Method and apparatus for predicting health state of power production equipment

Abstract

The application relates to a health state prediction method and device of a power production equipment. The method comprises the following steps: acquiring multi-class heterogeneous physical modal data and operation semantic data of the power production equipment; performing alignment processing on the multi-class heterogeneous physical modal data to obtain a modal alignment feature vector; performing cleaning processing and encoding processing on the operation semantic data to obtain an operation semantic vector; performing weighted fusion processing on the modal alignment feature vector based on the operation semantic vector to obtain a semantic enhanced modal vector; inputting the operation semantic vector and the semantic enhanced modal vector into a health state prediction model constructed in advance respectively to obtain a health state score and a fault type of the power production equipment; and acquiring a health state prediction result of the power production equipment according to the health state score, the fault type and the operation semantic vector. The method can improve the identification accuracy and prediction stability of the fault precursor of the power production equipment.

Description

Technical Field

[0001] This application relates to the field of power equipment fault prediction technology, and in particular to a method and apparatus for predicting the health status of power production equipment. Background Technology

[0002] Key equipment in power production systems, such as steam turbines, generators, main transformers, GIS equipment, and switchgear, operate under complex environments of high voltage, high temperature, and high load for extended periods. Their operational stability and health status directly affect the safety and economy of the entire power system.

[0003] However, current mainstream health status prediction methods are still mainly focused on analysis and shallow statistical modeling based on single-mode signals, such as current and voltage trend analysis, vibration spectrum diagnosis, or life regression prediction. These methods rely excessively on manual feature extraction, have poor model generalization, and are insufficient in representing nonlinear degradation behavior under complex working conditions, resulting in poor reliability and interpretability of prediction results. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and apparatus for predicting the health status of power production equipment that can improve the accuracy of prediction results, addressing the aforementioned technical problems.

[0005] In a first aspect, this application provides a method for predicting the health status of power generation equipment, the method comprising:

[0006] Acquire multi-type heterogeneous physical mode data and operational semantic data of power production equipment;

[0007] The heterogeneous physical modal data are aligned to obtain modal alignment feature vectors;

[0008] The runtime semantic data is cleaned and encoded to obtain runtime semantic vectors;

[0009] Based on the running semantic vector, the modality alignment feature vector is subjected to weighted fusion processing to obtain a semantically enhanced modality vector;

[0010] The operational semantic vector and the semantically enhanced modal vector are respectively input into a pre-built health status prediction model to obtain the health status score and fault type of the power production equipment;

[0011] Based on the health status score, the fault type, and the operational semantic vector, the health status prediction result of the power production equipment is obtained.

[0012] In one embodiment, the health status prediction model includes a semantic modality gating model and a master prediction network model;

[0013] The step of inputting the operational semantic vector and the semantically enhanced modal vector into a pre-built health status prediction model to obtain the health status score and fault type of the power production equipment includes:

[0014] The semantic modality gating model maps the running semantic vector to gating weights, and adjusts the semantic enhancement modality vector element by element to obtain the semantic control state vector.

[0015] The semantic control state vector and the running semantic vector are concatenated to obtain the overall state vector;

[0016] The overall state vector is processed by feature extraction and nonlinear transformation using the master prediction network model to obtain the health status score and fault type of the power production equipment.

[0017] In one embodiment, the method further includes:

[0018] Obtain training samples; the training samples include health status score labels and fault type labels;

[0019] The training samples are input into a pre-built health status prediction model for training, in order to obtain the semantic control state vector, health status score and fault type of the training samples;

[0020] Based on the semantic control state vector of the training samples, determine the modality distribution constraint regularization term of the model loss function;

[0021] The model loss function value is determined based on the health status score and fault type of the training samples, the modal distribution constraint regularization term, the health status score label value, and the fault label label value.

[0022] The model parameters of the health status prediction model are optimized based on the model loss function value.

[0023] In one embodiment, determining the modality distribution constraint regularization term of the model loss function based on the semantic control state vector of the training samples includes:

[0024] Obtain the variance of each modality dimension in the semantic control state vector of a preset number of training samples;

[0025] Obtain the mean variance of each modality dimension;

[0026] The modal distribution constraint regularization term is calculated based on the variance mean and the preset regularization term weight.

[0027] In one embodiment, obtaining the health status prediction result of the power production equipment based on the health status score, the fault type, and the operational semantic vector includes:

[0028] To determine the importance of the power generation equipment in the power system;

[0029] Perform a linear transformation on the operational semantic vector to obtain the operational condition risk level;

[0030] The health risk value of the power production equipment is determined based on the importance of the power production equipment, the risk level of the operating condition, and the health status score.

[0031] The alarm intensity value is determined based on the health risk value and the hazard weight corresponding to the fault type;

[0032] Based on the alarm intensity value, the fault level of the power production equipment is determined and corresponding maintenance recommendations are generated.

[0033] In one embodiment, determining the health risk value of the power production equipment based on the importance of the power production equipment, the operating condition risk level, and the health status score includes:

[0034] The health risk value is calculated using the following formula:

[0035]

[0036] in, Indicates health risk value; Indicates the importance of power generation equipment; Indicates a health status score; Indicates the risk level of the operating condition; and These represent the preset adjustment coefficients.

[0037] In one embodiment, the modality alignment feature vector includes feature vectors of multiple different physical modes;

[0038] The step of performing a weighted fusion process on the modality alignment feature vector based on the running semantic vector to obtain a semantically enhanced modality vector includes:

[0039] The semantic vector and each feature vector are subjected to similarity calculation, exponential transformation and normalization to determine the attention weight of each feature vector;

[0040] Based on the attention weights of each feature vector, the feature vectors are weighted and summed to obtain the semantic enhancement modality vector.

[0041] In one embodiment, the runtime semantic data is cleaned and encoded to obtain a runtime semantic vector, including:

[0042] The runtime semantic data is cleaned.

[0043] By using a gated recursive unit structure with two hidden state directions, each word in the cleaned semantic data is processed, and the mean of the word processing results is obtained.

[0044] Perform named entity-based regular expression operations on the cleaned semantic data to obtain the results of the named entity regular expression operations;

[0045] The semantic vector is determined based on the average of the named entity regular expression operation result and the word processing result.

[0046] In one embodiment, the multi-type heterogeneous physical mode data includes temperature image signals, structured acoustic signals, and electrical operation signals; the mode alignment feature vector includes thermal mode feature vectors, acoustic mode feature vectors, and electrical mode feature vectors.

[0047] The alignment process for the multi-class heterogeneous physical modal data to obtain modal alignment feature vectors includes:

[0048] Using the image frame of the temperature image signal as a reference clock, the structural acoustic signal and the electrical operation signal are pooled within a preset time window to obtain time-aligned temperature signal, acoustic signal and electrical signal;

[0049] The temperature signal, the acoustic signal, and the electrical signal are respectively subjected to structural encoding processing to obtain thermal mode feature vectors, acoustic mode feature vectors, and electrical mode feature vectors with the same vector length.

[0050] Secondly, this application also provides a health status prediction device for power generation equipment, comprising:

[0051] The data acquisition module is used to acquire various heterogeneous physical mode data and operational semantic data of power production equipment;

[0052] The encoding module is used to align the multi-type heterogeneous physical modal data to obtain modal alignment feature vectors; and to clean and encode the runtime semantic data to obtain runtime semantic vectors.

[0053] The weighted fusion module is used to perform weighted fusion processing on the modality alignment feature vector based on the running semantic vector to obtain a semantically enhanced modality vector;

[0054] The health status prediction module is used to input the operational semantic vector and the semantically enhanced modal vector into a pre-built health status prediction model to obtain the health status score and fault type of the power production equipment; and to obtain the health status prediction result of the power production equipment based on the health status score, the fault type and the operational semantic vector.

[0055] In the aforementioned method and apparatus for predicting the health status of power production equipment, multiple heterogeneous physical modal data and operational semantic data of the power production equipment are acquired. The heterogeneous physical modal data are then aligned to obtain modality alignment feature vectors, ensuring the collaborative expression of different physical modes within a unified temporal framework. This facilitates subsequent matching with operational semantic vectors. The operational semantic data is then cleaned and encoded to obtain operational semantic vectors. Based on these operational semantic vectors, the modality alignment feature vectors are weighted and fused to obtain semantically enhanced modality vectors. These generated semantically enhanced modality vectors better reflect the actual operating scenarios of the equipment, avoiding... This approach addresses the disconnect between physical characteristics and operating conditions, providing high-quality input for subsequent health status prediction models. Subsequently, the operational semantic vector and semantically enhanced modal vector are input into the pre-constructed health status prediction model to obtain health status scores and fault types for power production equipment. The weight allocation of the semantically enhanced modal vector is dynamically adjusted by running the semantic vector, accurately matching the changing patterns of modal importance under different operating conditions. Finally, based on the health status score, fault type, and operational semantic vector, a health status prediction result containing health level, maintenance requirements, and specific operational suggestions is generated, achieving a closed loop from data prediction to decision guidance. The health status prediction provided in this application not only improves the accuracy and stability of identifying precursors to catastrophic power equipment failures but also makes the prediction results structurally interpretable and engineering-executable. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart illustrating a method for predicting the health status of power generation equipment in one embodiment;

[0058] Figure 2 This is a flowchart illustrating the process of inputting semantic vectors and semantically enhanced modal vectors into a pre-built health status prediction model to obtain the health status score and fault type of power production equipment in one embodiment.

[0059] Figure 3 This is a flowchart illustrating the training process of a health status prediction model in one embodiment.

[0060] Figure 4 This is a flowchart illustrating the process of obtaining health status prediction results for power production equipment based on health status score, fault type, and operational semantic vector in one embodiment.

[0061] Figure 5 This is a flowchart illustrating the process of cleaning and encoding runtime semantic data to obtain runtime semantic vectors in one embodiment.

[0062] Figure 6 This is a flowchart illustrating the process of aligning multiple types of heterogeneous physical modal data to obtain modal alignment feature vectors in one embodiment.

[0063] Figure 7 This is a structural block diagram of a health status prediction device for power generation equipment in one embodiment;

[0064] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0066] In one embodiment, this application provides a method for predicting the health status of power generation equipment, such as... Figure 1 As shown, it includes the following S102-S112.

[0067] S102, acquire multi-type heterogeneous physical mode data and operational semantic data of power production equipment.

[0068] Power generation equipment includes steam turbines, generators, main transformers, GIS equipment, switchgear, and other power generation or supply equipment. Multi-type heterogeneous physical modal data refers to different types of physical sensing signals, such as temperature image signals, acoustic signals, and electrical operation signals. Temperature image signals can be acquired through infrared thermal imaging sensors installed on the surface of the power generation equipment; acoustic signals can be acquired through piezoelectric transducers near the equipment; and electrical operation signals can be acquired through voltage or current sampling modules installed at cable entry points. Operational semantic data refers to semantic data generated during the operation of power generation equipment, which may include scheduling text instructions, operation log fragments, and maintenance plan content. Operational semantic data can be obtained through the scheduling coefficients of the power generation equipment, log servers, and maintenance management platforms.

[0069] S104 performs alignment processing on multiple types of heterogeneous physical modal data to obtain modal alignment feature vectors.

[0070] Heterogeneous physical modal data differ in sampling frequency, physical properties, and data structure. For example, temperature image signals are sampled at 1 frame per second, sound signals at 24,000 times per second, and electrical operation signals at 1,000 times per second. Therefore, it is necessary to align these heterogeneous physical modal data both temporally and structurally to facilitate subsequent matching with operational semantic data. The modal alignment feature vector is a set of feature vectors from the temporally and structurally aligned heterogeneous physical modal data.

[0071] S106, perform cleaning and encoding processing on the runtime semantic data to obtain the runtime semantic vector.

[0072] Cleaning processes may include punctuation normalization, field standardization, word segmentation, and named entity recognition. Encoding can be performed using a pre-built semantic encoder. The resulting semantic vector represents the scheduling and operating status of power generation equipment within the current time window.

[0073] S108, based on the running semantic vector, performs weighted fusion processing on the modality alignment feature vector to obtain the semantically enhanced modality vector.

[0074] Semantic augmented modal vectors represent the device operating state after incorporating semantic bias information.

[0075] S110, the running semantic vector and semantically enhanced modal vector are respectively input into the pre-built health status prediction model to obtain the health status score and fault type of the power production equipment.

[0076] Health status prediction models are used to predict the health status score and fault type of power generation equipment. The health status score reflects the Euclidean distance or deviation of the current state of the power generation equipment from its normal state, while the fault type refers to the faults that the power generation equipment may experience within the current time window.

[0077] Due to the following characteristics in the operation of power generation equipment: faults are often triggered by early, weak anomalies in a few modes, while other modes respond with a lag; changes in operating conditions significantly affect mode weights, and the importance of modes varies with semantics under different scheduling states; different faults exhibit mode preference, for example, mechanical loosening is first manifested by abnormal acoustic signals, and overheating faults are reflected first by thermal modes. Therefore, the weights of the semantically enhanced mode vectors that align with the semantic background in the health status prediction model can be dynamically adjusted based on the operational semantic vector, thereby improving the accuracy of the model's prediction results.

[0078] S112, based on health status score, fault type and operation semantic vector, obtain the health status prediction results of power production equipment.

[0079] Health status prediction results include the current health status level of power production equipment, whether maintenance is required, and recommended operations if maintenance is needed. It is understandable that while health status scores and fault types can reflect the fault conditions of power production equipment, they cannot guide dispatchers in performing operation and maintenance operations. Therefore, it is necessary to convert health status scores and fault types into health status prediction results that can be used for decision-making.

[0080] In this embodiment, by acquiring multiple heterogeneous physical modal data and operational semantic data of power production equipment, the heterogeneous physical modal data is aligned to obtain modal alignment feature vectors. This ensures the collaborative expression of different physical modes within a unified temporal framework, facilitating subsequent matching with operational semantic vectors. The operational semantic data is cleaned and encoded to obtain operational semantic vectors. Based on these operational semantic vectors, the modal alignment feature vectors are weighted and fused to obtain semantically enhanced modal vectors. These generated semantically enhanced modal vectors better fit the actual operating scenarios of the equipment, avoiding the problem of physical features being disconnected from operating conditions, and providing high-quality input for subsequent health status prediction models. Subsequently, the operational semantic vectors and semantically enhanced modal vectors are input into a pre-constructed health status prediction model to obtain the health status score and fault type of the power production equipment. The weight allocation of the semantically enhanced modal vectors in the model is dynamically adjusted using the operational semantic vectors to accurately match the changing patterns of modal importance under different operating conditions. Finally, based on the health status score, fault type, and operational semantic vectors, a health status prediction result containing health level, maintenance requirements, and specific operational suggestions is generated, achieving a closed loop from data prediction to decision guidance. The health status prediction provided in this application not only improves the accuracy and stability of identifying precursors to catastrophic failures of power equipment, but also makes the prediction results structurally interpretable and engineering feasible.

[0081] In one embodiment, the health status prediction model includes a semantic modality gating model and a main prediction network model. The semantic modality gating model is a modality dimension gating module under semantic conditions, which can adaptively adjust the expression strength of each dimension in the semantically enhanced modality vector based on the running semantic vector. The main prediction network uses two layers of nonlinear transformation for feature extraction, each layer consisting of a fully connected structure, batch normalization, and an activation function (ReLU). The main prediction network has two branches: a health score branch and a fault type branch. The health score branch uses interval regression, outputting values ​​between [0,1]; the fault type branch outputs multi-class prediction results, representing the most likely fault type of the device under the current time window. The fault type can be in one-hot format; for example, if the output fault type is represented as [0,0,1,0], it indicates that the currently predicted fault type is "stator winding insulation aging".

[0082] In one embodiment, the running semantic vector and semantically enhanced modal vector are input into a pre-built health status prediction model to obtain the health status score and fault type of the power production equipment, including the following S202-S206.

[0083] S202 uses a semantic modality gating model to map the running semantic vector to gating weights, and adjusts the semantic enhancement modality vector element by element to obtain the semantic control state vector.

[0084] The mathematical expression of the semantic modality gating model can be:

[0085]

[0086] in, Represents the semantic control state vector. These are the trainable parameters. This is the bias term, where d represents the vector length of the running semantic vector and the semantic enhancement modality vector, both of which are d. This represents the semantic vector of execution. Represents a semantically enhanced modal vector. This indicates element-wise multiplication.

[0087] This structure can dynamically suppress modal dimensions unrelated to the current operating state and strengthen those consistent with the semantic context. For example, it strengthens the electrical signal dimension during "nighttime heavy load switching" and weakens the acoustic modal dimension during "post-maintenance test operation". The activation function uses "tanh" to provide adjustment capabilities for both strongly positively and strongly negatively correlated dimensions, adapting to situations where modal responses may exhibit strong fluctuations or significant suppression in real-world scenarios.

[0088] S204, concatenate the semantic control state vector and the running semantic vector to obtain the overall state vector.

[0089] S206 uses a master prediction network model to perform feature extraction and nonlinear transformation on the overall state vector to obtain the health status score and fault type of power production equipment.

[0090] In one embodiment, the training process for the health status prediction model includes the following steps S302-S310.

[0091] S302, Obtain training samples.

[0092] The training samples include health status score labels and fault type labels. The health status score labels are the known health status scores, and the fault type labels are the known fault types.

[0093] S304. Input the training samples into the pre-built health status prediction model for training, so as to obtain the semantic control state vector, health status score and fault type of the training samples.

[0094] Refer to S202-S206 mentioned above to obtain the semantic control state vector, health status score and fault type of the training sample.

[0095] S306. Based on the semantic control state vector of the training samples, determine the modal distribution constraint regularization term of the model loss function.

[0096] Modal distribution constraint regularization terms can be used to improve the overall discreteness of modal dimension representation, preventing the model from concentrating all discriminative power on a few modal dimensions, which would lead to a decrease in sensitivity to minor faults.

[0097] As shown in Equation (2), the variance of each modality dimension in the semantic control state vector of a preset number of training samples can be obtained, the mean variance of each modality dimension can be obtained, and the modality distribution constraint regularization term can be calculated based on the mean variance and the preset regularization term weight.

[0098]

[0099] in, As the weight of the regularization term, in the case of to Range of values, Indicates the first The variance of each dimension in a small batch of samples.

[0100] S308. Determine the model loss function value based on the health status score and fault type of the training samples, the modal distribution constraint regularization term, the health status score label value, and the fault type label value.

[0101] The model loss function can be expressed as:

[0102]

[0103] in, and These represent the health status score and fault type predicted by the model, respectively. and Label the health status score and the fault type.

[0104] S310, optimize the model parameters of the health status prediction model based on the model loss function value.

[0105] The model parameters of the health status prediction model are optimized based on the model loss function value to minimize the model loss function value. After the model training is completed, during normal use (i.e., in steps S202-S206 above), the model parameters are frozen, and there is no need to calculate the model loss function; the health status score and fault type can be directly output.

[0106] In one embodiment, the health status prediction result of the power production equipment is obtained based on the health status score, fault type and operation semantic vector, including the following S402-S410.

[0107] S402, Obtain the importance of power generation equipment in the power system.

[0108] The importance of equipment is determined by its basic equipment profile, such as the main transformer of a substation being configured as follows: The power distribution cabinet is set up Switchgear is set to These weights can be set uniformly during system deployment.

[0109] S404 performs a linear transformation on the runtime semantic vector to obtain the runtime risk level.

[0110] The operational risk level can be determined by the operational semantic vector. After linear transformation, it is calculated that, for example, when high-risk scheduling statements such as "reconnect to the grid after maintenance" or "load switching" appear in the current operating semantics, the risk level of the operating condition can be increased to [a higher level]. above.

[0111] S406 determines the health risk value of power production equipment based on its importance, operating condition risk level, and health status score.

[0112] The health risk score can be calculated using the following formula:

[0113]

[0114] in, Indicates health risk value; Indicates the importance of power generation equipment; Indicates a health status score; Indicates the risk level of the operating condition; and These represent preset adjustment coefficients. For example, and They can be respectively and This can be used to control the nonlinearity of risk response and ensure that... The risk value increases rapidly when it is around 0.5 and less than 0.5, which is consistent with the actual operation and maintenance logic of power equipment: "mild degradation is observable, severe degradation is intolerable".

[0115] S408 determines the alarm intensity value based on the health risk value and the hazard weight corresponding to the fault type.

[0116] The hazard weight indicates the degree of impact of a fault type on equipment operation and maintenance, and can be set based on a historical incident database and human experience. For example, the hazard weight corresponding to winding overheating is set to... The corresponding danger weight for core vibration is set as follows: The hazard weight corresponding to insulation aging is set as follows: .

[0117] The high static intensity value integrates the importance of health risk value and failure type. It can be based on equation (5), according to the health risk value. Risk weights corresponding to fault types Determine the alarm intensity value .

[0118]

[0119] S410 determines the fault level of power production equipment based on the alarm intensity value and generates corresponding maintenance suggestions.

[0120] Fault levels are classified into four grades based on alarm intensity values:

[0121] when At that time, it was assumed that the power production equipment was in normal condition and was only recorded in the archives.

[0122] when At that time, a routine alert is generated, prompting dispatchers to include the power production equipment in their daily inspection checklist.

[0123] when When an anomaly occurs, an early warning alert will be generated, and an on-site inspection will be recommended within the next 3 days, along with suggested inspection items.

[0124] when When necessary, a maintenance instruction alarm is generated, and a maintenance priority ranking is recommended, pushing the equipment to the high-priority queue of the maintenance schedule.

[0125] Furthermore, by combining the fault type and corresponding modal anomaly characteristics, and matching historical fault case databases with maintenance records, specific execution operations can be generated. For example, if the predicted fault type is "stator winding insulation aging," corresponding cases can be found and combined with the current modal anomaly source, such as significant thermal modal deviation, and suggestions can be automatically generated: "It is recommended to perform insulation resistance testing on the stator winding and review the local characteristics of the winding to detect whether there is localized heat distribution."

[0126] In this embodiment, by calculating the health risk value based on the health status score, fault type, and operational semantic vector, the accurate fusion of multi-dimensional factors is achieved. Then, by combining the health risk value and the degree of danger of the fault type, an alarm intensity value is generated. Based on the alarm intensity value, the corresponding danger level is divided and maintenance suggestions are generated, which can reduce operation and maintenance costs and improve decision-making efficiency.

[0127] In one embodiment, the modal alignment feature vector includes feature vectors of multiple different physical modes. For example, the modal alignment feature vector includes thermal modal feature vectors, acoustic modal feature vectors, and electrical modal feature vectors.

[0128] Based on the running semantic vector, the modality alignment feature vector is weighted and fused to obtain the semantically enhanced modality vector, including: as in equation (6), similarity calculation, exponential transformation and normalization are performed on the running semantic vector and each feature vector to determine the attention weight of each feature vector; as in equation (7), based on the attention weight of each feature vector, the feature vector is weighted and summed to obtain the semantically enhanced modality vector.

[0129]

[0130] in, The weights of the eigenvectors are represented by m. , To distinguish between different physical modes, the subscript h represents thermal modes, the subscript a represents acoustic modes, and the subscript e represents electrical modes. k is the index. W represents the aligned transformation torque. The superscript T indicates that the vector matrix has been transposed.

[0131]

[0132] in, This represents a semantically enhanced modal vector.

[0133] In this embodiment, the semantic vector is used as a contextual attention condition to guide the model to focus on the most relevant modal dimension, which is particularly suitable for situations in power scenarios where anomalies occur under specific operating conditions. For example, "the dispatch instruction is a heavy load transfer" may make the electrical mode more important than the acoustic mode, while "the shift worker reports abnormal equipment noise" should increase the proportion of the acoustic mode, thereby improving the overall model's ability to distinguish under different operating conditions.

[0134] In one embodiment, such as Figure 5 As shown, the runtime semantic data is cleaned and encoded to obtain the runtime semantic vector, including the following S502-S508.

[0135] S502 performs cleaning and processing on the semantic data.

[0136] S504 utilizes a gated recursive unit structure with two hidden state directions to process each word in the cleaned semantic data and obtain the average of the word processing results.

[0137] S506, Perform named entity-based regular expression operations on the cleaned semantic data to obtain the named entity regular expression operation results.

[0138] S508, determine the running semantic vector based on the average of the named entity regularization operation result and the word processing result.

[0139] Steps S502-S508 can be completed by a pre-built semantic encoder, which can consist of a word embedding lookup layer and a bidirectional recurrent neural network. The mathematical model of the semantic encoder can be expressed as:

[0140]

[0141] in, Indicates the semantic vector of execution; Indicates the length of the semantic data being processed; For indexing; This represents the runtime semantic data after normalized cleaning. This represents a gated recursive unit structure with two hidden state directions; This represents the embedding representation of the i-th word in the normalized and cleaned semantic data. Before use, it can be obtained from the training corpus. Training can be conducted using dispatch texts from the power industry, covering keywords such as equipment status descriptions, dispatch commands, and load change descriptions. This represents a regular expression term for named entities. This represents the regularization control coefficient, which forces the model to assign higher feature weights to words such as "device name", "operation action", and "operation status", thereby enhancing the semantic encoder's ability to focus on key semantic segments.

[0142] In one embodiment, such as Figure 6 As shown, alignment processing is performed on multiple types of heterogeneous physical modal data to obtain modal alignment feature vectors, including the following S602-S604.

[0143] S602 uses the image frame of the temperature image signal as a reference clock, and performs pooling processing on the structural acoustic signal and electrical operation signal respectively within a preset time window to obtain the time-aligned temperature signal, acoustic signal and electrical signal.

[0144] The process of pooling structural acoustic signals and electrical operation signals can be represented as follows:

[0145]

[0146]

[0147] in, and These represent time-aligned acoustic and electrical signals, respectively. and These represent structural acoustic signals and electrical operating signals, respectively. This indicates the length of the time window. The pooling process can also be understood as the process of averaging the structural acoustic signal and the electrical operating signal within the time window.

[0148] S604 performs structural encoding processing on temperature signals, acoustic signals, and electrical signals respectively to obtain thermal mode feature vectors, acoustic mode feature vectors, and electrical mode feature vectors with the same vector length.

[0149] Specialized structured coding networks can be constructed for each of the three types of time-aligned modal data. Temperature signal The data is fed into a two-dimensional feature extraction structure, and after being processed sequentially through two two-dimensional convolutional layers, a normalization layer, and an activation function, it is flattened into a vector to generate a thermal mode feature vector. . acoustic signal After processing by a one-dimensional convolutional network to extract local spectral modes, the data is compressed into acoustic modal feature vectors via a fully connected layer. Electrical signals The data is fed into a bidirectional recurrent network structure to extract the dynamic information contained in the time series, and the output is the state concatenation result at the last moment, i.e., the electrical mode feature vector. .

[0150] The mathematical model for this process can be represented as:

[0151]

[0152] in, This represents a two-dimensional structure encoder used for thermal modes, consisting of two convolutional, normalization, and flattening layers; This represents a one-dimensional convolutional encoder for acoustic modes, including feature extraction and compression transformation modules; This represents a bidirectional sequence modeler for electrical modes, internally using gated recursive units in two directions. The final output... , , All are of length The vector representation of , It is the set of alignment vectors for three types of modalities, i.e., modal alignment feature vectors.

[0153] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0154] Based on the same inventive concept, this application also provides a power production equipment health status prediction device for implementing the above-mentioned power production equipment health status prediction method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations of one or more power production equipment health status prediction device embodiments provided below can be found in the limitations of the power production equipment health status prediction method described above, and will not be repeated here.

[0155] In one exemplary embodiment, such as Figure 7 As shown, a health status prediction device for power generation equipment is provided, comprising:

[0156] The data acquisition module 702 is used to acquire various heterogeneous physical mode data and operational semantic data of power production equipment;

[0157] The encoding module 704 is used to align multi-type heterogeneous physical modal data to obtain modal alignment feature vectors; and to clean and encode runtime semantic data to obtain runtime semantic vectors.

[0158] The weighted fusion module 706 is used to perform weighted fusion processing on the modality alignment feature vector based on the running semantic vector to obtain the semantically enhanced modality vector;

[0159] The health status prediction module 708 is used to input the operating semantic vector and the semantic enhancement modal vector into the pre-built health status prediction model to obtain the health status score and fault type of the power production equipment; and to obtain the health status prediction result of the power production equipment based on the health status score, fault type and operating semantic vector.

[0160] In one embodiment, the health status prediction module is further configured to: map the running semantic vector to gating weights using a semantic modality gating model, adjust the semantic enhancement modality vector element by element to obtain a semantic control state vector; concatenate the semantic control state vector and the running semantic vector to obtain an overall state vector; and perform feature extraction and nonlinear transformation processing on the overall state vector using a master prediction network model to obtain the health status score and fault type of the power production equipment.

[0161] In one embodiment, the health status prediction device for power production equipment further includes a pre-training module, which is used to: acquire training samples; the training samples include health status score labels and fault type labels; input the training samples into a pre-built health status prediction model for training to obtain the semantic control state vector, health status score, and fault type of the training samples; determine the modal distribution constraint regularization term of the model loss function based on the semantic control state vector of the training samples; determine the model loss function value based on the health status score and fault type, modal distribution constraint regularization term, health status score labels, and fault type labels of the training samples; and optimize the model parameters of the health status prediction model based on the model loss function value.

[0162] In one embodiment, the pre-training module is further configured to: obtain the variance of each modality dimension in the semantic control state vector of a preset number of training samples; obtain the mean variance of each modality dimension; and calculate the modality distribution constraint regularization term based on the mean variance and the preset regularization term weight.

[0163] In one embodiment, the health status prediction module is further configured to: obtain the importance of power production equipment in the power system; perform linear transformation processing on the operation semantic vector to obtain the operation condition risk level; determine the health risk value of the power production equipment based on the importance of the power production equipment, the operation condition risk level, and the health status score; determine the alarm intensity value based on the health risk value and the hazard weight corresponding to the fault type; and determine the fault level of the power production equipment based on the alarm intensity value and generate corresponding maintenance suggestions.

[0164] In one embodiment, the weighted fusion module is further configured to: perform similarity calculation, exponential transformation and normalization on the running semantic vector and each feature vector to determine the attention weight of each feature vector; and perform weighted summation on each feature vector based on the attention weight of each feature vector to obtain the semantic enhancement modality vector.

[0165] In one embodiment, the encoding module is further configured to: clean the runtime semantic data; process each word in the cleaned runtime semantic data using a gated recursive unit structure with two hidden state directions, and obtain the mean of the word processing results; perform named entity-based regularization operations on the cleaned runtime semantic data, and obtain the named entity regularization operation results; and determine the runtime semantic vector based on the mean of the named entity regularization operation results and the word processing results.

[0166] In one embodiment, the weighted fusion module is further configured to: use the image frame of the temperature image signal as a reference clock, and within a preset time window length, perform pooling processing on the structure acoustic signal and the electrical operation signal respectively to obtain time-aligned temperature signal, acoustic signal and electrical signal; and perform structure encoding processing on the temperature signal, acoustic signal and electrical signal respectively to obtain thermal mode feature vector, acoustic mode feature vector and electrical mode feature vector with the same vector length.

[0167] Each module in the aforementioned power production equipment health status prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0168] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for predicting the health status of power production equipment. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0169] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0170] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the health status prediction method for power production equipment provided in any of the above embodiments.

[0171] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the health status prediction method for power production equipment provided in any of the above embodiments.

[0172] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the health status prediction method for power generation equipment provided in any of the above embodiments.

[0173] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0174] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0175] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting the health status of power generation equipment, characterized in that, The method includes: Acquire multi-type heterogeneous physical mode data and operational semantic data of power production equipment; The heterogeneous physical modal data are aligned to obtain modal alignment feature vectors; The runtime semantic data is cleaned and encoded to obtain runtime semantic vectors; Based on the running semantic vector, the modality alignment feature vector is subjected to weighted fusion processing to obtain a semantically enhanced modality vector; The operational semantic vector and the semantically enhanced modal vector are respectively input into a pre-constructed health status prediction model to obtain the health status score and fault type of the power production equipment. The health status prediction model includes a semantic modal gating model and a main prediction network model. Specifically, through the semantic modal gating model, the operational semantic vector is mapped to gating weights, and the semantically enhanced modal vector is adjusted element-wise to obtain a semantic control state vector. The semantic control state vector and the operational semantic vector are concatenated to obtain an overall state vector. Through the main prediction network model, feature extraction and nonlinear transformation processing are performed on the overall state vector to obtain the health status score and fault type of the power production equipment. Based on the health status score, the fault type, and the operational semantic vector, the health status prediction result of the power production equipment is obtained.

2. The method according to claim 1, characterized in that, The method further includes: Obtain training samples; the training samples include health status score labels and fault type labels; The training samples are input into a pre-built health status prediction model for training, in order to obtain the semantic control state vector, health status score and fault type of the training samples; Based on the semantic control state vector of the training samples, determine the modality distribution constraint regularization term of the model loss function; The model loss function value is determined based on the health status score and fault type of the training samples, the modal distribution constraint regularization term, the health status score label value, and the fault type label value. The model parameters of the health status prediction model are optimized based on the model loss function value.

3. The method according to claim 2, characterized in that, The step of determining the modality distribution constraint regularization term of the model loss function based on the semantic control state vector of the training samples includes: Obtain the variance of each modality dimension in the semantic control state vector of a preset number of training samples; Obtain the mean variance of each modality dimension; The modal distribution constraint regularization term is calculated based on the variance mean and the preset regularization term weight.

4. The method according to claim 1, characterized in that, The step of obtaining the health status prediction result of the power production equipment based on the health status score, the fault type, and the operational semantic vector includes: To determine the importance of the power generation equipment in the power system; Perform a linear transformation on the operational semantic vector to obtain the operational condition risk level; The health risk value of the power production equipment is determined based on the importance of the power production equipment, the risk level of the operating condition, and the health status score. The alarm intensity value is determined based on the health risk value and the hazard weight corresponding to the fault type; Based on the alarm intensity value, the fault level of the power production equipment is determined and corresponding maintenance recommendations are generated.

5. The method according to claim 4, characterized in that, The process of determining the health risk value of the power production equipment based on its importance, operating condition risk level, and health status score includes: The health risk value is calculated using the following formula: in, Indicates health risk value; Indicates the importance of power generation equipment; Indicates a health status score; Indicates the risk level of the operating condition; and These represent the preset adjustment coefficients.

6. The method according to claim 1, characterized in that, The modality alignment feature vector includes feature vectors of multiple different physical modes; The step of performing a weighted fusion process on the modality alignment feature vector based on the running semantic vector to obtain a semantically enhanced modality vector includes: The semantic vector and each feature vector are subjected to similarity calculation, exponential transformation and normalization to determine the attention weight of each feature vector; Based on the attention weights of each feature vector, the feature vectors are weighted and summed to obtain the semantic enhancement modality vector.

7. The method according to claim 1, characterized in that, The runtime semantic data is cleaned and encoded to obtain runtime semantic vectors, including: The runtime semantic data is cleaned. By using a gated recursive unit structure with two hidden state directions, each word in the cleaned semantic data is processed, and the mean of the word processing results is obtained. Perform named entity-based regular expression operations on the cleaned semantic data to obtain the named entity regular expression operation results; The semantic vector is determined based on the average of the named entity regular expression operation result and the word processing result.

8. The method according to claim 1, characterized in that, The heterogeneous physical modal data includes temperature image signals, structural acoustic signals, and electrical operation signals; the modal alignment feature vectors include thermal modal feature vectors, acoustic modal feature vectors, and electrical modal feature vectors. The alignment process for the multi-type heterogeneous physical modal data to obtain modal alignment feature vectors includes: Using the image frame of the temperature image signal as a reference clock, the structural acoustic signal and the electrical operation signal are pooled within a preset time window to obtain time-aligned temperature signal, acoustic signal and electrical signal; The temperature signal, the acoustic signal, and the electrical signal are respectively subjected to structural encoding processing to obtain thermal mode feature vectors, acoustic mode feature vectors, and electrical mode feature vectors with the same vector length.

9. A health status prediction device for power production equipment, characterized in that, include: The data acquisition module is used to acquire various heterogeneous physical mode data and operational semantic data of power production equipment; The encoding module is used to perform alignment processing on the multi-type heterogeneous physical modal data to obtain modal alignment feature vectors; The runtime semantic data is cleaned and encoded to obtain runtime semantic vectors; The weighted fusion module is used to perform weighted fusion processing on the modality alignment feature vector based on the running semantic vector to obtain a semantically enhanced modality vector; A health status prediction module is used to input the operational semantic vector and the semantically enhanced modal vector into a pre-constructed health status prediction model to obtain the health status score and fault type of the power production equipment. The health status prediction model includes a semantic modal gating model and a main prediction network model. Specifically, through the semantic modal gating model, the operational semantic vector is mapped to gating weights, and the semantically enhanced modal vector is adjusted element-wise to obtain a semantic control state vector. The semantic control state vector and the operational semantic vector are concatenated to obtain an overall state vector. Through the main prediction network model, feature extraction and nonlinear transformation processing are performed on the overall state vector to obtain the health status score and fault type of the power production equipment. Based on the health status score, the fault type, and the operational semantic vector, the health status prediction result of the power production equipment is obtained.

Images