Power plant equipment operation anomaly detection method based on deep learning

Abstract

The invention discloses a power plant equipment operation anomaly detection method based on deep learning, and belongs to the technical field of equipment operation monitoring. A first deep learning model is adopted to construct a feature extraction sub-model, and a second deep learning model is adopted to construct a feature recognition sub-model; according to the feature extraction sub-model and the feature recognition sub-model, obtaining an operation anomaly detection model; then training the operation anomaly detection model by adopting an intelligent optimization algorithm to obtain a trained operation anomaly detection model; and finally, the trained operation anomaly detection model is dispatched to identify the real-time multi-dimensional sample data, the operation anomaly detection result of the power plant equipment is determined, and the fault features implied in the operation data can be effectively analyzed through the processes of feature extraction and feature identification, so that the rapid and accurate detection of the operation anomaly of the power plant equipment is realized, and the detection efficiency is improved. False report and missing report are avoided, so that workers can timely maintain power plant equipment and remove faults.

Description

Technical Field

[0001] This invention belongs to the field of equipment operation monitoring technology, specifically relating to a method for detecting abnormal operation of power plant equipment based on deep learning. Background Technology

[0002] The safe, stable, and efficient operation of power plant equipment is crucial. Power plant equipment (such as steam turbines, boilers, and generators) has complex structures and operates in harsh environments, constantly under high temperature, high pressure, and high speed conditions, making it highly susceptible to various faults. Failure to detect and address these faults in a timely manner can lead to anything from degraded equipment performance and increased energy consumption to unplanned shutdowns, and even major safety accidents and economic losses. Traditional methods for detecting anomalies in power plant equipment mainly rely on threshold alarms and expert systems. Threshold alarms set fixed upper and lower limits for key operating parameters, triggering an alarm when the parameters exceed these limits. While simple and direct, this method cannot handle abnormal fluctuations in parameters within normal ranges and is insensitive to early, slowly changing faults, easily resulting in missed or false alarms. Expert systems rely on the knowledge and experience of domain experts to build rule bases for reasoning and judgment. This method is limited by the breadth and depth of expert knowledge, has high rule base maintenance costs, struggles to cover all complex fault modes, and is particularly ineffective against novel faults. Therefore, existing technologies cannot effectively identify abnormal operating conditions of power plant equipment, resulting in missed and false alarms, which prevents staff from performing timely maintenance and troubleshooting of power plant equipment. Summary of the Invention

[0003] This invention provides a deep learning-based method for detecting abnormal operation of power plant equipment, which solves the problem that existing technologies cannot effectively identify abnormal operating states of power plant equipment, resulting in missed or false alarms, and preventing staff from performing timely maintenance and troubleshooting of power plant equipment.

[0004] This application provides a deep learning-based method for detecting operational anomalies in power plant equipment, including: Obtain historical operation datasets for power plant equipment at historical operating times; wherein, the historical operation datasets include historical multidimensional sample data and their corresponding historical operating states; A feature extraction sub-model is constructed using a first deep learning model and a feature recognition sub-model is constructed using a second deep learning model. Based on the feature extraction sub-model and the feature recognition sub-model, an anomaly detection model is obtained. Based on the historical multidimensional sample data and its corresponding historical operating status, an intelligent optimization algorithm is used to train the operating anomaly detection model to obtain the trained operating anomaly detection model. Real-time multidimensional sample data corresponding to power plant equipment is collected, and the trained operation anomaly detection model is scheduled to identify the real-time multidimensional sample data to determine the power plant equipment operation anomaly detection result.

[0005] In one possible implementation, the historical multidimensional sample data includes temperature, voltage, vibration, and current of the power plant equipment collected at a preset data sampling frequency within a data sampling period; the historical operating status corresponding to the historical multidimensional sample data is the fault operating status type or normal operating status type of the power plant equipment within the same data sampling period.

[0006] In one possible implementation, the step of constructing a feature extraction sub-model using a first deep learning model and a feature recognition sub-model using a second deep learning model, and obtaining an anomaly detection model based on the feature extraction sub-model and the feature recognition sub-model, includes: A feature extraction sub-model is constructed using a first deep learning model; wherein, the first deep learning model is set as a convolutional neural network; A feature recognition sub-model is constructed using a second deep learning model; wherein, the second deep learning model is set as a long short-term memory network; The feature output layer of the feature extraction sub-model is connected to the input layer of the feature recognition sub-model to obtain the running anomaly detection model.

[0007] In one possible implementation, the step of training the operational anomaly detection model using an intelligent optimization algorithm based on the historical multidimensional sample data and its corresponding historical operating states to obtain the trained operational anomaly detection model includes: Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the feature extraction sub-model in the operation anomaly detection model to obtain the final hyperparameters of the feature extraction sub-model. The final hyperparameters of the feature extraction sub-model are then fixed to obtain the operation anomaly detection model with a fixed feature extraction sub-model. Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the operating anomaly detection model of the fixed feature extraction sub-model to obtain the trained operating anomaly detection model.

[0008] In one possible implementation, based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the feature extraction sub-model in the operational anomaly detection model to obtain the final hyperparameters of the feature extraction sub-model, including: The hyperparameters of the feature extraction sub-model in the anomaly detection model are initialized and encoded to obtain multiple different first encoding vectors; For any first encoding vector, the loss function value corresponding to the first encoding vector is obtained based on the historical multidimensional sample data and its corresponding historical running state; The first encoding vector with the minimum loss function value is determined as the first optimal vector, and the first historical optimal state vector corresponding to the first encoding vector is determined based on the loss function value corresponding to the first encoding vector during the historical training process. Based on the first optimal vector and the first historical optimal state vector, a nonlinear joint guidance strategy is used to locally update the first encoding vector to determine the first encoding vector after the local update. A historical oscillation learning strategy is used to perform local oscillation updates on the first encoding vector after the local update, and the first encoding vector after the local oscillation update is determined. A global adaptive mutation strategy is used to perform a global adaptive update on the first encoding vector after the local oscillation update, and the first encoding vector after the global adaptive update is determined. Determine whether the current number of training iterations meets the preset maximum number of training iterations. If so, determine the final hyperparameters of the feature extraction sub-model based on the first encoding vector after global adaptive update. Otherwise, return to the step of obtaining the loss function value corresponding to the first encoding vector.

[0009] In one possible implementation, the step of training the runtime anomaly detection model of the fixed feature extraction sub-model using an intelligent optimization algorithm based on the historical multidimensional sample data and its corresponding historical operating states to obtain the trained runtime anomaly detection model includes: The other hyperparameters of the anomaly detection model of the fixed feature extraction sub-model are initialized and encoded to obtain multiple different second encoding vectors; For any second encoding vector, the loss function value corresponding to the second encoding vector is obtained based on the historical multidimensional sample data and its corresponding historical running state; The second encoding vector with the minimum loss function value is determined as the second optimal vector, and the second historical optimal state vector corresponding to the second encoding vector is determined based on the loss function value corresponding to the second encoding vector during the historical training process. Based on the second optimal vector and the second historical optimal state vector, a nonlinear joint guidance strategy is used to locally update the second encoding vector to determine the second encoding vector after the local update. A historical oscillation learning strategy is used to perform local oscillation updates on the second encoding vector after the local update, and the second encoding vector after the local oscillation update is determined. A global adaptive mutation strategy is used to perform a global adaptive update on the second encoding vector after the local oscillation update, and the second encoding vector after the global adaptive update is determined. Determine whether the current number of training iterations meets the preset maximum number of training iterations. If so, determine the training anomaly detection model based on the second encoding vector after global adaptive update. Otherwise, return to the step of obtaining the loss function value corresponding to the second encoding vector.

[0010] In one possible implementation, the nonlinear joint guidance strategy includes: Based on the target's optimal vector and its historical optimal state vector, the target update amount corresponding to the target encoding vector is determined as follows:

[0011] Wherein, the target optimal vector is either the first optimal vector or the second optimal vector, the target historical optimal state vector is either the first historical optimal state vector or the second historical optimal state vector, and the target encoding vector is either the first encoding vector or the second encoding vector. Indicates the first t During the training process, the first i One target encoded vector, i =1,2,…,NP, where NP represents the total number of target encoding vectors. Represents the target optimal vector. Indicates the first t During the training process, the first i The target historical optimal state vector corresponding to each target encoding vector. Indicates the first t During the training process, the first i The target update amount corresponding to each target encoding vector. Indicates the first t +1 training session i The target update amount corresponding to each target encoding vector. Denotes the first joint update coefficient. Indicates the second joint update coefficient. e Represents the natural constant. This indicates the update range control coefficient. Represents the target encoding vector With the target's historical optimal state vector The Euclidean distance between them Represents the target encoding vector With the target optimal vector The Euclidean distance between them; The target encoding vector is locally updated based on the target update amount corresponding to the target encoding vector, resulting in the locally updated target encoding vector as follows:

[0012] in, Indicates the first i The target encoding vector after local updates.

[0013] In one possible implementation, the historical oscillation learning strategy includes:

[0014] in, Indicates the first t During the training process, the first j The first locally updated target encoding vector d Dimensional parameters, j =1,2,...,NP d =1,2,...,D; D represents the total dimension of the target encoding vector. Indicates the first j The target encoding vector after the local oscillation update is the first... d Dimensional parameters, Indicates the first t- During the first training session j The first locally updated target encoding vector d Dimensional parameters, Indicates the first t- During the 2nd training session, the first j The first locally updated target encoding vector d Dimensional parameters, Indicates the first t- During the 3 training sessions, the first j The first locally updated target encoding vector d Dimensional parameters, This represents the historical oscillation adjustment coefficient.

[0015] In one possible implementation, the global adaptive mutation strategy includes:

[0016]

[0017]

[0018] in, Indicates the first t During the training process, the first m The target encoding vector after the local oscillation update is the first...d Dimensional parameters, m =1,2,...,NP; Indicates the first m The target encoding vector after the global adaptive update of the nth global adaptive update d Dimensional parameters, The first element representing the global information vector d Dimensional parameters, Indicates the preset update step size upper limit. Compared with the preset update step size lower limit Random step size between This represents the update direction control factor between [-1, 1]. Represents pi (π). Represents a random number between [-1, 1]. Indicates the first intermediate parameter. Represents the first random encoded vector d Dimensional parameters, Indicates the first random update coefficient. Indicates the second random update coefficient. Represents a random number between (0, 2], and and It follows a normal distribution, that is , ; Indicates the second intermediate parameter, and , Represents the gamma function. This represents the maximum value of the first intermediate parameter. This represents the minimum value of the first intermediate parameter, and T represents the maximum number of training iterations.

[0019] In one possible implementation, after determining the detection result of the power plant equipment malfunction, the method further includes: If the power plant equipment operation anomaly detection result is any type of fault operation state, an alarm message carrying real-time multidimensional sample data and the corresponding power plant equipment operation anomaly detection result is generated. After the alarm message is associated with a timestamp and stored, the alarm message is transmitted to the device designated by the staff.

[0020] This invention provides a deep learning-based method for detecting operational anomalies in power plant equipment. It acquires historical multidimensional sample data and their corresponding historical operating states, constructs a feature extraction sub-model using a first deep learning model, and a feature recognition sub-model using a second deep learning model. Based on these models, an operational anomaly detection model is obtained. Then, using the historical multidimensional sample data and their corresponding historical operating states as a foundation, an intelligent optimization algorithm is employed to train the operational anomaly detection model, resulting in a trained model. Finally, the trained model is used to identify real-time multidimensional sample data to determine the detection result of operational anomalies in the power plant equipment. Through the feature extraction and recognition processes, the method effectively analyzes the fault characteristics hidden in the operating data, thereby achieving rapid and accurate detection of operational anomalies in power plant equipment, avoiding false alarms and missed alarms, and enabling staff to perform timely maintenance and troubleshooting of the power plant equipment. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0022] Figure 1 This is a flowchart illustrating a deep learning-based method for detecting operational anomalies in power plant equipment, as provided in an embodiment of the present invention.

[0023] Figure 2 The flowchart for obtaining the final hyperparameters of the feature extraction sub-model provided in this embodiment of the invention.

[0024] Figure 3 A flowchart of the trained anomaly detection model provided in this embodiment of the invention.

[0025] The accompanying drawings have illustrated specific embodiments of the invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0027] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figure 1 As shown, this embodiment of the invention provides a method for detecting abnormal operation of power plant equipment based on deep learning, including: S101. Obtain the historical operation dataset of the power plant equipment at historical operating times; wherein, the historical operation dataset includes historical multidimensional sample data and its corresponding historical operating status.

[0029] Historical multidimensional sample data can be data collected within a certain data sampling period at a preset data sampling frequency. A higher preset data sampling frequency indicates more data collection, leading to more accurate detection results. However, it also increases data processing volume and data transmission size. Therefore, when network conditions are good, edge computing or cloud service computing can be used to collect more data, enhancing the accuracy of power plant equipment anomaly detection. Conversely, a lower preset data sampling frequency indicates less data collection, but it also reduces the accuracy of detection results. However, for data processing equipment with poor network conditions or limited data processing capabilities, the preset data sampling frequency can be reduced to minimize data processing time. To facilitate subsequent identification and processing of historical multidimensional sample data, data collected at a single sampling time point can be used as one row of a data matrix, while data collected within a certain data sampling period at the preset data sampling frequency constitutes the entire data matrix.

[0030] The historical operating status corresponding to the historical multidimensional sample data can be the actual fault operating status type that occurred when the historical multidimensional sample data was collected. If no fault occurred, the historical operating status can be set to the normal operating status type.

[0031] For example, the historical multidimensional sample data includes the temperature, voltage, vibration, and current of the power plant equipment collected at a preset data sampling frequency within a data sampling period; the historical operating status corresponding to the historical multidimensional sample data is the fault operating status type or normal operating status type of the power plant equipment within the same data sampling period.

[0032] S102. A feature extraction sub-model is constructed using a first deep learning model and a feature recognition sub-model is constructed using a second deep learning model. An anomaly detection model is obtained based on the feature extraction sub-model and the feature recognition sub-model.

[0033] The feature extraction sub-model can be a model capable of processing multidimensional data. Therefore, the first deep learning model can be set as a convolutional neural network. Convolutional neural networks can recognize and process multidimensional data and data matrices, extract spatial features of historical multidimensional sample data, and output features with a fixed dimension. Therefore, the first deep learning model can be set as a convolutional neural network.

[0034] The feature recognition sub-model mainly identifies the fixed-dimensional features output by the feature extraction sub-model. Since the fixed-dimensional features output by the feature extraction sub-model are presented in vector form, the second deep learning model can be set as a long short-term memory network to further identify the temporal features of historical multidimensional sample data, thereby improving the recognition accuracy.

[0035] For example, a feature extraction sub-model is constructed using a first deep learning model and a feature recognition sub-model is constructed using a second deep learning model. Based on the feature extraction sub-model and the feature recognition sub-model, an anomaly detection model is obtained, including: A feature extraction sub-model is constructed using a first deep learning model. This first deep learning model is configured as a convolutional neural network.

[0036] A feature recognition sub-model is constructed using a second deep learning model. This second deep learning model is configured as a long short-term memory network.

[0037] The feature output layer of the feature extraction sub-model is connected to the input layer of the feature recognition sub-model to obtain an anomaly detection model. For example, in convolutional neural networks, fully connected layers typically output features of a fixed dimension. Therefore, the fully connected layer of the feature extraction sub-model can be used as the feature output layer and connected to the input layer of the feature recognition sub-model.

[0038] S103. Based on the historical multidimensional sample data and its corresponding historical operating status, the operating anomaly detection model is trained using an intelligent optimization algorithm to obtain the trained operating anomaly detection model. The anomaly detection model can learn the data relationship between the historical multidimensional sample data and its corresponding historical operating states. During the learning process, the hyperparameters (such as connection weights between network layers, thresholds, etc.) of the anomaly detection model can be optimized through intelligent optimization algorithms, so that the trained anomaly detection model can correctly classify the historical multidimensional sample data.

[0039] S104. Collect real-time multidimensional sample data corresponding to power plant equipment, and schedule the trained operation anomaly detection model to identify the real-time multidimensional sample data to determine the power plant equipment operation anomaly detection result.

[0040] The data composition, type, and dimensions of the real-time multidimensional sample data are the same as those of the historical multidimensional sample data, thus ensuring that the trained operation anomaly detection model can correctly identify the real-time multidimensional sample data and obtain the power plant equipment operation anomaly detection results.

[0041] Optionally, both real-time and historical multidimensional sample data can be normalized before being input into the anomaly detection model, thereby reducing data complexity and improving data recognition efficiency.

[0042] This invention provides a deep learning-based method for detecting operational anomalies in power plant equipment. Through feature extraction and recognition, it effectively analyzes the hidden fault features in operational data, thereby achieving rapid and accurate detection of operational anomalies in power plant equipment, avoiding false alarms and missed alarms, and enabling staff to perform timely maintenance and troubleshooting of power plant equipment. In one possible implementation, the method involves training the operational anomaly detection model using an intelligent optimization algorithm based on the historical multidimensional sample data and its corresponding historical operating states to obtain the trained operational anomaly detection model, including: Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the feature extraction sub-model in the operation anomaly detection model to obtain the final hyperparameters of the feature extraction sub-model. The final hyperparameters of the feature extraction sub-model are then fixed to obtain the operation anomaly detection model with a fixed feature extraction sub-model.

[0043] The training process of the feature extraction sub-model described above can be viewed as separate training of the convolutional neural network. Historical multidimensional sample data is used as input, and the historical operating states corresponding to the historical multidimensional sample data are used as the expected output. The feature extraction sub-model is trained normally, and then its hyperparameters are fixed before retraining. The anomaly detection model with the fixed feature extraction sub-model can extract data features from the historical multidimensional sample data through the feature extraction sub-model and output the extracted data features to the feature recognition sub-model through its fully connected layer.

[0044] Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the operating anomaly detection model of the fixed feature extraction sub-model to obtain the trained operating anomaly detection model.

[0045] The anomaly detection model of the fixed feature extraction sub-model can take historical multidimensional sample data as input. After the feature extraction sub-model extracts features, it is input into the feature recognition sub-model for recognition, thereby determining the actual output. At this time, the historical running state corresponding to the historical multidimensional sample data can be used as the expected output, which can realize the secondary training of the anomaly detection model of the fixed feature extraction sub-model.

[0046] like Figure 2As shown, based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the feature extraction sub-model in the operation anomaly detection model to obtain the final hyperparameters of the feature extraction sub-model, including: S201. Initialize and encode the hyperparameters of the feature extraction sub-model in the abnormal operation detection model to obtain multiple different first encoding vectors; For example, hyperparameters can be randomly initialized between the upper and lower bounds, and the resulting hyperparameter encoding vector can be used to obtain a first encoding vector. By repeating the initialization multiple times, multiple different first encoding vectors can be obtained.

[0047] S202. For any first encoding vector, obtain the loss function value corresponding to the first encoding vector based on the historical multidimensional sample data and its corresponding historical running state; The loss function value corresponding to the first encoding vector can be obtained using either the cross-entropy loss function or the root mean square loss function.

[0048] S203. Determine the first encoding vector with the minimum loss function value as the first optimal vector, and determine the first historical optimal state vector corresponding to the first encoding vector based on the loss function value corresponding to the first encoding vector during the historical training process. S204. Based on the first optimal vector and the first historical optimal state vector, a nonlinear joint guidance strategy is used to locally update the first encoding vector to determine the first encoding vector after the local update. S205. Using a historical oscillation learning strategy, perform local oscillation updates on the first encoding vector after the local update to determine the first encoding vector after the local oscillation update. S206. A global adaptive mutation strategy is used to perform a global adaptive update on the first encoding vector after the local oscillation update, and the first encoding vector after the global adaptive update is determined. S207. Determine whether the current number of training iterations meets the preset maximum number of training iterations. If so, determine the final hyperparameters of the feature extraction sub-model based on the first encoding vector after global adaptive update. Otherwise, return to the step of obtaining the loss function value corresponding to the first encoding vector.

[0049] Optionally, after each strategy is executed, the first optimal vector in the current training process can be maintained and the first encoding vector can be handled for out-of-bounds errors, thereby ensuring the effectiveness of training.

[0050] Existing technologies often get stuck in local optima during the training of model hyperparameters (such as in particle swarm optimization), leading to a failure of the trained model to accurately identify data. Therefore, this application provides a new training algorithm that improves the feature extraction sub-model's ability to extract features and the feature recognition sub-model's ability to recognize features, ensuring the accuracy of detecting anomalies in power plant equipment operation.

[0051] like Figure 3 As shown, the step of training the operational anomaly detection model of the fixed feature extraction sub-model using an intelligent optimization algorithm based on the historical multidimensional sample data and its corresponding historical operating states to obtain the trained operational anomaly detection model includes: S301. Initialize and encode other hyperparameters of the anomaly detection model of the fixed feature extraction sub-model to obtain multiple different second encoding vectors; Among them, other hyperparameters refer to the hyperparameters of the feature recognition sub-model.

[0052] S302. For any second encoding vector, obtain the loss function value corresponding to the second encoding vector based on the historical multidimensional sample data and its corresponding historical running state; S303. Determine the second encoding vector with the minimum loss function value as the second optimal vector, and determine the second historical optimal state vector corresponding to the second encoding vector based on the loss function value corresponding to the second encoding vector during historical training. S304. Based on the second optimal vector and the second historical optimal state vector, a nonlinear joint guidance strategy is used to locally update the second encoding vector to determine the second encoding vector after the local update. S305. Using a historical oscillation learning strategy, perform local oscillation updates on the second encoding vector after the local update, and determine the second encoding vector after the local oscillation update. S306. A global adaptive mutation strategy is used to perform a global adaptive update on the second encoding vector after the local oscillation update, and the second encoding vector after the global adaptive update is determined. S307. Determine whether the current number of training iterations meets the preset maximum number of training iterations. If yes, determine the training-based anomaly detection model based on the second encoding vector after global adaptive update. Otherwise, return to the step of obtaining the loss function value corresponding to the second encoding vector.

[0053] The methods described in steps S301-S307 are similar in principle and beneficial effects to those described in steps S201-S207, and will not be repeated here.

[0054] In one possible implementation, the nonlinear joint guidance strategy includes: Based on the target's optimal vector and its historical optimal state vector, the target update amount corresponding to the target encoding vector is determined as follows:

[0055] Wherein, the target optimal vector is either the first optimal vector or the second optimal vector, the target historical optimal state vector is either the first historical optimal state vector or the second historical optimal state vector, and the target encoding vector is either the first encoding vector or the second encoding vector. Indicates the first t During the training process, the first i One target encoded vector, i =1,2,…,NP, where NP represents the total number of target encoding vectors. Represents the target optimal vector. Indicates the first t During the training process, the first i The target historical optimal state vector corresponding to each target encoding vector. Indicates the first t During the training process, the first i The target update amount corresponding to each target encoding vector. Indicates the first t +1 training session i The target update amount corresponding to each target encoding vector. Denotes the first joint update coefficient. Indicates the second joint update coefficient. e Represents the natural constant. This indicates the update range control coefficient. Represents the target encoding vector With the target's historical optimal state vector The Euclidean distance between them Represents the target encoding vector With the target optimal vector The Euclidean distance between them; The target encoding vector is locally updated based on the target update amount corresponding to the target encoding vector, resulting in the locally updated target encoding vector as follows:

[0056] in, Indicates the first i The target encoding vector after local updates.

[0057] The nonlinear joint guidance strategy provided in this application embodiment enables the target encoding vector to effectively learn the target's optimal vector and historical optimal state vector. This not only improves the algorithm's training speed but also enhances its local convergence ability, ensuring effective training. Furthermore, the Euclidean distance-based search method allows the algorithm to possess a certain global search capability in the early stages and a strong local search capability in the later stages, thus balancing the algorithm's search performance.

[0058] In one possible implementation, the historical oscillation learning strategy includes:

[0059] in, Indicates the first t During the training process, the first j The first locally updated target encoding vector d Dimensional parameters, j =1,2,...,NP d =1,2,...,D; D represents the total dimension of the target encoding vector. Indicates the first j The target encoding vector after the local oscillation update is the first... d Dimensional parameters, Indicates the first t- During the first training session j The first locally updated target encoding vector d Dimensional parameters, Indicates the first t- During the 2nd training session, the first j The first locally updated target encoding vector d Dimensional parameters, Indicates the first t- During the 3 training sessions, the first j The first locally updated target encoding vector d Dimensional parameters, This represents the historical oscillation adjustment coefficient.

[0060] The historical oscillation learning strategy provided in this application embodiment can perform oscillation search based on the historical state of the target encoding vector. This can not only effectively avoid the algorithm getting stuck in local optima, but also improve the algorithm's ability to search unsearched regions, making it easier for the algorithm to find local optimal solutions.

[0061] In one possible implementation, the global adaptive mutation strategy includes:

[0062]

[0063]

[0064] in, Indicates the first t During the training process, the first m The target encoding vector after the local oscillation update is the first... d Dimensional parameters, m =1,2,...,NP; Indicates the first m The target encoding vector after the global adaptive update of the nth global adaptive update d Dimensional parameters, The first element representing the global information vector d Dimensional parameters, Indicates the preset update step size upper limit. Compared with the preset update step size lower limit Random step size between This represents the update direction control factor between [-1, 1]. Represents pi (π). Represents a random number between [-1, 1]. Indicates the first intermediate parameter. Represents the first random encoded vector d Dimensional parameters, Indicates the first random update coefficient. Indicates the second random update coefficient. Represents a random number between (0, 2], and and It follows a normal distribution, that is , ; Indicates the second intermediate parameter, and , Represents the gamma function. This represents the maximum value of the first intermediate parameter. This represents the minimum value of the first intermediate parameter, and T represents the maximum number of training iterations.

[0065] The global adaptive mutation strategy provided in this application can provide the algorithm with a more powerful global search capability, avoid the algorithm from getting stuck in local optima, and can also use simulated annealing algorithm to control the global search process, so as to ensure the training speed of the algorithm.

[0066] This application embodiment enhances the algorithm's ability to explore high-dimensional solution spaces by combining a nonlinear joint guidance strategy, a historical oscillation learning strategy, and a global adaptive mutation strategy. This not only improves the algorithm's global search capability but also ensures its convergence, thereby enhancing the feature extraction sub-model's ability to extract features and the feature recognition sub-model's ability to recognize features, thus ensuring the accuracy of detecting abnormal operation of power plant equipment.

[0067] In one possible implementation, after determining the detection result of the power plant equipment malfunction, the method further includes: If the power plant equipment operation anomaly detection result is any type of fault operation state, an alarm message carrying real-time multidimensional sample data and the corresponding power plant equipment operation anomaly detection result is generated. After the alarm message is associated with a timestamp and stored, the alarm message is transmitted to the device designated by the staff.

[0068] This invention provides a deep learning-based method for detecting operational anomalies in power plant equipment. It acquires historical multidimensional sample data and their corresponding historical operating states, constructs a feature extraction sub-model using a first deep learning model, and a feature recognition sub-model using a second deep learning model. Based on these models, an operational anomaly detection model is obtained. Then, using the historical multidimensional sample data and their corresponding historical operating states as a foundation, an intelligent optimization algorithm is employed to train the operational anomaly detection model, resulting in a trained model. Finally, the trained model is used to identify real-time multidimensional sample data to determine the detection result of operational anomalies in the power plant equipment. Through the feature extraction and recognition processes, the method effectively analyzes the fault characteristics hidden in the operating data, thereby achieving rapid and accurate detection of operational anomalies in power plant equipment, avoiding false alarms and missed alarms, and enabling staff to perform timely maintenance and troubleshooting of the power plant equipment.

[0069] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0070] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] Those skilled in the art will understand that all or part of the steps in the above facts and methods can be implemented by a program instructing related hardware. The program or the program described therein can be stored in a computer-readable storage medium. When the program is executed, it includes the following steps: at this time, the corresponding method steps are introduced. The storage medium can be ROM / RAM, magnetic disk, optical disk, etc.

[0074] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting operational anomalies in power plant equipment based on deep learning, characterized in that, include: Obtain historical operation datasets for power plant equipment at historical operating times; wherein, the historical operation datasets include historical multidimensional sample data and their corresponding historical operating states; A feature extraction sub-model is constructed using a first deep learning model and a feature recognition sub-model is constructed using a second deep learning model. Based on the feature extraction sub-model and the feature recognition sub-model, an anomaly detection model is obtained. Based on the historical multidimensional sample data and its corresponding historical operating status, an intelligent optimization algorithm is used to train the operating anomaly detection model to obtain the trained operating anomaly detection model. Real-time multidimensional sample data corresponding to power plant equipment is collected, and the trained operation anomaly detection model is scheduled to identify the real-time multidimensional sample data to determine the power plant equipment operation anomaly detection result.

2. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 1, characterized in that, The historical multidimensional sample data includes the temperature, voltage, vibration, and current of the power plant equipment collected at a preset data sampling frequency within a data sampling period; the historical operating status corresponding to the historical multidimensional sample data is the fault operating status type or normal operating status type of the power plant equipment within the same data sampling period.

3. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 1, characterized in that, The step of constructing a feature extraction sub-model using a first deep learning model and a feature recognition sub-model using a second deep learning model, and obtaining an anomaly detection model based on the feature extraction sub-model and the feature recognition sub-model, includes: A feature extraction sub-model is constructed using a first deep learning model; wherein, the first deep learning model is set as a convolutional neural network; A feature recognition sub-model is constructed using a second deep learning model; wherein, the second deep learning model is set as a long short-term memory network; The feature output layer of the feature extraction sub-model is connected to the input layer of the feature recognition sub-model to obtain the running anomaly detection model.

4. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 1, characterized in that, The step of training the operational anomaly detection model using an intelligent optimization algorithm based on the historical multidimensional sample data and its corresponding historical operating states to obtain the trained operational anomaly detection model includes: Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the feature extraction sub-model in the operation anomaly detection model to obtain the final hyperparameters of the feature extraction sub-model. The final hyperparameters of the feature extraction sub-model are then fixed to obtain the operation anomaly detection model with a fixed feature extraction sub-model. Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the operating anomaly detection model of the fixed feature extraction sub-model to obtain the trained operating anomaly detection model.

5. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 4, characterized in that, Based on the historical multidimensional sample data and its corresponding historical operating states, an intelligent optimization algorithm is used to train the feature extraction sub-model in the operation anomaly detection model to obtain the final hyperparameters of the feature extraction sub-model, including: The hyperparameters of the feature extraction sub-model in the anomaly detection model are initialized and encoded to obtain multiple different first encoding vectors; For any first encoding vector, the loss function value corresponding to the first encoding vector is obtained based on the historical multidimensional sample data and its corresponding historical running state; The first encoding vector with the minimum loss function value is determined as the first optimal vector, and the first historical optimal state vector corresponding to the first encoding vector is determined based on the loss function value corresponding to the first encoding vector during the historical training process. Based on the first optimal vector and the first historical optimal state vector, a nonlinear joint guidance strategy is used to locally update the first encoding vector to determine the first encoding vector after the local update. A historical oscillation learning strategy is used to perform local oscillation updates on the first encoding vector after the local update, and the first encoding vector after the local oscillation update is determined. A global adaptive mutation strategy is used to perform a global adaptive update on the first encoding vector after the local oscillation update, and the first encoding vector after the global adaptive update is determined. Determine whether the current number of training iterations meets the preset maximum number of training iterations. If so, determine the final hyperparameters of the feature extraction sub-model based on the first encoding vector after global adaptive update. Otherwise, return to the step of obtaining the loss function value corresponding to the first encoding vector.

6. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 5, characterized in that, The step involves training the operational anomaly detection model of the fixed feature extraction sub-model using an intelligent optimization algorithm based on the historical multidimensional sample data and its corresponding historical operating states, to obtain the trained operational anomaly detection model, including: The other hyperparameters of the anomaly detection model of the fixed feature extraction sub-model are initialized and encoded to obtain multiple different second encoding vectors; For any second encoding vector, the loss function value corresponding to the second encoding vector is obtained based on the historical multidimensional sample data and its corresponding historical running state; The second encoding vector with the minimum loss function value is determined as the second optimal vector, and the second historical optimal state vector corresponding to the second encoding vector is determined based on the loss function value corresponding to the second encoding vector during the historical training process. Based on the second optimal vector and the second historical optimal state vector, a nonlinear joint guidance strategy is used to locally update the second encoding vector to determine the second encoding vector after the local update. A historical oscillation learning strategy is used to perform local oscillation updates on the second encoding vector after the local update, and the second encoding vector after the local oscillation update is determined. A global adaptive mutation strategy is used to perform a global adaptive update on the second encoding vector after the local oscillation update, and the second encoding vector after the global adaptive update is determined. Determine whether the current number of training iterations meets the preset maximum number of training iterations. If so, determine the training anomaly detection model based on the second encoding vector after global adaptive update. Otherwise, return to the step of obtaining the loss function value corresponding to the second encoding vector.

7. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 6, characterized in that, The nonlinear joint guidance strategy includes: Based on the target's optimal vector and its historical optimal state vector, the target update amount corresponding to the target encoding vector is determined as follows: Wherein, the target optimal vector is either the first optimal vector or the second optimal vector, the target historical optimal state vector is either the first historical optimal state vector or the second historical optimal state vector, and the target encoding vector is either the first encoding vector or the second encoding vector. Indicates the first t During the training process, the first i One target encoded vector, i =1,2,…,NP, where NP represents the total number of target encoding vectors. Represents the target optimal vector. Indicates the first t During the training process, the first i The target historical optimal state vector corresponding to each target encoding vector. Indicates the first t During the training process, the first i The target update amount corresponding to each target encoding vector. Indicates the first t +1 training session i The target update amount corresponding to each target encoding vector. Denotes the first joint update coefficient. Indicates the second joint update coefficient. e Represents the natural constant. This indicates the update range control coefficient. Represents the target encoding vector With the target's historical optimal state vector The Euclidean distance between them Represents the target encoding vector With the target optimal vector The Euclidean distance between them; The target encoding vector is locally updated based on the target update amount corresponding to the target encoding vector, resulting in the locally updated target encoding vector as follows: in, Indicates the first i The target encoding vector after local updates.

8. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 7, characterized in that, The historical oscillation learning strategy includes: in, Indicates the first t During the training process, the first j The first locally updated target encoding vector d Dimensional parameters, j =1,2,...,NP d =1,2,...,D; D represents the total dimension of the target encoding vector. Indicates the first j The target encoding vector after the local oscillation update is the first... d Dimensional parameters, Indicates the first t- During the first training session j The first locally updated target encoding vector d Dimensional parameters, Indicates the first t- During the 2nd training session j The first locally updated target encoding vector d Dimensional parameters, Indicates the first t- During the 3 training sessions, the first j The first locally updated target encoding vector d Dimensional parameters, This represents the historical oscillation adjustment coefficient.

9. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 8, characterized in that, The global adaptive mutation strategy includes: in, Indicates the first t During the training process, the first m The target encoding vector after the local oscillation update is the first... d Dimensional parameters, m =1,2,...,NP; Indicates the first m The target encoding vector after the global adaptive update of the nth global adaptive update d Dimensional parameters, The first element representing the global information vector d Dimensional parameters, Indicates the preset update step size upper limit. Compared with the preset update step size lower limit Random step size between This represents the update direction control factor between [-1, 1]. Represents pi (π). Represents a random number between [-1, 1]. Indicates the first intermediate parameter. The first element of the random encoded vector is... d Dimensional parameters, Indicates the first random update coefficient. Indicates the second random update coefficient. Represents a random number between (0, 2], and and It follows a normal distribution, that is... , ; Indicates the second intermediate parameter, and , Represents the gamma function. This represents the maximum value of the first intermediate parameter. This represents the minimum value of the first intermediate parameter, and T represents the maximum number of training iterations.

10. The method for detecting abnormal operation of power plant equipment based on deep learning according to claim 1, characterized in that, After determining the abnormal operation detection results of the power plant equipment, the process also includes: If the power plant equipment operation anomaly detection result is any type of fault operation state, an alarm message carrying real-time multidimensional sample data and the corresponding power plant equipment operation anomaly detection result is generated. After the alarm message is associated with a timestamp and stored, the alarm message is transmitted to the device designated by the staff.

Images