Concrete production management method and system based on digital twin

Abstract

The present invention provides a concrete production management method and system based on digital twins, which relate to the field of digital twin technology, including real-time collection of health monitoring data of concrete production equipment through multiple sensors, and collection of full life cycle event data of concrete production equipment; pre-processing of the health monitoring data at the edge, and extraction of equipment status characteristics, based on the equipment status characteristics, using a pre-trained fault prediction model for real-time diagnosis, and at the same time building a multi-level digital twin model of the equipment on a cloud platform; based on the multi-level digital twin model, the pre-built physical degradation model and the data-driven model are integrated to build an equipment remaining life prediction model, and according to the obtained prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan, and the plan is converted into equipment control instructions to realize the automatic issuance and execution of the concrete production scheduling plan.

Description

Technical Field

[0001] The present invention relates to digital twin technology, and in particular to a concrete production management method and system based on digital twin. Background Art

[0002] As the most crucial foundational material in modern construction, concrete production management directly impacts project quality, schedule, and cost. However, traditional concrete production management relies primarily on manual experience and simple information technology. This leads to issues such as insufficient production status awareness, difficulty in process optimization, and low resource allocation efficiency, making it difficult to adapt to the increasingly complex demands of construction projects.

[0003] In recent years, the rise of digital twin technology has provided new insights for the intelligent transformation of concrete production management. By constructing virtual models of physical entities, digital twins enable real-time interaction and integration between the physical and digital worlds. These technologies have been widely applied in areas such as product design, manufacturing, and operations and maintenance. Introducing digital twins into concrete production management promises to overcome the limitations of traditional approaches, enabling global awareness, real-time simulation, dynamic optimization, and precise control of the production process. However, existing research primarily focuses on single steps or local applications in concrete production, lacking a comprehensive, multi-dimensional, and systematic approach. Numerous technical challenges remain in developing a digital twin model that covers the entire concrete production lifecycle, leverages its comprehensive capabilities in production monitoring, scheduling optimization, and quality control, and achieves dynamic, closed-loop optimization of manufacturing resource allocation and process execution. Therefore, there is an urgent need to explore new approaches to concrete production management based on digital twins to advance the intelligent and lean development of concrete production.

[0004] The purpose of this invention is to provide a concrete production management method and system based on digital twins to address the problems of low efficiency and serious waste of resources in traditional concrete production management, improve the intelligence level of concrete production management, and achieve the goals of cost reduction, efficiency improvement and green manufacturing. Summary of the Invention

[0005] The embodiments of the present invention provide a concrete production management method and system based on digital twins, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a concrete production management method based on digital twin, including:

[0008] The system collects health monitoring data of concrete production equipment in real time through various sensors, and uses IoT identification methods to collect event data throughout the entire life cycle of concrete production equipment. The collected health monitoring data and life cycle event data are transmitted to the edge and cloud platforms respectively.

[0009] At the edge, edge computing technology is used to pre-process the health monitoring data and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, and the annotation of the full life cycle event data is updated according to the diagnosis results. At the same time, the annotated full life cycle event data is used on the cloud platform to build a multi-level digital twin model of the equipment;

[0010] Based on the multi-level digital twin model, the pre-built physical degradation model and data-driven model are integrated to construct an equipment remaining life prediction model. The equipment remaining life prediction model is used to predict the remaining life of the equipment. According to the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan, which is then converted into equipment control instructions to realize the automatic issuance and execution of the concrete production scheduling plan.

[0011] In an optional embodiment,

[0012] At the edge, edge computing technology is used to pre-process the health monitoring data and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, including:

[0013] At the edge, we receive health monitoring data from concrete production equipment collected in real time by multiple sensors, extract multi-domain time-frequency features that represent the degradation status of the equipment, and use an attention mechanism to filter them to obtain a subset of key features.

[0014] Based on the obtained key feature subsets, a multi-view learning strategy is introduced to construct multi-physics field feature subspaces, multi-sensor location feature subspaces, and multi-degradation feature subspaces for different degradation modes from the perspectives of physical field, sensor location, and degradation mode.

[0015] In each subspace, a multi-view attention mechanism is used to screen the corresponding local key features. Different local key features are combined through an adaptive view weighting strategy to obtain a multi-view diagnostic feature set.

[0016] Based on the multi-view diagnostic feature set, a one-dimensional convolutional neural network is used at the edge to build an edge diagnostic model, which is used to determine whether the device is faulty.

[0017] If a fault occurs, a preliminary diagnosis of the fault mode is performed, and based on the preliminary diagnosis results, unknown complex faults with low confidence and the corresponding multi-view diagnostic features are uploaded to the cloud;

[0018] The cloud uses pre-trained fault diagnosis knowledge graphs to comprehensively consider equipment models, operating conditions, and degradation mechanisms to diagnose unknown complex faults. At the same time, the diagnosis results are used to optimize the fault diagnosis knowledge graphs, and the optimized fault diagnosis knowledge graphs are regularly embedded into the edge diagnosis model to guide edge diagnostic reasoning.

[0019] In an optional embodiment,

[0020] Using the annotated data from the entire life cycle of the equipment on the cloud platform, a multi-level digital twin model of the equipment is constructed, including:

[0021] Receiving annotated equipment lifecycle data in the cloud and obtaining 3D point cloud data therein, preprocessing the 3D point cloud data, converting the preprocessed 3D point cloud data into a triangular mesh surface model, and constructing a geometric model of the concrete production equipment using surface fitting technology;

[0022] Based on the geometric model of concrete production equipment, the finite element analysis method is used to perform numerical simulation and solve the stress state, vibration characteristics, and heat conduction characteristics of the concrete production equipment. The physical response data of the concrete production equipment under different working conditions is obtained, and the physical model of the equipment is constructed.

[0023] Obtain the condition monitoring parameters and performance parameters of concrete production equipment, use a gated recurrent unit neural network to establish a nonlinear mapping relationship between the condition monitoring parameters and performance parameters, and build an equipment degradation model for life prediction;

[0024] The equipment geometric model, equipment physical model and equipment degradation model are integrated to construct a multi-level digital twin model of concrete production equipment.

[0025] In an optional embodiment,

[0026] Based on the multi-level digital twin model, the pre-built equipment degradation model and data-driven model are integrated to build an equipment remaining life prediction model. The remaining life of the equipment is predicted using the equipment remaining life prediction model, including:

[0027] Based on the multi-level digital twin model, a data-driven model is constructed. The data-driven model uses a convolutional neural network as its basic structure and uses a multi-scale feature extraction mechanism to adaptively extract multi-scale degradation features related to the equipment degradation process from health monitoring data.

[0028] Based on the obtained multi-scale degradation features, a joint loss function combining fault diagnosis loss and remaining life prediction loss is constructed. The pre-built equipment degradation model is used as a regularization constraint term to train the data-driven model. The degradation trajectory output by the equipment degradation model is used as a priori constraint of the data-driven model to obtain prediction results that conform to the degradation trajectory. The prediction results are then used as feedback to optimize the equipment degradation model parameters. The iteration is repeated until the preset termination condition is met, and finally a trained equipment remaining life prediction model is obtained.

[0029] Based on the trained equipment remaining life prediction model, the fault diagnosis results are used as prior knowledge, and the Bayesian inference method is used to predict the remaining life of the component by combining physical priors and data-driven posteriori.

[0030] In an optional embodiment,

[0031] Based on the obtained multi-scale degradation characteristics, the calculation formula of the joint loss function combining fault diagnosis loss and remaining life prediction loss is constructed as follows:

[0032]

[0033] Among them, L represents the joint loss function, α represents the weight coefficient of the fault diagnosis model loss, N represents the number of devices, M represents the number of fault categories, and y ij Indicates the true label of whether the i-th device belongs to the j-th type of fault, represents the predicted value of the mean of the characteristic vector of the i-th device under the j-th type of fault, μ j represents the true value of the mean of the characteristic vector of the jth type of fault, β represents the weight coefficient of the remaining life prediction model loss, W p (·) represents the difference between the remaining life distributions, represents the predicted remaining life distribution of the i-th device, P ri represents the true remaining life distribution, γ represents the weight coefficient of the regularization term, W represents the parameter matrix of the model, and ||·||2 represents the L2 norm.

[0034] In an optional embodiment,

[0035] Based on the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan. This plan is then converted into equipment control instructions to automatically issue and execute the concrete production scheduling plan. This includes:

[0036] A multi-objective reinforcement learning optimization framework is constructed to model the concrete production process as a multi-objective Markov decision process. The multi-objective Markov decision process includes a state space, an action space, a state transition probability matrix, a multi-objective reward function, a discount factor, and an initial state distribution. The state space includes equipment states and process parameters, the action space includes equipment operation modes and production scheduling strategies, and the multi-objective reward function includes equipment remaining life, production efficiency, and energy consumption indicators.

[0037] Based on the constructed multi-objective Markov decision process, a hierarchical target learning algorithm is designed. A value function approximator and a policy function approximator are constructed for each optimization objective of the multi-objective reward function. The value function and policy function of each optimization objective are estimated using the value function approximator and the policy function approximator. A target improvement operator is obtained based on the obtained value function estimation value and policy function estimation value. The multi-objective reward function is updated based on the target improvement operator. Based on the updated multi-objective reward function, a non-dominated sorting genetic algorithm is used to search in the multi-objective Markov decision process to obtain an initial policy set.

[0038] Based on the obtained initial strategy set, in the digital twin model, by building an interactive interface between the intelligent agent and the virtual environment, the Monte Carlo tree search algorithm is used to select the optimal exploration path through an adaptive exploration-exploitation equilibrium strategy, generate large-scale optimization strategy sample data, and obtain the optimal strategy set;

[0039] The optimal strategy set is input into the interactive decision recommendation system of the target learning algorithm, and the feasibility of the strategy is evaluated by combining expert knowledge to generate the final optimal concrete production scheduling plan;

[0040] The optimal concrete production scheduling plan is converted into equipment control instructions, and the automatic issuance and execution of the concrete production scheduling plan is achieved through the interactive interface between the digital twin model and the actual production management system.

[0041] In an optional embodiment,

[0042] A value function approximator and a policy function approximator are constructed for each optimization objective of the multi-objective reward function. The value function and policy function of each optimization objective are estimated using the value function approximator and the policy function approximator. The target improvement operator is obtained based on the obtained value function estimation value and policy function estimation value. The calculation formula for updating the multi-objective reward function based on the target improvement operator is as follows:

[0043]

[0044] Among them, R′ k represents the updated multi-objective reward function, R krepresents the initial multi-objective reward function, λ represents the learning rate, A represents the action space, δ represents the discount factor, S represents the state space, P(·) represents the state transition probability matrix, represents the value function of the kth optimization objective estimated by the value function approximator at state s′, represents the value function of the kth optimization objective estimated by the value function approximator at state s, ω represents the adjustment factor that balances the reward function and the policy function, represents the probability of the kth optimization objective choosing action a′ in state s estimated by the policy function approximator, represents the probability of selecting action a in state s as estimated by the policy function approximator for the kth optimization objective.

[0045] According to a second aspect of the embodiments of the present invention,

[0046] Provide a concrete production management system based on digital twin, including:

[0047] The first unit is used to collect health monitoring data of concrete production equipment in real time through multiple sensors, and uses IoT identification methods to collect full life cycle event data of concrete production equipment, and transmit the collected health monitoring data and full life cycle event data to the edge and cloud platforms respectively;

[0048] The second unit is used to pre-process the health monitoring data using edge computing technology at the edge end, and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, and the annotation of the full life cycle event data is updated according to the diagnosis results. At the same time, the annotated full life cycle event data is used on the cloud platform to build a multi-level digital twin model of the equipment;

[0049] The third unit is used to integrate the pre-built physical degradation model and data-driven model based on the multi-level digital twin model, build an equipment remaining life prediction model, and use the equipment remaining life prediction model to predict the remaining life of the equipment. According to the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan, and the plan is converted into equipment control instructions to realize the automatic issuance and execution of the concrete production scheduling plan.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of the embodiments of the present invention,

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0057] In this embodiment, multiple sensors collect real-time equipment health monitoring data, and IoT identification methods are used to collect full lifecycle event data. This allows for comprehensive perception and recording of equipment operating status, fault events, maintenance records, and more, providing comprehensive data support for equipment health management. Edge computing technology enables real-time data preprocessing and fault diagnosis at the edge, close to the equipment, improving response speed and real-time performance while reducing computing pressure on the cloud. Using a Bayesian inference framework, a pre-built physical degradation model is integrated with a data-driven model based on monitoring data to form a hybrid model for remaining life prediction. This model leverages both prior knowledge and data posteriori, improving prediction accuracy and reliability. Based on the remaining life prediction results, combined with advanced technologies such as reinforcement learning and evolutionary algorithms, the concrete production scheduling strategy is dynamically optimized within the digital twin model. While satisfying multiple objective constraints, the optimal production scheduling solution is obtained, improving production efficiency, saving energy, and extending equipment life. This provides an intelligent, automated, and optimized solution for complex industrial processes such as concrete production. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the process of a concrete production management method based on digital twins according to an embodiment of the present invention;

[0059] Figure 2 This is a structural diagram of a concrete production management system based on digital twins according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0061] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a concrete production management method based on digital twins according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S 1 01. Collect health monitoring data of concrete production equipment in real time through multiple sensors, and use the Internet of Things identification method to collect full life cycle event data of concrete production equipment, and transmit the collected health monitoring data and full life cycle event data to the edge and cloud platforms respectively.

[0064] For example, multi-sensor fusion and industrial Internet of Things technologies are first used to collect multi-dimensional, full-life cycle data on the equipment's operating status. Vibration, temperature, pressure, strain and other sensors are deployed at key parts of the equipment, such as bearings, reducers, motors, etc., to collect dynamic response data of the equipment in real time. Using Internet of Things identification technologies such as radio frequency identification (RFID) and data matrix codes, data is recorded and associated with events throughout the equipment's life cycle, including manufacturing, assembly, and maintenance, to achieve interconnection and interoperability of data throughout the equipment's life cycle. At the same time, through communication technologies such as industrial Ethernet and wireless sensor networks, multi-source heterogeneous equipment status data is connected to the cloud platform to build a big data center for the equipment's full life cycle.

[0065] In this embodiment, equipment health monitoring data is collected in real time through multiple sensors, and full life cycle event data is obtained using IoT identification technology. This can fully perceive the operating status, failure mode, maintenance history, etc. of concrete production equipment, providing complete data support for equipment health management.

[0066] S102. Pre-process the health monitoring data at the edge using edge computing technology, extract the status characteristics of the equipment, perform real-time diagnosis based on the status characteristics of the equipment using a pre-trained fault prediction model, and update the annotation of the full life cycle event data based on the diagnosis results. At the same time, use the annotated full life cycle event data on the cloud platform to build a multi-level digital twin model of the equipment.

[0067] For example, edge computing technology is used to enable real-time data processing and feature extraction at the data acquisition end. Edge computing devices such as smart gateways and embedded systems are deployed at production sites to preprocess raw sensor data such as vibration and temperature, extracting health monitoring features in the time, frequency, and time-frequency domains. Signal processing methods such as wavelet transform and empirical mode decomposition are used to denoise and enhance vibration signals and extract features. Statistical indicators such as kurtosis and root mean square (RMS) are used to characterize the degradation trend of equipment health. Using deep learning-based end-side inference technology, pre-trained convolutional neural network models are deployed on gateway devices to achieve real-time diagnosis and early warning of equipment failure modes.

[0068] When the fault diagnosis model on the edge outputs a result, it is assigned a corresponding confidence score, indicating the degree of confidence in the fault type judgment. A confidence threshold is set. If the confidence level of the diagnosis result exceeds the threshold, it is considered a "reliable result" and can be directly used to update the event data annotation. If the confidence level is below the threshold, it is considered an "uncertain result" and needs to be uploaded to the cloud for further diagnosis and manual confirmation.

[0069] For "reliable results", the edge will associate the time point, fault type, location and other information of the fault with the device ID, construct a label string, upload the constructed label string to the cloud, and update the full life cycle event data file of the device stored in the cloud; at the same time, for "uncertain results", the edge will upload the corresponding feature data, preliminary diagnosis results, etc. to the cloud. The cloud uses more complex models for in-depth diagnosis and combines the experience of human experts to make judgments to obtain confirmed fault types.

[0070] The confirmed fault information is also constructed as a label string and updated to the full life cycle event data file. It can continuously record various events and faults occurring in the equipment in the full life cycle event data, laying a solid data foundation for building a high-fidelity digital twin model.

[0071] In an optional embodiment,

[0072] At the edge, edge computing technology is used to pre-process the health monitoring data and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, including:

[0073] At the edge, we receive health monitoring data from concrete production equipment collected in real time by multiple sensors, extract multi-domain time-frequency features that represent the degradation status of the equipment, and use an attention mechanism to filter them to obtain a subset of key features.

[0074] Based on the obtained key feature subsets, a multi-view learning strategy is introduced to construct multi-physics field feature subspaces, multi-sensor location feature subspaces, and multi-degradation feature subspaces for different degradation modes from the perspectives of physical field, sensor location, and degradation mode.

[0075] In each subspace, a multi-view attention mechanism is used to screen the corresponding local key features. Different local key features are combined through an adaptive view weighting strategy to obtain a multi-view diagnostic feature set.

[0076] Based on the multi-view diagnostic feature set, a one-dimensional convolutional neural network is used at the edge to build an edge diagnostic model, which is used to determine whether the device is faulty.

[0077] If a fault occurs, a preliminary diagnosis of the fault mode is performed, and based on the preliminary diagnosis results, unknown complex faults with low confidence and the corresponding multi-view diagnostic features are uploaded to the cloud;

[0078] The cloud uses pre-trained fault diagnosis knowledge graphs to comprehensively consider equipment models, operating conditions, and degradation mechanisms to diagnose unknown complex faults. At the same time, the diagnosis results are used to optimize the fault diagnosis knowledge graphs, and the optimized fault diagnosis knowledge graphs are regularly embedded into the edge diagnosis model to guide edge diagnostic reasoning.

[0079] For example, health monitoring data, including time series data and spectrum data, is collected in real time from various sensors installed on concrete production equipment (such as vibration sensors, temperature sensors, and pressure sensors). The raw data is preprocessed, such as through denoising and interpolation, to eliminate interference and invalid data.

[0080] Extract multi-domain time-frequency features that characterize the device degradation state. These include time-domain features such as mean, peak, peak-to-peak, and root-mean-square amplitude; frequency-domain features such as power spectral density, frequency center, and band power; and time-frequency features such as wavelet packet decomposition energy and wavelet energy entropy. An attention mechanism is used to perform weighted fusion on the extracted multi-domain features to obtain a subset of key features.

[0081] Based on a subset of key features, a multi-view learning strategy is introduced. From the perspective of physical fields, a multi-physics field feature subspace is constructed. Features corresponding to different physical fields (such as vibration, temperature, and pressure) are grouped and modeled and learned separately. Unsupervised methods such as autoencoders are used to learn feature representations for each physical field subspace. From the perspective of sensor position, a multi-sensor position feature subspace is constructed. Features corresponding to different sensor positions are grouped. Feature learning is performed on each position subspace to capture position-related degradation patterns. From the perspective of degradation patterns, multiple degradation feature subspaces are constructed. Features are grouped according to known degradation patterns (such as wear, cracks, and corrosion). A dedicated feature extraction model is trained for each degradation pattern subspace.

[0082] In each feature subspace, a multi-view attention mechanism is used to screen the corresponding local key features. Local key features from different viewpoints are combined using an adaptive view weighting strategy to obtain a multi-view diagnostic feature set. Specifically, for each feature subspace, a self-attention mechanism is used to capture long-range dependencies between features. The attention weights of each feature and other features are calculated, and high-weighted features are used as local key features. A gated recurrent unit (GRU) is used to perform temporal modeling of the local key features of each viewpoint. Adaptive weights are assigned to different viewpoints using the view attention mechanism. The weighted local key features are concatenated to obtain a multi-view diagnostic feature set.

[0083] At the edge, a one-dimensional convolutional neural network is used to build an edge diagnosis model. The input layer receives the multi-view diagnostic feature set, the convolution layer includes multiple layers of 1D convolution layers, each layer contains convolution, batch normalization and activation functions (such as ReLU), the pooling layer is used for feature downsampling and translation invariance, the fully connected layer performs nonlinear transformation and mapping on the features, and the output layer sets the number of nodes according to the number of fault types, such as 1 node for binary classification and multiple nodes for multi-classification. The multi-view diagnostic feature set is used as the input of the one-dimensional convolutional neural network, and after multiple layers of convolution and pooling operations, the fault diagnosis results are output. The fault diagnosis model is trained through supervised learning and the model is optimized using labeled data. The trained fault diagnosis model is used to perform online fault diagnosis on the real-time status of the equipment.

[0084] If the edge diagnosis model determines that a fault has occurred but the confidence level is low (i.e., it is an unknown complex fault), the multi-view diagnostic features are uploaded to the cloud. The cloud uses a pre-trained fault diagnosis knowledge graph, combined with information such as equipment model, operating conditions, and degradation mechanism, to diagnose unknown complex faults. The diagnostic knowledge graph contains knowledge such as equipment failure mode, failure characteristics, and diagnostic logic. Among them, node types include equipment model, component, operating condition, degradation mode, failure mode, etc. The relationship types of edges include "contains", "leads to", "manifests as", etc. Based on the multi-view features uploaded by the edge, the fault mode path that best matches it is found, and the knowledge graph reasoning module (such as GNNs) is used to infer the cause of the fault and the propagation path.

[0085] After reasoning is complete, new fault cases are added to the knowledge graph to continuously enrich the knowledge base. Relationship weights in the knowledge graph are adjusted based on diagnostic results. A federated transfer learning strategy is also employed to leverage diagnostic knowledge from different devices and scenarios in the cloud to continuously expand and optimize the knowledge graph. This knowledge graph is then regularly embedded into edge diagnostic models to guide edge diagnostic reasoning.

[0086] Optionally, for samples with low diagnostic confidence at the edge, the most valuable samples are adaptively labeled by maximizing diagnostic entropy, and incremental labeled samples are regularly uploaded to the cloud for fine-tuning of the diagnostic model. Under the federated learning framework, different edge nodes use mechanisms such as encrypted communication and differential privacy to achieve secure sharing of incremental samples and model parameters. Without exposing the original data, the labeled data of other nodes are used to enhance the discriminative ability of the local diagnostic model. The cloud uses the federated distillation algorithm to compress the globally aggregated incremental samples into a small-scale distillation set, selectively inheriting the optimization results of distributed incremental training, improving the generalization of the cloud diagnostic model while ensuring privacy and security, and transmitting the optimized diagnostic knowledge back to the edge to achieve the co-evolution of the end-cloud diagnostic model. This incremental training framework introduces active learning at the edge to achieve autonomous expansion of diagnostic samples; it integrates federated learning in the cloud to achieve distributed sharing of knowledge, improving the efficiency and generalization of distributed diagnostic model learning in complex industrial scenarios.

[0087] In this embodiment, edge computing technology is used to process monitoring data and diagnose faults at the nearest edge node, which greatly improves the real-time nature and response speed of diagnosis. Extracting key features based on the attention mechanism and multi-view learning strategy is conducive to accurately capturing the degradation state of the equipment and improving the accuracy of fault diagnosis. Most common faults are diagnosed at the edge, and key features are uploaded to the cloud only when unknown complex faults are encountered, which greatly reduces the amount of data transmission. The pre-trained fault diagnosis knowledge graph is used in the cloud, which integrates knowledge in the fields of equipment model, working conditions, degradation mechanism, etc., improves the diagnostic ability of unknown complex faults, and can give full play to the different advantages of the edge and cloud to achieve intelligent and efficient fault diagnosis.

[0088] In an optional embodiment,

[0089] Using the annotated data from the entire life cycle of the equipment on the cloud platform, a multi-level digital twin model of the equipment is constructed, including:

[0090] Receiving annotated equipment lifecycle data in the cloud and obtaining 3D point cloud data therein, preprocessing the 3D point cloud data, converting the preprocessed 3D point cloud data into a triangular mesh surface model, and constructing a geometric model of the concrete production equipment using surface fitting technology;

[0091] Based on the geometric model of concrete production equipment, the finite element analysis method is used to perform numerical simulation and solve the stress state, vibration characteristics, and heat conduction characteristics of the concrete production equipment. The physical response data of the concrete production equipment under different working conditions is obtained, and the physical model of the equipment is constructed.

[0092] Obtain the condition monitoring parameters and performance parameters of concrete production equipment, use a gated recurrent unit neural network to establish a nonlinear mapping relationship between the condition monitoring parameters and performance parameters, and build an equipment degradation model for life prediction;

[0093] The equipment geometric model, equipment physical model and equipment degradation model are integrated to construct a multi-level digital twin model of concrete production equipment.

[0094] For example, 3D point cloud data is first acquired from the equipment's full lifecycle data. Preprocessing of the raw point cloud data involves denoising, padding, and downsampling to eliminate outliers and redundant data. Using geometric processing algorithms such as spherical projection and RANSAC, the preprocessed point cloud data is segmented to extract the equipment's main components. For each component, a triangular mesh surface model is fitted using surface reconstruction algorithms such as moving least squares and Gaussian process regression. The surface models of each component are assembled to construct a 3D geometric model of the entire concrete production equipment.

[0095] Based on the geometric model, finite element analysis software is used to model the physical characteristics of the equipment. The equipment's material properties and boundary conditions are defined. Material properties include density, elastic modulus, and thermal conductivity. Static analysis is performed on the equipment's stress state, calculating stress and strain distributions. Modal analysis is also performed on the equipment's vibration characteristics to determine natural frequencies and modal shapes. Thermal analysis is performed on the equipment's heat conduction characteristics, calculating the temperature field distribution. The physical response data (such as stress, displacement, and temperature) under different operating conditions is then exported to construct a physical model of the equipment.

[0096] Collect the equipment's condition monitoring parameters (such as vibration, temperature, and pressure) and performance parameters (such as production capacity, energy consumption, and failure rate). Use a gated recurrent unit neural network to establish a mapping relationship between the condition monitoring parameters and performance parameters. The GRU network structure includes an input layer, multiple layers of GRU units, a fully connected layer, and an output layer. Use the condition monitoring parameter sequence as the GRU input and the performance parameters as supervisory labels for model training. The training goal is to minimize the prediction error of the performance parameters and obtain the optimal mapping model. Utilize the trained GRU network model to predict the equipment's performance parameters based on the condition monitoring parameters and construct a degradation model.

[0097] The device's geometric model, physical model, and degradation model are integrated into a unified model framework, and collaborative simulation technology is used to enable data interaction between the three sub-models. The geometric model provides 3D visualization and spatial information of the device, while the physical model calculates the device's physical response based on operating conditions. This serves as input to the degradation model, which predicts the device's performance parameters based on the physical response and provides feedback to the physical model to update the operating conditions. The three sub-models interact and iterate, forming a digital twin of the device's entire lifecycle.

[0098] In this embodiment, the accurate geometric shape of the equipment is reconstructed using three-dimensional point cloud data, and a detailed geometric digital twin model is constructed through surface fitting technology, laying the foundation for subsequent physical simulation and visualization analysis. Based on the finite element analysis method, the physical processes such as the stress state, vibration characteristics and heat conduction of the equipment are accurately simulated, which can truly reproduce the physical response of the equipment under different working conditions and provide support for safety and reliability analysis. Using deep learning technologies such as gated recurrent unit neural networks, a nonlinear mapping relationship between equipment status and performance parameters is established from monitoring data, which can predict the degree of aging and remaining life of the equipment, providing a basis for formulating preventive maintenance plans. By organically integrating geometric models, physical models and degradation models, a digital twin model of concrete production equipment that integrates multi-level knowledge is constructed. It not only has visualization and physical simulation functions, but also can predict equipment life and support equipment life cycle management.

[0099] S 1 03. Based on the multi-level digital twin model, the pre-built equipment degradation model and data-driven model are integrated to construct an equipment remaining life prediction model. The equipment remaining life prediction model is used to predict the remaining life of the equipment. According to the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan, which is then converted into equipment control instructions to realize the automatic issuance and execution of the concrete production scheduling plan.

[0100] In an optional embodiment,

[0101] Based on the multi-level digital twin model, the pre-built equipment degradation model and data-driven model are integrated to build an equipment remaining life prediction model. The remaining life of the equipment is predicted using the equipment remaining life prediction model, including:

[0102] Based on the multi-level digital twin model, a data-driven model is constructed. The data-driven model uses a convolutional neural network as its basic structure and uses a multi-scale feature extraction mechanism to adaptively extract multi-scale degradation features related to the equipment degradation process from health monitoring data.

[0103] Based on the obtained multi-scale degradation features, a joint loss function combining fault diagnosis loss and remaining life prediction loss is constructed. The pre-built equipment degradation model is used as a regularization constraint term to train the data-driven model. The degradation trajectory output by the equipment degradation model is used as a priori constraint of the data-driven model to obtain prediction results that conform to the degradation trajectory. The prediction results are then used as feedback to optimize the equipment degradation model parameters. The iteration is repeated until the preset termination condition is met, and finally a trained equipment remaining life prediction model is obtained.

[0104] Based on the trained equipment remaining life prediction model, the fault diagnosis results are used as prior knowledge, and the Bayesian inference method is used to predict the remaining life of the component by combining physical priors and data-driven posteriori.

[0105] For example, a data-driven model is first constructed based on the digital twin model. This data-driven model uses a convolutional neural network as its underlying structure, consisting of multiple convolutional layers, pooling layers, and fully connected layers, to automatically extract features from the raw input. A multi-scale feature extraction mechanism is introduced, capturing degradation features at different scales using convolution kernels of different sizes. Equipment health monitoring data (such as vibration and temperature) is used as input to the data-driven model. Through end-to-end training, the model adaptively learns multi-scale features related to physical degradation processes.

[0106] A joint loss function is designed, consisting of a fault diagnosis loss and a remaining life prediction loss. The fault diagnosis loss is based on a categorical cross-entropy loss, used to distinguish between faults and normal states. The remaining life prediction loss uses a regression loss function such as mean squared error or smoothed L1 loss. The output of a pre-built equipment degradation model is incorporated as a regularization constraint into the joint loss function. By minimizing the joint loss function, a data-driven model is trained to simultaneously meet the objectives of fault diagnosis and life prediction.

[0107] The device degradation model is iteratively integrated with the data-driven model. Specifically, the parameters of the device degradation model are initialized and pre-trained based on existing knowledge or historical data. The degradation trajectory output by the device degradation model serves as a priori constraints for the data-driven model. When training the data-driven model, the weights of the loss function are adjusted based on the degradation trajectory to optimize the model output. The prediction results of the data-driven model are then used as feedback to update the parameters of the device degradation model. This iterative process is repeated until the model converges or meets the preset termination criteria, ultimately resulting in a hybrid degradation model that integrates the physical model and the data-driven model.

[0108] The fault diagnosis module obtains information about the equipment's fault type and occurrence time, encoding this information into a probability distribution as one of the prior distributions in the Bayesian inference framework. For example, the fault type can be encoded using a one-hot encoding, and the fault occurrence time can be modeled using a Gaussian distribution. The equipment degradation model describes the degradation process of the equipment under different operating conditions and material properties. Based on the equipment model and process parameters, the parameter values or distribution of the degradation model are determined. These physical parameters and their distributions are used as another prior distribution in the Bayesian inference framework.

[0109] The trained data-driven model is used to predict the remaining life of new monitoring data. The output of the data-driven model is a point estimate of the remaining life or a parameterized distribution (such as a Gaussian distribution). The prediction results of the data-driven model are used as the likelihood function or posterior distribution of the Bayesian inference framework.

[0110] Based on Bayes' theorem, the fault prior, physical prior, and data posteriori are combined. Since the posterior distribution often has no analytical solution, sampling methods are needed to approximate it. Markov Monte Carlo sampling (MCMC) is a commonly used sampling method. Through MCMC, a large number of remaining life samples are drawn from the posterior distribution to approximately reconstruct the posterior distribution.

[0111] From the posterior distribution obtained by sampling, the mean or median of the remaining life is calculated as a point estimate. Based on the distribution of the sample, a confidence interval of the remaining life, such as a 95% credible interval, is determined. The confidence interval can quantify the uncertainty of the prediction and provide a basis for subsequent decision-making.

[0112] In an optional embodiment,

[0113] Based on the obtained multi-scale degradation characteristics, the calculation formula of the joint loss function combining fault diagnosis loss and remaining life prediction loss is constructed as follows:

[0114]

[0115] Among them, L represents the joint loss function, α represents the weight coefficient of the fault diagnosis model loss, N represents the number of devices, M represents the number of fault categories, and y ij Indicates the true label of whether the i-th device belongs to the j-th type of fault, represents the predicted value of the mean of the characteristic vector of the i-th device under the j-th type of fault, μ j represents the true value of the mean of the characteristic vector of the jth type of fault, β represents the weight coefficient of the remaining life prediction model loss, W p (·) represents the difference between the remaining life distributions, represents the predicted remaining life distribution of the i-th device, P ri represents the true remaining life distribution, γ represents the weight coefficient of the regularization term, W represents the parameter matrix of the model, and ||·||2 represents the L2 norm.

[0116] In this embodiment, a physics-based degradation model and a data-driven model are integrated, leveraging the strengths of both to more accurately describe and predict equipment degradation. The data-driven model adaptively extracts multi-scale degradation features from monitoring data using a convolutional neural network, helping to uncover complex nonlinear degradation patterns. The degradation model provides physical constraints, ensuring that predictions conform to the degradation trajectory and avoiding the deviations and outliers associated with purely data-driven models. Using a Bayesian inference framework, a pre-built physical degradation model is used as prior knowledge, and the output of the data-driven model serves as a posteriori evidence. By effectively integrating the physical prior with the data posteriori, the stability and reliability of the predictions can be effectively improved. By constructing a joint loss function that organically combines the fault diagnosis loss and the remaining life prediction loss, the resulting hybrid degradation model is able to describe complex nonlinear degradation processes and improve the generalization performance of the predictions. This technical solution provides an integrated life prediction solution, providing technical support for equipment lifecycle management, thereby promoting the intelligent upgrade of the manufacturing industry, improving equipment utilization efficiency and reliability, and reducing operation and maintenance costs.

[0117] In an optional embodiment,

[0118] Based on the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan. This plan is then converted into equipment control instructions to automatically issue and execute the concrete production scheduling plan. This includes:

[0119] A multi-objective reinforcement learning optimization framework is constructed to model the concrete production process as a multi-objective Markov decision process. The multi-objective Markov decision process includes a state space, an action space, a state transition probability matrix, a multi-objective reward function, a discount factor, and an initial state distribution. The state space includes equipment states and process parameters, the action space includes equipment operation modes and production scheduling strategies, and the multi-objective reward function includes equipment remaining life, production efficiency, and energy consumption indicators.

[0120] Based on the constructed multi-objective Markov decision process, a hierarchical target learning algorithm is designed. A value function approximator and a policy function approximator are constructed for each optimization objective of the multi-objective reward function. The value function and policy function of each optimization objective are estimated using the value function approximator and the policy function approximator. A target improvement operator is obtained based on the obtained value function estimation value and policy function estimation value. The multi-objective reward function is updated based on the target improvement operator. Based on the updated multi-objective reward function, a non-dominated sorting genetic algorithm is used to search in the multi-objective Markov decision process to obtain an initial policy set.

[0121] Based on the obtained initial strategy set, in the digital twin model, by building an interactive interface between the intelligent agent and the virtual environment, the Monte Carlo tree search algorithm is used to select the optimal exploration path through an adaptive exploration-exploitation equilibrium strategy, generate large-scale optimization strategy sample data, and obtain the optimal strategy set;

[0122] The optimal strategy set is input into the interactive decision recommendation system of the target learning algorithm, and the feasibility of the strategy is evaluated by combining expert knowledge to generate the final optimal concrete production scheduling plan;

[0123] The optimal concrete production scheduling plan is converted into equipment control instructions, and the automatic issuance and execution of the concrete production scheduling plan is achieved through the interactive interface between the digital twin model and the actual production management system.

[0124] Exemplarily, a multi-objective reinforcement learning optimization framework is first constructed to model the concrete production process as a multi-objective Markov decision process. Define the state space, which includes equipment status (such as temperature, vibration, etc.) and process parameters (such as material properties, environmental conditions, etc.). Define the action space, which includes equipment operation modes (such as on / off, load regulation, etc.) and production scheduling strategies (such as capacity adjustment, process optimization, etc.). Determine the state transition probability matrix, which describes the probability distribution of transitioning from the current state to the next state after taking a certain action. It can be estimated based on a physical model or historical data. Construct a multi-objective reward function, which includes indicators such as remaining life of equipment, production efficiency and energy consumption, and set weights according to priority. Set a discount factor to balance immediate rewards and long-term cumulative rewards. Determine the initial state distribution to describe the starting state of the production task.

[0125] Design a hierarchical multi-objective learning algorithm and construct corresponding value function approximators and policy function approximators for each optimization objective. The value function approximator can use a deep neural network structure, such as a fully connected network or a long short-term memory network. The policy function approximator can also use a deep neural network, and the output is an action probability distribution or a deterministic action.

[0126] The value and policy functions are estimated using temporal difference (TD) or policy gradient algorithms. Specifically, for each objective, a TD learning algorithm or policy gradient algorithm is used to update the parameters of the value and policy function approximators using the state-action-reward sequence data generated by interaction with the virtual environment. Techniques such as experience replay and target networks can also be used to improve training stability and convergence.

[0127] Based on the estimated value function and policy function, the target improvement operator (such as gradient or Monte Carlo estimation) corresponding to each target is calculated. The improvement operators of each target are linearly combined according to the weights to form a multi-target improvement operator. The multi-target improvement operator is used to update the multi-target reward function through policy gradient ascent or other optimization methods.

[0128] Multi-objective optimization is performed based on the updated reward function. Optionally, a non-dominated sorting genetic algorithm is used to optimize the updated multi-objective reward function. The non-dominated sorting genetic algorithm performs selection, crossover, and mutation operations based on the non-dominated sorting and crowding distance of individuals in the population. Through iterative evolution, it obtains an initial set of strategies that achieves a balance between multiple objectives.

[0129] The digital twin model serves as a virtual environment where intelligent agents can observe states and perform actions. State observation interfaces and action execution interfaces are defined to enable bidirectional data exchange. The agent selects an action based on the current state, and the environment updates the next state based on the action and returns a reward. A Monte Carlo tree search algorithm is embedded in the agent's decision-making mechanism. In the Monte Carlo tree, each node represents a state, and edges represent actions and their transition probabilities. An adaptive exploration-exploitation equilibrium strategy (such as Unified Combination or Progressive Knowledge Exploitation) is used to select the optimal path to expand the tree nodes. Through a large number of simulated interactions, sample data for optimized strategies is generated, and the optimal strategy set is selected.

[0130] The optimized strategy set is fed into an interactive decision-making recommendation system based on a target learning algorithm. Combined with domain knowledge from human experts, the feasibility, risks, and constraints of each strategy are evaluated. Based on the evaluation results, the optimal concrete production scheduling solution is finally selected from the strategy set.

[0131] The optimal scheduling plan is converted into a specific equipment control instruction sequence. Through the two-way interactive interface between the digital twin model and the actual production management system, the control instructions are automatically sent to the execution system at the production site, realizing real-time execution of the scheduling plan, real-time monitoring of the production process, and dynamic adjustment of the scheduling strategy based on feedback information.

[0132] In an optional embodiment,

[0133] A value function approximator and a policy function approximator are constructed for each optimization objective of the multi-objective reward function. The value function and policy function of each optimization objective are estimated using the value function approximator and the policy function approximator. The target improvement operator is obtained based on the obtained value function estimation value and policy function estimation value. The calculation formula for updating the multi-objective reward function based on the target improvement operator is as follows:

[0134]

[0135] Among them, R′ k represents the updated multi-objective reward function, R k represents the initial multi-objective reward function, λ represents the learning rate, A represents the action space, δ represents the discount factor, S represents the state space, P(·) represents the state transition probability matrix, represents the value function of the kth optimization objective estimated by the value function approximator at state s′, represents the value function of the kth optimization objective estimated by the value function approximator at state s, ω represents the adjustment factor that balances the reward function and the policy function, represents the probability of the kth optimization objective choosing action a′ in state s estimated by the policy function approximator, represents the probability of selecting action a in state s as estimated by the policy function approximator for the kth optimization objective.

[0136] In this embodiment, the concrete production process is modeled as a multi-objective Markov decision process, comprehensively considering multiple optimization objectives such as equipment remaining life, production efficiency, and energy consumption. This allows for simultaneous optimization of multiple objective reward functions, resulting in an optimal decision that balances these multiple objectives. By incorporating advanced algorithms such as Monte Carlo tree search and the non-dominated sorting genetic algorithm, large-scale optimization samples are generated through interactive exploration between the agent and the virtual environment, enabling the generation of highly intelligent optimal decision solutions. The Markov decision process state is dynamically updated based on the latest information, optimizing the scheduling plan in real time and achieving dynamic adaptive optimization of production scheduling. Dynamically optimized scheduling decision solutions can be directly converted into equipment control instructions, enabling automated management and control of the production process. Through modeling, simulation, optimization, and the issuance of control instructions, a highly intelligent digital twin model is integrated into the entire actual concrete production process, demonstrating the enormous potential of digital twin technology in intelligent manufacturing and promoting the intelligent upgrade of the manufacturing industry.

[0137] Figure 2 This is a structural diagram of a concrete production management system based on digital twins according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0138] The first unit is used to collect health monitoring data of concrete production equipment in real time through multiple sensors, and uses IoT identification methods to collect full life cycle event data of concrete production equipment, and transmit the collected health monitoring data and full life cycle event data to the edge and cloud platforms respectively;

[0139] The second unit is used to pre-process the health monitoring data using edge computing technology at the edge end, and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, and the annotation of the full life cycle event data is updated according to the diagnosis results. At the same time, the annotated full life cycle event data is used on the cloud platform to build a multi-level digital twin model of the equipment;

[0140] The third unit is used to integrate the pre-built physical degradation model and data-driven model based on the multi-level digital twin model, build an equipment remaining life prediction model, and use the equipment remaining life prediction model to predict the remaining life of the equipment. According to the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan, and the plan is converted into equipment control instructions to realize the automatic issuance and execution of the concrete production scheduling plan.

[0141] According to a third aspect of the embodiments of the present invention,

[0142] An electronic device is provided, comprising:

[0143] processor;

[0144] a memory for storing processor-executable instructions;

[0145] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0146] According to a fourth aspect of the embodiments of the present invention,

[0147] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0148] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A concrete production management method based on digital twin, characterized in that: include: The system collects health monitoring data of concrete production equipment in real time through various sensors, and uses IoT identification methods to collect event data throughout the entire life cycle of concrete production equipment. The collected health monitoring data and life cycle event data are transmitted to the edge and cloud platforms respectively. At the edge, edge computing technology is used to pre-process the health monitoring data and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, and the annotation of the full life cycle event data is updated according to the diagnosis results. At the same time, the annotated full life cycle event data is used on the cloud platform to build a multi-level digital twin model of the equipment; Based on a multi-level digital twin model, the pre-built physical degradation model and data-driven model are integrated to construct an equipment remaining life prediction model. This model is used to predict the remaining life of the equipment. Based on the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan. This plan is then converted into equipment control instructions to achieve the automatic issuance and execution of the concrete production scheduling plan. Using the annotated data from the entire life cycle of the equipment on the cloud platform, a multi-level digital twin model of the equipment is constructed, including: Receiving annotated equipment lifecycle data in the cloud and obtaining 3D point cloud data therein, preprocessing the 3D point cloud data, converting the preprocessed 3D point cloud data into a triangular mesh surface model, and constructing a geometric model of the concrete production equipment using surface fitting technology; Based on the geometric model of concrete production equipment, the finite element analysis method is used to perform numerical simulation and solve the stress state, vibration characteristics, and heat conduction characteristics of the concrete production equipment. The physical response data of the concrete production equipment under different working conditions is obtained, and the physical model of the equipment is constructed. Obtain the condition monitoring parameters and performance parameters of concrete production equipment, use a gated recurrent unit neural network to establish a nonlinear mapping relationship between the condition monitoring parameters and performance parameters, and build an equipment degradation model for life prediction; The equipment geometric model, the equipment physical model, and the equipment degradation model are integrated to construct a multi-level digital twin model of the concrete production equipment; Based on the multi-level digital twin model, the pre-built equipment degradation model and data-driven model are integrated to build an equipment remaining life prediction model. The remaining life of the equipment is predicted using the equipment remaining life prediction model, including: Based on the multi-level digital twin model, a data-driven model is constructed. The data-driven model uses a convolutional neural network as its basic structure and uses a multi-scale feature extraction mechanism to adaptively extract multi-scale degradation features related to the equipment degradation process from health monitoring data. Based on the obtained multi-scale degradation features, a joint loss function combining fault diagnosis loss and remaining life prediction loss is constructed. The pre-built equipment degradation model is used as a regularization constraint term to train the data-driven model. The degradation trajectory output by the equipment degradation model is used as a priori constraint of the data-driven model to obtain prediction results that conform to the degradation trajectory. The prediction results are then used as feedback to optimize the equipment degradation model parameters. The iteration is repeated until the preset termination condition is met, and finally a trained equipment remaining life prediction model is obtained. Based on the trained equipment remaining life prediction model, the fault diagnosis results are used as prior knowledge, and the Bayesian inference method is used to predict the remaining life of the component by combining physical priors and data-driven posteriori.

2. The method according to claim 1, characterized in that At the edge, edge computing technology is used to pre-process the health monitoring data and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, including: At the edge, we receive health monitoring data from concrete production equipment collected in real time by multiple sensors, extract multi-domain time-frequency features that represent the degradation status of the equipment, and use an attention mechanism to filter them to obtain a subset of key features. Based on the obtained key feature subsets, a multi-view learning strategy is introduced to construct multi-physics field feature subspaces, multi-sensor location feature subspaces, and multi-degradation feature subspaces for different degradation modes from the perspectives of physical field, sensor location, and degradation mode. In each subspace, a multi-view attention mechanism is used to screen the corresponding local key features. Different local key features are combined through an adaptive view weighting strategy to obtain a multi-view diagnostic feature set. Based on the multi-view diagnostic feature set, a one-dimensional convolutional neural network is used at the edge to build an edge diagnostic model, which is used to determine whether the device is faulty. If a fault occurs, a preliminary diagnosis of the fault mode is performed, and based on the preliminary diagnosis results, unknown complex faults with low confidence and the corresponding multi-view diagnostic features are uploaded to the cloud; The cloud uses pre-trained fault diagnosis knowledge graphs to comprehensively consider equipment models, operating conditions, and degradation mechanisms to diagnose unknown complex faults. At the same time, the diagnosis results are used to optimize the fault diagnosis knowledge graphs, and the optimized fault diagnosis knowledge graphs are regularly embedded into the edge diagnosis model to guide edge diagnostic reasoning.

3. The method according to claim 1, characterized in that Based on the obtained multi-scale degradation characteristics, the calculation formula of the joint loss function combining fault diagnosis loss and remaining life prediction loss is constructed as follows: ； Among them, L represents the joint loss function, α represents the weight coefficient of the fault diagnosis model loss, N represents the number of devices, M represents the number of fault categories, and y ij Indicates the true label of whether the i-th device belongs to the j-th type of fault, represents the predicted value of the mean of the characteristic vector of the i-th device under the j-th type of fault, μ j represents the true value of the mean of the characteristic vector of the jth type of fault, β represents the weight coefficient of the remaining life prediction model loss, W p (·) represents the difference between the remaining life distributions, represents the predicted remaining life distribution of the i-th device, P ri represents the true remaining life distribution, γ represents the weight coefficient of the regularization term, W represents the parameter matrix of the model, and ||·||2 represents the L2 norm.

4. The method according to claim 1, wherein Based on the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan. This plan is then converted into equipment control instructions to automatically issue and execute the concrete production scheduling plan. This includes: A multi-objective reinforcement learning optimization framework is constructed to model the concrete production process as a multi-objective Markov decision process. The multi-objective Markov decision process includes a state space, an action space, a state transition probability matrix, a multi-objective reward function, a discount factor, and an initial state distribution. The state space includes equipment states and process parameters, the action space includes equipment operation modes and production scheduling strategies, and the multi-objective reward function includes equipment remaining life, production efficiency, and energy consumption indicators. Based on the constructed multi-objective Markov decision process, a hierarchical target learning algorithm is designed. A value function approximator and a policy function approximator are constructed for each optimization objective of the multi-objective reward function. The value function and policy function of each optimization objective are estimated using the value function approximator and the policy function approximator. A target improvement operator is obtained based on the obtained value function estimation value and policy function estimation value. The multi-objective reward function is updated based on the target improvement operator. Based on the updated multi-objective reward function, a non-dominated sorting genetic algorithm is used to search in the multi-objective Markov decision process to obtain an initial policy set. Based on the obtained initial strategy set, in the digital twin model, by building an interactive interface between the intelligent agent and the virtual environment, the Monte Carlo tree search algorithm is used to select the optimal exploration path through an adaptive exploration-exploitation equilibrium strategy, generate large-scale optimization strategy sample data, and obtain the optimal strategy set; The optimal strategy set is input into the interactive decision recommendation system of the target learning algorithm, and the feasibility of the strategy is evaluated by combining expert knowledge to generate the final optimal concrete production scheduling plan; The optimal concrete production scheduling plan is converted into equipment control instructions, and the automatic issuance and execution of the concrete production scheduling plan is achieved through the interactive interface between the digital twin model and the actual production management system.

5. The method according to claim 4, characterized in that A value function approximator and a policy function approximator are constructed for each optimization objective of the multi-objective reward function. The value function and policy function of each optimization objective are estimated using the value function approximator and the policy function approximator. The target improvement operator is obtained based on the obtained value function estimation value and policy function estimation value. The calculation formula for updating the multi-objective reward function based on the target improvement operator is as follows: ； in, represents the updated multi-objective reward function, R k represents the initial multi-objective reward function, λ represents the learning rate, A represents the action space, δ represents the discount factor, S represents the state space, P(·) represents the state transition probability matrix, represents the value function of the kth optimization objective estimated by the value function approximator at state s', represents the value function of the kth optimization objective estimated by the value function approximator at state s, ω represents the adjustment factor that balances the reward function and the policy function, represents the probability of the kth optimization objective to select action a' in state s, estimated by the policy function approximator. represents the probability of selecting action a in state s as estimated by the policy function approximator for the kth optimization objective.

6. A concrete production management system based on digital twin, used to implement the method according to any one of claims 1 to 5, characterized in that: include: The first unit is used to collect health monitoring data of concrete production equipment in real time through multiple sensors, and uses IoT identification methods to collect full life cycle event data of concrete production equipment, and transmit the collected health monitoring data and full life cycle event data to the edge and cloud platforms respectively; The second unit is used to pre-process the health monitoring data using edge computing technology at the edge end, and extract the status characteristics of the equipment. Based on the status characteristics of the equipment, a pre-trained fault prediction model is used to perform real-time diagnosis, and the annotation of the full life cycle event data is updated according to the diagnosis results. At the same time, the annotated full life cycle event data is used on the cloud platform to build a multi-level digital twin model of the equipment; The third unit is used to integrate the pre-built physical degradation model and data-driven model based on the multi-level digital twin model, build an equipment remaining life prediction model, and use the equipment remaining life prediction model to predict the remaining life of the equipment. According to the prediction results, the concrete production strategy is dynamically adjusted in the digital twin model to obtain the optimal concrete production scheduling plan, and the plan is converted into equipment control instructions to realize the automatic issuance and execution of the concrete production scheduling plan.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images