Device anomaly sample generation method and device, device anomaly early warning method and device anomaly early warning device

Abstract

The application provides a device anomaly sample generation method and device, and a device anomaly early warning method, comprising: identifying the working condition of historical monitoring data of a device, constructing a physically feasible region of device operation based on a physical model and the identified working condition mode; within the feasible region, a sample generation strategy that fuses physical simulation and data-driven model generation is used to generate a target abnormal sample set; an abnormal sample set is applied to train a device anomaly early warning model, the sample generation strategy is feedback optimized based on the performance index of the trained model, the steps of generating a target abnormal sample set and feedback optimization are continuously executed until the model performance index meets the requirements, and an optimized device anomaly sample set is obtained. The application can generate high-quality device anomaly samples within the constructed physically feasible region, and the sample generation process is feedback adjusted through the performance of the device anomaly early warning model to obtain optimized device anomaly samples, thereby improving the generalization ability of the device anomaly early warning model.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology, and in particular to a method, device and method for generating equipment anomaly samples and an early warning method for equipment anomalies. Background Technology

[0002] In the field of industrial equipment operation monitoring and anomaly early warning, the performance of equipment anomaly early warning models is highly dependent on the quantity, quality, and coverage of historical monitoring data. However, in practical engineering applications, the following problems commonly exist: 1. Uneven distribution of historical monitoring samples During long-term operation, industrial equipment is in normal or routine operating conditions most of the time, resulting in a small number of abnormal operating condition samples. This leads to a serious imbalance in abnormal categories, and the early warning model does not learn enough about abnormal patterns.

[0003] 2. Incomplete coverage of operating conditions. Due to the influence of equipment operation strategies, safety restrictions and environmental conditions, some extreme or rare operating conditions are almost non-existent in historical data, but may still occur in actual operation. Existing models have poor generalization ability under such conditions.

[0004] 3. Purely data-driven sample generation methods lack credibility. In existing technologies, samples are often expanded by means of oversampling, noise perturbation or generative adversarial networks, but such methods usually lack physical constraints and the generated data may violate the basic physical laws of device operation, affecting the stability and reliability of model training.

[0005] 4. Separation between physical simulation and algorithm model Although dynamic or fluid dynamic models can describe the operating mechanism of equipment, existing technologies mostly treat physical simulation results as an independent analysis method, without forming an effective collaborative and closed-loop optimization mechanism with equipment anomaly early warning models. Summary of the Invention

[0006] The purpose of this application is to provide a method, apparatus and method for generating equipment anomaly samples, which can generate high-quality equipment anomaly samples within the constructed physical feasible domain, and adjust the sample generation process by feedback through the performance of the equipment anomaly early warning model to obtain optimized equipment anomaly samples, so as to alleviate the problems of uneven sample distribution, incomplete operating condition coverage and insufficient generalization ability of the equipment anomaly early warning model.

[0007] Firstly, this application provides a method for generating equipment anomaly samples. The method includes: identifying operating conditions from historical monitoring data of the equipment to determine multiple operating condition modes; constructing a physical feasible domain for equipment operation based on a physical model and the identified operating condition modes; the physical feasible domain is used to define the set of all operating states of the equipment that occur under specified operating conditions and conform to physical laws and safety requirements; within the physical feasible domain, a sample generation strategy that integrates physical simulation and data-driven generation model is adopted to generate a target anomaly sample set; the samples in the target anomaly sample set are used to train an equipment anomaly early warning model, and the sample generation strategy is optimized based on the performance indicators of the trained model. The steps of generating the target anomaly sample set and feedback optimization are continued until the performance indicators of the model meet the requirements, resulting in an optimized equipment anomaly sample set.

[0008] Furthermore, the steps described above for identifying operating conditions from historical equipment monitoring data and determining multiple operating condition modes include: acquiring historical equipment monitoring data; constructing a three-dimensional time series tensor based on the historical equipment monitoring data; the three dimensions of the three-dimensional time series tensor include the number of historical samples, the time step corresponding to each sample, and the dimensions of the equipment monitoring parameters; performing unsupervised operating condition clustering on the three-dimensional time series tensor to identify typical operating condition modes of equipment operation; determining the joint probability distribution of all monitoring parameters under each identified operating condition mode; and identifying operating condition modes with a joint probability distribution less than a threshold as rare operating condition modes.

[0009] Furthermore, the steps described above for constructing the physical feasible domain of equipment operation based on the physical model and the identified operating conditions include: for each operating condition, extracting a set of key parameters that determine the macroscopic operating state of the equipment to form an operating condition parameter vector; simulating the operating condition parameter vector using a high-fidelity parametric physical model to output key physical response indicators; the key physical response indicators include: maximum vibration displacement, isentropic efficiency, pressure pulsation coefficient, and critical speed margin; based on the key physical response indicators, constructing a constraint set consisting of multiple physical constraints as the physical feasible domain of equipment operation; the physical constraints include equality constraints and inequality constraints; and simplifying and visualizing the physical feasible domain.

[0010] Furthermore, the steps described above for generating a target anomalous sample set within the physically feasible domain using a sample generation strategy that integrates physical simulation and a data-driven generative model include: generating a first anomalous sample set based on constraints of the physically feasible domain through physical simulation, generating a second anomalous sample set through a physical constraint generative adversarial network; and filtering samples based on the first anomalous sample set to obtain the target anomalous sample set.

[0011] Furthermore, the steps described above for generating the first abnormal sample set through physical simulation based on the constraints of the physically feasible region include: determining a normal reference point within the physically feasible region; the normal reference point is the sample corresponding to the highest efficiency and best stability state; pre-defining a set of fault development direction vectors based on fault dynamics knowledge; and performing parameter scanning simulation along each fault development direction vector, starting from the normal reference point, with gradually increasing fault severity parameters to generate a continuous fault development sequence from normal to abnormal and from mild to severe, thus obtaining the first abnormal sample set; wherein each sample in the continuous fault development sequence carries a fault type label and a severity label.

[0012] Furthermore, the aforementioned physical constraint generative adversarial network includes a generator and a discriminator. The generator's inputs include random noise, a condition vector, operating parameters, and boundary parameters. The boundary parameters are used to input the boundary information of the physical feasible region into the generator to guide the generation process. The discriminator is used to determine whether a sample is within the physical feasible region. During the generator's training process, the model loss value is determined based on the adversarial loss and the physical consistency loss. The physical consistency loss forces the samples generated by the generator to satisfy the physical equality constraints without violating the inequality constraints.

[0013] Furthermore, the steps described above for selecting samples based on the first and second abnormal sample sets to obtain the target abnormal sample set include: merging the samples in the first and second abnormal sample sets to obtain a first candidate sample set; performing a physical constraint re-examination on the samples in the first candidate sample set, removing all samples that violate the physical feasible region constraint, to obtain a second candidate sample set; calculating the Mahalanobis distance between each candidate sample in the second candidate sample set and the corresponding historical data of the working condition, removing samples whose Mahalanobis distance exceeds a threshold, to obtain a third candidate sample set; and using the farthest point sampling algorithm, selecting a subset from the third candidate sample set that is evenly distributed in the feature space and can cover different abnormal patterns to the greatest extent, as the target abnormal sample set.

[0014] Furthermore, the above-mentioned steps for feedback optimization of the sample generation strategy based on the performance metrics of the trained model include: evaluating the model's performance metrics using multiple specified fault validation sets after each round of model training; performance metrics include precision, recall, and F1 score; identifying model performance weaknesses based on specified fault validation sets where performance metrics are below a threshold; model performance weaknesses include at least one of the following: low specified fault recognition rate, insufficient generalization ability in extreme operating conditions, and insufficient differentiation between minor and serious faults; and automatically adjusting the sample generation strategy for the next iteration based on the identified model performance weaknesses, specifically including: targeting specified... To address the low fault identification rate, simulation vectors for specified fault modes are added to the physical simulation path, and the parameter scanning resolution is improved. In the generative adversarial model path, more corresponding abnormal samples are generated in the condition vector. To address the insufficient generalization ability in extreme operating conditions, more intensive simulations are performed in the neighborhood of operating condition parameters during the construction of the physical feasible region to refine the feasible region boundary. The generator generates more samples simultaneously under this extreme condition to ensure more comprehensive coverage of extreme states. To address the insufficient distinction between minor and severe faults, the distribution of fault severity parameters in sample generation is adjusted, and the sample density is increased in the transition region to ensure that abnormal samples of different severity levels are adequately trained.

[0015] Secondly, this application also provides an equipment anomaly sample generation device, comprising: a feasible domain construction module, used to identify operating conditions from historical monitoring data of the equipment and determine multiple operating condition modes; based on a physical model and the identified operating condition modes, constructing a physical feasible domain for equipment operation; the physical feasible domain is used to define the set of all operating states of the equipment that occur under specified operating conditions and conform to physical laws and safety requirements; a sample generation module, used to generate a target anomaly sample set within the physical feasible domain using a sample generation strategy that integrates physical simulation and data-driven generation model; and a strategy optimization module, used to apply samples from the target anomaly sample set to train an equipment anomaly early warning model, perform feedback optimization of the sample generation strategy based on the performance indicators of the trained model, and continue to execute the steps of generating the target anomaly sample set and feedback optimization until the performance indicators of the model meet the requirements, thereby obtaining an optimized equipment anomaly sample set.

[0016] Thirdly, this application also provides a method for early warning of equipment anomalies, the method comprising: acquiring real-time equipment operation monitoring data; inputting the real-time equipment operation monitoring data into an equipment anomaly early warning model for prediction, and obtaining a corresponding anomaly early warning result; wherein, the equipment anomaly early warning model is obtained by training an optimized equipment anomaly sample set generated by the equipment anomaly sample generation method as described in the first aspect. The equipment anomaly sample generation method, apparatus, and equipment anomaly early warning method provided in this application first identify the operating conditions of historical equipment monitoring data. Then, based on a physical model and the identified multiple operating condition modes (including rare operating condition modes), a physically feasible domain for equipment operation is constructed. This physically feasible domain defines the set of all operating states of the equipment that occur under specified operating conditions and conform to physical laws and safety requirements. Within the physically feasible domain, a sample generation strategy that integrates physical simulation and data-driven generation models is used to generate a target anomaly sample set. Finally, the samples in the target anomaly sample set are used to train the equipment anomaly early warning model. Based on the performance indicators of the trained model, the sample generation strategy is optimized through feedback. The steps of generating the target anomaly sample set and feedback optimization are continued until the model's performance indicators meet the requirements, resulting in an optimized equipment anomaly sample set. This method can generate high-quality equipment anomaly samples within the constructed physically feasible domain and adjust the sample generation process through feedback adjustment of the equipment anomaly early warning model performance to obtain optimized equipment anomaly samples, thereby further improving the generalization ability of the equipment anomaly early warning model. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for generating device anomaly samples provided in this application embodiment; Figure 2 A structural block diagram of a device for generating device anomaly samples provided in an embodiment of this application; Figure 3 A flowchart of a device anomaly early warning method provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Existing technologies for generating equipment anomaly samples either rely on historical monitoring data for anomaly modeling, making it difficult to cover rare or unknown operating conditions; or they rely on physical simulation analysis of equipment mechanisms, making it difficult to form a collaborative optimization mechanism with equipment anomaly early warning models; or they generate anomaly samples through data-driven methods, but lack physical constraints, resulting in insufficient reliability and engineering applicability of the generated samples.

[0021] Therefore, there is an urgent need for a technical solution that can generate high-quality abnormal samples within a physically feasible range and adjust the sample generation process through feedback adjustment of the equipment anomaly early warning model performance, in order to solve the problems of uneven sample distribution, incomplete operating condition coverage, and insufficient generalization ability of equipment anomaly early warning model.

[0022] Based on this, this application provides a method, apparatus and method for generating equipment anomaly samples. To facilitate understanding of this embodiment, the method for generating equipment anomaly samples disclosed in this application will be described in detail first. Figure 1 A flowchart of a method for generating device anomaly samples provided in this application embodiment is shown. The method specifically includes the following steps: Step S102: Identify the operating conditions of the equipment's historical monitoring data and determine multiple operating condition modes; based on the physical model and the identified operating condition modes, construct the physical feasible domain for equipment operation; the physical feasible domain is used to define the set of all operating states of the equipment that occur under a specified operating condition and that conform to physical laws and safety requirements. Step S104: Within the physically feasible domain, a sample generation strategy that integrates physical simulation and data-driven generative model is adopted to generate a target anomaly sample set. Step S106: Use samples from the target abnormal sample set to train the device abnormal early warning model. Based on the performance indicators of the trained model, perform feedback optimization on the sample generation strategy. Continue to execute the steps of generating the target abnormal sample set and feedback optimization until the performance indicators of the model meet the requirements and obtain the optimized device abnormal sample set.

[0023] The equipment anomaly sample generation method provided in this application is based on the core concept of constructing an intelligent closed-loop process of "physical-data dual-drive, perception-generation-decision-optimization". This method uses the key concept of physical feasible domain to deeply couple mechanism-based physical simulation with statistical data-driven generation model. While ensuring the physical credibility of the generated samples, it effectively expands the sample coverage to various abnormal working conditions. Based on the performance feedback of the anomaly warning model, it dynamically optimizes the generation strategy, and finally significantly improves the generalization ability and reliability of the anomaly warning model under actual complex working conditions.

[0024] This application also provides another method for generating equipment anomaly samples, which is implemented based on the above embodiments. In this embodiment, the working condition determination process, feasible domain construction process, sample generation process, and strategy optimization process are described in detail.

[0025] In actual industrial settings, equipment often operates under a variety of different conditions, and the normal operating status and abnormal modes vary significantly under these conditions. Therefore, a refined understanding and classification of historical data to determine operating conditions is a crucial foundation for all subsequent steps.

[0026] That is, step S102 above, which involves identifying the operating conditions of historical equipment monitoring data and determining multiple operating mode patterns, includes: (1) Obtain historical monitoring data of the equipment; construct a three-dimensional time series tensor based on the historical monitoring data of the equipment; the three dimensions of the three-dimensional time series tensor include the number of historical samples, the time step corresponding to each sample, and the dimensions of the equipment monitoring parameters; In practice, the acquired historical monitoring data of the equipment is organized into a three-dimensional time-series tensor. The dimensions have the following meanings: : Total number of historical samples, representing the number of all data segments collected (e.g., in 10-minute segments).

[0027] The time step contained in each data segment is used to capture dynamic processes.

[0028] The dimensions of the monitoring parameters typically include various state parameters that can be directly measured or indirectly calculated, such as vibration acceleration (multi-directional), rotational speed, load, pressure, temperature, and flow rate.

[0029] (2) Perform unsupervised clustering of three-dimensional time series tensors to identify typical operating conditions of equipment; Considering that operating condition labels are often missing or incomplete in industrial data, this embodiment uses an unsupervised clustering algorithm to automatically identify typical operating condition patterns of equipment. The Gaussian Mixture Model (GMM) algorithm is preferred, as it can effectively handle high-dimensional, non-spherical distributed data.

[0030] In this embodiment, all historical data is divided into clusters. A mutually exclusive cluster of operating conditions: .

[0031] Each operating condition cluster This represents a stable operating mode, such as "stable operation under rated load", "low load operation", "start-up process", "shutdown process", etc.

[0032] (3) For each identified working condition mode, determine the joint probability distribution of all monitoring parameters under the working condition mode; For each identified cluster of operating conditions Calculate its prior probability This refers to the frequency of this operating condition occurring in historical operations. Estimate the joint probability distribution of all monitored parameters under this operating condition. It is usually approximated by a multidimensional Gaussian distribution or kernel density estimation.

[0033] (4) The working conditions with a joint probability distribution less than the threshold are identified as rare working conditions.

[0034] Rare operating conditions are defined as operating states that occur with extremely low frequency in historical data. The criteria for determining them are as follows:

[0035] in, For a preset threshold (e.g.) (or smaller). The identified rare operating conditions will be the focus of subsequent sample expansion and model performance validation, because the early warning model is most likely to fail under these conditions.

[0036] The physically feasible region is a key technical aspect of the embodiments of this application. It is strictly defined as the set of all operating states of the equipment that occur under specified operating conditions and conform to physical laws and safety requirements. Its construction process is a refined process that integrates physical simulation, engineering knowledge, and data fitting.

[0037] Furthermore, in step S102 above, the step of constructing the physically feasible domain for equipment operation based on the physical model and the identified operating conditions includes: (1) For each working condition mode, extract a set of key parameters that determine the macroscopic operating status of the equipment to form a working condition parameter vector; For the target operating condition cluster From its data, a set of key parameters that determine the macroscopic operating status of the equipment are extracted to form an operating condition parameter vector:

[0038] For a rotating machine (such as a centrifugal compressor). It can include: Rotor speed; : Load torque; Import pressure and temperature; : Critical gap value.

[0039] (2) The working condition parameter vector is simulated by a high-fidelity parametric physical model, and the key physical response indicators are output. The key physical response indicators include: maximum vibration displacement, isentropic efficiency, pressure pulsation coefficient, and critical speed margin. Calling a high-fidelity parametric physical model built for a specific device type ,in, These are the model's inherent parameters (such as mass, stiffness, damping matrix, and geometric dimensions). By changing... The controllable parameters or preset disturbances (simulated faults) are introduced to perform large-scale simulation calculations.

[0040] The simulation output consists of a series of key physical response indicators: .

[0041] These indicators are the direct basis for judging whether the equipment status is "feasible," such as: maximum vibration displacement. Determined by rotor dynamics and bearing characteristics; isentropic efficiency Determined by the aerodynamic design of the flow passage components; pressure pulsation coefficient : Reflects flow stability; critical speed margin : Ensure that you stay away from the safety boundary of resonance.

[0042] (3) Based on key physical response indicators, construct a constraint set consisting of multiple physical constraints as the physical feasible domain for equipment operation; physical constraints include equality constraints and inequality constraints. Based on simulation results and engineering specifications, the physically feasible region Formalized as a set of states satisfying a series of physical constraints, such as:

[0043] in, Indicates the first Operating parameters The physical feasible region is used to characterize the set of system states that satisfy all physical constraints. Represents the system state vector. As a state space dimension, the state vector consists of multiple physical quantities describing the operating state of the device, including but not limited to flow rate, vibration amplitude, pressure, temperature, and rotational speed. Indicates the first Each operating condition parameter vector is used to describe the operating conditions or input conditions of the equipment, including speed, inlet flow rate, inlet pressure, inlet temperature, and control parameters, etc. Indicates the first Inequality constraint functions, ,in This is a set of inequality constraint indices, which are used to describe the safety, stability, and performance boundary conditions of the system. Indicates the first An equality constraint function, ,in This is a set of equality constraint indices, which describe the physical conservation relationships or equilibrium conditions that the system must satisfy. and These represent the sets of indices for inequality constraints and equality constraints, respectively.

[0044] Example 1 (Vibration safety constraints): .in From state The effective value of vibration calculated in the middle, This is the maximum permissible vibration value under this operating condition.

[0045] Example 2 (efficiency constraint): .in, Indicates the system state The calculated isentropic efficiency, Indicates the current operating condition The preset efficiency threshold is set. When When the above constraints are not met, the corresponding state is judged as an abnormally inefficient state and does not belong to the physical feasible region.

[0046] Equality constraints This represents a physical conservation law or equilibrium relationship that must be strictly satisfied.

[0047] Example (power balance): .in, Indicates the system in state The input power at the following levels, Indicates output power. This represents the power loss caused by factors such as friction, leakage, and heat dissipation during system operation. The above constraints are used to ensure that the system satisfies the law of conservation of energy; the corresponding state is considered physically feasible only when the input power equals the sum of the output power and the power loss.

[0048] (4) Simplify and visualize the physical feasible domain.

[0049] In practical applications, for clarity, the high-dimensional feasible domain is often projected onto a subspace consisting of 2-3 key monitoring parameters.

[0050] For example, in "traffic" -vibration "On the plane, feasible region" It appears as a convex polygonal region bounded by multiple straight lines:

[0051] in, Indicates system traffic. Indicates the vibration amplitude; , and They represent the first The linear coefficients corresponding to the boundary constraints; This represents the number of constraints that constitute the boundary of the convex polygon. The coefficient... The convex hull representation can be obtained by performing convex hull calculations on a large number of physical simulation data points or by using linear fitting methods. This representation can be used to approximate the projected boundary of the original high-dimensional feasible region in a two-dimensional subspace. The two-dimensional feasible region representation obtained in this way has good geometric intuition and can be used for rapid boundary determination and feasibility screening in subsequent state generation or optimization processes.

[0052] Within the framework of the physically feasible domain, the embodiments of this application adopt a strategy of parallel and collaborative work of physical simulation and data-driven generation to balance the physical accuracy of the samples and the diversity of patterns.

[0053] Specifically, step S104 above, which involves generating a target anomaly sample set within the physically feasible domain using a sample generation strategy that integrates physical simulation and data-driven generative models, includes: (1) Based on the constraints of the physical feasible region, the first abnormal sample set is generated through physical simulation, and the second abnormal sample set is generated through physical constraint generative adversarial network. Path 1: Targeted anomaly sample generation based on physical simulation. This path focuses on simulating the development process of specific faults from a mechanistic perspective to generate anomaly samples with clear physical meaning and causal relationships.

[0054] The steps described above, which involve generating the first set of anomalous samples through physical simulation based on constraints within the physically feasible region, include: (1.1) Within the physically feasible region, determine the normal reference point; the normal reference point is the sample corresponding to the highest efficiency and best stability state; In practical implementation, within the physically feasible region Within, find the state corresponding to the highest efficiency, optimal stability, or design point. .

[0055] (1.2) Based on the knowledge of fault dynamics, a set of fault development direction vectors are predefined; Based on fault dynamics, a set of fault development direction vectors are predefined. .

[0056] Mass imbalance fault. Its characteristics include increased amplitude and phase change of power frequency vibration.

[0057] Shaft misalignment fault. Its characteristic is a significant increase in the second harmonic vibration component.

[0058] Fault caused by friction between moving and stationary parts. Its characteristic is the presence of abundant high-order harmonics and subharmonics in the frequency spectrum.

[0059] (1.3) Along each fault development direction vector, starting from the normal reference point, perform parameter scanning simulation with gradually increasing fault severity parameters to generate a continuous fault development sequence from normal to abnormal and from mild to severe, and obtain the first abnormal sample set; wherein, each sample in the continuous fault development sequence is labeled with fault type and severity.

[0060] In practice, along each fault direction, starting from the baseline point, the fault severity parameter is increased progressively. Perform a "parameter scan" simulation.

[0061]

[0062] This process generates a continuous fault progression sequence from normal to abnormal, and from minor to severe. Each sample comes with a precise fault type label. Severity labels .

[0063] Path Two: Multi-sample generation based on Physically Constrained Generative Adversarial Networks (PcGANs). This path focuses on learning from the data distribution level and generating more diverse, potentially unseen but physically plausible, anomaly patterns to compensate for the limitations of physical simulation in terms of coverage. The key lies in deeply embedding physical feasible domain constraints into the training process of the generative model.

[0064] The aforementioned physical constraint generative adversarial network includes a generator and a discriminator; A. Generator The input includes: random noise Provides generative diversity; conditional vectors : Specifies the expected anomaly type and approximate severity; operating parameters : Ensure the generated samples match the current operating conditions; boundary parameters Boundary parameters are used to store the boundary information of the physically feasible region (as described above). Input the value into the generator to guide the generation process.

[0065] B. The discriminator is used to determine whether a sample is within the physically feasible region; in this embodiment, a dedicated neural network is trained. To quickly identify any sample Is it within the physically feasible region? The discriminator is trained on a large number of physically simulated samples (feasible) and deliberately constructed violation samples (infeasible). After generation, only those samples that are... Only samples deemed "feasible" with high confidence will be retained.

[0066] C. During the generator training process, the model loss value is determined based on the adversarial loss and the physical consistency loss; the physical consistency loss forces the samples generated by the generator to satisfy the physical equality constraint and not violate the inequality constraint.

[0067] In practical implementation, in the generator In addition to traditional adversarial losses, the training objectives include... An additional physical consistency loss is added. :

[0068]

[0069] in, Represents the input noise vector. Represents a condition variable. This indicates that the generator operates under given parameters. The system state generated below; Indicates the first An equality constraint function, Indicates the first Inequality constraint functions; Represents a linear rectifier function used to handle violations of inequality constraints (i.e., Punishment will be imposed on the portion thereof.

[0070] Through the above construction, the physical consistency loss corresponds to the equality constraint. The deviation is penalized by square, and only in inequality constraints When a violation is detected, a penalty term is introduced to force the generator to produce samples that satisfy physical conservation relations and are within the physical feasible region, thereby improving the physical rationality and engineering usability of the generated results.

[0071] (2) Based on the first abnormal sample set and the first abnormal sample set, sample screening is performed to obtain the target abnormal sample set.

[0072] (2.1) Merge the samples in the first abnormal sample set and the second abnormal sample set to obtain the first candidate sample set; Preliminary samples generated from two paths and After merging, a rigorous fusion screening process is still required to ensure the final expanded sample set. High quality.

[0073] (2.2) Perform physical constraint re-examination on the samples in the first candidate sample set, remove all samples that violate the physical feasible region constraint, and obtain the second candidate sample set; Using a physical discriminator Alternatively, the constraint function value can be directly calculated to eliminate all samples that violate the physical feasible region constraints. This is the last line of defense to ensure the reliability of the samples.

[0074] (2.3) Calculate the Mahalanobis distance between each candidate sample and the corresponding historical data of the working condition in the second candidate sample set, and remove the samples whose Mahalanobis distance exceeds the threshold to obtain the third candidate sample set; To avoid generating overly bizarre samples, each candidate sample is calculated. Historical data corresponding to the operating conditions Mahalanobis distance:

[0075] in and This is the mean and covariance matrix of the historical data for this operating condition. Discard those that are too far apart (e.g., ) samples, A leniency factor slightly greater than 1 (e.g., 1.5) is used to allow the generated samples to deviate moderately from the historical distribution.

[0076] (2.4) Using the farthest point sampling algorithm, select a subset from the third candidate sample set that is evenly distributed in the feature space and can cover different anomaly patterns to the greatest extent. This serves as the target set of abnormal samples, ensuring the comprehensiveness of the training data.

[0077] The training and iterative feedback optimization mechanism for the anomaly warning model is as follows: Model training: Utilizing a high-quality, high-coverage augmented sample set It can train a high-performance device anomaly early warning model. The model can be a deep neural network (such as LSTM, Transformer), an ensemble learning model, or a combination thereof. The training objective is to minimize the regularized cross-entropy loss as is standard.

[0078] After model training, the sample generation strategy is optimized based on the performance metrics of the trained model, specifically including: (1) After each round of model training, evaluate the model’s performance metrics using multiple specified failure validation sets; the performance metrics include precision, recall and F1 score; In practice, after each round of model training, the model is first validated on an independent validation set, especially a set of rare test cases. The system calculates Recall, Precision, and F1 score for each type of failure.

[0079] (2) Based on the specified fault verification set with performance indicators less than the threshold, identify the weak points of model performance; the weak points of model performance include at least one of the following: low specified fault identification rate, insufficient generalization ability in extreme working condition range, and insufficient differentiation between minor faults and serious faults. In this embodiment, a comprehensive performance index can be defined:

[0080] Normal settings Furthermore, it emphasizes recall, i.e., reducing false negatives, which is particularly important for industrial safety early warning. Simultaneously, by analyzing the confusion matrix of the model across different fault types and severity ranges, it accurately identifies the weak points in the model's performance.

[0081] The system analyzes prediction performance under different fault types and severity levels using a confusion matrix (a matrix used to statistically correlate predicted and actual classes). Any model with a recall or F1-score below a preset threshold for a particular fault type or severity range is considered a weak point. These weak points are recorded and used to guide the next round of sample generation and training optimization.

[0082] (3) Based on the identified weaknesses in model performance, the system automatically adjusts the sample generation strategy for the next iteration, specifically including: To address the low recognition rate of specific faults (such as the rare fault "bearing cage fracture"), we add simulation vectors for the specified fault modes and improve parameter scanning resolution in the physical simulation path. In the generative adversarial model path, we generate more corresponding abnormal samples in the condition vector. To address the insufficient generalization capability in extreme operating conditions (such as ultra-low flow rates), the operating parameters should be adjusted during the construction of the physical feasible domain. The neighborhood is simulated more densely to refine the feasible region boundary; the generator generates more samples simultaneously under this extreme condition to ensure more comprehensive coverage of extreme states. To address the insufficient differentiation between minor and severe faults, the fault severity parameter in sample generation was adjusted. The distribution of samples is adjusted, and the sample density is increased in the transition region to ensure that abnormal samples of different severity levels are adequately trained.

[0083] In addition, it includes an adaptive correction process for the physical feasible region parameters: the closed-loop feedback also adjusts key parameters of the physical model, including the safety margin. Adjusting and updating the training sample density of the proxy physics model, as well as the operating parameters. The neighborhood coverage is expanded, thereby enabling bidirectional optimization of physical simulation and data-driven generative models.

[0084] Through the above measures, the sample generation strategy and the anomaly warning model continuously reinforce each other during the closed-loop iteration process, thereby continuously improving the model's identification ability under rare, extreme, and different severity conditions, while ensuring the physical credibility and engineering feasibility of the generated samples.

[0085] The adjusted generation strategy produces new, more targeted augmented samples for retraining or fine-tuning the anomaly warning model. This process is then evaluated again, and further adjustments are made based on feedback. This cycle repeats, forming a continuous "generation-evaluation-optimization" closed loop. As iterations continue, the sample generation process increasingly "targets" the model's actual needs, allowing the warning model's generalization ability to be continuously and purposefully improved until performance meets requirements, such as multiple evaluation metrics exceeding thresholds.

[0086] The method provided in this application is applicable to the monitoring and early warning of abnormalities in the operation of complex industrial equipment such as water turbines and rotating machinery. While ensuring physical consistency, it expands the abnormal sample and rare operating condition sample, improving the generalization ability and reliability of the abnormality early warning model under unknown and extreme operating conditions. Compared with the prior art, the embodiments of this application have the following beneficial effects: 1. By constructing a physically feasible region, invalid samples that violate physical laws are effectively avoided, thus improving sample credibility; 2. Expand anomalous and rare operating condition samples under physical constraints to alleviate the problems of uneven sample distribution and incomplete operating conditions; 3. Introduce a feedback optimization mechanism based on model performance to enable the sample generation process to adaptively adjust to improve the anomaly warning effect; 4. Significantly improves the generalization ability and stability of the anomaly early warning model under unknown and extreme working conditions; It is suitable for various industrial equipment scenarios and has good engineering feasibility and promotional value.

[0087] The system architecture upon which the method provided in this application relies mainly comprises three interrelated, progressively layered core layers: 1. Physical Mechanism Layer: Based on first-principles physical models such as equipment dynamics and fluid mechanics, and combined with safety specifications for actual engineering applications, a physically feasible domain is established to accurately describe the operating state of the equipment under different working conditions. This layer provides the fundamental guarantee of the physical reliability of the entire method.

[0088] 2. Data Generation Layer: Within the strict constraints of the physically feasible domain, physical simulation generation and physically constrained generative adversarial network (Physically-ConstrainedGAN, PcGAN) generation are carried out in parallel to generate high-quality anomalous samples from the two dimensions of mechanism deduction and distribution learning, respectively, and the final expanded sample set is formed through a fusion screening mechanism.

[0089] 3. Decision Optimization Layer: An advanced anomaly early warning model is trained using an expanded sample set, and a feedback optimization mechanism based on the model's generalization performance is established. By continuously monitoring the model's performance under rare, extreme, and other vulnerable conditions, the system guides the adjustment of physical simulation and data generation strategies, forming a self-evolving and adaptive closed-loop optimization system that ensures that sample generation always aims to improve the early warning effect.

[0090] This application also provides a specific application example: A mixed-flow turbine (model HL220) at a hydroelectric power station frequently experiences cavitation at the runner blade inlet edge under low head and high load conditions, causing material erosion and efficiency reduction. The existing monitoring system relies on a single threshold alarm for "tailrace tube pressure pulsation," but this indicator does not change significantly in the early stages of cavitation, leading to a delayed warning. Historical operating data shows very few severe cavitation samples, and these samples are poorly distinguishable from normal samples based on conventional monitoring parameters.

[0091] 2. Specific Implementation Process Step 1: Data Preparation and Operating Condition Locking The turbine's operating data from the past two years were collected, and five key parameters strongly correlated with cavitation were selected as the state vector. :

[0092] in, Water head (unit: m). Active power (unit: MW). Guide vane opening (unit: %) Pressure pulsation in the tailrace pipe (unit: kPa). Unit swing (unit: μm). Cluster analysis identified the "low head, high load" operating condition cluster. Its characteristics are and This operating condition is a cavitation-prone area, and as the target operating condition in this example, its operating condition parameter vector is denoted as... .

[0093] Step 2: Construct the physically feasible domain of "cavitation" First, the cavitation characteristic curve of this type of turbine is retrieved ( (Relationship) and CFD model of turbine fluid dynamics , operating parameters Input the model. By changing the cavitation coefficient in the model... By simulating cavitation development, simulations were obtained including pressure pulsation. and swing This includes a series of outputs. Based on this, the physically feasible region related to cavitation is formally defined. It is described by the following constraints: 1. Equality constraints Based on the power balance equation of the water turbine:

[0094] Where ρ is the density of water. The actual flow rate of the water turbine is represented by g, where g is the acceleration due to gravity. It is a model The given efficiency function is the cavitation coefficient. and guide vane opening The function. The generated sample state. This relationship must be approximately satisfied, that is .

[0095] 2. Inequality constraints Define the vacuolation initial boundary:

[0096] in It depends on the current state. Inversely calculated cavitation coefficient, It is the theoretical cavitation initiation coefficient under this operating condition. This is for engineering safety margin. Violation of this constraint (i.e....) This means that the operating state has entered the cavitation region.

[0097] 3. Inequality constraints Used to limit unit sway and ensure vibration remains within safe limits, its mathematical form is:

[0098] in, This represents the unit's runout, with 200 μm being the maximum permissible value set according to the unit's safety regulations. This constraint ensures that the unit's vibration will not exceed the structural limits when generating abnormal samples, thereby guaranteeing equipment safety.

[0099] Cavitation constraints are described by defining the following inequality:

[0100] Where σ(x) is the cavitation coefficient calculated from the current state x. Let be the theoretical cavitation initiation coefficient under the target operating condition, and δ be the engineering safety margin. When ≥ 0, it means σ(x) < +δ, the equipment status enters the cavitation risk zone; when When the value is less than 0, the state is in the cavitation safety zone, which is a slightly cavitation state within a controllable range.

[0101] Comprehensive power balance equation constraints ≈ 0 and the above inequality constraints and It can determine the cavitation-related physical feasible domain under the target operating condition. Defined as:

[0102] In this definition, ≤ 0 indicates that the state is within a controllable cavitation range (it can occur but is tolerable), while If the value is ≤ 0, it ensures that the unit vibration remains within the structurally permissible range. Through this constraint definition, the physical feasible region is defined. It can simultaneously describe the safety boundaries and structural limitations of the equipment during the cavitation development process, providing clear physical constraints for the subsequent generation of abnormal samples.

[0103] Step 3: Generate anomalous samples within the physically feasible region A dual-path collaborative strategy of physical simulation and data-driven generation is adopted. In the physical simulation path, a normal reference point is first set under the target operating condition. (corresponding to satisfy) (The state of no vacuolation). Then, the fault direction vector is defined. Its physical meaning is the cavitation coefficient. Decrease accompanied by pressure pulsation With swing The direction of parameter change as it increases. The cavitation severity parameter increases progressively along this direction. Perform parameter scanning, run the CFD model, and generate a sequence of states from normal to abnormal. ,satisfy and ,in Representing the physical simulation process, each generated sample carries a precise label indicating the severity of cavitation. .

[0104] In the PcGAN (Physically Constrained Generative Adversarial Network) approach, a historical normal dataset under the target operating condition is used. Using positive samples, a small number of initial cavitation samples generated by physical simulation are used as guides to train the generator. Its loss function Incorporating combat losses loss of physical consistency :

[0105] here As a weighting factor, The function ensures that the constraint is only violated if the constraint is not violated. A penalty is applied when the generator outputs samples that satisfy the physical constraints. After training, the generator operates on the condition parameter vector. A large number of particles are generated in the cavitation region (satisfying) abnormal sample set .

[0106] Step 4: Fusion screening and model training Sample sets generated by physical simulation Sample sets generated by physical constraint generative adversarial networks After merging, the first step is to use a fast physical discriminator. Alternatively, the constraint function values ​​can be calculated directly to eliminate those that clearly violate the power balance equation. Or other physical constraints on the samples to ensure that the final samples satisfy the basic physical laws.

[0107] Subsequently, a statistical rationality screening was conducted to avoid generating samples that resembled historical normal data for the target operating condition. The distribution is too skewed. Specifically, the method involves calculating the distribution of each candidate sample. Compared with historical data average Mahalanobis distance:

[0108] in The covariance matrix after shrinkage estimation or diagonal regularization:

[0109] The regularization coefficient is... The identity matrix is ​​used. This processing improves statistical robustness and avoids ill-conditioned or non-invertible covariance matrices in high-dimensional data. The distance threshold does not depend on the theoretical value of the chi-square distribution under the multivariate Gaussian assumption, but is adaptively determined through cross-validation or historical data quantiles. If the value exceeds the predetermined threshold, the sample is considered a statistically significant outlier and is removed.

[0110] After the above dual screening using both physical and statistical methods, a high-quality, physically reliable cavitation sample augmentation set was formed. Its size is approximately 10 times that of the original historical anomaly samples. Subsequently, using... Training a one-dimensional convolutional neural network model The model uses state data from a 60-minute time series window. As input, the corresponding cavitation risk level is output, where 0 represents normal, 1 represents slight cavitation, and 2 represents severe cavitation.

[0111] This integrated screening and model training process ensures that the generated samples conform to physical laws and have statistical rationality, thereby providing highly reliable and comprehensive training data for the anomaly early warning model and improving the model's generalization ability under rare and extreme conditions.

[0112] Step 5: Closed-loop feedback optimization The trained early warning model was deployed on three months of test data, and the evaluation found that it was effective against "initial cavitation" (i.e., Slightly lower , The recall rate was low (a slight increase). The system identified this as a weak link in performance and automatically initiated feedback optimization: First, it guided the physical simulation path at the cavitation inception boundary. The parameter scanning step size is doubled nearby to generate more refined samples of transition states; secondly, the PcGAN path is guided by the conditional vector. Increase the sampling weight of the "slight cavitation" category. Optimize the model using a newly generated dataset rich in initial cavitation samples. After fine-tuning, the model improved recall for the pattern by 35% in the next evaluation.

[0113] The performance comparison between the system method of this embodiment and the traditional method during the 6-month comparative test is shown in Table 1 below: Table 1

[0114] This example fully demonstrates the application process and significant value of the method described in this embodiment in the early warning of critical faults in hydraulic turbines. By constructing a physically feasible domain with cavitation initiation boundary as the core, strict physical consistency constraints are provided for sample generation. The dual-path generation strategy of physical simulation and PcGAN takes into account both the causal clarity of fault development and the diversity of abnormal mode generation. Crucially, the closed-loop optimization mechanism based on model performance feedback enables the system to automatically identify early warning weaknesses and directionally enhance the generation of corresponding samples, thereby achieving continuous and adaptive improvement in anomaly early warning capabilities.

[0115] The performance comparison results of different methods on a real test set containing 5 rare faults are shown in Table 2 below: Table 2

[0116] Experimental results show that the method in this embodiment is significantly better than the comparative method in all indicators, especially in terms of recall (anomaly detection capability).

[0117] Based on the above method embodiments, this application also provides a device for generating device anomaly samples, see [link to relevant documentation]. Figure 2 As shown, the device includes: a feasible domain construction module 22, used to identify operating conditions from historical monitoring data of the equipment and determine multiple operating mode patterns; based on the physical model and the identified operating mode patterns, it constructs a physical feasible domain for the equipment operation; the physical feasible domain is used to define the set of all operating states of the equipment that occur under specified operating conditions and conform to physical laws and safety requirements; a sample generation module 24, used to generate a target anomaly sample set within the physical feasible domain using a sample generation strategy that integrates physical simulation and data-driven generation model; and a strategy optimization module 26, used to train the equipment anomaly early warning model using samples from the target anomaly sample set, optimize the sample generation strategy based on the performance indicators of the trained model, and continue to execute the steps of generating the target anomaly sample set and feedback optimization until the performance indicators of the model meet the requirements, thus obtaining an optimized equipment anomaly sample set.

[0118] Furthermore, the aforementioned feasible domain construction module 22 is used to acquire historical monitoring data of the equipment; construct a three-dimensional time series tensor based on the historical monitoring data of the equipment; the three dimensions of the three-dimensional time series tensor include the number of historical samples, the time step corresponding to each sample, and the dimensions of the equipment monitoring parameters; perform unsupervised operating condition clustering on the three-dimensional time series tensor to identify typical operating condition modes of equipment operation; for each identified operating condition mode, determine the joint probability distribution of all monitoring parameters under the operating condition mode; and identify operating condition modes with a joint probability distribution less than a threshold as rare operating condition modes.

[0119] Furthermore, the aforementioned feasible domain construction module 22 is used to extract a set of key parameters that determine the macroscopic operating state of the equipment for each operating condition mode, forming an operating condition parameter vector; to simulate the operating condition parameter vector through a high-fidelity parametric physical model, and output key physical response indicators; the key physical response indicators include: maximum vibration displacement, isentropic efficiency, pressure pulsation coefficient, and critical speed margin; based on the key physical response indicators, a constraint set consisting of multiple physical constraints is constructed as the physical feasible domain of the equipment operation; the physical constraints include equality constraints and inequality constraints; and the physical feasible domain is simplified and visualized in engineering.

[0120] Furthermore, the aforementioned sample generation module 24 is used to generate a first abnormal sample set through physical simulation based on the constraints of the physical feasible region, and to generate a second abnormal sample set through a physical constraint generative adversarial network; and to perform sample screening based on the first abnormal sample set and the first abnormal sample set to obtain a target abnormal sample set.

[0121] Furthermore, the aforementioned sample generation module 24 is used to determine a normal reference point within the physically feasible domain; the normal reference point is the sample corresponding to the highest efficiency and best stability state; based on fault dynamics knowledge, a set of fault development direction vectors are predefined; along each fault development direction vector, starting from the normal reference point, parameter scanning simulation is performed with gradually increasing fault severity parameters to generate a continuous fault development sequence from normal to abnormal and from mild to severe, thus obtaining the first abnormal sample set; wherein, each sample in the continuous fault development sequence carries a fault type label and a severity label.

[0122] Furthermore, the aforementioned physical constraint generative adversarial network includes a generator and a discriminator. The generator's inputs include random noise, a condition vector, operating parameters, and boundary parameters. The boundary parameters are used to input the boundary information of the physical feasible region into the generator to guide the generation process. The discriminator is used to determine whether a sample is within the physical feasible region. During the generator's training process, the model loss value is determined based on the adversarial loss and the physical consistency loss. The physical consistency loss forces the samples generated by the generator to satisfy the physical equality constraints without violating the inequality constraints.

[0123] Furthermore, the sample generation module 24 is used to merge the samples in the first abnormal sample set and the second abnormal sample set to obtain the first candidate sample set; to perform physical constraint re-examination on the samples in the first candidate sample set, and to remove all samples that violate the physical feasible region constraint to obtain the second candidate sample set; to calculate the Mahalanobis distance between each candidate sample in the second candidate sample set and the corresponding working condition historical data, and to remove samples whose Mahalanobis distance exceeds the threshold to obtain the third candidate sample set; and to use the farthest point sampling algorithm to select a subset in the third candidate sample set that is evenly distributed in the feature space and can cover different abnormal patterns to the greatest extent as the target abnormal sample set.

[0124] Furthermore, the aforementioned strategy optimization module 26 is used to evaluate the model's performance metrics after each round of model training using multiple specified fault validation sets. These performance metrics include precision, recall, and F1 score. Based on the specified fault validation sets where performance metrics are below a threshold, the module identifies model performance weaknesses. These weaknesses include at least one of the following: low specified fault recognition rate, insufficient generalization ability in extreme operating conditions, and insufficient differentiation between minor and severe faults. Based on the identified model performance weaknesses, the system automatically adjusts the sample generation strategy for the next iteration, specifically including: addressing the low specified fault recognition rate by adjusting the physical simulation... The simulation vectors for specified fault modes are added to the true path, and the parameter scanning resolution is improved. In the generative adversarial model path, more corresponding abnormal samples are generated in the condition vector. To address the insufficient generalization ability in extreme operating conditions, more intensive simulations are performed in the neighborhood of the operating condition parameters during the construction of the physical feasible region to refine the feasible region boundary. The generator generates more samples simultaneously under the extreme operating conditions to ensure more comprehensive coverage of extreme states. To address the insufficient distinction between minor and severe faults, the distribution of fault severity parameters in sample generation is adjusted, and the sample density is increased in the transition region to ensure that abnormal samples of different severity levels are fully trained.

[0125] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts of the device embodiment not mentioned can be referred to the corresponding content in the aforementioned method embodiment.

[0126] Based on the above method embodiments, this application also provides a device anomaly early warning method, see [link to relevant documentation]. Figure 3 As shown, the method includes the following steps: Step S302: Obtain real-time equipment operation monitoring data; Step S304: Input the real-time equipment operation monitoring data into the equipment anomaly early warning model for prediction, and obtain the corresponding anomaly early warning result; wherein, the equipment anomaly early warning model is obtained by training the model with an optimized equipment anomaly sample set generated by the equipment anomaly sample generation method. The method provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any part of the method not mentioned in the embodiment section can be referred to the corresponding content in the aforementioned method embodiment.

[0127] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to implement the above-described method. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0128] The computer program products of the methods, apparatus, and electronic devices provided in the embodiments of this application include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementations, please refer to the method embodiments, which will not be repeated here.

[0129] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0130] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0132] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for generating equipment anomaly samples, characterized in that, The method includes: The equipment's historical monitoring data is used to identify operating conditions and determine multiple operating modes. Based on the physical model and the identified operating modes, a physical feasible domain for the equipment's operation is constructed. The physical feasible domain is used to define the set of all operating states of the equipment that occur under a specified operating condition and that conform to physical laws and safety requirements. Within the physically feasible domain, a sample generation strategy that integrates physical simulation and data-driven generative models is adopted to generate a target anomaly sample set. The device anomaly early warning model is trained using samples from the target anomaly sample set. Based on the performance metrics of the trained model, the sample generation strategy is optimized by feedback. The steps of generating the target anomaly sample set and the feedback optimization are continued until the performance metrics of the model meet the requirements, resulting in an optimized device anomaly sample set.

2. The method according to claim 1, characterized in that, The steps for identifying operating conditions from historical equipment monitoring data and determining multiple operating mode patterns include: Obtain historical monitoring data from the equipment; A three-dimensional time series tensor is constructed based on the historical monitoring data of the equipment; the three dimensions of the three-dimensional time series tensor include the number of historical samples, the time step corresponding to each sample, and the dimensions of the equipment monitoring parameters. Unsupervised clustering of the three-dimensional time series tensor is performed to identify typical operating conditions of the equipment. For each identified operating condition mode, determine the joint probability distribution of all monitored parameters under that operating condition mode; Operating conditions with a joint probability distribution less than a threshold are identified as rare operating conditions.

3. The method according to claim 1, characterized in that, The steps for constructing the physically feasible domain for equipment operation based on the physical model and identified operating conditions include: For each operating mode, a set of key parameters that determine the macroscopic operating status of the equipment are extracted to form an operating mode parameter vector; The working condition parameter vector is simulated using a high-fidelity parametric physical model, and key physical response indicators are output. The key physical response indicators include: maximum vibration displacement, isentropic efficiency, pressure pulsation coefficient, and critical speed margin. Based on the aforementioned key physical response indicators, a constraint set consisting of multiple physical constraints is constructed as the physical feasible domain for device operation; the physical constraints include equality constraints and inequality constraints. The physical feasible domain is simplified and visualized through engineering.

4. The method according to claim 1, characterized in that, Within the physically feasible domain, the steps for generating a target anomaly sample set using a sample generation strategy that integrates physical simulation and data-driven generative models include: Based on the constraints of the physical feasible region, a first set of anomalous samples is generated through physical simulation, and a second set of anomalous samples is generated through a physical constraint generative adversarial network. Based on the first abnormal sample set and the first abnormal sample set, sample screening is performed to obtain the target abnormal sample set.

5. The method according to claim 4, characterized in that, Based on the constraints of the physically feasible region, the steps for generating the first set of anomalous samples through physical simulation include: Within the physically feasible region, a normal reference point is determined; the normal reference point is the sample corresponding to the highest efficiency and best stability state; based on the knowledge of fault dynamics, a set of fault development direction vectors are predefined. Along each of the fault development direction vectors, starting from the normal reference point, parameter scanning simulation is performed with progressively increasing fault severity parameters to generate a continuous fault development sequence from normal to abnormal and from mild to severe, resulting in a first abnormal sample set; wherein each sample in the continuous fault development sequence is labeled with a fault type and a severity.

6. The method according to claim 4, characterized in that, The physical constraint generative adversarial network includes a generator and a discriminator. The generator's inputs include random noise, a condition vector, operating parameters, and boundary parameters. The boundary parameters are used to input the boundary information of the physical feasible region into the generator to guide the generation process. The discriminator is used to determine whether a sample is within the physical feasible region. During the training process of the generator, the model loss value is determined based on the adversarial loss and the physical consistency loss. The physical consistency loss forces the samples generated by the generator to satisfy physical equality constraints without violating inequality constraints.

7. The method according to claim 4, characterized in that, The step of filtering samples based on the first abnormal sample set and the second abnormal sample set to obtain the target abnormal sample set includes: The samples in the first abnormal sample set and the second abnormal sample set are merged to obtain the first candidate sample set; The samples in the first candidate sample set are subjected to physical constraint re-examination, and all samples that violate the physical feasible region constraint are removed to obtain the second candidate sample set; Calculate the Mahalanobis distance between each candidate sample in the second candidate sample set and the corresponding historical data of the working condition, and remove samples whose Mahalanobis distance exceeds the threshold to obtain the third candidate sample set. Using the farthest point sampling algorithm, a subset that is evenly distributed in the feature space and can cover different abnormal patterns to the greatest extent is selected from the third candidate sample set as the target abnormal sample set.

8. The method according to claim 1, characterized in that, The step of optimizing the sample generation strategy based on the performance metrics of the trained model includes: After each round of model training, the model's performance metrics are evaluated using multiple specified failure validation sets; these performance metrics include precision, recall, and F1 score. Based on a specified fault validation set where performance indicators are less than a threshold, weak points in model performance are identified; the weak points in model performance include at least one of the following: low identification rate of specified faults, insufficient generalization ability in extreme operating conditions, and insufficient differentiation between minor and serious faults. Based on the identified weaknesses in model performance, the system automatically adjusts the sample generation strategy for the next iteration, specifically including: To address the low recognition rate of specified faults, simulation vectors for specified fault modes are added to the physical simulation path, and the parameter scanning resolution is improved. In the generative adversarial model path, more corresponding abnormal samples are generated in the condition vector. To address the insufficient generalization capability in extreme operating conditions, more intensive simulations are performed in the neighborhood of the operating condition parameters during the construction of the physical feasible region, thereby refining the feasible region boundary; the generator simultaneously generates more samples under this extreme operating condition to ensure more comprehensive coverage of extreme states; To address the insufficient differentiation between minor and severe faults, the distribution of fault severity parameters in sample generation was adjusted, and the sample density was increased in the transition region to ensure that abnormal samples of different severity levels were adequately trained.

9. A device for generating equipment anomaly samples, characterized in that, The device includes: The feasible domain construction module is used to identify the operating conditions of the equipment based on historical monitoring data and determine multiple operating condition modes; based on the physical model and the identified operating condition modes, it constructs the physical feasible domain of the equipment operation; the physical feasible domain is used to define the set of all operating states of the equipment that occur under a specified operating condition and that conform to physical laws and safety requirements. The sample generation module is used to generate a target abnormal sample set within the physical feasible domain by employing a sample generation strategy that integrates physical simulation and data-driven generation model. The strategy optimization module is used to train the device anomaly early warning model using samples from the target anomaly sample set, optimize the sample generation strategy based on the performance indicators of the trained model, and continue to execute the steps of generating the target anomaly sample set and the feedback optimization until the performance indicators of the model meet the requirements, thereby obtaining an optimized device anomaly sample set.

10. A method for early warning of equipment malfunctions, characterized in that, The method includes: Obtain real-time equipment operation monitoring data; The real-time equipment operation monitoring data is input into the equipment anomaly early warning model for prediction to obtain the corresponding anomaly early warning result; wherein, the equipment anomaly early warning model is obtained by training the optimized equipment anomaly sample set generated by the equipment anomaly sample generation method as described in any one of claims 1-8.

Images