Multi-sensor fusion system and method for controlling moisture content of cold recycled aggregate

Abstract

This invention relates to the field of automated control technology in road engineering, and discloses a multi-sensor fusion-based system and method for controlling the moisture content of cold recycled aggregates. The system includes: a multimodal sensing module; a data preprocessing and tensor construction module; a tensor analysis and state recognition module; a moisture content prediction module; an intelligent decision-making module; and a collaborative control execution module. The method includes: fusing and collecting multimodal parameters of the aggregate, preprocessing them to construct a high-dimensional time-series tensor; identifying the internal state and decoupling interference through tensor analysis, extracting pure signals, and predicting future moisture content; based on the predicted value and internal state, determining the amount of water added through reinforcement learning, and precisely executing closed-loop control of moisture content. This invention solves the problems of inaccurate measurement and control lag in traditional methods. Through prediction and state recognition, it achieves forward-looking and precise dynamic control of moisture content, significantly improving the stability and engineering quality of the cold recycling process.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for road engineering, specifically to a multi-sensor fusion system and method for controlling the moisture content of cold recycled mixtures. Background Technology

[0002] Cold recycling technology, as an economical and environmentally friendly road maintenance method, is widely used in pavement repair projects. In this technology, the moisture content of the cold recycled mixture is a core process parameter determining the final project quality, directly affecting the mixing effect, compaction performance, and structural strength of the mixture. Precise and stable control of the moisture content is crucial to ensuring the performance of cold recycled pavements.

[0003] Currently, moisture content control at construction sites largely relies on traditional methods. The mainstream approach involves feedback control based on operator experience, combined with single-point sensors such as microwaves or infrared sensors deployed on the production line. This means manually or semi-automatically adjusting the amount of water added based on a single reading provided by the sensor. This model, based on real-time measurement and experience-based judgment, is the most commonly used technical solution at present.

[0004] However, the aforementioned existing technologies have inherent limitations in application. The aggregate, temperature, and other components of cold recycled mixtures are complex and variable, easily interfering with measurements from a single sensor, leading to inaccurate readings and poor stability. Simultaneously, the control process from measurement to adjustment has a significant delay, making it difficult to cope with real-time fluctuations in raw materials, resulting in low accuracy in moisture content control. More importantly, these methods struggle to perceive key physicochemical states within the mixture, such as the demulsification of emulsified asphalt; their control strategies remain superficial, limiting the improvement of process refinement.

[0005] Therefore, this invention proposes a multi-sensor fusion-based system and method for controlling the moisture content of cold recycled mixtures to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a multi-sensor fusion system and method for controlling the moisture content of cold recycled aggregates. This solves the problems of low moisture content control accuracy, poor robustness, and untimely response caused by the susceptibility of single-point measurement to interference, delayed control feedback, and difficulty in understanding the internal physicochemical state of materials in traditional control methods.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-sensor fusion-based cold recycling mixture moisture content control system, comprising:

[0008] Multimodal sensing module: used to collect parameter information of at least two different modes of cold recycled mixture in real time through multiple sensors. The parameter information includes dielectric constant data, spectral reflectance data, surface temperature distribution data, volume profile and flow rate data, as well as ambient temperature and humidity.

[0009] Data preprocessing and tensor construction module: used to clean, normalize and impute missing values ​​of the real-time collected parameter information, and construct a fourth-order temporal tensor including the parameter information;

[0010] Tensor analysis and state recognition module: used to perform tensor representation learning on the fourth-order time series tensor to extract latent features, identify the internal component state of the cold recycled mixture based on the latent features, and decouple the interference caused by non-moisture content factors in the parameter information to obtain pure moisture content related signals;

[0011] Moisture content prediction module: used to adaptively predict the future moisture content of the cold recycled mixture based on the pure moisture content related signal;

[0012] Intelligent decision-making module: Based on the predicted future moisture content and the state of the internal components, it dynamically determines the water addition adjustment command required for the cold recycled mixture through causal inference and reinforcement learning;

[0013] Collaborative control execution module: used to adjust the amount of water added according to the water addition adjustment command to dynamically control the moisture content of the cold recycled mixture.

[0014] Preferably, the multimodal sensing module collects the parameter information in real time through the following sensors:

[0015] A broadband microwave sensor array deployed at the conveyor belt or mixer outlet of the cold recycled mixture is used to collect the dielectric constant data of the mixture.

[0016] A multispectral near-infrared sensor deployed above the surface of the cold recycled mixture is used to collect the spectral reflectance data of the mixture;

[0017] A high-precision thermal imaging sensor is used to monitor the surface temperature distribution of the cold recycled mixture;

[0018] A high-resolution laser displacement sensor is used to measure the volume profile and flow rate of the cold recycled mixture;

[0019] An ambient temperature and humidity sensor is used to monitor the ambient temperature and humidity at the construction site of cold recycled mixtures.

[0020] Preferably, the data preprocessing and tensor construction module includes:

[0021] The collected parameter information is cleaned, including outlier removal and noise filtering;

[0022] The parameter information is then subjected to Min-Max normalization.

[0023] The parameter information is temporally interpolated to fill in missing data points and constructed as a fourth-order temporal tensor. ;

[0024] In the formula, Indicates the number of sensor modes; Indicates the number of spatial sampling points of the sensor; Indicates the length of the time series; This represents the feature dimension corresponding to each sampling point.

[0025] Preferably, the tensor analysis and state recognition module includes:

[0026] Tensor representation learning is performed on the fourth-order time series tensor, and the core tensor and factor matrix are obtained through non-negative tensor factor decomposition. The core tensor and factor matrix represent the potential characteristics of free water, bound water, aggregate, asphalt content and temperature components in the cold recycled mixture.

[0027] Based on the core tensor and factor matrix, the internal component state of the cold recycled mixture is identified by tensor neural network. The internal component state includes the real-time demulsification degree of emulsified asphalt, free water content, and bound water content.

[0028] Based on the internal component state, tensor independent component analysis or orthogonal projection method is used to decouple the interference caused by non-moisture content factors from the fourth-order time series tensor to obtain the pure moisture content related signal.

[0029] Preferably, the step of obtaining the core tensor and factor matrix through nonnegative tensor factorization includes:

[0030] For the fourth-order temporal tensor Perform a nonnegative Tucker decomposition, which is expressed as:

[0031] ;

[0032] In the formula, For the core tensor; , , and These are factor matrices corresponding to sensor modes, spatial sampling points, time series, and feature dimensions, respectively. Indicates the tensor along the first Modulus product, modal index These correspond to sensor modes, spatial sampling points, time series, and feature dimensions, respectively.

[0033] Preferably, the moisture content prediction module includes:

[0034] The purified moisture content-related signal is used to construct an input time series.

[0035] The input time series is processed by a gated recurrent unit network with an integrated attention mechanism to learn temporal dependencies and generate weighted hidden state representations;

[0036] The weighted hidden state representation is input to a fully connected output layer to generate and output the predicted future moisture content of the cold recycled mixture at one or more future time points.

[0037] Preferably, the weighted hidden state representation Calculated using the following formula:

[0038] ;

[0039] In the formula, For the gated cyclic unit network at time step The hidden state of the output; The hidden state calculated for the attention mechanism The weights; The length of the input time series.

[0040] Preferably, the intelligent decision-making module includes:

[0041] A structural causal model is constructed and updated using a causal inference method. The structural causal model is used to quantify the causal effects of the internal component state, historical water addition, and environmental parameters on the water content.

[0042] The reinforcement learning agent is based on a state composed of the predicted future moisture content, the internal component state, and the causal effect. With the optimization objective of minimizing the deviation between the predicted future moisture content and the preset target moisture content, it dynamically makes decisions and outputs the water addition adjustment command through a deep Q network.

[0043] Preferably, the collaborative control execution module includes:

[0044] Receive the water addition adjustment command and parse the water addition adjustment command into a target flow rate setting value;

[0045] The actuator control signal is generated based on the deviation between the target flow rate setpoint and the actual flow rate value fed back by the water flow meter in real time using a proportional-integral-derivative controller.

[0046] According to the actuator control signal, the electronic control valve or variable frequency water pump is driven to precisely adjust the real-time water addition to the cold recycled mixture, so as to dynamically regulate the moisture content.

[0047] This invention also provides a method for controlling the moisture content of cold recycled mixtures based on multi-sensor fusion, comprising the following steps:

[0048] The parameters of at least two different modes of cold recycled mixture are collected in real time by multiple sensors. The parameters include dielectric constant data, spectral reflectance data, surface temperature distribution data, volume profile and flow rate data, as well as ambient temperature and humidity.

[0049] The collected parameter information is cleaned, normalized, and missing values ​​are imputed. A fourth-order temporal tensor including the parameter information is then constructed.

[0050] Tensor representation learning is performed on the fourth-order time series tensor to extract latent features. Based on the latent features, the internal component state of the cold recycled mixture is identified, and the interference caused by non-moisture content factors in the parameter information is decoupled to obtain a pure moisture content related signal.

[0051] Based on the pure moisture content-related signal, the future moisture content of the cold recycled mixture is adaptively predicted.

[0052] Based on the predicted future moisture content and the internal component state, the required water addition adjustment command for the cold recycled mixture is dynamically determined through causal inference and reinforcement learning.

[0053] According to the water addition adjustment instruction, the water addition is adjusted to dynamically control the moisture content of the cold recycled mixture.

[0054] This invention provides a multi-sensor fusion-based system and method for controlling the moisture content of cold recycled aggregates. It offers the following advantages:

[0055] 1. This invention, by setting up a multimodal sensing module and a tensor analysis and state recognition module, utilizes data such as dielectric constant and spectral reflectance collected from multiple sensors to construct a high-dimensional time-series tensor, and employs tensor analysis methods for in-depth data processing. This approach effectively decouples interference from non-moisture content factors such as aggregates and temperature in the mixture, extracting pure moisture content-related signals. It fundamentally solves the problems of traditional single-point or single-modal measurement methods being susceptible to complex on-site conditions and having poor robustness, significantly improving the accuracy and reliability of moisture content monitoring in cold recycled mixtures.

[0056] 2. This invention constructs a moisture content prediction module and an intelligent decision-making module. Based on pure moisture content-related signals, it adaptively predicts the future moisture content of the mixture and formulates water addition adjustment instructions by combining causal inference and reinforcement learning. This achieves a shift from a "passive response" to an "active prediction" control mode, effectively overcoming the inherent delays caused by material transportation and moisture penetration in the production line. This makes dynamic control more forward-looking and accurate, thereby ensuring that the moisture content of the final mixture can remain stable within the target range in the long term.

[0057] 3. This invention identifies the internal component states of the mixture through tensor analysis and state recognition modules, enabling online monitoring of key process indicators such as the real-time demulsification degree of emulsified asphalt and the content of free and bound water. This allows the system to not only adjust the total moisture content but also adaptively optimize and control based on the inherent physicochemical changes of the mixture. This control strategy based on deep state understanding significantly improves the level of refined management of the cold recycling process, ensuring the final performance of the mixture and the quality of the project. Attached Figure Description

[0058] Figure 1 This is a system architecture diagram of the present invention;

[0059] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0060] The technical solutions in 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.

[0061] Please see the appendix Figure 1 This invention provides a multi-sensor fusion-based cold recycling mixture moisture content control system, comprising:

[0062] Multimodal sensing module: used to collect parameter information of at least two different modes of cold recycled mixture in real time through multiple sensors. The parameter information includes dielectric constant data, spectral reflectance data, surface temperature distribution data, volume profile and flow velocity data, as well as ambient temperature and humidity.

[0063] In this embodiment, the core task of the multimodal sensing module is to construct a comprehensive, three-dimensional, and high-dimensional real-time data stream, providing rich and reliable input for subsequent data processing, state recognition, and intelligent decision-making. The design concept of the multimodal sensing module is that the moisture content of cold recycled aggregates is not an isolated physical quantity, but rather the result of dynamic coupling with multiple factors such as the material's internal components, physical morphology, chemical reaction processes, and external environment. No single sensing method can fully and accurately characterize its true state. Therefore, this invention employs a multi-sensor information fusion strategy to achieve a deep understanding of the aggregate's state.

[0064] Specifically, the multimodal sensing module collects parameter information of the cold recycled mixture in real time through the coordinated deployment of a series of heterogeneous sensors. Preferably, the roles of these sensors and their functions in the system are described below:

[0065] First, to obtain the overall moisture content information of the mixture, a broadband microwave sensor array is deployed in this embodiment. This array is preferably installed below the conveyor belt of the cold recycled mixture or at the outlet channel of the mixer to achieve non-contact, penetrating measurement of the material flow. The basic principle is that the complex dielectric constant of water is much higher than that of other components such as aggregates and asphalt in the microwave frequency band. The complex dielectric constant is expressed as:

[0066] ;

[0067] In the formula, It is the complex permittivity; is the real part of the dielectric constant, reflecting the polarization and energy storage capacity of the material; Dielectric loss factor represents the material's ability to absorb microwave energy; It is the imaginary unit.

[0068] Real part of dielectric constant and dielectric loss factor All of these are closely related to the moisture content and form (free water or bound water) in the material. By emitting a broadband microwave signal and detecting its attenuation and phase shift after passing through the material, the system can calculate the dielectric constant data characterizing the overall volumetric moisture content, providing a basis for macroscopic control of moisture content.

[0069] Secondly, to compensate for the insufficient sensitivity of microwave sensors to surface moisture and to supplement the identification of water morphology, a multispectral near-infrared sensor is installed above the surface of the cold recycled mixture in this embodiment. The multispectral near-infrared sensor operates based on the principle that water molecules have strong characteristic absorption peaks in the near-infrared spectral region (preferably including the vicinity of 1450 nm and 1940 nm). By emitting near-infrared light of a specific wavelength onto the mixture surface and measuring the spectral reflectance data of its diffuse reflection, the content of free water on the material surface can be accurately quantified. This data, combined with the overall moisture content data measured by the microwave sensor, helps the system distinguish between internal water and surface water, providing crucial information for subsequent assessment of the internal component state, such as the demulsification process of emulsified asphalt.

[0070] Furthermore, the temperature state of the mixture is a crucial window into its internal chemical and physical processes. Therefore, this embodiment integrates a high-precision thermal imaging sensor. This sensor is used to monitor the continuous surface temperature distribution of the cold recycled mixture in real time during conveying or mixing. Temperature distribution information has a dual significance: firstly, the demulsification process of emulsified asphalt is an exothermic reaction; by monitoring areas of abnormal temperature increases, the location and severity of demulsification can be indirectly determined. Secondly, the evaporation of moisture is an endothermic process, leading to surface cooling, the rate of which is directly related to free water content, material temperature, wind speed, and ambient temperature and humidity. Therefore, surface temperature distribution data provides an indispensable thermodynamic perspective for the system to understand the dynamic changes in moisture content and energy balance.

[0071] Furthermore, to achieve precise proportional control of the water addition, it is essential to know the real-time output or flow rate of the mixture. This embodiment employs a high-resolution laser displacement sensor. This sensor is typically mounted above the conveyor belt and acquires the dynamic three-dimensional volume profile of the moving mixture by high-speed scanning of its surface. By integrating the profile data for each frame, the instantaneous cross-sectional area of ​​the material flow can be calculated. Combined with the real-time conveyor belt speed obtained by the speed measuring device The system can then accurately calculate the instantaneous volumetric flow rate of the mixture. The volume profile and flow velocity data are the core basis for the subsequent collaborative control execution module to calculate the target water injection flow rate, ensuring the accuracy of control.

[0072] Finally, the external environment is a boundary condition that cannot be ignored in affecting moisture content changes. This embodiment also includes environmental temperature and humidity sensors to monitor the ambient temperature and humidity at the construction site. These parameters directly affect the evaporation and heat dissipation rate of the mixture and are important input variables for environmental compensation and disturbance suppression in the moisture content prediction model and intelligent decision-making model.

[0073] The multimodal sensing module integrates and synchronizes the various sensors mentioned above, enabling it to capture comprehensive information about the state of cold recycled mixtures in real time and synchronously from multiple dimensions such as electromagnetic properties, spectrochemistry, thermodynamics, physical morphology, and environmental conditions, forming a high-dimensional time-series dataset.

[0074] Data preprocessing and tensor construction module: used to clean, normalize and impute missing values ​​of the real-time collected parameter information, and construct a fourth-order temporal tensor including parameter information;

[0075] In this embodiment, the data preprocessing and tensor construction module aims to transform the heterogeneous and noisy raw data stream collected by the upstream multimodal sensing module into a structured, high-quality data format suitable for advanced algorithm analysis. In actual industrial environments, directly collected sensor data often suffers from problems such as signal glitches, random noise, numerical drift, asynchrony, and data packet loss. If used directly without processing, it will seriously affect the accuracy and stability of subsequent analysis models.

[0076] Specifically, to achieve the above objectives, the data preprocessing and tensor construction module integrates a series of data processing functions. First, the module performs data cleaning on the real-time acquired parameter information. This step includes outlier removal and noise filtering. Outliers may arise from momentary sensor malfunctions or severe external electromagnetic interference. Therefore, the system can preferably employ statistical identification methods, such as those based on the 3σ criterion or interquartile range (IQR), to identify and remove abnormal data points that significantly deviate from the normal data distribution. After outlier removal, to further suppress high-frequency random noise (such as interference introduced by mechanical vibration or circuit thermal noise), the system can employ time-domain filtering algorithms, such as moving average filtering or, more preferably, Kalman filtering, to smooth the time-series signals of each sensor, thereby extracting more stable underlying trend signals.

[0077] After data cleaning, because the parameters output by different sensors have vastly different physical dimensions and numerical ranges (e.g., temperature is in degrees Celsius, and dielectric constant is dimensionless), the module performs Min-Max normalization on the parameter information to eliminate the potential problems of slow convergence or model bias towards features with larger numerical values ​​caused by these scale differences in subsequent model training. This process linearly maps the data of each modality to a uniform numerical range, such as [0, 1] or [-1, 1]. This ensures that features with different physical meanings have equal contribution weights in model analysis.

[0078] Considering the instability of industrial field network communication or the potential data loss due to intermittent sensor maintenance, this module further performs time-series interpolation on the parameter information to fill in the missing data points. To ensure the continuity and dynamic characteristics of the time series, an interpolation method suitable for time-series data can be used. In a simple implementation, linear interpolation or forward / backward filling can be used; in a preferred embodiment where higher data fidelity is required, methods based on spline interpolation or autoregressive models (such as ARIMA) can be used to estimate missing values ​​based on trend information before and after the data points, thereby maximizing the recovery of the dynamic characteristics of the original signal.

[0079] Finally, after the aforementioned cleaning, normalization, and interpolation processes, the data preprocessing and tensor construction module fuses and structures these high-quality multimodal data streams, constructing a fourth-order temporal tensor. As a high-order data structure, the tensor can naturally and losslessly preserve the inherent correlations of multi-source information across various physical dimensions, avoiding information loss that may occur with traditional matrix or vectorization methods. The fourth-order temporal tensor is specifically represented as follows:

[0080] ;

[0081] In the formula, Indicates the number of sensor modes; This indicates the number of sensor sampling points in space, reflecting the distribution of data in the spatial dimension; It represents the length of the time series and defines the size of the time window used for one analysis. The system processes the data stream in units of this window. This represents the feature dimension corresponding to each sampling point. For some sensors, a single sampling outputs multiple related feature values, and this dimension is used to characterize these features.

[0082] The data preprocessing and tensor construction module ultimately outputs a well-formed, clean, and highly condensed fourth-order temporal tensor. .

[0083] Tensor Analysis and State Recognition Module: Used to perform tensor representation learning on fourth-order time series tensors to extract latent features, identify the internal component state of cold recycled mixtures based on latent features, and decouple the interference caused by non-moisture content factors in parameter information to obtain pure moisture content related signals.

[0084] In this embodiment, the tensor analysis and state recognition module receives a structured fourth-order temporal tensor generated by the "data preprocessing and tensor construction module". The tensor analysis and state recognition module uses advanced tensor calculation methods to deeply mine the intrinsic structure of high-dimensional data, aiming to achieve two closely related goals: first, to extract and identify the internal component states that are difficult to measure directly from complex data and characterize the essential properties of cold recycled mixtures; second, based on a deep understanding of the mixture state, to accurately decouple the pure signals directly related to moisture content changes from highly coupled multimodal signals, thus laying a solid foundation for subsequent accurate prediction and decision-making.

[0085] To achieve the above objectives, this module first processes the fourth-order temporal tensor. Tensor representation learning is performed. In a preferred embodiment, this step is accomplished through non-negative tensor factorization, specifically using the non-negative Tucker decomposition model. This decomposition aims to approximate a complex high-order tensor as a product of a compact core tensor and a series of factor matrices, with the following mathematical expression:

[0086] ;

[0087] In the formula, As the core tensor, the size of its elements represents the strength of the interaction between potential features (or principal components) in different dimensions, and can be regarded as the core and hub of the entire data structure; , , and These are factor matrices corresponding to sensor modes, spatial sampling points, time series, and feature dimensions, respectively. Indicates the tensor along the first Modulus product, modal index These correspond to sensor modes, spatial sampling points, time series, and feature dimensions, respectively.

[0088] In this invention, nonnegativity constraints (i.e., requirements) are employed. and all The fact that all elements are non-negative is crucial because it makes the decomposed potential features clearly physically interpretable, directly corresponding to the contribution of actual physical quantities such as free water, bound water, aggregate, bitumen content, and temperature components.

[0089] Obtaining the core tensor that can characterize the intrinsic nature of the mixture sum factor matrix Subsequently, based on these extracted latent features, the module uses a pre-trained Tensor Neural Network (TNN) to identify the internal component states of the cold recycled mixture. Unlike traditional neural networks, TNNs can directly process inputs in tensor form, fully preserving the structural information of latent features in multi-dimensional space, thus enabling more effective learning of the complex nonlinear mapping relationship between these features and the macroscopic state of the mixture. The internal component states output by this network are the key basis for subsequent intelligent decision-making. Preferably, these states specifically include key performance indicators that are difficult to obtain directly through a single sensor, such as the real-time demulsification degree of emulsified asphalt, free water content, and bound water content.

[0090] Meanwhile, the tensor analysis and state identification module performs another crucial task: obtaining a pure moisture content-related signal. Because the original multimodal parameter information contains signal changes caused by non-moisture content factors (such as fluctuations in aggregate gradation, changes in recycled material sources, or fine-tuning of asphalt addition) that overlap with the signal caused by moisture content changes, directly using this information would lead to significant errors in moisture content prediction. Based on the internal component states identified in the previous step, this module can clearly identify which components are currently the main sources of interference. Therefore, the system preferably employs signal processing methods such as Tensor Independent Component Analysis (TICA) or orthogonal projection. The principle is to consider the total signal as a linear mixture of multiple statistically independent source signals (moisture content signal, asphalt interference signal, aggregate interference signal, etc.), and to maximize the non-Gaussianity of the output signal through algorithms or utilize the orthogonality principle to extract the components representing non-moisture content interference from the original fourth-order time-series tensor. Precisely separate or project and remove from the middle.

[0091] After the decoupling process described above, the tensor analysis and state recognition module ultimately outputs a one-dimensional time series, namely a pure moisture content-related signal. This signal eliminates interference from other physicochemical factors to the greatest extent possible, and its fluctuations can more realistically and sensitively reflect the actual changes in the moisture content of the cold recycled mixture. Thus, the tensor analysis and state recognition module successfully transforms a high-dimensional, complex, and interference-laden input data into two clear and high-value information outputs: one is the internal state understanding for macro-level decision-making, and the other is a pure core signal for accurate prediction.

[0092] Moisture content prediction module: Used to adaptively predict the future moisture content of cold recycled mixtures based on pure moisture content-related signals;

[0093] In this embodiment, the core task of the moisture content prediction module is to accurately and adaptively predict the moisture content of the cold recycled mixture in one or more future time steps, based on the pure moisture content-related signal provided by the upstream "tensor analysis and state recognition module" after interference has been eliminated. In actual production, there are unavoidable physical delays (such as water flow and material mixing) and chemical delays (such as emulsified asphalt breaking down and absorbing water) between the addition of water and the corresponding change in the moisture content of the mixture.

[0094] The input to the moisture content prediction module is the purified moisture content-related signal obtained after processing by the previous module. This signal is first constructed as an input time series, serving as the historical basis for the prediction model. In this embodiment, a Gated Recurrent Unit (GRU) network with an integrated attention mechanism is preferably used as the core prediction model.

[0095] The Gated Recurrent Unit (GRU) network was chosen because, as an advanced variant of the Recurrent Neural Network (RNN), its unique "update gate" and "reset gate" structures effectively capture long-term dependencies in time-series data, while significantly mitigating the gradient vanishing or exploding problems that traditional RNNs may encounter when processing long sequences. This allows the model to deeply understand the complex dynamic patterns of water content signal evolution over time.

[0096] However, in a sequence of moisture content changes, not all historical points contribute equally to future predictions. Data from certain critical events (such as a major material change or a sudden environmental change) are far more valuable than data from periods of stable operation. Therefore, this invention further integrates an attention mechanism into the GRU network. This mechanism endows the model with an adaptive "attention" capability, enabling it to dynamically assign different importance weights to each time step in the input time series when making predictions.

[0097] Specifically, the prediction process of this module is as follows: the input time series is sequentially fed into the GRU network. At each time step... The GRU unit combines the current input with the hidden state from the previous time step to generate the hidden state for the current time step. After processing the entire length of After the input sequence, the attention mechanism will process all generated hidden states. An evaluation was conducted, and a set of corresponding attention weights was calculated. .

[0098] Subsequently, the moisture content prediction module generates a single, highly condensed context vector by weighted summation of all hidden states, i.e., the weighted hidden state representation. The calculation process can be expressed by the following formula:

[0099] ;

[0100] In the formula, For gated cyclic unit networks at time steps The hidden state of the output; The hidden state calculated for the attention mechanism The weight of reflects the importance of the information at that moment to the final prediction, and satisfies . ; The length of the input time series.

[0101] Finally, this includes a weighted hidden state representation that dynamically focuses on key information throughout the entire historical sequence. The input is fed into a fully connected output layer. This output layer acts as a decoder, mapping and decoding this high-dimensional, abstract feature representation into specific physical quantities, thereby generating and outputting the predicted future moisture content of the cold recycled mixture at one or more future time points.

[0102] The moisture content prediction module can make full use of the inherent patterns and key nodes of historical information to provide reliable judgments on future trends.

[0103] Intelligent decision-making module: Based on the predicted future moisture content and the state of internal components, it dynamically decides the water addition adjustment instructions required for cold recycled mixtures through causal inference and reinforcement learning.

[0104] In this embodiment, the fundamental task of the intelligent decision-making module is to comprehensively utilize multi-dimensional information provided by upstream modules, including forward-looking predictions of the future and deep insights into the intrinsic state of materials, and dynamically determine the optimal water addition adjustment command through an advanced algorithm framework that transcends traditional control logic. The core idea behind its design is that moisture content control of cold recycled mixtures is a complex, non-linear, and dynamic process influenced by multiple coupled factors. Simple controllers based on fixed rules or models are ill-suited to adapt to real-time changes in operating conditions. This module, by integrating causal inference and reinforcement learning, aims to enable the system to understand, learn, and autonomously optimize.

[0105] To achieve this advanced decision-making function, the intelligent decision-making module first employs causal inference methods to construct and continuously update a structural causal model (SCM) online. The purpose of introducing causal inference is to enable the system to move from "correlation" to "causation," meaning it not only knows which variables change along with moisture content, but also understands to what extent these variables "cause" the change in moisture content. The structural causal model uses a directed acyclic graph (DAG) to represent the causal relationships between variables. Its nodes include internal component states (such as demulsification level), historical water addition, and environmental parameters identified by upstream modules, while the target node is moisture content. By learning from historical data, the model can quantitatively calculate the causal effect of each upstream variable on changes in moisture content. This step is crucial because it helps the system eliminate the interference of spurious correlations. For example, when a new batch of recycled aggregate has a high moisture content, the system can identify that the increase in moisture content is due to changes in the aggregate itself, rather than the amount of water added, thus avoiding erroneous adjustments.

[0106] After gaining a deep understanding of the causal relationships among the system's variables, the core of this module, a reinforcement learning agent, begins making decisions. This invention preferably employs a Deep Q-Network (DQN) as the algorithmic implementation for this agent. The essence of reinforcement learning lies in enabling the agent to learn the optimal action strategy through trial and error in interaction with the environment (or its precise digital twin model).

[0107] Specifically, the decision-making process of a reinforcement learning agent follows the framework below:

[0108] State Construction: The agent's decisions are based on its observations of the current environment, i.e., its "state." This invention constructs a rich and comprehensive state representation for the agent. This state is a high-dimensional vector, which is not merely the instantaneous value of the current water content, but is composed of the following three key pieces of information:

[0109] The predicted future moisture content is provided by the "Moisture Content Prediction Module";

[0110] The internal component states are provided by the "Tensor Analysis and State Recognition Module";

[0111] The causal effect is quantified by the causal inference part of this module. This state definition gives the agent unprecedented "insight": it can "see" the future (predicted value), "see through" appearances (internal state), and "understand" the real impact of its behavior (causal effect).

[0112] Setting the optimization objective: The learning direction of the agent is derived from its optimization objective. In this embodiment, the objective is set as minimizing the deviation between the predicted future moisture content and a preset target moisture content. The system defines a reward function based on this deviation; the closer the predicted value is to the target value, the higher the reward the agent receives.

[0113] Decision Making and Learning: Based on the aforementioned high-dimensional state, a deep Q-network (a deep neural network) is used to approximate an optimal action-value function. This function is used to evaluate the current state. Next, take a certain water volume adjustment action. The expected value of the long-term cumulative reward. During decision-making, the agent inputs its current state into the network, and the network outputs the corresponding values ​​for all possible water-adding actions. The agent then selects the value that enables... The optimal action is selected and parsed into a specific water addition adjustment command, which is then output. Through mechanisms such as experience replay and target networks, DQN can continuously optimize its network parameters through ongoing interaction with the environment, making its assessment of action value increasingly accurate, thereby learning the optimal control strategy.

[0114] The intelligent decision-making module deeply integrates causal inference (understanding physical laws) with reinforcement learning (learning optimal behavior), creating a decision-making core capable of autonomous thinking and evolution. It no longer relies on static rules or models, but rather dynamically and intelligently generates each control command based on comprehensive and profound state cognition, aiming to achieve long-term optimal control.

[0115] Collaborative control execution module: used to adjust the water addition amount according to the water addition adjustment command to dynamically control the moisture content of cold recycled mixture;

[0116] In this embodiment, the collaborative control execution module is the final physical execution end of the entire dynamic control system. Its core responsibility is to accurately and flawlessly translate the abstract, high-level water addition adjustment commands output by the upstream "intelligent decision-making module" into physical adjustment actions for the real-time water addition of the cold recycled mixture. This module serves as a bridge connecting intelligent decision-making and the physical world, and is the fundamental guarantee for ensuring that the entire closed-loop control system can "act as instructed" and achieve precise control.

[0117] Specifically, to achieve the above functions, the module's workflow strictly follows a closed-loop feedback control logic. First, the module receives a water volume adjustment command. This command is a logical instruction output by the intelligent decision-making module based on its complex algorithm model; for example, a relative value indicating "increase the flow rate by 5%" or an absolute value indicating "adjust to 15 liters / minute". The module's first task is to parse this command, transforming it into a clear physical quantity usable by the engineering control system: the target flow rate setpoint.

[0118] Once the target flow rate setpoint is obtained, the core of the module, a closed-loop control unit preferably employing a proportional-integral-derivative (PID) controller, begins to operate. This PID controller is a classic and robust control algorithm in industrial control, designed to make a controlled variable (in this case, the actual water flow rate) approach its setpoint infinitely through continuous feedback and adjustment.

[0119] To this end, the coordinated control execution module collects the current actual flow rate in real time and at high frequency through a water flow meter installed in the water supply pipeline. The controller continuously compares this real-time feedback actual flow rate with the aforementioned target flow rate setpoint, and the deviation between the two is recorded. This is used as the basis for controller calculations. This deviation... It is the basis for all actions of the controller.

[0120] The proportional-integral-derivative controller is based on this deviation. It generates an actuator control signal through its control law. Its classic control law can be expressed by the following formula:

[0121] ;

[0122] In the formula, In time The generated actuator control signals; In time deviation, that is ; This is the proportional gain coefficient, which is used to quickly adjust the gain proportionally to the magnitude of the current deviation. The integral gain coefficient is used to accumulate historical deviations, aiming to eliminate steady-state errors in the system and ensure that the actual flow rate can be accurately stabilized at the target value after long-term operation. The differential gain coefficient is adjusted according to the rate of change of the deviation, aiming to predict the future trend of the deviation, thereby suppressing the oscillation of the system and improving the stability of the response. It is the integral variable.

[0123] Finally, the coordinated control execution module adjusts the actuator control signal. This drives the final physical actuator. In a preferred embodiment, the actuator can be an electronically controlled valve or a variable frequency water pump. If an electronically controlled valve is used, the control signal... It will be converted into a standard control current or voltage to precisely adjust the valve opening; if a variable frequency pump is used, the control signal will instruct the frequency converter to adjust the operating frequency of the pump motor, thereby changing the pump speed and output flow.

[0124] In this way, the collaborative control module drives the actuator to precisely adjust the real-time water addition to the cold recycling mixture until the deviation between the actual flow rate value fed back by the water flow meter and the target flow rate setting value is corrected. Approaching zero.

[0125] In summary, the collaborative control and execution module, through a complete and rigorous closed-loop control process of "instruction parsing - feedback comparison - PID calculation - drive execution," ensures that every intention of the upper-level intelligent decision-making can be realized in the physical world with high fidelity, stability, and speed, ultimately achieving dynamic control of moisture content.

[0126] Please see the appendix Figure 2 The present invention also provides a method for controlling the moisture content of cold recycled mixtures using multi-sensor fusion, comprising the following steps:

[0127] S1. Real-time acquisition of parameter information of at least two different modes of cold recycled mixture through multiple sensors, the parameter information including dielectric constant data, spectral reflectance data, surface temperature distribution data, volume profile and flow rate data, as well as ambient temperature and humidity;

[0128] S2. Clean, normalize, and impute missing values ​​for the parameter information collected in real time, and construct a fourth-order temporal tensor including the parameter information.

[0129] S3. Perform tensor representation learning on the fourth-order time series tensor to extract latent features, identify the internal component state of the cold recycled mixture based on the latent features, and decouple the interference caused by non-moisture content factors in the parameter information to obtain a pure moisture content related signal.

[0130] S4. Based on the pure moisture content related signal, adaptively predict the future moisture content of the cold recycled mixture;

[0131] S5. Based on the predicted future moisture content and the state of the internal components, dynamically decide the water addition adjustment command required for the cold recycled mixture through causal inference and reinforcement learning.

[0132] S6. Adjust the amount of water added according to the water addition adjustment instruction to dynamically control the moisture content of the cold recycled mixture.

[0133] The method in this embodiment can be used to execute the above system embodiment, and its principle and technical effect are similar, so it will not be described again here.

[0134] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor fusion-based cold recycling mixture moisture content control system, characterized in that, include: Multimodal sensing module: used to collect parameter information of at least two different modes of cold recycled mixture in real time through multiple sensors. The parameter information includes dielectric constant data, spectral reflectance data, surface temperature distribution data, volume profile and flow rate data, as well as ambient temperature and humidity. Data preprocessing and tensor construction module: used to clean, normalize and impute missing values ​​of the real-time collected parameter information, and construct a fourth-order temporal tensor including the parameter information; Tensor analysis and state recognition module: used to perform tensor representation learning on the fourth-order time series tensor to extract latent features, identify the internal component state of the cold recycled mixture based on the latent features, and decouple the interference caused by non-moisture content factors in the parameter information to obtain pure moisture content related signals; Moisture content prediction module: used to adaptively predict the future moisture content of the cold recycled mixture based on the pure moisture content related signal; Intelligent decision-making module: Based on the predicted future moisture content and the state of the internal components, it dynamically determines the water addition adjustment command required for the cold recycled mixture through causal inference and reinforcement learning; Collaborative control execution module: used to adjust the amount of water added according to the water addition adjustment command to dynamically control the moisture content of the cold recycled mixture; The tensor analysis and state recognition module includes: Tensor representation learning is performed on the fourth-order time series tensor, and the core tensor and factor matrix are obtained through non-negative tensor factor decomposition. The core tensor and factor matrix represent the potential characteristics of free water, bound water, aggregate, asphalt content and temperature components in the cold recycled mixture. Based on the core tensor and factor matrix, the internal component state of the cold recycled mixture is identified by tensor neural network. The internal component state includes the real-time demulsification degree of emulsified asphalt, free water content, and bound water content. Based on the internal component state, tensor independent component analysis or orthogonal projection method is used to decouple the interference caused by non-moisture content factors from the fourth-order time series tensor to obtain the pure moisture content related signal; For the fourth-order temporal tensor Perform a nonnegative Tucker decomposition, which is expressed as: ； In the formula, For the core tensor; , , and These are factor matrices corresponding to sensor modes, spatial sampling points, time series, and feature dimensions, respectively. Indicates the tensor along the first Modulus product, modal index These correspond to sensor modes, spatial sampling points, time series, and feature dimensions, respectively.

2. The multi-sensor fusion-based cold recycling mixture moisture content control system according to claim 1, characterized in that, The multimodal sensing module collects the parameter information in real time through the following sensors: A broadband microwave sensor array deployed at the conveyor belt or mixer outlet of the cold recycled mixture is used to collect the dielectric constant data of the mixture. A multispectral near-infrared sensor deployed above the surface of the cold recycled mixture is used to collect the spectral reflectance data of the mixture; A high-precision thermal imaging sensor is used to monitor the surface temperature distribution of the cold recycled mixture; A high-resolution laser displacement sensor is used to measure the volume profile and flow rate of the cold recycled mixture; An ambient temperature and humidity sensor is used to monitor the ambient temperature and humidity at the construction site of cold recycled mixtures.

3. The multi-sensor fusion-based cold recycling mixture moisture content control system according to claim 1, characterized in that, The data preprocessing and tensor construction module includes: The collected parameter information is cleaned, including outlier removal and noise filtering; The parameter information is then subjected to Min-Max normalization. The parameter information is temporally interpolated to fill in missing data points and constructed as a fourth-order temporal tensor. ; In the formula, For constructing a fourth-order temporal tensor; Represents the real number field; Indicates the number of sensor modes; Indicates the number of spatial sampling points of the sensor; Indicates the length of the time series; This represents the feature dimension corresponding to each sampling point.

4. The multi-sensor fusion-based cold recycling mixture moisture content control system according to claim 1, characterized in that, The moisture content prediction module includes: The purified moisture content-related signal is used to construct an input time series. The input time series is processed by a gated recurrent unit network with an integrated attention mechanism to learn temporal dependencies and generate weighted hidden state representations. The weighted hidden state representation is input to a fully connected output layer to generate and output the predicted future moisture content of the cold recycled mixture at one or more future time points.

5. The multi-sensor fusion-based cold recycling mixture moisture content control system according to claim 4, characterized in that, The weighted hidden state representation Calculated using the following formula: ； In the formula, For the gated cyclic unit network at time step The hidden state of the output; The hidden state calculated for the attention mechanism The weights; The length of the input time series.

6. The multi-sensor fusion-based cold recycling mixture moisture content control system according to claim 1, characterized in that, The intelligent decision-making module includes: A structural causal model is constructed and updated using a causal inference method. The structural causal model is used to quantify the causal effects of the internal component state, historical water addition, and environmental parameters on the water content. The reinforcement learning agent is based on the state composed of the predicted future moisture content, the internal component state, and the causal effect. With the optimization objective of minimizing the deviation between the predicted future moisture content and the preset target moisture content, it dynamically makes decisions and outputs the water addition adjustment command through a deep Q network.

7. The multi-sensor fusion-based cold recycling mixture moisture content control system according to claim 1, characterized in that, The coordinated control execution module includes: Receive the water addition adjustment command and parse the water addition adjustment command into a target flow rate setting value; The actuator control signal is generated based on the deviation between the target flow rate setpoint and the actual flow rate value fed back by the water flow meter in real time using a proportional-integral-derivative controller. According to the actuator control signal, the electronic control valve or variable frequency water pump is driven to precisely adjust the real-time water addition to the cold recycled mixture, so as to dynamically regulate the moisture content.

8. A multi-sensor fusion method for controlling the moisture content of cold recycled aggregates, applied to the multi-sensor fusion system for controlling the moisture content of cold recycled aggregates as described in any one of claims 1-7, characterized in that, Includes the following steps: The parameters of at least two different modes of cold recycled mixture are collected in real time by multiple sensors. The parameters include dielectric constant data, spectral reflectance data, surface temperature distribution data, volume profile and flow rate data, as well as ambient temperature and humidity. The collected parameter information is cleaned, normalized, and missing values ​​are imputed. A fourth-order temporal tensor including the parameter information is then constructed. Tensor representation learning is performed on the fourth-order time series tensor to extract latent features. Based on the latent features, the internal component state of the cold recycled mixture is identified, and the interference caused by non-moisture content factors in the parameter information is decoupled to obtain a pure moisture content related signal. Based on the pure moisture content-related signal, the future moisture content of the cold recycled mixture is adaptively predicted. Based on the predicted future moisture content and the internal component state, the required water addition adjustment command for the cold recycled mixture is dynamically determined through causal inference and reinforcement learning. According to the water addition adjustment instruction, the water addition is adjusted to dynamically control the moisture content of the cold recycled mixture.

Images