prediction results. After adaptive correction based on different maintenance stages of the precast beam, maintenance instructions are generated and executed. At the same time, a maintenance effect feedback and model adaptive update mechanism is established to form a closed-loop maintenance process.

[0005] Furthermore, the multi-source data acquisition in step one specifically includes: collecting environmental parameters using a master-slave sensor redundancy deployment method, the environmental parameters including ambient temperature, ambient humidity, ultraviolet intensity, and ambient wind speed, wherein the ambient temperature is collected by a platinum resistance master sensor and a thermocouple slave sensor, and the effective temperature data is obtained by fusion using a weighted average algorithm, and the ambient wind speed is converted into the equivalent wind speed on the beam surface using a height correction formula; collecting beam state parameters using a combination of pre-embedded sensors, surface monitoring sensors, and non-destructive testing equipment, the beam state parameters including the internal humidity of the beam, beam strain, surface and internal temperature of the beam, and the width of cracks on the beam surface.

[0006] Furthermore, in step one, data transmission adopts a wired + wireless dual-mode transmission strategy. Near-range sensors use LoRa wireless transmission, while long-range or critical data uses RS485 wired transmission. Before data transmission, it is encoded using the LZ77 compression algorithm and a CRC-16 checksum is added. Data storage adopts a local edge storage + cloud backup mode. The local storage stores real-time data for the past 7 days, while the cloud uses a time-series database cluster to store historical data. The storage format includes the acquisition timestamp, sensor number, parameter type, original value, correction value, reliability, and transmission status information.

[0007] Furthermore, the data preprocessing in step two specifically includes: using the 3σ criterion combined with the isolated forest algorithm for dual outlier detection; processing missing values ​​through a hierarchical interpolation strategy; removing sensor noise using a combination of wavelet threshold denoising and Kalman filtering; standardizing parameters that follow a normal distribution using Z-score and non-normally distributed parameters using min-max standardization, followed by uniform normalization calibration; constructing a three-dimensional feature system of basic features, coupled features, and trend features; obtaining the optimal input feature set using a three-level screening strategy of Pearson correlation coefficient method, mutual information method, and L1 regularization; dividing the training set, validation set, and test set in a 7:2:1 ratio; and performing data augmentation using time series augmentation and noise perturbation strategies.

[0008] Furthermore, the neural network model construction in step three specifically includes: the number of input layer nodes equals the dimension of the optimal input feature set, and the input vector is processed by batch normalization; a channel attention mechanism is introduced, and attention weights are generated through global average pooling and two fully connected layers to perform weighted processing on the input features; a three-layer hidden layer structure is adopted, using ReLU, ELU, and Sigmoid activation functions respectively, and the number of hidden layer nodes is determined by empirical formulas and grid search; the fuzzy processing layer uses the output of the third hidden layer as fuzzy input, adopts an improved Gaussian membership function, optimizes the fuzzy rule base through a genetic algorithm, and obtains the fuzzy output result using the TS fuzzy inference method; the number of output layer nodes corresponds to the maintenance control parameters, and an output constraint layer is added to limit the range of values ​​of each output parameter using a saturation function.

[0009] Furthermore, the model training and optimization in step four specifically includes: employing a weighted hybrid loss function of mean squared error, mean absolute percentage error, and maintenance effect penalty term, wherein the maintenance effect penalty term takes effect when the maintenance effect evaluation index fails to meet the standard; using an improved AdamW optimization algorithm with a cosine annealing learning rate strategy; employing a phased training strategy, in which the first phase fixes the parameters of the fuzzy processing layer to train the neural network part, and in the second phase unlocks the parameters of the fuzzy processing layer to jointly train the entire hybrid model, combined with an early stopping strategy to prevent overfitting; and employing multiple regularization methods such as Dropout, L2 regularization, and gradient clipping to avoid model overfitting and gradient explosion.

[0010] Furthermore, the model performance evaluation in step four adopts a multi-dimensional evaluation system, including prediction accuracy indicators, generalization ability indicators, and maintenance effect indicators. The prediction accuracy indicators include root mean square error, mean absolute error, and mean absolute percentage error. The generalization ability indicators include the error difference between the test set and the validation set and the K-fold cross-validation accuracy. The maintenance effect indicators include the beam strength growth rate, the probability of crack formation, and the concrete carbonation depth. When all indicators meet the preset threshold, the model is deemed to have passed the training.

[0011] Furthermore, step five, model data processing and maintenance strategy generation, specifically includes: building a real-time data preprocessing pipeline; using TensorRT to optimize the model to improve inference speed; introducing an inference confidence assessment mechanism; triggering a backup model when the confidence level is lower than a preset threshold; refining the maintenance stage into five stages: initial setting stage, final setting stage, early strength growth stage, mid-term strength growth stage, and late strength growth stage; using support vector machines to train and obtain dynamic correction coefficients; performing stage-adaptive correction on the model output results; converting the corrected maintenance control parameters into standardized PLC control instructions; the maintenance execution module using closed-loop control logic to execute the instructions; and collecting maintenance effect data in real time and feeding it back to the intelligent decision-making module.

[0012] Furthermore, the adaptive update and iterative optimization of the model in step five specifically includes: periodically using an incremental learning algorithm to incrementally train the model, updating the model parameters and fuzzy rule base; using Git version control tools to manage model versions, replacing the currently running model when the performance of the new model improves by a preset percentage; and automatically triggering an emergency training process for the model when encountering extreme environmental conditions, quickly training a temporary adaptive model using historical extreme condition data and real-time collected data to ensure the maintenance effect under extreme conditions.

[0013] Secondly, a smart curing system for precast beams suitable for complex environmental conditions includes: The data acquisition module is configured to acquire multi-source data, collect environmental parameters related to the curing of precast beams and the beam's own state parameters, and perform data fusion and preliminary correction. The preprocessing module is configured to clean, standardize, perform feature engineering on, and partition the collected raw data. The neural network model building module is configured to build a triple hybrid intelligent model based on an improved BP neural network, fuzzy control, and attention mechanism. The model training and optimization module is configured to train the model using a phased training strategy combined with an improved optimization algorithm, and to prevent overfitting of the model through multiple regularization techniques. The model data processing and maintenance strategy generation module is configured to input real-time preprocessed data into a qualified trained model to obtain prediction results, perform adaptive corrections based on different maintenance stages of the precast beam, generate maintenance instructions and execute them, and establish a maintenance effect feedback and model adaptive update mechanism to form a closed-loop maintenance process.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the intelligent maintenance method for precast beams applicable to complex environmental conditions.

[0015] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for intelligent maintenance of precast beams under complex environmental conditions.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This method constructs a triple hybrid intelligent model of "improved BP neural network + fuzzy control + attention mechanism". The introduction of channel attention mechanism adaptively strengthens the weight of key features, which solves the problem that a single model is difficult to handle the uncertainty of complex environment and the redundancy of multiple features. The fuzzy rule base is optimized by genetic algorithm and combined with TS fuzzy reasoning to improve the reasoning accuracy and efficiency. The integration of expert experience and data-driven advantages significantly improves the reliability of maintenance decisions.

[0017] 2. In the data processing stage, a full-process optimization strategy of "redundant acquisition + dual anomaly detection + combined denoising + three-dimensional feature engineering + multiple feature screening" was adopted to construct a high-quality feature dataset; new data augmentation technology and weighted label design were added to effectively improve the model's generalization ability and adaptability to maintenance priorities.

[0018] 3. In terms of model training and optimization, an innovative phased training strategy and an improved AdamW optimization algorithm are adopted, combined with multiple regularization techniques and early stopping strategies, to solve problems such as gradient vanishing, overfitting, and insufficient generalization ability in traditional training; a multi-dimensional performance evaluation system is established to take into account prediction accuracy, generalization ability, and actual maintenance effect, so as to ensure the engineering practicality of the model.

[0019] 4. A closed-loop curing system of "data acquisition → model reasoning → command execution → effect feedback → model update" was constructed, refining the concrete curing stages and introducing dynamic correction coefficients to achieve dynamic optimization of the curing strategy throughout its entire life cycle; an emergency adaptation mechanism for extreme working conditions was added, which solved the shortcomings of traditional curing methods in extreme environments and greatly improved the curing quality and durability of precast beams. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an intelligent maintenance method for precast beams applicable to complex environmental conditions according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the model architecture of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the experimental results of Embodiment 1 of the present invention; Figure 4 This is another experimental effect diagram of Embodiment 1 of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Example 1 Reference Figure 1 This embodiment provides an intelligent maintenance method for precast beams suitable for complex environmental conditions, comprising: Step 1: Acquiring multi-source data The core of this step is to comprehensively and accurately collect environmental parameters related to the curing of precast beams and the beams' own state parameters, providing fundamental data support for subsequent model analysis. Specifically, this includes: 1. Environmental Parameter Acquisition and Redundancy Deployment: Real-time data acquisition is achieved through a high-precision sensor array deployed by the environmental sensing module. A master-slave sensor redundancy design is adopted to improve data reliability. Specific parameters, acquisition methods, and redundancy strategies are as follows: Ambient temperature T e The main sensor uses a platinum resistance temperature sensor with an accuracy of ±0.5℃, and the slave sensors use thermocouple sensors with an accuracy of ±0.8℃. Both are deployed at three different locations around the precast beam (top, side, and bottom), with one main and one slave sensor at each location. The data acquisition frequency is once every 5 minutes, and the data range is [-10℃, 40℃]. Data fusion uses a weighted average algorithm. α is the weighting coefficient (dynamically adjusted according to sensor accuracy, α=0.7). When the data deviation between the master and slave sensors is greater than 2℃, an abnormal alarm is triggered and the backup sensor is activated. 2. Ambient humidity H e The main sensor uses a capacitive humidity sensor with an accuracy of ±3%, and the slave sensor uses a resistive humidity sensor with an accuracy of ±5%. They are deployed in the same location as the temperature sensor, with a data acquisition frequency of once every 5 minutes, and a data range of [30%, 95%]. Data validity criteria: When the humidity change rate is >10% / min, it is considered abnormal data and corrected using a moving average filtering method. ; 3. Ultraviolet Intensity U: Two identical ultraviolet radiation sensors (accuracy ±2W / m²) were deployed on the top of the precast beam in an unobstructed area (2m apart), with a sampling frequency of once every 10 minutes, and the value range was [0, 1000W / m²] (0 indicates no ultraviolet radiation, and the maximum value is determined according to the local solar radiation intensity); Data consistency verification: When the data deviation between the two sensors is >5%, the average of the two is taken and marked as low confidence data; 4. Ambient Wind Speed ​​V: An additional wind speed sensor is installed (to supplement and optimize the environmental perception dimension). A cup-type wind speed sensor (measurement range 0-20m / s, accuracy ±0.3m / s) is used, deployed at three heights (1m, 3m, and 5m above the ground) on the side of the beam. The data is collected once every 10 minutes, and converted into the equivalent wind speed on the beam surface using a height correction formula. , where k is the von Kármán constant (k=0.41), Z is the sensor deployment height, and Z0 is the surface roughness (Z0=0.5m at the construction site).

[0023] 5. Beam structure condition parameter acquisition and multi-dimensional monitoring: Data is collected through a combination of pre-embedded sensors, surface monitoring sensors, and non-destructive testing equipment, covering both macroscopic and microscopic conditions of the beam. Specific parameters are as follows: internal humidity H of the beam β During the casting of the precast beam, three sets of humidity sensors (accuracy ±2%) were pre-embedded in the core area (≥20cm from the surface), the surface area (5cm from the surface), and the transition area (12cm from the surface) of the beam. The sampling frequency was once every 10 minutes, and the value range was [40%, 100%]. A spatial interpolation algorithm was used to construct the humidity field distribution of the beam. , where φi is a shape function and (x,y,z) are arbitrary spatial coordinates of the beam; 6. Beam strain ε: Strain gauges (accuracy ±1με) are pre-embedded on the top, bottom, and side surfaces of key sections such as mid-span, supports, and 1 / 4 span of the beam. At the same time, fiber optic strain sensors (measurement range -1500με~1500με) are attached to the beam surface for redundant monitoring. The acquisition frequency is 1 time / 5min, and the value range is [-100με, 100με] (tensile strain is positive, compressive strain is negative). 7. Monitoring of beam surface and internal temperature ( , A dual monitoring scheme of "surface measurement + internal pre-embedded" is adopted to simultaneously acquire surface and internal temperature data of the beam, ensuring the directness and accuracy of temperature field monitoring. The specific implementation is as follows: (1) Surface temperature of the beam ( Monitoring: An infrared temperature sensor (measurement range -20℃~100℃, accuracy ±0.5℃) was selected. Eight monitoring areas were evenly divided on the beam surface (two on the top, three on each side, and two on the bottom, covering the key stress sections and areas prone to temperature fluctuations). The sensor was deployed 50cm away from the beam surface (unobstructed and without direct sunlight). The sampling frequency was once every 5 minutes, and the measured value range was [-10℃, 40℃]. The surface temperature data of each area was recorded in real time.

[0024] (2) Internal temperature of the beam ( Monitoring: During the precast beam casting process, temperature sensors are pre-embedded according to the principle of "layered placement + key section reinforcement". Location of sensors: core area of ​​beam (≥20cm from surface), transition area (12cm from surface), surface area (5cm from surface). In each area, one sensor is pre-embedded at each of the three key sections: mid-span, support, and 1 / 4 span. A total of nine temperature sensors are pre-embedded in the entire beam (3 areas × 3 sections). Sensor selection: A PT1000 platinum resistance temperature sensor (measuring range -50℃~200℃, accuracy ±0.3℃) is used. It has the characteristics of being resistant to concrete corrosion and having strong compressive strength. The sensor probe is encapsulated in a stainless steel sleeve and fixed on the steel reinforcement cage before pouring to ensure close contact with the concrete. Data acquisition: Synchronized with the surface temperature sensor, the acquisition frequency is 1 time / 5min, and the measured value range is [-5℃, 50℃] (covering the internal temperature rise scenario caused by the release of heat of hydration of concrete). Data consistency verification: The difference between the data of the pre-embedded sensor in the surface area of ​​the same cross section and the data of the surface infrared sensor in the corresponding area is verified. When the absolute value of the difference is >2℃, the sensor fault alarm is triggered (check whether the pre-embedded sensor package is damaged or whether the surface sensor is affected by environmental interference), and the backup pre-embedded sensor data in the same area is activated (one backup sensor is pre-embedded in the surface area of ​​each critical cross section). Temperature field construction: Combining measured data from 8 surface monitoring points and 9 internal pre-embedded points, a spatial interpolation algorithm was used to construct the three-dimensional temperature field distribution of the beam, which intuitively presents the temperature differences at different locations and depths, providing direct data support for subsequent temperature difference control and maintenance strategies.

[0025] 8. Monitoring of surface cracks in beams: A new high-definition industrial camera (1920×1080 resolution) is deployed on both sides of the beam. The beam surface image is captured once every 30 minutes. The crack width w is extracted by image recognition algorithm (recognition accuracy 0.02mm), with a value range of [0,2mm], as a feedback indicator of maintenance effect.

[0026] 9. Data Transmission and Storage Optimization: Data collected by each sensor adopts a dual-mode transmission strategy of "wired + wireless" to ensure data stability in complex construction environments: ① Near-range sensors (beam surface, surrounding environment) use LoRa wireless transmission (communication distance ≥ 3km, rate 1200bps), while long-range or critical data uses RS485 wired transmission; ② Data is compressed and encoded before transmission: LZ77 compression algorithm is used to reduce data volume, with a compression ratio ≥ 3:1, and a CRC-16 checksum (polynomial x) is added. 16 +x 15 +x 2 +1), ensuring data transmission integrity; ③ Storage adopts a "local edge storage + cloud backup" mode: local storage uses industrial-grade SD cards (capacity 128GB, read / write speed ≥100MB / s) to store real-time data for the past 7 days, and cloud storage uses a time-series database InfluxDB cluster (3-node redundancy) to store historical data. The storage format is "collection timestamp-sensor number-parameter type-raw value-corrected value-reliability-transmission status", where data reliability is comprehensively evaluated by sensor accuracy and data consistency verification results: , For the inherent reliability of the sensor, To verify the credibility of data consistency.

[0027] The raw dataset collected in this step is represented as D0= Where ti is the timestamp of the i-th collection, Ci is the overall credibility of the i-th data set (range [0,1], 1 indicates complete credibility), and n is the total amount of collected data; after data fusion and correction, the effective dataset before preprocessing is obtained. , where n'≤n is the effective data volume.

[0028] Step 2: Data Preprocessing The raw collected data contains noise, missing values, and inconsistencies in units. Preprocessing is required to improve data quality and provide reliable input for model training. The specific steps are as follows: 1. Data cleaning and anomaly detection: Missing value handling: A hierarchical interpolation strategy is adopted, and an appropriate method is selected for different missing value scenarios: ① For data with ≤3 consecutive missing values, a linear interpolation formula is used for calculation: ,in ① Let x(t-1) and x(t+1) be the estimated missing values ​​at time t, where x(t-1) and x(t+1) are the adjacent valid data points before and after time t, respectively; ② For data with consecutive missing values ​​of 3 < k ≤ 10, cubic spline interpolation is used to solve for the coefficients through four adjacent valid data points. ③ For data with more than 10 consecutive missing occurrences, remove the data segment and mark it as an abnormal acquisition period to trigger the sensor fault troubleshooting process.

[0029] 2. Outlier Detection: A dual detection method of "3σ criterion + Isolation Forest algorithm" is adopted: ① First, the 3σ criterion is used for preliminary screening: For parameters x that follow a normal distribution, if |x-μ|>3σ, they are identified as outliers, where μ is the mean of the parameter and σ is the standard deviation; ② For non-normally distributed parameters or outliers missed by the 3σ criterion, the Isolation Forest algorithm is used for further detection: 100 isolated trees are constructed, and the outlier score S(x) of each data point is calculated. When S(x)>0.7, it is identified as an outlier; ③ Outlier correction: Local weighted regression scatter smoothing method (LOWESS) is used for correction. The window size is set to 15, and the weight function is W(i)=(1-|(x-xi) / d|²)² (d is the maximum distance between data points within the window).

[0030] 3. Noise Removal: A combined algorithm of "wavelet threshold denoising + Kalman filtering" is used to process sensor noise. Taking temperature data as an example, the specific steps are as follows: ① Perform db4 wavelet decomposition on the original temperature sequence T, with a decomposition level of 3, to obtain 1 low-frequency component and 3 high-frequency components; ② Calculate the threshold λ=σ×√(2lnn) for each high-frequency component (detail coefficient), where σ=median(|w|) / 0.6745, w is the detail coefficient, and n is the data length; ③ Perform improved threshold processing on the detail coefficients (avoiding the step of hard threshold): w'=sign(w)×(|w|-λ)×exp(-λ / |w|) (when |w|>λ), w'=0 (when |w|≤λ); ④ Perform inverse transform on the processed wavelet coefficients to obtain the preliminary denoised sequence T1; ⑤ Use Kalman filtering for further smoothing: the state equation is T=A·T -1 +B·u -1 +w -1 The observation equation is Z = H·T + v, where A = 1 (state transition matrix), B = 0 (control matrix), H = 1 (observation matrix), and w -1 v represents the process noise (variance Q=0.01) and v represents the observation noise (variance R=0.02), ultimately yielding the denoised temperature sequence T'.

[0031] 4. Data Standardization and Normalization Integration: Due to significant differences in the dimensions and value ranges of various parameters, and the differences in distribution characteristics of some parameters, a "standardization + normalization" integration process is adopted: ① For parameters that follow a normal distribution, Z-score standardization is used: ,in The mean of the parameters, ① Standard deviation; ② For parameters that are not normally distributed (such as UV intensity U, crack width w), min-max standardization is used to map them to the [0,1] interval, and the formula is: in The value after processing the j-th parameter. This is the original value of the parameter. , These are the minimum and maximum values ​​of the parameter in the historical data; ③ The standardized parameters are uniformly normalized and calibrated: To avoid the vanishing gradient during model training due to parameter values ​​approaching 0, the final standardized parameters include: .

[0032] 5. Feature Engineering: Feature extraction: Construct a three-dimensional feature system of "basic features + coupled features + trend features" to enhance the model's ability to represent maintenance needs: ① Basic features: original standardized parameters ② Coupling characteristics: a. Temperature and humidity coupling characteristic F1=Te×(1-He) (characterizing the extreme degree of high temperature and low humidity or low temperature and high humidity); b. Temperature difference between the inside and outside of the beam F2=|Tβ-Te|; c. Ultraviolet-humidity-wind speed coupling characteristic F3=U×(1-He)×V (characterizing the combined effect of ultraviolet radiation, humidity, and wind speed on the evaporation of moisture on the beam surface); d. Strain-temperature coupling characteristic F4=ε×Tβ (characterizing the effect of temperature change on the strain of the beam); ③ Trend characteristics: a. Strain change rate characteristics (Characterizing the beam's shrinkage / expansion rate); b. Characteristics of humidity change trends (Characterizing the short-term trend of humidity variation in the beam); c. Crack development rate characteristics (Characterizes the crack propagation, non-negative).

[0033] 6. Feature Selection: A three-level screening strategy of "Pearson correlation coefficient method + mutual information method + L1 regularization" is adopted to screen and maintain the effectiveness of maintenance (based on beam strength growth rate γ and crack occurrence probability). Concrete carbonation depth For evaluation indicators, the characteristics with high correlation are: ① First level: Pearson correlation coefficient method, the correlation coefficient calculation formula is: Where x represents candidate features and y represents the maintenance effect evaluation index. , Given the average values ​​of x and y respectively, filter out... The first level is the mutual information method, which calculates the mutual information value I(x,y) between the selected features and the evaluation index, and selects features with I(x,y)≥0.2 (the larger the mutual information value, the stronger the correlation between the feature and the index); the second level is the mutual information method, which calculates the mutual information value I(x,y) between the selected features and the evaluation index, and selects features with I(x,y)≥0.2 (the larger the mutual information value, the stronger the correlation between the feature and the index); the third level is the L1 regularization method, which uses the L1 regularization coefficient of the logistic regression model to remove features with coefficients close to 0, and finally obtains the optimal input feature set. (A total of 11 features).

[0034] 7. Label Construction and Weighting: The core control parameters of the maintenance execution module are used as labels, and weights are assigned to the labels in conjunction with maintenance effect evaluation indicators to form a weighted label set: ① Core Control Parameters: a. Spray flow rate Q (range [0.5, 2] L / min); b. Heating / cooling power P (range [0, 3] kW, 0 indicates no operation); c. Shading facility opening / closing degree S (range [0, 1], 0 indicates completely closed, 1 indicates completely open); d. Spray interval time τ (range [5, 30] min, new label added); ② Label Weighting: Considering the differences in maintenance focus at different maintenance stages, weight coefficients are introduced. , , , ( The weighting coefficients are determined by the Analytic Hierarchy Process (AHP), for example, during the initial setting period. , , , Intensity growth period , , , The final label set is represented as Y= Q, ·P, ·S, ·τ].

[0035] 8. Dataset Partitioning and Augmentation: ① Partitioning Strategy: The preprocessed dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used for adjusting model hyperparameters, and the test set is used to evaluate the final model performance. ② Data Augmentation: A "time series augmentation + noise perturbation" augmentation strategy is adopted for the training set to improve the model's generalization ability: a. Time series augmentation: New training samples are generated using the sliding window method (window size of 30 data points, step size of 5 data points); b. Noise perturbation: Gaussian noise (mean 0, standard deviation 0.01) is added to the augmented samples to simulate the small fluctuations in the actual collected data and enhance the model's robustness to noise.

[0036] Step 3: Neural Network Model Construction (Improved Hybrid Intelligent Model) A triple hybrid intelligent model based on "improved BP neural network + fuzzy control + attention mechanism" is constructed. The attention mechanism is introduced to enhance the influence of key features on maintenance decisions, and to achieve accurate mapping from the input feature set X to the maintenance control parameters Y. The model structure consists of five parts: input layer, attention layer, hidden layer, fuzzy processing layer and output layer, as detailed below: 1. Input Layer Design: The number of nodes in the input layer is equal to the dimension of the final input feature set X, i.e., 11 nodes, corresponding to... These 11 features, the input vector is Where x (k=1,2,…,11) are the values ​​of each feature after processing; to improve the stability of model training, the input vector needs to be processed by batch normalization: in The mean of features within the batch. The characteristic variance within the batch, ε=10⁻ 5 To prevent tiny values ​​with a denominator of 0.

[0037] 2. Attention Layer Design: A channel attention mechanism (SENet) is introduced to adaptively adjust the weights of each input feature, strengthening features that significantly influence maintenance decisions and suppressing interference from ineffective features. Specific steps include: ① Performing global average pooling on the batch-normalized features of the input layer: Where N is the batch size. This represents the batch normalized value of the k-th feature of the i-th sample. ① The global pooling result of the k-th feature; ② Construct an attention weight generator through two fully connected layers (FC): 2·σ(W1· 1) + b2, where W1 is the weight of the first fully connected layer (dimension 11×6), W2 is the weight of the second fully connected layer (dimension 6×11), b1 and b2 are biases, and σ is the Sigmoid activation function; ③ Generate attention weights (Value range [0,1]), weighting the input features: · The attention layer output X' is obtained. .

[0038] 3. Hidden Layer Design: A three-layer hidden layer structure is adopted, using different activation functions to improve the model's nonlinear fitting ability: ① First hidden layer (nodes 11→20): ReLU activation function is used (to solve the gradient vanishing problem), and batch normalization is added; ② Second hidden layer (nodes 20→12): ELU activation function is used (to avoid the death problem of ReLU in the negative interval); ③ Third hidden layer (nodes 12→6): Sigmoid activation function is used (output range [0,1], adapting to the input of subsequent blurring layers). The number of hidden layer nodes is determined by "empirical formula + grid search". in Let l be the number of hidden nodes in the l-th layer. This represents the number of nodes in the previous layer. This represents the number of nodes in the next layer. Using empirical coefficients (range [1, 8]), the optimal node combinations are determined to be 20, 12, and 6 through grid search (search range 1-30). Output formulas for each layer: ① Output of the first hidden layer: in Let be the weights from the k-th node in the attention layer to the i-th node in the first hidden layer. The bias is the bias of the i-th node in the first hidden layer, and BN is the batch normalization operation, i=1,2,…,20; ② Output of the second hidden layer: in Let be the weights from the i-th node in the first hidden layer to the j-th node in the second hidden layer. This is the bias of the j-th node in the second hidden layer, where j = 1, 2, ..., 12; ③ Output of the third hidden layer: in Let be the weights from the j-th node in the second hidden layer to the m-th node in the third hidden layer. This is the bias of the m-th node in the third hidden layer, where m = 1, 2, ..., 6.

[0039] 4. Fuzzy Processing Layer Design (Enhanced Version): An adaptive fuzzy control algorithm is introduced, combined with a genetic algorithm to optimize the fuzzy rule base, to handle uncertainties in complex environments. Specifically: ① Fuzzy Input: The output of the third hidden layer... ① As fuzzy input variables (6 in total); ② Fuzzy subset partitioning: The fuzzy subset of each input variable is divided into five categories: "Very Small (VS), Small (S), Medium (M), Large (L), and Maximum (VL)" to improve the accuracy of fuzzy inference; ③ Membership function: An improved Gaussian membership function is adopted (to solve the sensitivity problem of the traditional Gaussian function at the endpoints of the interval): Where A is the fuzzy subset (VS / S / M / L / VL), c is the membership function center, and σ is the width; ④ Fuzzy rule base optimization: The initial rule base is constructed based on expert experience and historical maintenance data (total 5 6 =15625 potential rules), and the rule base was optimized using a genetic algorithm: the fitness function was the minimum fuzzy inference error, the selection operator was roulette wheel selection, the crossover operator was single-point crossover, and the mutation operator was uniform mutation. The number of iterations was set to 100 generations, and the population size was set to 50. Finally, 120 effective rules were selected. For example: "If F1 (temperature and humidity coupling feature) is L, F5 (strain change rate feature) is M, and F7 (crack development rate feature) is VS, then the spray flow rate Q is M, the heating / cooling power P is S, the shading opening degree S is M, and the spray interval time τ is L"; ⑤ Fuzzy inference: The TS fuzzy inference method was used (to improve inference efficiency and accuracy). For the k-th rule, the inference output is ,in 0~ 6 represents the consequent parameters of the rules; ⑥ Fuzzy output fusion: The inference results of all rules are fused using a weighted average method. · ,in The activation degree of the k-th rule (calculated from the membership function of the input variable).

[0040] 5. Output layer design: Output the fuzzy inference results. After clarification processing, the final maintenance control parameters are obtained. The number of output layer nodes is 4, which correspond to the weighted sprinkler flow rate Q, heating / cooling power P, shading facility opening degree S, and sprinkler interval time τ, respectively. To ensure that the output parameters meet the actual engineering constraints, an output constraint layer is added: a saturation function is used to limit the value range of each output parameter.

[0041] in This is the final result of the m-th output node (m=1 corresponds to Q, m=2 corresponds to P, m=3 corresponds to S, and m=4 corresponds to τ). , These are the minimum and maximum values ​​of the m-th output parameter, respectively.

[0042] The hybrid intelligent model is trained using the training set, and the hyperparameters are adjusted using the validation set. The model weights and biases are optimized using gradient descent. The specific steps are as follows: Loss function design (weighted hybrid loss): A weighted hybrid loss function is adopted, consisting of "mean squared error (MSE) + mean absolute percentage error (MAPE) + maintenance effect penalty term", which balances the model's prediction accuracy and the actual maintenance effect. The formula is as follows: Where ω1=0.5, ω2=0.3, and ω3=0.2 are weighting coefficients (satisfying ω1+ω2+ω3=1); the calculation method for each part is as follows: ①Mean Squared Error (MSE): Measures the deviation between the model's predicted output and the actual maintenance control parameters. Where N is the number of training set samples and M is the number of output nodes (M=4). Let be the model's predicted value for the m-th label of the i-th sample. The true value of the m-th label for the i-th sample (determined based on historical best maintenance data); ②Mean Absolute Percentage Error (MAPE): Reduces the proportion of error from large numerical labels, improving the fairness of the loss function. Where ε = 10⁻³ is used to avoid tiny values ​​where the denominator is 0; ③ Penalty for Maintenance Effectiveness: When the maintenance strategy predicted by the model results in unsatisfactory maintenance effectiveness, a penalty term is added to strengthen the model's focus on maintenance effectiveness. in The maintenance effect evaluation index for the i-th sample (k=1 corresponds to the beam strength growth rate γ, k=2 corresponds to the probability of crack formation) k=3 corresponds to the concrete carbonation depth ), The threshold values ​​for each evaluation indicator are (γ≥0.05% / d). , ),when < The penalty will take effect at that time.

[0043] 6. Algorithm Selection and Improvement: An improved Adam optimization algorithm (AdamW) is adopted, introducing a weight decay term to address the insufficient generalization ability of the traditional Adam algorithm. This algorithm combines the advantages of momentum gradient descent, RMSProp algorithm, and weight decay, effectively improving training efficiency, stability, and generalization ability. The parameter update formula is: in , These are the first and second moment estimates of the gradient, respectively, and β1 and β2 are the exponential decay rates (taken as 0.9 and 0.999, respectively). η is the gradient of the loss function with respect to parameters θ (weights w, bias b, fuzzy membership function parameter c, σ, fuzzy rule consequent parameter p), and η is the learning rate (using a cosine annealing learning rate strategy). cos(π·t / T)), where , (T is the total number of iterations), ε is a small value to prevent the denominator from being zero (taken as 10). -8 ), Let t be the parameter value at time t, and λ be the weight decay coefficient (λ=0.0001).

[0044] 7. Model Training Process (Phased Training): A phased training strategy is adopted to improve the accuracy and efficiency of model training: ① First Phase (Pre-training): The parameters of the fuzzy processing layer are fixed, and only the neural network part (input layer, attention layer, hidden layer, output layer) is trained. The number of iterations is set to 50 epochs, and the batch size is... The goal is to enable the neural network to quickly learn the basic mapping relationship between features and labels; ② Second stage (fine-tuning training): Unlock the parameters of the fuzzing layer, jointly train the entire hybrid model, set the number of iterations to 150 epochs, and the batch size... ③ Training process: a. Input training set samples into the model in batches; b. Calculate the model's predicted output through forward propagation. c. Calculate the weighted mixed loss function L; d. Update all model parameters (weights w, bias b, fuzzy membership function parameters c, σ, and fuzzy rule consequent parameters p) through backpropagation using the AdamW algorithm; e. Use an early stopping strategy to prevent overfitting: stop training when the validation set loss does not decrease for 15 consecutive epochs, and save the model parameters with the minimum loss; ④ Learning rate adjustment: during the fine-tuning training phase, perform a learning rate decay every 30 epochs (decay coefficient of 0.8) to further improve training stability.

[0045] 8. Model Regularization Optimization (Multiple Regularization): To prevent overfitting, a multiple regularization technique of "Dropout + L2 Regularization + Gradient Clipping" is adopted: ① Dropout Regularization: The Dropout rates are set to 0.2 and 0.1 in the first and second hidden layers respectively (i.e., randomly discarding 20% ​​and 10% of the hidden layer nodes), as shown in the formula: in The dropout rate for the l-th layer. A random matrix with a 0-1 distribution (the probability that an element is 1 is 1). ), This is the original output of the l-th hidden layer. ① Output after Dropout processing; ② L2 regularization: Add L2 regularization constraints to all neural network weights, with a weight decay coefficient λ=0.0001 to avoid excessively large weights; ③ Gradient clipping: Use a norm clipping strategy, setting the gradient norm threshold to 1.0. When the gradient norm exceeds the threshold, the gradient is clipped proportionally. ·(threshold / To prevent gradient explosion.

[0046] 9. Model Performance Evaluation (Multi-dimensional Evaluation): The performance of the trained model is evaluated from three dimensions—prediction accuracy, generalization ability, and maintenance effect—using a test set. Evaluation metrics include: ① Prediction accuracy metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE); ② Generalization ability metrics: Error difference between the test set and validation set (difference ≤ 0.02), K-fold cross-validation accuracy (K=5, average accuracy ≥ 0.93); ③ Maintenance effect metrics: Beam strength growth rate γ corresponding to the model's output maintenance strategy (≥ 0.05% / d), and crack occurrence probability. (≤5%), Concrete carbonation depth (≤0.5mm); Specific calculation formula: in M represents the number of test set samples, and M represents the number of output nodes (M=4). The average value of the m-th label in the test set is used. When the model satisfies the following conditions: RMSE≤0.04, MAE≤0.02, MAPE≤2%, K-fold cross-validation mean accuracy≥0.93, and maintenance effect indicators meet the standards, the model is considered to be qualified for training and can be used for actual maintenance decision-making.

[0047] Step 5: Model Data Processing and Maintenance Strategy Generation The real-time collected and preprocessed environmental and beam condition data are input into the trained model. The maintenance control parameters are obtained through model inference. The maintenance strategy is then dynamically adjusted in conjunction with the precast beam maintenance stage. The specific steps are as follows: 1. Real-time Data Input and Model Inference Optimization: ① Data Preprocessing Pipeline: Build a real-time data preprocessing pipeline to transform the real-time raw data collected in step one into a model input feature set through an automated process of "data fusion → anomaly detection → noise removal → standardization → feature extraction → attention weighting". ② Accelerated Model Inference: TensorRT is used to optimize the trained hybrid model (quantization accuracy of FP16) to improve inference speed (inference latency ≤50ms) and meet the needs of real-time maintenance decision-making; ③ Reliability Verification of Inference Results: An inference confidence evaluation mechanism is introduced to calculate the confidence of the model output results. · ( For attention weights, (for fuzzy rule activation), when When the confidence level is less than 0.6, the backup model (traditional fuzzy control model) is triggered to output the maintenance strategy, and the batch of data is marked as low confidence data for subsequent incremental training of the model.

[0048] 2. Adaptive Adjustment of Curing Stages (Refined Stages and Dynamic Coefficients): Based on the entire lifecycle of concrete curing, the curing stages are refined into five phases: initial setting (0-12 hours after pouring), final setting (12-72 hours after pouring), early strength growth (72 hours-7 days after pouring), mid-term strength growth (7-14 days after pouring), and late strength growth (14-28 days after pouring). Combining the mechanical properties of concrete and curing requirements at each stage, dynamic correction coefficients are used to adjust the model output. These correction coefficients are obtained by training a Support Vector Machine (SVM) based on historical curing data. Initial setting period (0-12 hours after pouring): The core requirement is to keep the beam surface moist, avoid rapid moisture evaporation and structural disturbance, and the dynamic correction coefficient is as follows: (t represents the time after pouring, which gradually increases over time). , , The corrected parameter is Q1 P1 S1 ,τ1 ; 3. Final setting period (12-72 hours after pouring): The core requirement is precise control of temperature and humidity to promote cement hydration reaction. Dynamic correction coefficient: , 2| / 2) (F2 represents the temperature difference characteristics inside and outside the beam). , The corrected parameter is Q2. P2 S3 ,τ2 ; 4. Early strength development period (72h-7d after pouring): The core requirement is to control the temperature difference between the inside and outside of the beam and prevent shrinkage cracks. Dynamic correction coefficient: Hᵦ / 100), 2|), , The corrected parameter is Q3. P3 S3 ,τ3 ; 5. Mid-term strength growth period (7-14 days after pouring): The core requirements are maintaining suitable humidity and ensuring stable strength growth. Dynamic correction coefficient: Hᵦ / 100), 2| / 3), , The corrected parameter is Q4. P4 S4 ,τ4 ; 6. Later strength development period (14-28 days after pouring): The core requirement is to gradually reduce humidity and improve beam durability. Dynamic correction coefficient: Hᵦ / 100), , , The corrected parameter is Q5. P5 S5 ,τ5 .

[0049] 7. Maintenance Instruction Generation, Execution, and Feedback Closed Loop: ① Instruction Generation: The corrected maintenance control parameters (Q, P, S, τ, m=1-5 corresponding to different maintenance stages) are converted into standardized PLC control instructions. The instruction format adopts the Modbus-RTU protocol and includes fields such as device address, instruction code, parameter value, and checksum; ② Instruction Execution: After receiving the instruction, the maintenance execution module executes it according to the following logic: a. Automatic Sprinkler Device: Operates according to flow rate Q and interval time τ, using a variable frequency pump to control flow accuracy (error ≤ ±0.05L / min), while simultaneously feeding back the actual flow rate in real time through a flow sensor. a. Flow closed-loop control; b. Heating / cooling equipment: Started by power P, the output power of the equipment is controlled by a PID regulation algorithm, with the goal of controlling the ambient temperature around the beam within the range of [15℃, 25℃], with a temperature control accuracy of ≤±1℃; c. Shading facilities: The angle and area are adjusted by a stepper motor according to the opening degree S (angle adjustment accuracy ≤±1°), and the actual opening degree is fed back by a displacement sensor to form a position closed-loop control; ③ Effect feedback: During the maintenance process, the beam state parameters (strain ε, internal humidity Hᵦ, surface crack width w) and environmental parameters are collected in real time to calculate the maintenance effect evaluation index (γ, , The data is then fed back to the intelligent decision-making module, forming a closed-loop maintenance system of "data acquisition → model reasoning → instruction execution → effect feedback".

[0050] 8. Model Adaptive Update and Iterative Optimization: Establish a full lifecycle management mechanism for the model to ensure its adaptability and accuracy under different complex environments: ① Incremental Training: Regularly (every 3 months) collect actual data from the maintenance closed-loop system (including environmental parameters, beam state parameters, maintenance control parameters, and maintenance effect evaluation data), and use the incremental learning algorithm (Finetune) to train the model, update model parameters and fuzzy rule base, and avoid the waste of resources caused by retraining; ② Model Version Management: Use Git version control tools to manage model versions at different times, record training data, hyperparameters, and performance indicators of each version, and replace the currently running model when the performance of the new model is improved by ≥5%; ③ Adaptation to Abnormal Working Conditions: When encountering extreme environmental conditions (such as temperature <-5℃ or >35℃, humidity <40% or >90%), automatically trigger the model emergency training process, and quickly train a temporary adaptive model using historical extreme working condition data and real-time collected data to ensure maintenance effectiveness under extreme environments.

[0051] like Figure 3As shown, the strength growth rate (% / d) of beams under different curing stages (initial setting period, final setting period, early strength, etc.) is compared between traditional curing and this curing scheme: the strength growth rate of traditional curing fluctuates little in each stage, with a maximum of only 0.03% / d (early strength stage); this scheme reaches 0.055% / d in the critical "early strength growth period", which is about 83% higher than traditional curing; the strength growth rate of this scheme is higher than that of traditional curing in each stage, which reflects the scheme's ability to accurately control the hydration reaction of concrete.

[0052] Figure 4 The study compared the crack incidence rates of precast beams under four different curing methods: traditional manual curing had the highest crack incidence rate (28%), followed by timed spray curing (15%); the crack incidence rate of this curing method was only 2%, which is about 93% lower than traditional manual curing and 75% lower than ordinary intelligent curing. This clearly demonstrates that this method effectively avoids shrinkage cracks in the beams caused by uncontrolled temperature and humidity through real-time environmental perception and dynamic adjustment of curing strategies.

[0053] Example 2 This embodiment provides an intelligent curing system for precast beams suitable for complex environmental conditions, including: The data acquisition module is configured to acquire multi-source data, collect environmental parameters related to the curing of precast beams and the beam's own state parameters, and perform data fusion and preliminary correction. The preprocessing module is configured to clean, standardize, perform feature engineering on, and partition the collected raw data. The neural network model building module is configured to build a triple hybrid intelligent model based on an improved BP neural network, fuzzy control, and attention mechanism. The model training and optimization module is configured to train the model using a phased training strategy combined with an improved optimization algorithm, and to prevent overfitting of the model through multiple regularization techniques. The model data processing and maintenance strategy generation module is configured to input real-time preprocessed data into a qualified trained model to obtain prediction results, perform adaptive corrections based on different maintenance stages of the precast beam, generate maintenance instructions and execute them, and establish a maintenance effect feedback and model adaptive update mechanism to form a closed-loop maintenance process.

[0054] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned intelligent maintenance method for precast beams under complex environmental conditions.

[0055] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for intelligent maintenance of precast beams under complex environmental conditions.

[0056] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A smart curing method for precast beams suitable for complex environmental conditions, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition, collecting environmental parameters related to the curing of precast beams and the beam's own state parameters, and performing data fusion and preliminary correction; Step 2, data preprocessing, involves cleaning, standardizing, performing feature engineering on the collected raw data, and partitioning the dataset; Step 3: Neural network model construction, building a triple hybrid intelligent model based on improved BP neural network, fuzzy control and attention mechanism; Step 4, Model Training and Optimization: A phased training strategy combined with an improved optimization algorithm is used to train the model, and multiple regularization techniques are used to prevent overfitting. Step 5: Model data processing and maintenance strategy generation. Real-time preprocessed data is input into the trained model to obtain prediction results. After adaptive correction based on different maintenance stages of the precast beam, maintenance instructions are generated and executed. At the same time, a maintenance effect feedback and model adaptive update mechanism is established to form a closed-loop maintenance process.

2. The method according to claim 1, characterized in that, The multi-source data acquisition in step one specifically includes: collecting environmental parameters using a master-slave sensor redundancy deployment method. These environmental parameters include ambient temperature, ambient humidity, ultraviolet intensity, and ambient wind speed. The ambient temperature is collected by a platinum resistance master sensor and a thermocouple slave sensor, and the effective temperature data is obtained by fusion using a weighted average algorithm. The ambient wind speed is converted into the equivalent wind speed on the beam surface using a height correction formula. The beam state parameters are collected using a combination of pre-embedded sensors, surface monitoring sensors, and non-destructive testing equipment. These beam state parameters include the internal humidity of the beam, beam strain, surface and internal temperature of the beam, and the width of cracks on the beam surface.

3. The method according to claim 1, characterized in that, In step one, data transmission adopts a wired + wireless dual-mode transmission strategy. Near-range sensors use LoRa wireless transmission, while long-range or critical data uses RS485 wired transmission. Before data transmission, it is encoded using the LZ77 compression algorithm and a CRC-16 checksum is added. Data storage adopts a local edge storage + cloud backup mode. The local storage stores real-time data for the past 7 days, while the cloud uses a time-series database cluster to store historical data. The storage format includes the acquisition timestamp, sensor number, parameter type, original value, correction value, reliability, and transmission status information.

4. The method according to claim 1, characterized in that, The data preprocessing in step two specifically includes: using the 3σ criterion combined with the isolated forest algorithm for dual outlier detection; handling missing values ​​through a hierarchical interpolation strategy; removing sensor noise using a combination of wavelet threshold denoising and Kalman filtering; standardizing parameters that follow a normal distribution using Z-score and non-normally distributed parameters using min-max standardization, followed by uniform normalization calibration; constructing a three-dimensional feature system of basic features, coupled features, and trend features; obtaining the optimal input feature set using a three-level screening strategy of Pearson correlation coefficient method, mutual information method, and L1 regularization; dividing the training set, validation set, and test set in a 7:2:1 ratio; and performing data augmentation using time series augmentation and noise perturbation strategies.

5. The method according to claim 1, characterized in that, The neural network model construction in step three specifically includes: the number of nodes in the input layer equals the dimension of the optimal input feature set, and the input vector is processed by batch normalization; a channel attention mechanism is introduced, and attention weights are generated through global average pooling and two fully connected layers to perform weighted processing on the input features; a three-layer hidden layer structure is adopted, using ReLU, ELU, and Sigmoid activation functions respectively, and the number of nodes in the hidden layer is determined by empirical formulas and grid search; the fuzzy processing layer uses the output of the third hidden layer as the fuzzy input, adopts an improved Gaussian membership function, optimizes the fuzzy rule base through a genetic algorithm, and obtains the fuzzy output result using the TS fuzzy inference method; the number of nodes in the output layer corresponds to the maintenance control parameters, and an output constraint layer is added to limit the value range of each output parameter using a saturation function.

6. The method according to claim 1, characterized in that, Step four, model training and optimization, specifically includes: employing a weighted hybrid loss function of mean squared error, mean absolute percentage error, and maintenance effect penalty term, whereby the maintenance effect penalty term takes effect when the maintenance effect evaluation index fails to meet the standard; using an improved AdamW optimization algorithm with a cosine annealing learning rate strategy; adopting a phased training strategy, in which the first phase fixes the parameters of the fuzzy processing layer to train the neural network part, and in the second phase unlocks the parameters of the fuzzy processing layer to jointly train the entire hybrid model, combined with an early stopping strategy to prevent overfitting; and employing multiple regularization methods such as Dropout, L2 regularization, and gradient clipping to avoid model overfitting and gradient explosion.

7. The method according to claim 1, characterized in that, In step four, the model performance evaluation adopts a multi-dimensional evaluation system, including prediction accuracy indicators, generalization ability indicators, and maintenance effect indicators. The prediction accuracy indicators include root mean square error, mean absolute error, and mean absolute percentage error. The generalization ability indicators include the error difference between the test set and the validation set and the K-fold cross-validation accuracy. The maintenance effect indicators include the beam strength growth rate, the probability of crack formation, and the concrete carbonation depth. When all indicators meet the preset threshold, the model is deemed to have passed the training.

8. The method according to claim 1, characterized in that, Step five, model data processing and maintenance strategy generation, specifically includes: building a real-time data preprocessing pipeline; using TensorRT to optimize the model to improve inference speed; introducing an inference confidence assessment mechanism, triggering a backup model when the confidence level is lower than a preset threshold; refining the maintenance stage into five stages: initial setting stage, final setting stage, early strength growth stage, mid-term strength growth stage, and late strength growth stage; using support vector machines to train and obtain dynamic correction coefficients; performing stage-adaptive correction on the model output results; converting the corrected maintenance control parameters into standardized PLC control instructions; the maintenance execution module using closed-loop control logic to execute the instructions; and collecting maintenance effect data in real time and feeding it back to the intelligent decision-making module.

9. The method according to claim 1, characterized in that, Step five, model adaptive updating and iterative optimization, specifically includes: periodically using incremental learning algorithms to incrementally train the model, updating model parameters and fuzzy rule base; using Git version control tools to manage model versions, replacing the currently running model when the performance of the new model improves by a preset percentage; and automatically triggering an emergency model training process when encountering extreme environmental conditions, quickly training a temporary adaptive model using historical extreme condition data and real-time collected data to ensure maintenance effectiveness under extreme environments.

10. A smart curing system for precast beams suitable for complex environmental conditions, characterized in that, include: The data acquisition module is configured to acquire multi-source data, collect environmental parameters related to the curing of precast beams and the beam's own state parameters, and perform data fusion and preliminary correction. The preprocessing module is configured to clean, standardize, perform feature engineering on, and partition the collected raw data. The neural network model building module is configured to build a triple hybrid intelligent model based on an improved BP neural network, fuzzy control, and attention mechanism. The model training and optimization module is configured to train the model using a phased training strategy combined with an improved optimization algorithm, and to prevent overfitting of the model through multiple regularization techniques. The model data processing and maintenance strategy generation module is configured to input real-time preprocessed data into a qualified trained model to obtain prediction results, perform adaptive corrections based on different maintenance stages of the precast beam, generate maintenance instructions and execute them, and establish a maintenance effect feedback and model adaptive update mechanism to form a closed-loop maintenance process.

Images