Power generation equipment fault monitoring method and system based on big data

Abstract

This invention discloses a method and system for monitoring power generation equipment faults based on big data. The method includes power generation equipment data acquisition, equipment status index construction, equipment deviation adaptation loss construction, equipment status identification model construction, and equipment status anomaly monitoring. This invention belongs to the field of data processing, specifically referring to a method and system for monitoring power generation equipment faults based on big data. This solution effectively avoids missing high-risk anomalies by constructing a risk-oriented dynamic anchor point system; simultaneously, it introduces the standard deviation vector of sample features as an objective fluctuation benchmark, significantly reducing misjudgments of normal states; by quantifying the sample fitting complexity and allocating a larger dynamic penalty coefficient, it strengthens the fitting strength for critical and slightly anomalous samples; and it constructs an equipment deviation adaptation loss function and embeds a noise removal mechanism to achieve intelligent removal of false-anomaly equipment noise data, thereby improving the reliability of equipment status anomaly monitoring and reducing missed detections of early fault precursors.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for monitoring power generation equipment faults based on big data. Background Technology

[0002] Power generation equipment fault monitoring methods involve collecting equipment operating parameters, analyzing their changing patterns and abnormal characteristics, and determining in real time whether the equipment deviates from its normal state, identifying fault types, and issuing early warnings. However, general equipment condition anomaly monitoring systems suffer from fixed status judgment anchor points, making it impossible to perceive fluctuations in power generation equipment data, thus leading to poor anomaly monitoring accuracy. Furthermore, these systems often have poor ability to detect critical and minor anomalies, and false anomaly data can interfere with equipment condition monitoring, easily causing missed fault precursors. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a method and system for monitoring power generation equipment faults based on big data. Addressing the problem that general equipment condition anomaly monitoring systems often rely on fixed state judgment anchor points, failing to detect fluctuations in power generation equipment data and thus leading to poor anomaly monitoring accuracy, this solution constructs a risk-oriented dynamic anchor point system. The equipment state anchor point is decomposed into a state baseline and an anchor point correction term. The anchor point correction term is positively correlated with the equipment anomaly occurrence rate; the higher the anomaly occurrence rate, the larger the anchor point correction term, resulting in stricter anchor point judgment criteria and greater sensitivity to minor deviations from normal states, effectively avoiding missed detections of high-risk anomalies. Simultaneously, a sample feature standard deviation vector is introduced as an objective fluctuation baseline, objectively reflecting the natural and reasonable fluctuation range of power generation equipment monitoring data under the corresponding state. The more stable the data fluctuation, the smaller the anchor point correction, significantly reducing misjudgments of normal states and thus improving equipment condition anomaly monitoring. To improve accuracy, this solution addresses the shortcomings of general equipment condition anomaly monitoring systems, such as poor perception of critical and minor anomalies, and the interference of false anomaly data, which can easily lead to missed fault precursors. Specifically, for critical and minor anomaly samples, this solution calculates the similarity difference between the sample and normal state anchor points, as well as various anomaly state anchor points, to accurately quantify the sample fitting complexity. A larger dynamic penalty coefficient is then allocated based on the fitting complexity to strengthen the fitting strength for critical and minor anomaly samples. An equipment deviation adaptation loss function is constructed, embedding a clutter removal mechanism to intelligently filter out clutter data from false anomalies. Simultaneously, during the model benchmark training phase, a loss function with a fixed benchmark penalty coefficient is used to freeze the clutter removal function, ensuring the stability of the equipment state anchor point fitting base and avoiding the unstable anchor point distribution caused by random parameters in the early training stage. This improves the reliability of equipment condition anomaly monitoring and reduces missed fault precursors.

[0004] The technical solution adopted by this invention is as follows: The fault monitoring method for power generation equipment based on big data provided by this invention includes the following steps:

[0005] Step S1: Data acquisition from power generation equipment;

[0006] Step S2: Constructing equipment status indicators;

[0007] Step S3: Constructing equipment deviation adaptation loss;

[0008] Step S4: Construction of equipment status recognition model;

[0009] Step S5: Monitoring abnormal equipment status.

[0010] Further, in step S1, the data acquisition of the power generation equipment involves obtaining historical time-series monitoring data of the generator set, including the main unit's operating parameters, key auxiliary machine status parameters, operating condition parameters, and environmental medium parameters; and performing preprocessing, including missing value handling and standardization; and using a sliding window to extract samples from the preprocessed time-series monitoring data, and labeling them with equipment status tags, including normal operation and various fault states, to obtain the power generation equipment monitoring set.

[0011] Further, in step S2, the construction of the device status indicators specifically includes:

[0012] The equipment status anchor point definition decomposes the equipment status anchor point into a status baseline and an anchor point correction item;

[0013] Calculate state similarity based on device state anchor points.

[0014] Furthermore, in step S3, the construction of the device deviation adaptation loss specifically includes:

[0015] The difficulty of monitoring samples is calibrated by measuring the difference in similarity between the sample and the anchor points of normal and fault states of the equipment, thereby calibrating the fitting complexity of each monitoring sample.

[0016] Dynamic penalty coefficient allocation is performed based on the fitting complexity value.

[0017] The equipment deviation adaptation loss is constructed based on the characteristic distribution pattern of normal equipment samples, amplifying the characteristic differences between normal equipment samples and pseudo-abnormal noise, and filtering out pseudo-abnormal samples; at the same time, a noise filtering mechanism is added to construct the equipment deviation adaptation loss function.

[0018] Further, in step S4, the construction of the device state recognition model specifically includes:

[0019] The model architecture design takes the power generation equipment monitoring set as input and adopts an end-to-end deep learning architecture with temporal feature extraction and state anchor point classification. The overall architecture is divided into three parts: temporal feature extraction sub-network, state anchor point mapping sub-network, and fault discrimination output layer.

[0020] The temporal feature extraction subnetwork adopts a hybrid structure combining gated recurrent units and one-dimensional convolutions;

[0021] The state anchor mapping subnetwork, with fully connected layers as its main body, maps the high-dimensional features output by the temporal feature extraction subnetwork to the device-specific normal state feature space. At the same time, it realizes the similarity calibration between sample features and each state anchor point through state similarity calculation.

[0022] The fault discrimination output layer serves as the model's output layer. Based on the state similarity results of the state anchor mapping subnetwork and combined with the dynamic discrimination threshold, it outputs the generator set state identification results and specific fault categories.

[0023] Model training employs a benchmark training and adaptive fine training strategy, optimizing and updating model parameters based on a power generation equipment monitoring set. In the benchmark training phase, a loss function with a fixed benchmark penalty coefficient is used as the training objective function, noise filtering is frozen, and the basic parameters of the sub-network and the state anchor mapping sub-network are optimized only through classification loss optimization of temporal feature extraction. In the adaptive fine training phase, based on a stable state anchor distribution, the fitting complexity of the equipment monitoring samples is accurately calibrated, penalty coefficients are adaptively allocated, and the equipment deviation adaptation loss function is used as the training objective function, ultimately completing the establishment of the equipment state recognition model.

[0024] Furthermore, in step S5, the abnormal equipment status monitoring is based on the established equipment status recognition model. The time-series monitoring data of the generator set operation is acquired in real time. After preprocessing, sliding window truncation and status feature vector construction consistent with offline training, the data is input into the equipment status recognition model. The equipment status type output by the model is used as the final monitoring result. If the monitoring result is a fault status, the operation and maintenance management personnel are given a graded warning according to the preset alarm strategy, and the specific fault type, occurrence time and confidence level are automatically reported to provide a basis for maintenance decision-making.

[0025] The power generation equipment fault monitoring system based on big data provided by this invention includes a power generation equipment data acquisition module, an equipment status index construction module, an equipment deviation adaptation loss construction module, an equipment status identification model construction module, and an equipment status anomaly monitoring module.

[0026] The power generation equipment data acquisition module acquires historical time-series monitoring data of generator sets and constructs a power generation equipment monitoring set.

[0027] The equipment status index construction module defines equipment status anchor points for the power generation equipment monitoring set, defines risk-oriented dynamic anchor points through status benchmarks and anchor point correction terms, and calculates status similarity.

[0028] The equipment deviation adaptation loss construction module calibrates the sample fitting complexity based on the state similarity difference, adaptively allocates dynamic penalty coefficients, and constructs the equipment deviation adaptation loss function in combination with the noise removal mechanism.

[0029] The equipment status recognition model construction module constructs an equipment status recognition model based on the equipment deviation adaptation loss function.

[0030] The equipment status anomaly monitoring module performs real-time equipment status anomaly monitoring based on the equipment status recognition model.

[0031] The beneficial effects achieved by the present invention using the above solution are as follows:

[0032] (1) To address the problem that general equipment status anomaly monitoring systems have fixed status judgment anchor points and cannot perceive the fluctuation range of power generation equipment data, resulting in poor anomaly monitoring accuracy, this solution constructs a risk-oriented dynamic anchor point system: the equipment status anchor point is decomposed into a status benchmark and an anchor point correction term, where the anchor point correction term is positively correlated with the equipment anomaly occurrence rate. The higher the equipment anomaly occurrence rate, the larger the anchor point correction term, the stricter the anchor point judgment standard, and the more sensitive to small deviations in normal status, effectively avoiding missed detection of high-risk anomalies; at the same time, the standard deviation vector of sample features is introduced as an objective fluctuation benchmark, which objectively reflects the natural and reasonable fluctuation range of power generation equipment monitoring data under the corresponding status. The more stable the data fluctuation, the smaller the anchor point correction amount, which can significantly reduce the misjudgment of normal status, thereby improving the accuracy of equipment status anomaly monitoring.

[0033] (2) In view of the problems that general equipment status anomaly monitoring systems have poor perception of critical and minor anomalies, and that false anomaly data interferes with equipment status anomaly monitoring, which can easily lead to missed detection of fault precursors, this solution calculates the similarity difference between the sample and the normal state anchor point and various abnormal state anchor points for critical and minor anomaly samples, accurately quantifies the sample fitting complexity, and assigns a larger dynamic penalty coefficient based on the fitting complexity to strengthen the fitting strength of critical and minor anomaly samples; constructs an equipment deviation adaptation loss function and embeds a noise removal mechanism to realize intelligent removal of noise data from false anomaly equipment; at the same time, in the model benchmark training stage, a loss function with a fixed benchmark penalty coefficient is used to freeze the noise removal function to ensure the stability of the fitting basis of the equipment status anchor points, avoid the problem of unstable anchor point distribution caused by random parameters in the early stage of training, thereby improving the reliability of equipment status anomaly monitoring and reducing missed detection of fault precursors. Attached Figure Description

[0034] Figure 1 A flowchart illustrating the big data-based fault monitoring method for power generation equipment provided by this invention;

[0035] Figure 2 This is a schematic diagram of a big data-based power generation equipment fault monitoring system provided by the present invention.

[0036] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system 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 invention.

[0039] Example 1, see Figure 1 The present invention provides a big data-based fault monitoring method for power generation equipment, which includes the following steps:

[0040] Step S1: Data acquisition of power generation equipment, obtaining historical time-series monitoring data of generator sets, and constructing a monitoring set for power generation equipment;

[0041] Step S2: Equipment status index construction, defining equipment status anchor points for the power generation equipment monitoring set, defining risk-oriented dynamic anchor points through status benchmarks and anchor point correction terms; and calculating status similarity;

[0042] Step S3: Construct the equipment deviation adaptation loss function. Based on the difference in state similarity, the sample fitting complexity is calibrated, the dynamic penalty coefficient is adaptively allocated, and the equipment deviation adaptation loss function is constructed in combination with the noise removal mechanism.

[0043] Step S4: Equipment status identification model construction, constructing an equipment status identification model based on the equipment deviation adaptation loss function;

[0044] Step S5: Equipment status anomaly monitoring, real-time equipment status anomaly monitoring based on equipment status recognition model.

[0045] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the power generation equipment data acquisition involves obtaining historical time-series monitoring data of the generator set, including main unit operating parameters, key auxiliary machine status parameters, operating condition parameters, and environmental medium parameters. The main unit operating parameters include turbine generator set vibration intensity, bearing metal temperature, lubricating oil pressure, and axial displacement. The key auxiliary machine status parameters include induced draft fan bearing temperature, feedwater pump outlet pressure, and condenser vacuum. The operating condition parameters include generator active power, main steam pressure, and main steam temperature. The environmental medium parameters include atmospheric temperature and circulating cooling water inlet temperature. Preprocessing is performed, including missing value processing (linear interpolation, direct removal of time-series segments with more than 5 consecutive missing values) and standardization processing (maximum and minimum normalization) to eliminate dimensional differences. The preprocessed time-series monitoring data is subjected to sliding window truncation to form samples, and equipment status labels are added, including normal operation and various fault states (rotor imbalance, bearing wear, blade scaling, condenser leakage), to obtain the power generation equipment monitoring set.

[0046] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the construction of equipment status indicators involves defining equipment status anchor points for the power generation equipment monitoring set and calculating status similarity. The specific operation is as follows:

[0047] The equipment state anchor point definition decomposes the equipment state anchor point into a state baseline and an anchor point correction item, represented as follows: ; ;in, It is the state anchor point for device state type y; It is the sample mean vector of the device state type y, which serves as the state baseline; It is the anchor point dynamic adjustment factor, which is obtained by linear transformation of the historical anomaly proportion of state types. The higher the anomaly rate, the stricter the anchor point requirements. It is the total number of historical outlier samples of state type y. It is the total number of historical anomaly monitoring samples; It is the standard deviation vector of sample features of state type y; It is an abnormal state; It is an anchor point correction item. It provides an objective fluctuation benchmark for anchor point correction, reflecting the natural and reasonable fluctuation range of power generation equipment monitoring under different conditions; As a correction coefficient, it is a quantitative representation of the historical anomaly risk of the state type, giving the anchor point adjustment a clear risk orientation;

[0048] State similarity is calculated based on device state anchor points and expressed as follows: ; ;in, It is the k-th dimension eigenvalue of the anchor point in the normal state; It is the k-th eigenvalue of the anchor point of the j-th type of abnormal state; d is the k-th dimension feature value of the i-th monitored sample; d is the total number of features; It is the state similarity between the i-th monitoring sample and the normal state anchor point (the larger the value, the closer it is to the corresponding state). It is the state similarity between the i-th monitoring sample and the j-th abnormal state anchor point (the larger the value, the closer to the abnormal state). It represents the true state of the i-th monitored sample.

[0049] By performing the above operations, this solution addresses the problem of fixed status judgment anchor points in general equipment status anomaly monitoring systems, which fail to perceive the fluctuation range of power generation equipment data and thus lead to poor anomaly monitoring accuracy. Instead, it constructs a risk-oriented dynamic anchor point system: decomposing equipment status anchor points into status benchmarks and anchor point correction terms. The anchor point correction term is positively correlated with the equipment anomaly occurrence rate; the higher the anomaly occurrence rate, the larger the anchor point correction term, resulting in stricter anchor point judgment criteria and greater sensitivity to minor deviations from normal states, effectively avoiding missed detections of high-risk anomalies. Simultaneously, it introduces the standard deviation vector of sample features as an objective fluctuation benchmark, objectively reflecting the natural and reasonable fluctuation range of power generation equipment monitoring data under the corresponding state. The more stable the data fluctuation, the smaller the anchor point correction amount, significantly reducing misjudgments of normal states and thus improving the accuracy of equipment status anomaly monitoring.

[0050] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the construction of the equipment deviation adaptation loss addresses the problems of sample fitting complexity differences (difficulty in fitting critical faults / minor faults) and equipment noise data (transient interference data interfering with fault monitoring) in equipment condition anomaly monitoring. It constructs an equipment deviation adaptation loss function to improve the fitting strength for critical fault samples and automatically filters out false abnormal equipment noise data, thereby improving the accuracy of equipment fault state identification. Specifically, it includes:

[0051] In equipment monitoring, difficult samples are defined as those prone to misjudgment, such as critical fault samples (vibration values ​​close to the alarm line but not exceeding the limit) and samples whose normal state is disturbed (brief load fluctuations). Therefore, the fitting complexity of each monitoring sample is determined by the difference in similarity between the sample and the equipment's normal state anchor points and fault state anchor points, expressed as: ;in, It is the fitting complexity value of the i-th monitored sample; It is a smoothing term, with a value of 10. -8 ~10 -7 C represents the total number of categories of abnormal equipment conditions. It is an indicator function; ; This is the baseline penalty coefficient, ranging from 0.05 to 0.12;

[0052] Dynamic penalty coefficient allocation assigns larger dynamic penalty coefficients to critical and minor equipment faults, resulting in a more accurate fit to the boundary characteristics between normal and fault conditions; baseline penalty coefficients are reserved for typical normal and typical fault conditions; the dynamic penalty coefficient allocation is based on the fitting complexity value, expressed as: ;in, is the dynamic penalty coefficient of the i-th monitoring sample; t is the penalty factor, which takes a value of 0.5 to 1.0 and controls the growth rate of the penalty coefficient.

[0053] The equipment deviation adaptation loss is constructed based on the characteristic distribution pattern of normal equipment samples, amplifying the characteristic differences between normal samples and pseudo-abnormal noise, and filtering out pseudo-abnormal samples; at the same time, a noise removal mechanism is added to construct the equipment deviation adaptation loss function, which is expressed as: ;in, It is a noise removal indicator function; ; It is the feature distance between the i-th sample and the normal state anchor point, using Euclidean distance; This is the screening threshold, with a value of 10. -3 ~10 -2 ; It is a Gaussian kernel function; It is adaptive kernel bandwidth. ; is the standard deviation of the sample feature distance vector; IQR is the interquartile range of the feature distance vector; N is the total number of samples; s is the scaling factor, which takes a value of 2 to 20 to enhance the discriminative power of feature similarity. This is the weight of the regularization term, with a value ranging from 0.01 to 0.2;

[0054] Automatically filter out false and abnormal noise data in equipment monitoring, eliminate the impact of sensor electromagnetic interference and mechanical vibration on equipment monitoring; improve the feature boundary fitting ability between normal and fault conditions; force critical faults to move closer to the corresponding fault state anchor point, strengthen the feature boundary fitting ability between normal and critical faults, and improve the critical fault identification accuracy.

[0055] Example 5, see Figure 1This embodiment is based on the above embodiment. In step S4, the equipment status identification model is constructed based on the power generation equipment monitoring set, status feature vector, and equipment deviation adaptation loss function. This completes the customized design and phased training optimization of the model architecture, enabling the model to accurately identify and classify the generator set's operating status and fault types. Specifically, this includes:

[0056] The model architecture design takes the power generation equipment monitoring set as input and adopts an end-to-end deep learning architecture with temporal feature extraction and state anchor point classification. The overall architecture is divided into three parts: temporal feature extraction sub-network, state anchor point mapping sub-network, and fault discrimination output layer.

[0057] The temporal feature extraction sub-network serves as the basic feature fitting layer of the model, employing a hybrid structure combining gated recurrent units and one-dimensional convolutions. One-dimensional convolutions are used to extract local correlation features from the equipment's temporal monitoring data (impact components in bearing vibration signals, instantaneous pulsation features of feedwater flow); gated recurrent units are used to capture long-term temporal dependency features of the data (continuous change trends of axial displacement and bearing temperature, gradual evolution features of boiler combustion conditions).

[0058] The state anchor mapping subnetwork serves as the core classification layer of the model. It is mainly composed of fully connected layers and maps the high-dimensional features output by the temporal feature extraction subnetwork to the device-specific normal state feature space. At the same time, it realizes the similarity calibration between sample features and each state anchor point through state similarity calculation.

[0059] The fault discrimination output layer serves as the model's output layer. Based on the state similarity results of the state anchor mapping subnetwork and combined with the dynamic discrimination threshold, it outputs the generator set's state identification result (normal / fault) and the specific fault category (if it is a fault).

[0060] To avoid instability in the distribution of state anchor points and easy misjudgment in difficulty calibration and noise removal caused by random parameters in the early stage of training, a training strategy of benchmark training and adaptive fine training is adopted, and the model parameters are optimized and updated based on the power generation equipment monitoring set.

[0061] During the baseline training phase, a loss function with a fixed baseline penalty coefficient is used as the training objective function. Clutter filtering is frozen, and only the basic parameters of the temporal feature extraction subnetwork and the state anchor mapping subnetwork are optimized using classification loss. When the model's accuracy in recognizing normal device states on the validation set stabilizes above 90% and the loss function value tends to plateau, baseline training ends, and the current model parameters are saved as initial parameters for adaptive fine-tuning. The loss function is expressed as: ;

[0062] In the adaptive fine training phase, based on a stable distribution of state anchor points, the fitting complexity of the equipment monitoring samples is accurately calibrated, and the penalty coefficient is adaptively allocated. The equipment deviation adaptation loss function is used as the training objective function. A mini-batch stochastic gradient descent optimizer is adopted, and the learning rate is dynamically decayed using a cosine annealing strategy. When the model achieves an equipment fault identification accuracy of ≥98% and a critical fault identification accuracy of ≥95% on the test set, and there is no significant improvement in model performance in 10 consecutive training rounds, training is terminated, the final model parameters are saved, and the establishment of the equipment state identification model is completed.

[0063] By performing the above operations, this solution addresses the problems of poor perception of critical and minor anomalies in general equipment condition anomaly monitoring systems, as well as the interference of false anomaly data in equipment condition anomaly monitoring, which easily leads to missed detection of early faults. For critical and minor anomaly samples, this solution calculates the similarity difference between the sample and normal state anchor points and various anomaly state anchor points to accurately quantify the sample fitting complexity. Based on the fitting complexity, a larger dynamic penalty coefficient is allocated to strengthen the fitting strength for critical and minor anomaly samples. An equipment deviation adaptation loss function is constructed, embedding a clutter removal mechanism to achieve intelligent removal of clutter data from false anomaly equipment. Simultaneously, during the model benchmark training phase, a loss function with a fixed benchmark penalty coefficient is used to freeze the clutter removal function, ensuring the stability of the fitting basis of the equipment condition anchor points and avoiding the problem of unstable anchor point distribution caused by random parameters in the early training stage. This improves the reliability of equipment condition anomaly monitoring and reduces missed detection of early faults.

[0064] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the abnormal equipment status monitoring is based on the established equipment status recognition model. The time-series monitoring data of the generator set operation is acquired in real time. After preprocessing, sliding window truncation and status feature vector construction consistent with offline training, the data is input into the equipment status recognition model. The equipment status type (normal or specific fault category) output by the model is used as the final monitoring result. If the monitoring result is a fault status, the operation and maintenance management personnel are given a graded warning according to the preset alarm strategy, and the specific fault type, occurrence time and confidence level are automatically reported to provide a basis for maintenance decision-making.

[0065] Example 7, see Figure 2 Based on the above embodiments, the big data-based power generation equipment fault monitoring system provided by the present invention includes a power generation equipment data acquisition module, an equipment status index construction module, an equipment deviation adaptation loss construction module, an equipment status identification model construction module, and an equipment status anomaly monitoring module.

[0066] The power generation equipment data acquisition module acquires historical time-series monitoring data of generator sets and constructs a power generation equipment monitoring set.

[0067] The equipment status index construction module defines equipment status anchor points for the power generation equipment monitoring set, defines risk-oriented dynamic anchor points through status benchmarks and anchor point correction terms, and calculates status similarity.

[0068] The equipment deviation adaptation loss construction module calibrates the sample fitting complexity based on the state similarity difference, adaptively allocates dynamic penalty coefficients, and constructs the equipment deviation adaptation loss function in combination with the noise removal mechanism.

[0069] The equipment status recognition model construction module constructs an equipment status recognition model based on the equipment deviation adaptation loss function.

[0070] The equipment status anomaly monitoring module performs real-time equipment status anomaly monitoring based on the equipment status recognition model.

[0071] 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, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0073] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for monitoring power generation equipment faults based on big data, characterized in that: The method includes the following steps: Step S1: Data acquisition of power generation equipment, obtaining historical time-series monitoring data of generator sets, and constructing a monitoring set for power generation equipment; Step S2: Equipment status index construction, defining equipment status anchor points for the power generation equipment monitoring set, defining risk-oriented dynamic anchor points through status benchmarks and anchor point correction terms; and calculating status similarity; Step S3: Construct the equipment deviation adaptation loss function. Based on the difference in state similarity, the sample fitting complexity is calibrated, the dynamic penalty coefficient is adaptively allocated, and the equipment deviation adaptation loss function is constructed in combination with the noise removal mechanism. Step S4: Constructing the equipment status recognition model, based on the equipment deviation adaptation loss function; Step S5: Equipment status anomaly monitoring, real-time equipment status anomaly monitoring based on equipment status recognition model; In step S4, the construction of the device status identification model specifically includes: Model architecture design: Taking the power generation equipment monitoring set as input, an end-to-end deep learning architecture with temporal feature extraction and state anchor point classification is adopted. The overall architecture is divided into three parts: temporal feature extraction sub-network, state anchor point mapping sub-network, and fault discrimination output layer. The temporal feature extraction subnetwork adopts a hybrid structure combining gated recurrent units and one-dimensional convolutions; The state anchor mapping subnetwork, with fully connected layers as its main body, maps the high-dimensional features output by the temporal feature extraction subnetwork to the device-specific normal state feature space. At the same time, it realizes the similarity calibration between sample features and each state anchor point through state similarity calculation. The fault discrimination output layer serves as the model's output layer. Based on the state similarity results of the state anchor mapping subnetwork and combined with the dynamic discrimination threshold, it outputs the generator set state identification results and specific fault categories. Model training employs a benchmark training and adaptive fine training strategy, optimizing and updating model parameters based on a power generation equipment monitoring set. In the benchmark training phase, a loss function with a fixed benchmark penalty coefficient is used as the training objective function, noise is frozen and filtered out, and the basic parameters of the sub-network and the state anchor mapping sub-network are optimized only through classification loss. In the adaptive fine training phase, based on a stable state anchor distribution, the fitting complexity of the equipment monitoring samples is accurately calibrated, penalty coefficients are adaptively allocated, and the equipment deviation adaptation loss function is used as the training objective function, ultimately completing the establishment of the equipment state recognition model.

2. The method for monitoring power generation equipment faults based on big data according to claim 1, characterized in that: In step S2, the construction of the device status indicators specifically includes: The equipment status anchor point definition decomposes the equipment status anchor point into a status baseline and an anchor point correction item; Calculate state similarity based on device state anchor points.

3. The method for monitoring power generation equipment faults based on big data according to claim 1, characterized in that: In step S3, the construction of the device deviation adaptation loss specifically includes: Monitoring sample difficulty calibration: The fitting complexity of each monitoring sample is calibrated by the difference in state similarity between the sample and the equipment's normal state anchor point and fault state anchor point. Dynamic penalty coefficient allocation; larger dynamic penalty coefficients are allocated for critical and minor equipment faults; baseline penalty coefficients are reserved for typical normal and typical faults; dynamic penalty coefficient allocation is based on fitting complexity values; Equipment deviation adaptation loss construction.

4. The method for monitoring power generation equipment faults based on big data according to claim 3, characterized in that: In step S3, the equipment deviation adaptation loss is constructed based on the characteristic distribution law of the normal state samples of the equipment, amplifying the characteristic differences between the real normal samples and the pseudo-abnormal noise, and screening out the pseudo-abnormal samples; at the same time, a noise screening mechanism is added to construct the equipment deviation adaptation loss function.

5. The method for monitoring power generation equipment faults based on big data according to claim 1, characterized in that: In step S1, the data acquisition of the power generation equipment is to obtain historical time-series monitoring data of the generator set, including the main unit's operating parameters, key auxiliary equipment status parameters, operating condition parameters, and environmental medium parameters. Preprocessing is performed, including handling missing values ​​and standardization. The preprocessed time-series monitoring data is sampled by sliding window extraction and labeled with equipment status tags, including normal operation and various fault states, to obtain the power generation equipment monitoring set.

6. The method for monitoring power generation equipment faults based on big data according to claim 1, characterized in that: In step S5, the abnormal equipment status monitoring is based on the established equipment status recognition model. The time-series monitoring data of the generator set operation is acquired in real time. After preprocessing, sliding window truncation and status feature vector construction consistent with offline training, the data is input into the equipment status recognition model. The equipment status type output by the model is used as the final monitoring result. If the monitoring result is a fault status, the operation and maintenance management personnel are given a graded warning according to the preset alarm strategy, and the specific fault type, occurrence time and confidence level are automatically reported to provide a basis for maintenance decision-making.

7. A big data-based power generation equipment fault monitoring system, used to implement the big data-based power generation equipment fault monitoring method as described in any one of claims 1-6, characterized in that: It includes a power generation equipment data acquisition module, an equipment status index construction module, an equipment deviation adaptation loss construction module, an equipment status identification model construction module, and an equipment status anomaly monitoring module; The power generation equipment data acquisition module acquires historical time-series monitoring data of generator sets and constructs a power generation equipment monitoring set. The equipment status index construction module defines equipment status anchor points for the power generation equipment monitoring set, and defines risk-oriented dynamic anchor points through status benchmarks and anchor point correction items. And calculate state similarity; The equipment deviation adaptation loss construction module calibrates the sample fitting complexity based on the state similarity difference, adaptively allocates dynamic penalty coefficients, and constructs the equipment deviation adaptation loss function in combination with the noise removal mechanism. The equipment status recognition model construction module constructs an equipment status recognition model based on the equipment deviation adaptation loss function. The equipment status anomaly monitoring module performs real-time equipment status anomaly monitoring based on the equipment status recognition model.

Images