Early warning and self-healing system for supercritical CO2 equipment based on multimodal trend fusion

Abstract

This invention discloses a supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion, comprising a sensing and computing layer, a hybrid model fusion diagnostic layer, a multi-level hierarchical self-healing control layer, and a shadow-following evaluation module. The sensing and computing layer collects operational data and constructs a private knowledge base. The hybrid model fusion diagnostic layer constructs a multi-dimensional residual feature space, generates early warning scores, and outputs early warning information. The multi-level hierarchical self-healing control layer executes a three-level strategy, optimizing control parameters through a cost function. The three-level strategy includes performance degradation self-healing, equipment fault self-healing, and emergency protection self-healing. This invention constructs a sensing and computing layer, a hybrid model fusion diagnostic layer, and a multi-level hierarchical self-healing control layer. It generates early warning scores through multi-dimensional residual trend analysis and matching with the private knowledge base, executes a three-level self-healing strategy, and introduces a shadow-following model to quantify the net self-healing benefit and dynamically optimize decision thresholds, thereby achieving early warning prediction for supercritical CO2 equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and safety control technology of energy and power systems, specifically to a supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion. Background Technology

[0002] With the advancement of the "dual carbon" goal, Brayton cycle power generation technology, using supercritical carbon dioxide (sCO2) as the working fluid, has become a leading direction for replacing traditional steam cycles due to its significant advantages such as compact system, high thermal efficiency, and good flexibility. A typical sCO2 power generation system includes a series of core equipment operating under high temperature, high pressure, and near-critical conditions, such as a turbine, a regenerative heat exchanger, a compressor pump, and a high-pressure storage tank. Because the physical properties of sCO2 change drastically near the critical point (31.1℃, 7.38MPa), any slight parameter disturbance can trigger cascaded instability in the system, posing a great threat to equipment safety and operational continuity.

[0003] Currently, the operation and maintenance of such complex thermal systems mainly suffer from the following technical deficiencies: 1) Single threshold alarm, lacking early warning detection capability: Existing systems generally use fixed threshold alarm logic, that is, an alarm is triggered when a single parameter such as temperature, pressure, or vibration exceeds a set limit. This method only responds when the fault has developed to a certain extent, lacking the ability to identify early warning signs of equipment operating trend decline and early micro-anomalies. The time window for operators to respond and make decisions is extremely short, which can easily lead to unplanned downtime accidents. For example, micro-wear of internal turbine seals or slow decline in heat exchanger efficiency are difficult to detect in time under a fixed threshold.

[0004] 2) Data silos and knowledge fragmentation lead to poor diagnostic accuracy: Existing diagnostic methods largely rely on real-time data from distributed control systems (DCS) and simple expert rules, failing to integrate heterogeneous information from multiple sources such as historical fault case libraries, maintenance records, and equipment mechanism models. These "data silos" result in weak root cause analysis capabilities for complex faults, easily leading to false alarms and missed alarms. For example, a decline in the performance of a regenerative heat exchanger may be caused by multiple factors such as scaling, internal leaks, or impurities in the working fluid; relying solely on pressure data cannot accurately identify the cause.

[0005] 3) Fault response relies on manual intervention, and the system lacks self-healing capabilities: From fault diagnosis to the execution of protective actions, the traditional model requires a lengthy process of "system alarm -> manual judgment -> strategy determination -> manual operation." Under the conditions of high temperature, high pressure, and high dynamic response requirements of CO2, the speed and accuracy of manual decision-making and operation cannot meet the needs of emergency fault handling. Shutdown is the only "safe" option, but this results in huge energy consumption and economic losses during start-up and shutdown. Current technology lacks a closed-loop control system capable of automatic degradation, bypass, and self-healing without shutdown within safety boundaries.

[0006] Therefore, there is an urgent need for a new method that can integrate operational trends, expert knowledge, and historical data to achieve early warning of precursory faults and enable rapid online self-healing control when faults occur, so as to ensure the safe, stable, and economical operation of sCO2 power generation systems. Summary of the Invention

[0007] To address the aforementioned issues, the present invention aims to propose an early warning and self-healing system for supercritical CO2 equipment based on multimodal trend fusion. This system constructs an intelligent sensing and edge computing layer, a hybrid model fusion diagnostic layer, and a multi-level hierarchical self-healing control layer. It generates early warning scores through multi-dimensional residual trend analysis and matching with a private knowledge base, implements a three-level self-healing strategy encompassing performance degradation, equipment failure, and emergency protection, and introduces a shadow-following model to quantify the net self-healing benefits and dynamically optimize decision thresholds. This enables supercritical CO2 equipment to transition from "post-event alarm" to "early warning prediction," significantly reducing unplanned downtime while ensuring safety.

[0008] This was achieved through the following technical solutions: A supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion includes a sensing and computing layer, a hybrid model fusion diagnostic layer, a multi-level hierarchical self-healing control layer, and a shadow following evaluation module. Among them, the perception computing layer is used to collect the operating data of each key device through the sensor network, and to build a private knowledge base for each key device; The hybrid model integrates a diagnostic layer to construct a multi-dimensional residual feature space, generate early warning scores and remaining effective time predictions by matching degradation indices with case similarity, and output early warning information. The multi-dimensional residual feature space includes basic residuals, dynamic residual cumulative sum, and residual change rate. The early warning information includes faulty devices, fault modes, early warning confidence, and expected deterioration time. A multi-level hierarchical self-healing control layer is used to execute a three-level strategy based on the early warning score and the expected deterioration time, and to optimize the control parameters of each key device through a cost function; the three-level strategy includes performance degradation self-healing, equipment failure self-healing and emergency protection self-healing.

[0009] Optionally, it also includes: a shadow following evaluation module, used to generate shadow models of each key component in the supercritical CO2 device without self-healing control; when a warning is issued due to a precursory warning, the shadow model starts from the same state of the supercritical CO2 device and injects a fault evolution model to simulate the uninterrupted path, which is used to compare the economic loss cost of the real path and the shadow path, and to optimize the cost function.

[0010] Optionally, the key components include a supercritical CO2 turbine, a regenerative heat exchanger, a circulating water cooling subsystem, a hot water lithium bromide refrigeration unit, a low-pressure CO2 storage tank, a CO2 compression pump, a high-pressure CO2 heater, and a high-pressure supercritical CO2 storage tank; a private knowledge base stores historical failure cases, maintenance records, mechanism models, and expert rules for each key component in a structured manner.

[0011] Optionally, the early warning score is a weighted composite of the degradation index and the maximum case similarity score; wherein, the degradation index is obtained by linear regression of the basic residuals through a sliding window; and the maximum case similarity score is obtained by calculating the cosine similarity between the real-time abnormal pattern vector and the feature vectors corresponding to historical fault cases and taking the maximum value.

[0012] Optionally, when executing the three-level strategy, if the early warning score is lower than the preset value A1 and the remaining effective time meets the preset value A2, then performance degradation self-healing is triggered; if the early warning score meets the preset value D1 and the remaining effective time is greater than D2, then equipment fault self-healing is triggered; D2 is the preset minimum time required for seamless switching; if the preset interlocking conditions of the supercritical CO2 turbine and the high-pressure supercritical CO2 storage tank are triggered, then emergency protection self-healing is triggered.

[0013] Optionally, the self-healing mechanism for performance degradation includes: adjusting the pump frequency in the circulating water cooling subsystem, or adjusting the cooling capacity of the lithium bromide unit, or linearly reducing the load setpoint of the supercritical CO2 turbine; the self-healing mechanism for equipment failure includes: performing online switching for the standby pump of the CO2 compressor pump to shut down the faulty CO2 compressor pump and start the standby pump; the self-healing mechanism for emergency protection includes: closing the quick-closing valve at the inlet of the supercritical CO2 turbine and opening the emergency relief valve of the high-pressure supercritical CO2 storage tank.

[0014] Optionally, after triggering the early warning information, a cost function is obtained by weighted summation based on the power loss and the preset safety margin.

[0015] Optionally, in the shadow following evaluation module, the simulation endpoint of the fault evolution model is determined as follows: if the fault mode corresponding to the early warning information can predict the trip time, then the predicted trip time is used as the simulation endpoint; if the fault mode is a slow performance degradation, then a fixed simulation duration is set as the simulation endpoint.

[0016] Optionally, the economic losses may include power generation losses, device repair or replacement costs, and start-up and shutdown costs.

[0017] Optionally, after executing the three-level strategy, the corresponding self-healing net benefit is also calculated, and the weights of any threshold and cost function in the three-level strategy are dynamically adjusted based on the positive or negative value and magnitude of the self-healing net benefit.

[0018] The beneficial effects of this invention compared to the prior art are: 1) Construct a multi-dimensional residual space and cumulative sum detection algorithm to capture early micro-anomalies in equipment, significantly advance the early warning window, and realize the early detection of fault precursors; 2) Integrate real-time data with private knowledge bases such as historical cases and maintenance records, and use cosine similarity to match similar faults to provide traceable evidence, improve accuracy and reduce false alarms and missed alarms; 3) Design a multi-level self-healing architecture with online adjustment, redundant switching and safety interlocking. Optimize control parameters through cost function, utilize the buffering capacity of high-pressure storage tank to support bumpless switching, and have the ability to self-heal with no downtime or fewer machines, significantly reducing adverse conditions such as unplanned downtime. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the framework of a supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion; Figure 2 A flowchart for shadow following evaluation and strategy optimization; Figure 3 This is a verification diagram for Case 1; Figure 4 This is a verification diagram for Case 2. Detailed Implementation

[0020] The following will be combined with the present invention Figures 1 to 4 The technical solutions in the embodiments of the present invention will be described in detail below.

[0021] like Figure 1 The diagram shown is a framework schematic of a supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion; combined with Figure 1 As shown, this scheme constructs a perception computing layer, a hybrid model fusion diagnostic layer, a multi-level hierarchical self-healing control layer, and a shadow-following evaluation module. It generates early warning scores by matching multi-dimensional residual trend analysis with a private knowledge base, executes a three-level self-healing strategy, and introduces a shadow-following model to quantify the net self-healing benefits and dynamically optimize decision thresholds. This is used for early warning prediction of supercritical CO2 equipment in supercritical compressed carbon dioxide power generation systems.

[0022] In this embodiment, the sensing and computing layer collects operational data from key components in the supercritical compressed carbon dioxide power generation system through a sensor network, and constructs a private knowledge base for each key component. The private knowledge base stores historical fault cases, maintenance records, thermodynamic mechanism models, and other data for each key component in a structured manner.

[0023] Key components include a supercritical CO2 turbine (referred to as the turbine), a regenerative heat exchanger, a circulating water cooling subsystem, a hot water lithium bromide chiller unit, a low-pressure CO2 storage tank, a CO2 compressor pump, a high-pressure CO2 heater, and a high-pressure supercritical CO2 storage tank. Supercritical CO2 turbine: Collects data on radial / axial vibration of the turbine head, bearing temperature, rotational speed, and inlet / outlet temperature and pressure; this unit is centripetal and includes the turbine head and generator. Regenerative heat exchanger: Collects inlet and outlet temperatures and pressures of the working fluid on the hot / cold side to calculate the heat exchange difference and efficiency. Circulating water cooling subsystem: Collects data on the start / stop status, current, outlet pressure, and flow rate of two circulating water pumps; this subsystem is located outside the container and connected via a pipe coupling interface. Hot water lithium bromide chiller unit: Collects data on the inlet and outlet temperatures of chilled water, the temperature and flow rate of the hot water source (approximately 115°C), and the cooling capacity; this unit is switched via a three-way valve and can be bypassed in winter. Low-pressure CO2 storage tank (8MPa, approximately 20℃): Collects liquid level, pressure, and temperature data; employs vacuum insulation and is equipped with a safety valve. CO2 compressor pumps (3 main, 1 backup variable frequency design): Collects operating frequency, current, outlet pressure, bearing temperature, and vibration data for each pump; used to achieve the pressurization process from 8MPa liquid to 14MPa liquid. High-pressure CO2 heater: Collects CO2 side outlet temperature and pressure, and hot water side inlet temperature and flow rate; used to heat 14MPa CO2 to 115℃. High-pressure supercritical CO2 storage tank (14MPa, 115℃): Collects temperature, pressure, and liquid level data; the storage capacity of the high-pressure supercritical CO2 storage tank meets the system's rated operation for 20 minutes and supports black start.

[0024] In this embodiment, the hybrid model integrates a diagnostic layer to construct a multi-dimensional residual feature space, generate a warning score and remaining effective time prediction by matching the degradation index with case similarity, and output warning information. The warning score is a weighted composite of the degradation index and the maximum case similarity score; the degradation index is obtained through linear regression of the basic residuals using a sliding window; the maximum case similarity score is calculated by taking the maximum value of the cosine similarity between the real-time abnormal pattern vector and the corresponding feature vectors of historical fault cases.

[0025] The multidimensional residual feature space defines a series of characteristic indicators based on physical operating principles for each key component, and its essence is various forms of residuals. Specifically, the multidimensional residual feature space includes basic residuals (performance residuals, vibration residuals, etc.), cumulative sum of dynamic residuals, and residual change rate.

[0026] any t The basic residual at time e base (t) , is the measured value of each sensor in the sensor network. y actual (t) Compared with the predicted value y predicted (t) The absolute value of the difference. For example: for a regenerative heat exchanger, calculate the performance residual. , η actual To measure energy efficiency, η pred To predict energy efficiency, a continuous, unidirectional increase in the efficiency residual is a strong early warning indicator of scaling or micro-leakage in the regenerator heat exchanger. For turbines / CO2 compressor pumps, vibration prediction values ​​can be obtained by establishing a conventional vibration autoregressive prediction model in existing technology. Combined with measured values v(t) Then the vibration residual is calculated. A sustained, non-steady increase is a typical characteristic of early damage to rotor components in turbines / CO2 compressor pumps. (Predicting energy efficiency) η pred The data is obtained through predictions using known existing models, such as existing thermodynamic mechanism models, autoregressive or regression-based data-driven models, and hybrid models that combine mechanism and data-driven approaches. The purpose is to provide reference values ​​for energy efficiency under normal equipment operation.

[0027] Dynamic residual cumulative sum S CUSUM It is used to sensitively capture the slight upward drift of the residual mean. ,in, μ 0 represents the mean of all basic residuals under normal operating conditions. k This is a preset allowable small drift amount (which can be half the average residual value under normal operating conditions). When S CUSUM If the control limit H (the preset critical value) is exceeded, it is determined that there is a statistically significant small deviation.

[0028] residual change rate Reres By calculating T s The magnitude of change over a time period is used to assess the rate of failure deterioration. When Reres When the value is consistently positive and continuously increasing, it indicates that the damage has entered an accelerated phase.

[0029] Early warning information includes the faulty device, fault mode, early warning confidence level, and estimated deterioration time, which is the remaining effective time. Fault modes are divided into faults with predictable trip times and faults with unpredictable trip times. Faults with predictable trip times, such as a turbine bearing vibration continuously rising to the trip threshold, can have their trip time predicted by trend analysis. The corresponding characteristics of faults with predictable trip times are: equipment condition parameters (such as vibration, residuals, temperature, pressure, etc.) exhibit a continuous, stable, and fitable deterioration trend; the corresponding judgment method is to calculate the deterioration slope through sliding window linear regression or exponential regression and extrapolate it to the safety limit.

[0030] Faults with unpredictable trip times include, for example, sudden leaks, transient short circuits, and nonlinear sudden damage. The characteristics of faults with unpredictable trip times are: random occurrence, rapid jumps, and difficulty in predicting the specific trip time through trends; the corresponding judgment method is: residuals or performance indicators do not show a clear and continuous deterioration trend, or the failure is abnormally sudden.

[0031] For early warning information, the degradation index needs to be calculated first: apply a sliding window linear regression of width N to the basic residuals to obtain the degradation slope. k trend and goodness of fit Deterioration Index D degrade Defined as: Where f and g are normalization functions, and α and β are their respective weights. Simultaneously, case similarity matching calculation is also required: the real-time anomaly pattern vector P... anom Calculate the cosine similarity between the feature vectors of the same historical failure cases in the private knowledge base and the feature vectors of the same cases. P hist,i Let be the feature vector of the i-th identical historical fault case; take the maximum similarity as the matching score. Then, calculate the early warning score: Where γ is the weight, I For threshold exceedance indicator function, The maximum value of the cosine similarity; when PWS(t) >Preset values ​​for warnings θ warn At that time, a preliminary warning was officially issued. It should be noted that the real-time anomaly pattern vector P... anom It is calculated from real-time data of key equipment collected by sensor network. The specific method is to calculate the basic residual, dynamic residual cumulative sum, residual change rate and other features for each key parameter, and then combine these features into a vector for similarity matching with the feature vector of historical failure cases.

[0032] When predicting the remaining effective time, the degradation is extrapolated based on the degradation slope, and the degradation to the shutdown limit is expected. e limit time , e(t) Indicators used for extrapolating and predicting fault development can be basic residuals. e_base(t) (Such as heat exchanger performance residuals), or vibration residuals. e_vib(t) (Such as the rotor vibration residual of a turbine or compressor pump), both are used to calculate the degradation trend and predict the remaining effective time. This allows for the output of a four-tuple of early warning information containing the faulty equipment, failure mode, precursor confidence level, and expected deterioration time.

[0033] In this embodiment, a multi-level hierarchical self-healing control layer is used to execute a three-level strategy based on the early warning score and the expected deterioration time, and to optimize the control parameters of each key device through a cost function; the three-level strategy includes performance degradation self-healing, equipment failure self-healing and emergency protection self-healing.

[0034] The following is an example of the operating logic of a 1.5MW supercritical compressed carbon dioxide power generation system. Normal operating logic: Liquid CO2 in the low-pressure CO2 storage tank is pressurized to 14MPa by a compressor pump, enters the high-pressure CO2 heater, and is heated to 115℃ to become supercritical, then temporarily stored in the high-pressure supercritical CO2 storage tank. The sCO2 output from the high-pressure supercritical CO2 storage tank enters the hot side of the regenerative heat exchanger, recovers heat from the exhaust gas at the supercritical CO2 turbine outlet for preheating, and then enters the high-pressure CO2 heater to be heated to the operating temperature before re-entering the supercritical CO2 turbine to generate electricity. Exhaust gas refers to the low-temperature, low-pressure working gas discharged from the supercritical CO2 turbine or turbine outlet, i.e., gas that has already done work and released energy. The high-temperature, low-pressure sCO2 at the turbine outlet sequentially passes through the cold side of the regenerative heat exchanger, a hot-water lithium bromide refrigeration unit (in summer) or directly through a bypass (in winter), and a circulating water cooling system, where it is cooled and condensed into liquid, finally returning to the low-pressure CO2 storage tank, completing a closed-loop cycle.

[0035] The triggering logic for multi-level hierarchical self-healing is as follows: Define a strategy selection function Φ, based on the current state S(t) = { PWS (t) , t RUL (t) Fault mode output control strategy level L: L(t) = Φ (PWS(t) , t RUL (t)) .

[0036] When implementing the Level 3 strategy, if the early warning score is lower than the preset value A1, and the remaining effective time meets the preset value A2, then performance degradation self-healing is triggered (L=1). A1 is... θwarn，L1 A2 is for t RUL (t) Preset values. Self-healing for performance degradation includes adjusting the pump frequency in the circulating water cooling subsystem, adjusting the cooling capacity of the lithium bromide unit, or linearly reducing the load setpoint of the supercritical CO2 turbine, thereby enabling timely adjustments in the early stages of performance degradation.

[0037] If the early warning score meets the preset value D1 and the remaining effective time is greater than D2, then equipment fault self-healing (L=2) is triggered, where D2 is the preset minimum time required for seamless switching. Executing equipment fault self-healing includes: performing online switching of the backup pump for the CO2 compressor pump to reduce the load on the faulty CO2 compressor pump, then shutting down the faulty CO2 compressor pump and starting the backup pump. During equipment fault self-healing, a weighted summation based on power loss and a preset safety margin can be performed to solve the economic cost function (referred to as the cost function, used to determine the economic loss cost) J in real time to optimize the control parameter u. ,in, P loss The power loss is represented by SafetyMargin, which is the safety margin. y ref Set values ​​for key parameters (e.g., control the penalty weight to be more than 10 times, to strengthen safety priority in the cost function and limit over-adjustment of key parameters). y (u) is the actual parameter value corresponding to u (measured by the sensor). , , They are three different penalty weights and Safety margin is a preset safety boundary that is determined by engineers during the system design phase based on equipment specifications and operational safety requirements. It is used in the cost function to penalize control operations to ensure that safety takes priority. y ref Key parameter settings, also known as preset values, correspond to the rated operating parameters or design target values ​​of key components, such as the rated frequency of a pump, turbine load, and temperature difference of a heat exchanger. This is achieved by controlling the penalty weights. (For example, more than 10 times) to achieve safety first and take into account the economy, without stopping the machine for self-healing.

[0038] The preset value D2 represents the minimum uninterrupted handover time required for equipment fault self-healing. This parameter is fixed during the system design phase, not calculated in real-time. The system will only trigger equipment fault self-healing when the early warning score reaches D1 and the remaining effective time is greater than D2, ensuring a safe handover process without interrupting normal operation.

[0039] If the preset interlock conditions of the supercritical CO2 turbine and the high-pressure supercritical CO2 storage tank are triggered (referring to a combination of critical components reaching dangerous thresholds, such as excessive turbine vibration or excessive pressure in the high-pressure storage tank; if these conditions are met, emergency protection self-healing is triggered), then emergency protection self-healing is triggered (L=3). Executing emergency protection self-healing includes: closing the quick-closing valve at the inlet of the supercritical CO2 turbine, opening the emergency relief valve of the high-pressure supercritical CO2 storage tank, and prioritizing safe shutdown.

[0040] like Figure 2 The diagram shows a flowchart of shadow-following evaluation and strategy optimization. This system also includes a shadow-following evaluation module to generate shadow models of key components in a supercritical CO2 device that are not subject to self-healing control. When a warning is issued based on an early warning signal, the shadow model is injected into the fault evolution model from the same state of the supercritical CO2 device to simulate the uninterrupted path. This is used to compare the economic losses of the actual path and the shadow path, and to optimize the cost function. Economic losses include power generation loss, component repair or replacement costs, and start-up / shutdown costs. The simulation endpoint of the fault evolution model is determined as follows: if the fault mode corresponding to the early warning signal can predict the trip time, the predicted trip time is used as the simulation endpoint; if the fault mode is a slow performance degradation, a fixed simulation duration is set as the simulation endpoint.

[0041] When the system is in a stable and healthy operating state (meaning the PWS score is below the preset minimum warning threshold), a dynamically aligned shadow model is automatically created. This shadow model is implemented as a state vector X. shadow It includes key thermodynamic parameters and performance indicators that are completely consistent with the real system, such as rotational speed, pressure, temperature, heat transfer efficiency, and vibration amplitude residuals. X real The state values ​​of the real system. t 0 means t At time 0, n It is the dimension or number of elements of the state vector, that is, the number of key thermodynamic parameters and performance indicators contained in the state vector of the shadow model or the real system. T out The superscript represents the transpose operation of a vector or matrix, used in mathematical expressions to transform a column vector into a row vector, or vice versa; T This indicates transpose; p out This indicates the outlet pressure (e.g., the outlet pressure of a turbine, CO2 compressor pump, etc.). η eff This indicates heat exchange efficiency (e.g., the heat exchange efficiency of a heat exchanger).

[0042] When a real system detects an anomaly and issues a warning (i.e.) PWS ≥ θ warn When the warning time is [time], [the warning will be issued]. t warn As a branching point, a shadow-following simulation is initiated. This simulation is divided into two different paths: the real path (with the self-healing strategy already implemented): will evolve according to the aforementioned self-healing process; the shadow path (assuming no intervention): will then proceed from... t warn Starting from the same initial state at any given time, the diagnosed fault evolution model is injected to simulate the laissez-faire trajectory without intervention.

[0043] Fault evolution models use mathematical equations to describe the degradation process of performance or vibration indicators. Fault evolution models include: linear degradation sub-models: Exponential degradation sub-model: ; k trend The deterioration slope when applying a sliding window of width N to linear regression based on the base residuals. and They represent t Time and t warn The performance metrics at any given time, here e Let λ be the base of the natural logarithm, and let λ represent the basic residual over a time period with a sliding window width of N. e base (t) The degradation slope obtained from exponential regression is used to quantify the rate of accelerated failure deterioration, predict remaining effective time, and determine the failure evolution path. It is applicable to exponential degradation models and differs from the degradation slope in linear degradation models. k trend Correspondingly, two models, linear degradation and exponential degradation, are used, which complement each other and cover the degradation trends of different types of faults. If the fault is a slow and stable performance degradation, linear degradation is used; if the fault is accelerated or nonlinear, exponential degradation is used.

[0044] The shadow model is set with a clearly defined evaluation time window endpoint. t end If the failure mode allows for a predictable trip time, then the predicted trip time will be used. As the simulation endpoint, t end = t trip If the failure mode is a gradual performance degradation (without a significant shutdown time), then a fixed simulation duration should be set. T sim (e.g., 24 hours) t end = t warn + Tsim .exist[ t warn , t end Within the interval, dual-track parallel computation is performed: Under self-healing control, the real system samples X once every Δt time interval. real The shadow model is based on the degradation equation and simulates X once every Δt time interval. shadow .

[0045] In this embodiment, after executing the three-level strategy, the corresponding self-healing net benefit is also calculated. Based on the sign and magnitude of the self-healing net benefit, the weights of any threshold and cost function in the three-level strategy are dynamically adjusted. t end At time t, calculate the combined economic loss cost for both paths. The loss cost function C is defined as follows: .

[0046] Real path cost C real middle, C energy_loss The power generation loss during the execution of the self-healing strategy can be obtained by integrating the power derating curve over time. C equip_damage The potential damage to various components due to reduced load operation can be disregarded; C startup_penalty This is the start-up and shutdown cost included when the self-healing strategy involves downtime (L=3); if no downtime is involved, this item is zero.

[0047] Shadow path cost C shadow middle, C energy_loss To simulate to t end At that time, the cumulative efficiency loss due to performance degradation, plus from t trip Losses due to complete power outage during the period until repairs are completed; C equip_damage The cost of repairing or replacing the main components required to prevent the failure from progressing to a shutdown (such as severe bearing seizure or impeller damage); C startup_penalty This refers to the cost of restarting the system to full capacity after a shutdown.

[0048] Therefore, the net benefit of the self-healing strategy is quantified as follows: ; B net This refers to the direct economic loss avoided by this self-healing control, calculated based on the shadow-following assessment. Table 1 below shows the results based on B... netThe magnitude and sign of the value trigger different judgment conclusions and subsequent strategy optimization actions: Table 1:

[0049] It should be noted that when the system executes its self-healing strategy, it creates a shadow-following model that runs parallel to the original system. This model synchronizes the system state in real time, but simulates the future evolution path without any intervention. By comparing this shadow path with the actual path after self-healing intervention, the direct economic losses avoided by this self-healing control can be quantitatively assessed. The assessment feedback can then be used to optimize future self-healing decision thresholds and strategies, achieving the goal of prioritizing safety while ensuring economic efficiency.

[0050] like Figure 3 and Figure 4 The diagrams shown are verification diagrams for Case 1 and Case 2, respectively. Case 1 demonstrates the self-healing control results for the early signs of rotor imbalance in a supercritical carbon dioxide turbine, while Case 2 demonstrates the self-healing control results for the early signs of fouling in a supercritical carbon dioxide regenerative heat exchanger. Both cases effectively verify the economic benefits of implementing self-healing control.

[0051] In summary, this invention constructs a multi-dimensional residual space and a cumulative sum detection algorithm to capture early microscopic anomalies in equipment, significantly advancing the warning window and enabling the identification of early signs of faults. It also integrates real-time data with a private knowledge base including historical cases and maintenance records, using cosine similarity to match similar faults, providing traceable evidence, improving accuracy, and reducing false alarms and missed alarms. Furthermore, it designs a multi-level self-healing architecture with online adjustment, redundancy switching, and safety interlocks. By optimizing control parameters through a cost function and utilizing the buffering capacity of high-pressure storage tanks to support seamless switching, it possesses the ability to self-heal with or without downtime, significantly reducing unplanned downtime and other adverse conditions, demonstrating significant progress.

[0052] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion, characterized in that, It includes a perception computing layer, a hybrid model fusion diagnostic layer, a multi-level hierarchical self-healing control layer, and a shadow-following assessment module; Among them, the perception computing layer is used to collect the operating data of each key device through the sensor network, and to build a private knowledge base for each key device; The hybrid model integrates a diagnostic layer to construct a multi-dimensional residual feature space, generate early warning scores and remaining effective time predictions by matching degradation indices with case similarity, and output early warning information. The multi-dimensional residual feature space includes basic residuals, dynamic residual cumulative sum, and residual change rate. The early warning information includes faulty devices, fault modes, early warning confidence, and expected deterioration time. A multi-level hierarchical self-healing control layer is used to execute a three-level strategy based on the early warning score and the expected deterioration time, and to optimize the control parameters of each key device through a cost function; the three-level strategy includes performance degradation self-healing, equipment failure self-healing and emergency protection self-healing.

2. The supercritical CO2 equipment early warning and self-healing system based on multi-modal trend fusion according to claim 1, characterized in that, Also includes: The shadow following evaluation module is used to generate shadow models of key components in supercritical CO2 equipment without self-healing control. When a warning is issued due to a precursory warning, the shadow model starts from the same state of the supercritical CO2 equipment and injects a fault evolution model to simulate the uninterrupted path. This is used to compare the economic loss cost of the real path and the shadow path, and to optimize the cost function. 3.The supercritical CO2 device early warning and self-healing system based on multi-modal trend fusion of claim 1, wherein, Key components include supercritical CO2 turbine, regenerative heat exchanger, circulating water cooling subsystem, hot water lithium bromide refrigeration unit, low-pressure CO2 storage tank, CO2 compressor pump, high-pressure CO2 heater, and high-pressure supercritical CO2 storage tank; a private knowledge base stores historical failure cases, maintenance records, mechanism models, and expert rules for each key component in a structured manner.

4. The supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion according to claim 3, characterized in that, The early warning score is a weighted composite of the degradation index and the maximum case similarity score. The degradation index is obtained by linear regression of the basic residuals through a sliding window. The maximum case similarity score is obtained by calculating the cosine similarity between the real-time abnormal pattern vector and the feature vectors corresponding to historical fault cases and taking the maximum value.

5. The supercritical CO2 equipment early warning and self-healing system based on multi-modal trend fusion according to claim 3, characterized in that, When executing the Level 3 strategy, if the early warning score is lower than the preset value A1 and the remaining effective time meets the preset value A2, then the performance degradation self-healing will be triggered. If the early warning score meets the preset value D1 and the remaining effective time is greater than D2, the device fault self-healing will be triggered; D2 is the preset minimum time required for seamless handover. If the preset interlocking conditions of the supercritical CO2 turbine and the high-pressure supercritical CO2 storage tank are triggered, the emergency protection self-healing will be activated.

6. The supercritical CO2 equipment early warning and self-healing system based on multi-modal trend fusion according to claim 5, characterized in that, Self-healing of performance degradation includes: adjusting the pump frequency in the circulating water cooling subsystem, or adjusting the cooling capacity of the lithium bromide unit, or linearly reducing the load setpoint of the supercritical CO2 turbine; The equipment fault self-healing includes: performing online switching of the backup pump for the CO2 compressor pump to deactivate the faulty CO2 compressor pump and start the backup pump; Implementing emergency protection and self-healing measures includes: closing the quick-closing valve at the inlet of the supercritical CO2 turbine and opening the emergency relief valve of the high-pressure supercritical CO2 storage tank.

7. The supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion according to claim 1, characterized in that, After triggering the early warning information, the cost function is obtained by weighted summation based on the power loss and the preset safety margin. 8.The supercritical CO2 device early warning and self-healing system based on multi-modal trend fusion of claim 2, wherein, In the shadow follower evaluation module, the simulation endpoint of the fault evolution model is determined as follows: if the fault mode corresponding to the early warning information can predict the trip time, then the predicted trip time is used as the simulation endpoint; if the fault mode is a slow performance degradation, then a fixed simulation duration is set as the simulation endpoint. 9.The supercritical CO2 device early warning and self-healing system based on multi-modal trend fusion of claim 2, wherein, The economic losses include power generation loss, equipment repair or replacement costs, and start-up and shutdown costs.

10. The supercritical CO2 equipment early warning and self-healing system based on multimodal trend fusion according to claim 1, characterized in that, After implementing the three-level strategy, the corresponding self-healing net benefit is also calculated. Based on the positive or negative value and magnitude of the self-healing net benefit, the weights of any threshold and cost function in the three-level strategy are dynamically adjusted.

Images