Textile setting machine flue gas waste heat recovery self-adaptive control system and method

Abstract

The invention discloses a textile setting machine flue gas waste heat recovery adaptive control system and method, and the method comprises the steps: firstly carrying out the global state estimation of multi-mode sensor data through a graph neural network, so as to precisely sense the hidden states, such as the scaling distribution, in a heat exchanger, and then introducing a causal inference technology to carry out the active risk assessment, a dynamic safety control constraint is generated by analyzing causal association between variables, so that traditional passive threshold alarm is replaced. In the safety boundary, the system utilizes multi-target reinforcement learning to formulate a macroscopic operation strategy giving consideration to both energy efficiency and process stability, and realizes accurate trajectory tracking and execution of a set point in combination with model prediction control. In this way, deep perception and active risk defense of complex flue gas working conditions can be achieved, the waste heat recovery energy efficiency is maximized on the premise that the intrinsic safety of the system is guaranteed, and the operation and maintenance cost and fire hazards are remarkably reduced.

Description

Technical Field

[0001] This application relates to the field of intelligent control, and more specifically, to an adaptive control system and method for waste heat recovery from flue gas in textile setting machines. Background Technology

[0002] The textile industry is an energy-intensive industry. Stenter machines, as key energy-consuming equipment, emit large amounts of waste gas at temperatures reaching 160°C to 200°C during fabric heat setting, containing significant potential for heat recovery. However, the composition of stenter flue gas is extremely complex. Fabric lint and volatile oil mist easily condense, adhere, and accumulate as they flow through heat exchangers and pipes, forming a high-thermal-resistance oil scale layer. This scaling not only leads to a sharp decrease in heat exchange efficiency and increased energy consumption, but the accumulated flammable deposits can also easily cause serious fires under high-temperature conditions. Therefore, developing an adaptive control scheme for waste heat recovery that can adapt to complex flue gas conditions is crucial for achieving energy conservation, emission reduction, and inherently safe production in textile enterprises.

[0003] However, existing waste heat recovery control schemes for textile setting machines generally suffer from insufficient intelligence. Current system operating parameter settings largely rely on operator experience or simple PID control strategies. This crude approach cannot dynamically adapt to fluctuations in operating conditions caused by changes in fabric material, weight, and machine speed, making it difficult for the system to maintain optimal operation in diverse production environments. Furthermore, existing safety management mechanisms primarily rely on threshold-based passive alarms (such as temperature over-limit alarms), lacking deep perception of the system's internal state. They cannot accurately monitor scale distribution inside heat exchangers, which is difficult to measure directly, and are even less capable of conducting root cause analysis and proactive prediction of risks such as spontaneous combustion caused by increased oil mist concentration. This lagging control and safety response model often leaves enterprises facing multiple challenges, including reduced energy efficiency, high maintenance costs, and frequent fire hazards.

[0004] Therefore, an optimized adaptive control method for waste heat recovery from textile setting machines is needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an adaptive control system and method for waste heat recovery from flue gas in textile setting machines.

[0006] According to one aspect of this application, an adaptive control method for waste heat recovery from flue gas in a textile setting machine is provided, comprising: Acquire raw data from multimodal sensors; Global state estimation based on graph neural network is performed on the raw data from multimodal sensors to obtain the enhanced system state matrix; Active risk assessment based on causal inference is performed on the enhanced system state matrix to obtain dynamic security control constraints; Macroscopic policy optimization based on multi-objective reinforcement learning is performed on the enhanced system state matrix and dynamic security control constraints to obtain the macroscopic operating setpoint; Trajectory tracking and execution under model predictive control are performed on the macroscopic operating setpoint, dynamic safety control constraints, and enhanced system state matrix to obtain the actuator instruction sequence.

[0007] According to another aspect of this application, an adaptive control system for waste heat recovery from textile setting machine flue gas is provided, comprising: The raw data acquisition module is used to acquire raw data from the multimodal sensor. The global state estimation module is used to perform global state estimation based on graph neural networks on the raw data from multimodal sensors to obtain the enhanced system state matrix; The proactive risk assessment module is used to perform proactive risk assessment on the enhanced system state matrix based on causal inference to obtain dynamic security control constraints; The macro-policy optimization module is used to perform macro-policy optimization based on multi-objective reinforcement learning on the enhanced system state matrix and dynamic security control constraints to obtain the macro-operation setpoint. The trajectory tracking and execution module is used to track and execute the trajectory under model predictive control based on the macroscopic operating setpoint, dynamic safety control constraints, and enhanced system state matrix to obtain the actuator instruction sequence.

[0008] Compared with existing technologies, this application provides an adaptive control system and method for waste heat recovery from flue gas in textile setting machines. First, it utilizes graph neural networks to perform global state estimation on multimodal sensor data to accurately perceive hidden conditions such as fouling distribution inside the heat exchanger. Then, it introduces causal inference technology for proactive risk assessment, generating dynamic safety control constraints by analyzing causal relationships between variables, thereby replacing traditional passive threshold alarms. Within this safety boundary, the system uses multi-objective reinforcement learning to formulate macroscopic operating strategies that balance energy efficiency and process stability, and combines model predictive control to achieve precise trajectory tracking and execution of the setpoint. In this way, it achieves deep perception and proactive risk defense of complex flue gas conditions, maximizing waste heat recovery efficiency while ensuring the inherent safety of the system, and significantly reducing operation and maintenance costs and fire hazards. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a framework diagram of an adaptive control system for waste heat recovery from textile setting machine flue gas according to an embodiment of this application; Figure 2 This is a flowchart of an adaptive control method for waste heat recovery from textile setting machine flue gas according to an embodiment of this application; Figure 3 This is a schematic diagram of the data flow of the adaptive control method for waste heat recovery from textile setting machine flue gas according to an embodiment of this application; Figure 4 This is a block diagram of an adaptive control system for waste heat recovery from flue gas in a textile setting machine, according to an embodiment of this application. Detailed Implementation

[0011] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention. Below, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] As mentioned in the background section, existing safety management mechanisms mainly rely on threshold-based passive alarms (such as temperature over-limit alarms), lacking deep perception of the system's internal state. They cannot accurately monitor the scale distribution inside heat exchangers, which is difficult to measure directly, and are even less capable of conducting root cause analysis and forward-looking prediction of risks such as spontaneous combustion caused by increased oil mist concentration. This paper proposes an adaptive control system for waste heat recovery from textile setting machine flue gas, the specific framework of which is shown in the figure below. Figure 1 As shown, specifically, firstly, graph neural networks are used to perform global state estimation on multimodal sensor data to accurately perceive hidden states such as fouling distribution inside the heat exchanger; then, causal inference technology is introduced to build a safety engine, which performs proactive risk assessment and generates dynamic safety control constraints through forward inference, realizing the transformation from passive alarm to pre-warning; on this basis, a two-layer intelligent control architecture is adopted, using a multi-objective reinforcement learning (MORL) agent trained offline in an executable digital twin (xDT) environment to formulate macro-operation strategies, and combining model predictive control (MPC) to achieve precise trajectory tracking and execution of the setpoint within the dynamic safety boundary; in addition, the system also integrates interpretable artificial intelligence (XAI) technology to enhance decision-making transparency, thereby maximizing the energy efficiency of waste heat recovery while ensuring the inherent safety of the system.

[0016] Figure 2 This is a flowchart of an adaptive control method for waste heat recovery from textile setting machine flue gas according to an embodiment of this application. Figure 3 This is a system architecture diagram of the adaptive control method for waste heat recovery from textile setting machine flue gas according to an embodiment of this application. Figure 2 and Figure 3As shown, the adaptive control method for waste heat recovery from textile setting machine flue gas according to an embodiment of this application includes the following steps: S1, acquiring raw data from multimodal sensors; S2, performing global state estimation based on graph neural networks on the raw data from multimodal sensors to obtain an enhanced system state matrix; S3, performing active risk assessment based on causal inference on the enhanced system state matrix to obtain dynamic safety control constraints; S4, performing macroscopic strategy optimization based on multi-objective reinforcement learning on the enhanced system state matrix and dynamic safety control constraints to obtain a macroscopic operating setpoint; S5, performing trajectory tracking and execution under model predictive control on the macroscopic operating setpoint, dynamic safety control constraints, and enhanced system state matrix to obtain an actuator instruction sequence.

[0017] Specifically, S1 involves acquiring raw data from multimodal sensors. This raw data refers to a collection of unprocessed heterogeneous signals collected in real-time by a multi-source sensing layer deployed on the physical system (i.e., the textile setting machine, heat exchange and piping subsystem, and cleaning subsystem). In embodiments of this application, this specifically includes temperature, pressure, flow rate, acoustic signals, oil mist concentration, and infrared image pixel values. These data are not merely single numerical records, but multidimensional physical representations reflecting system heat exchange efficiency, fluid dynamics characteristics, equipment mechanical state, and the degree of contaminant accumulation. It should be understood that traditional single-dimensional threshold alarms (such as those relying solely on temperature) cannot effectively monitor the scale distribution inside heat exchangers, which is difficult to measure directly, nor can they capture the risk of spontaneous combustion caused by increased oil mist concentration. Therefore, in the technical solution of this application, by acquiring raw data from multimodal sensors and deploying a rich network of heterogeneous sensors, subtle changes in system operation can be comprehensively captured from multiple physical dimensions such as heat, force, sound, light, and chemistry. This provides redundant and complementary information input for state estimation and lays a solid data foundation for subsequent intelligent control and safety decisions.

[0018] In practical implementation, firstly, regarding the acquisition of thermal parameters, the system densely deploys high-precision temperature sensors (such as PT100 RTDs) and differential pressure sensors at the inlet and outlet of the heat exchanger on both the flue gas and fresh air sides, as well as at key intermediate locations. These sensors are responsible for collecting temperature and pressure data in real time, forming the basis for calculating the basic heat transfer coefficient and pressure drop. Secondly, regarding the acquisition of flow data, the system installs thermal or Pitot tube-type anemometers / flow sensors in the main and bypass ducts to accurately measure airflow. Thirdly, to capture more subtle changes in system status, the system introduces an acoustic signal acquisition mechanism, namely, installing broadband microphones or vibration acceleration sensors near the exhaust fan and at key bends in the ductwork. These acoustic sensors not only reflect the operating status of the fan (such as bearing wear and blade dust accumulation), but more importantly, when scaling causes changes in the airflow channel morphology, its acoustic characteristics also change, thus allowing the system to capture changes in airflow turbulence intensity through acoustic signals. In addition, for pollutant monitoring, the system installs particulate matter / oil mist concentration sensors based on light scattering or light absorption principles in the flue gas duct to directly measure the pollutant content in the flue gas, serving as a key input variable for scaling rate analysis. Finally, to achieve visualized condition monitoring, the system installs infrared thermal imagers or high-temperature resistant industrial cameras in key areas of the heat exchanger to acquire infrared image pixel values. This image data provides a detailed two-dimensional temperature distribution map of the heat exchange surface, thus intuitively identifying "cold spots" or "hot spots" caused by severe local scaling. The raw signals collected by all the above sensors will form a complete multimodal sensor raw dataset.

[0019] Specifically, S2 involves performing global state estimation based on a graph neural network on the raw data from the multimodal sensors to obtain an enhanced system state matrix. It should be understood that traditional methods often rely on a single physical model or simple data-driven approaches, making it difficult to effectively handle the heterogeneity of multimodal data, the spatial correlation of system structures, and the embedding of physical constraints. Therefore, in the technical solution of this application, by introducing a graph neural network to model the system as a graph structure and utilizing the message passing mechanism between nodes, the spatial correlation information in the multimodal data can be fully mined, achieving accurate estimation of the system's global state, thereby providing precise decision-making basis for subsequent risk assessment and intelligent control.

[0020] In practical implementation, the first step is to clean and extract features from the raw multimodal sensor data to obtain a standardized feature vector set. Considering that raw multimodal sensor data collected in industrial settings often suffers from various quality issues, directly using this data would severely impact the performance and reliability of subsequent models. Specifically, industrial sensor data typically presents the following problems: First, the data may contain significant outliers and noise, which could originate from sensor malfunctions, signal interference, or environmental interference; second, data from different sensors have different dimensions and magnitudes, and direct use could lead to certain features dominating model training, affecting model convergence and performance; third, the raw data may contain redundant information or require further extraction of meaningful physical features. Therefore, in this application's technical solution, the raw multimodal sensor data undergoes data cleaning and feature extraction to provide clean and standardized input for subsequent graph neural network state estimation. Specifically, outliers and noise are removed through data cleaning, meaningful physical quantities are calculated from the raw signals through feature extraction, and feature values ​​of different dimensions and magnitudes are normalized to a unified numerical range through standardization, ultimately forming a standardized feature vector set, laying the foundation for efficient processing of subsequent models.

[0021] In this process, firstly, the raw data from the multimodal sensors is cleaned to remove outliers and noise, ensuring data quality and reliability. Specifically, for the raw data sequence of each sensor channel, statistical methods are used to identify and remove outliers; a common method is based on the three-standard-deviation principle. For the removed data points, linear interpolation or moving averages can be used to fill in the gaps. Additionally, sliding window techniques, such as using a moving average filter, can be employed for smoothing. Secondly, feature extraction is performed on the cleaned data to calculate meaningful physical quantities from the raw sensor signals, thus better reflecting the system's operating state. Specifically, different feature extraction methods are required for different types of sensors. For temperature sensors, the raw data might be resistance or voltage values, which need to be converted into temperature values ​​using calibration curves or formulas. For differential pressure sensors, the raw data is a differential pressure signal, which needs to be converted into actual pressure values ​​based on the sensor's calibration parameters. For flow sensors, such as thermal or pitot tube anemometers, the raw signal needs to be converted into flow rate or velocity based on the sensor's characteristic curves or physical formulas. For acoustic sensors, frequency domain features need to be extracted from the time-domain waveform, for example, by calculating the spectrum using Fast Fourier Transform (FFT) and then extracting the amplitude and frequency of the main frequency components. For optical sensors, such as particulate matter concentration sensors, the raw light intensity signal needs to be converted into particulate matter concentration values ​​based on the principles of light scattering or absorption. For infrared thermal imagers, pixel values ​​need to be converted into temperature values, and statistical features of temperature distribution (such as mean, variance, maximum, and minimum values) need to be extracted. Through feature extraction, the raw sensor signals are converted into physically meaningful feature values, providing more valuable information for subsequent state estimation. Furthermore, the extracted features are standardized to normalize feature values ​​of different dimensions and magnitudes to a unified numerical range. Specifically, for each feature dimension, Z-score standardization can be used. Through standardization, all features are transformed into a distribution with a mean of 0 and a standard deviation of 1, making different features comparable and facilitating subsequent model training and convergence.

[0022] Next, initial graph state data is constructed based on a standardized feature vector set. It should be understood that the waste heat recovery system for textile setting machine flue gas consists of multiple physical components (such as heat exchanger tube bundles, inlets and outlets, key monitoring points, etc.), which are physically connected through pipes, ducts, etc., forming a complex network structure. Traditional data processing methods often ignore this spatial correlation, while graph neural networks can effectively capture the spatial dependencies between nodes. Therefore, in the technical solution of this application, the system is modeled as a graph structure, where nodes represent key locations in the system, and edges represent physical connections or spatial proximity relationships between nodes, thus providing structured input to the graph neural network and achieving a deep perception of the system's global state. By constructing initial graph state data, the system can utilize the message passing mechanism of the graph neural network to aggregate the features of each node with the features of its neighboring nodes, thereby capturing the global spatial correlation of the system and providing a more accurate state representation for subsequent state estimation, risk assessment, and intelligent control.

[0023] In this process, firstly, the entire waste heat recovery system is abstracted into a mathematical graph model, where nodes represent key physical locations such as heat exchanger tube bundles and inlet / outlet monitoring points, and edges represent the actual physical connections or spatial proximity relationships between these locations. The system employs a distance threshold-based adjacency determination method, calculating the Euclidean distance between node feature vectors to determine spatial proximity. Secondly, node features are assigned, matching corresponding feature data from a standardized feature vector set based on the physical location of each node. For locations with direct sensor monitoring, the corresponding feature vector is directly extracted as the node's initial features; for locations without direct monitoring points, a linear interpolation algorithm is used to calculate the node's feature value based on the feature data of adjacent monitoring points and distance weights. This interpolation method ensures that all nodes in the graph structure obtain reasonable initial feature representations. Next, edge connections are constructed, using the k-nearest neighbor algorithm or a distance threshold method to establish connections between nodes. By calculating the feature spatial distance between all node pairs, the k nearest neighbors are selected for each node to establish connection edges, or node pairs with distances less than a preset threshold are connected. The weights of edges can be calculated based on the reciprocal of the distance or other similarity measures to form a weighted adjacency matrix, thereby fully constructing the initial graph state data containing node features and connection relationships.

[0024] Furthermore, message-passing-based GNN feature aggregation is performed on the initial graph state data to obtain a high-dimensional set of node embedding vectors. It is understandable that traditional machine learning methods often neglect the connections between nodes when processing graph data, while graph neural networks, through message passing mechanisms, can effectively integrate graph structure information into node feature learning. Specifically, in the state estimation task of a waste heat recovery system for textile finishing machines, the system's operating state depends not only on the sensor readings of individual monitoring points but also on the interactions between adjacent monitoring points and upstream and downstream equipment. Through multi-layer message passing in graph neural networks, the feature representation of each node can progressively aggregate information from its hops, two-hops, and even more distant neighbors, thus forming a richer and more robust node embedding. This embedding representation not only includes the multimodal sensor features of the node but also encodes the node's position information in the graph structure, its similarity relationship with neighboring nodes, and the statistical features of the global graph structure, providing a strong feature representation foundation for subsequent fouling state estimation, risk assessment, and intelligent control.

[0025] In this process, message generation begins. Each node receives information from its neighbors, achieved through a learnable message function. This function combines the features of its neighbors with those of its current node, generating an information flow from neighbors to the target node. Message functions typically employ multilayer perceptrons or linear transformations to achieve non-linear feature combinations. Next, information aggregation occurs. The target node integrates all messages sent by its neighbors. This aggregation process must satisfy the mathematical requirement of permutation invariance, meaning the aggregation result does not depend on the order of the neighbors. Common aggregation methods include summation aggregation, mean aggregation, and maximum aggregation. Summation aggregation preserves the complete contribution of all neighbor information, mean aggregation eliminates the influence of differences in the number of neighbors, and maximum aggregation focuses on the strongest response signal in each feature dimension. Finally, node updates occur. The target node fuses the aggregated neighbor information with its current features, generating a new node representation through an update function. The update process typically uses a join operation to concatenate its own features with the aggregated information, then uses a transform layer with a non-linear activation function to generate a higher-dimensional embedded representation. Through a multi-layered stacked graph neural network structure, each node can gradually aggregate information from its one-hop neighbors, two-hop neighbors, and even more distant nodes to form a high-dimensional embedding vector containing rich contextual semantics.

[0026] Subsequently, soft-sensor regression and parameter inversion are performed on the high-dimensional node embedding vector set to obtain the fouling distribution map. It should be understood that in the waste heat recovery system of textile finishing machine flue gas, the fouling state of the heat exchanger surface is a key factor affecting energy efficiency. Traditional soft-sensor regression models, when decoding the high-dimensional node embedding vectors output by the graph neural network (GNN), typically use a standard multilayer perceptron (MLP) to predict two core physical quantities in parallel and independently: fouling thickness and heat transfer coefficient attenuation factor. This approach ignores the inherent, deterministic physical causal relationship between these two target physical quantities. Specifically, in thermophysics, fouling thickness is the direct physical cause of decreased heat transfer performance. As an added thermal resistance layer, the thickness of the fouling layer directly determines the magnitude of the thermal resistance, and thus the degree of attenuation in heat transfer efficiency. A general function approximator does not explicitly enforce this physical constraint in the model structure, but implicitly learns this correlation from massive amounts of data. This not only reduces learning efficiency, but may also lead to the model outputting physically inconsistent predictions in edge cases not covered by the training data. For example, extremely high fouling thickness may be accompanied by extremely low heat transfer attenuation, thus reducing the robustness and interpretability of the model. To address this, this application proposes a physical causal-guided sequential regression readout layer framework, which directly embeds the physical causal chain into the network structure to replace the traditional parallel regression approach.

[0027] Among them, soft measurement regression refers to using computer algorithms and mathematical models to estimate process-dominant variables that are difficult to measure directly (such as the thickness of scale inside a heat exchanger) based on easily measurable auxiliary variables (such as the embedded vectors of temperature and pressure sensor data after GNN processing); parameter inversion refers to the process of deriving the internal physical parameters of the system from the observed system response characteristics.

[0028] Specifically, firstly, basic physical parameter regression is performed on the high-dimensional node embedding vectors to obtain the predicted fouling thickness. This step aims to extract the fundamental parameters, i.e., fouling thickness, as the starting point of the physical causal chain from the high-dimensional features aggregated by the graph neural network (GNN). Specifically, a dedicated regression head is used to decode the high-dimensional embedding vector of each node output by the GNN. From the complex multimodal features aggregated by the GNN, the most direct physical state variables that can serve as the basis for subsequent inference are extracted and quantified first, thus establishing a solid and reliable physical anchor for subsequent predictions. This process is expressed by the formula: in, This represents the predicted fouling thickness for node v; It is the high-dimensional node embedding vector output by node v in the last layer of the GNN; and These represent the learnable weight matrix and bias vector in the regression head, respectively. As a nonlinear activation function, it ensures that the predicted thickness is non-negative, which is consistent with physical reality, thereby achieving the goal of accurately quantifying the fundamental physical parameters.

[0029] Secondly, based on the predicted fouling thickness, physical prior-guided feature enhancement is applied to the high-dimensional node embedding vector to obtain a physically enhanced embedding vector. That is, the physical quantities obtained in the previous step are used as strong prior knowledge to guide the regression of subsequent related physical quantities. Specifically, a physical attention gating mechanism is constructed to... As a strong physical prior, the original node embedding features are adaptively enhanced. In this process, firstly, a physical attention gating value is calculated, representing the model's confidence or attention to the physical prior of fouling thickness. This process is expressed by the formula: in, Indicates the physical attention gating value; The high-dimensional node embedding vector output by the GNN; and These are the weight matrix and bias vector of the gated unit; The Sigmoid activation function is used to ensure that the gate value is between 0 and 1. Subsequently, this gate value is used to perform weighted fusion of the original features and the transformed features incorporating the fouling thickness, in order to generate a physically enhanced embedding vector. This process is expressed by the formula: in, It is a physically enhanced embedding vector that incorporates physical priors; The above-mentioned gate values; It is the original high-dimensional embedding vector; This is the predicted scale thickness; and These are the weights and biases used for feature fusion; This represents a vector concatenation operation; Represents element-wise multiplication; This is the Tanh activation function. If the gate value is close to 1, the model will rely more on the newly fused features; otherwise, it will rely more on the original features, thus achieving dynamic optimization of the features.

[0030] Next, a conditional sequential regression is performed on the physical enhancement embedding vector to obtain the predicted attenuation factor. This step aims to complete the closed loop of the physical causal chain, that is, to accurately predict the effect (heat transfer attenuation) given the cause (fouling thickness). Specifically, a second specialized regression head is used to regress the physical enhancement embedding vector generated in the previous step to predict the heat transfer coefficient attenuation factor as a result of the causal chain. This regression process is a conditional prediction, no longer requiring learning the relationship between thickness and attenuation from scratch, but focusing on learning, at a given thickness, the attenuation factor due to other factors (such as the dense structure of the fouling layer, etc.), whose information is implied in the original... The change in heat transfer attenuation caused by the (middle) factor allows for an accurate regression of the heat transfer coefficient attenuation factor. This process is expressed by the formula: in,, The attenuation factor represents the predicted heat transfer coefficient. It is the input physical augmentation embedding vector; and These are the learnable weights and biases in the regression head. They are activated using the Sigmoid activation function. This ensures that the output attenuation factor value ranges between 0 and 1, ultimately achieving an accurate assessment of the system's energy efficiency status. Subsequently, the predicted attenuation factor and predicted scale thickness are combined to obtain a scale distribution map. This scale distribution map is a data matrix that maps the degree of scale (thickness) and performance impact (attenuation factor) at each physical node of the heat exchanger, providing a digital perspective on the internal state of the heat exchanger.

[0031] Furthermore, the scale distribution map and the standardized feature vector set are fused to obtain an enhanced system state matrix. It should be understood that traditional control systems often rely solely on directly readable sensor values ​​(such as temperature and pressure), which is called explicit state. However, the key factors truly determining system energy efficiency and safety, such as the thickness of ash accumulation inside the heat exchanger and the attenuation of heat transfer efficiency, are often implicit states that cannot be directly measured. If only explicit state is used for control, the system will be unable to perceive the degradation trend of the equipment, while using only implicit state lacks a reference for real-time operating conditions. Therefore, in the technical solution of this application, by fusing the scale distribution map and the standardized feature vector set, the standardized feature vector set (explicit observation), which reflects real-time operating conditions, is organically combined with the scale distribution map (implicit inference) obtained through deep learning inversion at the data level. This generates an enhanced system state matrix that includes both real-time operating parameters and internal health status. This matrix serves as a unified data foundation for subsequent causal inference safety engines to conduct risk assessments and for multi-objective reinforcement learning agents to formulate macro-level strategies, achieving full-dimensional perception of the system's physical entities.

[0032] In this process, the system traverses each node in the graph structure, concatenating the standardized feature vector of that node with the corresponding inverted physical parameters along the channel dimension. This operation not only preserves the time-varying information of the original sensors but also enhances the expressive power of the state by explicitly injecting physical priors (thickness and attenuation). Next, the system reorganizes the enhanced state vectors of all nodes according to the graph's topology, ultimately forming a comprehensive matrix, namely the enhanced system state matrix.

[0033] Specifically, S3 involves performing a proactive risk assessment based on causal inference on the enhanced system state matrix to obtain dynamic safety control constraints. It should be understood that traditional safety systems typically rely on static rules and thresholds set by experts, essentially based on correlations observed in historical experience rather than causality. This mechanism exhibits significant lag, often triggering alarms only after a dangerous state has already formed or even occurred; simultaneously, because the system cannot distinguish between true warning signs and normal operating condition fluctuations, it is prone to false alarms or missed alarms due to failure to detect the root cause. Therefore, in the technical solution of this application, a proactive risk assessment based on causal inference is performed on the enhanced system state matrix to transform safety management from passive, post-event alarms to proactive avoidance based on root cause analysis through causal inference technology. That is, by understanding the deep causal relationships between variables, risk propagation paths are proactively identified and blocked, thereby generating dynamic safety boundaries to ensure the inherent safety of the system.

[0034] In practice, firstly, the enhanced system state matrix is ​​analyzed and mapped based on the system safety causal graph to obtain an instantiated causal graph. During this process, the system receives the enhanced system state matrix generated in the previous stage and maps it to a pre-learned system safety causal graph. Specifically, the instantiated causal graph can be obtained offline from massive historical and simulation data using time-series causal discovery algorithms (such as PCMCI). Figure 1 A directed acyclic graph (DAG) is formed, in which nodes represent system variables and directed edges represent direct causal relationships and time delays between variables [1]. By assigning the values ​​in the real-time state matrix to the corresponding nodes in the causal graph, the system generates an instantiated causal graph that reflects the specific operating conditions at the current moment.

[0035] Next, the instantiated causal graph is subjected to forward inference-based risk propagation probability calculation to obtain the causal risk probability. This step aims to use causal inference techniques to perform forward inference on the future based on the current system state, predict the causal propagation chain of abnormal states, and thus quantify the probability of a safety incident occurring before it happens, i.e., the causal risk probability, so that the system can take action in advance.

[0036] In this process, the causal inference safety engine first initiates a forward inference procedure. During this process, the engine utilizes pre-learned directed edges (representing causal directions) and causal delay times in the graph to simulate the evolution of the system state from the current state along the path of the causal graph. Specifically, if the state of a node in the graph (e.g., "oil mist concentration") is marked as abnormal or shows a deteriorating trend, the algorithm calculates how this state change will be transmitted to downstream nodes (e.g., "local temperature" or "heat exchange efficiency") through causal paths within a specific future time window. The system comprehensively considers the superposition effect of multiple causal paths, quantifying the probability that the propagation of these abnormal states will ultimately trigger critical safety event nodes in the system (such as "spontaneous combustion" or "explosion"). This process, based on the conditional independence and temporal dependencies inherent in the causal graph, can infer the likelihood of a specific catastrophic consequence occurring in the future; this value is the causal risk probability.

[0037] Furthermore, the causal risk probability is correlated with constraints and its constraint margin is calculated to obtain a set of safety constraints, which constitutes dynamic safety control. It should be understood that traditional safety mechanisms typically rely on fixed thresholds, but when high-risk conditions have emerged (e.g., increased oil mist concentration leading to increased flammability) but the fixed threshold has not yet been triggered, the system is often in a blind zone. Therefore, in the technical solution of this application, by correlating and calculating the causal risk probability with constraints, the safety boundary becomes an elastic barrier that can dynamically shrink or adjust based on the real-time calculated causal risk probability. In this way, it is ensured that even when the lower-level controller attempts to approach physical limits in pursuit of economic benefits (e.g., maximizing heat recovery rate), its actual operating trajectory will never enter the high-risk area determined by causal analysis, thereby avoiding dangerous conditions from the source.

[0038] In this process, firstly, the system identifies the associated constraints of causal risk probabilities. Specifically, the system analyzes the specific causal path that leads to an increased risk probability and identifies controllable process variables that can cut off the risk propagation path. For example, if the risk originates from "oil mist accumulation," the associated constraint variables might involve exhaust temperature and fresh air ratio; if the risk originates from "excessive back pressure," the associated variable might involve fan frequency. Secondly, the system calculates the constraint margin. During this process, the engine quantitatively adjusts the operational margins of these key variables based on the magnitude of the risk probability. The calculation logic follows an inverse proportionality principle: the higher the risk probability, the narrower the safety margin and the tighter the operational boundary. It is worth noting that this is not a simple on / off control, but a dynamic adjustment in numerical value. For example, when a causal path is activated and the risk is high, the system calculates a safety offset, compressing the originally broad physical boundary towards the safety side. Furthermore, all calculated and adjusted variable boundaries are integrated into a unified data package, namely the safety constraint set. This set specifically includes the maximum allowable flue gas temperature, the maximum allowable local surface temperature of the heat exchanger, the minimum fresh air supply ratio, the minimum total air volume or minimum fan frequency, the maximum allowable system back pressure or differential pressure, and the maximum allowable fouling level or minimum heat transfer efficiency. These calculated numerical sets together constitute dynamic safety control and are sent in real time to the lower-level Model Predictive Controller (MPC) as insurmountable hard constraints.

[0039] Specifically, S4 involves performing macroscopic policy optimization based on multi-objective reinforcement learning on the enhanced system state matrix and dynamic safety control constraints to obtain the macroscopic operating setpoint. It should be understood that traditional control methods often rely on human experience or simple PID rules, making it difficult to adapt to nonlinear fluctuations caused by changes in different materials, processes, and environments, and typically focusing only on immediate benefits while neglecting long-term risks. Therefore, in the technical solution of this application, the long-term policy planning capability of multi-objective reinforcement learning (MORL) is utilized to find the globally optimal operating strategy that maximizes long-term cumulative rewards within the safety boundary defined by the causal inference engine, thereby achieving an intelligent closed loop from local optimization to global coordination.

[0040] In practice, firstly, the enhanced system state matrix is ​​input into the trained conservative Q-network model to obtain a set of state-action values. The conservative Q-network model is trained offline based on millions of state-action-reward transition samples generated by an executable digital twin. It penalizes overestimation of unseen distributed actions by adding a regularization term to the objective function, thereby safely estimating the long-term value (Q-value) of taking various potential actions (such as adjusting fan frequency or fresh air ratio) in the current state, resulting in a set of state-action values.

[0041] Furthermore, based on dynamic safety control constraints, the optimal action selection and setpoint generation for the state-action value set with constraints are performed to obtain the macroscopic operating setpoint. The macroscopic operating setpoint includes the target exhaust temperature, target system back pressure / differential pressure, target fresh air / return air ratio, target scaling tolerance, and cleaning cycle trigger command. In this process, dynamic safety control constraints generated in real-time by the causal inference safety engine are introduced as filtering conditions. The candidate action space is traversed, and actions that would cause the system's operating trajectory to touch safety boundaries (e.g., violating the maximum allowable exhaust temperature or minimum total airflow constraints) are eliminated, even if these actions theoretically might bring extremely high short-term energy efficiency gains. Within the feasible domain that satisfies the safety constraints, the system selects the action combination with the highest Q value and decodes it into a physical-level control objective. Finally, the system outputs a set of macroscopic operating setpoints.

[0042] Specifically, in step S5, trajectory tracking and execution under model predictive control are performed on the macroscopic operating setpoint, dynamic safety control constraints, and enhanced system state matrix to obtain the actuator command sequence. It should be understood that the macroscopic setpoint generated by the upper-level multi-objective reinforcement learning (MORL) agent is static and discontinuous, and cannot directly drive the smooth operation of physical equipment such as fans and valves. Simultaneously, physical systems possess inertial, nonlinear, and time-varying characteristics, making it difficult for simple feedback control to handle complex dynamic constraints. Therefore, in the technical solution of this application, model predictive control (MPC) technology is introduced, utilizing its powerful real-time optimization and multivariate constraint handling capabilities to solve the execution problem. Specifically, by calculating the optimal control sequence that allows the system output to accurately approximate the macroscopic setpoint, while strictly adhering to the dynamic safety boundary defined by the causal inference engine, the control behavior is ensured to be both accurate and safe.

[0043] In practical implementation, firstly, based on the system dynamic model and the enhanced system state matrix, the future state trajectory of potential future control input sequences is predicted using the dynamic model to obtain the predicted state trajectory. During this process, the MPC controller receives the enhanced system state matrix from the GNN as the initial state at the current moment and uses the built-in system dynamic model (such as a reduced-order model based on physical mechanisms or a simplified version of xDT) to simulate and deduce the dynamic behavior of the system under potential future control input sequences in the prediction time domain. Through this step, the system can predict how the future exhaust gas temperature or back pressure will evolve if a certain combination of fan frequency or valve opening is adopted, thus obtaining the predicted state trajectory.

[0044] Next, constrained rolling time-domain optimization is performed on the predicted state trajectory, macroscopic operating setpoint, and dynamic safety control constraints to obtain the optimal control sequence. In this step, the MPC controller constructs and solves an optimization problem within a finite control time domain, aiming to find a set of optimal control input sequences (such as control increments in the next few tens of seconds) so that the system's future output trajectory can track the macroscopic operating setpoint given by the upper-level MORL at the lowest cost. The core of this optimization problem lies in minimizing an objective function that includes tracking error and control energy. This process can be expressed by the following formula: in, At the current moment Predicted future The system output at any given time, It is the target setpoint trajectory given by the upper-level MORL. It is the future Increment of control quantity at any given time and This is the weight matrix. Specifically, the optimization process must strictly satisfy a set of dynamic constraints: in, and It is not a fixed value, but a dynamic security control constraint generated in real time by the causal inference security engine in the preceding steps.

[0045] Furthermore, the optimal control sequence is extracted and executed in the rolling time domain to obtain the actuator instruction sequence. Specifically, after solving for the optimal control sequence for a future time period, the system does not execute the entire sequence, but only extracts the first control element in the sequence (i.e., the best action to be taken at the current moment), converts it into a specific voltage or current signal, and sends it to physical devices such as the frequency converter and actuator drive board. In the next control cycle, the system will use the latest state feedback to repeat the above prediction, optimization, and execution process, forming a closed-loop control.

[0046] Taking the scheme in this application as an example, assume that the upper-level MORL agent issues a macroscopic setpoint of "maintaining the flue gas temperature at 175°C". After receiving this setpoint, the MPC controller, combined with the current measured temperature of 170°C (from the enhanced state matrix), uses its built-in model to predict future trends. The controller calculates that to rapidly raise the temperature to 175°C, the fan frequency needs to be significantly increased. However, the causal engine simultaneously issues a dynamic safety constraint of "maximum permissible system pressure difference < 500Pa". When solving the optimization problem, the MPC automatically finds a compromise path: increasing the fan frequency at a relatively gentle rate to ensure that the predicted future pressure difference never exceeds 500Pa, while the temperature gradually approaches 175°C. The first optimal action calculated is "increasing the fan frequency by 0.5Hz", and this instruction is then sent to the frequency converter for execution. This process is repeated every second to ensure that the system precisely adheres to the target within the safety boundary.

[0047] In summary, the adaptive control method for waste heat recovery from textile setting machine flue gas according to the embodiments of this application is explained. First, it utilizes a graph neural network to perform global state estimation on multimodal sensor data to accurately perceive hidden conditions such as fouling distribution inside the heat exchanger. Then, it introduces causal inference technology for proactive risk assessment, generating dynamic safety control constraints by analyzing the causal relationships between variables, thereby replacing traditional passive threshold alarms. Within this safety boundary, the system uses multi-objective reinforcement learning to formulate a macroscopic operating strategy that balances energy efficiency and process stability, and combines model predictive control to achieve precise trajectory tracking and execution of the setpoint. In this way, it can achieve deep perception and proactive risk defense of complex flue gas conditions, maximizing waste heat recovery efficiency while ensuring the inherent safety of the system, and significantly reducing operation and maintenance costs and fire hazards.

[0048] Furthermore, an adaptive control system for waste heat recovery from flue gas in textile setting machines is also provided.

[0049] Figure 4 This is a block diagram of an adaptive control system for waste heat recovery from textile setting machine flue gas according to an embodiment of this application. Figure 4As shown, the adaptive control system 300 for waste heat recovery from textile setting machine flue gas according to an embodiment of this application includes: a raw data acquisition module 310 for acquiring raw data from multimodal sensors; a global state estimation module 320 for performing global state estimation based on graph neural networks on the raw data from multimodal sensors to obtain an enhanced system state matrix; an active risk assessment module 330 for performing active risk assessment based on causal inference on the enhanced system state matrix to obtain dynamic safety control constraints; a macro-strategy optimization module 340 for performing macro-strategy optimization based on multi-objective reinforcement learning on the enhanced system state matrix and dynamic safety control constraints to obtain a macro-operation setpoint; and a trajectory tracking and execution module 350 for performing trajectory tracking and execution under model predictive control on the macro-operation setpoint, dynamic safety control constraints, and enhanced system state matrix to obtain an actuator command sequence.

[0050] As described above, the adaptive control system 300 for waste heat recovery from textile setting machine flue gas according to embodiments of this application can be implemented in various wireless terminals, such as servers with adaptive control algorithms for waste heat recovery from textile setting machine flue gas. In one possible implementation, the adaptive control system 300 for waste heat recovery from textile setting machine flue gas according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the adaptive control system 300 for waste heat recovery from textile setting machine flue gas can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the adaptive control system 300 for waste heat recovery from textile setting machine flue gas can also be one of many hardware modules of the wireless terminal.

[0051] Alternatively, in another example, the adaptive control system 300 for waste heat recovery from the textile setting machine flue gas and the wireless terminal can also be separate devices, and the adaptive control system 300 for waste heat recovery from the textile setting machine flue gas can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0052] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A self-adaptive control method for flue gas waste heat recovery of a textile setting machine, characterized in that, The method comprises the following steps: obtaining multi-modal sensor raw data; performing global state estimation on the multi-modal sensor raw data based on a graph neural network to obtain an enhanced system state matrix; performing active risk assessment on the enhanced system state matrix based on causal inference to obtain dynamic safety control constraints; performing macroscopic strategy optimization on the enhanced system state matrix and the dynamic safety control constraints based on multi-objective reinforcement learning to obtain a macroscopic operation set point; performing trajectory tracking and execution under model predictive control on the macroscopic operation set point, the dynamic safety control constraints and the enhanced system state matrix to obtain an actuator instruction sequence.

2. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 1, characterized in that, The multi-modal sensor raw data includes temperature, pressure, flow, acoustic signal, oil mist concentration and infrared image pixel value.

3. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 2, characterized in that, The global state estimation on the multi-modal sensor raw data based on the graph neural network to obtain the enhanced system state matrix comprises the following steps: performing data cleaning and feature extraction on the multi-modal sensor raw data to obtain a standardized feature vector set; based on the standardized feature vector set, constructing initial graph state data; performing message passing-based GNN feature aggregation on the initial graph state data to obtain a high-dimensional node embedding vector set; performing soft measurement regression and parameter inversion on the high-dimensional node embedding vector set to obtain a fouling distribution map; performing state fusion on the fouling distribution map and the standardized feature vector set to obtain the enhanced system state matrix.

4. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 3, characterized in that, The soft measurement regression and parameter inversion on the high-dimensional node embedding vector set to obtain the fouling distribution map comprises the following steps: performing basic physical parameter regression on the high-dimensional node embedding vector to obtain a predicted scale layer thickness; based on the predicted scale layer thickness, performing physical prior guided feature enhancement on the high-dimensional node embedding vector to obtain a physically enhanced embedding vector; performing conditional sequential regression on the physically enhanced embedding vector to obtain a predicted attenuation factor; combining the predicted attenuation factor and the predicted scale layer thickness to obtain the fouling distribution map.

5. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 1, characterized in that, The active risk assessment on the enhanced system state matrix based on causal inference to obtain dynamic safety control comprises the following steps: performing state analysis and mapping on the enhanced system state matrix based on a system safety causal graph to obtain an instantiated causal graph; performing risk propagation probability calculation on the instantiated causal graph based on forward reasoning to obtain a causal risk probability; performing correlation constraint identification and constraint margin calculation on the causal risk probability to obtain a safety constraint set, wherein the safety constraint set constitutes the dynamic safety control.

6. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 5, characterized in that, The safety constraint set includes the maximum allowed flue gas temperature, the maximum allowed local heat exchanger surface temperature, the minimum fresh air supplement ratio, the minimum total air volume / lowest fan frequency, the maximum allowed system back pressure / differential pressure, and the maximum allowed fouling level / lowest heat transfer efficiency.

7. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 1, characterized in that, The macroscopic strategy optimization on the enhanced system state matrix and the dynamic safety control constraints based on multi-objective reinforcement learning to obtain a macroscopic operation set point comprises the following steps: inputting the enhanced system state matrix into a trained conservative Q network model to obtain a state-action value set; Based on the dynamic safety control constraints, optimal action selection with constraints on the state-action value set and set point generation are performed to obtain macro operation set points, wherein the macro operation set points include target flue gas temperature, target system back pressure / differential pressure, target fresh air / return air ratio, target fouling tolerance, and cleaning cycle trigger instruction.

8. The textile setting machine flue gas waste heat recovery adaptive control method according to claim 1, characterized in that, Trajectory tracking and execution under model predictive control are performed on the macro operation set points, dynamic safety control constraints, and enhanced system state matrix to obtain an actuator instruction sequence, including: Based on the system dynamic model and the enhanced system state matrix, future state trajectory prediction based on the dynamic model is performed on the potential future control input sequence to obtain a predicted state trajectory. Constrained rolling horizon optimization is performed on the predicted state trajectory, macro operation set points, and dynamic safety control constraints to obtain an optimal control sequence. The optimal control sequence is subjected to instruction extraction and rolling horizon execution to obtain an actuator instruction sequence.

9. A textile setting machine flue gas waste heat recovery adaptive control system characterized by, It includes: An original data acquisition module for acquiring multi-modal sensor original data; A global state estimation module for global state estimation based on a graph neural network to obtain an enhanced system state matrix from the multi-modal sensor original data; An active risk assessment module for active risk assessment based on causal inference on the enhanced system state matrix to obtain dynamic safety control constraints; A macro strategy optimization module for macro strategy optimization based on multi-objective reinforcement learning on the enhanced system state matrix and dynamic safety control constraints to obtain macro operation set points; A trajectory tracking and execution module for trajectory tracking and execution under model predictive control on the macro operation set points, dynamic safety control constraints, and enhanced system state matrix to obtain an actuator instruction sequence.

Images