Fault prediction and health management method for compressed air energy storage power station

Abstract

The invention relates to the technical field of compressed air energy storage, and discloses a fault prediction and health management method for a compressed air energy storage power station, which comprises the following steps: establishing an accurate physical model of a key component of the compressed air energy storage power station, constructing a fault simulation injection mechanism to simulate a real fault, and obtaining a fault mode library; setting different severity levels for each fault mode in the fault mode library; collecting multi-source data in a simulation process under multiple working conditions; extracting sensitive features from the multi-source data and associating the sensitive features with corresponding fault tags; constructing a hybrid deep learning fault prediction model to output fault classification and residual service life; and evaluating the health state of the equipment based on multi-source information fusion. According to the method, massive fault data are generated through the high-fidelity simulation model, the hybrid deep learning fault prediction model is trained, fault prediction and health assessment of the compressed air energy storage power station are achieved, and the requirements of the large compressed air energy storage power station for high reliability, high availability and intelligent operation and maintenance are met.

Description

Technical Field

[0001] This invention relates to the field of compressed air energy storage technology, and in particular to a method for fault prediction and health management of compressed air energy storage power plants. Background Technology

[0002] Compressed air energy storage (CAES), as a large-scale, long-term energy storage technology, plays a crucial role in building new power systems and promoting the integration of renewable energy. With the advancement of dual-carbon goals, the construction and application of CAES power plants are accelerating. However, CAES systems have complex structures, and their core equipment (such as multi-stage compressors, high-parameter expanders, and large gas storage devices) operate under harsh conditions of high pressure, high temperature, and high speed for extended periods, resulting in a high risk of failure. Once a failure occurs, it can not only lead to unplanned shutdowns and significant power generation losses but may also trigger serious equipment damage and safety accidents.

[0003] Currently, the operation and maintenance of CAES power plants mainly relies on traditional technologies, which have significant shortcomings in prognostics and health management (PHM). Existing technical solutions mainly include: 1. Periodic maintenance mode: Equipment is shut down for inspection and disassembly repair based on fixed time cycles (such as quarterly or annually). This mode lacks awareness of the actual health status of the equipment, which may lead to "over-maintenance" or "under-maintenance". 2. Threshold-based alarm monitoring: Upper and lower limits are set for alarms on key parameters (such as temperature, pressure, and vibration) through distributed control systems (DCS) or supervisory data acquisition and monitoring systems (SCADA). This method can only issue alarms when a fault has occurred or is about to occur, which is a reactive or in-process control and cannot achieve early warning. 3. Traditional fault diagnosis expert systems: Based on a limited expert experience rule base, specific fault phenomena are matched for diagnosis. Their knowledge base is difficult to update and cannot cope with complex and ever-changing new and compound faults.

[0004] While the aforementioned technologies played a role at a certain historical stage, as CAES power plants develop towards larger capacities and higher parameters, the existing technologies have revealed the following major defects and shortcomings in terms of reliability, economy, and intelligence: 1. Outdated and unpredictable operation and maintenance models lead to significant losses from unplanned downtime. Passive response and reactive maintenance: Current technology is essentially a passive "fault occurrence-alarm-downtime-repair" model. It fails to identify potential risks in the early stages of a fault or early degradation, missing the optimal maintenance window. Statistics show that in large industrial rotating machinery using traditional operation and maintenance models, unplanned downtime accounts for over 15%, resulting in extremely high production losses and emergency repair costs. "One-size-fits-all" scheduled maintenance is unscientific: Scheduled maintenance does not consider the equipment's actual operating history and current health status. For equipment in good working condition, unnecessary disassembly and inspection may introduce new fault risks; while for equipment with potential defects, fixed maintenance cycles may not cover its fault development period, causing it to fail between maintenance periods.

[0005] 2. Fault prediction technology faces bottlenecks, with insufficient accuracy and early warning capabilities. Modeling complex fault mechanisms is difficult: CAES systems are strongly coupled systems involving aerodynamics, thermodynamics, mechanics, and electrical systems, resulting in complex fault mechanisms (such as compressor surge, expander blade fatigue fracture, and gas storage chamber sealing failure). Traditional methods based on single physical models struggle to accurately describe their dynamic characteristics and fault evolution processes. Data-driven methods face the challenge of data scarcity: Data-driven methods, such as machine learning, heavily rely on large amounts of historical fault data for model training. However, for large-scale critical facilities like CAES, which have high reliability, sample data of severe faults in actual operation are extremely scarce, leading to the risk of overfitting in data-driven models, poor generalization ability, and difficulty in applying them to practical engineering. Weak multi-source information fusion capabilities: While existing monitoring systems can collect multi-source data such as vibration, acoustics, temperature, pressure, and performance parameters, they lack effective multi-source information fusion and analysis methods. Each parameter is often monitored independently, failing to capture early, subtle signs of faults from a systemic and interconnected perspective.

[0006] In summary, existing technologies cannot meet the urgent needs of large-scale compressed air energy storage power plants for high reliability, high availability, and intelligent operation and maintenance. Developing an integrated intelligent system capable of early fault prediction and accurate condition assessment has become a key issue that urgently needs to be addressed for the development and large-scale application of CAES technology. Summary of the Invention

[0007] This invention provides a fault prediction and health management method for compressed air energy storage power stations. It generates massive amounts of fault data through a high-fidelity simulation model, trains a hybrid deep learning fault prediction model, and realizes fault prediction and health assessment of compressed air energy storage power stations, meeting the urgent needs of large-scale compressed air energy storage power stations for high reliability, high availability and intelligent operation and maintenance.

[0008] This invention provides a fault prediction and health management method for compressed air energy storage power stations, comprising: S1. Establish accurate physical models of key components of the compressed air energy storage power station, including compressor model, air storage device model, expander model, heat exchange system model, and control system model; S2. Construct a fault simulation injection mechanism to simulate real faults by perturbing parameters and modifying boundary conditions of the precise physical model, including progressive fault simulation and sudden fault simulation, to obtain a fault mode library. S3. Set different severity levels for each fault mode in the fault mode library; wherein, the severity levels include minor, moderate and severe; S4. Under multiple operating conditions, collect multi-source data at a set frequency during the simulation process based on the accurate physical model, including process parameters, mechanical parameters, electrical parameters and performance parameters; S5. Extract sensitive features that can characterize the trend of equipment condition degradation from the multi-source data, and associate the sensitive features with the corresponding fault labels according to the fault mode library to construct the original dataset for model training. S6. Construct a hybrid deep learning fault prediction model to output fault classification and remaining service life based on the hybrid deep learning fault prediction model; S7. Health status assessment based on multi-source information fusion.

[0009] Furthermore, in S1, the compressor model is based on thermodynamic and fluid dynamic equations to establish a variable operating condition model of a multi-stage axial / centrifugal compressor to simulate flow-head characteristics, efficiency MAP, and surge boundary. The gas storage device model is designed for salt caverns, abandoned mines, or high-pressure storage tanks. An unsteady-state model considering geomechanical characteristics and thermodynamics is established to simulate pressure changes, temperature stratification, and gas leakage. The expander model is a multi-stage expander model to simulate its work capacity, efficiency characteristics, and partial load performance. The heat exchange system model is a distributed parameter model of the regenerator and intercooler to simulate the decrease in heat exchange efficiency and scaling. The control system model simulates the power plant's start-up and shutdown logic, load regulation logic, and safety interlock protection system.

[0010] Furthermore, in S2, the progressive fault simulation includes performance degradation faults and clearance increase faults. The performance degradation fault is simulated by gradually modifying the efficiency coefficient and flow coefficient in the precise physical model over time to simulate the slow development of faults such as compressor blade fouling and expander flow passage wear. The clearance increase fault simulates the change in rotor dynamic characteristics caused by bearing wear, which is achieved by changing the mass, stiffness, and damping matrix. The sudden failure simulation includes leakage failure, sensor failure, and control system failure. The leakage failure is to introduce an additional mass flow outlet in the gas storage chamber or pipeline model to simulate leakage at different locations and sizes. The sensor failure simulates sensor drift, jamming, and decreased accuracy, which is achieved by modifying the corresponding measurement output value. The control system failure simulates actuator jamming and controller failure.

[0011] Furthermore, in S4, the process parameters include pressure, temperature, and flow rate at each stage; the mechanical parameters include X / Y axis vibration, axis displacement, bearing temperature, and noise; the electrical parameters include motor current, voltage, and power; and the performance parameters include efficiency, pressure ratio, and output.

[0012] Furthermore, S5 specifically includes: S501. Extract sensitive features that characterize the degradation trend of equipment status from the multi-source data, including time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features; wherein, the time-domain features are mean, variance, peak value, kurtosis, and waveform factor; the frequency-domain features are obtained by FFT transformation to obtain the spectrum and extract the dominant frequency, harmonic amplitude, and centroid frequency; the time-frequency-domain features are obtained by wavelet transform or empirical mode decomposition to extract non-stationary signal features such as energy entropy and frequency band energy distribution; the deep learning features are obtained by using a one-dimensional convolutional neural network to automatically learn deep features directly from the original vibration signal; S502. Associate all sensitive features with corresponding fault labels according to the fault mode library to construct an original dataset for model training; wherein, the fault labels are extracted from the fault mode library and include fault type, severity, and remaining service life.

[0013] Furthermore, in S6, the hybrid deep learning fault prediction model includes a 1D-CNN branch, an LSTM branch, an attention mechanism, and an output layer; The hybrid deep learning fault prediction model preprocesses and extracts features from the collected multi-source data. Mechanical parameters are input into the 1D-CNN branch to automatically learn local fault features, while time-series data such as process parameters and performance parameters are input into the LSTM branch to capture long-term degradation trends. The feature vectors extracted by the two branches are concatenated and dynamically weighted and fused through an attention mechanism, so that the model focuses on the key information most relevant to the current fault. The fused features are fed into the multi-task output layer, which simultaneously completes the accurate classification of fault type and severity and the regression prediction of remaining service life, so as to realize the analysis from multi-source data input to fault warning and service life assessment.

[0014] Furthermore, S7, the health status assessment based on multi-source information fusion, specifically includes: S701, Health Index Calculation: Integrating multiple sensitive features extracted from multi-source data and deep state representations learned from a hybrid deep learning fault prediction model, and using a fusion algorithm of fuzzy inference or deep belief network, the above multi-dimensional indicators are mapped to a comprehensive health index between 0 and 1. This index quantifies the overall health of the equipment, where 1 represents a brand new state and 0 represents complete failure. S702. Remaining service life prediction: Taking the current health index calculated in S701 and its historical change sequence as input, the particle filter or deep learning sequence prediction model is used to predict the future trajectory of the health index, deduce the time when the equipment reaches the failure threshold, output the predicted value of the remaining service life of the equipment, and provide a confidence interval to characterize the uncertainty of the prediction.

[0015] The present invention also provides a fault prediction and health management device for compressed air energy storage power stations, comprising: A module is established to create accurate physical models of key components of a compressed air energy storage power station, including compressor models, air storage device models, expander models, heat exchange system models, and control system models. The simulation module is used to construct a fault simulation injection mechanism to simulate real faults by perturbing parameters and modifying boundary conditions of the precise physical model, including progressive fault simulation and sudden fault simulation, to obtain a fault mode library. The setting module is used to set different severity levels for each fault mode in the fault mode library; wherein the severity levels include minor, moderate and severe. The acquisition module is used to acquire multi-source data, including process parameters, mechanical parameters, electrical parameters and performance parameters, at a set frequency during the simulation process based on the accurate physical model under multiple operating conditions. An extraction module is used to extract sensitive features that can characterize the trend of equipment condition degradation from the multi-source data, and associate the sensitive features with the corresponding fault labels according to the fault mode library to construct the original dataset for model training. The output module is used to construct a hybrid deep learning fault prediction model to output fault classification and remaining service life based on the hybrid deep learning fault prediction model. The assessment module is used for health status assessment based on multi-source information fusion.

[0016] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0017] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0018] The beneficial effects of this invention are as follows: This invention provides an integrated solution for fault prediction and health management of compressed air energy storage power stations through the fusion of multidisciplinary technologies. Specifically, it constructs a multiphysics simulation model to generate a rich dataset with multiple operating conditions and fault types, safely simulating extreme faults that are not allowed in reality, thus fundamentally solving the problem of scarce fault data. It employs a CNN-LSTM-attention hybrid model to achieve accurate early fault identification, quantify equipment health status, and assess remaining service life, providing a scientific basis for predictive maintenance. This invention can significantly reduce unplanned downtime of power stations, extend the life of critical equipment, reduce operation and maintenance costs and safety risks, and can be easily adapted to compressed air energy storage power stations of different scales and technical routes. It has broad application prospects and comprehensively improves the operation and maintenance level and overall benefits of power stations. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0022] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] This invention constructs an integrated solution that combines virtual and real technologies, is data-driven, provides intelligent early warning, and enables immersive training. It generates massive amounts of fault data through high-fidelity simulation models, trains advanced artificial intelligence algorithms, and ultimately achieves three core functions for compressed air energy storage power stations: fault prediction, health assessment, and personnel training. The specific solution is as follows: like Figure 1 As shown, the present invention provides a fault prediction and health management method for compressed air energy storage power stations, comprising: S1. Establish accurate physical models of key components of the compressed air energy storage power station, including compressor model, air storage device model, expander model, heat exchange system model, and control system model, as the basic simulation environment.

[0025] (1) Compressor model: Based on thermodynamic and fluid dynamic equations, a variable operating condition model of a multi-stage axial / centrifugal compressor is established, which can simulate flow-head characteristics, efficiency MAP, surge boundary, etc.

[0026] (2) Gas storage device model: For salt caverns, abandoned mines or high-pressure storage tanks, establish unsteady-state models that consider geomechanical characteristics and thermodynamics to simulate pressure changes, temperature stratification, gas leakage, etc.

[0027] (3) Expander model: Establish a multi-stage expander model to simulate its work capacity, efficiency characteristics and partial load performance.

[0028] (4) Heat exchange system model: Establish distributed parameter models of regenerator and intercooler to simulate the state of heat exchange efficiency decline and scaling.

[0029] (5) Control system model: Simulate the start-up and shutdown logic, load regulation logic and safety interlock protection system of the power plant.

[0030] S2. Construct a fault simulation injection mechanism to simulate real faults by perturbing parameters and modifying boundary conditions of the precise physical model without changing the core structure of the model. This includes progressive fault simulation and sudden fault simulation, resulting in a fault mode library.

[0031] (1) Progressive Failure Simulation: a. Performance degradation: By gradually modifying the efficiency coefficient, flow coefficient, etc. in the model over time, the model simulates slowly developing faults such as compressor blade fouling and expander flow path wear.

[0032] b. Increased clearance: This simulates the change in rotor dynamic characteristics caused by bearing wear, achieved by altering the mass, stiffness, and damping matrix.

[0033] (2) Simulation of sudden failure: a. Leakage Faults: Introduce additional mass flow outlets in the gas storage chamber or pipeline model to simulate leaks at different locations and sizes.

[0034] b. Sensor failure: Simulate sensor drift, jamming, and decreased accuracy, which can be achieved by modifying the corresponding measurement output value.

[0035] c. Control system failure: jamming of the analog actuator, controller failure, etc.

[0036] S3. Set different severity levels for each fault mode in the fault mode library; wherein the severity levels include minor, moderate and severe.

[0037] Specifically, for each failure mode, different severity levels and different development rates are set, and simulations are performed on different severity levels and different development rates. The results are recorded and compared with data in the failure mode library. Finally, different severity levels and different development rates are set for each failure mode in the failure mode library.

[0038] S4. Under multiple operating conditions, collect multi-source data during the simulation process based on the accurate physical model at a set frequency (e.g., 1kHz), including process parameters, mechanical parameters, electrical parameters, and performance parameters.

[0039] The multi-condition setting is a combination of different operating conditions, covering different load levels (25%, 50%, 75%, 100%), ambient temperature, start-up and shutdown processes, etc.

[0040] Multi-source data specifically includes: (1) Process parameters: pressure, temperature and flow rate at each stage.

[0041] (2) Mechanical parameters: shaft vibration (X / Y direction), shaft displacement, bearing temperature, noise.

[0042] (3) Electrical parameters: motor current, voltage, power.

[0043] (4) Performance parameters: efficiency, pressure ratio, output.

[0044] S5. Extract sensitive features that can characterize the degradation trend of equipment status from the multi-source data, and associate the sensitive features with the corresponding fault labels according to the fault mode library to construct the original dataset for model training.

[0045] S501. Extract sensitive features that can characterize the degradation trend of the device from the multi-source data, including time-domain features, frequency-domain features, time-frequency-domain features, and deep learning features.

[0046] (1) Time-domain characteristics: mean, variance, peak value, kurtosis, waveform factor, etc.

[0047] (2) Frequency domain characteristics: The spectrum is obtained by FFT transformation, and the main frequency, harmonic amplitude, centroid frequency, etc. are extracted.

[0048] (3) Time-frequency domain features: Wavelet transform or empirical mode decomposition (EMD) is used to extract non-stationary signal features such as energy entropy and frequency band energy distribution.

[0049] (4) Deep learning features: One-dimensional convolutional neural network (1D-CNN) is used to automatically learn deep features directly from the original vibration signal.

[0050] S502. Associate all sensitive features with corresponding fault labels according to the fault mode library to construct a large-scale, high-quality raw dataset for AI model training; wherein, the fault labels are extracted from the fault mode library and include fault type, severity, and remaining service life.

[0051] S6. Construct a hybrid deep learning fault prediction model to output fault classification and remaining service life based on the hybrid deep learning fault prediction model.

[0052] The hybrid deep learning fault prediction model includes a 1D-CNN branch, an LSTM branch, an attention mechanism, and an output layer.

[0053] (1) 1D-CNN branch: Input high-dimensional vibration signals (such as mechanical parameters such as shaft vibration and noise) as one-dimensional time series data. Vibration signals are usually sampled at high frequencies (such as 1kHz) and contain rich fault features.

[0054] The structure includes: a. Multiple one-dimensional convolutional layers, using convolutional kernels of different sizes (e.g., 3, 5, 7) to scan the input signal and extract local features (e.g., peaks, waveform abrupt changes). Each convolutional layer is followed by a ReLU activation function to introduce non-linearity. b. Pooling layers: using max pooling or average pooling to reduce feature dimensionality, retaining key information and reducing computation. c. Batch normalization layers: accelerating the training process and improving model stability. d. Fully connected layers: flattening the features after convolution and pooling, converting them into high-level feature vectors.

[0055] (2) LSTM branch: Input: process parameters (such as pressure, temperature, flow rate) and performance parameters (such as efficiency, pressure ratio, output). These parameters are time-varying sequence data with long-term dependence characteristics.

[0056] The structure includes: a. Multiple LSTM layers: Each LSTM unit contains an input gate, a forget gate, and an output gate to control the information flow and remember long-term trends. Bidirectional LSTMs may be used to capture contextual information. b. Dropout layers: Prevent overfitting and improve generalization ability. c. Fully connected layers: Convert the LSTM output into feature vectors.

[0057] (3) The attention mechanism is located after the feature fusion of the 1D-CNN and LSTM branches. The feature vectors of the two branches are concatenated to form a fused feature vector.

[0058] It includes an attention layer that calculates the importance score of each feature (using the softmax function), and then performs a weighted summation of the fused features. The attention weights are dynamically adjusted based on the current input, and multi-head attention is used to capture feature relationships between different subspaces.

[0059] (4) Output Layer: Simultaneously outputs fault classification (what type of fault) and remaining service life regression (how much longer it can operate). Includes: Multi-task learning framework: Simultaneously handles fault classification and RUL regression.

[0060] Fault classification branch: Use the softmax activation function to output the probability distribution of fault types (such as compressor surge, leakage, sensor failure, etc.), including severity levels (minor, moderate, severe).

[0061] RUL regression branch: Uses a linear activation function to output the remaining lifetime value (in hours or number of cycles).

[0062] The specific processing steps are as follows: First, the collected multi-source data is preprocessed and features are extracted. Mechanical parameters such as vibration signals are input into the 1D-CNN branch to automatically learn local fault features, while time-series data such as process parameters and performance parameters are input into the LSTM branch to capture long-term degradation trends. Then, the feature vectors extracted by the two branches are concatenated and dynamically weighted and fused through an attention mechanism, so that the model focuses on the key information most relevant to the current fault. Finally, the fused features are sent to the multi-task output layer, which simultaneously completes the accurate classification of fault type and severity as well as the regression prediction of remaining service life, thereby realizing end-to-end intelligent analysis from multi-source data input to fault warning and life assessment.

[0063] S7. Health status assessment based on multi-source information fusion.

[0064] S701, Health Index Calculation: Integrating multiple sensitive features extracted from multi-source data and deep state representations learned from a hybrid deep learning fault prediction model, and using a fusion algorithm of fuzzy inference or deep belief network, the above multi-dimensional indicators are mapped to a comprehensive health index between 0 and 1. This index quantifies the overall health of the equipment, where 1 represents a brand new state and 0 represents complete failure. S702. Remaining service life prediction: Taking the current health index calculated in S701 and its historical change sequence as input, the particle filter or deep learning sequence prediction model is used to predict the future trajectory of the health index, deduce the time when the equipment reaches the failure threshold, output the predicted value of the remaining service life of the equipment, and provide a confidence interval to characterize the uncertainty of the prediction.

[0065] In addition, after the health status assessment is completed, a simulation training platform can be developed, including designing the platform architecture and training mode, and an intelligent coaching system.

[0066] (1) Platform Architecture. A B / S (Browser / Server) architecture is adopted for easy deployment and access. The front end uses WebGL (such as Three.js) to realize the 3D visualization of the main equipment of the power plant, showing the internal status of the equipment and the dynamic development of faults. The back end adopts a microservice architecture, decoupling the simulation engine, AI analysis engine, user management, scene management and other modules to improve the system's flexibility and scalability.

[0067] (2) Training mode.

[0068] a. Normal operation training mode: Trainees perform standard operations in a virtual environment, such as cold start-up of power plants, normal shutdown, load regulation, etc. The system records the standardization of the operation and scores it.

[0069] b. Fault Reproduction and Diagnosis Mode: The system randomly or the instructor specifies a fault. Students need to analyze and locate the cause of the fault based on the alarm information and parameter changes using the virtual diagnostic tools provided by the system.

[0070] c. Emergency Response Assessment Mode: Simulates emergency failures (such as severe leaks or surges) to assess trainees' rapid response, decision-making, and operational capabilities under pressure. The system provides a comprehensive evaluation based on the order of operations, response time, and effectiveness of the response.

[0071] (3) Intelligent Coaching System. The platform has a built-in expert knowledge base that can provide intelligent guidance during training, including: a. Operation prompts: Provide prompt suggestions when students hesitate or make mistakes in their operations.

[0072] b. Fault Analysis Guidance: Guide trainees to follow a scientific diagnostic process and avoid blind spots.

[0073] c. Automated evaluation report: After the training is completed, a detailed evaluation report is automatically generated, pointing out the strengths and areas for improvement.

[0074] Finally, system deployment and application are carried out, including offline preparation, online deployment, and data integration. Offline preparation: In a laboratory environment, simulation model construction, fault injection, data generation, and AI model training are completed. Online deployment: The trained AI model and simulation platform are deployed to the power plant's on-site data center or cloud platform. Data integration: The model is lightweightly integrated with the power plant's real DCS / SCADA system, utilizing real-time data for model fine-tuning and online early warning. The system analyzes data periodically or in real-time, generating health reports and early warning information to guide maintenance decisions. Maintenance personnel undergo regular training and assessments on the simulation platform to maintain and improve their skill levels.

[0075] In summary, this invention deeply integrates simulation technology, artificial intelligence, and modern training concepts, providing a complete technical solution for the safe, reliable, and efficient operation of compressed air energy storage power stations.

[0076] like Figure 2 As shown, the present invention also provides a fault prediction and health management device for compressed air energy storage power stations, comprising: A module is established to create accurate physical models of key components of a compressed air energy storage power station, including compressor models, air storage device models, expander models, heat exchange system models, and control system models. The simulation module is used to construct a fault simulation injection mechanism to simulate real faults by perturbing parameters and modifying boundary conditions of the precise physical model, including progressive fault simulation and sudden fault simulation, to obtain a fault mode library. The setting module is used to set different severity levels for each fault mode in the fault mode library; wherein the severity levels include minor, moderate and severe. The acquisition module is used to acquire multi-source data, including process parameters, mechanical parameters, electrical parameters and performance parameters, at a set frequency during the simulation process based on the accurate physical model under multiple operating conditions. An extraction module is used to extract sensitive features that can characterize the trend of equipment condition degradation from the multi-source data, and associate the sensitive features with the corresponding fault labels according to the fault mode library to construct the original dataset for model training. The output module is used to construct a hybrid deep learning fault prediction model to output fault classification and remaining service life based on the hybrid deep learning fault prediction model. The assessment module is used for health status assessment based on multi-source information fusion.

[0077] Each of the above modules is used to perform the respective steps in the above-described method for fault prediction and health management of compressed air energy storage power stations. The specific implementation methods are as described in the above-described method embodiments, and will not be repeated here.

[0078] Compared with existing technologies, the fault prediction and health management method and system for compressed air energy storage power stations provided by this invention have the following significant advantages: 1. Solved the problem of scarce fault data.

[0079] This invention generates massive amounts of fault data through high-fidelity simulation: The multiphysics simulation model constructed by this invention can accurately simulate the complete evolution process of a compressed air energy storage power station from normal state to various fault states, generating a rich dataset covering various operating conditions, different fault types, and severity levels. This fundamentally solves the problem of difficult and low-accuracy data-driven model training caused by the scarcity of fault samples in actual operation and maintenance.

[0080] Supports simulation of complex and extreme fault conditions: The system can safely and cost-effectively simulate severe faults and extreme conditions (such as severe surge and deflagration) that are absolutely not allowed to occur in actual power plants, providing irreplaceable data support for studying fault mechanisms and developing emergency response plans.

[0081] 2. Enables early, accurate, and explainable prediction of faults, significantly improving early warning capabilities.

[0082] The hybrid intelligent model has complementary advantages: the CNN-LSTM-attention mechanism hybrid model can effectively extract local fault features of signals such as vibration through CNN, capture the long-term degradation trend of system parameters through LSTM, and focus on key information through attention mechanism, so as to achieve keen perception and accurate identification of early weak fault features.

[0083] High prediction accuracy and foresight: The average early warning time of this invention for typical faults such as compressor surge and bearing wear is more than 48 hours earlier than that of traditional threshold alarm methods, and the fault type identification accuracy reaches 96.2%, which is significantly higher than the prediction method based on a single model (about 85%).

[0084] Provides quantitative assessment of remaining useful life: Not only predicting whether a failure will occur, but also quantitatively assessing the health status (Health Index) of the equipment and predicting the remaining useful life (RUL), providing a direct, quantitative scientific basis for predictive maintenance decisions and avoiding over-maintenance or under-maintenance.

[0085] 3. Build an immersive, efficient, and low-risk training system for operations and maintenance personnel.

[0086] Provides a safe and risk-free fault diagnosis training environment: Trainees can repeatedly practice fault diagnosis and emergency response in a highly realistic virtual environment without worrying about damage to real equipment due to misoperation, which greatly reduces training risks and costs.

[0087] The training modes are diverse and highly targeted: the system supports comprehensive training from normal operation to fault reproduction and emergency assessment, and can customize training content according to the trainees' positions and skill levels, improving training efficiency by more than 50% compared to the traditional shadowing learning model.

[0088] Built-in intelligent coach enables personalized teaching: the system can automatically evaluate trainees' operations, generate detailed competency assessment reports, and intelligently recommend training courses for weak areas, effectively accelerating the growth of operation and maintenance personnel.

[0089] 4. Form a closed-loop empowerment system of "simulation-prediction-training" to comprehensively improve the operation and maintenance level of power plants.

[0090] This system has established a value chain from data to decision-making: it uses simulation-generated data for AI model training, uses the trained models for fault prediction and health management, and solidifies management knowledge into the training platform to improve personnel skills, forming a virtuous cycle of mutual promotion and continuous optimization.

[0091] Significantly improves the economic efficiency and reliability of power plants: By implementing predictive maintenance, unplanned downtime can be reduced by more than 70%, the lifespan of critical equipment can be extended by 15%-20%, and maintenance costs and safety risks can be significantly reduced.

[0092] The system is highly scalable and easy to promote and apply: The modular design and standard data interface adopted by this invention make the system easy to adapt to compressed air energy storage power stations of different scales and different technical routes, and has broad prospects for promotion and application.

[0093] In summary, this invention, through the deep integration of multidisciplinary technologies, solves several key challenges in fault prediction and health management of compressed air energy storage power stations, providing a complete integrated solution for the safe, reliable, efficient, and intelligent operation and maintenance of such power stations. The technological progress is significant, and the overall benefits are outstanding.

[0094] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores all data required for the process of fault prediction and health management methods for compressed air energy storage power plants. The network interface is used for communication with external terminals via a network connection. The computer program is executed by the processor to implement the fault prediction and health management methods for compressed air energy storage power plants.

[0095] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0096] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described fault prediction and health management methods for compressed air energy storage power stations.

[0097] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0098] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0099] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for failure prognosis and health management of a compressed air energy storage power plant, characterized in that, The method comprises the following steps: S1, establishing an accurate physical model of key components of a compressed air energy storage power station, including a compressor model, a gas storage device model, an expander model, a heat exchange system model, and a control system model; S2, constructing a fault simulation injection mechanism to simulate real faults by parameter perturbation and boundary condition modification of the accurate physical model, including gradual fault simulation and sudden fault simulation, to obtain a fault mode library; S3, setting different severity levels for each fault mode in the fault mode library; wherein the severity levels include slight, moderate, and severe; S4, collecting multi-source data during the simulation process based on the accurate physical model under multiple working conditions at a set frequency, including process parameters, mechanical parameters, electrical parameters, and performance parameters; S5, extracting sensitive features that can represent the degradation trend of the equipment state from the multi-source data, and associating the sensitive features with corresponding fault labels according to the fault mode library to construct an original data set for model training; S6, constructing a hybrid deep learning fault prediction model to output fault classification and remaining useful life according to the hybrid deep learning fault prediction model; S7, performing health state evaluation based on multi-source information fusion.

2. The method for prognostics and health management of compressed air energy storage power plants according to claim 1, characterized in that, In S1, the compressor model is based on thermodynamic and fluid mechanics equations to establish a variable working condition model of a multi-stage axial / centrifugal compressor to simulate flow-head characteristics, efficiency MAP, and surge boundary; The gas storage device model is for salt caverns, abandoned mines, or high-pressure storage tanks, and establishes a non-steady-state model considering geomechanical properties and thermodynamics to simulate pressure changes, temperature stratification, and gas leakage; The expander model is a multi-stage expander model to simulate its power output, efficiency characteristics, and part-load performance; The heat exchange system model is a distributed parameter model of regenerators and intercoolers to simulate heat exchange efficiency decline and fouling state; The control system model simulates the start-stop logic, load regulation logic, and safety interlock protection system of the power station.

3. The method for prognostics and health management of compressed air energy storage power plants according to claim 1, characterized in that, In S2, the gradual fault simulation includes performance degradation fault and gap increase fault, the performance degradation fault modifies the efficiency coefficient and flow coefficient in the accurate physical model over time to simulate the slow development of compressor blade fouling and expander flow passage wear; the gap increase fault simulates the change of rotor dynamic characteristics caused by bearing wear by changing the mass, stiffness, and damping matrix; The sudden fault simulation includes leakage fault, sensor fault, and control system fault, the leakage fault introduces an additional mass flow outlet in the gas chamber or pipeline model to simulate different locations and sizes of leakage; the sensor fault simulates sensor drift, sticking, and accuracy decline by modifying the corresponding measurement output value; the control system fault simulates actuator sticking and controller failure.

4. The method for prognostics and health management of compressed air energy storage power plants of claim 1, wherein, In S4, the process parameters include stage pressure, temperature, and flow rate, the mechanical parameters include X / Y direction shaft vibration, shaft displacement, bearing temperature, and noise, the electrical parameters include motor current, voltage, and power, and the performance parameters include efficiency, pressure ratio, and output.

5. The method for prognostics and health management of compressed air energy storage power plants of claim 1, wherein, The S5 specifically includes: S501, extracting sensitive features capable of representing equipment state degradation trends from the multi-source data, including time domain features, frequency domain features, time-frequency domain features and deep learning features; wherein the time domain features are mean, variance, peak value, kurtosis, waveform factor, the frequency domain features are frequency spectrum obtained by FFT transformation, and the main frequency, frequency amplitude and center of gravity frequency are extracted; the time-frequency domain features are non-stationary signal features such as energy entropy and frequency band energy distribution extracted by wavelet transform or empirical mode decomposition; and the deep learning features are automatically learned from the original vibration signal by using a one-dimensional convolutional neural network; S502, associating all sensitive features with corresponding fault labels according to the fault mode library to construct an original data set for model training; wherein the fault labels are extracted from the fault mode library, including fault type, severity and remaining useful life.

6. The method for prognostics and health management of compressed air energy storage power plants of claim 1, wherein, In the S6, the mixed deep learning fault prediction model includes a 1D-CNN branch, an LSTM branch, an attention mechanism and an output layer; The mixed deep learning fault prediction model pre-processes and extracts features from the collected multi-source data, wherein the mechanical parameters input the 1D-CNN branch to automatically learn local fault features, and the process parameters and performance parameters and other time series data input the LSTM branch to capture long-term degradation trends; the feature vectors extracted by the two branches are spliced and dynamically weighted and fused by the attention mechanism, so that the model focuses on the key information most relevant to the current fault; the fused features are input into the multi-task output layer to simultaneously complete accurate classification of fault type and severity and regression prediction of remaining useful life, so as to realize analysis from multi-source data input to fault warning and life assessment.

7. The method for prognostics and health management of compressed air energy storage power plants of claim 1, wherein, The S7, health state evaluation based on multi-source information fusion specifically includes: S701, health index calculation: integrating multiple sensitive features extracted from multi-source data and deep state representations learned from the mixed deep learning fault prediction model, using a fuzzy reasoning or deep belief network fusion algorithm to map the above multi-dimensional indexes to a comprehensive health index between 0 and 1, which quantifies the overall health of the equipment, wherein 1 represents a brand new state and 0 represents complete failure; S702, remaining useful life prediction: taking the current health index calculated in S701 and its historical change sequence as input, using particle filtering or deep learning sequence prediction model to predict the future trajectory of the health index, deducing the time when the equipment reaches the failure threshold, outputting the remaining useful life prediction value of the equipment, and providing a confidence interval to represent the uncertainty of the prediction.

8. A device for failure prognosis and health management of a compressed air energy storage power plant, characterized in that It includes: A building module for building an accurate physical model of key components of a compressed air energy storage power station, including a compressor model, a gas storage device model, an expander model, a heat exchange system model, and a control system model; A simulation module for constructing a fault simulation injection mechanism to simulate real faults by parameter perturbation and boundary condition modification on the accurate physical model, including gradual fault simulation and sudden fault simulation, to obtain a fault mode library; The setting module is configured to set different severity levels for each fault mode in the fault mode library; wherein the severity levels include slight, medium, and severe; The acquisition module is configured to acquire multi-source data based on the precise physical model during the simulation process at a set frequency under multiple working conditions, including process parameters, mechanical parameters, electrical parameters, and performance parameters; The extraction module is configured to extract sensitive features capable of representing the degradation trend of the equipment state from the multi-source data, and associate the sensitive features with corresponding fault labels according to the fault mode library to construct an original data set for model training; The output module is configured to construct a hybrid deep learning fault prediction model to output fault classification and remaining useful life according to the hybrid deep learning fault prediction model; The evaluation module is configured to perform health state evaluation based on multi-source information fusion. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images