Temperature sensor fault diagnosis method and system

Abstract

The present invention relates to the field of fault diagnosis technology, and discloses a temperature sensor fault diagnosis method and system. The temperature sensor fault diagnosis method includes the following steps: processing multi-source sensor data of an industrial furnace to obtain a standardized multi-physics field feature data set; processing the standardized multi-physics field feature data set to obtain a multi-physics field coupling feature vector; processing the multi-physics field coupling feature vector and the temperature sensor network topology to obtain a temperature sensor health status assessment result; processing the temperature sensor health status assessment result and historical fault data to obtain a sensor fault prediction report; and processing faulty sensor data and surrounding healthy sensor data to obtain a data compensation result for the faulty sensor. The present invention can significantly improve system reliability and safety under temperature sensor network monitoring in complex industrial environments, and reduce downtime losses and safety risks caused by sensor failures.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and more particularly to a temperature sensor fault diagnosis method and system. Background Art

[0002] In some high-temperature industrial furnace system temperature monitoring scenarios, such as those in steel smelting, glass manufacturing, or ceramic firing, the temperature sensor network is simultaneously subject to the combined influence of high-temperature thermal fields, molten metal flow fields, electromagnetic induction fields, and structural vibration fields. In such complex environments, temperature sensors have a high failure rate and complex failure modes. Relying solely on temperature data for fault diagnosis has low accuracy. Once a sensor fails, the loss of critical temperature data can lead to product quality fluctuations or safety accidents.

[0003] In existing technologies, fault diagnosis of temperature sensors in industrial environments mainly relies on single temperature data analysis or simple threshold detection. These methods have low accuracy under the interference of complex physical environments. Traditional fault diagnosis methods have the following main shortcomings:

[0004] Dependence on a single data source: Most existing methods only use temperature data for fault diagnosis, ignoring the interaction between multiple physical fields, resulting in low diagnostic accuracy in complex physical environments.

[0005] Weak ability to identify physical environment interference: Unable to effectively distinguish between data anomalies caused by physical environment interference (such as electromagnetic interference and vibration) and actual sensor hardware failures, resulting in a high false alarm rate.

[0006] Lack of predictive analysis: Most methods only detect faults after they occur, lacking the ability to predict the health status of sensors and providing early warning of potential faults.

[0007] Low data compensation accuracy: After a sensor failure, the accuracy of traditional compensation methods is limited. Especially under conditions with large temperature gradients and complex environments, the compensation data deviates greatly from the actual value.

[0008] Poor model adaptability: Existing methods have weak generalization capabilities for different industrial scenarios and are difficult to adapt to changing working conditions and the special needs of different industries. Summary of the Invention

[0009] The present invention provides a temperature sensor fault diagnosis method and system to solve the technical problems in the above-mentioned related technologies.

[0010] The present invention provides a temperature sensor fault diagnosis method, comprising:

[0011] The multi-source sensor data of the industrial furnace is processed through a multi-physics hierarchical preprocessing network to obtain a standardized multi-physics feature data set.

[0012] Processing the standardized multi-physics field feature data set through a multi-field coupling feature extraction network enhanced by physical knowledge to obtain a multi-physics field coupling feature vector;

[0013] Processing the multi-physics field coupling eigenvector and the temperature sensor network topology through a physical consistency graph neural network to obtain a temperature sensor health status assessment result;

[0014] Processing the temperature sensor health status assessment result and historical fault data through a multi-physics field driven sensor fault prediction model to obtain a sensor fault prediction report;

[0015] The faulty sensor data and the surrounding healthy sensor data are processed through a sensor data compensation network assisted by a physical model to obtain the data compensation result of the faulty sensor.

[0016] Furthermore, in the step of processing the multi-source sensor data of the industrial furnace through a multi-physics field hierarchical preprocessing network, the mathematical expression of the multi-physics field hierarchical preprocessing network is:

[0017] ;

[0018] in:

[0019] Represents raw multiphysics data;

[0020] represents the data acquisition function;

[0021] represents the noise filtering function;

[0022] represents the feature extraction function;

[0023] Represents the output normalized multiphysics feature dataset.

[0024] Furthermore, in the step of processing the standardized multi-physics field feature data set by the multi-field coupling feature extraction network enhanced by physical knowledge, the mathematical expression of the multi-field coupling feature extraction network enhanced by physical knowledge is:

[0025] ;

[0026] in:

[0027] represents a standardized multiphysics feature dataset;

[0028] represents the physical field-specific feature extraction function;

[0029] Represents the physical field interaction characteristic analysis function;

[0030] represents the feature concatenation function;

[0031] Represents the multi-field fusion self-attention function;

[0032] represents the multiphysics coupling eigenvector.

[0033] Furthermore, in the step of processing the multi-physics field coupling eigenvector and the temperature sensor network topology structure through the physical consistency graph neural network, the mathematical expression of the physical consistency graph neural network is:

[0034] ;

[0035] in:

[0036] represents the multi-physics coupling eigenvector;

[0037] Represents sensor location and topological relationship data;

[0038] Represents the graph construction function;

[0039] represents the graph convolution function;

[0040] represents the health status assessment function;

[0041] Indicates the temperature sensor health status assessment result.

[0042] Furthermore, the sensor network graph structure generated by the graph construction function in the physical consistency graph neural network is represented as:

[0043] ;

[0044] in:

[0045] Represents a node set consisting of N sensor nodes;

[0046] represents a set of edges, connecting related sensor nodes;

[0047] Represents the optimized adjacency matrix, which defines the connection relationship and weight between nodes.

[0048] Furthermore, the graph convolution function calculation formula in the physical consistency graph neural network is:

[0049] ;

[0050] in:

[0051] represents the adjacency matrix after adding the self-loop;

[0052] Represents the corresponding degree matrix, whose diagonal elements ;

[0053] represents the N-dimensional identity matrix;

[0054] represents the weight matrix of the lth layer;

[0055] Represents the node feature matrix of the lth layer;

[0056] represents a non-linear activation function.

[0057] Furthermore, in the step of processing the temperature sensor health status assessment result and historical fault data by a multi-physics field driven sensor fault prediction model, the mathematical expression of the multi-physics field driven sensor fault prediction model is:

[0058] ;

[0059] in:

[0060] Indicates the health status assessment result of the temperature sensor;

[0061] Represents multi-physics environment parameters;

[0062] represents the multi-physics coupling eigenvector;

[0063] represents the health feature extraction function;

[0064] represents the fault prediction function;

[0065] represents the fault analysis function;

[0066] Indicates a sensor failure prediction report.

[0067] Furthermore, in the step of processing the faulty sensor data and the surrounding healthy sensor data through the physical model-assisted sensor data compensation network, the mathematical expression of the physical model-assisted sensor data compensation network is:

[0068] ;

[0069] in:

[0070] Represents temperature sensor network data;

[0071] Indicates fault sensor information;

[0072] represents the physical model parameters;

[0073] represents the relevant sensor selection function;

[0074] represents the data compensation function;

[0075] Indicates the data compensation result of the faulty sensor.

[0076] Furthermore, the calculation formula of the physical constraint graph attention compensation function in the physical model-assisted sensor data compensation network is:

[0077] ;

[0078] in:

[0079] Represents the subgraph containing the faulty sensor and related sensors;

[0080] Represents the timing characteristics of related sensors;

[0081] Represents the parameters of the heat conduction physical model;

[0082] represents the physical constraint graph attention compensation function;

[0083] Indicates the data compensation result of the faulty sensor.

[0084] A temperature sensor fault diagnosis system, comprising:

[0085] A multi-physics hierarchical preprocessing network is used to process multi-source sensor data of industrial furnaces and obtain standardized multi-physics feature data sets;

[0086] A multi-field coupling feature extraction network enhanced by physical knowledge is used to process the standardized multi-physics field feature data set to obtain a multi-physics field coupling feature vector;

[0087] A physical consistency graph neural network is used to process the multi-physical field coupling eigenvector and the temperature sensor network topology to obtain a temperature sensor health status assessment result;

[0088] A multi-physics field driven sensor fault prediction model is used to process the temperature sensor health status assessment results and historical fault data to obtain a sensor fault prediction report;

[0089] The physical model-assisted sensor data compensation network is used to process the faulty sensor data and the surrounding healthy sensor data to obtain the data compensation result of the faulty sensor.

[0090] The beneficial effects of the present invention are:

[0091] Improved fault prediction performance: Compared with traditional methods that only use temperature data, fault prediction accuracy is improved and the average warning lead time is extended, providing a sufficient time window for preventive maintenance.

[0092] Improved data compensation accuracy: In the event of sensor failure, the average relative error between the compensated data and the actual value is reduced, and high accuracy can be maintained in environments with large temperature fluctuations, meeting industrial process control requirements.

[0093] Enhanced false fault identification capabilities: Through multi-physical field information fusion, the system can accurately distinguish between anomalies caused by physical interference and actual sensor hardware failures, reducing the false alarm rate and significantly reducing unnecessary maintenance costs.

[0094] Improved system robustness: The system can maintain a high level of fault prediction accuracy in the presence of strong vibration, electromagnetic interference, and extreme temperature gradients, demonstrating excellent adaptability to complex operating conditions.

[0095] Computing efficiency optimization: Through edge deployment optimization, the system can run in real time on edge servers at industrial sites, reducing average processing latency to meet real-time monitoring needs.

[0096] Enhanced maintenance decision support: Comprehensive failure prediction reports provide detailed analysis of failure risk, type, and physical cause, providing maintenance personnel with actionable decision-making recommendations and shortening the mean time to troubleshoot. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 The present invention is a flow chart of a temperature sensor fault diagnosis method. DETAILED DESCRIPTION

[0098] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0099] At least one embodiment of the present invention discloses a temperature sensor fault diagnosis method, such as Figure 1 As shown, including:

[0100] Step 1: Process the multi-source sensor data of the industrial furnace through a multi-physics hierarchical preprocessing network to obtain a standardized multi-physics feature dataset;

[0101] The multi-physics hierarchical preprocessing network consists of a data acquisition module, a noise filtering module, and a feature extraction module. The raw data collected by the data acquisition module is transmitted to the noise filtering module for noise reduction. The noise-reduced data is then passed to the feature extraction module to generate standardized features.

[0102] The mathematical expression of the multi-physics hierarchical preprocessing network is:

[0103] ;

[0104] in: Represents raw multiphysics data; represents the data acquisition function; represents the noise filtering function; represents the feature extraction function; represents the output standardized multiphysics feature dataset;

[0105] Step 1 includes the following sub-steps:

[0106] Sub-step 1.1: Collect multi-physics data of the industrial furnace system using a multi-scale adaptive sampling algorithm to obtain an original multi-physics data set;

[0107] Substep 1.1 uses a multiscale adaptive sampling algorithm to handle the data collection requirements of different physical fields in the industrial furnace system. The input is real-time signals from various sensors, and the output is a raw multi-physics dataset. This algorithm automatically adjusts the sampling frequency based on the rate of change of different physical fields, using high-frequency sampling for rapidly changing physical quantities and low-frequency sampling for slowly changing physical quantities, thereby optimizing resource allocation.

[0108] The mathematical expression of the multi-scale adaptive sampling algorithm is:

[0109] ;

[0110] in: represents the sampling frequency of the i-th physical field at time t; Indicates the basic sampling frequency; Represents the basic sampling coefficient of the i-th physical field; Represents the adaptive adjustment coefficient; Represents the change of the i-th physical field signal per unit time; Indicates the maximum allowed change in the signal.

[0111] The acquired raw multiphysics dataset includes:

[0112] Temperature field data: contains multiple point temperature values ​​and their timestamps;

[0113] Fluid field data: including flow velocity, flow direction, and pressure parameters;

[0114] Electromagnetic field data: including electric field strength, magnetic field strength, and electromagnetic interference spectrum parameters;

[0115] Vibration field data: including acceleration, amplitude, and frequency parameters;

[0116] Environmental parameter data: including humidity, air pressure, and ambient temperature parameters;

[0117] Sub-step 1.2: Process the original multiphysics dataset using a multiphysics adaptive noise filtering algorithm to obtain a denoised multiphysics dataset;

[0118] Sub-step 1.2 uses a multi-physics adaptive noise filtering algorithm to process the original multi-physics dataset obtained in sub-step 1.1, outputting a denoised multi-physics dataset. This algorithm combines wavelet transforms and Kalman filters to perform adaptive noise reduction based on the noise characteristics of different physical fields.

[0119] The mathematical expression of the multiphysics adaptive noise filtering algorithm is:

[0120] ;

[0121] in: Represents the original data of the i-th physical field; represents the wavelet transform function; represents the inverse wavelet transform function; represents the threshold function, is the threshold parameter of the i-th physical field; represents the Kalman filter function; represents the denoised data of the i-th physical field;

[0122] For different physical fields, the threshold parameter Adaptively calculated using the following formula:

[0123] ;

[0124] in: represents the standard deviation estimate of the i-th physical field noise; Indicates the number of data points;

[0125] Sub-step 1.3: Process the denoised multi-physics dataset using a physical knowledge fusion feature extraction network to obtain a standardized multi-physics feature dataset;

[0126] In substep 1.3, a physics-based feature extraction network is used to process the denoised multiphysics dataset obtained in substep 1.2, outputting a standardized multiphysics feature dataset. This network combines physics model knowledge with deep learning techniques to extract key features related to temperature sensor failures from multi-source heterogeneous data.

[0127] The physical knowledge fusion feature extraction network consists of a feature extraction module, a physical knowledge constraint module, and a feature normalization module. The initial features extracted by the feature extraction module are passed to the physical knowledge constraint module for screening and optimization. The optimized features are then passed to the feature normalization module to generate the final standardized features.

[0128] The mathematical expression of the network is:

[0129] ;

[0130] in: represents the multiphysics data after noise reduction; represents the feature extraction function; Represents physical knowledge constraint function; represents the feature normalization function; represents a standardized multiphysics feature dataset;

[0131] The feature extraction function is implemented based on the multi-layer perceptron:

[0132] ;

[0133] in: Represents the activation function ReLU; denote the first and second weight matrices respectively; Represents the bias vector.

[0134] The physical knowledge constraint function adjusts the features through physical rules:

[0135] ;

[0136] in: Represents element-wise multiplication; represents the physical correlation mask matrix; represents the physical consistency correction function; Represents the physical knowledge correction coefficient.

[0137] The standardized multiphysics feature dataset contains the following feature categories:

[0138] Temperature field characteristics: temperature distribution, gradient, and rate of change;

[0139] Fluid field characteristics: velocity distribution, turbulence intensity, pressure gradient;

[0140] Electromagnetic field characteristics: electromagnetic interference intensity, spectrum characteristics;

[0141] Vibration field characteristics: vibration mode and frequency characteristics;

[0142] Cross-coupling features: features that describe the interactions between different physical fields;

[0143] Step 2: Process the standardized multi-physics field feature data set through the multi-field coupling feature extraction network enhanced by physical knowledge to obtain the multi-physics field coupling feature vector;

[0144] In step 2, a multi-field coupling feature extraction network augmented with physics knowledge processes the standardized multi-physics feature dataset obtained in step 1 and outputs a multi-physics coupling feature vector. This network consists of a physics-specific feature extraction module, a physics interaction feature analysis module, and a multi-field fusion self-attention module. The single-field features extracted by the physics-specific feature extraction module are passed to the physics interaction feature analysis module for cross-field correlation analysis. The outputs of these two modules are then fed into the multi-field fusion self-attention module to generate the final multi-physics coupling feature vector.

[0145] The mathematical expression of the network is:

[0146] ;

[0147] in: represents a standardized multiphysics feature dataset; represents the physical field-specific feature extraction function; Represents the physical field interaction characteristic analysis function; represents the feature concatenation function; Represents the multi-field fusion self-attention function; represents the multiphysics coupling eigenvector.

[0148] Step 2 includes the following sub-steps:

[0149] Sub-step 2.1: Process the standardized multi-physics field feature dataset through a physics-specific feature extraction network to obtain a single-field feature vector set;

[0150] Substep 2.1 uses a physics-specific feature extraction network to process the standardized multi-physics feature dataset obtained in step 1, outputting a set of single-field feature vectors. This network consists of four parallel feature extraction branches, each dedicated to processing the characteristics of a specific physics field. This network uses a specifically designed network structure to capture the characteristic patterns of different physics fields.

[0151] The mathematical expression of the physics-specific feature extraction network is:

[0152] ;

[0153] ;

[0154] in: , , , Respectively represent the standardized characteristic data of temperature field, electromagnetic field, vibration field and fluid field; , , , Represent the feature extraction functions of temperature field, electromagnetic field, vibration field and fluid field respectively; , , , Respectively represent the extracted characteristic vectors of temperature field, electromagnetic field, vibration field and fluid field; Represents a single-field eigenvector set, which contains the eigenvectors of four physical fields.

[0155] The feature extraction function for the temperature field uses a time and frequency dual-path feature extractor:

[0156] ;

[0157] in: represents a time domain convolutional neural network; represents a frequency domain convolutional neural network; represents fast Fourier transform; Represents the adaptive weight coefficient, which is learned through the attention mechanism.

[0158] The feature extraction function of the electromagnetic field adopts the method of combining wavelet packet transform and CNN:

[0159] ;

[0160] in: represents wavelet packet transform; Represents a convolutional neural network that processes wavelet coefficients.

[0161] The improved long short-term memory network is used for the feature extraction function of the vibration field:

[0162] ;

[0163] in: represents a bidirectional long short-term memory network.

[0164] The feature extraction function for the fluid field adopts a recursive neural network based on phase space reconstruction:

[0165] ;

[0166] in: represents the phase space reconstruction function; represents a gated recurrent unit network.

[0167] Sub-step 2.2: Process the single-field eigenvector set using a cross-field correlation analysis algorithm constrained by physical laws to obtain the physical field interaction characteristic matrix;

[0168] Substep 2.2 uses a cross-field correlation analysis algorithm constrained by physical laws to process the single-field eigenvector set obtained in substep 2.1 and output a physical field interaction feature matrix. This algorithm, based on physics domain knowledge, analyzes the coupling relationships and interactions between different physical fields and generates a feature matrix that represents the interaction effects between the physical fields.

[0169] The mathematical expression of the cross-field correlation analysis algorithm constrained by physical laws is:

[0170] ;

[0171] ;

[0172] in: Represents the eigenvector of the i-th physical field; Represents the interaction feature extraction function between the i-th and j-th physical fields; Represents the interaction characteristics between the i-th and j-th physical fields; represents the physics interaction characteristic matrix.

[0173] The interactive feature extraction function between the temperature field and the electromagnetic field is designed based on the Joule heating effect and the principle of electromagnetic induction:

[0174] ;

[0175] in: represents the time derivative of the temperature characteristic; Represents material-related conductivity parameters.

[0176] The interactive feature extraction function between temperature field and vibration field is designed based on thermal expansion and thermoelastic effects:

[0177] ;

[0178] in: Represents the spatial gradient of temperature characteristics; Indicates the coefficient of thermal expansion.

[0179] Similarly, the interaction feature extraction functions between other physical field pairs are also designed based on the corresponding physical laws, and finally form a complete physical field interaction feature matrix R.

[0180] Sub-step 2.3: Process the single-field feature vector set and the physical field interaction feature matrix through the multi-field fusion self-attention network to obtain the multi-physics field coupling feature vector;

[0181] In substep 2.3, a multi-field fusion self-attention network is used to process the single-field feature vectors obtained in substep 2.1 and the physical field interaction feature matrix obtained in substep 2.2, outputting a multi-field coupled feature vector. This network uses a self-attention mechanism to capture the complex relationships between different features and ensures that the features conform to physical laws through physical consistency constraints.

[0182] The multi-field fusion self-attention network consists of a feature initialization fusion layer, a multi-head self-attention layer, and a physical consistency constraint layer. The feature initialization fusion layer fuses the single-field feature vector and the interaction feature matrix into a unified feature. The multi-head self-attention layer captures the interrelationships between different features. The physical consistency constraint layer ensures that the feature representation conforms to physical laws.

[0183] The mathematical expression of the network is:

[0184] ;

[0185] in: Represents the output of the last self-attention layer; Represents the physical law constraint function; represents the self-attention balance parameter; represents the multiphysics coupling eigenvector.

[0186] The initial feature fusion process is expressed as:

[0187] ;

[0188] in: Represents vector concatenation operation; Represents the operation of converting a matrix to a vector.

[0189] The calculation formula of the multi-head self-attention layer is:

[0190] ;

[0191] in: Represents the characteristics of the i-th node in the l-th layer; and Represent the weight matrix and bias vector of the lth layer respectively; represents the attention weight from node j to node i in layer l, which is calculated as:

[0192] ;

[0193] ;

[0194] in: represents the learnable self-attention vector; Represents vector concatenation operation; Represents the ReLU activation function.

[0195] The physical consistency constraint function includes the physical models of the first law of thermodynamics, the heat conduction equation, and the law of electromagnetic induction to ensure that the feature representation conforms to physical laws:

[0196] ;

[0197] in: represents the kth physical constraint function; represents the weight of the k-th physical constraint; Indicates the total number of physical constraints.

[0198] The final output multi-physics coupling feature vector It is a d-dimensional vector, where d is usually set to 128 or 256, which contains rich information about the complex interactions between physical fields and will be used for subsequent sensor health status assessment.

[0199] Step 3: Process the multi-physics field coupling feature vector and the temperature sensor network topology through the physical consistency graph neural network to obtain the temperature sensor health status assessment result;

[0200] In step 3, a physical consistency graph neural network processes the multi-physics coupling eigenvectors obtained in step 2 and the topological structure of the temperature sensor network to output the health status assessment results of the temperature sensors. This network models sensors as nodes in a graph and the physical and data relationships between sensors as edges. Through message passing and physical constraints, it achieves highly accurate health status assessment.

[0201] The physical consistency graph neural network consists of a graph construction module, a graph convolution module, and a health status assessment module. The sensor network graph structure generated by the graph construction module is passed to the graph convolution module for feature propagation and update. The output of the graph convolution module is passed to the health status assessment module to generate the final health status assessment result.

[0202] The mathematical expression of the network is:

[0203] ;

[0204] in: represents the multi-physics coupling eigenvector; Represents sensor location and topological relationship data; Represents the graph construction function; represents the graph convolution function; represents the health status assessment function; Indicates the temperature sensor health status assessment result.

[0205] Step 3 includes the following sub-steps:

[0206] Sub-step 3.1: Process the sensor spatial layout information and physical association relationships through a dynamic adaptive graph construction algorithm to obtain the sensor network graph structure;

[0207] Sub-step 3.1 uses a dynamic adaptive graph construction algorithm to process the spatial layout information and physical relationships of the sensors and output the sensor network graph structure. This algorithm comprehensively considers the physical distance between sensors, data correlation, and physical environmental factors to construct a graph structure that accurately reflects the complex relationships between sensors.

[0208] The mathematical expression of the dynamic adaptive graph construction algorithm is:

[0209] ;

[0210] in: Represents a node set consisting of N sensor nodes; represents a set of edges, connecting related sensor nodes; Represents the optimized adjacency matrix, which defines the connection relationship and weight between nodes.

[0211] The algorithm first calculates the physical distance matrix based on the spatial position of the sensor:

[0212] ;

[0213] in: and Represent the three-dimensional space coordinates of sensors i and j respectively; represents the Euclidean distance between two sensors; represents the Gaussian kernel width parameter; Indicates the distance threshold.

[0214] The correlation matrix is ​​then calculated based on the sensor data:

[0215] ;

[0216] in: and represent the temperature time series data of sensors i and j respectively; Represents the Pearson correlation coefficient calculation function.

[0217] The physical distance matrix and the correlation matrix are fused to generate the initial adjacency matrix:

[0218] ;

[0219] in: represents a balance parameter that adjusts the relative importance of physical distance and data relevance.

[0220] Finally, an adaptive threshold is applied to optimize the adjacency matrix:

[0221] ;

[0222] in: Represents the adaptive threshold, which is determined by the graph connectivity optimization algorithm.

[0223] Sub-step 3.2: Process the sensor network graph structure and multi-physics field coupling feature vectors through the physical constraint graph convolutional network to obtain the node health feature matrix.

[0224] In substep 3.2, a physically constrained graph convolutional network (GCN) processes the sensor network graph structure obtained in substep 3.1 and the multi-physics coupling feature vectors obtained in step 2, outputting a node health feature matrix. This network uses graph convolution operations to propagate and update features between nodes and introduces physical constraints to ensure that feature evolution conforms to physical laws such as heat conduction.

[0225] The physical constraint graph convolutional network consists of a node feature initialization module, a multi-layer graph convolution module, and a physical constraint module. The node feature initialization module assigns multi-physics coupling features to corresponding sensor nodes, the multi-layer graph convolution module performs message passing on the graph, and the physical constraint module introduces physical constraints for heat conduction.

[0226] The mathematical expression of the network is:

[0227] ;

[0228] in: Represents the node feature matrix of the lth layer; Indicates the feature update amount of the lth layer; represents a nonlinear activation function; Represents the node health feature matrix output by the last layer.

[0229] The node feature initialization process is:

[0230] ;

[0231] in: represents the multi-physics coupling eigenvector; and Represent the initialized weight matrix and the initialized bias vector respectively; represents the initial features of the i-th sensor node.

[0232] The calculation formula of the graph convolution layer is:

[0233] ;

[0234] in: represents the adjacency matrix after adding the self-loop; Represents the corresponding degree matrix, whose diagonal elements ; represents the N-dimensional identity matrix; represents the weight matrix of layer l.

[0235] The heat conduction physical constraint is introduced through the following loss function:

[0236] ;

[0237] in: represents the discrete Laplace operator; represents the virtual time step; represents the thermal diffusivity; represents the Frobenius norm.

[0238] Sub-step 3.3: Process the node health feature matrix through a multi-level threshold health status classifier to obtain the temperature sensor health status assessment result;

[0239] Substep 3.3 uses a multi-level threshold health status classifier to process the node health feature matrix obtained in substep 3.2 and output the temperature sensor health status assessment result. This classifier uses a multi-level threshold judgment mechanism, combined with time persistence verification and interference elimination strategies, to accurately assess the sensor health status.

[0240] The multi-level threshold health status classifier consists of a health score calculation module, a dynamic threshold generation module, and a status determination module. The health score calculation module converts node health characteristics into numerical health scores. The dynamic threshold generation module calculates judgment thresholds based on historical data and current operating conditions. The status determination module determines the health status of the sensor based on the health score and threshold.

[0241] The mathematical expression of the classifier is:

[0242] ;

[0243] in: represents the node health feature matrix; represents the health score calculation function; represents a dynamic threshold set; represents the state determination function; Indicates the temperature sensor health status assessment result.

[0244] The health score calculation formula is:

[0245] ;

[0246] in: represents the health feature vector of the i-th sensor node; represents a multilayer perceptron; represents the health score of the i-th sensor, 1 represents completely healthy and 0 represents completely faulty.

[0247] The dynamic threshold calculation formula is:

[0248] ;

[0249] in: and represent the mean and standard deviation of historical health scores, respectively; represents the working condition adjustment function; Indicates the current working condition parameters.

[0250] The status determination rules are:

[0251] ;

[0252] in: and Indicates the minimum duration thresholds for the fault state and warning state respectively.

[0253] Interference elimination is achieved by analyzing the correlation between health score anomalies and physical field disturbances:

[0254] ;

[0255] in: Represents the time series correlation analysis function; A time series representing a specific physical field disturbance; Represents the correlation threshold.

[0256] The final output of the temperature sensor health status assessment result is:

[0257] ;

[0258] in: Indicates the sensor index; Indicates health status (healthy, warning, or faulty); Indicates the confidence level of the state judgment.

[0259] Step 4: Process the temperature sensor health status assessment results and historical failure data through the multi-physics field-driven sensor failure prediction model to obtain a sensor failure prediction report;

[0260] In step 4, a multiphysics-driven sensor failure prediction model processes the temperature sensor health assessment results and historical failure data obtained in step 3 to produce a sensor failure prediction report. This model, combining the sensor health evolution characteristics with multiphysics environmental parameters, uses deep learning technology to predict the risk of future sensor failures and analyze the possible failure types and causes.

[0261] The multi-physics-driven sensor fault prediction model consists of a health feature extraction module, a fault prediction module, and a fault analysis module. The health feature extraction module extracts health evolution features that are fed into the fault prediction module for risk prediction. The prediction results from the fault prediction module, along with multi-physics coupling features, are then fed into the fault analysis module to generate the final fault prediction report.

[0262] The mathematical expression of the model is:

[0263] ;

[0264] in: Indicates the health status assessment result of the temperature sensor; Represents multi-physics environment parameters; represents the multi-physics coupling eigenvector; represents the health feature extraction function; represents the fault prediction function; represents the fault analysis function; Indicates a sensor failure prediction report.

[0265] Step 4 includes the following sub-steps:

[0266] Sub-step 4.1: Process the time series of the temperature sensor health status assessment results using a time series health feature extraction algorithm to obtain the sensor health evolution feature vector;

[0267] Sub-step 4.1 uses a time series health feature extraction algorithm to process the time series of temperature sensor health assessment results obtained in step 3 and output a sensor health evolution feature vector. This algorithm uses time series analysis to extract features that reflect sensor performance degradation trends and patterns from the historical data of health assessment results.

[0268] The mathematical expression of the time series health feature extraction algorithm is:

[0269] ;

[0270] in: Represents trend characteristics and describes the long-term trend of health status; Represents the fluctuation characteristics, describing the short-term fluctuation pattern of health status; Represents periodic characteristics and describes the periodic changes in health status; Represents event characteristics, describing the pattern of how health status is affected by specific events; Represents the sensor health evolution feature vector.

[0271] The algorithm first performs time series preprocessing, including standardization and denoising:

[0272] ;

[0273] in: represents the health score of the i-th sensor at time t; and denote the mean and standard deviation of sensor health scores, respectively; represents the standardized health score.

[0274] Then the improved seasonal decomposition algorithm is applied to extract trend features:

[0275] ;

[0276] in: Indicates trend items, reflecting long-term change trends; represents the seasonal term, reflecting periodic changes; Represents the residual term, reflecting random fluctuations.

[0277] The trend characteristic calculation formula is:

[0278] ;

[0279] in: Indicates the rate of change of the trend; Indicates the acceleration of the trend.

[0280] The calculation formula for fluctuation characteristics is:

[0281] ;

[0282] in:

[0283] represents the standard deviation of the residual term;

[0284] represents the skewness of the residual term;

[0285] Represents the kurtosis of the residual term.

[0286] Periodic features and event features are extracted through wavelet analysis and autocorrelation function, and finally all features are integrated into the sensor health evolution feature vector.

[0287] Sub-step 4.2: Process the sensor health evolution feature vector and multi-physics environment parameters through the attention-enhanced long short-term memory prediction model to obtain the sensor failure risk prediction result;

[0288] Substep 4.2 uses an attention-enhanced long-short-term memory prediction model to process the sensor health evolution feature vector and multi-physics environmental parameters obtained in substep 4.1, outputting a sensor failure risk prediction result. This model, combining an attention mechanism with a long-short-term memory network, can capture failure risk patterns at different time scales and account for the influence of physical environmental factors.

[0289] The mathematical expression of the attention-enhanced long-term and short-term memory prediction model is:

[0290] ;

[0291] in: represents the sensor health evolution feature vector; Represents multi-physics environment parameters; Long short-term memory networks representing enhanced attention; Indicates the sensor failure risk prediction result.

[0292] The calculation formula of the physical field attention mechanism is:

[0293] ;

[0294] in: represents the jth physical field environment parameter; Represents vector concatenation operation; and Represents the weights and biases of the attention network; Represents the attention weight of the j-th physical field environment parameter.

[0295] The update formula of the long short-term memory network is:

[0296] ;

[0297] in: and Represent the hidden state and cell state of the LSTM unit at time t respectively; and Represent the hidden state and cell state of the LSTM unit at time t;1 respectively.

[0298] The final failure risk prediction results include failure risk probabilities and uncertainty estimates within different time frames:

[0299] ;

[0300] in: Indicates that sensors i in the future The risk probability of failure within a certain time period; Expresses uncertainty estimates for risk predictions; Indicates different forecast time ranges (such as 24 hours, 72 hours, 168 hours).

[0301] Sub-step 4.3: Process the sensor fault risk prediction results and the multi-physics field coupling feature vector through the multi-modal fault analysis engine to obtain a sensor fault prediction report;

[0302] Sub-step 4.3 uses a multimodal fault analysis engine to process the sensor fault risk prediction results obtained in sub-step 4.2 and the multi-physics coupling feature vectors obtained in step 2, and outputs a sensor fault prediction report. This engine uses a combination of machine learning and rule-based reasoning to analyze potential fault types, physical causes, and key evolution time points to generate a comprehensive fault prediction report.

[0303] The mathematical expression of the multimodal fault analysis engine is:

[0304] ;

[0305] in: represents the sensor failure risk prediction result; represents the multi-physics coupling eigenvector; Indicates the health status assessment result of the temperature sensor; represents the multimodal fault analysis function; Indicates a sensor failure prediction report.

[0306] The fault type classifier is implemented based on support vector machine:

[0307] ;

[0308] in: represents the failure risk vector of sensor i; Represents vector concatenation operation; represents a support vector machine classifier; Indicates the predicted failure type.

[0309] The physical field contribution analysis is achieved through gradient calculation:

[0310] ;

[0311] in: Represents the failure risk versus physical field parameters The partial derivative of Indicates the contribution of the j-th physical field to the failure risk.

[0312] The calculation formula for key time point markers is:

[0313] ;

[0314] in: represents the health score of sensor i at time t; Indicates the change threshold.

[0315] The resulting sensor failure prediction report contains the following:

[0316] Basic sensor information (ID, location, type);

[0317] The probability and uncertainty of failure risks in each future time period;

[0318] Predicted fault types and probability ranking;

[0319] Analysis of the contribution of each physical field to failure risk;

[0320] Marking key time points in the evolution of health status;

[0321] Failure prevention recommendations and optimal maintenance time windows.

[0322] Step 5: Process the faulty sensor data and surrounding healthy sensor data through the physical model-assisted sensor data compensation network to obtain the data compensation result of the faulty sensor.

[0323] In step 5, a physical model-assisted sensor data compensation network processes the faulty sensor data and surrounding healthy sensor data, outputting the data compensation result for the faulty sensor. This network, combined with a physical heat transfer model and a graph attention mechanism, can infer the actual value of the failed sensor based on physical laws and adjacent sensor data in the event of a sensor failure, achieving high-precision data compensation.

[0324] The physical model-assisted sensor data compensation network consists of a fault detection module, a relevant sensor selection module, and a data compensation module. The fault signal detected by the fault detection module is transmitted to the relevant sensor selection module, which then selects the relevant sensor data and inputs it into the data compensation module to generate the final data compensation result.

[0325] The mathematical expression of the network is:

[0326] ;

[0327] in: Represents temperature sensor network data; Indicates fault sensor information; represents the physical model parameters; represents the relevant sensor selection function; represents the data compensation function; Indicates the data compensation result of the faulty sensor.

[0328] Step 5 includes the following sub-steps:

[0329] Sub-step 5.1: Process the real-time sensor data and fault prediction report through the real-time anomaly detection trigger to obtain a data compensation start signal;

[0330] Sub-step 5.1 uses a real-time anomaly detection trigger to process the real-time sensor data and the fault prediction report obtained in step 4, and outputs a data compensation start signal. This trigger combines statistical methods with prediction information to quickly detect sensor data anomalies or interruptions and trigger the data compensation process.

[0331] The mathematical expression of the real-time anomaly detection trigger is:

[0332] ;

[0333] in: Represents real-time sensor data; Indicates a failure prediction report; represents the anomaly detection scoring function; Indicates triggering decision function; Indicates the data compensation start signal.

[0334] The anomaly detection scoring function is calculated based on a combination of statistical deviation and predicted risk:

[0335] ;

[0336] in: Indicates the Z score of the sensor value; represents the failure risk probability of sensor i; and Indicates the first and second score weight coefficients.

[0337] The trigger decision function determines whether to start data compensation through threshold judgment:

[0338] ;

[0339] in: Indicates the trigger threshold; Indicates a situation where sensor data is interrupted.

[0340] The data compensation start signal contains the following information: fault sensor ID, fault detection timestamp, fault type indication, and anomaly score value.

[0341] Sub-step 5.2: Process the sensor network topology and fault sensor information through a physically guided relevant sensor selection algorithm to obtain the optimal relevant sensor set;

[0342] Substep 5.2 uses a physics-based, relevant sensor selection algorithm to process the sensor network topology and the faulty sensor information obtained in substep 5.1, outputting the optimal set of relevant sensors. This algorithm, based on a physical heat transfer model and data correlation analysis, selects a subset of sensors from the sensor network that are highly correlated with the faulty sensor, providing reliable input data for subsequent data compensation.

[0343] The mathematical expression of the physical-oriented sensor selection algorithm is:

[0344] ;

[0345] in: Represents the sensor network graph structure; Indicates the fault sensor index; Represents the parameters of the heat conduction physical model; represents the sensor ranking function; represents a function that selects the first K elements; represents the optimal set of correlated sensors.

[0346] The sensor ranking function is calculated based on the composite relevance score:

[0347] ;

[0348] in: represents the physical distance between sensors i and j; represents the time delay mutual correlation coefficient between sensors i and j; represents the time delay corresponding to the maximum cross-correlation; represents the correlation score based on the heat flow propagation path; 、 and represents the first, second and third correlation weight coefficients; Represents the distance constant.

[0349] The heat flow propagation path analysis is based on the heat conduction equation:

[0350] ;

[0351] in: represents the thermal diffusivity; represents the Laplace operator; Represents a heat source term.

[0352] The optimal correlated sensor set finally selected contains K sensors with the highest correlation with the faulty sensor, and K is usually set to 5-10.

[0353] Sub-step 5.3: Process the optimal relevant sensor set data through the physical constraint graph attention compensation network to obtain the data compensation result of the faulty sensor;

[0354] Sub-step 5.3 uses a physical constraint graph attention compensation network to process the optimal set of relevant sensors obtained in sub-step 5.2 and output the data compensation results for the faulty sensor. This network combines the graph attention mechanism with physical constraints to capture the complex relationships between different sensors and ensure that the generated compensation values ​​conform to physical laws.

[0355] The mathematical expression of the physical constraint graph attention compensation network is:

[0356] ;

[0357] in: Represents the subgraph containing the faulty sensor and related sensors; Represents the timing characteristics of related sensors; Represents the parameters of the heat conduction physical model; represents the physical constraint graph attention compensation function; represents the data compensation result of the i-th faulty sensor.

[0358] The timing characteristics of the relevant sensors are expressed as:

[0359] ;

[0360] in: represents the temperature value of sensor j at time t;k△t; Indicates the length of the time window; Indicates the time step.

[0361] The calculation formula of the graph attention layer is:

[0362] ;

[0363] ;

[0364] in: and Represents the characteristics of nodes i and j; represents the neighbor set of node i; represents the third weight matrix; Represents the graph attention vector; Represents vector concatenation operation; represents the activation function; represents the attention weight from node j to node i.

[0365] The physical constraint layer introduces the heat conduction law through the following loss function:

[0366] ;

[0367] in: represents the prediction function based on the physical model; represents the spatial distance between nodes i and j; represents the time delay between nodes i and j.

[0368] The final data compensation result includes the compensation value and its uncertainty estimation:

[0369] ;

[0370] in: represents the compensation value of fault sensor i at time t; represents an estimate of the standard deviation of the compensation value; Represents the confidence interval for the compensation value.

[0371] This implementation is suitable for temperature sensor network monitoring in complex industrial environments such as steel smelting, glass manufacturing, and power generation. It can significantly improve system reliability and safety, and reduce downtime losses and safety risks caused by sensor failures.

[0372] A temperature sensor fault diagnosis system, comprising:

[0373] A multi-physics hierarchical preprocessing network is used to process multi-source sensor data of industrial furnaces and obtain standardized multi-physics feature data sets;

[0374] A multi-field coupling feature extraction network enhanced by physical knowledge is used to process the standardized multi-physics field feature data set to obtain a multi-physics field coupling feature vector;

[0375] A physical consistency graph neural network is used to process the multi-physical field coupling eigenvector and the temperature sensor network topology to obtain a temperature sensor health status assessment result;

[0376] A multi-physics field driven sensor fault prediction model is used to process the temperature sensor health status assessment results and historical fault data to obtain a sensor fault prediction report;

[0377] The physical model-assisted sensor data compensation network is used to process the faulty sensor data and the surrounding healthy sensor data to obtain the data compensation result of the faulty sensor.

[0378] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A temperature sensor fault diagnosis method, characterized in that: The following steps are involved: The multi-source sensor data of the industrial furnace is processed through a multi-physics hierarchical preprocessing network to obtain a standardized multi-physics feature data set. The multi-physics fields include high-temperature thermal field, molten metal fluid field, electromagnetic induction field, and structural vibration field. Processing the standardized multi-physics field feature data set through a multi-field coupling feature extraction network enhanced by physical knowledge to obtain a multi-physics field coupling feature vector; Processing the multi-physics field coupling eigenvectors and the temperature sensor network topology through a physical consistency graph neural network to obtain a temperature sensor health status assessment result, wherein the physical consistency constraint function includes the first law of thermodynamics, the heat conduction equation, and the electromagnetic induction law physical model; The temperature sensor health status assessment results and historical fault data are processed by a multi-physics field driven sensor fault prediction model to obtain a sensor fault prediction report. The fault prediction model uses an attention-enhanced long short-term memory network combined with a physical field attention mechanism. The faulty sensor data and surrounding healthy sensor data are processed through a physical model-assisted sensor data compensation network to obtain the data compensation result of the faulty sensor. The data compensation network adopts a graph attention compensation mechanism based on the heat conduction physical model. The physical constraint graph in the physical model-assisted sensor data compensation network is based on the physical law of heat conduction and is calculated as follows: ； in: Represents the subgraph containing the faulty sensor and related sensors; Represents the temporal characteristics of related sensors. Related sensors are selected based on three dimensions: physical distance correlation, data correlation, and heat flow propagation path correlation; Represents the parameters of the heat conduction physical model; represents the physical constraint graph attention compensation function; represents the data compensation result of the i-th faulty sensor.

2. The temperature sensor fault diagnosis method according to claim 1, characterized in that: In the step of processing the multi-source sensor data of the industrial furnace through the multi-physics field hierarchical preprocessing network, a multi-scale adaptive sampling algorithm is used to automatically adjust the sampling frequency according to the change rate of different physical fields. High-frequency sampling is used for rapidly changing physical quantities, and low-frequency sampling is used for slowly changing physical quantities. The mathematical expression of the multi-physics field hierarchical preprocessing network is: ； in: Represents raw multiphysics data; represents the data acquisition function; represents the noise filtering function; represents the feature extraction function; Represents the output normalized multiphysics feature dataset.

3. The temperature sensor fault diagnosis method according to claim 2, characterized in that: In the step of processing the standardized multi-physics field feature data set through the multi-field coupling feature extraction network enhanced by physical knowledge, a physical correlation mask matrix and a physical consistency correction function are introduced. The mathematical expression of the multi-field coupling feature extraction network enhanced by physical knowledge is: ； in: represents a standardized multiphysics feature dataset; represents the physical field-specific feature extraction function; Represents the physical field interaction characteristic analysis function; represents the feature concatenation function; Represents the multi-field fusion self-attention function; represents the multiphysics coupling eigenvector.

4. The temperature sensor fault diagnosis method according to claim 3, characterized in that: In the step of processing the multi-physics field coupling eigenvectors and the temperature sensor network topology structure through the physical consistency graph neural network, the physical model constraints of the first law of thermodynamics, the heat conduction equation, and the electromagnetic induction law are introduced. The mathematical expression of the physical consistency graph neural network is: ； in: represents the multi-physics coupling eigenvector; Represents sensor location and topological relationship data; Represents the graph construction function; represents the graph convolution function; represents the health status assessment function; Indicates the temperature sensor health status assessment result.

5. The temperature sensor fault diagnosis method according to claim 4, characterized in that: The graph construction function in the physical consistency graph neural network adopts a dynamic adaptive graph construction algorithm, which comprehensively considers the physical distance, data correlation and physical environment factors of the sensors. The generated sensor network graph structure is expressed as follows: ； in: Represents a node set consisting of N sensor nodes; represents a set of edges, connecting related sensor nodes; Represents the optimized adjacency matrix, which defines the connection relationship and weight between nodes.

6. The temperature sensor fault diagnosis method according to claim 5, characterized in that: The graph convolution function in the physical consistency graph neural network introduces the physical constraint of heat conduction, and the calculation formula is: ； in: represents the adjacency matrix after adding the self-loop; Represents the corresponding degree matrix, whose diagonal elements , i and j represent the row index and column index in the matrix respectively; represents the N-dimensional identity matrix; represents the weight matrix of the lth layer; Represents the node feature matrix of the lth layer; represents a non-linear activation function.

7. The temperature sensor fault diagnosis method according to claim 6, characterized in that: In the step of processing the temperature sensor health status assessment results and historical fault data through the multi-physics field driven sensor fault prediction model, an attention-enhanced long short-term memory network is combined with a physical field attention mechanism. The mathematical expression of the multi-physics field driven sensor fault prediction model is: ； in: Indicates the health status assessment result of the temperature sensor; Represents multi-physics environment parameters; represents the multi-physics coupling eigenvector; represents the health feature extraction function; represents the fault prediction function; represents the fault analysis function; Indicates a sensor failure prediction report.

8. The temperature sensor fault diagnosis method according to claim 7, characterized in that: In the step of processing faulty sensor data and surrounding healthy sensor data through a physical model-assisted sensor data compensation network, a physical-oriented related sensor selection algorithm is adopted. The mathematical expression of the physical model-assisted sensor data compensation network is: ； in: Represents temperature sensor network data; Indicates fault sensor information; represents the physical model parameters; represents the relevant sensor selection function; represents the data compensation function; Indicates the data compensation result of the faulty sensor.

9. A temperature sensor fault diagnosis system, characterized in that: include: A multi-physics hierarchical preprocessing network is used to process multi-source sensor data of industrial furnaces to obtain a standardized multi-physics feature data set. The multi-physics fields include high-temperature thermal field, molten metal fluid field, electromagnetic induction field, and structural vibration field. A multi-field coupling feature extraction network enhanced by physical knowledge is used to process the standardized multi-physics field feature data set through the multi-field coupling feature extraction network enhanced by physical knowledge to obtain a multi-physics field coupling feature vector; A physical consistency graph neural network is used to process the multi-physics field coupling eigenvector and the temperature sensor network topology through the physical consistency graph neural network to obtain a temperature sensor health status assessment result, wherein the physical consistency constraint function includes the first law of thermodynamics, the heat conduction equation, and the electromagnetic induction law physical model; A multi-physics field driven sensor fault prediction model is used to process the temperature sensor health status assessment results and historical fault data through the multi-physics field driven sensor fault prediction model to obtain a sensor fault prediction report. The fault prediction model uses an attention-enhanced long short-term memory network combined with a physical field attention mechanism; A physical model-assisted sensor data compensation network is used to process faulty sensor data and surrounding healthy sensor data through the physical model-assisted sensor data compensation network to obtain data compensation results for the faulty sensor. The data compensation network adopts a graph attention compensation mechanism based on the heat conduction physical model. The physical constraint graph in the physical model-assisted sensor data compensation network is based on the physical law of heat conduction and is calculated as follows: ； in: Represents the subgraph containing the faulty sensor and related sensors; Represents the temporal characteristics of related sensors. Related sensors are selected based on three dimensions: physical distance correlation, data correlation, and heat flow propagation path correlation; Represents the parameters of the heat conduction physical model; represents the physical constraint graph attention compensation function; represents the data compensation result of the i-th faulty sensor.

Images