A method and system for intelligent fault diagnosis of heat exchange equipment based on digital twins

Abstract

This invention discloses an intelligent fault diagnosis method and system for heat exchange equipment based on digital twins, belonging to the field of digital twin intelligent diagnosis technology. It addresses the difficulty in quickly and accurately locating and safely handling concurrent valve jamming and multi-device coupled faults. Design configuration, maintenance history, and real-time operating data are semantically mapped and synchronized with a unified timeframe, and written into a twin database while maintaining virtual-real synchronization. Secondly, flow rate, differential pressure, and temperature sequences are fused with color-infrared images, and a health vector is generated using a dimensionality reduction network for real-time early warning. Upon an early warning, a cloned twin copy of the abnormal equipment and regulating valve is used, injecting valve jamming disturbances into parallel simulation, and identifying concurrent faults through comprehensive similarity analysis. Finally, a knowledge base is invoked to evaluate candidate handling strategies, the optimal solution is selected and executed, and residual monitoring is used for dynamic correction, achieving a closed loop of diagnosis-decision-execution. This method improves equipment reliability and energy utilization.

Description

Technical Field

[0001] This invention relates to the field of digital twin intelligent diagnostic technology, and more specifically, to a method and system for intelligent fault diagnosis of heat exchange equipment based on digital twins.

[0002] In petrochemical plants, tube heat exchangers, reboilers, and heat tracing pumps are subjected to high loads due to steam-material coupling for extended periods. Valve wear, scaling, and throttling fluctuations easily lead to a decline in heat exchange efficiency. Frontline operation often relies on static alarms based on differential pressure and temperature limits, followed by manual troubleshooting. This method cannot distinguish between system-wide power supply fluctuations and individual unit degradation, resulting in frequent false alarms and missed alarms. Data-driven residual monitoring lacks physical constraints and is prone to inaccuracy in the face of flow-thermal oscillations. Offline CFD simulations are computationally intensive, allowing only post-event analysis and failing to meet hourly diagnostic windows. Shared main pipes lead to high equipment coupling, causing traditional single-unit algorithms to fail directly in scenarios with surging correlations. While multivariate statistical methods can capture collaborative behavior, they struggle to provide precise root causes and solutions. Standards such as API682, TEMA, and IEC61511 emphasize hardware redundancy and lack rapid identification methods for concurrent valve-equipment collaborative failures. Although OPC-UA and edge AI have been deployed on-site, the lack of unified semantic mapping and cross-source time synchronization mechanisms leads to severe model drift; large cloud-based models are also limited by bandwidth and latency. The open-loop "stop / open valve + bypass switching" strategy in DCS ignores the energy recovery rate, which can easily lead to secondary thermal shock and steam waste.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for intelligent fault diagnosis of heat exchange equipment based on digital twins, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a preferred embodiment, it includes:

[0007] Collect design configuration, maintenance history and real-time operating condition data from three sources, unify fields and time scales and write them into the digital twin library, and use parallel coupled calculation to keep the virtual and real states synchronized at all times;

[0008] The flow-pressure-temperature sequence is aligned and fused with color-infrared images. After anomalies are removed, the data is fed into a dimensionality reduction network to generate a single health vector in real time and trigger an early warning flag.

[0009] Based on the cloned twins of abnormal equipment sets and control valve combinations, batch injection of jamming disturbances is performed and solved in parallel. The best matching fault combination is quickly identified by matching the field curve.

[0010] Multiple disposal strategies are generated by calling the knowledge base. After evaluating their merits in a twin environment, the optimal solution is selected and executed. The data-model-decision closed loop is achieved by using residual monitoring and iterative correction.

[0011] In a preferred embodiment, a digital twin model is constructed, a shared digital twin library is established and the device's design configuration file, maintenance history and real-time operating condition curves are integrated, and then the physical object's state is virtually mapped.

[0012] In a preferred embodiment, comprehensive perception and monitoring of the device's operating status converts the fused device data within the digital twin model into a unique health vector.

[0013] In a preferred embodiment, a health vector sequence of a continuous time window is retrieved from a digital twin database and dimensionality reduced, and then the equipment degradation status is assessed to quantify the health status and realize anomaly detection and early warning.

[0014] In a preferred embodiment, when an early warning occurs, the equipment operating parameters are retrieved to determine the coupling strength, the coupling information is output, and then possible fault propagation chains are traced and recorded as hidden features.

[0015] The coupled information and hidden features are combined and then input into the diagnostic engine for fault diagnosis.

[0016] In a preferred embodiment, during the steady-state operation of the high-pressure steam header heat exchange network, the flow rate, shell-side pressure difference and outlet temperature data are sampled. Then, all original sampling frames are rewritten and mapped to a unified sampling window to generate a three-component tensor. After preprocessing, a standardized sequence is obtained. Then, the synchronization amplitude and pairwise correlation coefficients of each device are solved.

[0017] Calculate the synchronization amplitude threshold and correlation threshold, compare the thresholds of newly arrived sequences within the unified sampling window, and determine the power supply resonance synchronization anomaly set.

[0018] In a preferred embodiment, a concurrent fault hypothesis is generated based on the power supply resonance synchronization anomaly set and the main pipe inlet regulating valve entity, and a virtual twin of the concurrent fault hypothesis is cloned.

[0019] A throttling angle disturbance is applied to the main pipe inlet regulating valve body and injected into a three-dimensional multi-field solver. The three-dimensional multi-field solver dynamically subdivides the iteration steps according to the local pressure gradient and performs block sparse solution, controlling the calculation error to be lower than the stability constraint when compressing the total amount of calculation.

[0020] The virtual twins output the simulation response curves in real time and upload them to the simulation results buffer. Then, a comprehensive similarity is generated and sorted in ascending order. A matching threshold is set to filter out the concurrent valve failure hypothesis with the lowest comprehensive similarity and lock the corresponding virtual twin for root cause analysis.

[0021] In a preferred embodiment, the locked valve disturbance amplitude, energy loss rate curve, and flow-pressure differential coupling stage increment are extracted, and a joint fitting is performed to output a fault severity scalar, which is then labeled with the fault type. Differentiated collaborative handling schemes are then constructed, and the optimal collaborative handling scheme is selected.

[0022] Calculate the key feature consistency of other similar concurrent valve failure hypotheses, the energy imbalance exposure, and the combined confidence level of the second-order sensitivity to the power supply stability of the main pipe. Rank them in descending order to generate a list. When the concurrent valve failure hypothesis with the lowest overall similarity becomes abnormal, automatically select the concurrent valve failure hypothesis with the highest confidence level to replace the concurrent valve failure hypothesis with the lowest overall similarity.

[0023] In a preferred embodiment, after the root cause is identified, a candidate treatment sequence is automatically generated. The twin simulation performs multi-objective optimization to select the optimal solution, and compiles and issues safety instructions for execution. The residual monitor continuously compares the real-time curves on site with the twin prediction curves. When the residual exceeds the limit, the backup solution is switched. The execution results are written back to the knowledge graph and trigger the model's self-learning.

[0024] In a preferred embodiment, it includes: a data synchronization and fusion module, a state feature generation module, a concurrent fault simulation and screening module, and a collaborative handling closed-loop module, with signal connections between the modules;

[0025] The data synchronization and fusion module is mainly used to collect three sources of data: design configuration, maintenance history and real-time operating conditions. After unifying the fields and time scale, the data is written into the digital twin library and parallel coupled calculation is used to maintain the real and virtual states in real time.

[0026] The status feature generation module is mainly used to align and fuse the flow-pressure-temperature sequence with the color-infrared image, clean up anomalies, and send them into the dimensionality reduction network to generate a single health vector in real time and trigger an early warning flag.

[0027] The concurrent fault simulation screening module is mainly used to clone twin copies of abnormal equipment sets and control valve combinations, inject jamming disturbances in batches for parallel solution, and quickly lock the most matching fault combination by matching field curves.

[0028] The collaborative handling closed-loop module is mainly used to call the knowledge base to generate multiple handling strategies, select the optimal solution after evaluating its merits in the twin environment, and execute it. It also uses residual monitoring and iterative correction to achieve a data-model-decision closed loop.

[0029] The technical effects and advantages of the intelligent fault diagnosis method and system for heat exchange equipment based on digital twins of the present invention are as follows:

[0030] This invention constructs a closed-loop system of "data fusion - health vector - concurrent simulation - collaborative handling" to achieve millisecond-level early warning and minute-level root cause localization. The three-component synchronous amplitude-correlation adaptive threshold significantly reduces system-level power supply wave misjudgments; TwinCopy parallel simulation combined with Markovnikov-DTW composite similarity reduces concurrent valve fault screening time to one-tenth of traditional methods. A multi-objective optimizer automatically selects the handling scheme with the highest energy recovery rate and lowest downtime cost while ensuring SIL safety level; the residual monitor can switch backup strategies online to prevent risk propagation. Plant-level tests show that the false alarm rate is reduced to 0.8%, the total diagnosis-handling time is reduced by 60%, the annual steam saving rate is increased by 3.7%, and maintenance costs are reduced by 28%. Furthermore, relying on incremental learning to continuously optimize thresholds and rules, it achieves self-evolving fault self-healing capabilities. Attached Figure Description

[0031] Figure 1 This invention provides a timing diagram of an intelligent fault diagnosis method and system for heat exchange equipment based on digital twins.

[0032] Figure 2 This is a schematic diagram of the intelligent fault diagnosis method and system structure for heat exchange equipment based on digital twins according to the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example

[0035] This invention discloses an intelligent fault diagnosis method for heat exchange equipment based on digital twins, such as... Figure 1 As shown, it includes:

[0036] Collect design configuration, maintenance history and real-time operating condition data from three sources, unify fields and time scales and write them into the digital twin library, and use parallel coupled calculation to keep the virtual and real states synchronized at all times;

[0037] The flow-pressure-temperature sequence is aligned and fused with color-infrared images. After anomalies are removed, the data is fed into a dimensionality reduction network to generate a single health vector in real time and trigger an early warning flag.

[0038] Based on the cloned twins of abnormal equipment sets and control valve combinations, batch injection of jamming disturbances is performed and solved in parallel. The best matching fault combination is quickly identified by matching the field curve.

[0039] Multiple disposal strategies are generated by calling the knowledge base. After evaluating their merits in a twin environment, the optimal solution is selected and executed. The data-model-decision closed loop is achieved by using residual monitoring and iterative correction.

[0040] First, the equipment's design configuration files, maintenance history, and real-time operating condition curves are used as the sole source of original data. Fields, units, and time scales are precisely aligned using unified semantic mapping rules. Then, the three types of data are aggregated into a shared digital twin database. Subsequently, a parallel physics field solver is used to solve the thermo-fluid-mechanical coupling equations at fixed intervals and update the database records in real time, so that the virtual objects and physical devices maintain fine-grained synchronization across the entire time domain.

[0041] Next, the flow, pressure, and temperature sequences collected by the field sensors, along with the color images and infrared temperature matrix acquired by industrial vision, are injected into the digital twin model.

[0042] Specifically, hardware clock calibration is first used to ensure that the data from the sensor, industrial vision, and infrared temperature matrix are in the same time window. Then, the visible light pixels and thermal image pixels are reprojected onto the digital twin coordinate system using the camera calibration results. Subsequently, the false temperature rise caused by oil or shadows is removed by combining temperature difference threshold with texture gradient. The abnormal temperature area on the surface of the extracted equipment is directly depicted in the virtual mapping, so as to achieve accurate early warning of equipment overheating without generating false alarms.

[0043] After completing the space-temperature fusion within the digital twin model, state features are extracted from the fused tensor. Specifically, an improved normalization-hyperbolic tangent strategy is first used to suppress outlier noise, and then the processed multidimensional data is fed into a stacked autoencoder network for automatic dimensionality reduction.

[0044] After the network converges, the bottleneck layer output is extracted as the unique health vector, and a continuous feature flow is formed by a sliding time window and an exponential smoother and written back to the digital twin database.

[0045] Furthermore, a stacked autoencoder network (SAE) is used to retrieve health vector sequences from a digital twin database over continuous time windows. Nonlinear dimensionality reduction is achieved through layer-by-layer unsupervised pre-training combined with overall fine-tuning and reconstruction. During the training phase, reconstruction error is used as the loss metric, and each hidden layer suppresses collinearity interference through weight sparsity constraints, enabling the network to automatically filter noise and redundancy while preserving key dynamics.

[0046] Since the bottleneck layer vector extracted after convergence not only condenses the joint evolution law of the temperature, pressure and flow ternary curves, but also amplifies the micro-amplitude drift signal, the early degradation characteristics of the equipment can be captured without manual feature selection.

[0047] Next, the hidden features are input into a support vector data descriptor using a Morlet wavelet kernel; an adaptive optimization strategy using a whale-optimized WOA is performed in the joint space of the penalty coefficient and kernel bandwidth, and after iterations, the parameter combination that minimizes the hypersphere volume and the false alarm rate is automatically locked.

[0048] During the online evaluation phase, the distance from each new feature point to the surface of the hypersphere is mapped in real time to the equipment reliability index and compared with a threshold that is updated by an exponentially weighted moving average. Once the index falls below the threshold, an anomaly warning is triggered.

[0049] When an early warning occurs, the operating parameters of each device recorded in the same time window by the digital twin network model are retrieved, and the coupling strength is determined by the correlation coefficient matrix. Then, the possible fault propagation chain is traced by the path search algorithm and recorded as hidden features.

[0050] Then, the hidden features and coupling information are combined and input into the diagnostic engine, and the specific diagnosis is as follows:

[0051] If the feature pattern satisfies the expert rules in the knowledge base, immediately output the corresponding fault mechanism and cause;

[0052] If the feature pattern does not meet the expert rules in the knowledge base, the offline trained artificial intelligence model is invoked to infer the abnormal pattern and locate the fault source using the graph attention mechanism.

[0053] It should be noted that in a heat exchange network sharing a high-pressure steam header, when the header inlet regulating valve intermittently jams and generates unsteady pressure waves for a short period, the steam-side power supply will simultaneously cause highly synchronized sawtooth oscillations in the side flow rate, shell-side pressure differential, and outlet temperature curves of multiple shell-and-tube heat exchangers, reboilers, and heat tracing pumps on the same branch within the same sampling window. This leads the digital twin network model to misjudge the multi-point synchronous oscillations as load fluctuations of a single heat exchanger; it is even more difficult to distinguish whether the system-level power supply wave is caused by valve jamming or a fault in the downstream equipment itself, resulting in missing alarms and incorrect handling. Therefore, in this embodiment,

[0054] During the steady-state operation of the high-pressure steam header heat exchange network, when analyzing the dependencies between equipment using a digital twin network model, the high-precision Coriolis flowmeter, differential pressure transmitter, and platinum resistance temperature probe pre-deployed on each tube heat exchanger, reboiler, and heat tracing pump are first invoked to sequentially perform millisecond-level cyclic sampling of the flow rate F(t), shell-side pressure difference ΔP(t), and outlet temperature T(t).

[0055] Next, to eliminate the impact of cross-device clock drift on synchronization analysis, using the pulse time base T0 output by the main control PLC, all original sampling frames are rewritten with hardware timestamps and mapped to a unified sampling window W1, and a three-component tensor Draw is generated in the digital twin monitoring buffer. Subsequently, an adaptive bandpass filter chain is used to suppress background noise, remove valve position opening and closing spikes, and correct low-drift segmented baselines on the three-component tensor Draw, resulting in a standardized sequence Dstd after voltage regulation. Then, the synchronization amplitude index Ai and the pairwise correlation coefficient ρij of each device are solved in real time using a sliding window method.

[0056] The solution results are written to the statistical queue Qsta immediately, and the threshold updater reads the mean μ and standard deviation σ of the most recent N healthy operating cycles. The synchronization amplitude threshold Athr and the correlation threshold ρthr are dynamically calculated without manual intervention, thus ensuring that the threshold pairs always reflect the latest coordinated fluctuation benchmark of the equipment group.

[0057] Furthermore, the monitoring logic is used to perform threshold comparison on the newly arrived sequences within the unified sampling window W1. If and only if the conditions Ai≥Athr and ρij≥ρthr are met simultaneously, the device subset corresponding to the current sampling window is immediately marked as the power supply resonance synchronization anomaly set Ω, and written into the digital twin database together with the trigger threshold snapshot.

[0058] Subsequently, to avoid false alarms triggered by brief noise, a minimum duration constraint is embedded in the marking process. Only when the power supply resonance synchronization anomaly set Ω remains in the threshold over-limit state for two consecutive sampling windows will it proceed to the next stage of processing.

[0059] Next, based on the physical topology map maintained by the digital twin platform, the power supply resonance synchronization anomaly set Ω and the main pipe inlet regulating valve entity Vin are extended by Cartesian expansion to automatically generate multiple valve jamming-equipment response concurrent fault hypothesis hk, and call the twin copy replication instruction TwinCopy to clone a virtual twin Twk in the distributed computing power pool Cpool for each concurrent fault hypothesis hk that is completely consistent with the field working conditions;

[0060] Meanwhile, to ensure the numerical isomorphism between the virtual twin Twk and the physical body in terms of the coupling relationship of the three fields of heat, fluid and force, the current transient boundary conditions and the input of each driving source are copied simultaneously while copying, fundamentally eliminating the error of secondary modeling.

[0061] Once the virtual twin Twk is ready, a throttling angle perturbation is applied to Vin according to the preset Kase perturbation curve δk, and this perturbation is injected into the three-dimensional multi-field solver as a time-dependent boundary.

[0062] During parallel computing, the three-dimensional multi-field solver dynamically subdivides the iteration steps according to the local pressure gradient using an adaptive step size strategy, and performs block sparse solution of the thermal-fluid-mechanical control equations through GPU thread bundles, thereby compressing the total computation while ensuring that the computational error is lower than the stability constraint εs.

[0063] Furthermore, each virtual twin Twk outputs the simulation response curve Rk(t) in real time, and immediately uploads it to the simulation result buffer Rbuf after the calculation is completed. Normalized time alignment τ is performed through feature alignment to eliminate the difference in response time delay between scenes. Then, the comprehensive similarity Sk is generated by combining the weighted Mahalanobis distance DM and the dynamic time warping distance DDTW.

[0064] After the comprehensive similarity calculation is completed, all comprehensive similarity Sk values ​​are sorted in ascending order at once, and the concurrent valve failure hypothesis h with the smallest Sk value and Sk≤Sthr is selected using the set matching threshold Sthr as the boundary. The corresponding virtual twin Tw is then locked and entered into the root cause analysis process.

[0065] Within the h framework, for each device qj involved in the anomaly, features such as valve disturbance amplitude δj, energy loss rate curve Ej(t), and flow-pressure differential coupling stage increments ΔF and ΔPj(t) are extracted from the virtual twin Tw. A multinomial regression model Rfit with L2 regularization is used to jointly fit δj and Ej(t) to output a fault severity scalar gj. At the same time, a fault type label lj is assigned based on the valve opening-flow nonlinear response curve matching results.

[0066] All fault severity scalars gj are sorted in descending order to form a fault priority matrix G = [g1, g2, ..., gm], and then sent to the knowledge base interface.

[0067] The knowledge base interface retrieves the built-in valve-heat exchange coupling treatment rule set, comprehensively considers the heat load dependence between equipment and the energy supply balance of the pipeline network, and dynamically constructs a candidate set of differentiated collaborative treatment schemes P = {P1, P2, ...}. Subsequently, the twin simulation verifier Simeval is called to perform virtual simulation again in the virtual twin Tw for each scheme Pi, and the candidate schemes are selected for Pareto optimal screening using energy recovery rate ηe, reliability improvement ΔR and treatment time cost τc as three objective vectors fi. Finally, the scheme P* with the highest comprehensive score is selected and output to the DCS control execution layer, while generating a detailed operation instruction sequence and safety checklist.

[0068] Subsequently, to ensure continuous control of diagnostic uncertainty, the set H′ of secondary concurrent hypotheses with the top k similarity scores and Sk > Sthr is saved into the hypothesis monitoring pool. The key feature fit Kh′, energy imbalance exposure Eh′, and second-order sensitivity σh′ to the power supply stability of the main pipe are calculated to synthesize a confidence score πh′. These confidence scores πh′ are then arranged in descending order to generate a list Π. When P* is executed in the field, the residual ε(t) = S(t)new - R(t) is used as the closed-loop driving variable. If ε(t) deviates unacceptably within the set observation window, the highest confidence score h′ in Π is automatically selected to replace h in a new simulation alignment process, and the threshold pairs are updated synchronously to complete self-correction periodically. If ε(t) converges to zero, h and its simulation-reality alignment features are written into the twin knowledge graph and used as incremental samples for subsequent case inference.

[0069] After the most matching valve failure combination is identified as the root cause through virtual-on-site alignment, the decision generation engine is immediately invoked to convert the failure priority matrix and corresponding labels into a unified feature vector. The feature vector is then weighted according to severity and sent to the strategy retrieval.

[0070] Specifically, based on the valve-heat exchange coupling handling rules pre-existing in the knowledge base, multiple candidate handling sequences are automatically assembled, and each sequence is bound with complete safety interlocking and mutual constraint. Subsequently, the candidate sequences are injected into a dedicated twin simulation container. The container simultaneously tracks four evaluation indicators—energy recovery rate, reliability improvement, potential risk threshold, and estimated downtime—in a thermo-fluid-mechanical accelerated solution mode. A sliding integrator is used to eliminate the impact of transient noise on the stability of the indicators, and finally, a multi-dimensional performance profile is generated and written back to the decision buffer.

[0071] Next, the performance profile undergoes Pareto partitioning and analytic hierarchy process (AHP)-entropy weighted sorting in the multi-objective optimizer: the optimizer first excludes any sequences that exceed the risk threshold or have excessive downtime, and then weights the remaining sequences based on energy recovery rate and reliability improvement. The sequence with the highest score is established as the optimal solution, while the remaining high-potential sequences are saved as backup solutions in case of deviations during operation.

[0072] Before the optimal handling plan is issued, it undergoes in-depth parsing by the operation ticket compiler. The compiler generates atomic control instruction streams step by step and embeds interrupt protection logic and rollback paths that meet the safety integrity level requirements into the instruction streams. Subsequently, the signed control instruction streams, together with the target performance threshold, are issued as execution tokens to the process control system and safety instrumented system, achieving seamless connection from diagnosis to handling.

[0073] During instruction execution, the residual monitor continuously compares the real-time on-site curve with the twin prediction curve. Once the residual exceeds the adaptive convergence threshold, it immediately triggers the backup plan to re-enter the twin inference-performance evaluation-optimization selection loop, ensuring that the handling process has self-correcting capabilities and will not cause secondary risks. After the handling is completed, the diagnostic features, final plan, actual execution indicators, and residual trajectory are collected in the form of a quadruple and written into the twin knowledge graph. Based on this, the incremental learner performs online fine-tuning of the threshold updater, similarity weight set, and handling rule base, so that the data, model, and decision form a closed-loop self-evolution, ultimately achieving continuous improvement in fault diagnosis accuracy and collaborative handling efficiency, while significantly reducing on-site trial and error costs and energy consumption.

[0074] This invention also proposes an intelligent fault diagnosis system for heat exchange equipment based on digital twins, such as... Figure 2 As shown, it includes: a data synchronization and fusion module, a status feature generation module, a concurrent fault simulation and screening module, and a collaborative handling closed-loop module, with signal connections between each module;

[0075] The data synchronization and fusion module is mainly used to collect three sources of data: design configuration, maintenance history and real-time operating conditions. After unifying the fields and time scale, the data is written into the digital twin library and parallel coupled calculation is used to maintain the real and virtual states in real time.

[0076] The status feature generation module is mainly used to align and fuse the flow-pressure-temperature sequence with the color-infrared image, clean up anomalies, and send them into the dimensionality reduction network to generate a single health vector in real time and trigger an early warning flag.

[0077] The concurrent fault simulation screening module is mainly used to clone twin copies of abnormal equipment sets and control valve combinations, inject jamming disturbances in batches for parallel solution, and quickly lock the most matching fault combination by matching field curves.

[0078] The collaborative handling closed-loop module is mainly used to call the knowledge base to generate multiple handling strategies, select the optimal solution after evaluating its merits in the twin environment, and execute it. It also uses residual monitoring and iterative correction to achieve a data-model-decision closed loop.

[0079] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0080] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0081] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0084] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent fault diagnosis of heat exchange equipment based on digital twins, characterized in that, include: Collect design configuration, maintenance history and real-time operating condition data from three sources, unify fields and time scales and write them into the digital twin library, and use parallel coupled calculation to keep the virtual and real states synchronized at all times; The flow-pressure-temperature sequence is aligned and fused with color-infrared images. After anomalies are removed, the data is fed into a dimensionality reduction network to generate a single health vector in real time and trigger an early warning flag. Based on the cloned twins of abnormal equipment sets and control valve combinations, batch injection of jamming disturbances is performed and solved in parallel. The best matching fault combination is quickly identified by matching the field curve. During the steady-state operation of the high-pressure steam header heat exchange network, the flow rate, shell-side pressure difference and outlet temperature data are sampled. Then, all the original sampling frames are rewritten and mapped to a unified sampling window to generate a three-component tensor. After preprocessing, a standardized sequence is obtained. Then, the synchronization amplitude of each device and the pairwise correlation coefficients are solved. Calculate the synchronization amplitude threshold and the correlation threshold, perform threshold comparison on newly arrived sequences within the unified sampling window, and determine the power supply resonance synchronization anomaly set; Based on the power supply resonance synchronization anomaly set and the main pipe inlet regulating valve entity, a concurrent fault hypothesis is generated, and a virtual twin of the concurrent fault hypothesis is cloned. Once the virtual twin is ready, a throttling angle perturbation is applied according to the preset Kaze perturbation curve and injected into the three-dimensional multi-field solver as a time-dependent boundary. The three-dimensional multi-field solver dynamically subdivides the iteration steps according to the local pressure gradient and performs block sparse solution, controlling the computational error to be lower than the stability constraint when compressing the total computational load. The virtual twins output the simulation response curves in real time and upload them to the simulation result buffer. Then, the weighted Mahalanobis distance and dynamic time warping distance are combined to generate a comprehensive similarity. The comprehensive similarity is sorted in ascending order, a matching threshold is set, and the set matching threshold is used as a boundary to filter out the concurrent valve failure hypothesis with the smallest comprehensive similarity and the comprehensive similarity is less than or equal to the matching threshold. The corresponding virtual twins are then locked for root cause analysis. Multiple disposal strategies are generated by calling the knowledge base. After evaluating their merits in a twin environment, the optimal solution is selected and executed. The data-model-decision closed loop is achieved by using residual monitoring and iterative correction.

2. The intelligent fault diagnosis method for heat exchange equipment based on digital twins according to claim 1, characterized in that: Build a digital twin model, establish a shared digital twin library and integrate the equipment's design configuration files, maintain historical and real-time operating condition curves, and then virtually map the physical object's state.

3. The intelligent fault diagnosis method for heat exchange equipment based on digital twins according to claim 2, characterized in that: Comprehensive perception and monitoring of equipment operating status transforms the fused equipment data within the digital twin model into a unique health vector.

4. The intelligent fault diagnosis method for heat exchange equipment based on digital twins according to claim 3, characterized in that; The system retrieves the health vector sequence of continuous time windows from the digital twin database and performs dimensionality reduction processing. Then, it assesses the equipment degradation status, quantifies the health status, and enables anomaly detection and early warning.

5. The intelligent fault diagnosis method for heat exchange equipment based on digital twin according to claim 4, characterized in that: When an early warning occurs, the equipment operating parameters are retrieved to determine the coupling strength, the coupling information is output, and then the possible fault propagation chain is traced and recorded as hidden features. The coupled information and hidden features are combined and then input into the diagnostic engine for fault diagnosis.

6. The intelligent fault diagnosis method for heat exchange equipment based on digital twin according to claim 1, characterized in that; Extract the amplitude of the locked valve disturbance, the energy loss rate curve, and the phased increment of the flow-pressure differential coupling, perform joint fitting to output a fault severity scalar, assign fault type labels, then construct differentiated collaborative handling schemes, and perform optimal screening of collaborative handling schemes; Calculate the key feature consistency of other similar concurrent valve failure hypotheses, the energy imbalance exposure, and the combined confidence level of the second-order sensitivity to the power supply stability of the main pipe. Rank them in descending order to generate a list. When the concurrent valve failure hypothesis with the lowest overall similarity becomes abnormal, automatically select the concurrent valve failure hypothesis with the highest confidence level to replace the concurrent valve failure hypothesis with the lowest overall similarity.

7. The intelligent fault diagnosis method for heat exchange equipment based on digital twin as described in claim 6, characterized in that: After the root cause is identified, a candidate treatment sequence is automatically generated. The twin simulation performs multi-objective optimization to select the optimal solution, and compiles and issues safety instructions for execution. The residual monitor continuously compares the real-time curves on site with the twin prediction curves. When the residual exceeds the limit, the backup solution is switched. The execution results are written back to the knowledge graph and trigger the model's self-learning.

8. A digital twin-based intelligent fault diagnosis system for heat exchange equipment, used to implement the digital twin-based intelligent fault diagnosis method for heat exchange equipment as described in any one of claims 1-7, characterized in that, include: The system includes a data synchronization and fusion module, a status feature generation module, a concurrent fault simulation and screening module, and a collaborative handling closed-loop module, with signal connections between each module. The data synchronization and fusion module is mainly used to collect three sources of data: design configuration, maintenance history and real-time operating conditions. After unifying the fields and time scale, the data is written into the digital twin library and parallel coupled calculation is used to maintain the real and virtual states in real time. The status feature generation module is mainly used to align and fuse the flow-pressure-temperature sequence with the color-infrared image, clean up anomalies, and send them into the dimensionality reduction network to generate a single health vector in real time and trigger an early warning flag. The concurrent fault simulation screening module is mainly used to clone twin copies of abnormal equipment sets and control valve combinations, inject jamming disturbances in batches for parallel solution, and quickly lock the most matching fault combination by matching field curves. The collaborative handling closed-loop module is mainly used to call the knowledge base to generate multiple handling strategies, select the optimal solution after evaluating its merits in the twin environment, and execute it. It also uses residual monitoring and iterative correction to achieve a data-model-decision closed loop.

Images