Concrete mixing performance prediction method based on deep learning

Abstract

The invention relates to the technical field of intelligent sensors, in particular to a concrete stirring performance prediction method based on deep learning, which comprises the following steps: synchronously acquiring a concrete diffuse reflection spectrum and a friction acoustic emission signal in a pneumatic optical coupling flow environment; the chemical components of the expanding agent are represented by analyzing the differential form of a spectral absorption peak, and the particle rheological state is analyzed by combining the time-frequency energy entropy evolution law of the sound wave signal; performing physical verification and dynamic fusion on the chemical component characteristics according to the rheological state analysis result to generate an intermediate index reflecting the mixing uniformity; and thermodynamic dissipation equilibrium constraint of particle fluid dynamics is introduced to carry out energy consistency correction on the index, and finally the restricted expansion rate representing the volume stability of the material is calculated. According to the invention, real-time online prediction and analysis of the restrained expansion rate of concrete in a severe environment are realized through multi-dimensional collaborative perception of the intelligent sensor and heterogeneous signal analysis under physical constraints.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, specifically to a method for predicting concrete mixing performance based on deep learning. Background Technology

[0002] In modern underground engineering and ultra-long structure construction, ready-mixed shrinkage-compensating concrete is widely used because it can generate pre-stress by introducing an expansive agent to counteract shrinkage cracks. The core mechanism of this material relies on the microscopic and uniform distribution of the expansive agent in the cement matrix. In the industrial production of dry-mixed concrete, the expansive agent usually accounts for only a small proportion of the total dry mass, and its particle density and particle size differ significantly from ordinary cement and aggregates. It is very prone to segregation during the mixing process, resulting in localized areas of insufficient and abundant expansion within the structure, thereby causing cracking or poor stability risks.

[0003] Existing testing technologies for concrete exhibit significant lag. Traditional restricted expansion rate tests require specimen preparation and curing for several weeks to obtain results, making them post-testing methods that cannot meet the real-time quality control needs of modern batching plants. While chemical titration can analyze component content, its single-point sampling method introduces statistical bias and cannot represent the overall homogeneity of dry-mix concrete. Furthermore, existing mixer current monitoring only reflects macroscopic rheological properties and cannot detect the chemical distribution of trace amounts of expansion agents.

[0004] In recent years, although near-infrared spectroscopy and acoustic emission technology have been attempted for materials analysis, single-modal smart sensors struggle to simultaneously capture microscopic chemical composition and macroscopic physical motion characteristics in the harsh environment of concrete mixing with high dust and strong vibration. In particular, establishing a mapping relationship between multidimensional signals in the dry powder state and the volume stability after hardening remains an unsolved problem in the industry.

[0005] Therefore, in order to solve the technical problem that the limited expansion rate of premixed shrinkage-compensating concrete under harsh mixing conditions cannot be predicted online in real time, this invention proposes a concrete mixing performance prediction method based on deep learning. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting concrete mixing performance based on deep learning. By using multi-dimensional collaborative perception of intelligent sensors and heterogeneous signal analysis under physical constraints, the method can achieve real-time online prediction and analysis of the restricted expansion rate of concrete under harsh environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Deep learning-based methods for predicting concrete mixing performance include: The diffuse reflectance spectral response data of premixed shrinkage-compensating concrete under aero-optical coupled flow environment were obtained, and the frictional acoustic emission signal of the concrete during the mixing process was also obtained simultaneously. The diffuse reflectance spectral response data and the triboacoustic emission signal are input into a dual-channel feature extraction model for parallel signal analysis. The first channel of the dual-channel feature extraction model extracts the peak shape of the diffuse reflectance spectral response data to generate a chemical component fingerprint characterizing the distribution of the expanding agent. The second channel performs rheological state evolution analysis on the triboacoustic emission signal to generate rheological dynamic parameters characterizing the particle motion state. By using rheological dynamic parameters as state gating signals to dynamically weight the chemical component fingerprint, a dry mixing characteristic index reflecting the real-time mixing uniformity of materials is generated. Energy consistency correction is performed on the dry mixing characteristic index based on the thermodynamic dissipation equilibrium constraint of particle fluid dynamics to generate effective expansion potential energy characteristics that characterize material properties; multi-parameter nonlinear inversion is performed on the effective expansion potential energy characteristics to obtain the predicted value of the restricted expansion rate of concrete; based on the predicted value of the restricted expansion rate, the judgment result characterizing the mixing performance of concrete is output.

[0008] Preferably, the acquisition process of the diffuse reflectance spectral response data and the triboacoustic emission signal includes: controlling the sampling clock of the high-pressure air curtain generator and the near-infrared spectral intelligent sensor to perform synchronous triggering, constructing a local laminar air shield in the acquisition optical path that can eliminate dust interference, forming a transient transparent window in the aero-optical coupled flow environment; driving the near-infrared spectral intelligent sensor to perform photoelectric detection through the transient transparent window, acquiring the original light intensity signal, and using the built-in edge computing module to perform dark current compensation and standardization processing on the original light intensity signal to generate diffuse reflectance spectral response data containing material reflection characteristics and atmospheric medium absorption characteristics of the optical path; simultaneously driving the acoustic emission intelligent sensor with edge computing capabilities to perform acoustic wave acquisition, using the built-in front-end filtering module to perform structural waveguide signal separation, filtering out aerodynamic noise and low-frequency mechanical vibration under the air coupling path, retaining the particle contact friction signal in the ultrasonic frequency band transmitted through the mixer wall, and generating a triboacoustic emission signal.

[0009] Preferably, the dual-channel feature extraction model is constructed as a heterogeneous signal parallel analysis architecture, including: a first channel, which constructs a spatial morphology analysis flow path based on a one-dimensional convolutional neural network, performs multi-scale sliding convolution operation on the diffuse reflectance spectral response data, and extracts the spatial distribution characteristics of the expanding agent in the dry mix; and a second channel, which constructs a temporal dynamic analysis flow path based on a long short-term memory network, performs time series dependency analysis on the triboacoustic emission signal, and captures the evolution law of particle motion state with stirring time.

[0010] Preferably, the first channel of the dual-channel feature extraction model extracts the peak morphology of the diffuse reflectance spectral response data to generate a chemical component fingerprint characterizing the distribution of the expanding agent. Specifically, this includes: locating the first characteristic absorption band corresponding to the characteristic chemical bonds of the active component of the expanding agent and the second characteristic absorption band corresponding to the matrix material in the diffuse reflectance spectral response data; performing differential operations on the first characteristic absorption band to calculate the slope of the first derivative and the curvature of the second derivative of the spectral response curve, generating morphological parameters characterizing the chemical bond strength; comparing and analyzing the morphological parameters with the absorbance baseline value of the second characteristic absorption band to eliminate chromatic background interference and generate a chemical component fingerprint corrected for baseline drift.

[0011] Preferably, the step of performing rheological state evolution analysis on the triboacoustic emission signal in the second channel to generate rheological dynamic parameters characterizing the particle motion state specifically includes: performing a short-time Fourier transform on the triboacoustic emission signal to construct a time-frequency distribution matrix reflecting the frequency change over time; extracting the energy proportion of the ultrasonic frequency band friction component from the time-frequency distribution matrix and calculating the signal energy entropy within the sliding time window; monitoring the time series change rate of the signal energy entropy, and when the change rate approaches zero, extracting the current signal energy entropy value as a rheological dynamic parameter characterizing the particle's transition from a collision state to a steady-state shear state.

[0012] Preferably, the step of dynamically weighting the chemical component fingerprint using rheological kinetic parameters as state gating signals to generate a dry mixing characteristic index reflecting the real-time mixing uniformity of the material specifically includes: constructing a mixing uniformity confidence interval based on the rheological kinetic parameters; if the current rheological kinetic parameters are outside the mixing uniformity confidence interval, it is determined to be an unsteady stirring period, and a signal attenuation weight coefficient is generated; and feature fusion is performed, wherein the feature fusion includes: constructing a feature suppression gating vector adapted to the dimension of the chemical component fingerprint using the signal attenuation weight coefficient; performing an element-wise multiplication weighting operation on the chemical component fingerprint using the feature suppression gating vector to suppress spectral fluctuation noise during the unsteady stirring period; and concatenating the data dimensions of the chemical component fingerprint and the rheological kinetic parameters to generate a dry mixing characteristic index.

[0013] Preferably, the step of performing energy consistency correction on the dry mixing characteristic index based on the thermodynamic dissipation equilibrium constraint of particle fluid dynamics to generate an effective expansion potential energy characteristic characterizing material properties specifically includes: constructing a thermodynamic dissipation equilibrium constraint function, which defines the energy conversion residual between the input mechanical stirring work, the inverted theoretical chemical potential energy, and the frictional heat dissipation; performing gradient descent iteration during the feature mapping process to minimize the energy conversion residual, so that the characteristic distribution of the dry mixing characteristic index conforms to the energy flow boundary constraint of the non-equilibrium thermodynamic system, thereby generating an effective expansion potential energy characteristic corrected by physical laws.

[0014] Preferably, the step of performing multi-parameter nonlinear inversion on the effective expansion potential energy characteristics to obtain the predicted value of the restricted expansion rate of the premixed shrinkage-compensating concrete specifically includes: inputting the effective expansion potential energy characteristics corrected by the energy conservation constraint into the nonlinear regression mapping space; performing multi-objective solution in the nonlinear regression mapping space, and simultaneously outputting the predicted value of the restricted expansion rate corresponding to the standard curing age and the mixing uniformity index characterizing the dispersion degree of the premixed shrinkage-compensating concrete components.

[0015] Preferably, the method further includes the step of environmental adaptive calibration using the diffuse reflectance spectral response data: extracting the atmospheric carbon dioxide characteristic absorption peak corresponding to the atmospheric medium absorption characteristics of the optical path in the diffuse reflectance spectral response data, and using the atmospheric carbon dioxide characteristic absorption peak as the wavelength drift calibration point; calculating the offset between the actual detection wavelength of the characteristic absorption peak and the standard physical wavelength, and using the offset to dynamically calibrate the wavelength axis of the diffuse reflectance spectral response data.

[0016] Preferably, the calculation basis for the predicted value of the restricted expansion rate further includes: pre-constructing a ternary association database between the amount of expansion agent, the water-cement ratio, and the restricted expansion rate; performing similarity matching on the effective expansion potential energy feature in the ternary association database to retrieve reference association nodes with similar features; calculating the weight based on the similarity between the effective expansion potential energy feature and the reference association node, and performing weighted interpolation on the correction coefficient corresponding to the reference association node to obtain the correction coefficient of the predicted value of the restricted expansion rate.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing an aero-optical coupled flow environment through a high-pressure air curtain and combining it with an edge computing filtering strategy from intelligent sensors, a transient transparent window that eliminates dust interference is formed in the acquisition optical path. Simultaneously, edge computing is used to separate the structural waveguide signal and aerodynamic noise in the acoustic wave in real time. This combined design of "active environment construction + edge signal cleaning" enables high-fidelity in-situ acquisition of weak diffuse reflection spectra and triboacoustic emission signals in industrial stirring scenarios with high dust and strong vibration.

[0018] 2. By constructing a dual-channel feature extraction model and dynamically weighting the chemical component fingerprint using rheological parameters as state-gated signals, the effectiveness of chemical component distribution (spectral) is verified and screened using particle motion state (rheology), suppressing noise interference during the unsteady mixing period. This heterogeneous signal gating fusion mechanism enables the synchronous coupled analysis of the concrete's "physical motion state" and "chemical dispersion degree," effectively solving the problem that a single modal sensor cannot comprehensively characterize the uniformity of complex dry-mix materials.

[0019] 3. By introducing thermodynamic dissipation equilibrium constraints based on particle fluid dynamics to perform energy consistency correction on deep learning features, the feature extraction process of the black-box model is restricted to the energy flow boundary of the non-equilibrium thermodynamic system. This feature correction method driven by physical laws achieves accurate nonlinear inversion from dry mixed energy features to the hardened restricted expansion rate, overcoming the poor generalization ability of pure data-driven models under small sample conditions, and replacing the lagging traditional physical experiments.

[0020] 4. By utilizing the characteristic absorption peaks of the atmospheric medium (carbon dioxide) contained in the diffuse reflectance spectral data as in-situ wavelength drift calibration points, adaptive wavelength axis calibration is performed within the intelligent sensor. This technique of self-correction using environmental background information achieves automatic zero-point drift compensation of the sensor without the need for external standard components, ensuring long-term online detection accuracy in environments with drastic temperature variations in mixing plants. Attached Figure Description

[0021] Figure 1 This is a flowchart of the deep learning-based concrete mixing performance prediction method of the present invention. Figure 2 This is a schematic diagram of the architecture of the dual-channel feature extraction model and dynamic gating weighting logic in an embodiment of the present invention; Figure 3 This is a schematic diagram of the energy consistency correction logic based on thermodynamic dissipation balance constraints in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Other embodiments obtained by those skilled in the art based on the ideas in this specification without creative effort all fall within the protection scope of this invention.

[0023] Reference Figure 1 As shown in the figure, this invention provides a method for predicting concrete mixing performance based on deep learning, and the specific technical solution is as follows: The diffuse reflectance spectral response data of premixed shrinkage-compensating concrete under aero-optical coupled flow environment were obtained, and the frictional acoustic emission signal of the concrete during the mixing process was also obtained simultaneously. The diffuse reflectance spectral response data and the triboacoustic emission signal are input into a dual-channel feature extraction model for parallel signal analysis. The first channel of the dual-channel feature extraction model extracts the peak shape of the diffuse reflectance spectral response data to generate a chemical component fingerprint characterizing the distribution of the expanding agent. The second channel performs rheological state evolution analysis on the triboacoustic emission signal to generate rheological dynamic parameters characterizing the particle motion state. By using rheological dynamic parameters as state gating signals to dynamically weight the chemical component fingerprint, a dry mixing characteristic index reflecting the real-time mixing uniformity of materials is generated. Energy consistency correction is performed on the dry mixing characteristic index based on the thermodynamic dissipation equilibrium constraint of particle fluid dynamics to generate effective expansion potential energy characteristics that characterize material properties; multi-parameter nonlinear inversion is performed on the effective expansion potential energy characteristics to obtain the predicted value of the restricted expansion rate of concrete; based on the predicted value of the restricted expansion rate, the judgment result characterizing the mixing performance of concrete is output.

[0024] Example 1: This embodiment takes the production and preparation of C50 premixed shrinkage-compensating concrete segments used in underground rail transit engineering as the application scenario. This scenario requires high-precision real-time monitoring of the restricted expansion rate of concrete to prevent microcracks from forming in the segments after curing.

[0025] Reference Figure 1In the deep learning-based concrete mixing performance prediction method described in this embodiment, the execution process covers the complete path from signal acquisition in a complex multiphase flow environment to physical sensing feature calculation. During the method execution, a transient optical window is constructed inside the mixer using a high-pressure air curtain to simultaneously acquire diffuse reflectance spectral response data excluding dust interference and triboelectric acoustic emission signals excluding aerodynamic noise. These heterogeneous signals are transmitted in parallel to a physically constrained dual-channel feature extraction model. The model obtains the chemical composition fingerprint of the expansion agent by analyzing the spectral peak morphology, and simultaneously obtains particle rheological dynamic parameters by analyzing the time-frequency evolution law of the acoustic emission signal. The rheological dynamic parameters are used to dynamically gate and weight the chemical composition fingerprint to generate a dry mixing characteristic index. Based on the thermodynamic dissipation balance rule of particle fluid dynamics, the index is corrected for energy consistency. Finally, the predicted value of the restricted expansion rate of the target concrete at the standard curing age is obtained through nonlinear inversion, realizing real-time quantitative evaluation of the mixing quality.

[0026] Furthermore, the acquisition process of the diffuse reflectance spectral response data and triboacoustic emission signal includes: controlling the sampling clock of the high-pressure air curtain generator and the near-infrared spectral intelligent sensor to perform synchronous triggering, constructing a local laminar air shield in the acquisition optical path that can eliminate dust interference, forming a transient transparent window in the aero-optical coupled flow environment; driving the near-infrared spectral intelligent sensor to perform photoelectric detection through the transient transparent window, acquiring the original light intensity signal, and using the built-in edge computing module to perform dark current compensation and standardization processing on the original light intensity signal to generate diffuse reflectance spectral response data containing material reflection characteristics and atmospheric medium absorption characteristics of the optical path; simultaneously driving the acoustic emission intelligent sensor with edge computing capabilities to perform acoustic wave acquisition, using the built-in front-end filtering module to perform structural waveguide signal separation, filtering out aerodynamic noise and low-frequency mechanical vibration under the air coupling path, retaining the particle contact friction signal in the ultrasonic frequency band transmitted through the mixer wall, and generating triboacoustic emission signal.

[0027] Specifically, in order to obtain high-quality signals in the harsh environment of high dust and high noise inside the concrete mixer, the system adopts a strategy that combines active environmental control with hardware-level signal preprocessing.

[0028] In acquiring diffuse reflectance spectral response data, the control unit sends a trigger command to the high-pressure air curtain generator, driving it to spray clean, dry compressed air at a preset pressure (e.g., 0.6 MPa to 0.8 MPa). This high-speed airflow forms a local laminar air shield in front of the near-infrared spectral intelligent sensor probe, using fluid dynamics principles to forcibly displace suspended cement dust and water mist in the optical path, thereby constructing a transient transparent window in an aero-optical coupled flow environment with a duration of approximately 500 milliseconds to 1 second.

[0029] Simultaneously, upon receiving the synchronization trigger signal, the near-infrared spectral intelligent sensor performs a spectral scan through the transient transparent window, acquiring raw light intensity signals covering a wavelength range of 900nm to 1700nm. The edge computing module integrated within the sensor (such as an embedded FPGA chip) then preprocesses the raw signal: it reads pre-stored dark current noise data (i.e., the sensor background response when the light source is off) and subtracts it from the raw light intensity signal to eliminate thermal noise interference; subsequently, it performs a ratio calculation using the reflection intensity data of a reference whiteboard to complete the standardization process. As a preferred embodiment, the step of generating diffuse reflectance spectral response data containing material reflection characteristics and atmospheric medium absorption characteristics of the optical path further includes: connecting a gas heating and drying module in series in the gas path of the high-pressure gas curtain generator; utilizing the Joule heating effect to perform active thermal intervention on the jet gas flow, eliminating the risk of localized supercooling and condensation in the aero-optical coupled flow environment; and constructing an anti-fog transparent channel using the hot gas flow to obtain diffuse reflectance spectral response data with a high signal-to-noise ratio. Specifically, considering that the rapid temperature drop caused by the Joule-Thomson effect when high-pressure gas is released instantaneously into a normal-pressure environment, which can easily lead to water vapor condensation and form water mist that obstructs the light path in the high-temperature and high-humidity environment of concrete mixing, the system preheats the compressed air to a temperature higher than the ambient dew point (e.g., 50 to 60 degrees Celsius) using a gas heating and drying module (e.g., the hot end of a vortex tube or an electric heater). This preheated high-speed airflow not only counteracts the expansion cooling effect and prevents the formation of condensation mist, but also utilizes the evaporation capacity of the hot airflow to further remove any tiny droplets that may be present in the light path. The edge computing module then focuses on the gas density changes caused by the hot airflow and uses a preset temperature-air refractive index lookup table to perform baseline correction on the original light intensity signal. This step resolves the inherent contradiction between pneumatic dust removal and condensation misting. By introducing an active thermal intervention mechanism, it eliminates secondary environmental interference caused by the high-pressure air curtain, ensuring the authenticity and accuracy of spectral data under extreme temperature variations. After the above processing, the output data is the diffuse reflectance spectral response data. This data not only contains the diffuse reflectance spectral characteristics of the material, but also retains the absorption characteristics of the atmospheric medium (such as carbon dioxide and water vapor) in the optical path, thus reserving a data foundation for subsequent environmental calibration.

[0030] Regarding the acquisition of triboacoustic emission signals, an acoustic emission smart sensor installed on the outer wall of the mixer works synchronously with spectral acquisition. Due to the large amount of aerodynamic noise (air coupling path) and low-frequency mechanical vibration (typically below 20kHz) from the motor drive inside the mixer, the signal-to-noise ratio of the directly acquired signal is extremely low. Therefore, this embodiment utilizes the metal wall of the mixer as a structural waveguide for sound wave transmission. The acoustic emission smart sensor is tightly attached to the wall through a coupling agent to receive elastic waves transmitted through the wall. The front-end filtering module built into the sensor performs structural waveguide signal separation, specifically using a bandpass filter to block low-frequency components below 20kHz, allowing only ultrasonic signals in the 50kHz to 300kHz frequency band to pass. This frequency band mainly corresponds to the stress wave energy released during microscopic contact, friction, and collision between aggregate particles and between aggregate and the wall. The signal after hardware filtering is digitally sampled (e.g., sampling rate of 1MHz) to generate a high-fidelity triboacoustic emission signal, which can sensitively reflect the microscopic evolution of the rheological state of concrete materials.

[0031] This technical solution effectively solves two major acquisition challenges in concrete mixing scenarios: "dust obscuring the spectrum" and "environmental noise drowning out sound waves" by constructing aero-optical transient windows and utilizing structured waveguide filtering technology. This ensures that the data input to the model has an extremely high signal-to-noise ratio and feature fidelity, laying a physical foundation for subsequent high-precision prediction.

[0032] Furthermore, the dual-channel feature extraction model is constructed as a heterogeneous signal parallel analysis architecture, including: a first channel, which constructs a spatial morphology analysis flow path based on a one-dimensional convolutional neural network, performs multi-scale sliding convolution operations on the diffuse reflectance spectral response data, and extracts the spatial distribution characteristics of the expanding agent in the dry mix; and a second channel, which constructs a temporal dynamic analysis flow path based on a long short-term memory network, performs time series dependency analysis on the triboacoustic emission signal, and captures the evolution law of particle motion state with stirring time.

[0033] Specifically, refer to Figure 2 In this embodiment, the dual-channel feature extraction model is designed with differentiated deep learning network topologies for two types of input data with drastically different physical properties.

[0034] For the first channel, the model constructs a one-dimensional convolutional neural network based on the local correlation and translation invariance of the spectral waveform of diffuse reflectance spectral response data (e.g., a spectral vector of dimension 1×1024). This channel performs multi-scale sliding convolution operations, specifically designed with parallel convolutional branches with different receptive fields: the first branch is configured with a smaller convolutional kernel (e.g., kernel size 3 or 5) to capture sharp characteristic absorption peaks on the spectral curve; these high-frequency details directly correspond to the vibrational modes of specific chemical bonds in the expanding agent. The second branch is configured with a larger convolutional kernel (e.g., kernel size 15 or 25) to extract broadband baseline bulges or scattering slopes; these low-frequency features reflect particle packing density and microscopic surface roughness. Through channel-dimensional splicing of multi-scale feature maps, the network can simultaneously perceive the "presence" (characteristic peaks) and "distribution" (scattering background) of chemical components, thereby generating a high-dimensional chemical component fingerprint. The spectral morphology distribution features mentioned here do not refer to macroscopic geometric coordinates, but rather to the response intensity distribution of the microscopic dispersion uniformity of expanding agent particles within the sampling spot volume along the spectral wavelength axis.

[0035] For the second channel, the model constructs a Long Short-Term Memory (LSTM) network to address the strong time dependence of triboacoustic emission signals (e.g., feature sequences divided into 50 time steps). This channel utilizes the forget gate, input gate, and output gate mechanisms within the LSTM unit to recursively process the time series of the acoustic emission signal. At each time step, the forget gate discards the historical information of particle collision noise from the previous moment, while the input gate updates the current frictional energy information. This gating mechanism enables the network to capture the long-term evolution trend of particle motion during stirring (e.g., from the violent disordered collisions at the beginning of dry mixing to the viscous shear flow after water is added), rather than just focusing on the instantaneous sound pressure amplitude. The second channel ultimately outputs the hidden layer state vector at the last moment, serving as rheological dynamic parameters characterizing the current stage of rheological state evolution.

[0036] By adopting a heterogeneous parallel architecture, the physical properties of the input data are precisely matched: the topological feature extraction capability of CNN is used to analyze the static spectral "fingerprint", and the temporal memory capability of LSTM is used to analyze the dynamic acoustic "story", thereby realizing a three-dimensional perception of the concrete mixing process from microscopic chemical composition to macroscopic physical rheology.

[0037] Furthermore, the first channel of the dual-channel feature extraction model extracts the peak morphology of the diffuse reflectance spectral response data to generate a chemical component fingerprint characterizing the distribution of the expanding agent. Specifically, this includes: locating the first characteristic absorption band corresponding to the characteristic chemical bonds of the active component of the expanding agent and the second characteristic absorption band corresponding to the matrix material in the diffuse reflectance spectral response data; performing differential operations on the first characteristic absorption band to calculate the slope of the first derivative and the curvature of the second derivative of the spectral response curve, generating morphological parameters characterizing the chemical bond strength; comparing and analyzing the morphological parameters with the absorbance baseline value of the second characteristic absorption band to eliminate chromatic background interference and generate a chemical component fingerprint corrected for baseline drift.

[0038] Specifically, in this step, the data processing of the first channel aims to extract pure chemical features from the raw spectrum, which is heavily affected by environmental interference. First, based on a pre-established spectral knowledge base, the system locks specific index intervals in the input diffuse reflectance spectral response data (e.g., a 1024-dimensional vector). For example, for commonly used calcium sulfoaluminate expansive agents, the system identifies the 1440nm to 1460nm band as the first characteristic absorption band, which mainly corresponds to the first overtone absorption peaks of hydroxyl (-OH) or water of crystallization in the expansive agent's crystal structure; simultaneously, it identifies the 1100nm to 1150nm band as the second characteristic absorption band, which is located in the non-sensitive region of the cement matrix material, and its absorbance mainly reflects the overall physical scattering level of the material (baseline background).

[0039] Subsequently, numerical differentiation operations are performed on the discrete spectral data points within the first characteristic absorption band. Specifically, a Savitzky-Golay convolution filter with a window width of 5 to 9 sampling points is used to calculate the first and second derivatives of the spectral curves. The first derivative extracts the slope change of the curve, effectively removing the constant additive baseline shift in the spectrum; the second derivative extracts the curvature of the curve (i.e., the sharpness of the peaks), which can not only remove linear baseline drift but also separate overlapping peaks, generating a set of high-dimensional morphological parameter vectors.

[0040] Finally, a comparative analysis is performed to eliminate chromatic background interference (i.e., multiplicative optical path error caused by the dryness or color depth of the material). The system calculates the arithmetic mean of absorbance within the second characteristic absorption band and uses it as the current dynamic baseline value. Each element in the previously generated morphological parameter vector is divided by this dynamic baseline value, and a normalization operation is performed. This process is similar to the "internal standard method," which offsets the influence of overall reflectance fluctuations by comparing the relative changes between the characteristic band and the matrix band. The vector output after this processing is the baseline-drift corrected chemical composition fingerprint. The value of this fingerprint is only related to the chemical bonding strength (i.e., effective concentration) of the expanding agent and is decoupled from the light conditions of the stirring environment.

[0041] By introducing differential morphological analysis and a dual-band comparison mechanism, mathematical methods are cleverly used to eliminate baseline drift and multiplicative scattering noise commonly found in spectral data. This enables the model to accurately capture trace chemical characteristics of expansion agents even in complex environments such as concrete mixing where there is high dust and unstable lighting.

[0042] Furthermore, the second channel performs rheological state evolution analysis on the triboacoustic emission signal to generate rheological dynamic parameters characterizing the particle motion state. Specifically, this includes: performing a short-time Fourier transform on the triboacoustic emission signal to construct a time-frequency distribution matrix reflecting frequency changes over time; extracting the energy proportion of the ultrasonic frequency band friction component from the time-frequency distribution matrix and calculating the signal energy entropy within the sliding time window; monitoring the time-series change rate of the signal energy entropy, and when the change rate approaches zero, extracting the current signal energy entropy value as a rheological dynamic parameter characterizing the particle's transition from a collision state to a steady-state shear state.

[0043] Specifically, this step aims to quantify the physical process of concrete material transitioning from "disordered collisions of dry materials" to "ordered flow of slurry" through acoustic characteristics. First, the processing unit performs a short-time Fourier transform on the input time-domain triboacoustic emission signal (e.g., a discrete sequence with a sampling rate of 1 MHz). In practice, a Hamming window with 256 sampling points is used as the window function, with a 50% window overlap rate, mapping the one-dimensional time-domain signal into a two-dimensional time-frequency distribution matrix. The row indices of this matrix correspond to frequencies (0 Hz to 500 kHz), the column indices correspond to mixing time steps, and the values ​​of the matrix elements represent the energy spectral density at a specific time-frequency point.

[0044] Subsequently, a row index range of 50kHz to 300kHz is defined in the time-frequency distribution matrix, corresponding to the ultrasonic friction component excited by rigid friction between aggregate particles. The system first calculates the energy value of each frequency component within this band and normalizes it to construct an energy probability distribution sequence with a sum of 1. Next, the signal energy entropy within the sliding time window is calculated based on information theory principles. The calculation process does not directly use formulas but performs the following logical operations: for each probability value in the above energy probability distribution sequence, its base-2 logarithm is calculated, the probability value is multiplied by the corresponding logarithm to obtain a product term, and finally, all product terms are summed and their negatives are taken to obtain the instantaneous signal energy entropy. At the physical level, this entropy value reflects the degree of disorder in the particle motion system. In the initial stage of mixing, the aggregate undergoes random collisional motion, resulting in a chaotic and broad spectral distribution, leading to a high calculated entropy value. As the hydration reaction proceeds, the slurry encapsulates the aggregate to form a viscous fluid, and the particle motion transforms into directional laminar shear, with the spectral energy concentrated in a specific frequency band, causing the entropy value to gradually decrease.

[0045] Finally, the system constructs a differential observer to monitor the time series change rate of signal energy entropy. Specifically, it calculates the difference in signal energy entropy between adjacent time steps. If the absolute value of this difference is consistently lower than a preset minimum threshold within a continuous preset time period (e.g., 5 seconds), and the synchronously monitored mixer motor load current is within a preset effective load range (to exclude pseudo-steady-state signals caused by mixing blade slippage or idling), it is determined that the change rate approaches zero and the mixing is effective. At this point, it indicates that the mixing process has crossed the unsteady-state region and entered the rheological steady state. The system extracts the current signal energy entropy value as a rheological dynamic parameter. This parameter not only marks the physical endpoint of mixing, but its value also implies the viscosity information of the concrete (generally, the higher the viscosity, the stronger the fluid damping effect, and the lower the steady-state entropy value).

[0046] By using time-frequency energy entropy as a fingerprint of rheological state, the blindness of relying solely on mixing time control is overcome. By capturing the inflection point of entropy convergence, the phase transition moment of concrete from "collision state" to "shear state" in non-Newtonian fluid behavior can be accurately identified, ensuring that subsequent feature fusion is carried out under the premise of stable material rheological state.

[0047] Furthermore, the step of dynamically weighting the chemical component fingerprint using rheological kinetic parameters as state gating signals to generate a dry-state mixing characteristic index reflecting the real-time mixing uniformity of materials specifically includes: constructing a mixing uniformity confidence interval based on the rheological kinetic parameters; if the current rheological kinetic parameters are outside the mixing uniformity confidence interval, it is determined to be an unsteady-state stirring period, and a signal attenuation weighting coefficient is generated; and feature fusion is performed, the feature fusion including: constructing a feature suppression gating vector adapted to the dimension of the chemical component fingerprint using the signal attenuation weighting coefficient, performing an element-wise multiplication weighting operation on the chemical component fingerprint using the feature suppression gating vector to suppress spectral fluctuation noise during the unsteady-state stirring period, and concatenating the data dimensions of the chemical component fingerprint and the rheological kinetic parameters to generate a dry-state mixing characteristic index.

[0048] Specifically, the core logic of this step lies in establishing a physical perception gating mechanism, using the reliability of the rheological state in the acoustic dimension to control the input amount of spectral features in the chemical dimension. In the data processing flow, the system first retrieves the standard mixing rheological benchmark value stored in the database and sets the confidence interval for mixing uniformity. This interval is constructed based on the statistical distribution of the rheological dynamic parameters (i.e., the aforementioned steady-state signal energy entropy) of a large number of qualified batches of concrete at the end of mixing, for example, setting the interval range to [0.1, 0.3].

[0049] Subsequently, the system maps the real-time calculated rheological parameters (scalar values) to this interval for discrimination. If the current parameter value is 0.8, which clearly falls outside the confidence interval, the system determines that the mixer is currently in a period of intense "unsteady mixing," at which point the material distribution is extremely uneven and the spectral data contains a large amount of random noise. Based on this determination, the system generates a signal attenuation weighting coefficient with a value between 0 and 1. The calculation of this coefficient adopts a distance-based attenuation mapping logic: first, the absolute value of the difference between the current rheological parameter and the boundary value of the confidence interval is calculated, and this absolute value is used as the denominator to construct an attenuation function. The further the rheological parameter deviates from the confidence interval, the smaller the function output value becomes and the closer it is to zero, and the closer the generated weighting coefficient is to 0 (e.g., 0.05). Conversely, if the parameter falls within the interval, the weighting coefficient is set to 1.

[0050] In the feature fusion stage, to achieve dimensional alignment, the system replicates and extends the scalar signal attenuation weight coefficients in terms of dimensions to construct a feature suppression gating vector with a length completely consistent with the chemical component fingerprint (e.g., 1024 dimensions). Then, element-wise multiplication is performed, multiplying the corresponding elements of the 1024-dimensional chemical component fingerprint vector with the corresponding elements of the 1024-dimensional feature suppression gating vector. Through this operation, the spectral feature values ​​collected during the unsteady stirring period are significantly compressed, thereby mathematically suppressing the influence of low-confidence data on the model. Finally, the weighted chemical component fingerprint vector is concatenated with the original rheological dynamic parameters (1-dimensional) in the channel dimension to generate a dry mixing feature index with a total dimension of 1025.

[0051] A cross-modal "soft attention" mechanism was constructed, which uses acoustic signals to determine the physical state of stirring and then adaptively adjusts the weight of spectrochemical signals. This effectively avoids the "dirty data" pollution of the model caused by uneven mixing of materials in the early stage of stirring, and significantly improves the robustness of feature representation.

[0052] Furthermore, the step of correcting the dry mixing characteristic index for energy consistency based on the thermodynamic dissipation equilibrium constraint of particle fluid dynamics to generate an effective expansion potential energy characteristic that characterizes the material properties specifically includes: constructing a thermodynamic dissipation equilibrium constraint function, which defines the energy conversion residual between the input mechanical stirring work, the inverted theoretical chemical potential energy, and the frictional heat dissipation; performing gradient descent iteration during the feature mapping process to minimize the energy conversion residual, so that the characteristic distribution of the dry mixing characteristic index conforms to the energy flow boundary constraint of the non-equilibrium thermodynamic system, thereby generating an effective expansion potential energy characteristic corrected by physical laws.

[0053] Specifically, refer to Figure 3This step aims to eliminate the "physical illusion" that pure data-driven models may produce in sparse sample regions, that is, to prevent the model from outputting predictions that violate the law of conservation of energy.

[0054] First, this function aims to quantify the degree of deviation from the physical law of energy conservation. Its value is obtained by calculating the square of the difference between "total energy input" and "total energy conversion and consumption." The specific calculation logic is as follows: Integrating the product over time yields the input mechanical stirring work, which is then multiplied by a preset mechanical transmission efficiency coefficient to compensate for mechanical losses and heat dissipation in the non-closed system of the mixer, resulting in the effective input work. Second, the dry-state mixing characteristic index vector is input into a preset linear mapping layer, and the theoretical chemical potential energy is derived through weighted summation. Third, the sum of squares of the amplitude values ​​of the triboacoustic emission signal in the ultrasonic band is calculated, and this sum is multiplied by a preset acoustic-thermal conversion coefficient to obtain the frictional heat dissipation. Finally, the square of the difference between the effective input work and the sum of the theoretical chemical potential energy and the frictional heat dissipation is calculated, and this squared value is defined as the energy conversion residual. Gradient descent iterations are performed during the characteristic mapping process to minimize the aforementioned energy conversion residual, adjusting the value of the dry-state mixing characteristic index until the residual is less than a preset convergence threshold, generating an effective expansion potential energy characteristic corrected by physical laws.

[0055] As a preferred embodiment, the step of constructing the thermodynamic dissipation equilibrium constraint function specifically includes: establishing a nonlinear mapping relationship between the frictional heat dissipation coefficient and the rheological kinetic parameters; dynamically adjusting the energy conversion efficiency weight in the thermodynamic dissipation equilibrium constraint function according to the current rheological kinetic parameters, increasing the weight of the solid-phase frictional dissipation term when the rheological kinetic parameters indicate a collision state, and increasing the weight of the liquid-phase viscous dissipation term when the rheological kinetic parameters indicate a steady-state shear state, thereby generating a rheologically adaptive energy flow boundary constraint.

[0056] Specifically, to achieve the aforementioned dynamic adjustment, the system labels the rheological parameters output from the second channel as λ and normalizes them to a range of zero to one. The system pre-configures a dynamic adjustment logic, setting a solid-state friction coefficient constant and a fluid viscosity coefficient constant. Using the rheological parameter λ as a weight, it performs a weighted average operation on the solid-state friction coefficient constant and the fluid viscosity coefficient constant to calculate the current dynamic dissipation coefficient. Specifically, the dynamic dissipation coefficient equals the difference between the fluid viscosity coefficient constant multiplied by λ and the solid-state friction coefficient constant multiplied by one minus λ. The constraint function uses this dynamic dissipation coefficient to calculate the current theoretical upper limit of frictional heat dissipation in real time, no longer using a single fixed physical boundary, but constructing a dynamic energy envelope that evolves with the stirring process.

[0057] This step addresses the problem of mismatched physical model parameters caused by different mixing stages by associating the parameters in the physical constraint function with the real-time measured rheological state, enabling the energy correction process to accurately adapt to the continuous change of concrete from dry to wet state.

[0058] In the phase of generating effective expansion potential energy features, the system does not directly output the results generated by the forward propagation of the neural network. Instead, it performs an optimization-based correction operation. The system inputs the dry-state mixed feature index generated in the previous step into a feature mapping layer to obtain the feature vector of the initial stage. The energy conversion residual corresponding to this initial vector is calculated. If the residual exceeds the preset physically acceptable range, the system starts the gradient descent iterative algorithm. In this iteration phase, the system calculates the gradient direction of the residual relative to each element of the feature vector and fine-tunes the value of the feature vector along the direction of residual reduction. This is equivalent to applying a physical force field to the feature space, forcibly pulling the feature points back to the manifold surface that follows the law of conservation of energy. After several iterations (e.g., 5 to 10 times), when the energy conversion residual converges to a minimum value, the feature vector at this time is the effective expansion potential energy feature adjusted by physical laws. This feature is consistent with the distribution of the training samples in terms of data statistics and truly adheres to the boundary constraints of energy flow in terms of physical properties, that is, it ensures that the predicted expansion potential energy will not exceed the amount of actual input effective mechanical work.

[0059] This method introduces a thermodynamic physical filter to restrict the solution space of the black-box model to the range allowed by physical laws, effectively eliminating abnormal features caused by data noise or overfitting, and significantly improving the model's generalization ability and prediction reliability under unseen conditions.

[0060] As a preferred embodiment, before performing multi-parameter nonlinear inversion on the effective expansion potential energy characteristics, the method further includes performing pre-hydration activity loss compensation: performing time differentiation on the chemical component fingerprints output by the dual-channel feature extraction model at continuous time steps to calculate the decay rate of the expansion agent characteristic peak intensity; integrating the decay rate over the stirring time to estimate the pre-consumed chemical potential energy lost due to early hydration reactions during the stirring cycle; subtracting the pre-consumed chemical potential energy from the effective expansion potential energy characteristics to generate net effective expansion potential energy characteristics, and inputting them into the nonlinear regression mapping space.

[0061] Specifically, the expansive agent begins a chemical reaction the moment it comes into contact with water, and some of its chemical potential energy is consumed within the mixer, failing to contribute to the limiting expansion rate of the hardened concrete. The system tracks the changes in the chemical component fingerprint (i.e., the spectral feature vector) over time, calculating the negative first derivative of the intensity of characteristic absorption peaks (e.g., 1450 nm). This derivative directly reflects the hydration consumption rate of the expansive agent. The system integrates this consumption rate throughout the entire wet-mixing cycle, quantifying the lost "dead potential energy" (e.g., calculating that approximately 5% of the expansive agent has reacted prematurely). Then, the system performs a compensation calculation, subtracting this pre-consumed chemical potential energy from the physically corrected effective expansion potential energy characteristics to obtain the "net potential energy" truly used for later expansion. This step effectively prevents overestimation bias caused by the model's failure to identify "ineffective reserves."

[0062] This method, by exploring the time-varying characteristics of material chemical reactions, introduces an activity loss compensation mechanism, eliminates the systematic error in prediction caused by early hydration during the mixing process, and significantly improves the accuracy of prediction for highly reactive special concrete.

[0063] Furthermore, the step of performing multi-parameter nonlinear inversion on the effective expansion potential energy characteristics to obtain the predicted value of the restricted expansion rate of the premixed shrinkage-compensating concrete specifically includes: inputting the effective expansion potential energy characteristics corrected by the energy conservation constraint into the nonlinear regression mapping space; performing multi-objective solution in the nonlinear regression mapping space, and simultaneously outputting the predicted value of the restricted expansion rate corresponding to the standard curing age and the mixing uniformity index characterizing the dispersion degree of the premixed shrinkage-compensating concrete components.

[0064] Specifically, this step utilizes the general approximation capability of deep neural networks to establish a nonlinear mapping relationship from microscopic features to macroscopic performance. The system first inputs the effective expansion potential energy features (e.g., a dense vector of 1025 dimensions) generated in the previous step and verified by physical energy into the nonlinear regression mapping space. This mapping space is specifically constructed as a multi-layer fully connected neural network containing several hidden layers (e.g., 3 layers with 512, 256, and 128 neurons respectively). The layers are connected by nonlinear activation functions (such as Leaky ReLU) to simulate the complex physicochemical nonlinear evolution mechanism during the hardening process of concrete materials.

[0065] At the network output, a parallel branch structure for multi-objective solution was designed: the first branch focuses on performance prediction, outputting a single floating-point value through a regression layer, namely the predicted value of the restricted expansion rate of the premixed shrinkage-compensating concrete. This predicted value corresponds to the longitudinal restricted expansion rate under a specific curing age (e.g., 14 days of water curing) specified by national standards, expressed as a percentage (%). The model establishes a potential correlation between the characteristics of the mixing state and the performance of the hardened state by learning from a large amount of historical batch data. The second branch focuses on quality assessment, simultaneously outputting a mixing uniformity index, which is a confidence scalar between 0 and 1, used to quantify the degree of dispersion of each component (especially the expansion agent) in the current mixing batch on a macro scale. This index is not a direct measurement value, but a quality evaluation index obtained by the model based on the noise level and feature consistency in the input acoustic-optical fusion features.

[0066] In actual operation, these two objectives are jointly solved under the constraints of the loss function. For example, if the model predicts a low mixing uniformity index (indicating uneven mixing), the model will automatically adjust and limit the output range of the expansion rate prediction value, or output a prediction result with high uncertainty. This multi-parameter nonlinear inversion mechanism ensures that the prediction result not only contains numerical values, but also implicitly includes a self-evaluation of the credibility of the numerical value.

[0067] By adopting a multi-objective collaborative inversion strategy, not only can the key performance index of concrete (limited expansion rate) be predicted in advance, but auxiliary index (mixing uniformity) reflecting production quality is also provided simultaneously. This provides operators with a dual basis for decision-making on "whether the performance meets the standard" and "whether the production is under control", avoiding prediction deviations caused by fluctuations in mixing quality.

[0068] Furthermore, it also includes the step of using the diffuse reflectance spectral response data for environmental adaptive calibration: extracting the atmospheric carbon dioxide characteristic absorption peak corresponding to the atmospheric medium absorption characteristics of the optical path in the diffuse reflectance spectral response data, and using the atmospheric carbon dioxide characteristic absorption peak as the wavelength drift calibration point; calculating the offset between the actual detection wavelength of the characteristic absorption peak and the standard physical wavelength, and using the offset to dynamically calibrate the wavelength axis of the diffuse reflectance spectral response data.

[0069] Specifically, this method does not rely on external etalons (such as holmium oxide glass), but instead utilizes the atmospheric medium inherent in the optical path as a constant physical reference. During each acquisition of diffuse reflectance spectral response data, the system performs superposition and averaging of multiple acquisitions of diffuse reflectance spectral response data during the mixer's idling phase or low-load intervals. This enhances the signal-to-noise ratio and extracts the atmospheric carbon dioxide characteristic absorption peak corresponding to the atmospheric medium absorption characteristics in the background optical path, using this characteristic absorption peak as the wavelength drift calibration point. Instead of calculating the extrema of the original discrete pixels acquired in a single acquisition, the system performs Gaussian function fitting on the superimposed and averaged spectral curve within the sensitive region to calculate the center wavelength of the currently measured characteristic absorption peak. To accurately capture the position of this peak, the system does not calculate the extrema of the original discrete pixels, but instead performs Gaussian function fitting or centroid peak finding on the spectral curve within the sensitive region, thereby calculating the center wavelength of the currently measured characteristic absorption peak with sub-pixel accuracy (e.g., a calculated result of 1573.8 nm).

[0070] The system retrieves a pre-stored database of standard physical wavelengths and searches for the theoretical center wavelength of carbon dioxide in this band (for example, the standard physical value is 1572.5 nm). The system calculates the algebraic difference between the measured center wavelength and the standard physical wavelength to obtain the current wavelength shift (in this example, +1.3 nm).

[0071] Finally, the wavelength drift is used to dynamically calibrate the wavelength axis of the diffuse reflectance spectral response data. In practice, the system does not change the hardware settings of the spectrometer, but performs a reverse translation operation on the wavelength index array of the spectral data, subtracting the wavelength drift from the calibration wavelength corresponding to each pixel, and re-establishing the mapping relationship between the pixel index and the wavelength value. For nonlinear drift cases, the system can also select the water vapor absorption peak in the optical path (such as around 1380nm) as the second calibration point to perform multi-point linear regression calibration.

[0072] By leveraging the ubiquitous nature of air, the measurement environment itself is transformed into a calibration tool, enabling all-time, maintenance-free, and self-healing wavelength correction of the spectrometer. This ensures that the spectral characteristics of the input model remain strictly aligned with the wavelength reference used during model training, even under harsh operating conditions.

[0073] Furthermore, the calculation basis for the predicted value of the restricted expansion rate also includes: pre-constructing a ternary association database between the expansion agent dosage, water-cement ratio and the restricted expansion rate; performing similarity matching on the effective expansion potential energy feature in the ternary association database to retrieve reference association nodes with similar features; calculating the weight based on the similarity between the effective expansion potential energy feature and the reference association node, and performing weighted interpolation on the correction coefficient corresponding to the reference association node to obtain the correction coefficient of the predicted value of the restricted expansion rate.

[0074] Specifically, this step introduces a memory-based reinforcement learning strategy that uses historical operating data to post-process and correct the prediction results of the deep learning model.

[0075] First, the system maintains a dynamically updated ternary relational database in the local storage unit. Each data node in this database corresponds to a historical production batch, and each node contains three sets of core data: the first set is physical parameters, namely the actual recorded amount of expanding agent in that batch (kg / m³). 3 The first group consists of three sets of data: the water-cement ratio; the second group is the true performance value, which is the actual limited expansion rate measured in the laboratory after 14 days of standard curing; the third group is the feature index, which is the effective expansion potential energy feature vector generated by the model of this invention at the end of stirring for this batch. In addition, each node also stores a correction coefficient, which is defined as the ratio between the historical true value and the model's prediction at that time.

[0076] During the real-time prediction phase, the system uses the currently generated effective expansion potential energy features (e.g., a 1025-dimensional vector) as the query key and performs a similarity search in the database. Specifically, the k-nearest neighbor algorithm is used to calculate the cosine similarity between the query vector and all historical feature index vectors in the database. The system sorts the nodes based on similarity and selects the top k (e.g., k=5) most similar reference nodes. These nodes represent the historical conditions that are most physically similar to the current stirring state.

[0077] Subsequently, the system performs a weighted interpolation operation. The similarity values ​​corresponding to these k reference nodes are normalized to obtain a weight vector that sums to 1. This weight vector is then used to calculate the weighted average of the historical correction coefficients stored in these k nodes, yielding the correction coefficient for the current confined expansion rate prediction. Finally, the original prediction value output from the aforementioned nonlinear inversion step is multiplied by this correction coefficient to obtain the final output confined expansion rate.

[0078] By introducing a "historical memory retrospective" mechanism, known historical true values ​​are used to calibrate unknown current predictions, effectively compensating for systematic deviations caused by batch fluctuations in raw materials (such as changes in cement activity). This ensures that the prediction results not only conform to the logical deduction of the model, but also to long-term engineering statistical laws.

[0079] The deep learning-based concrete mixing performance prediction method provided in this embodiment successfully acquires high-quality dual-modal data that accurately reflects the microscopic chemical and physical state of materials in the extreme environment of high dust and high noise inside a concrete mixer by constructing an aero-optical transient window and utilizing structural waveguide acoustic filtering. This overcomes the shortcomings of traditional sensors that are easily obscured by dust and interfered with by environmental noise. Simultaneously, this method designs a heterogeneous dual-channel gating architecture, using the rheological dynamic state sensed by acoustic signals to dynamically weight the spectrochemical signals, ensuring that the model extracts chemical fingerprints only within a high-confidence window of uniform material mixing, effectively eliminating random noise caused by the discrete distribution of materials in the early stages of mixing. Furthermore, unlike traditional black-box models, this method introduces thermodynamic dissipation equilibrium constraints based on particle fluid dynamics, forcing the feature mapping process to obey the law of energy conservation, ensuring the physical interpretability of the prediction results and improving generalization ability. Finally, through a multi-objective nonlinear inversion and historical data correction mechanism, this method advances the limitation expansion rate index, which originally required waiting 14 days, to the mixing stage for real-time inversion, achieving a fundamental shift from "post-event verification" to "process control."

[0080] Example 2: This embodiment applies the aforementioned deep learning-based concrete mixing performance prediction method to the prefabrication production scenario of CRTSⅢ type slab track concrete for high-speed railways. In this specific application scenario, the concrete design strength grade is C60, and the process specifications have extremely stringent requirements for limiting the expansion rate, which must be strictly controlled within a narrow window of 0.020% to 0.035% to prevent micro-cracks or arching deformation of the track slab due to temperature difference loads during later service.

[0081] In the specific implementation process, the system first deploys a high-pressure air curtain generator and a multimodal sensing terminal on a twin-shaft forced mixer. When the mixing cycle enters the steady-state mixing stage after water is added, the control unit triggers a synchronous acquisition command, driving the high-pressure air curtain generator to spray dried compressed air at a pressure of 0.75MPa, constructing a local laminar flow air shield with a thickness of about 50mm in front of the observation window where the material is tumbling violently. This air shield can instantly clear the dust cloud that obstructs the view, forming a transient transparent window in the aero-optical coupled flow environment. Meanwhile, the near-infrared spectral intelligent sensor collects raw light intensity signals in the 900nm to 1700nm band at a frequency of 10Hz through the transient transparent window, and performs dark current compensation and whiteboard calibration through the built-in edge computing module to generate a 1024-dimensional diffuse reflectance spectral response data vector. Simultaneously, the acoustic emission intelligent sensor coupled to the bottom of the mixer with a magnetic base performs sound wave acquisition, uses an analog bandpass filter to filter out mechanical vibrations and aerodynamic noise below 20kHz, retains only the signal in the 50kHz to 300kHz frequency band, and extracts a 4096-point triboacoustic emission signal sequence aligned with the time of the spectral acquisition window.

[0082] The acquired heterogeneous data is transmitted in real time to an edge computing workstation and input into a pre-trained dual-channel feature extraction model for parallel analysis. In the first channel of the model, a three-layer one-dimensional convolutional neural network performs multi-scale sliding convolution operations on the input diffuse reflectance spectral response data, focusing on extracting the characteristic absorption peak sharpness of key components of the expansion agent (such as calcium sulfoaluminate) at 1450 nm and the background scattering baseline of the matrix cement at 1100 nm. Through global max pooling, the high-dimensional spectral data is compressed into a 256-dimensional chemical composition fingerprint, which accurately characterizes the chemical bonding density of the expansion agent within the current micro-element. In the second channel of the model, a bidirectional long short-term memory network performs time-series analysis on the triboacoustic emission signal, calculating the energy entropy evolution curve of the signal in the ultrasonic band. When the rate of change of energy entropy is detected to be lower than a preset threshold for 10 consecutive time steps, it is determined that the aggregate motion has changed from disordered collision to ordered laminar flow, and a normalized scalar value is output as a rheological dynamics parameter to characterize the stability of the current material rheological state.

[0083] Subsequently, the system performs feature fusion and physical correction operations. Using the rheological parameters output from the second channel as state gating signals, a feature suppression gating vector adapted to the chemical component fingerprint dimension is generated, and an element-wise multiplication weighted operation is performed on the chemical component fingerprint. If the rheological parameters indicate that the current state is unsteady stirring, the gating vector will automatically attenuate the weights of the chemical features, thereby suppressing spectral fluctuation noise caused by uneven material distribution. The generated dry-state mixing feature index is then sent to the physical constraint module. In this module, the system constructs a thermodynamic dissipation equilibrium constraint function based on particle fluid dynamics, calculates the difference between the mechanical stirring work input by the motor and the frictional heat dissipation, and sets it as the physical boundary of the theoretical chemical potential energy. If the energy characteristics mapped by several state mixing feature indices violate this energy conservation boundary, the system iteratively fine-tunes the feature values ​​through gradient descent to generate an effective expansion potential energy feature corrected by physical laws.

[0084] Finally, the system performs multi-parameter nonlinear inversion on the effective expansion potential energy characteristics. The physically corrected feature vector is input into a fully connected neural network, which simultaneously calculates the predicted value of the limiting expansion rate (e.g., 0.028%) corresponding to the standard curing period of 14 days, as well as the mixing uniformity index characterizing the degree of component dispersion. To further improve the prediction accuracy, the system automatically retrieves reference nodes with similar raw material sources and the same mix proportions from the historical database. Based on the similarity of the effective expansion potential energy characteristics, it calculates weights and performs weighted interpolation on the historical correction coefficients. This coefficient is then used to perform the final calibration of the model's output limiting expansion rate prediction value. The prediction result is transmitted to the central control room in real time, serving as a key basis for determining whether this batch of C60 concrete meets the CRTS III type precast requirements, thus realizing a closed-loop process from microscopic physical perception to macroscopic engineering decision-making.

[0085] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the scope of protection defined in the claims.

Claims

1. A method for predicting concrete mixing performance based on deep learning, characterized in that, include: The diffuse reflectance spectral response data of premixed shrinkage-compensating concrete under aero-optical coupled flow environment were obtained, and the frictional acoustic emission signal of the concrete during the mixing process was also obtained simultaneously. The diffuse reflectance spectral response data and the triboacoustic emission signal are input into a dual-channel feature extraction model for parallel signal analysis; the first channel of the dual-channel feature extraction model extracts the peak shape of the diffuse reflectance spectral response data to generate a chemical component fingerprint characterizing the distribution of the expanding agent; The second channel performs rheological state evolution analysis on the triboacoustic emission signal to generate rheological dynamic parameters characterizing the particle motion state; By using rheological dynamic parameters as state gating signals to dynamically weight the chemical component fingerprint, a dry mixing characteristic index reflecting the real-time mixing uniformity of materials is generated. By performing energy consistency correction on the dry mixing characteristic index based on the thermodynamic dissipation equilibrium constraint of particle fluid dynamics, an effective expansion potential energy characteristic that characterizes material properties is generated. A multi-parameter nonlinear inversion is performed on the effective expansion potential energy characteristics to obtain the predicted value of the restricted expansion rate of concrete; based on the predicted value of the restricted expansion rate, the judgment result characterizing the concrete mixing performance is output.

2. The method for predicting concrete mixing performance based on deep learning according to claim 1, characterized in that, The process of acquiring the diffuse reflectance spectral response data and the triboacoustic emission signal includes: The sampling clocks of the high-pressure air curtain generator and the near-infrared spectral intelligent sensor are synchronized and triggered to construct a local laminar air shield that can eliminate dust interference in the acquisition optical path, forming a transient transparent window in the aero-optical coupled flow environment. The near-infrared spectral intelligent sensor is driven to perform photoelectric detection through the transient transparent window to acquire the original light intensity signal. The built-in edge computing module is used to perform dark current compensation and standardization processing on the original light intensity signal to generate diffuse reflectance spectral response data that includes material reflection characteristics and atmospheric medium absorption characteristics of the optical path. At the same time, the acoustic emission intelligent sensor with edge computing capabilities is driven to perform acoustic wave acquisition. The built-in front-end filtering module is used to perform structural waveguide signal separation to filter out aerodynamic noise and low-frequency mechanical vibration in the air coupling path, and retain the particle contact friction signal in the ultrasonic frequency band transmitted through the mixer wall to generate a friction acoustic emission signal.

3. The method for predicting concrete mixing performance based on deep learning according to claim 1, characterized in that, The dual-channel feature extraction model is constructed as a heterogeneous signal parallel analysis architecture, including: The first channel constructs a spatial morphology analysis flow path based on a one-dimensional convolutional neural network, performs multi-scale sliding convolution operations on the diffuse reflectance spectral response data, and extracts the spatial distribution characteristics of the expanding agent in the dry mix. The second channel constructs a temporal dynamic analysis flow path based on a long short-term memory network, performs time series dependency analysis on the triboacoustic emission signal, and captures the evolution law of particle motion state with stirring time.

4. The method for predicting concrete mixing performance based on deep learning according to claim 3, characterized in that, The first channel of the dual-channel feature extraction model extracts the peak shape of the diffuse reflectance spectral response data to generate a chemical component fingerprint characterizing the distribution of the expanding agent. Specifically, this includes the following steps: In the diffuse reflectance spectral response data, the first characteristic absorption band corresponding to the characteristic chemical bonds of the active component of the expanding agent and the second characteristic absorption band corresponding to the matrix material are located; differential operation is performed on the first characteristic absorption band to calculate the slope of the first derivative and the curvature of the second derivative of the spectral response curve, generating morphological parameters characterizing the chemical bond strength; the morphological parameters are compared and analyzed with the absorbance baseline value of the second characteristic absorption band to eliminate chromatic background interference and generate a chemical component fingerprint corrected for baseline drift.

5. The method for predicting concrete mixing performance based on deep learning according to claim 3, characterized in that, The second channel performs rheological state evolution analysis on the triboacoustic emission signal to generate rheological dynamic parameters characterizing the particle motion state. The specific steps include: A short-time Fourier transform is performed on the frictional acoustic emission signal to construct a time-frequency distribution matrix reflecting the frequency change over time. The energy proportion of the frictional component in the ultrasonic band is extracted from the time-frequency distribution matrix, and the signal energy entropy within the sliding time window is calculated. The time series change rate of the signal energy entropy is monitored, and when the change rate approaches zero, the current signal energy entropy value is extracted as a rheological dynamic parameter characterizing the transition of particles from a collision state to a steady-state shear state.

6. The method for predicting concrete mixing performance based on deep learning according to claim 5, characterized in that, The steps of dynamically weighting the chemical component fingerprint using rheological kinetic parameters as state gating signals to generate a dry-state mixing characteristic index reflecting the real-time mixing uniformity of materials specifically include: A mixing uniformity confidence interval is constructed based on the rheological parameters. If the current rheological parameters are outside the mixing uniformity confidence interval, it is determined to be an unsteady stirring period, and a signal attenuation weighting coefficient is generated. Feature fusion is then performed, which includes: constructing a feature suppression gating vector adapted to the chemical component fingerprint dimension using the signal attenuation weighting coefficient; performing an element-wise multiplication weighted operation on the chemical component fingerprint using the feature suppression gating vector to suppress spectral fluctuation noise during the unsteady stirring period; and concatenating the data dimensions of the chemical component fingerprint and the rheological parameters to generate a dry mixing feature index.

7. The method for predicting concrete mixing performance based on deep learning according to claim 1, characterized in that, The step of performing energy consistency correction on the dry mixing characteristic index based on the thermodynamic dissipation equilibrium constraint of particle fluid dynamics to generate an effective expansion potential energy characteristic characterizing material properties specifically includes: A thermodynamic dissipation equilibrium constraint function is constructed, which defines the energy conversion residual between the input mechanical stirring work, the inverted theoretical chemical potential energy, and the frictional heat dissipation. During the feature mapping process, gradient descent iteration is performed to minimize the energy conversion residual, so that the feature distribution of the dry mixing feature index conforms to the energy flow boundary constraint of the non-equilibrium thermodynamic system, and an effective expansion potential energy feature corrected by physical laws is generated.

8. The method for predicting concrete mixing performance based on deep learning according to claim 7, characterized in that, The step of performing multi-parameter nonlinear inversion on the effective expansion potential energy characteristics to obtain the predicted value of the restricted expansion rate of the premixed shrinkage-compensated concrete specifically includes: The effective expansion potential energy characteristics after correction by the energy conservation constraint are input into the nonlinear regression mapping space; multi-objective solution is performed in the nonlinear regression mapping space, and the predicted value of the restricted expansion rate corresponding to the standard curing age and the mixing uniformity index characterizing the dispersion degree of the premixed shrinkage-compensating concrete components are output simultaneously.

9. The method for predicting concrete mixing performance based on deep learning according to claim 1, characterized in that, It also includes the step of performing environmental adaptive calibration using the diffuse reflectance spectral response data: Extract the atmospheric carbon dioxide characteristic absorption peak corresponding to the atmospheric medium absorption characteristics of the optical path from the diffuse reflectance spectral response data, and use the atmospheric carbon dioxide characteristic absorption peak as the wavelength drift calibration point. The offset between the actual detection wavelength of the characteristic absorption peak and the standard physical wavelength is calculated, and the wavelength axis of the diffuse reflectance spectral response data is dynamically calibrated using the offset.

10. The method for predicting concrete mixing performance based on deep learning according to claim 1, characterized in that, The calculation basis for the predicted value of the restricted expansion rate also includes: A ternary association database is pre-constructed to link the expansion agent dosage, water-cement ratio, and restricted expansion rate. The effective expansion potential energy features are similarly matched in the ternary association database to retrieve reference association nodes with similar features. The weights are calculated based on the similarity between the effective expansion potential energy features and the reference association nodes. Weighted interpolation is then performed on the correction coefficients corresponding to the reference association nodes to obtain the correction coefficients for the predicted restricted expansion rate.

Images