Plate heat exchanger fault diagnosis method based on blind source separation and SVM fusion

Abstract

The invention discloses a plate heat exchanger fault diagnosis method based on blind source separation and SVM fusion, and relates to the technical field of industrial equipment intelligent operation and maintenance and fault prediction, and the method comprises the steps: extracting a pure fault feature source signal from a strong noise environment through a multi-channel vibration sensing and independent component analysis technology; combining multi-domain features such as power spectrum density and envelope entropy with process parameters; a radial basis kernel and Sigmoid kernel compounded support vector machine model is adopted, a complex nonlinear relation between fault features and types is accurately fitted, and high-precision classification and real-time state evaluation of various faults such as blockage, leakage and scaling are achieved; according to the method, the early fault recognition capability and diagnosis robustness under complex working conditions are remarkably improved, and an effective solution is provided for intelligent operation and maintenance of the plate heat exchanger.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault prediction technology for industrial equipment, and in particular to a fault diagnosis method for plate heat exchangers based on blind source separation and SVM fusion. Background Technology

[0002] Currently, plate heat exchangers are key equipment in industrial processes such as petroleum, chemical, and nuclear power. Their fault diagnosis primarily relies on traditional methods such as thermal parameter monitoring, periodic disassembly and inspection, and threshold judgment based on expert experience. These traditional methods have significant limitations in practice: thermal parameter monitoring is slow to respond to early, minor faults, often triggering alarms only when the fault has progressed to the middle or late stages and caused significant performance degradation; periodic disassembly and inspection requires shutdown, affecting production continuity and failing to achieve real-time status awareness; and threshold settings based on human experience are heavily influenced by subjective factors, making it difficult to accurately identify complex states of concurrent or multiple faults. Overall, existing technologies are insufficient in terms of real-time performance, early warning capabilities, and objective accuracy, failing to meet the urgent needs of modern industry for high equipment reliability, long-term operation, and predictive maintenance.

[0003] Against this backdrop, the operation and maintenance management of plate heat exchangers will face severe challenges. The direct result is that early, subtle fault characteristics cannot be effectively identified due to being drowned out by strong environmental and fluid noise, allowing equipment to continue operating with defects until the fault worsens and triggers significant performance degradation or a chain reaction. This not only leads to unplanned downtime, causing production interruptions and economic losses, but may also cause safety accidents due to the failure to address leaks and blockages in a timely manner. Furthermore, relying on traditional methods makes it difficult to accurately distinguish and classify different fault modes such as blockages, leaks, and scaling, easily leading to misjudgments and missed reports. This results in a lack of targeted maintenance decisions, ultimately restricting equipment operating efficiency and lifespan, and hindering the construction of intelligent operation and maintenance systems. Summary of the Invention

[0004] This application provides a plate heat exchanger fault diagnosis method based on blind source separation and SVM fusion, which solves the problem of difficulty in identifying early faults in plate heat exchangers under strong noise in the prior art. Through blind source separation and SVM fusion technology, it achieves highly robust and accurate intelligent diagnosis and early warning.

[0005] This application provides a plate heat exchanger fault diagnosis based on blind source separation and SVM fusion, including:

[0006] S1, under standard operating conditions of the plate heat exchanger, the equipment status data is collected synchronously by multi-channel vibration sensors and process parameter sensors to obtain multi-dimensional time-series vibration signals that reflect the operating status of the equipment;

[0007] S2. After processing the multidimensional time series vibration signal, the independent component analysis algorithm is used to separate blind sources to extract pure intrinsic vibration fault source signals. Multi-domain features are extracted from these signals and fused with process parameters in a multi-modal manner to construct a high-dimensional feature vector.

[0008] S3. Construct a support vector machine classification model that combines radial basis kernel and sigmoid kernel, train it with labeled high-dimensional feature vectors, fit the complex nonlinear relationship between fault features and fault types, and obtain the optimal fault classification accuracy.

[0009] S4. After processing the online monitoring data in the aforementioned steps, input it into the trained composite kernel support vector machine diagnostic model, output the fault type identification results and diagnostic confidence in real time, and display the status and provide multi-level early warnings through a visual interface.

[0010] Preferably, the construction of the high-dimensional feature vector includes: purifying the preprocessed multi-channel vibration signal; using an independent component analysis algorithm to find a demixing matrix, thereby estimating statistically independent source signals from the mixed observation signals, thus separating the intrinsic vibration source signal characterizing the equipment fault from the interference and obtaining a pure fault source signal; extracting features from the pure fault source signal obtained after blind source separation to characterize the fault state; performing multimodal fusion of all multidimensional features extracted from the vibration signal with synchronously acquired equipment operating process parameters; and concatenating all feature values ​​and process parameter values ​​in sequence to jointly constitute a high-dimensional feature vector.

[0011] Preferably, obtaining the optimal fault classification accuracy specifically involves: constructing a support vector machine classification model, which uses a composite kernel function consisting of a linear combination of a radial basis function kernel function and a sigmoid kernel function to fit the complex nonlinear classification boundary between fault features and fault types; using a dataset composed of high-dimensional feature vectors and their corresponding fault labels to train the composite kernel support vector machine model; evaluating the model performance using K-fold cross-validation and optimizing hyperparameters within a preset parameter space using a grid search method to determine the optimal hyperparameter combination; and iteratively solving the Lagrange multipliers using a serialized minimum optimization algorithm during the model solution process to finally obtain a fault diagnosis model with optimal classification accuracy.

[0012] Preferably, the visualization interface displays the status and provides multi-level early warnings, specifically as follows: Vibration signals and process parameters acquired through online monitoring are preprocessed and input into a trained composite kernel support vector machine diagnostic model for real-time status identification; post-processing technology based on Platt calibration is used to map the model's decision function values ​​to diagnostic confidence levels; the fault type identification results and diagnostic confidence levels are visualized on the monitoring terminal; a multi-level early warning mechanism is activated based on preset confidence thresholds: a definite early warning is triggered when the confidence level is higher than the first threshold, an attention warning is triggered when the confidence level is between the first and second thresholds, and a status uncertainty is determined when the confidence level is lower than the second threshold.

[0013] Preferably, the fitted complex nonlinear relationship between fault features and fault type includes:

[0014] S31, through multi-physics coupling modeling integrating fluid dynamics, heat transfer and solid mechanics, and parameterizing fault variables of fault characteristics, constructs a virtual digital twin that simulates the vibration, temperature and pressure response of a plate heat exchanger under different fault states.

[0015] S32 utilizes real-time monitoring data as observational evidence and employs a Bayesian inference algorithm to probabilistically estimate and continuously calibrate the hidden state parameters in the digital twin, thereby dynamically synchronizing the virtual model with the physical entity.

[0016] S33, based on the calibrated twin state, sets the failure development law and extrapolates the performance degradation process. Combining Monte Carlo simulation and failure threshold definition, it provides probabilistic prediction and uncertainty assessment of the remaining useful life of plate heat exchangers.

[0017] Preferably, the multiphysics coupling modeling includes: a fluid dynamics model that simulates the medium flow state and pressure field in the flow channel by solving the Navier-Stokes equations describing momentum conservation; a heat transfer model that calculates the temperature field distribution and heat transfer efficiency by solving the governing equations based on energy conservation; and a solid mechanics model that calculates the vibration modes and dynamic response of the solid body under fluid pressure and thermal stress loads by solving the structural dynamics equations.

[0018] Preferably, the virtual digital twin is specifically: the physical variables characterizing the fault are embedded as key parameters into the control equations of fluid dynamics, heat transfer, and solid mechanics; and a virtual digital twin that can dynamically reflect the vibration, temperature, and pressure response of equipment during different fault states and their evolution is generated by iteratively solving the coupled partial differential equations after parameterization.

[0019] Preferably, the probabilistic prediction of the remaining useful life of the plate heat exchanger specifically involves: obtaining the optimal estimate and uncertainty information of key fault state parameters in a digital twin calibrated by Bayesian inference, as the initial conditions for fault evolution simulation; setting a model for the development law of key fault parameters over time; using the digital twin to perform forward extrapolation of fault evolution based on the set parameter evolution law, simulating the performance index change trend under future time series, and judging whether the equipment has failed based on a preset performance failure threshold; using the Monte Carlo simulation method to handle parameter uncertainty, randomly sampling from the probability distribution of key fault parameters, calculating the parameter evolution trajectory for each sample and driving the digital twin to simulate the corresponding performance index decay process, recording the future time point corresponding to the performance index decaying to the failure threshold, and calculating the difference between this time point and the current time as the remaining useful life of this simulation.

[0020] Preferably, the step of dynamically synchronizing the virtual model with the physical entity includes:

[0021] S321 defines the entire fluid loop, including the target plate heat exchanger and its key associated equipment, and abstracts the entire fluid system into a topological map.

[0022] S322 is designed with devices as nodes and connections as edges. By defining the message passing and state update mechanism between nodes and using historical system fault data for training, it simulates the timing and intensity evolution of fault characteristics in fluid propagation.

[0023] S323 initializes the real-time status of each device as the characteristics of graph network nodes, runs the network to perform multi-step forward simulation, simulates the fault propagation process, and finally outputs a collaborative remaining useful life and fault propagation path risk heatmap that takes into account the impact between devices.

[0024] Preferably, the abstraction of the entire fluid system into a topological graph specifically involves defining each independent device in the fluid as a graph node, with each node associated with all states and parameters of its corresponding device twin, and using directed edges to connect these nodes. Each edge represents the actual physical connection between devices and the direction of the causal interaction it carries. Each edge is assigned a set of attributes that define its physical characteristics, as well as coupling equations describing how the states of adjacent devices are mutually constrained through this connection.

[0025] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0026] 1. Identify the core defects of existing technologies and their causes:

[0027] Existing fault diagnosis methods for plate heat exchangers mainly rely on traditional thermal parameter analysis (such as temperature and pressure monitoring), periodic disassembly and inspection, or threshold judgment based on expert experience. These methods have inherent drawbacks such as perception lag, inability to provide early warnings, and the necessity of shutdown for implementation. Furthermore, experience-based judgments are heavily influenced by subjective factors, making it difficult to accurately identify concurrent faults and their evolution. The core deficiency lies in the fact that under complex operating conditions, the weak characteristic signals caused by faults are completely obscured by background noise, resulting in an extremely low signal-to-noise ratio. The acquired vibration signals are a linear mixture of multiple sources, including equipment vibration, fluid noise, and environmental interference. Conventional time-frequency analysis of this mixed signal alone is insufficient to directly and stably separate and identify the essential characteristics representing specific fault modes, leading to missed or misjudged early faults. Current technologies are limited to single signal processing or simple classification strategies, failing to fundamentally address the challenge of insufficient accuracy in fault feature extraction and classification under noise interference.

[0028] 2. Propose new technological approaches

[0029] Based on the understanding of the above problems, this invention addresses the inherent deficiency of traditional fault diagnosis methods in effectively identifying early, subtle faults under strong noise interference. This method first uses multi-channel vibration sensing and independent component analysis (ICA) to adaptively separate noise and fault source signals in mixed observation signals, extracting the pure intrinsic vibration components of the fault. Based on this, multi-domain features, including power spectral density and envelope entropy, are extracted from the separated signals and fused with process parameters in a multi-modal manner to construct a high-dimensional feature vector comprehensively representing the equipment state. Finally, a support vector machine model combining radial basis function (RBF) kernel and sigmoid kernel is used to accurately fit the complex nonlinear mapping relationship between fault features and categories, achieving precise classification of multiple fault types. This technical approach transforms the inherent physical correlation between internal fault evolution, changes in system dynamics, and external vibration response into a complete digital processing chain, constructing an enhanced analysis model from signal perception to intelligent diagnosis.

[0030] 3. Overcoming technical obstacles in new pathways:

[0031] Vibration signals from the equipment are synchronously acquired using multi-channel vibration sensors, and blind source separation is performed using the Independent Component Analysis (ICA) algorithm to effectively eliminate environmental and fluid noise interference and extract pure fault source signals. Subsequently, multi-domain features such as power spectral density and envelope entropy are extracted from the separated signals and fused with process parameters to construct a high-dimensional feature vector. Finally, an SVM classifier combining radial basis function kernels and sigmoid kernels is used to train and identify the features, achieving high-precision classification of various faults such as blockage, leakage, and scaling.

[0032] 4. Provide the optimal solution.

[0033] The ICA-SVM collaborative framework enables robust and accurate real-time identification of typical faults in plate heat exchangers.

[0034] This method can effectively suppress noise interference in noisy industrial environments, significantly improve the accuracy of identifying early and subtle faults such as blockage, leakage, and scaling, and has strong generalization ability, supporting multi-level early warning and visual decision-making. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the process for fault diagnosis of plate heat exchangers based on blind source separation and SVM fusion according to the present invention. Detailed Implementation

[0036] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terminology used herein includes any and all combinations of one or more of the associated listed items.

[0038] Example 1: Currently, plate heat exchangers often fail to provide early warnings due to noise masking early fault characteristics. Traditional methods can only issue alarms in the middle and late stages of a fault based on hysteresis parameters, which can easily lead to irreversible damage and unplanned downtime. The operation and maintenance management of plate heat exchangers will revert to a traditional model relying on hysteresis parameters and manual judgment, failing to meet the high standards of reliability, safety, and intelligence required by modern industry. Therefore, this application proposes a plate heat exchanger fault diagnosis method based on blind source separation and SVM fusion.

[0039] like Figure 1 As shown, a fault diagnosis method for plate heat exchangers based on blind source separation and SVM fusion includes the following steps:

[0040] S1, under standard operating conditions of the plate heat exchanger, acquires equipment status data synchronously through multi-channel vibration sensors and process parameter sensors to obtain multi-dimensional time-series vibration signals reflecting the equipment's operating status.

[0041] Specifically, when the plate heat exchanger is under various standard operating conditions, including normal, blockage, leakage and scaling, multi-channel vibration signal sensors arranged on the surface of the equipment shell and key points of the pipeline are used to synchronously collect vibration signals at the same sampling frequency.

[0042] Key process parameters such as the working temperature, pressure, and flow rate of the medium are collected synchronously. The acquired raw vibration signal is preprocessed, including: eliminating linear or slowly changing trend terms in the signal through linear fitting to complete detrending processing; and adjusting the signal amplitude to a uniform scale (such as the range of [0, 1]) using the minimum-maximum method to complete normalization processing and eliminate the influence of dimensions.

[0043] Based on finite element simulation results, historical maintenance records, or domain expert experience, each preprocessed vibration data sample is labeled with a fault state, categorized as normal, blocked, leaking, or scaled. The training sample set is artificially expanded by adding Gaussian white noise to the one-dimensional vibration signal data or by applying small-amplitude random variations such as amplitude scaling or time stretching.

[0044] S2. After preprocessing the multidimensional time series vibration signal, the independent component analysis algorithm is used to separate blind sources to extract pure intrinsic vibration fault source signals. Multi-domain features are extracted from these signals and fused with process parameters in a multi-modal manner to construct a high-dimensional feature vector.

[0045] Specifically, the preprocessed multi-channel vibration signals are purified. Independent component analysis (ICA) is used to find a demixing matrix that allows statistically independent source signals to be estimated from the mixed observation signals. This separates the intrinsic vibration source signals characterizing equipment faults from interference such as environmental noise and fluid pulsation, resulting in pure fault source signals.

[0046] From the clean fault source signal obtained after blind source separation, features from multiple domains are extracted to characterize the fault state. In the time domain, higher-order statistics of the signal, such as kurtosis, are calculated; in the frequency domain, the Welch method is used to estimate the power spectral density of the signal; simultaneously, nonlinear features are extracted, and the signal envelope is obtained through Hilbert transform, and its envelope entropy is calculated.

[0047] All multi-dimensional features extracted from the vibration signal are fused with synchronously acquired equipment operating process parameters (such as temperature, pressure, and flow rate) using a multimodal method. All these feature values ​​and process parameter values ​​are then sequentially concatenated to form a high-dimensional feature vector that comprehensively describes the equipment's health status.

[0048] S3. Construct a support vector machine classification model that combines radial basis function kernel and sigmoid kernel. Train the model using labeled high-dimensional feature vectors to fit the complex nonlinear relationship between fault features and fault types, and obtain the optimal fault classification accuracy.

[0049] Specifically, a support vector machine (SVM) classification model is constructed. This model employs a composite kernel function, which is a linear combination of a radial basis function (RBF) kernel and a sigmoid kernel function. Its mathematical expression is as follows:

[0050]

[0051] in, and Represents two high-dimensional feature vector samples. It is the fusion coefficient that controls the weights of the two kernel functions (the value range is usually [0,1]). It is the bandwidth parameter of the radial basis kernel function. and These are the parameters of the Sigmoid kernel function. This composite kernel aims to combine the local learning capability of the radial basis function kernel with the global generalization performance of the Sigmoid kernel to accurately fit the complex nonlinear classification boundary between fault features and fault types. The model's decision function is... ,in These are sample labels. is a Lagrange multiplier, b is a bias term, and its output corresponds to discrete fault categories such as normal, blockage, leakage, and scaling.

[0052] The dataset consisting of the obtained high-dimensional feature vectors and their corresponding fault labels is used to train the composite kernel support vector machine model. K-fold cross-validation is used to evaluate the model's performance under different hyperparameter combinations. Simultaneously, a grid search method is employed within a predefined parameter space (e.g., the penalty factor C can range from [1e-3, 1e3], and the fusion coefficient...). The step size of the parameters can be 0.1) to perform system optimization to determine the optimal combination of hyperparameters that maximizes classification accuracy. During the model solution process, a serialized minimum optimization algorithm is used to efficiently iteratively solve for the Lagrange multipliers. Ultimately, a fault diagnosis model with optimal classification accuracy is obtained.

[0053] S4. After processing the online monitoring data in the aforementioned steps, input it into the trained composite kernel support vector machine diagnostic model, output the fault type identification results and diagnostic confidence in real time, and display the status and provide multi-level early warnings through a visual interface.

[0054] Specifically, the vibration signals and process parameters acquired through online monitoring are preprocessed and then used for real-time state identification through a trained composite kernel support vector machine model. This process first extracts high-dimensional feature vectors in real time... Substitute the decision function into the model Perform calculations, where They are support vectors. It is its category label, These are the corresponding Lagrange multipliers, and b is the bias term. It is a composite kernel function. The direct calculation result of this function is used to determine the fault type.

[0055] A post-processing technique based on Platt calibration is employed to map the decision function values ​​to a probability estimate. By optimizing the maximum likelihood function, the parameters A and B of the sigmoid function are obtained, and then the probability that a sample belongs to class k is calculated. This probability value is the diagnostic confidence level.

[0056] The fault type and confidence level are visualized on the monitoring terminal, for example, as real-time data curves, fault status indicators, and confidence level percentages. The system uses a preset confidence level threshold. (First threshold) and (Second threshold) (its specific value is determined during model training through grid search and cross-validation, aiming to balance accuracy and false positive rate) triggers multi-level alerts: when the confidence level is higher than... A certain warning is triggered when the time is right; a warning is triggered when the time is between the two; and a warning is triggered when the time is below the specified level. If the state is uncertain, then it is determined to be uncertain.

[0057] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0058] By effectively eliminating noise interference under complex operating conditions through blind source separation technology, pure fault features are extracted, and a high-precision diagnostic model is constructed by combining a composite kernel support vector machine. Its advantage lies in its ability to accurately identify early and subtle faults in plate heat exchangers (such as blockage, leakage, and scaling), achieving accurate classification and real-time status assessment of multiple types of faults, and significantly improving the robustness, accuracy, and intelligent operation and maintenance level of diagnosis.

[0059] Example 2: The intelligent diagnostic method provided in Example 1 effectively solves the problem of detecting and classifying early faults in plate heat exchangers under complex operating conditions. However, in the field of predictive maintenance for high-end equipment, simply determining whether a fault exists at present is insufficient. Maintenance personnel urgently need to understand the evolution trend of faults and the remaining useful life of the equipment in order to make forward-looking maintenance decisions and resource allocation. Existing technologies lack deep integration of dynamic modeling and physical mechanisms of fault development processes, and cannot transform a single fault identification result into a quantitative prediction of the long-term operating status of the equipment. This limits the evolution of maintenance strategies from reactive and periodic maintenance to condition-based maintenance.

[0060] In some embodiments, to fit the complex nonlinear relationship between fault characteristics and fault type, step S3 further includes:

[0061] S31 constructs a virtual digital twin of a plate heat exchanger under different fault conditions by integrating multi-physics coupling modeling of fluid dynamics, heat transfer and solid mechanics, and parameterizing fault variables that incorporate fault characteristics.

[0062] Specifically, in the process of constructing a high-fidelity multiphysics digital twin, it is necessary to accurately input all geometric dimensions, material physical properties, and operating boundary conditions of the physical plate heat exchanger.

[0063] The numerical computational model based on multiphysics coupling comprises a fluid dynamics model (simulating the fluid flow state and pressure field within the channel by solving the Navier-Stokes equations describing momentum conservation), a heat transfer model (calculating the temperature field distribution and heat transfer efficiency by solving the governing equations based on energy conservation), and a solid mechanics model (calculating the vibration modes and dynamic response of the solid under fluid pressure and thermal stress loads by solving the structural dynamics equations). To achieve interaction between the physical fields, bidirectional coupling boundary conditions need to be set at the physical interface between the fluid and solid (i.e., the fluid-solid interface, mainly including the two sides of the plate and the surface of the gasket in contact with the medium). Specifically, within each solution time step, the interface pressure and temperature data calculated by the fluid model are transferred as loads to the solid mechanics model; simultaneously, the structural deformation or displacement data calculated by the solid mechanics model is fed back to the fluid and heat transfer models to update the channel geometry and computational mesh. Through this real-time bidirectional transfer of pressure, temperature, and displacement data, a fully coupled iterative solution between fluid flow, heat transfer, and structural deformation is achieved until the calculation results converge.

[0064] Physical variables characterizing the fault are incorporated as key parameters into the governing equations of fluid dynamics, heat transfer, and solid mechanics. For example, fouling thickness is introduced into heat transfer and flow resistance calculations to simulate its impact on heat transfer efficiency and pressure drop; leakage area is incorporated as a source term or a modified boundary condition into the fluid mass conservation equation to simulate flow loss and local flow field disturbance caused by leakage. By iteratively solving the parameterized coupled partial differential equations, a virtual digital twin is generated that dynamically reflects the equipment vibration, temperature, and pressure responses during different fault states and their evolution.

[0065] S32 utilizes real-time monitoring data as observational evidence and employs a Bayesian inference algorithm to probabilistically estimate and continuously calibrate the hidden state parameters in the digital twin, thereby dynamically synchronizing the virtual model with the physical entity.

[0066] S33, based on the calibrated twin state, sets the failure development law and extrapolates the performance degradation process. Combining Monte Carlo simulation and failure threshold definition, it provides probabilistic prediction and uncertainty assessment of the remaining useful life of plate heat exchangers.

[0067] Specifically, the optimal estimates of current fault state parameters and their uncertainties are obtained from the digital twin that has completed Bayesian inference calibration, serving as initial conditions for fault evolution simulation. Examples include the mean and variance of scale thickness and the probability distribution characteristics of leakage area. Development models for key fault parameters over time are then established. For instance, assuming scale thickness grows linearly, a probability distribution based on historical data or physical mechanisms is set for its growth rate. Similarly, corresponding evolutionary patterns are established for other fault parameters.

[0068] Digital twins are used for forward extrapolation of fault evolution. Based on predefined parameter evolution patterns, the digital twin, through its built-in multiphysics coupled solver, simulates the changing trends of performance indicators over future time series. Simultaneously, a performance failure threshold is defined as a criterion for end-of-life assessment. For example, the failure threshold for heat exchange efficiency is set to... ,in This represents the efficiency value of the equipment under its initial healthy state. When the simulated performance index degrades to this threshold, the equipment is considered to have failed.

[0069] Monte Carlo simulation is employed to handle parameter uncertainties. A large number of random samples are taken from the probability distribution of key failure parameters, with each sample representing a possible failure development scenario. For each sample, its complete parameter evolution trajectory is calculated, and a digital twin is used to simulate the corresponding performance index decay process. The future time point corresponding to the performance index decaying to a preset failure threshold is recorded, and the difference between this time point and the current time is calculated, which represents the remaining usable lifetime predicted by this simulation.

[0070] Repeat the above sampling simulation process a sufficient number of times to obtain a large set of predicted remaining useful life values. Based on this set, perform statistical analysis, calculate its mean as a point estimate of the remaining useful life, calculate its variance to measure the uncertainty of the prediction, and determine the confidence interval at a specific confidence level. Finally, output the probabilistic prediction of the remaining useful life, including the point estimate, confidence interval, and overall distribution pattern.

[0071] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0072] By introducing a combination of multiphysics digital twins, Bayesian inference, and Monte Carlo simulation, a leap from fault identification to remaining useful life prediction has been achieved. This solution can dynamically extrapolate the performance degradation process of plate heat exchangers, outputting probabilistic predictions of remaining life and their uncertainty ranges. This elevates maintenance decisions from post-event diagnosis to pre-event prediction, significantly enhancing the predictability of equipment condition management. It provides crucial quantitative evidence for achieving condition-based predictive maintenance, optimizing maintenance strategies, and avoiding unplanned downtime.

[0073] Example 3: In Example 2, when the failure prognosis of the plate heat exchanger is only performed by state inference and life prediction through its own digital twin, the plate heat exchanger, as a core component of the fluid system, is closely coupled with equipment such as pumps and valves. Its failure evolution will trigger a systemic chain effect through pipeline connections. The single equipment prognosis model cannot capture the dynamic interaction between equipment, resulting in the prognosis results being limited to a local perspective and making it difficult to assess the impact of failure propagation on the overall reliability of the fluid system.

[0074] In some embodiments, step S32, which dynamically synchronizes the virtual model with the physical entity, further includes:

[0075] S321 defines the entire fluid loop, including the target plate heat exchanger and its key associated equipment, and abstracts the entire fluid system into a topological map.

[0076] In this topology graph, each independent device in the fluid is defined as a node; for example, node A represents a plate heat exchanger, and node B represents a feed pump. Each node is associated with all the states and parameters of its corresponding device twin. These nodes are connected by directed edges, each representing the actual physical connection between devices (such as a pipe section) and the direction of the causal interaction it carries. Each edge is assigned a set of properties defining its physical characteristics, including the connection type (series or parallel), the key physical quantities transmitted along the connection (such as flow rate, pressure, and temperature), and the coupling equations describing how the states of adjacent devices are mutually constrained through this connection, such as the flow continuity equation and the pressure drop calculation equation.

[0077] S322 is designed as a graph neural network with devices as nodes and connections as edges. By defining the message passing and state update mechanism between nodes and using historical system fault data for training, it simulates the timing and intensity evolution of fault characteristics in fluid propagation.

[0078] Specifically, based on the constructed system topology graph, each device node is instantiated as a computational node in the graph network. The prognostic feature vector output by each device's digital twin at the initial moment is concatenated with its real-time operating parameter vector. The concatenated vector is then passed through a trainable linear transformation layer to generate the initial hidden state vector of each node, which represents the initial health and operating status of the device.

[0079] Within each simulation time step, for each directed edge in the graph, a trainable multilayer perceptron function is used to calculate and generate the corresponding edge message vector based on the hidden states of the source node, the target node, and the edge's attribute feature vector. For each target node, an attention mechanism is used to calculate the weights of all message vectors from its surrounding neighbor nodes, and these weighted sums are performed to obtain the node's aggregated message vector. Each node inputs its previous time step hidden state and its current aggregated message vector into a gated recurrent unit to calculate and output its updated hidden state for the current time step. This process iterates for a preset number of time steps, ultimately ensuring that the hidden states of each node encode the impact of the fault propagation and evolution within the system.

[0080] Using historical plate heat exchanger running sequence data, network nodes are initialized with the device states at the start of the sequence. The network is run for forward propagation to simulate node state sequences at a series of time steps. A composite loss function is defined, which simultaneously calculates the error between the network's predicted intermediate node state sequences and the actual health indicator sequences, as well as the error between the network's predicted remaining device lifetime at the final moment and the actual lifetime label. Through backpropagation, a gradient descent optimizer is used to adjust all trainable parameters in the network (including the linear transformation layer, message generation multilayer perceptron, attention mechanism, and weights and biases in the gated recurrent unit) to minimize the composite loss function, thereby enabling the network to learn and master the propagation dynamics of faults within the system topology.

[0081] S323 initializes the real-time status of each device as the characteristics of graph network nodes, runs the network to perform multi-step forward simulation, simulates the fault propagation process, and finally outputs a collaborative remaining useful life and fault propagation path risk heatmap that takes into account the impact between devices.

[0082] Specifically, the real-time prognostic feature vectors output by the digital twins of each device are concatenated and fused with their corresponding real-time operating parameter vectors. Then, the fused vectors are transformed and dimensionally aligned by a linear projection layer with trainable parameters to generate the initial hidden state vectors of each graph neural network node, thereby completing the embedding mapping from physical device states to network node features.

[0083] The trained network parameters are loaded, and starting from the initialized node states, iterative computation is performed across multiple time steps. Within each time step, three operations are performed sequentially: message generation, message aggregation, and state update. For each directed edge in the graph, a parameterized multilayer perceptron is used to calculate the corresponding edge message vector based on the current hidden state vectors of the source and target nodes at both ends of the edge, as well as the edge's intrinsic attribute vector. For each target node, an attention mechanism is used to weight the message vectors from all its incoming neighbor nodes, and the aggregated message vector for that node is obtained by weighted summation. In the state update operation, each node inputs its hidden state vector from the previous time step along with the current aggregated message vector into a gated recurrent unit, outputting its updated hidden state vector for the current time step. This iterative process continues until a preset inference step size is reached, ultimately obtaining the final hidden state vectors of all nodes. These state vectors encode the cumulative dynamic impact of faults propagating through multiple steps in the system topology.

[0084] Using the node's final hidden state vector obtained in the previous step as input, a fully connected linear regression layer is used to calculate the collaborative remaining useful life of each device after considering the impact of fault interactions within the system. Simultaneously, the dynamic association weight sequence generated between all nodes through the attention mechanism during the entire simulation process is extracted. Through statistical aggregation over time (such as calculating the mean or maximum value), an adjacency matrix characterizing the fault impact intensity between any two devices within the system is constructed. Based on this matrix, a fault propagation path risk heatmap is generated, intuitively revealing the critical propagation paths and their risk level distribution.

[0085] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0086] By abstracting the entire fluid system into a topological map and constructing a fault propagation model based on graph neural networks, a significant breakthrough has been achieved, moving from single-device diagnosis to system-level prognostic capabilities. This scheme can accurately simulate the propagation path and intensity evolution of faults among interconnected devices, outputting a collaborative remaining useful life and a visualized risk heatmap that considers the dynamic coupling effects between devices. This significantly improves the predictive ability of systemic fault chains and provides crucial technical support for the overall reliability assessment and collaborative maintenance decision-making of complex industrial systems.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A plate heat exchanger fault diagnosis method based on blind source separation and SVM fusion, characterized in that, include: S1, under standard operating conditions of the plate heat exchanger, the equipment status data is collected synchronously by multi-channel vibration sensors and process parameter sensors to obtain multi-dimensional time-series vibration signals that reflect the operating status of the equipment; S2. After preprocessing the multidimensional time series vibration signal, the independent component analysis algorithm is used to separate blind sources to extract pure intrinsic vibration fault source signals. Multi-domain features are extracted from these signals and fused with process parameters in a multi-modal manner to construct a high-dimensional feature vector. S3. Construct a support vector machine classification model that combines radial basis kernel and sigmoid kernel, train it with labeled high-dimensional feature vectors, fit the complex nonlinear relationship between fault features and fault types, and obtain the optimal fault classification accuracy. S4. After processing the online monitoring data in the aforementioned steps, input it into the trained composite kernel support vector machine diagnostic model, output the fault type identification results and diagnostic confidence in real time, and display the status and provide multi-level early warnings through a visual interface.

2. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 1, characterized in that, The construction of the high-dimensional feature vector includes: purifying the pre-processed multi-channel vibration signal; using independent component analysis (ICA) to find an unmixing matrix, thereby estimating statistically independent source signals from the mixed observation signals, thus separating the intrinsic vibration source signal characterizing the equipment fault from the interference and obtaining a pure fault source signal; extracting features from the pure fault source signal obtained after blind source separation to characterize the fault state; performing multimodal fusion of all multidimensional features extracted from the vibration signal with synchronously acquired equipment operating process parameters; and concatenating all feature values ​​and process parameter values ​​in sequence to form a high-dimensional feature vector.

3. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 1, characterized in that, To achieve optimal fault classification accuracy, the following steps are taken: A support vector machine (SVM) classification model is constructed. This model employs a composite kernel function, consisting of a linear combination of a radial basis function (RBF) kernel and a sigmoid kernel function, to fit the complex nonlinear classification boundary between fault features and fault types. A dataset composed of high-dimensional feature vectors and their corresponding fault labels is used to train the composite kernel SVM model. K-fold cross-validation is used to evaluate the model's performance, and a grid search method is employed to optimize hyperparameters within a predefined parameter space to determine the optimal hyperparameter combination. During the model solution process, a serialized minimum optimization algorithm is used to iteratively solve for the Lagrange multipliers, ultimately obtaining a fault diagnosis model with optimal classification accuracy.

4. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 1, characterized in that, The visualization interface displays the status and provides multi-level early warnings. Specifically, the vibration signals and process parameters acquired by online monitoring are preprocessed and then input into the trained composite kernel support vector machine diagnostic model for real-time status identification. Post-processing technology based on Platt calibration is used to map the model's decision function values ​​to diagnostic confidence scores. The fault type identification results and diagnostic confidence scores are then visualized on the monitoring terminal. A multi-level early warning mechanism is activated based on a preset confidence threshold. When the confidence level is higher than the first threshold, a definite early warning is triggered. When the confidence level is between the first and second thresholds, an attention warning is triggered. When the confidence level is lower than the second threshold, the status is determined to be uncertain.

5. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 1, characterized in that, The complex nonlinear relationship between the fitted fault features and the fault type includes: S31, through multi-physics coupling modeling integrating fluid dynamics, heat transfer and solid mechanics, and parameterizing fault variables of fault characteristics, constructs a virtual digital twin that simulates the vibration, temperature and pressure response of a plate heat exchanger under different fault states. S32 utilizes real-time monitoring data as observational evidence and employs a Bayesian inference algorithm to probabilistically estimate and continuously calibrate the hidden state parameters in the digital twin, thereby dynamically synchronizing the virtual model with the physical entity. S33, based on the calibrated twin state, sets the failure development law and extrapolates the performance degradation process. Combining Monte Carlo simulation and failure threshold definition, it provides probabilistic prediction and uncertainty assessment of the remaining useful life of plate heat exchangers.

6. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 5, characterized in that, The multiphysics coupling modeling includes: a fluid dynamics model that simulates the flow state and pressure field of the medium in the flow channel by solving the Navier-Stokes equations describing momentum conservation; a heat transfer model that calculates the temperature field distribution and heat transfer efficiency by solving the governing equations based on energy conservation; and a solid mechanics model that calculates the vibration modes and dynamic response of the solid body under fluid pressure and thermal stress loads by solving the structural dynamics equations.

7. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 5, characterized in that, The virtual digital twin is specifically defined as follows: physical variables characterizing the fault are incorporated as key parameters into the control equations of fluid dynamics, heat transfer, and solid mechanics; and a virtual digital twin that can dynamically reflect the equipment vibration, temperature, and pressure responses during different fault states and their evolution is generated by iteratively solving a set of parameterized coupled partial differential equations.

8. The plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 5, characterized in that, The probabilistic prediction of the remaining useful life of the plate heat exchanger specifically involves: obtaining the optimal estimate and uncertainty information of key fault state parameters in a digital twin calibrated by Bayesian inference, as the initial conditions for fault evolution simulation; setting a model for the development law of key fault parameters over time; using the digital twin to perform forward extrapolation of fault evolution based on the set parameter evolution law, simulating the performance index change trend under future time series, and judging whether the equipment has failed based on a preset performance failure threshold; using the Monte Carlo simulation method to handle parameter uncertainty, randomly sampling from the probability distribution of key fault parameters, calculating the parameter evolution trajectory for each sample and driving the digital twin to simulate the corresponding performance index decay process, recording the future time point corresponding to the performance index decaying to the failure threshold, and calculating the difference between this time point and the current time as the remaining useful life of this simulation.

9. A plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 8, characterized in that, The dynamic synchronization of the virtual model with the physical entity includes: S321 defines the entire fluid loop, including the target plate heat exchanger and its key associated equipment, and abstracts the entire fluid system into a topological map. S322 is designed with devices as nodes and connections as edges. By defining the message passing and state update mechanism between nodes and using historical system fault data for training, it simulates the timing and intensity evolution of fault characteristics in fluid propagation. S323 initializes the real-time status of each device as the characteristics of graph network nodes, runs the network to perform multi-step forward simulation, simulates the fault propagation process, and finally outputs a heat map of collaborative remaining useful life and fault propagation path risk that takes into account the impact between devices.

10. A plate heat exchanger fault diagnosis based on blind source separation and SVM fusion as described in claim 9, characterized in that, The abstraction of the entire fluid system into a topological graph is as follows: the topological graph defines each independent device in the fluid as a graph node, each node is associated with all the states and parameters of its corresponding device twin, and these nodes are connected by directed edges. Each edge represents the actual physical connection between devices and the direction of the causal action it carries; each edge is assigned a set of attributes that define its physical characteristics, as well as coupling relationship equations that describe how the states of adjacent devices are mutually constrained through this connection.

Images