Artificial intelligence based method for predicting concrete performance

Abstract

The application discloses a kind of based on artificial intelligence's prediction method of concrete performance.The prediction method drives sensor to collect temperature and dual-frequency resistivity background data with initial low frequency;Real-time calculation window temperature evolution slope, when reaching early warning threshold, control sampling frequency switch to high frequency mode;High-frequency data is derived to determine to lock hydration acceleration period starting anchor point and pore connectivity cut-off anchor point, and executes sampling hardware reset;Calculate the time delay difference between anchor points and perform tolerance comparison, when exceeding threshold, use spatial distance inverse square weighting algorithm to perform interpolation reconstruction to eliminate aggregate distortion;Integral and difference operation is carried out in phase change interval, and one-dimensional dynamic characteristic tensor containing multidimensional dynamic index is constructed;Characteristic tensor is normalized and input into multilayer perception machine model for forward inference, and the final mechanical property prediction value of concrete is output after inverse normalization.

Description

Technical Field

[0001] This invention relates to the field of concrete performance testing and artificial neural network application technology, specifically to an artificial intelligence-based method for predicting concrete performance. Background Technology

[0002] As the most widely used structural material in modern engineering construction, concrete's mechanical properties and durability directly determine the quality and safety of building projects. In concrete engineering, early and accurate prediction of physical properties such as concrete compressive strength not only provides data support for formwork removal, prestressing tensioning, and subsequent procedures during construction, but also enables dynamic optimization and quality pre-control of concrete mix design.

[0003] Traditional concrete performance prediction mainly relies on pressure tests of standard cured specimens or uses purely data-driven modeling methods, taking static mix proportion data as input to fit the later mechanical strength. In recent years, with the development of sensor technology, the industry has begun to use physical sensors embedded in concrete to synchronously collect time-series parameters such as temperature field or resistivity during the early hydration process of concrete. Microprocessors are then used to analyze the time-series monitoring data to reflect the internal microscopic evolution process, thereby establishing a neural network mapping relationship between multi-physics characteristics and the later mechanical properties of concrete.

[0004] However, in existing technologies for multiphysics data acquisition and performance prediction using embedded sensors, the system typically employs a constant, fixed sampling frequency, which is difficult to adapt to the differences in kinetic rates at different stages of concrete hydration. During the slow phase of hydration, fixed high-frequency sampling leads to increased power consumption of the underlying hardware and ineffective use of storage resources; if the sampling frequency is uniformly lowered, there is a lack of sufficient temporal resolution during the active hydration heat release phase, resulting in limited accuracy in extracting key phase transition features. Furthermore, the heterogeneity of concrete aggregate distribution makes single-point sensors susceptible to local signal interference, making it difficult for measurement data to accurately represent the evolution of the overall physical field; and existing prediction models mostly rely on static mix proportions as the main input, making it difficult to adaptively adjust based on the dynamic fluctuations of actual material activity, thus limiting the model's generalization and prediction capabilities in complex construction environments. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide an artificial intelligence-based method for predicting concrete performance, which addresses the issues of data redundancy and feature loss caused by using a fixed sampling frequency in existing early performance monitoring of concrete, as well as the poor generalization ability of static mix proportion prediction models.

[0006] This invention provides a method for predicting concrete performance based on artificial intelligence, comprising the following steps:

[0007] Step S100: Collect multi-physics field measurement data inside the concrete to be tested at an initial low-frequency sampling frequency. The multi-physics field measurement data includes discrete temperature field signals and discrete high and low frequency resistivity signals. Store the multi-physics field measurement data as a low-frequency background data time-series array in a time sequence into a circular data buffer.

[0008] Step S200: Perform evolution slope calculation on the multiphysics field measurement data within the sliding observation window, and when the absolute value of the first-order evolution slope reaches the warning slope threshold, control the sampling frequency to switch to a high-frequency sampling frequency and collect data to obtain high-frequency sequence data and write it into the circular data buffer.

[0009] Step S300: Differentiate the high-frequency sequence data, extract the extreme value anchor point of hydration heat release rate and the abrupt change anchor point of resistivity ratio, and simultaneously trigger sampling frequency reset to restore the sampling frequency to the initial low-frequency sampling frequency.

[0010] Step S400: Extract data slices within the phase transition interval, extract physical feature parameters, and construct a one-dimensional dynamic feature tensor;

[0011] Step S500: Receive the normalized one-dimensional dynamic feature tensor and input it into the pre-deployed multilayer perceptron model for forward inference, output the predicted value of concrete physical properties, and perform inverse normalization restoration operation to obtain the final predicted value of physical properties with actual engineering dimensions.

[0012] Furthermore, the evolution slope is calculated by calling the discrete temperature field signal within the sliding observation window, using the discrete difference formula to obtain the physical temperature difference between the two ends of the sliding observation window, dividing this physical temperature difference by the product of the total number of sampling points within the window and the low-frequency sampling period, and taking the average evolution slope of multiple consecutive sampling points as the comprehensive judgment basis to control the hardware switching of the underlying sampling period. After entering the high-frequency sampling stage, the high-frequency temperature sequence is subjected to moving average filtering for noise reduction and differential operation to obtain the second derivative sequence. Discrete data points with zero second derivative and positive first derivative are located as extreme value anchor points of hydration heat release rate. The ratio parameter of high-frequency resistivity to low-frequency resistivity at the same absolute timestamp is calculated simultaneously. Forward difference operation is performed on the ratio parameter sequence to obtain the first-order evolution slope. Discrete sampling points where the first-order evolution slope of the ratio reaches the step judgment threshold are extracted as anchor points of resistivity ratio change.

[0013] Furthermore, to identify and eliminate distortion in physical sensor data caused by uneven local aggregate distribution, extreme anchor points of hydration heat release rate and abrupt changes in resistivity ratio are extracted at the same spatial monitoring node, and the time delay difference parameter between them is calculated. This time delay difference parameter is then compared with a preset tolerance threshold. When the time delay difference parameter is greater than or equal to the preset tolerance threshold, the underlying data sequence is determined to be physically distorted due to local aggregate interference. At this point, the three-dimensional physical coordinates and effective phase transition anchor point timestamps of adjacent, undistorted spatial monitoring nodes are read. Coordinate interpolation weights are calculated using a spatial distance inverse square weighting algorithm to reconstruct the new extreme anchor points of hydration heat release rate and new abrupt changes in resistivity ratio at the current monitoring node, replacing the original distorted data.

[0014] Furthermore, in the characteristic parameter construction stage, the extreme anchor point of the reconstructed new hydration heat release rate and the abrupt change anchor point of the new resistivity ratio are used as the time boundary conditions for calculus operations, and the time delay difference parameter is recalculated. Data slices within the time boundary conditions are extracted, and discrete integration is performed using the trapezoidal rule numerical integration algorithm to obtain the temperature-time integral features. The low-frequency resistivity measurement values ​​at both ends of the time boundary conditions are extracted to calculate the difference, and this difference is divided by the time delay difference parameter to obtain the average slope of resistivity change. The time delay difference parameter, temperature-time integral features, and average slope of resistivity change are physically concatenated in memory into a one-dimensional dynamic feature tensor. Finally, this feature tensor is loaded into the multilayer perceptron model to perform forward matrix multiplication and activation function mapping, outputting continuous real values ​​representing mechanical properties. Inverse normalization is then performed using the maximum and minimum values ​​of performance index labels in the historical training set to output the final predicted values ​​of concrete mechanical properties.

[0015] The present invention, by adopting the above technical solution, can bring the following beneficial effects:

[0016] 1. This invention, by predicting the slope of the physical field evolution and dynamically switching the sampling frequency of the analog-to-digital converter in the frequency conversion control module, enables the system to maintain low-frequency acquisition during the hydration plateau period and automatically switch to high-frequency mode during the active period. This ensures accurate extraction of the starting anchor point and pore connectivity cutoff anchor point during the critical reaction stage, while eliminating the accumulation of invalid redundant data generated by high-frequency sampling throughout the entire cycle. This achieves the effect of reducing the power consumption of the underlying hardware and the consumption of internal storage resources while ensuring the integrity of the acquisition of microscopic feature data.

[0017] 2. This invention diagnoses the response distortion of a single-point sensor caused by local aggregate heterogeneity interference from the perspective of the asynchronous nature of physical field evolution by calculating the time delay difference between the temperature field and the electrochemical phase transition and comparing it with the judgment threshold. When the distortion is judged, the system calls the coordinates of the surrounding effective nodes and the distance inverse weighted algorithm to perform spatial interpolation replacement, which effectively enhances the system anti-interference capability and reliability of multi-physics feature data sources.

[0018] 3. This invention extracts data slices within the phase transition boundary conditions, discretizes the temperature-time integral and the average resistivity slope, and uses dynamic physical parameters reflecting the evolution of microstructure as the core input tensor of the network. This expands the input dimension of the prediction model from purely static mix proportions to dynamic evolution features, reduces the model's single dependence on static prior information, and improves the predictive model's adaptive generalization ability to material activity fluctuations in actual construction environments. Attached Figure Description

[0019] Figure 1 This is a flowchart of the concrete performance prediction method based on artificial intelligence according to the present invention.

[0020] Figure 2 This is a diagram of the concrete performance prediction system based on multi-physics feature mapping according to the present invention.

[0021] Figure 3 This is a flowchart of the frequency conversion triggering mechanism based on the prediction of the slope of physical field evolution in this invention;

[0022] Figure 4 This is a flowchart of the abnormal physical distortion identification and spatial interpolation reconstruction process of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please refer to the appendix. Figure 1-4 This invention provides a method for predicting concrete performance based on artificial intelligence. In specific implementation, this method uses a concrete performance prediction system based on multi-physics feature mapping to predict concrete performance. The system includes:

[0025] The data sensing layer 10, signal conditioning layer 20, and core computing layer 30 are connected by cables and a data bus to form a closed-loop control and data flow topology.

[0026] The data sensing layer 10 is arranged inside the concrete to be tested, including a distributed and embedded temperature sensor 11 and a dual-frequency AC impedance sensor 12. The temperature sensor 11 adopts a resistance temperature detector or a thermocouple to obtain the time-series change signal of the local temperature field. The dual-frequency AC impedance sensor 12 includes a pre-embedded stainless steel electrode array.

[0027] The signal conditioning layer 20 is communicatively connected to the data sensing layer 10. The signal conditioning layer 20 includes an analog-to-digital converter 21 and a programmable amplifier 22. The signal conditioning layer 20 is configured to receive the analog signals output by the temperature sensor 11 and the dual-frequency AC impedance sensor 12, perform filtering and noise reduction and digital conversion, and transmit the converted discrete digital signals to the core computing layer 30 through a serial communication interface. In order to adapt to the strong electromagnetic interference environment of the concrete construction site, the above-mentioned serial communication interface specifically adopts RS485 industrial bus or step SPI hardware interface.

[0028] The core computing layer 30 includes a microprocessor 31 and a memory 32. The microprocessor 31 adopts an industrial-grade microcontroller or a field-programmable gate array. The memory 32 stores computer program instructions. When the microprocessor 31 executes the computer program instructions, it realizes the frequency conversion control module 101, the phase change anchor point extraction module 102, the feature tensor reconstruction module 103 and the neural network prediction module 104.

[0029] The frequency conversion control module 101 is connected to the signal conditioning layer 20 through a programmable timer. The frequency conversion control module 101 is configured to adjust the sampling period interval of the analog-to-digital converter 21 according to the evolution slope of the received data. The frequency conversion control module 101 realizes the switching of the hardware sampling frequency between low-frequency mode and high-frequency mode by writing the frequency division coefficient value to the programmable timer.

[0030] The phase change anchor point extraction module 102 communicates with the data buffer of the frequency conversion control module 101. The phase change anchor point extraction module 102 has a built-in calculus operation logic unit. By performing second-order difference operation on the data in the high-frequency sampling interval, it extracts the time nodes when physical parameters undergo extreme value changes or step jumps. The phase change anchor point extraction module 102 marks the extracted time nodes as phase change anchor point timestamps, which are used to calibrate the absolute time position of the early hydration reaction kinetic state switching.

[0031] The feature tensor reconstruction module 103 receives the timestamp calibration information output by the phase transition anchor point extraction module 102. The feature tensor reconstruction module 103 extracts multiphysics data slices between adjacent phase transition anchor point timestamps, performs definite integral and slope calculation operations, and reduces the dimensionality of the continuous slice data in the time dimension into a one-dimensional structured array containing accumulated energy, rate of change, and time span to construct a dynamic feature tensor. ;

[0032] The neural network prediction module 104 is connected to the output of the feature tensor reconstruction module 103. The neural network prediction module 104 is equipped with a pre-trained artificial neural network model. As a specific implementation of the artificial neural network, this embodiment adopts a multilayer perceptron architecture, including at least one input layer, two hidden layers, and one output layer. The neural network prediction module 104 converts the dynamic feature tensor... The input node signals of the multilayer perceptron are used to perform forward matrix multiplication and activation function mapping operations to calculate the predicted values ​​characterizing the macroscopic physical properties of concrete. :

[0033] In the formula, This refers to a nonlinear mapping function within a multilayer perceptron, consisting of nested activation functions. The weight matrix obtained during network training. This is the bias vector.

[0034] Furthermore, relying on the aforementioned concrete performance prediction system based on multiphysics feature mapping, the entire process from acquiring the original signal to performance prediction is completed through information flow and instruction interaction between software and hardware. The specific execution flow of the above-mentioned artificial intelligence-based concrete performance prediction method includes the following steps:

[0035] In step S100, the core computing layer 30 initializes the system control parameters, drives the signal conditioning layer 20 to collect multi-physics field measurement data at an initial low-frequency cycle, and the microprocessor 31 writes a preset low-frequency division constant into the programmable timer in the signal conditioning layer 20 to set the initial low-frequency sampling frequency. The temperature sensor 11 and the dual-frequency AC impedance sensor 12 embedded in the data sensing layer 10 synchronously collect discrete temperature field signals and discrete high and low frequency resistivity signals inside the concrete at the low-frequency sampling frequency. The collected multi-physics field measurement data is discretized by the analog-to-digital converter and stored in the circulating data buffer area inside the microprocessor 31 in time sequence.

[0036] In step S200, the frequency conversion control module 101 performs evolution slope calculation on the data within the sliding time window, and controls the signal conditioning layer 20 to switch to high-frequency sampling state based on the calculation result. The system sets the time width as the sliding observation window within the cyclic data buffer. The frequency conversion control module 101 calls the discrete temperature field signal within the sliding observation window and calculates the first-order evolution slope using the following discrete difference formula:

[0037] In the formula, For the end of the sliding window The average evolution slope calculated at each sampling point For the first Temperature measurement values ​​at each sampling point This represents the temperature measurement value corresponding to the starting point of the sliding window. This represents the total number of sampling points within the sliding window. As the initial low-frequency sampling period, when the absolute value of the first-order evolution slope reaches the warning slope threshold, the frequency conversion control module 101 writes a high-frequency division constant to the programmable timer, switching the sampling frequency from the initial low frequency to a high frequency, so as to record the microscopic physical field change signal with high time resolution.

[0038] In step S300, the phase transition anchor point extraction module 102 performs derivative determination on the high-frequency sequence data, locks the phase transition anchor point timestamp, and triggers the frequency conversion control module 101 to execute sampling frequency reset through the state reset flag. The phase transition anchor point extraction module 102 performs moving average filtering and noise reduction on the high-frequency temperature sequence, and then calculates the high-resolution second derivative sequence. The phase transition anchor point extraction module 102 traverses the second derivative sequence, locates discrete data points that satisfy the condition that the second derivative is zero and the first derivative is positive, and extracts the located discrete data points. The time attribute corresponding to the data point is calibrated as the starting anchor point of the hydration acceleration period. The phase change anchor point extraction module 102 synchronously calculates the ratio data of the high and low frequency resistivity sequences, extracts the extreme points of the step characteristics in the first derivative sequence of the ratio data, and calibrates them as the pore connectivity cutoff anchor points. After completing the timestamp calibration of the above two phase change anchor points, the phase change anchor point extraction module 102 sends a status confirmation interrupt signal to the frequency conversion control module 101. The frequency conversion control module 101 resets the programmable timer and restores the sampling frequency to the initial low frequency period.

[0039] In step S400, the feature tensor reconstruction module 103 extracts data slices within the phase transition interval, extracts physical feature parameters, and constructs a dynamic feature tensor. To identify the physical field response distortion caused by local aggregate interference, the feature tensor reconstruction module 103 calculates the time delay difference parameter between the timestamps of the two phase transition anchor points with different physical states extracted above. The calculation formula is as follows:

[0040] In the formula, The timestamp value of the anchor point that cuts off the pore connectivity. The feature tensor reconstruction module 103 uses the time delay difference parameter to represent the timestamp value of the starting anchor point of the hydration acceleration period. Combining the interval temperature-time integral feature score with the average slope value of resistivity change, these values ​​are sequentially written into a continuous memory space and concatenated into a one-dimensional dynamic feature tensor. To eliminate the numerical weight allocation offset introduced by different physical dimensions, the feature tensor reconstruction module 103 uses the following maximum-minimum normalization formula to perform linear mapping on the elements within the one-dimensional dynamic feature tensor:

[0041] In the formula, These are the original values ​​of the physical characteristic parameters before normalization. and These are the minimum and maximum values ​​set for the corresponding physical characteristic parameters in the historical training set, respectively. This represents the normalized value after mapping and compression to the 0-1 range;

[0042] In step S500, the neural network prediction module 104 receives the normalized one-dimensional dynamic feature tensor, calculates and outputs the predicted values ​​of concrete performance. The neural network prediction module 104 loads the one-dimensional dynamic feature tensor into the input layer node of the pre-deployed multilayer perceptron model. The data undergoes forward matrix multiplication and nonlinear activation function mapping between the hidden layer nodes of the multilayer perceptron to complete the spatial feature transformation. The output layer node of the multilayer perceptron performs weighted summation based on the feature vector passed to the end, and outputs continuous real values ​​representing the mechanical properties of concrete.

[0043] Furthermore, in a specific embodiment of the present invention, the specific configuration of physical sensing nodes and the underlying data flow logic for the execution process of the aforementioned multiphysics sensing system initialization and low-frequency environment monitoring are detailed below:

[0044] Step S110: Before the concrete component to be tested is poured into the formwork, the physical space deployment of the sensing monitoring nodes is performed. The temperature sensor 11 and the dual-frequency AC impedance sensor 12 are arranged at the set monitoring anchor points inside the structure. In order to prevent the exposed electrodes of the dual-frequency AC impedance sensor 12 from making direct electrical contact with the metal steel reinforcement skeleton inside the building and causing a short circuit in the test circuit, the operator uses polymer insulating cable ties to fix the dual-frequency AC impedance sensor 12 to the steel reinforcement mesh node. The insulating fixing operation cuts off the physical conduction path of external stray current and ensures the independence of the physical field boundary of the acquisition area of ​​the dual-frequency AC impedance sensor 12.

[0045] In step S120, the core computing layer 30 is powered on and performs parameter initialization configuration of the underlying hardware peripherals. The microprocessor 31 sends the physical range and operating parameters of the excitation signal generator to the signal conditioning layer 20. For the dual-frequency AC impedance sensor 12, the microprocessor 31 configures two sets of sinusoidal excitation signals of different frequency bands. The two sets of sinusoidal excitation signals of different frequency bands include a low-frequency excitation frequency and a high-frequency excitation frequency. The value range of the low-frequency excitation frequency is set to 1kHz to 10kHz. The technical purpose of setting the low-frequency excitation frequency is to characterize the long-range migration conductivity of free ions in the pore solution inside the concrete. The value range of the high-frequency excitation frequency is set to 1MHz to 10MHz. The technical purpose of setting the high-frequency excitation frequency is to penetrate the double layer and characterize the polarization capacitance characteristics of the early hydration product generation phase interface.

[0046] Step S130, microprocessor 31 configures the initial low-frequency sampling period. And construct the corresponding data storage structure. Based on the initial induction period of silicate cement material after mixing with water, the hydration reaction rate of the microstructure in the initial induction period is slow. The microprocessor 31 locks the initial sampling state of the frequency conversion control module 101 to low frequency mode to reduce the redundant acquisition of invalid data. Simultaneously, the microprocessor 31 divides a fixed-length continuous physical address range in the internal static random access memory and instantiates the circular data cache area.

[0047] In step S140, the prediction system continues to synchronously acquire and store background environmental data according to the initial low-frequency mode, and the programmable timer in the signal conditioning layer 20 operates according to the low-frequency sampling period. At regular intervals, a hardware analog-to-digital conversion interrupt is periodically triggered. The microprocessor 31 receives the interrupt request and drives the multiplexer to sequentially select the analog signal channels of the temperature sensor 11 and the dual-frequency AC impedance sensor 12, performing digital sampling. The microprocessor 31 packages the converted discrete temperature measurement values ​​and discrete high and low frequency resistivity measurement values, adds the current system absolute timestamp, and constructs a single sampling data packet. :

[0048] In the formula, For the first Discrete data packets sampled subsampled. For the first The absolute timestamp of the next sample. For the first Temperature measurement value from the second sampling, For the first High-frequency resistivity measurement values ​​from the second sampling. For the first The microprocessor 31 processes the low-frequency resistivity measurement value from the second sample into a single-sample data packet. Write to the tail address of the circular data buffer; the prediction system uses consecutive single-sample data packets. A low-frequency background data time series array was established to support the calculation of the evolution slope.

[0049] Furthermore, after the low-frequency data array of S140 is established, frequency conversion judgment needs to be introduced. In a specific embodiment of the present invention, the specific logic judgment and underlying frequency conversion execution steps for the control process of the aforementioned prediction system switching from low-frequency monitoring state to high-frequency sampling state are detailed as follows:

[0050] In step S210, the frequency converter control module 101 sets a sliding observation window with a fixed time width within the cyclic data buffer of the microprocessor 31. The sliding observation window contains the total number of sampling points. To ensure that high-frequency hardware noise is filtered out without missing the actual physical evolution, the total number of sampling points is... According to the thermal response time constant setting of the temperature sensor 11, as the data sensing layer 10 continuously samples, the sliding observation window moves synchronously backward on the time axis according to the first-in-first-out principle. The frequency conversion control module 101 periodically extracts the temperature measurement value and resistivity measurement value sequence within the sliding observation window according to the set operation rhythm, and loads the extracted sequence data into the high-speed register of the microprocessor 31 to perform differential operation.

[0051] In step S220, the frequency conversion control module 101 calls the sequence data in the high-speed register and calculates the comprehensive evolution slope at the sampling point at the end of the sliding observation window as the comprehensive judgment basis. The frequency conversion control module 101 calculates the comprehensive evolution slope at the sampling point at the end of the sliding observation window according to the discrete difference formula. The calculation formula is as follows:

[0052] ;

[0053] In step S230, the frequency converter control module 101 executes the slope threshold comparison and phase change prediction control logic, and the frequency converter control module 101 integrates the evolution slope. The absolute value and the system's preset warning slope threshold Compare and determine the warning slope threshold. Characterizing the boundary of the rate of change of physical quantities within concrete during the transition from the initial induction period to the accelerated hydration period, and the warning slope threshold. The data was obtained from the historical adiabatic temperature rise test data of the benchmark specimen with the same mix proportion of concrete, and the comprehensive evolution slope was calculated. The absolute value is less than the warning slope threshold At this time, the frequency conversion control module 101 maintains the current control state, and the signal conditioning layer 20 maintains the initial low-frequency sampling mode. When the comprehensive evolution slope... The absolute value is greater than or equal to the warning slope threshold. At that time, the frequency converter control module 101 determines that the hydration heat release rate has increased sharply and triggers a frequency converter interrupt request;

[0054] In step S240, the microprocessor 31 responds to the frequency conversion interrupt request and sends a high-frequency sampling instruction to the signal conditioning layer 20 via the data bus to complete the switching of the underlying sampling period. The microprocessor 31 locates the physical address of the programmable timer that controls the working cycle of the analog-to-digital converter 21, and overwrites the low-frequency division constant in the internal control register of the programmable timer with the high-frequency division constant. The sampling period of the analog-to-digital converter 21 is changed from the initial low-frequency sampling period. Change to high-frequency sampling period Signal conditioning layer 20 during high-frequency sampling period Driven by this, the changes in the microscopic physical field during the hydration acceleration stage are recorded with high temporal resolution, and the acquired high-frequency sequence data is written into the circular data buffer.

[0055] Furthermore, in a specific embodiment of the present invention, the specific logic and execution steps of the digital filtering and extraction algorithm executed by the phase transition anchor point extraction module 102 for the processing of multi-physics field data under high-frequency sampling state of the aforementioned prediction system are detailed below:

[0056] In step S310, the phase change anchor extraction module 102 calls the high-frequency temperature sequence data and high-frequency resistivity sequence data from the internal loop data buffer of the microprocessor 31 and performs digital smoothing filtering. To eliminate hardware white noise interference introduced by the analog-to-digital converter in high-frequency sampling mode and strictly meet the causality requirements of the real-time monitoring system, the phase change anchor extraction module 102 performs causal moving average filtering on the high-frequency sequence data. In specific implementation, the filtering window width is set to... The phase transition anchor point extraction module 102 iterates through the original high-frequency sequence data according to the time series, and outputs the filtered smooth data sequence. The moving average filtering calculation formula for any physical quantity is as follows:

[0057] ;

[0058] In step S320, the phase transition anchor point extraction module 102 calculates the first-order temperature derivative at the corresponding sampling point using the first-order forward difference method on the smoothed high-frequency temperature sequence. The formula for calculating the first-order temperature derivative is as follows:

[0059] In the formula, For the first The first derivative of temperature at each high-frequency sampling point To predict the discrete sampling period when the system is in high-frequency sampling mode, the phase change anchor point extraction module 102, to ensure the causality of real-time monitoring, simultaneously uses the second-order backward difference approximation method to solve the temperature second-order derivative sequence at the discrete sampling points. The approximate calculation formula for the second-order derivative is as follows:

[0060] In the formula, For the first The second derivative of temperature at each high-frequency sampling point , and These are the smoothed temperature data at the current and previous two high-frequency sampling points, respectively;

[0061] In step S330, the phase transition anchor point extraction module 102 traverses the constructed temperature first derivative sequence and temperature second derivative sequence, determines and extracts the starting anchor point of the hydration acceleration period. Due to the floating-point precision error in the determination of absolute zero in the discrete digital system, the phase transition anchor point extraction module 102 equivalently replaces the determination condition that the second derivative is equal to zero with finding zero-crossing crossover points in the second derivative sequence. The phase transition anchor point extraction module 102 determines whether adjacent discrete sampling points simultaneously satisfy the following zero-crossing crossover logic condition:

[0062] Phase transition anchor extraction module 102 extracts the time attributes corresponding to discrete sampling points that satisfy the above zero-crossing crossover logic condition, and marks the extracted time attributes as extreme anchor points of hydration heat release rate. ;

[0063] In step S340, the phase transition anchor extraction module 102 performs a ratio calculation on the smoothed high- and low-frequency resistivity sequences to calculate the ratio parameter of the dual-frequency resistivity. Under the same absolute timestamp, the phase transition anchor extraction module 102 extracts the smoothed high-frequency resistivity data and the smoothed low-frequency resistivity data. The formula for calculating the ratio parameter is as follows:

[0064] In the formula, For the first Dual-frequency resistivity ratio parameter at each sampling point For the first Smoothed high-frequency resistivity data at each sampling point For the first Smoothed low-frequency resistivity data at each sampling point, phase transition anchor point extraction module 102 based on high-frequency sampling period Perform a forward difference operation on the dual-frequency resistivity ratio parameter sequence to calculate the first-order evolution slope of the ratio parameter. :

[0065] In step S350, the phase transition anchor point extraction module 102 traverses the first-order evolution slope sequence of the ratio parameter, performs the determination of the pore connectivity cutoff anchor point, and the phase transition anchor point extraction module 102 extracts the first-order evolution slope of the ratio parameter at each sampling point. Compared with the preset step detection threshold Perform numerical comparison and step detection threshold. The value is calibrated using AC impedance spectrum test data from historical benchmark specimens. The system detects that the first-order evolution slope of the ratio parameter satisfies the criterion inequality. When the phase transition anchor point extraction module 102 determines that the microporous network structure has changed significantly, the phase transition anchor point extraction module 102 extracts the time attribute corresponding to the discrete sampling point that triggers the determination inequality and marks it as the resistivity ratio mutation anchor point. Phase change anchor point extraction module 102 hydration exothermic rate extreme anchor point Anchor points with abrupt changes in resistivity ratio Output to feature tensor reconstruction module 103.

[0066] Furthermore, in a specific embodiment of the present invention, the specific execution steps for the hardware state reset process after the prediction system completes the high-frequency phase transition feature capture, and the identification and reconstruction mechanism for local data distortion, are detailed below:

[0067] In step S360, after the phase change anchor extraction module 102 completes the extraction of the initial anchor point and the pore communication cut-off anchor point during the hydration acceleration period, it generates a state reset flag. The phase change anchor extraction module 102 sends the state reset flag to the frequency conversion control module 101 through the data bus inside the microprocessor 31. The frequency conversion control module 101 reads the state reset flag and locates the physical address of the programmable timer that controls the working cycle of the analog-to-digital converter 21. The frequency conversion control module 101 erases the high-frequency division constant in the internal control register of the programmable timer and rewrites the low-frequency division constant. The sampling period of the analog-to-digital converter 21 is restored from the high-frequency sampling period to the initial low-frequency sampling period. This hardware reset operation only reduces the acquisition frequency of subsequent background data and does not affect the integrity of the high-frequency sequence data already stored in the internal loop data buffer of the microprocessor, thereby providing an independent data foundation for subsequent anomaly identification and spatial reconstruction.

[0068] In step S370, the feature tensor reconstruction module 103 executes an abnormal physical distortion identification procedure. Although there is an inherent intrinsic time delay between the temperature field phase transition and the electrochemical field phase transition during normal concrete hydration, due to the spatial heterogeneity of the internal coarse aggregate distribution, when a single-point sensor encounters significant aggregate interference in a local area, the asynchronous misalignment between the two will be abnormally aggravated. The feature tensor reconstruction module 103 extracts the extreme value anchor point of the hydration heat release rate at the same spatial monitoring node. Anchor with a sudden change in resistivity ratio And calculate the time delay difference parameter between the two. ;

[0069] In step S380, the feature tensor reconstruction module 103 calls the tolerance judgment threshold stored in memory 32. Time delay difference parameter With tolerance judgment threshold Perform numerical comparison and tolerance judgment threshold. The values ​​are obtained statistically from historical phase transition anchor point time delay misalignment sample data of interference-free test blocks, and the tolerance judgment threshold is determined. The specific calculation boundary is the sample mean plus three times the sample standard deviation, when the time delay difference parameter... Less than the tolerance threshold When the feature tensor reconstruction module 103 determines that the multiphysics data acquired by the current space monitoring node is valid, the feature tensor reconstruction module 103 retains the multiphysics data acquired by the current space monitoring node, and when the time delay difference parameter Greater than or equal to the tolerance threshold When the feature tensor reconstruction module 103 determines that the data sequence of the current spatial monitoring node is physically distorted by local aggregate interference, the feature tensor reconstruction module 103 triggers the abnormal data reconstruction logic.

[0070] In step S390, the feature tensor reconstruction module 103 performs spatial interpolation reconstruction on the phase transition anchor point data that has undergone physical distortion using a spatial distance inverse square weighted algorithm. The feature tensor reconstruction module 103 reads the coordinates of adjacent valid spatial monitoring nodes that have not undergone distortion. The adjacent valid spatial monitoring nodes are data nodes that have passed the tolerance judgment threshold screening. The effective phase transition anchor point timestamps of the adjacent valid spatial monitoring nodes are extracted. The feature tensor reconstruction module 103 calculates the starting anchor point of the new hydration acceleration period of the distorted node. and new pore connection cut-off anchor points The interpolation reconstruction calculation formula is as follows:

[0071] In the formula, This serves as the starting anchor point for the accelerated hydration phase after reconstruction. To cut off the anchor points for reconstructed pore connectivity, The total number of adjacent valid spatial monitoring nodes invoked. For the first The starting anchor point of the hydration acceleration period of each adjacent effective spatial monitoring node. For the first The pore connectivity cutoff anchor point of adjacent effective spatial monitoring nodes For the first The spatial distance between adjacent effective spatial monitoring nodes is inversely proportional to the weight;

[0072] Step S395: Feature tensor reconstruction module 103 establishes a spatial rectangular coordinate system and calculates the spatial distance inverse weights. To avoid division by zero singularities, a smoothing constant is introduced into the calculation formula. :

[0073] In the formula, The three-dimensional coordinates of the target monitoring node. For the first The three-dimensional coordinates of each of the adjacent valid nodes. As a minimal positive real number, this spatial interpolation reconstruction algorithm is based on the presupposition that the macroscopic physical field inside concrete is continuous and isotropic in local space. It can effectively smooth single-point random abrupt changes caused by local aggregate aggregation. The feature tensor reconstruction module 103 uses the calculated extreme value anchor point of the new hydration heat release rate. Anchor point with abrupt change in the ratio of new resistivity It replaces the original distorted data and completes the repair of abnormal physical distortion data.

[0074] Furthermore, in a specific embodiment of the present invention, the numerical calculation and memory splicing execution steps of the feature tensor reconstruction module 103 at the bottom layer for the data dimensionality reduction and feature reconstruction process after multiphysics feature extraction are detailed as follows:

[0075] In step S410, the feature tensor reconstruction module 103 allocates an independent tensor construction cache in the continuous static random access memory of the microprocessor 31, and the feature tensor reconstruction module 103 calls the reconstructed hydration acceleration period starting anchor point after anomaly identification and spatial reconstruction processing. Interconnection with reconstructed pores and cut-off anchor points The feature tensor reconstruction module 103 will reconstruct the starting anchor point of the hydration acceleration period. Interconnection with reconstructed pores and cut-off anchor points The feature tensor reconstruction module 103 calculates the time delay difference parameter of the time boundary condition, which is set as the time boundary condition for calculus operations. Time delay difference parameter The calculation formula is:

[0076] Feature tensor reconstruction module 103 instructs microprocessor 31 to convert the time delay difference parameter Write it to the first physical address of the tensor-built buffer;

[0077] In step S420, the feature tensor reconstruction module 103 extracts continuous data slices within the time boundary conditions from the circular data buffer and performs discretization of the thermodynamic definite integral. The data object processed by the microprocessor 31 is a discretely sampled digital sequence. The feature tensor reconstruction module 103 uses the trapezoidal rule numerical integration algorithm to perform discrete integration operations. The feature tensor reconstruction module 103 traverses the temperature sequence in the continuous data slice, calculates the area of ​​the discrete trapezoid formed by adjacent sampling points, and performs accumulation. The accumulated temperature-time integral feature is obtained. The calculation formula is as follows:

[0078] In the formula, To reconstruct the starting anchor point of the accelerated hydration period The corresponding storage index value in the circular data cache. To reconstruct pore connectivity, cut off anchor points The corresponding storage index value in the circular data cache. To store index values Temperature measurement values ​​at the location, To store index values Temperature measurement values ​​at the location, To store index values The absolute timestamp value at that location. To store index values The absolute timestamp value at that location, the feature tensor reconstruction module 103 will solve the temperature-time integral feature obtained from the solution. Write tensors to build a buffer;

[0079] In step S430, the feature tensor reconstruction module 103 calls the low-frequency resistivity sequence data within the time boundary conditions, performs a differential solution of the electrochemical evolution slope, and extracts and reconstructs the pore connectivity cutoff anchor points. The corresponding low-frequency resistivity measurement values ​​at the end, and the starting anchor point of the reconstructed hydration acceleration period. The corresponding initial low-frequency resistivity measurement value is used to calculate the physical difference between the two resistivity measurement values. The characteristic tensor reconstruction module 103 then divides the physical difference by the time delay difference parameter. Solve for the average slope of resistivity change within the interval. Average slope of resistivity change The calculation formula is as follows:

[0080] In the formula, To store index values Low-frequency resistivity measurement values ​​at the location. To store index values The low-frequency resistivity measurement values ​​at the location are used by the characteristic tensor reconstruction module 103 to average the slope of the resistivity change. Write the third physical address of the tensor to construct the buffer;

[0081] In step S440, the feature tensor reconstruction module 103 instructs the microprocessor 31 to sequentially access the first three consecutive physical addresses in the tensor construction cache through memory pointer increment operations, completing the physical splicing of discrete scalar data. The physical splicing operation assembles scalar values ​​of different dimensions into a one-dimensional dynamic feature tensor. Due to absolute timestamps Due to the influence of initial batch conditions, cross-batch comparability is not possible. The dynamic feature tensor removes absolute time and retains relative duration. The structural formula is modified as follows:

[0082] In the formula, This is the time delay difference parameter between the two phase change anchor points. This is a temperature-time integral characteristic. This represents the average slope of the resistivity change.

[0083] Furthermore, in a specific embodiment of the present invention, regarding the performance mapping process after the feature tensor reconstruction is completed, the underlying model topology construction and forward computation steps of the neural network prediction module 104 are as follows:

[0084] In step S510, the neural network prediction module 104 statically allocates memory space in the memory 32 to construct the network topology of a multilayer perceptron. The multilayer perceptron consists of a cascaded input layer, hidden layer, and output layer, combined with a one-dimensional dynamic feature tensor. In terms of data dimensions, the neural network prediction module 104 statically configures the number of neurons in the input layer to 3. These 3 neurons correspond to the receiving time delay difference parameter, the temperature-time integral feature, and the average slope of resistivity change, respectively. The hidden layer is configured as a two-layer fully connected network structure, with the first hidden layer configured as follows: Each neuron node, configured in the second hidden layer. Each layer has a single neuron node, and the output layer is configured with a single neuron node to output the predicted scalar value.

[0085] Step S520, the neural network prediction module 104 converts the one-dimensional dynamic feature tensor The input layer's memory register is loaded, and forward inference matrix operations from the input layer to the first hidden layer are performed. The neural network prediction module 104 calls the first weight matrix and the first bias vector pre-stored in memory 32, performs a linear weighted summation on the input data, and maps it to the first hidden layer through a non-linear activation function. The first hidden layer outputs a vector. The calculation formula is as follows:

[0086] In the formula, This is the column vector output by the solution of the first hidden layer. For the set nonlinear activation function, It is the transpose column vector of a one-dimensional dynamic feature tensor; The first weight matrix is ​​the first weight matrix from the input layer to the first hidden layer. The matrix dimension is ; The first bias column vector, the first bias column vector The vector dimension is Nonlinear activation function Specifically, a modified linear unit activation function is used;

[0087] Step S530: The neural network prediction module 104 extracts the output vector of the first hidden layer. The output vector of the first hidden layer The data is passed to the second hidden layer for deep feature dimension transformation. The forward inference calculation logic of the second hidden layer is consistent with the structure of the first hidden layer. The output vector of the second hidden layer is... The calculation formula is:

[0088] In the formula, This is the column vector output from the solution of the second hidden layer; The second weight matrix is ​​the weight matrix from the first hidden layer to the second hidden layer. The matrix dimension is ; The second bias column vector, the second bias column vector The vector dimension is ;

[0089] Step S540, the neural network prediction module 104 predicts based on the output vector of the second hidden layer. The output layer performs regression calculations. Each neuron in the output layer performs a linear combination and accumulation of the input feature sequence, followed by boundary constraints imposed by the Sigmoid nonlinear activation function. This strictly confines the continuous predicted output values ​​to the (0,1) interval. The physical performance prediction values... The solution formula is as follows:

[0090] In the formula, The predicted physical properties of the output layer are calculated. The output weight matrix is ​​the output weight matrix from the second hidden layer to the output layer. The matrix dimension is ; This is the bias scalar of the output layer;

[0091] Step S550, the neural network prediction module 104 predicts the physical performance values. Performing an inverse normalization reduction operation maps the dimensionless values ​​output by the network to the final predicted physical properties with actual engineering dimensions. Final physical performance prediction value The inverse normalization reduction calculation formula is as follows:

[0092] In the formula, The output is the final predicted physical performance value. This represents the maximum value of the physical performance index label in the historical training set. The microprocessor 31 extracts the final predicted physical performance value from the minimum value of the physical performance indicator label in the historical training set. The final physical performance prediction value The microprocessor 31 writes the data into the transmit register of the system communication bus and transmits it to an external display terminal or host computer control system through a hardware interface.

Claims

1. A method for predicting concrete performance based on artificial intelligence, characterized in that, Includes the following steps: Step S100: Collect multi-physics field measurement data inside the concrete to be tested at an initial low-frequency sampling frequency. The multi-physics field measurement data includes discrete temperature field signals and discrete high and low frequency resistivity signals. Store the multi-physics field measurement data as a low-frequency background data time-series array in a time sequence into a circular data buffer. Step S200: Perform evolution slope calculation on the multiphysics field measurement data within the sliding observation window, and when the absolute value of the first-order evolution slope reaches the warning slope threshold, control the sampling frequency to switch to a high-frequency sampling frequency and collect data to obtain high-frequency sequence data and write it into the circular data buffer. Step S300: Differentiate the high-frequency sequence data, extract the extreme value anchor point of hydration heat release rate and the abrupt change anchor point of resistivity ratio, and simultaneously trigger sampling frequency reset to restore the sampling frequency to the initial low-frequency sampling frequency. Step S400: Extract data slices within the phase transition interval, extract physical feature parameters, and construct a one-dimensional dynamic feature tensor; Step S500: Receive the normalized one-dimensional dynamic feature tensor and input it into the pre-deployed multilayer perceptron model for forward inference, output the predicted value of concrete physical properties, and perform inverse normalization restoration operation to obtain the final predicted value of physical properties with actual engineering dimensions.

2. The method for predicting concrete performance based on artificial intelligence according to claim 1, characterized in that, Step S200 specifically includes: Step S210: Set a sliding observation window with a fixed time width within the circular data buffer of the microprocessor. The sliding observation window contains the total number of sampling points. To ensure that high-frequency hardware noise is filtered out without missing the actual physical evolution, the total number of sampling points is [number missing]. Based on the macroscopic thermodynamic evolution timescale of the initial induction period of the concrete to be tested, as the data sensing layer continues to sample, the sliding observation window moves synchronously backward on the time axis according to the first-in-first-out principle. The frequency conversion control module periodically extracts the temperature measurement value and resistivity measurement value sequence within the sliding observation window according to the set operation rhythm, and loads the extracted sequence data into the high-speed register of the microprocessor 31 to perform differential operation. In step S220, the frequency conversion control module calls the sequence data in the high-speed register to calculate the comprehensive evolution slope at the sampling point at the end of the sliding observation window as the comprehensive judgment basis. The frequency conversion control module calculates the comprehensive evolution slope at the sampling point at the end of the sliding observation window according to the discrete difference formula. The calculation formula is as follows: In the formula, For the end of the sliding window The average evolution slope calculated at each sampling point For the first Temperature measurement values ​​at each sampling point This represents the temperature measurement value corresponding to the starting point of the sliding window. This represents the total number of sampling points within the sliding window. The initial low-frequency sampling period is used; and the arithmetic mean of the first-order evolution slopes of multiple consecutive sampling points is calculated as the comprehensive evolution slope. Step S230: Compare the absolute value of the comprehensive evolution slope with the system's preset warning slope threshold. Step S240: When the absolute value of the comprehensive evolution slope is greater than or equal to the warning slope threshold, a frequency conversion interruption request is triggered and a high-frequency sampling command is sent to the signal conditioning layer through the data bus to complete the physical switching of the underlying sampling cycle.

3. The method for predicting concrete performance based on artificial intelligence according to claim 1, characterized in that, The step of extracting the extreme value anchor point of the hydration heat release rate in step S300 includes: Step S310: Call the high-frequency temperature sequence data and high-frequency resistivity sequence data of the cyclic data buffer, and perform moving average filtering and noise reduction to output a smooth data sequence; Step S320: The first-order forward difference method is used to calculate the first-order temperature derivative sequence at the corresponding sampling point for the smoothed high-frequency temperature sequence, and the second-order backward difference operation is performed in combination with the temperature data to construct a high-resolution second-order temperature derivative sequence. Step S330: Traverse the constructed temperature first derivative sequence and temperature second derivative sequence, determine whether adjacent discrete sampling points simultaneously satisfy the logical condition that the second derivative crosses zero and the first derivative is positive, and mark the time attribute corresponding to the discrete sampling point that satisfies the logical condition as the extreme value anchor point of hydration heat release rate.

4. The method for predicting concrete performance based on artificial intelligence according to claim 1, characterized in that, The step of extracting the anchor point of the resistivity ratio change in step S300 includes: Step S340: Extract smoothed high-frequency resistivity data and smoothed low-frequency resistivity data under the same absolute timestamp to perform ratio operation, calculate the ratio parameter of dual-frequency resistivity, and perform forward difference operation on the ratio parameter sequence of dual-frequency resistivity to obtain the first-order evolution slope sequence of ratio parameter. Step S350: Traverse the first-order evolution slope sequence of the ratio parameter, compare the first-order evolution slope of the ratio parameter at each sampling point with the preset step judgment threshold, and determine that the micro-pore network has a topological connectivity blockage when the judgment inequality is satisfied, and mark the time attribute corresponding to the discrete sampling point that triggers the judgment inequality as the resistivity ratio mutation anchor point.

5. The method for predicting concrete performance based on artificial intelligence according to claim 1, characterized in that, Step S300 also includes a step of performing abnormal physical distortion identification: Step S370: Extract the extreme value anchor point of the hydration heat release rate and the abrupt change anchor point of the resistivity ratio at the same spatial monitoring node, and calculate the time delay difference parameter between the two. Step S380: The time delay difference parameter is compared with the preset tolerance judgment threshold in the memory. When the time delay difference parameter is greater than or equal to the tolerance judgment threshold, it is determined that the data sequence of the current spatial monitoring node is physically distorted by local aggregate interference, triggering the abnormal data reconstruction logic.

6. The method for predicting concrete performance based on artificial intelligence according to claim 5, characterized in that, The steps following the triggering of the abnormal data reconstruction logic in step S300 include: Step S390: Read the coordinates of adjacent valid spatial monitoring nodes that have not undergone physical distortion and the timestamps of valid phase transition anchor points. Use the spatial distance inverse square weighted algorithm to perform spatial interpolation reconstruction on the phase transition anchor point data that has undergone physical distortion. Calculate the new extreme value anchor point of the hydration heat release rate and the new abrupt change anchor point of the resistivity ratio of the spatial monitoring node. Step S395: Establish a spatial rectangular coordinate system and calculate the spatial distance inverse weight by combining the three-dimensional coordinates of the spatial monitoring node. Use the new extreme value anchor point of the hydration heat release rate and the new abrupt change anchor point of the resistivity ratio to replace the original distorted data.

7. The method for predicting concrete performance based on artificial intelligence according to claim 6, characterized in that, Step S400 specifically includes: Step S410: In the continuous static random access memory, an independent tensor construction buffer is opened, and the reconstructed new hydration heat release rate extreme value anchor point and the new resistivity ratio change anchor point are set as time boundary conditions. The time delay difference parameter of the time boundary conditions is calculated and written into the first physical address of the tensor construction buffer. Step S420: Extract continuous data slices within the time boundary conditions from the circular data buffer, perform discrete integration using the trapezoidal rule numerical integration algorithm, solve for the temperature-time integral features, and write them to the second physical address of the tensor construction buffer. Step S430: Extract the initial low-frequency resistivity measurement value and the final low-frequency resistivity measurement value corresponding to the time boundary condition and calculate the physical difference. Divide the physical difference by the time delay difference parameter to solve the average slope of resistivity change and write it into the third physical address of the tensor construction buffer. Step S440: The first three consecutive physical addresses in the tensor construction cache are accessed sequentially through memory pointer increment operation to complete the physical splicing of discrete scalar data to assemble it into the one-dimensional dynamic feature tensor.

8. The method for predicting concrete performance based on artificial intelligence according to claim 7, characterized in that, In step S420, the temperature-time integral feature is solved. The calculation formula is as follows: In the formula, The storage index value in the cyclic data buffer corresponds to the extreme anchor point of the new hydration heat release rate. The storage index value in the cyclic data buffer corresponds to the anchor point of the new resistivity ratio mutation. To store index values Temperature measurement values ​​at the location, To store index values Temperature measurement values ​​at the location, To store index values The absolute timestamp value at that location. To store index values The absolute timestamp value at that location.

9. The method for predicting concrete performance based on artificial intelligence according to claim 1, characterized in that, The multilayer perceptron model consists of a cascaded input layer, a hidden layer, and an output layer. Step S500, which involves forward inference and outputting the predicted values ​​of the concrete physical properties, includes: Step S510: Statically allocate memory space in the memory, construct the network topology of the multilayer perceptron, and statically configure the number of neuron nodes in the input layer to 3. Step S520: Load the one-dimensional dynamic feature tensor into the memory register of the input layer, call the preset first weight matrix and the first bias vector to perform forward inference matrix operation, and map it to the first hidden layer through a non-linear activation function; Step S530: Extract the output vector of the first hidden layer and pass it to the second hidden layer for deep feature dimension transformation; Step S540: Perform regression calculation operation of the output layer based on the output vector of the second hidden layer, and generate continuous predicted values ​​representing macroscopic mechanical properties through linear combination and accumulation as the predicted values.

10. The method for predicting concrete performance based on artificial intelligence according to claim 1, characterized in that, The step of performing the inverse normalization reduction operation in step S500 includes: Step S550: Using the maximum and minimum values ​​of the physical performance index labels in the historical training set, perform an inverse normalization restoration operation on the predicted value to restore it to the final physical performance prediction value with actual engineering dimensions, and transmit it to the external terminal through the hardware interface.

Images