Industrial equipment fault prediction method and system based on neural network model

Abstract

The application discloses an industrial equipment fault prediction method and system based on a neural network model, and relates to the technical field of intelligent monitoring and fault diagnosis of industrial equipment; the application obtains and models a physical mechanism atlas and a degradation mechanism model of equipment through physical mechanism knowledge, generates a physically consistent feature set through physical constraint feature extraction, constructs a physical and data fusion network, encodes physical priori into the network structure, designs a multi-objective loss function for training, and performs fault prediction and model verification, so that three-stage fault prediction, physical consistency verification and model online iteration are realized; the application effectively solves the problems of poor physical interpretability of a traditional fault prediction model, weak generalization ability under complex working conditions and prediction results that easily violate physical laws, and significantly improves the precision and timeliness of industrial equipment fault prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault diagnosis technology for industrial equipment, specifically a fault prediction method for industrial equipment based on a neural network model. Background Technology

[0002] With the rapid development of intelligent manufacturing and the industrial internet, industrial equipment is showing a trend of integration, high speed and intelligence. The operating status of equipment is directly related to production safety, operation and maintenance efficiency and economic benefits. Fault prediction, as a core link of equipment health management, can identify potential anomalies and predict fault trends in advance, providing a basis for preventive maintenance. Current mainstream fault prediction methods mostly rely on traditional statistical models, shallow machine learning, or pure data-driven neural networks. Although they can achieve anomaly detection in specific scenarios, they have significant drawbacks: First, the physical interpretability is poor, the models are mostly black box structures, the decision-making process is disconnected from the equipment degradation mechanism, and it is difficult to meet the reliability requirements of industrial scenarios. Second, the generalization ability for complex working conditions is weak. When faced with dynamic working conditions such as variable load, variable speed, and strong interference, the model is prone to overfitting or insufficient feature adaptation, resulting in a sharp drop in the effectiveness of cross-device and cross-scenario deployment. Third, the prediction results are prone to violating the laws of physics. Pure data learning does not incorporate physical constraints and may output results that contradict the operating rules of equipment, energy conservation, and material properties, resulting in low engineering practicality. Existing technologies cannot simultaneously achieve prediction accuracy, generalization and physical consistency, and cannot meet the health management needs of high-end equipment throughout its entire life cycle. Therefore, it is urgent to construct a fault prediction method that combines strong feature learning ability with physical interpretability, break through the bottleneck of traditional models, and improve the accuracy and reliability of fault early warning in complex industrial environments.

[0003] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of poor physical interpretability, weak generalization ability under complex working conditions, and easy violation of physical laws in traditional fault prediction models, and to propose a fault prediction method for industrial equipment based on neural network models.

[0005] The objective of this invention can be achieved through the following technical solutions: Industrial equipment fault prediction methods based on neural network models include: S1. Physical Mechanism Knowledge Acquisition and Modeling: The physical mechanism of the equipment is acquired through the physical mechanism analysis unit, the physical mechanism map of the equipment is constructed, and the causal relationship between physical parameters is established. S2. Physical constraint feature extraction: Based on the physical mechanism map, physical constraint-guided feature extraction is performed on the equipment operation data to generate a physical consistency feature set; S3. Construction of a Physics and Data Fusion Network: Design an embedded neural network architecture based on physical mechanisms, encode prior physical knowledge into the network structure, and construct a feature extraction layer driven by both physics and data. S4. Fault Prediction and Model Validation Execution: The trained physical and data fusion network is deployed to the online monitoring equipment. Through the dual-driven reasoning mechanism of physical and data, the fault prediction, result verification and model iteration update of industrial equipment are completed, realizing a closed-loop fault prediction and health management throughout the entire process.

[0006] Furthermore, the specific operation process of S1 is as follows: Physical mechanism map construction: Extract the physical operation mechanism of the equipment from industrial equipment technical documents, design manuals and operation and maintenance records, decompose the equipment into multiple core functional units and correspond to a physical node, and construct a physical mechanism map containing a set of physical nodes, a set of relationships between nodes, and a relationship strength matrix; Construction of causal chain of physical parameters: Identify the input and output physical quantities of each physical node, and derive the physical transfer function based on the input and output physical quantities; Physical constraint boundary: Defines the physical feasible region for each physical quantity, including minimum and maximum physical limits and physical law constraints; Degradation mechanism model construction: A degradation mechanism model based on physical wear accumulation is established for vulnerable components, and the cumulative damage degree of the component at any time is calculated by integration; The physical mechanism map definition includes physical quantity identifiers, sets of operational operators, and physical feature extraction templates based on characteristic time scales.

[0007] Furthermore, the specific operation steps of S2 include: A multi-physics field monitoring network is formed by deploying multiple types of sensors. Raw signals are collected at a preset sampling frequency. The sensor deployment locations match the physical node distribution of the physical mechanism map. The raw signals are synchronously stored to the data acquisition terminal. The acquired raw signals are mapped to physical quantities defined in the physical mechanism map. For directly measurable physical quantities, they are associated with the corresponding target physical quantities in the physical mechanism map after noise reduction preprocessing. For physical quantities that cannot be directly measured, they are obtained by inversion based on the physical transfer function and the causal relationship of physical nodes. Based on physical feature extraction templates, features that are adapted to the physical operation mechanism of equipment are extracted for different types of physical quantities, such as thermodynamic, mechanical, and electrical ones.

[0008] Furthermore, the specific operation steps of S2 also include: Perform a physical feasibility test on all extracted features and calculate a physical consistency index. When the physical consistency index is less than a preset threshold, it is determined that there is a physical violation, and the data marking process is started to mark the corresponding original signal, the mapped physical quantity and the extracted features. Based on the multi-scale temporal evolution characteristics of the physical process of the device, a multi-scale feature pyramid is constructed; a set of feature scales is defined and local physical features are extracted at each scale through a sliding window to form a multi-scale feature tensor. By combining the correlation strength matrix of the physical mechanism map and the degradation mechanism model, the physical importance weights of each physical quantity are obtained. All features verified by physical constraints, extracted at multiple scales and quantified by importance are integrated to generate a physical consistency feature set.

[0009] Furthermore, the specific operation steps of S3 include: A physical embedding layer is inserted after the input layer of the neural network. The initial weight matrix of the physical embedding layer is initialized based on the correlation strength matrix of the physical mechanism map. The physical correlation information is encoded into the feature representation through the forward propagation of the physical embedding layer. Physically constrained residual blocks are constructed to maintain physical conservation relationships. Physically mechanism-guided self-attention units are constructed and incorporated into a physical causal relationship mask matrix to obtain the attention matrix. A dual-branch network structure for feature learning based on physical mechanisms and data is constructed. The physical mechanism-driven branch is built based on physical mechanism maps and degradation mechanism models, and uses a differentiable physical simulation model to learn features. The data-driven branch uses a neural network framework combining convolutional neural networks and Transformers to capture nonlinear correlations. The output features of the physical mechanism-driven branch and the data-driven branch are integrated through a dynamic weighting mechanism of adaptive learning of gating network; a multi-objective loss function including prediction, physical consistency, monotonicity and causal consistency loss is constructed, a learnable mechanism for physical parameters is set and physical boundary constraints are applied; the constructed physical and data fusion network is initialized and configured, and the physical consistency feature set is divided into training, validation and test sets to complete the construction of the feature extraction layer driven by both physical and data for network training.

[0010] Furthermore, the specific operation steps of S4 are as follows: Real-time data collection of industrial equipment operation is performed, real-time features are generated through a physical constraint feature extraction process, and corresponding physical consistency indices are calculated. When the physical consistency index is lower than the preset threshold, the sensor fault detection process is triggered, and the missing physical quantity is reconstructed based on the physical transfer function and the causal relationship of the physical nodes. Real-time features are input into the trained physical and data fusion network to perform short-term, medium-term, and long-term fault prediction. The contribution of each physical feature to the fault category is obtained based on the ensemble gradient algorithm. Based on the physical mechanism map, the contribution is mapped to the physical cause, forming a reasoning chain in the order of feature anomaly, physical quantity anomaly, physical mechanism deviation and fault type, and generating a fault diagnosis report. Physical consistency verification is performed on the prediction results. The network output features are substituted into the physical transfer function and the physical feasible region for verification. If the verification fails, the anomaly correction mechanism is activated to increase the weight of physical constraints in the inference process. Feature fusion and state prediction are re-executed until the output results meet the physical rationality requirements. After the verification passes, the prediction results are visualized. At the same time, an online model iteration mechanism is established to update the network parameters.

[0011] A second aspect of the present invention provides an industrial equipment fault prediction system based on a neural network model, comprising: Physical mechanism modeling module: Extract physical knowledge from equipment documents and operation and maintenance data, construct equipment physical mechanism map, physical parameter causal chain and degradation mechanism model, and form a physical prior knowledge base; Physical constraint feature module: Based on the multi-physics monitoring network, it collects operational data, completes physical quantity mapping and multi-scale feature extraction based on the physical mechanism spectrum, performs physical consistency verification on the features, and generates a consistent feature set; Fusion network building module: Construct a physical embedding layer, a physical constraint residual block and a physical guided self-attention unit to form a dual-branch fusion network structure. Use a multi-objective loss function for model training to achieve deep feature learning driven by both physics and data. Fault prediction execution module: Based on real-time data, it generates real-time features and performs three-stage fault prediction. At the same time, it performs physical consistency verification and results visualization, establishes an online iterative mechanism to update network parameters, and realizes closed-loop fault prediction and health management.

[0012] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves accurate mining and quantification of physical prior knowledge by constructing a physical mechanism map and degradation mechanism model of equipment; it completes multi-scale feature extraction guided by physical constraints to generate a physical consistency feature set that conforms to the operating rules of the equipment; it designs a physical embedding layer and constraint residual blocks to encode the physical mechanism into a neural network architecture, builds a physical and data dual-driven fusion network, and completes network training by combining a multi-objective loss function; it realizes short, medium and long-term three-stage fault prediction, ensures the rationality of prediction through physical consistency verification and anomaly correction mechanisms, and establishes an online model iteration mechanism to achieve adaptive parameter updates; it significantly improves the accuracy and physical interpretability of industrial equipment fault prediction, enhances the model's generalization ability and physical stability under complex working conditions, enhances the timeliness and reliability of fault prediction, realizes a closed loop of equipment fault prediction and health management, and effectively adapts to the actual needs of equipment condition migration and long-term operation. Attached Figure Description

[0013] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example: like Figure 1 As shown, the industrial equipment fault prediction method based on neural network model includes physical mechanism knowledge acquisition and modeling, physical constraint feature extraction, physical and data fusion network construction, and fault prediction and model verification execution.

[0016] S1. Physical Mechanism Knowledge Acquisition and Modeling: The physical mechanism of the equipment is acquired through the physical mechanism analysis unit, a physical mechanism map of the equipment is constructed, and causal relationships between physical parameters are established. The specific implementation process is as follows: Construction of physical mechanism map: The physical mechanism analysis unit extracts the physical operating mechanism from industrial equipment technical documents, design manuals and operation and maintenance records, including thermodynamic relationships, mechanical wear laws, electrical characteristic curves and fluid dynamic characteristics. Industrial equipment is broken down into core functional units, including power drive units, transmission execution units, load-bearing units, cooling and lubrication units, monitoring and sensing units, and auxiliary support units; each functional unit corresponds to a physical node, constructing a physical mechanism diagram of the equipment:

[0017] in, , representing the set of physical nodes Indicates the number of physical nodes; This represents the set of physical relationships between physical nodes. Indicates the first The physical node and the first The physical relationships that exist between physical nodes; , representing the correlation strength matrix, Indicates the first The physical node pairs the The physical influence intensity of each physical node; Construction of causal chains for physical parameters: For each physical node, identify the node's input and output physical quantities using formulas. The derivation yields the first Physical transfer function corresponding to each physical node ,in, Indicates the first The output physical quantity of each physical node. Indicates the first Input physical quantities of each physical node Indicates the first The physical parameter vector corresponding to each physical node; Physical constraint boundaries: For each physical quantity, define the physical feasible region of the physical quantity using the formula. Get the first The physical feasible region corresponding to each physical node, where... Indicates the first The minimum physical limit of the physical quantity corresponding to each physical node. Indicates the first The maximum physical limit of the physical quantity corresponding to each physical node. Indicates the first The set of physical law constraints that each physical node must satisfy for its corresponding physical quantity; Construction of degradation mechanism model: For vulnerable components of equipment, a degradation mechanism model based on physical wear accumulation is established; through formulas... Calculations are performed to obtain the component at time [time]. The cumulative damage, of which, This indicates the initial degree of damage, referring to the inherent damage that already exists in the equipment at the time of manufacture, assembly, or the start of operation. This represents the damage evolution function of multiphysics coupling. Indicates time The stress time history under the following conditions Indicates time The temperature over time, Indicates time The strain rate time history is as follows. This represents the time integral from time 0 to the current time t. Represents an integral infinitesimal element;

[0018] Based on the physical mechanism map, a physical feature extraction template is defined. The template includes a physical quantity identifier, a set of operation operators for the physical quantity, and the characteristic time scale of the physical quantity. The set of operation operators includes time-domain differentiation, frequency-domain transformation, and phase space reconstruction.

[0019] S2. Physical Constraint Feature Extraction: Based on the physical mechanism map, physical constraint-guided feature extraction is performed on the equipment operation data to generate a physical consistency feature set; the specific implementation process is as follows: Vibration sensors, temperature sensors, pressure sensors, and current sensors are deployed to form a multi-physics field monitoring network, and raw signals are collected at a preset sampling frequency. The sensor deployment locations are planned in conjunction with the distribution of physical nodes in the physical mechanism map. Each physical node corresponding to each core functional unit is covered by at least one corresponding sensor, thereby realizing comprehensive monitoring of the equipment's multi-physics field parameters. The raw signals obtained are synchronously stored in the data acquisition terminal. The raw signals collected by the sensors are mapped to physical quantities defined in the physical mechanism map. For directly measurable physical quantities, signal noise reduction preprocessing is performed and then associated with the corresponding target physical quantity in the physical mechanism map. For physical quantities that cannot be directly measured, they are obtained by inversion based on the physical transfer function and the causal relationship of physical nodes. For example, local stress is obtained by inversion from vibration acceleration signals and equipment geometric parameters. The inversion process follows the causal relationship of physical nodes in the physical mechanism map. Based on the physical feature extraction template, features that are adapted to the physical operation mechanism of the equipment are extracted for each physical quantity: For thermodynamic physical quantities, through formulas The characteristics of the heat flux density gradient are obtained. Indicates the first The characteristics of the heat flux density gradient corresponding to each thermodynamic physical quantity Indicates the spatial location of the contents of the equipment components. ,time The heat flux density gradient below, This indicates the length of the time interval selected for this heat flux density gradient characteristic calculation; For mechanical vibration, nonlinear dynamic characteristics are extracted: based on the phase space reconstruction method, through formulas... The maximum Lyapunov exponent is obtained through calculation, where, Indicates the time at which Limit operations as the value approaches infinity Represents the natural logarithm operation. In phase space, time represents... The magnitude of the deviation vector between adjacent trajectories. It represents the magnitude of the initial deviation vector of neighboring trajectories in the phase space at the initial time. At the same time, through the formula Calculate the number of related dimensions ,in, The correlation integral indicates that the distance between two points in phase space is less than the neighborhood radius. The proportion of point pairs to the total number of point pairs; For electrical physical quantities, extract power spectral entropy features: through the formula The power spectral entropy is calculated, where, The total number of frequency components after frequency domain decomposition is determined by both the signal sampling rate and the frequency domain analysis resolution. Indices representing frequency domain components. , indicating the first Normalized power spectral density corresponding to each frequency component Represents frequency components The original power spectral density at that location, This represents the sum of the power spectral densities of all frequency components, i.e., the total energy of the electrical signal in the frequency domain. Indicates the frequency component traversal index; Power spectral entropy characteristics are used to characterize the frequency distribution uniformity of electrical signals and reflect abnormal operation of motors and electrical circuits; Perform a physical feasibility test on the extracted features using the formula. The physical consistency index is calculated, where, Indicates the first The feature vector extracted from the input physical quantities corresponding to each physical node. Indicates the first The feature vector extracted from the actual output physical quantities corresponding to each physical node. This indicates an indicator function that returns 1 if the relative error between the physical transfer function output value and the actual characteristic value is less than a preset threshold, and 0 otherwise. This indicates that for all physical nodes from 1 to Perform a series of multiplication operations; When the physical consistency index is less than the preset threshold, it is determined that there is a physical violation, and the data marking process is initiated to mark the corresponding original signal, the mapped physical quantity and the extracted features. Based on the multi-scale time evolution characteristics of equipment physical processes, a multi-scale feature pyramid is constructed to adapt to the dynamic change rate of different physical quantities. Among them, the field quantity changes, including temperature and pressure, are usually slow-changing processes, while vibration and shock, and transient stress fluctuations are fast-changing processes. The feature pyramid is used to capture and comprehensively characterize the equipment operating status at different time scales in multiple dimensions. Define the feature scale set Among them, each feature scale satisfies , The feature pyramid represents the first The time scale value corresponding to the layer, Indicates the basic time scale. This represents the scale factor, which is a constant greater than 1. This indicates the total number of layers in the feature pyramid; At each feature scale, local physical features are extracted through a pre-defined sliding window to form a multi-scale feature tensor. Based on the correlation strength matrix and degradation mechanism model in the physical mechanism map, through the formula The physical importance weights are calculated, where, Representing the physical mechanism diagram and physical nodes The set of associated nodes, This represents the partial derivative of the cumulative damage of the equipment with respect to the physical quantity.

[0020] Integrate all features that have undergone physical constraint verification, multi-scale extraction, and importance quantification to generate a physically consistent feature set.

[0021] S3. Construction of a Physics-Data Fusion Network: An embedded neural network architecture based on physical mechanisms is designed, encoding prior physical knowledge into the network structure, and constructing a feature extraction layer driven by both physics and data. The specific implementation process is as follows: A physical embedding layer is inserted after the input layer of the neural network. The transformation matrix of the physical embedding layer is not randomly initialized, but rather initialized based on the correlation strength matrix of the physical mechanism map, using the formula... The initial weight matrix of the physical embedding layer is obtained, where, This represents the physical prior strength coefficient, used to adjust the degree of influence of physical prior knowledge on network initialization. This represents a random weight matrix that follows a Xavier initialization distribution; The forward propagation formula for the physical embedding layer is as follows:

[0022] in, This represents the output feature vector of the physical embedding layer. This represents the input feature vector of the physical embedding layer. This represents the bias vector of the physical embedding layer. Represents a non-linear activation function; Physical association information is encoded into the feature representation based on the physical embedding layer; Construct physically constrained residual blocks. The structure of the physically constrained residual blocks is forced to maintain physical conservation relationships, so as to avoid feature mappings that violate physical laws during the network learning process. To address the physical constraint of energy conservation, an energy conservation residual block is designed. The output feature of the energy conservation residual block is formed by superimposing the input features and the output result of the residual function. The input features are the preceding feature information received by the residual block, and the residual function is the feature transformation function built into the energy conservation residual block. The output result of the residual function follows the law of energy conservation, specifically, the sum of the feature values ​​corresponding to all energy input channels is equal to the sum of the feature values ​​corresponding to all energy output channels and the sum of the dissipated energy generated during the operation of the equipment. Constructing self-attention units guided by physical mechanisms, through formulas We obtain the attention matrix, where, Represents the normalized activation function. The query matrix and key matrix are derived from the physical consistency feature set based on a linear transformation. Represents the transpose of a matrix. The square root of the dimension of the key matrix is ​​the classic scaling factor for attention mechanisms. This represents the physical guidance hyperparameters. The physical causal relationship mask matrix is ​​extracted from the physical parameter causal chain constructed by S1; A dual-branch network structure is constructed to form a dual-drive fusion mode of feature learning driven by physical mechanisms and data. The physical mechanism-driven branch is built based on the physical mechanism map and degradation mechanism model constructed by S1, and uses a differentiable physical simulation model to realize feature learning. The input of the physical mechanism-driven branch is the physical consistency feature set generated by S2, and the output is a feature representation that conforms to the physical operation law of the equipment. The data-driven branch adopts a neural network architecture combining convolutional neural network and Transformer. The input is also the physical consistency feature set generated by S2. It captures the complex nonlinear correlation between features through a multi-layer network structure. The output features of the physical mechanism-driven branch and the data-driven branch are fused through a dynamic weighting mechanism. The fusion weights are obtained through adaptive learning by a gating network, and the specific calculation formula is as follows:

[0023] in, Indicates time Dynamic fusion weights under the following conditions This represents the weight matrix of the gated network. These represent the output feature vectors of the physical mechanism-driven branch and the data-driven branch, respectively. This represents the bias term of the gated network; Construct a multi-objective loss function, the specific form of which is as follows:

[0024] in, This represents the total loss value. The prediction loss is represented by the average of the squared deviations between the network predictions and actual values ​​of all collected samples. The physical consistency loss is calculated by summing all physical constraint terms. Each constraint term is the squared value of the deviation between the output result corresponding to the input feature of the physical transfer function and the actual output feature of the network. The monotonicity loss is calculated by summing all physical features that satisfy the monotonicity requirement. Each feature's corresponding calculation term is processed by a non-linear activation function. The processed item is the negative product of the partial derivative of the physical feature with respect to the network's prediction output and the feature's monotonic direction indicator value. The feature's monotonic direction indicator value takes the value of... ; The loss of causal consistency is calculated by summing all causal pairs in the physical causal chain. These represent the weighting factors for the impact of physical consistency loss, monotonicity loss, and causal consistency loss, respectively. A learnable mechanism for physical parameters is constructed by setting the physical parameters in the physical transfer function as trainable parameters and imposing physical boundary constraints on these trainable parameters; this is achieved through parameterization methods. It means that, among them, This refers to the nominal values ​​of physical parameters obtained from the equipment's technical documentation. Indicates the allowable range of relative deviation. Represents the unconstrained variables of online learning. Represents the hyperbolic tangent activation function; The constructed physical and data fusion network is initialized and configured, and hyperparameters including the number of iterations and learning rate are set for network training. The learning rate adopts an adaptive decay strategy, which gradually decreases as the number of training iterations increases. The physical consistency feature set generated by S2 is divided into training set, validation set and test set for network training, parameter optimization and performance verification, and finally the construction of the feature extraction layer driven by both physics and data is completed.

[0025] S4. Fault Prediction and Model Validation: The trained physical and data fusion network is deployed to the online equipment monitoring device. Through a dual-driven inference mechanism of physical and data, fault prediction, result verification, and model iterative updates for industrial equipment are completed, achieving a closed-loop fault prediction and health management process. The specific implementation process is as follows: Real-time acquisition of equipment operation data; based on the S3 physical constraint feature extraction process, real-time features are generated and corresponding physical consistency indicators are calculated. When the physical consistency index of real-time features is lower than a preset threshold, the sensor fault detection process is triggered, and the missing physical quantity is reconstructed by virtual sensing based on the physical transfer function. The reconstruction process follows the physical node causal relationship of S1. Real-time features are input into the trained physical and data fusion network to perform short-term, medium-term, and long-term fault predictions. Short-term predictions cover the next 1 to 24 hours, with the physical and data fusion network directly outputting the device status. Medium-term predictions cover the next 1 to 7 days, with the physical and data fusion network combining temporal features and cumulative damage features to extrapolate the status. Long-term predictions cover the next 1 to 3 months, obtained by extrapolating from the physical degradation trajectory. Based on the degradation rate and degradation acceleration parameters of the device's cumulative damage, and constrained by the degradation mechanism model constructed by S1, the remaining service life on a long-term scale is estimated. The physical and data fusion network uses the softmax activation function to obtain the probability distribution of each fault category; for each fault category, the contribution of each physical feature to the fault category is obtained based on the ensemble gradient algorithm; based on the physical mechanism map, the contribution is mapped to the physical cause, forming an inference chain in the order of feature anomaly, physical quantity anomaly, physical mechanism deviation, and fault type, and a fault diagnosis report is generated. Physical consistency checks are performed on the predicted failure probability, failure time, and remaining service life. The network output features are then substituted back into the physical transfer function and the physical feasible region for verification, including determination of the physical feasible region range, verification of energy conservation relationships, and verification of causal time series relationships. For example, it verifies whether the predicted physical quantities fall within the physical feasible region, checks whether the energy conservation relationship is satisfied, and ensures that the causal time series relationships between physical quantities are reasonable. If a prediction result exceeds the constraint boundary or violates the physical mechanism, an anomaly correction mechanism is activated to increase the weight of physical constraints in the reasoning process, and feature fusion and state prediction are re-executed until the output result meets the physical rationality requirements. The verified prediction results are visualized, including the equipment health curve, the confidence level of the fault type, the fault warning level, and the remaining service life range. At the same time, an online model iteration mechanism is established to compare the actual fault occurrence time, operation and maintenance records, and repair results with the model's historical prediction results to form a prediction error dataset; Using the prediction error dataset as incremental samples, without changing the backbone architecture of the network, the gating fusion weights, physical guidance hyperparameters and learnable physical parameters of the physical and data fusion network are adjusted and updated so that the model can adapt to the performance degradation, working condition migration and environmental changes during the long-term operation of the equipment. By continuously introducing newly collected monitoring data and physical constraint verification rules, the prediction accuracy, generalization ability and physical stability of the model under complex working conditions are continuously improved, realizing a closed loop of the entire process from physical modeling, feature extraction, network construction to fault prediction, result verification and model iteration.

[0026] An industrial equipment fault prediction system based on a neural network model includes: Physical mechanism modeling module: Extract physical mechanism knowledge from equipment technical documents and operation and maintenance records, construct equipment physical mechanism map, build physical parameter causal chain and physical constraint boundary, establish wear accumulation degradation mechanism model for vulnerable parts, and define physical feature extraction template to provide physical prior support for subsequent processes; Physical constraint feature module: Deploy multiple types of sensors to form a monitoring network to collect raw signals, map the raw signals to physical quantities corresponding to the physical mechanism map, extract adaptation features for different physical quantities, and after physical feasibility verification, multi-scale feature construction, and physical importance weight calculation, integrate to generate a physical consistency feature set; Fusion network construction module: embed physical layers into the neural network and initialize weights based on the physical mechanism map, construct physical constraint residual blocks and physical guided self-attention units, build a physical mechanism and data-driven dual-branch network, fuse dual-branch features through a gating network, design a multi-objective loss function to train the network, and complete the construction of the dual-drive feature extraction layer; Fault prediction execution module: Real-time data collection generates real-time features and calculates physical consistency indicators. The data is then input into the fusion network to achieve short, medium and long-term three-stage fault prediction, generate a fault diagnosis report, and perform physical consistency verification on the prediction results. After the verification is passed, the prediction results are visualized. At the same time, an online model iteration mechanism is established to update network parameters based on actual operation and maintenance data, realizing closed-loop management of the entire process.

[0027] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting industrial equipment faults based on a neural network model, characterized in that, include: S1. Physical Mechanism Knowledge Acquisition and Modeling: The physical mechanism of the equipment is acquired through the physical mechanism analysis unit, the physical mechanism map of the equipment is constructed, and the causal relationship between physical parameters is established. The specific operation process of S1 is as follows: Physical mechanism map construction: Extract the physical operation mechanism of the equipment from the technical documents, design manuals and operation and maintenance records of industrial equipment, decompose the equipment into multiple core functional units and each corresponding to a physical node, and construct a physical mechanism map containing a set of physical nodes, a set of relationships between nodes, and a relationship strength matrix; Construction of causal chain of physical parameters: Identify the input and output physical quantities of each physical node, and derive the physical transfer function based on the input and output physical quantities; Physical constraint boundary: Defines the physical feasible region for each physical quantity, including minimum and maximum physical limits and physical law constraints; Degradation mechanism model construction: A degradation mechanism model based on physical wear accumulation is established for vulnerable components, and the cumulative damage degree of the component at any time is calculated by integration; S2. Physical constraint feature extraction: Based on the physical mechanism map, after extracting features guided by physical constraints from the equipment operation data, the extracted features are subjected to physical feasibility verification to generate a physical consistency feature set. S3. Construction of a Physics and Data Fusion Network: Design an embedded neural network architecture based on physical mechanisms, encode prior physical knowledge into the network structure, and construct a feature extraction layer driven by both physics and data. The specific operation steps of S3 include: A physical embedding layer is inserted after the input layer of the neural network. The initial weight matrix of the physical embedding layer is initialized based on the correlation strength matrix of the physical mechanism map. The physical correlation information is encoded into the feature representation through the forward propagation of the physical embedding layer. Physically constrained residual blocks are constructed to maintain physical conservation relationships. Physically mechanism-guided self-attention units are constructed and incorporated into a physical causal relationship mask matrix to obtain the attention matrix. A dual-branch network structure for feature learning driven by physical mechanisms and data is constructed. The physical mechanism-driven branch is built based on the physical mechanism map and degradation mechanism model, and features are learned using a differentiable physical simulation model. The data-driven branch uses a neural network framework that combines convolutional neural networks and Transformers to capture nonlinear correlations; The output features of the physical mechanism-driven branch and the data-driven branch are fused through a dynamic weighting mechanism of adaptive learning using a gated network; a multi-objective loss function is constructed, a learnable mechanism for physical parameters is set, and physical boundary constraints are applied; the constructed physical and data fusion network is initialized and configured, and the physical consistency feature set is divided into training, validation, and test sets to complete the network training; S4. Fault Prediction and Model Validation Execution: The trained physical and data fusion network is deployed to the online monitoring equipment. Through the dual-driven reasoning mechanism of physical and data, the fault prediction, result verification and model iteration update of industrial equipment are completed, realizing a closed-loop fault prediction and health management throughout the entire process.

2. The industrial equipment fault prediction method based on a neural network model according to claim 1, characterized in that, The specific operation steps of S2 include: A multi-physics field monitoring network is formed by deploying multiple types of sensors. Raw signals are collected at a preset sampling frequency. The sensor deployment locations match the physical node distribution of the physical mechanism map. The raw signals are synchronously stored to the data acquisition terminal. The acquired raw signals are mapped to physical quantities defined in the physical mechanism map. For directly measurable physical quantities, they are associated with the corresponding target physical quantities in the physical mechanism map after noise reduction preprocessing. For physical quantities that cannot be directly measured, they are obtained by inversion based on the physical transfer function and the causal relationship of physical nodes. The definition of the physical mechanism map includes physical quantity identifiers, a set of computational operators for calculating physical quantities, and physical feature extraction templates for the selected characteristic time scales of the calculated physical quantities. Based on physical feature extraction templates, features that are adapted to the physical operation mechanism of equipment are extracted for different types of physical quantities, such as thermodynamic, mechanical, and electrical ones.

3. The industrial equipment fault prediction method based on a neural network model according to claim 2, characterized in that, The specific operation steps of S2 also include: Perform a physical feasibility test on all extracted features and calculate a physical consistency index. When the physical consistency index is less than a preset threshold, it is determined that there is a physical violation, and the data marking process is started to mark the corresponding original signal, the mapped physical quantity and the extracted features. Based on the multi-scale temporal evolution characteristics of the physical process of the device, a multi-scale feature pyramid is constructed; a set of feature scales is defined and local physical features are extracted at each scale through a sliding window to form a multi-scale feature tensor. By combining the correlation strength matrix of the physical mechanism map and the degradation mechanism model, the physical importance weights of each physical quantity are obtained. All features verified by physical constraints, extracted at multiple scales and quantified by importance are integrated to generate a physical consistency feature set.

4. The industrial equipment fault prediction method based on a neural network model according to claim 1, characterized in that, The specific operation steps of S4 are as follows: Real-time data collection of industrial equipment operation is performed, real-time features are generated through a physical constraint feature extraction process, and corresponding physical consistency indices are calculated. When the physical consistency index is lower than the preset threshold, the sensor fault detection process is triggered, and the missing physical quantity is reconstructed based on the physical transfer function and the causal relationship of the physical nodes. Real-time features are input into the trained physical and data fusion network to perform short-term, medium-term, and long-term fault prediction. The contribution of each physical feature to the fault category is obtained based on the ensemble gradient algorithm. Based on the physical mechanism map, the contribution is mapped to the physical cause, forming a reasoning chain in the order of feature anomaly, physical quantity anomaly, physical mechanism deviation and fault type, and generating a fault diagnosis report. Physical consistency verification is performed on the prediction results. The network output features are substituted into the physical transfer function and the physical feasible region for verification. If the verification fails, the anomaly correction mechanism is activated to increase the weight of physical constraints in the inference process. Feature fusion and state prediction are re-executed until the output results meet the physical rationality requirements. After the verification passes, the prediction results are visualized. At the same time, an online model iteration mechanism is established to update the network parameters.

5. The industrial equipment fault prediction method based on a neural network model according to claim 1, characterized in that, The specific construction process of the multi-objective loss function is as follows: Construct a multi-objective loss function that includes prediction, physical consistency, monotonicity, and causal consistency losses; The prediction loss is obtained by averaging the squared deviations between the network prediction values ​​and the actual values ​​of all collected samples. The physical consistency loss is calculated by summing all physical constraint terms, where each constraint term is the square of the deviation between the output result corresponding to the input feature of the physical transfer function and the actual output feature of the network. The monotonicity loss is calculated by summing all physical features that satisfy the monotonicity requirement. The calculation term for each feature is processed by a nonlinear activation function. The processed item is the negative product of the partial derivative of the physical feature with respect to the network prediction output and the feature's monotonic direction indicator value. The feature's monotonic direction indicator value is set to... ; The causal consistency loss is calculated by summing all causal pairs in the physical causal chain.

6. A system applied to the industrial equipment fault prediction method based on a neural network model as described in any one of claims 1-5, comprising: Physical mechanism modeling module: Extract physical knowledge from equipment documents and operation and maintenance data, construct equipment physical mechanism map, physical parameter causal chain and degradation mechanism model, and form a physical prior knowledge base; Physical constraint feature module: Based on the multi-physics monitoring network, it collects operational data, completes physical quantity mapping and multi-scale feature extraction based on the physical mechanism spectrum, performs physical consistency verification on the features, and generates a consistent feature set; Fusion network building module: Construct a physical embedding layer, a physical constraint residual block and a physical guided self-attention unit to form a dual-branch fusion network structure. Use a multi-objective loss function for model training to achieve deep feature learning driven by both physics and data. Fault prediction execution module: Based on real-time data, it generates real-time features and performs three-stage fault prediction. At the same time, it performs physical consistency verification and results visualization, establishes an online iterative mechanism to update network parameters, and realizes closed-loop fault prediction and health management.

Images