Transfer learning optimization system and method for predicting early-age strength of concrete

Abstract

The invention relates to the field of civil engineering and artificial intelligence, in particular to a transfer learning optimization system and method for predicting the early-age strength of concrete, and the system comprises a multi-scale data sensing module, a physical information constrained deep neural network module, a formula adaptive transfer learning module, a Bayesian optimization prediction module and a federated learning feedback module. The whole process from data collection to model optimization is achieved, through the system, the concrete strength prediction errors of the extremely early age and the standard age are remarkably reduced to + / -5% and + / -3% respectively, meanwhile, the number of concrete test pieces for testing is reduced by 85%, the material and labor cost is greatly saved, and the method not only improves the prediction precision, but also reduces the construction cost. And through continuous learning and feedback, the prediction model is continuously optimized, and an efficient and economical concrete strength prediction solution is provided for actual engineering.

Description

Technical Field

[0001] The present invention relates to the fields of civil engineering and artificial intelligence, and in particular to a transfer learning optimization system and method for predicting early-age strength of concrete. Background Art

[0002] The prediction of early-age strength of concrete is of great significance for the construction progress and quality control of construction projects. Traditional concrete strength prediction mainly adopts the reserved test block method, that is, standard test pieces are reserved when pouring concrete, and compressive strength tests are carried out at different ages to analyze their changing patterns. In the prior art, a common method is based on temperature monitoring and maturity theory, which calculates the equivalent age by the temperature change inside the concrete, and then infers its strength. For example, Chinese patent CN118350489A discloses a method for predicting the very early-age strength of concrete. This method determines the strength of concrete within 12-24 hours under standard curing conditions, establishes the relationship formula between age and strength, and uses equivalent age calculation to predict the actual concrete strength.

[0003] However, existing technologies have the following shortcomings: First, they rely heavily on a single temperature parameter and cannot fully reflect the complex changes in the concrete hydration process; second, they use preset empirical formulas for fitting and lack the ability to adapt to concrete with different mix proportions; third, a separate prediction model needs to be established for each mix proportion and each project, making knowledge reuse difficult; in addition, these methods provide single-point predictions, lack uncertainty quantification, and cannot provide risk assessment for decision-making; finally, the models are static and fixed, lacking adaptability, and cannot accumulate experience from multiple engineering projects to improve prediction accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a transfer learning optimization system and method for predicting the early-age strength of concrete, so as to solve the problems in the prior art such as single data source, poor model adaptability, difficulty in knowledge reuse, and limited prediction accuracy.

[0005] The present invention proposes a transfer learning optimization system for predicting early-age strength of concrete, including:

[0006] Multi-scale hydration process data perception module, used to collect multiple physical parameters inside concrete to form multi-modal time series data;

[0007] A physical information constrained deep neural network module is connected to the multi-scale hydration process data perception module, and is used to receive the multimodal time series data and generate a concrete strength prediction model based on a combination of physical law constraints and data-driven methods;

[0008] A transfer learning module for formula adaptation is connected to the deep neural network module for physical information constraints, and is used to transfer the model knowledge trained based on standard test blocks to actual engineering concrete to generate a specific formula adaptability model;

[0009] A Bayesian optimized multi-time scale prediction module, connected to the formula adaptive transfer learning module, is used to generate a concrete strength evolution curve and confidence interval based on the specific formula adaptability model;

[0010] The continuously evolving federated learning feedback module is connected to the Bayesian optimized multi-time scale prediction module to integrate the prediction results and measured data of multiple engineering projects, update the global knowledge base and optimize the concrete strength prediction model.

[0011] Preferably, the multi-scale hydration process data perception module includes:

[0012] Micro-scale sensing unit, used to collect data on changes in hydration microstructure;

[0013] Mesoscale sensing unit, used to measure temperature field distribution;

[0014] Macro-scale sensing units to monitor wave velocity changes;

[0015] integral scale sensing unit for monitoring resistivity changes;

[0016] The data preprocessing unit is connected to the above-mentioned sensing units and is used to perform noise removal, outlier filtering and smoothing on the collected data.

[0017] Preferably, the micro-scale sensing unit includes an ultrasonic probe with a frequency of 20-100kHz, the meso-scale sensing unit includes a distributed fiber grating sensor with a resolution of 10mm, the macro-scale sensing unit includes a surface wave velocity measurement device, and the overall scale sensing unit includes a resistivity array.

[0018] Preferably, the physical information constrained deep neural network module includes:

[0019] Data-driven path unit, used to process multimodal time series data through a multi-head attention mechanism;

[0020] Physical constraint path unit, used to embed hydration reaction kinetics equation, heat conduction equation and strength development equation;

[0021] The loss function calculation unit is used to comprehensively consider data fitting loss, physical law constraint loss, and consistency loss of different modal predictions.

[0022] Preferably, the loss function used by the loss function calculation unit is:

[0023] L total =α·L data +β·L physics +γ·L consistency ,

[0024] Where: L data is the mean square error loss between the measured data and the predicted value, L physics is the physical law constraint loss, L consistency is the consistency loss of prediction results of different modalities, and α, β, and γ are weight coefficients.

[0025] Preferably, the recipe adaptive transfer learning module includes:

[0026] Basic knowledge layer, used to store pre-trained models based on a large number of standard test blocks;

[0027] Recipe adaptation layer, used to fine-tune parameters for specific concrete mixes;

[0028] Environmental adaptation layer, used for dynamic correction according to environmental conditions;

[0029] The domain adaptation unit is used to align the source and target domain features using the maximum average difference metric.

[0030] Preferably, the Bayesian optimized multi-time scale prediction module includes:

[0031] Very early age prediction unit, used to generate fine-grained predictions at 10-minute intervals within 0-24 hours;

[0032] Early-stage forecast unit, used to generate medium-grained forecasts at 2-hour intervals within 1-7 days;

[0033] Standard age prediction unit, used to generate coarse-grained predictions of 1-day intervals within 7-28 days;

[0034] The uncertainty quantification unit is used to distinguish between aleatory uncertainty and epistemic uncertainty and calculate the confidence interval of the prediction results.

[0035] Preferably, the continuously evolving federated learning feedback module includes:

[0036] Edge node units are deployed at each project site to perform local model training;

[0037] Coordination node unit, used to aggregate model parameters and update the global model;

[0038] Secure channel unit, used to transmit model parameter gradients based on homomorphic encryption technology;

[0039] Incremental learning unit, used to prevent catastrophic forgetting by adopting elastic weight merging algorithm.

[0040] Preferably, the modules are connected via an encrypted two-way communication channel, and data interaction is performed using an asynchronous communication method in a request-response mode, supporting real-time data streaming and batch data transmission.

[0041] Transfer learning optimization method for early-age concrete strength prediction, including:

[0042] Through the multi-scale sensor network, multiple physical parameters inside the concrete are collected to form multi-modal time series data;

[0043] Based on a deep neural network that combines physical law constraints and data-driven, the multimodal time series data is processed to generate a concrete strength prediction model;

[0044] Using transfer learning technology, the model knowledge trained based on standard test blocks is transferred and adapted to actual engineering concrete to generate a specific formula adaptability model;

[0045] Using a Bayesian optimization method, concrete strength evolution curves and confidence intervals at multiple time scales are generated based on the specific formulation adaptability model;

[0046] Through federated learning, the prediction results and measured data of multiple engineering projects are integrated to continuously update the global knowledge base and optimize the concrete strength prediction model.

[0047] Beneficial effects of the present invention

[0048] The beneficial effects of the present invention include: the strength prediction error at very early ages (12-24 hours) is reduced from ±15% of the traditional method to ±5%, and the strength prediction error at standard ages (28 days) is reduced from ±10% to ±3%; the number of concrete specimens used in the test is reduced by 85%, saving material and labor costs; it is applicable to various types of concrete and different environmental conditions; through accurate prediction, the construction period is shortened by an average of 15-20%, improving construction safety; the system performance continues to improve with usage time and data accumulation, and has the ability to self-evolve. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a diagram showing the overall architecture of a transfer learning optimization system for predicting early-age strength of concrete provided by an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of the structure of a multi-scale hydration process data perception module provided by an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of the structure of a deep neural network module with physical information constraints provided by an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of the structure of a recipe-adaptive transfer learning module provided in an embodiment of the present invention;

[0053] Figure 5 A schematic diagram of the structure of a Bayesian optimized multi-time scale prediction module provided by an embodiment of the present invention;

[0054] Figure 6 A schematic diagram of the structure of a continuously evolving federated learning feedback module provided by an embodiment of the present invention;

[0055] Figure 7 This is a flowchart of a transfer learning optimization method for predicting early-age strength of concrete provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] Please refer to the attached Figure 1-7 , the specific implementation of the present invention is further described in detail below with reference to the accompanying drawings.

[0057] Example 1

[0058] like Figure 1 As shown, the transfer learning optimization system for predicting the early-age strength of concrete provided by the present invention includes a multi-scale hydration process data perception module 1, a physical information constrained deep neural network module 2, a formula-adaptive transfer learning module 3, a Bayesian optimized multi-time scale prediction module 4 and a continuously evolving federated learning feedback module 5.

[0059] The Multi-Scale Hydration Process Data Perception Module 1 collects multiple physical parameters within concrete to generate multimodal time-series data. This data comprehensively characterizes changes in key indicators such as temperature, ultrasonic propagation velocity, and resistivity during the early hydration phase of concrete. This module utilizes a layered sensor network to capture key information on the evolution of concrete properties at multiple scales, from microscopic to macroscopic.

[0060] The physical information-constrained deep neural network module 2 is connected to the multi-scale hydration process data perception module 1. It receives multimodal time series data and generates a concrete strength prediction model based on a combination of physical law constraints and data-driven methods. This module uses an innovative dual-pathway network architecture that considers both data patterns and the physical laws of hydration reactions, improving the model's prediction accuracy and generalization capabilities.

[0061] The recipe-adaptive transfer learning module 3 is connected to the physical information-constrained deep neural network module 2 to transfer model knowledge trained on standard test blocks to actual concrete projects, generating a recipe-specific adaptability model. This module, through a three-layer transfer architecture, bridges the knowledge gap from standard testing to engineering applications, significantly reducing the number of samples required for each new project.

[0062] The Bayesian-optimized multi-timescale prediction module 4 is connected to the recipe-adaptive transfer learning module 3 to generate concrete strength evolution curves and confidence intervals based on a specific recipe-adaptive model. This module designs prediction strategies for different timescales and quantifies the uncertainty of the prediction results using a Bayesian framework, providing a reliable basis for construction decision-making.

[0063] The continuously evolving federated learning feedback module 5 is connected to the Bayesian-optimized multi-timescale prediction module 4 to integrate prediction results and measured data from multiple engineering projects, updating the global knowledge base and optimizing the concrete strength prediction model. This module, through a federated learning architecture, enables secure knowledge sharing and continuous model evolution, and system performance continues to improve as the scope of application expands.

[0064] Each module is connected via an encrypted, bidirectional communication channel, using a request-response asynchronous communication model for data exchange. The system workflow primarily encompasses data collection, model training, migration adaptation, strength prediction, and model updates, automating the entire process from data acquisition to decision support.

[0065] Example 2

[0066] like Figure 2 As shown, the multi-scale hydration process data perception module 1 includes a micro-scale sensing unit 11, a meso-scale sensing unit 12, a macro-scale sensing unit 13, an overall scale sensing unit 14 and a data preprocessing unit 15.

[0067] The microscale sensing unit 11 is used to collect data on hydration microstructural changes. This unit primarily consists of an ultrasonic probe operating at a frequency of 20-100 kHz, capable of detecting the evolution of microstructure during concrete hydration. The ultrasonic probe is installed at a predetermined location within the concrete. By transmitting and receiving ultrasonic signals, it measures the propagation characteristics of sound waves in different media, thereby reflecting the microscopic pore structure and degree of hardening.

[0068] Mesoscale sensing unit 12 measures temperature distribution. This unit, primarily composed of distributed fiber Bragg grating (FBG) sensors with a 10mm resolution, enables high-precision, three-dimensional monitoring of the temperature field within concrete. Fiber optic sensors are resistant to electromagnetic interference, corrosion-resistant, and compact, making them suitable for long-term, continuous monitoring buried in concrete. The temperature sensor has an accuracy of ±0.1°C and a sampling rate of 10 minutes.

[0069] The macroscale sensing unit 13 is used to monitor wave velocity changes. This unit primarily consists of a surface wave velocity measurement device. By analyzing the ratio of shear wave to compression wave velocities, it obtains information on the evolution of the concrete's overall stiffness. The wave velocity measurement device includes a transmitter and a receiver. The transmitter generates mechanical waves of a specific frequency, while the receiver detects the signal after the wave propagates. The system determines the wave velocity by calculating the propagation time. The wave velocity sensor has an accuracy of ±0.1 km / s, and the sampling rate is set at 30 minutes.

[0070] The integral scale sensing unit 14 monitors resistivity changes. This unit primarily consists of a resistivity array, which measures ion migration and pore solution evolution to reflect changes in the overall conductivity of the concrete. Resistivity measurements utilize a four-electrode method, where current is applied to an outer pair of electrodes, and the potential difference is measured between the inner pair of electrodes to calculate the resistivity. The resistivity sensor has an accuracy of ±0.1 Ω·m, and the sampling rate is set to 1 hour.

[0071] The data preprocessing unit 15 is connected to each of the aforementioned sensing units and is responsible for denoising, filtering outliers, and smoothing the collected data. This unit comprises three main functional modules: a wavelet-based denoising algorithm module to suppress the influence of environmental noise; an adaptive threshold filtering module to filter out anomalous data points; and a Gaussian smoothing filter module to preserve data trends. The preprocessed multimodal time series data is then fed into a deep neural network module with physical information constraints for further analysis.

[0072] Optimally, the placement of various sensors within the concrete follows the principle of a four-level monitoring network: "point-line-surface-volume," enabling comprehensive monitoring from local to global scale. For large concrete structures, the density of sensor placement can be adjusted based on structural complexity, with the number of sensors appropriately increased in key areas.

[0073] Example 3

[0074] like Figure 3 As shown, the physical information constrained deep neural network module 2 includes a data-driven path unit 21, a physical constraint path unit 22 and a loss function calculation unit 23.

[0075] The data-driven path unit 21 is used to process multimodal time series data using a multi-head attention mechanism. This unit consists of three layers of self-attention modules, each containing eight attention heads, which can effectively capture the spatiotemporal correlations between different sensor signals. The core of the multi-head attention mechanism is to simultaneously learn representations of multiple feature subspaces, calculated using the following formula:

[0076]

[0077] Where: Q is the query matrix, K is the key matrix, V is the value matrix, d kis the dimension of the key vector, and softmax is the normalization function. Multi-head attention repeats the above attention mechanism multiple times and concatenates the results:

[0078] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O ,

[0079] in: and W O is the parameter matrix and h is the number of attention heads.

[0080] The physical constraint path unit 22 is used to embed the hydration reaction kinetics equation, the heat conduction equation, and the strength development equation. This unit is designed based on the principles of physical information neural network (PINN), integrating physical laws as constraints into the neural network structure. The typical hydration heat equation can be expressed as:

[0081]

[0082] Where: ρ is the density of concrete, c is the specific heat capacity, T is the temperature, t is the time, k is the thermal conductivity, Q h (t) is the hydration heat release rate function. The physical constraint path unit approximates the solution of this equation through a specific neural network structure and fuses the result with the output of the data driven path unit.

[0083] The loss function calculation unit 23 is used to comprehensively consider the data fitting loss, physical law constraint loss and different modal prediction consistency loss. The loss function is expressed as:

[0084] L total =α·L data +β·L physics +γ·L consistency ,

[0085] Where: L data is the mean square error loss between the measured data and the predicted value, L physics is the physical law constraint loss, L consistency is the consistency loss of prediction results for different modes, and α, β, and γ are weight coefficients. Experience shows that the optimal value ranges of α, β, and γ are 0.4-0.6, 0.2-0.4, and 0.1-0.3, respectively.

[0086] Preferably, the network is trained using a gradient descent method, with the learning rate initially set to 0.001 and dynamically adjusted using a cosine annealing strategy. To prevent overfitting, the network also incorporates dropout and L2 regularization techniques, with the dropout rate set to 0.2 and the L2 regularization coefficient set to 0.0001.

[0087] This dual-path network architecture, which combines data-driven and physical constraints, not only utilizes the laws contained in the data but also incorporates prior physical knowledge, significantly improving the model's generalization ability in data-sparse areas. The prediction error is reduced by 40% compared to a purely data-driven model.

[0088] Example 4

[0089] like Figure 4 As shown, the recipe adaptation transfer learning module 3 includes a basic knowledge layer 31 , a recipe adaptation layer 32 , an environment adaptation layer 33 and a domain adaptation unit 34 .

[0090] The basic knowledge layer 31 stores pre-trained models based on a large number of standard test blocks. This layer, pre-trained on a large dataset of standard test blocks, incorporates general patterns in concrete strength development. Pre-training utilizes supervised learning, using age as input and strength as output to construct a basic mapping relationship. The basic knowledge layer's network structure consists of a four-layer fully connected neural network, with 64, 128, 64, and 32 nodes per layer, respectively, and uses Reinforced Luminance (ReLU) as the activation function.

[0091] The recipe adaptation layer 32 is used to fine-tune parameters for specific concrete mixes. This layer receives limited sample data for a specific concrete mix and uses transfer learning techniques to fine-tune model parameters to adapt to the strength development characteristics of different mixes. Recipe adaptation primarily focuses on the impact of factors such as cement type, water-cement ratio, coarse-to-fine aggregate ratio, and admixture type and dosage on strength development. The fine-tuning process uses a small learning rate (0.0001), freezing the base layer parameters and updating only the adaptation layer parameters.

[0092] The environmental adaptation layer 33 is used to dynamically adjust predictions based on environmental conditions. This layer considers the impact of environmental factors such as temperature and humidity on concrete strength development and dynamically adjusts predictions through an environmental compensation mechanism. The environmental adaptation layer employs an attention mechanism to automatically adjust the weights of different features based on changes in environmental conditions. In practical applications, this layer can adapt to a wide range of environmental conditions, including temperatures from -10°C to 40°C and humidity from 30% to 95%.

[0093] The domain adaptation unit 34 is used to align the source domain and target domain features using the maximum mean difference (MMD) metric. This unit addresses the distribution difference between the standard test block data (source domain) and the actual engineering concrete (target domain). The MMD metric can be expressed as:

[0094]

[0095] Where: X s and X t are the sample sets of the source domain and the target domain respectively, n s and n tare the number of samples in the source domain and the target domain respectively, φ is the feature mapping function, The domain adaptation unit minimizes the MMD loss to align the feature distribution and improve the performance of the model in the target domain.

[0096] Preferably, the transfer learning process also introduces knowledge distillation technology to compress the knowledge of the complex model (teacher model) into a lightweight model (student model) for easy deployment on edge devices. The knowledge distillation loss function is:

[0097] L KD =α·L CE (,σ(z s ))+(1-α)·τ 2 ·L KL (σ(z t / τ),σ(z s / τ)),

[0098] Where: L CE is the cross entropy loss, L KL is the KL divergence loss, y is the true label, z s and z t are the outputs of the student model and the teacher model respectively, σ is the softmax function, α is the balance parameter, and τ is the temperature parameter.

[0099] Through this three-layer migration architecture and domain adaptation technology, this module achieves knowledge transfer from standard test blocks to actual engineering concrete, allowing the model to adapt to new formulas with only 3-5 engineering specimens, reducing the amount of testing by 85% compared to traditional methods.

[0100] Example 5

[0101] like Figure 5 As shown, the Bayesian optimized multi-time scale prediction module 4 includes a very early age prediction unit 41 , an early age prediction unit 42 , a standard age prediction unit 43 and an uncertainty quantification unit 44 .

[0102] The very early age prediction unit 41 generates fine-grained predictions at 10-minute intervals within the first 24 hours. This unit provides high-frequency predictions for the strength development of concrete within the first 24 hours after placement, a period during which concrete strength changes rapidly and significantly impacts subsequent strength development. This unit utilizes a recurrent neural network (LSTM) architecture, effectively capturing short-term temporal dependencies. The number of LSTM units is set to 64, with the input being the current multimodal sensor data and the output being a sequence of strength predictions for the next 10-minute intervals.

[0103] Early-age prediction unit 42 generates medium-grained predictions at two-hour intervals within 1-7 days. This unit focuses on the early strength development phase of concrete, when the concrete has initially hardened but is still rapidly gaining strength. This unit employs an attention-enhanced temporal convolutional network (TCN) architecture, capable of capturing pattern changes over longer timeframes. The TCN uses expansion factors set to [1, 2, 4, 8], a kernel size of 3, and 32 channels, enabling it to effectively handle multi-scale temporal dependencies.

[0104] The standard age prediction unit 43 generates coarse-grained predictions at daily intervals within 7-28 days. This unit focuses on the concrete standard strength formation stage, during which concrete strength growth gradually slows and stabilizes. This unit uses a feedforward neural network structure to predict long-term strength based primarily on early strength development trends and material properties. The network consists of three layers, with 32, 16, and 8 nodes per layer, respectively. The activation function uses the Sigmoid function, which is suitable for fitting S-shaped strength development curves.

[0105] Uncertainty quantification unit 44 is used to distinguish between aleatory uncertainty and epistemic uncertainty and calculate the confidence interval of the prediction results. This unit is based on the principle of Bayesian neural network and uses Monte Carlo random dropout (MC Dropout) technology to achieve uncertainty estimation. Aleatory uncertainty reflects the randomness of the data itself, while epistemic uncertainty reflects the uncertainty of the model in the unknown area. The confidence interval calculation formula is:

[0106]

[0107] in: is the predicted mean, z α / 2 is the z score with a confidence level of 1-α, For epistemic uncertainty, is the chance uncertainty. These two uncertainties can be estimated by multiple forward propagations and calculating statistics.

[0108] The forecast is preferably updated dynamically based on real-time sensor data. The system uses a sliding window strategy with a 24-hour window size and a one-hour step size, re-forecasting the intensity development for the next period after each sliding window. For critical time points (such as mold removal), the system also activates an adaptive sampling strategy, increasing the prediction frequency to improve accuracy.

[0109] This multi-timescale prediction strategy, combined with Bayesian optimization methods, has achieved continuous prediction of concrete strength throughout its entire life cycle for the first time, and provided reliable uncertainty quantification, enabling key construction decision-making to be accurate to ±1.5 hours, 12-24 hours earlier than existing technologies.

[0110] Example 6

[0111] like Figure 6 As shown, the continuously evolving federated learning feedback module 5 includes an edge node unit 51 , a coordination node unit 52 , a secure channel unit 53 and an incremental learning unit 54 .

[0112] Edge node units 51 are deployed at each project site to perform local model training. These units run on on-site edge computing devices, processing locally collected sensor data and performing model inference. The edge devices utilize a quad-core ARM Cortex-A72 processor, 4GB of LPDDR4 memory, and 32GB of eMMC storage, sufficient to run lightweight models. Local training uses only data collected for this project, maintaining data privacy. Training utilizes mini-batch gradient descent with a batch size of 32 and a number of 5 iterations.

[0113] The coordination node unit 52 is used to aggregate model parameters and update the global model. This unit is deployed on the central server and is responsible for coordinating the parameter updates of multiple edge nodes. The global model aggregation adopts a weighted average method, and the weight is proportional to the amount of data of each node:

[0114]

[0115] Where: W global is the global model parameter, W i is the local model parameter of the i-th node, |D i | is the data volume of the i-th node, |D| is the total data volume of all nodes, and n is the number of nodes. The coordinating node is also responsible for model version management and supports rollback to the best performing version in history.

[0116] The secure channel unit 53 is used to transmit model parameter gradients based on homomorphic encryption technology. This unit ensures model parameter sharing without leaking the original data. Homomorphic encryption allows operations to be performed directly on encrypted data without decryption, greatly improving the security of data sharing. Encryption uses a semi-homomorphic encryption scheme with a 2048-bit key length, which is resistant to cracking attempts under current computing power. The encrypted gradients are transmitted via the HTTPS protocol to ensure communication security.

[0117] The incremental learning unit 54 is used to prevent catastrophic forgetting by using the elastic weight consolidation (EWC) algorithm. This unit solves the problem of old knowledge loss that may occur during the model update process. The EWC algorithm protects important parameters related to old tasks from being over-updated by introducing an important parameter penalty term:

[0118]

[0119] Where: L(θ) is the total loss function, L B(θ) is the loss of the current task, λ is the importance coefficient, F i is the parameter importance measure, The incremental learning unit also implements an experience replay mechanism to retain key historical cases for new model training.

[0120] Preferably, the system also establishes an anomaly detection mechanism by monitoring model performance indicators (such as MAE, RMSE, R 2 The system automatically identifies and isolates abnormal data based on the performance of the node and its changing trends. When a significant decline in model performance is detected, the system triggers an alarm and suspends parameter updates for that node to prevent bad data from contaminating the global model.

[0121] Through the federated learning architecture and incremental learning strategy, this module builds a self-evolving knowledge base, realizes experience sharing among multiple engineering projects, and the system performance continues to improve with the increase of usage time. The average prediction accuracy is 32% higher than that of traditional methods.

[0122] Example 7

[0123] like Figure 7 As shown, the transfer learning optimization method for predicting early age strength of concrete provided by the present invention includes the following steps:

[0124] Step S1: Collect multiple physical parameters inside the concrete through a multi-scale sensor network to form multimodal time series data.

[0125] This step begins by deploying multiple types of sensors within the concrete according to a four-level monitoring network principle: "point-line-surface-volume." These sensors include ultrasonic probes, fiber Bragg grating temperature sensors, surface wave velocity measurement devices, and resistivity measurement arrays. Each sensor collects raw data at a preset frequency: microscale sensors (ultrasonic) every 30 minutes, mesoscale sensors (temperature) every 10 minutes, macroscale sensors (wave velocity) every 30 minutes, and bulk sensors (resistivity) every hour. Edge computing devices receive and store the raw data, performing preliminary noise filtering and anomaly detection. The processed data forms a multimodal time series feature vector, containing information such as temperature field gradient, temperature change rate, ultrasonic velocity, ultrasonic amplitude, and resistivity variation pattern.

[0126] Step S2: Based on a deep neural network that combines physical law constraints and data-driven, multimodal time series data is processed to generate a concrete strength prediction model.

[0127] This step first constructs a dual-path network architecture, consisting of a data-driven path and a physical constraint path. The data-driven path processes multimodal time series data through a multi-head attention mechanism, capturing the spatiotemporal correlations between different sensor signals. The physical constraint path embeds the hydration reaction kinetics equation, the heat conduction equation, and the strength development equation to ensure that the model conforms to physical laws. A comprehensive loss function is designed, simultaneously considering data fitting loss, physical law constraint loss, and consistency loss of predictions across different modalities. The network is trained using gradient descent, with the learning rate initially set to 0.001 and dynamically adjusted using a cosine annealing strategy. Dr-opout and L2 regularization techniques are used to prevent overfitting, with a dropout rate of 0.2 and an L2 regularization coefficient of 0.0001.

[0128] Step S3: Using transfer learning technology, the model knowledge trained based on standard test blocks is transferred and adapted to actual engineering concrete to generate a specific formula adaptability model.

[0129] This step first pre-trains the basic model on a large-scale standard test block dataset to form a basic knowledge layer. Receive limited sample data (usually 3-5) of concrete with a specific formula, and fine-tune the model parameters through transfer learning technology to adapt to the strength development characteristics of different formulas. Considering the impact of environmental factors such as temperature and humidity on the strength development of concrete, the prediction results are dynamically adjusted through the environmental compensation mechanism. The maximum mean difference (MMD) measurement method is used to achieve feature alignment between the source domain (standard test block) and the target domain (actual engineering concrete). The knowledge distillation technology is introduced to compress the knowledge of the complex model (teacher model) into a lightweight model (student model) for easy deployment on edge devices.

[0130] Step S4: Using the Bayesian optimization method, the concrete strength evolution curves and confidence intervals at multiple time scales are generated based on the specific formulation adaptability model.

[0131] This step designs corresponding forecasting strategies for different time scales: For the very early age (0-24 hours), fine-grained forecasts with 10-minute intervals are used, using an LSTM network structure to capture short-term temporal dependencies; for the early age (1-7 days), medium-grained forecasts with 2-hour intervals are used, using an attention-enhanced TCN structure to capture medium-term pattern changes; and for the standard age (7-28 days), coarse-grained forecasts with 1-day intervals are used, using a feedforward neural network structure to predict long-term intensity trends. Based on the principles of Bayesian neural networks, uncertainty estimation is achieved through the Monte Carlo random dropout technique, distinguishing between aleatory and epistemic uncertainty, and calculating confidence intervals for the forecast results. Forecast results are dynamically updated based on real-time sensor data, using a sliding window strategy with a 24-hour window size and a 1-hour step size. After each sliding window, the intensity development for the future period is re-forecasted.

[0132] Step S5: Through federated learning, the prediction results and measured data of multiple engineering projects are integrated to continuously update the global knowledge base and optimize the concrete strength prediction model.

[0133] This step performs local model training on edge devices at each engineering site, using only local data to maintain data privacy. Homomorphic encryption technology is used to securely transmit model parameter gradients, ensuring that model parameters are shared without leaking the original data. Parameter updates from multiple edge nodes are aggregated on a central server to update the global model. The elastic weight consolidation (EWC) algorithm is used to prevent catastrophic forgetting and protect important parameters related to old tasks from being over-updated. An experience replay mechanism is implemented to retain key historical cases for new model training. An anomaly detection mechanism is established to monitor model performance indicators and their changing trends, and automatically identify and isolate abnormal data.

[0134] Example 8

[0135] This example describes the application of this system in the construction of a spillway at a large hydropower station. This project uses C40 concrete, which has strict requirements for early strength development and requires precise control of formwork removal time to ensure construction progress and quality.

[0136] System Deployment: A network of 12 multi-scale sensors was deployed in the spillway concrete pouring area, covering key structural areas. Four edge computing devices were deployed on-site and connected to a central management platform. The system was initially trained using 200 sets of standard test block data and adapted to the on-site concrete mix through transfer learning.

[0137] Application effect: The system's strength prediction error at very early ages (12-24h) is ±4.2%, significantly better than the traditional method's ±14.8%; the key construction time point (demolding time) is predicted accurately to ±1.3 hours, providing a decision-making basis 18 hours earlier than the traditional method; by accurately predicting the demolding time, the construction period was shortened by 16.5%, saving a lot of manpower and material costs; as the system is used for a longer time, the prediction accuracy continues to improve. After 3 months of use, the prediction error was further reduced to ±3.6%.

[0138] The system also analyzed variations in concrete strength development across different areas, identifying and promptly adjusting local curing conditions to ensure consistent overall structural performance. Upon project completion, the knowledge accumulated within the system was securely integrated into the global knowledge base, providing valuable experience for subsequent similar projects.

[0139] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A transfer learning optimization system for predicting early-age strength of concrete, characterized by: include: Multi-scale hydration process data perception module, used to collect multiple physical parameters inside concrete to form multi-modal time series data; A physical information constrained deep neural network module is connected to the multi-scale hydration process data perception module, and is used to receive the multimodal time series data and generate a concrete strength prediction model based on a combination of physical law constraints and data-driven methods; A transfer learning module for formula adaptation is connected to the deep neural network module for physical information constraints, and is used to transfer the model knowledge trained based on standard test blocks to actual engineering concrete to generate a specific formula adaptability model; A Bayesian optimized multi-time scale prediction module, connected to the formula adaptive transfer learning module, is used to generate a concrete strength evolution curve and confidence interval based on the specific formula adaptability model; The continuously evolving federated learning feedback module is connected to the Bayesian optimized multi-time scale prediction module to integrate the prediction results and measured data of multiple engineering projects, update the global knowledge base and optimize the concrete strength prediction model.

2. The system according to claim 1, wherein: The multi-scale hydration process data perception module includes: Micro-scale sensing unit, used to collect data on changes in hydration microstructure; Mesoscale sensing unit, used to measure temperature field distribution; Macro-scale sensing units to monitor wave velocity changes; integral scale sensing unit for monitoring resistivity changes; The data preprocessing unit is connected to the above-mentioned sensing units and is used to perform noise removal, outlier filtering and smoothing on the collected data.

3. The system according to claim 2, characterized in that The micro-scale sensing unit includes an ultrasonic probe with a frequency of 20-100kHz, the meso-scale sensing unit includes a distributed fiber grating sensor with a resolution of 10mm, the macro-scale sensing unit includes a surface wave velocity measurement device, and the overall scale sensing unit includes a resistivity array.

4. The system according to claim 1, wherein: The physical information constrained deep neural network module includes: Data-driven path unit, used to process multimodal time series data through a multi-head attention mechanism; Physical constraint path unit, used to embed hydration reaction kinetics equation, heat conduction equation and strength development equation; The loss function calculation unit is used to comprehensively consider data fitting loss, physical law constraint loss, and consistency loss of different modal predictions.

5. The system according to claim 4, characterized in that The loss function used by the loss function calculation unit is: L total =α·L data +β·L physics +γ·L consistency , Where: L data is the mean square error loss between the measured data and the predicted value, L physics is the physical law constraint loss, L consistency is the consistency loss of prediction results of different modalities, and α, β, and γ are weight coefficients.

6. The system according to claim 1, wherein: The recipe adaptive transfer learning module includes: Basic knowledge layer, used to store pre-trained models based on a large number of standard test blocks; Recipe adaptation layer, used to fine-tune parameters for specific concrete mixes; Environmental adaptation layer, used for dynamic correction according to environmental conditions; The domain adaptation unit is used to align the source and target domain features using the maximum average difference metric.

7. The system according to claim 1, wherein: The Bayesian optimized multi-time scale prediction module includes: Very early age prediction unit, used to generate fine-grained predictions at 10-minute intervals within 0-24 hours; Early-stage forecast unit, used to generate medium-grained forecasts at 2-hour intervals within 1-7 days; Standard age prediction unit, used to generate coarse-grained predictions of 1-day intervals within 7-28 days; The uncertainty quantification unit is used to distinguish between aleatory uncertainty and epistemic uncertainty and calculate the confidence interval of the prediction results.

8. The system according to claim 1, wherein: The continuously evolving federated learning feedback module includes: Edge node units are deployed at each project site to perform local model training; Coordination node unit, used to aggregate model parameters and update the global model; Secure channel unit, used to transmit model parameter gradients based on homomorphic encryption technology; Incremental learning unit, used to prevent catastrophic forgetting by adopting elastic weight merging algorithm.

9. The system according to claim 1, wherein: The modules are connected through encrypted two-way communication channels, and use request-response asynchronous communication to exchange data, supporting real-time data streaming and batch data transmission.

10. A transfer learning optimization method for predicting early-age strength of concrete, using the system according to any one of claims 1 to 9, characterized in that: include: Through the multi-scale sensor network, multiple physical parameters inside the concrete are collected to form multi-modal time series data; Based on a deep neural network that combines physical law constraints and data-driven, the multimodal time series data is processed to generate a concrete strength prediction model; Using transfer learning technology, the model knowledge trained based on standard test blocks is transferred and adapted to actual engineering concrete to generate a specific formula adaptability model; Using a Bayesian optimization method, concrete strength evolution curves and confidence intervals at multiple time scales are generated based on the specific formulation adaptability model; Through federated learning, the prediction results and measured data of multiple engineering projects are integrated to continuously update the global knowledge base and optimize the concrete strength prediction model.

Images