A dry-type air-core reactor indoor operation fault prediction method, medium and system

Abstract

The application provides a dry-type air-core reactor indoor operation fault prediction method, medium and system, and belongs to the technical field of reactors, and comprises the following: the method can monitor the operation state of the dry-type air-core reactor in real time, extracts the features of various signals by using a pre-trained feature extraction model through multi-source information such as electrical parameters, thermal imaging, acoustic imaging and vibration signals, and inputs the features into a fault diagnosis model to obtain the fault type and occurrence probability, so as to predict the fault condition of the reactor. The multi-source information fusion method can more comprehensively evaluate the operation state of the reactor, help to discover and diagnose faults in time, and improve the operation reliability of the reactor. The method provides a feasible solution for intelligent monitoring and fault diagnosis of the dry-type air-core reactor, and solves the technical problem that existing methods usually rely on the experience of professionals to judge faults and are difficult to predict future fault development trends.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric reactors, and particularly relates to a dry-type air-core electric reactor indoor operation fault prediction method, medium and system. BACKGROUND

[0002] The dry-type air-core electric reactor is one of the important power equipment in the power system, and is widely used in the fields of power transmission, motor driving, power electronics and the like. Compared with the liquid-immersed electric reactor, the dry-type electric reactor has the advantages of small volume, light weight, oil-free cooling and the like, and can be directly installed indoors or overhead, and can stably operate in a relatively harsh environment.

[0003] The core components of the dry-type electric reactor include a core, a winding, a cooling system and the like. In a long-term operation process, various fault problems caused by insulation aging, local overheating, bearing wear and the like may occur in these key components, such as phase failure, overload failure, insulation breakdown and the like. Once these faults occur, not only serious equipment damage will be caused, but also a cascading failure of the power system may be caused, and even the safe and stable operation of the power grid is threatened. Therefore, how to accurately and effectively diagnose and predict the operation state of the dry-type electric reactor is crucial for avoiding major accidents and ensuring the safety of the power grid.

[0004] At present, the state monitoring and fault diagnosis technology of the dry-type electric reactor generally adopts a periodic inspection method: personnel are regularly dispatched to perform appearance inspection, insulation test, partial discharge test and the like on the electric reactor, and the fault hidden danger is found through manual observation and simple measurement means. This method depends on the experience of the inspection personnel, and has certain limitations and hysteresis.

[0005] Although the above-mentioned prior art has realized the state monitoring and fault diagnosis of the dry-type electric reactor to a certain extent, the fault judgment still depends on experience, and it is difficult to realize automatic fault prediction: the existing method generally depends on the experience of professional personnel to judge the fault, and it is difficult to predict the future fault development trend. SUMMARY

[0006] Therefore, the application provides a dry-type air-core electric reactor indoor operation fault prediction method, medium and system, which can solve the technical problem that the existing method generally depends on the experience of professional personnel to judge the fault, and it is difficult to predict the future fault development trend.

[0007] The application is implemented in the following manner:

[0008] A first aspect of the application provides a dry-type air-core electric reactor indoor operation fault prediction method, which comprises the following steps:

[0009] S10, real-time acquisition of input electrical parameters, output electrical parameters, infrared thermal imaging, acoustic imaging and vibration signal group under the indoor running state of the to-be-tested dry-type air core reactor, wherein the vibration signal group comprises vibration signals collected by a plurality of vibration sensors at the detection points on the surface of the to-be-tested dry-type reactor;

[0010] S20, extracting features of the input electrical parameters and the output electrical parameters by using a pre-trained first feature extraction model for analyzing electrical parameters, as first features;

[0011] S30, extracting features of the infrared thermal imaging by using a pre-trained second feature extraction model for analyzing heat sources, as second features;

[0012] S40, extracting features of the acoustic imaging by using a pre-trained third feature extraction model for analyzing noise, as third features;

[0013] S50, extracting features of the vibration signal group by using a pre-trained fourth feature extraction model for analyzing vibration, as fourth features;

[0014] S60, inputting the first features, the second features, the third features and the fourth features into a pre-trained fault diagnosis model to obtain a fault vector comprising a fault type and a fault occurrence probability;

[0015] S70, predicting a fault of the to-be-tested dry-type air core reactor according to the fault vector.

[0016] The step S10 comprises: real-time acquisition of input electrical parameters, output electrical parameters, infrared thermal imaging, acoustic imaging and vibration signal group under the indoor running state of the to-be-tested dry-type air core reactor. The input electrical parameters and the output electrical parameters are collected by voltage and current sensors, including voltage, current, phase and other parameters. The infrared thermal imaging is obtained by scanning the surface of the reactor by an infrared thermal imager. The acoustic imaging is obtained by acoustic wave detection on the surface and inside of the reactor by an acoustic wave sensor. The vibration signal group is obtained by a plurality of vibration sensors arranged on the surface of the reactor. The above multi-modal operation data comprehensively reflect the running state of the reactor, providing basic data for subsequent fault prediction.

[0017] The step S20 comprises: using a pre-trained first feature extraction model to perform feature extraction on the input electrical parameters and the output electrical parameters to obtain a feature vector as a first feature.

[0018] The step S30 comprises: using a pre-trained second feature extraction model to perform feature extraction on the infrared thermal imaging data to obtain a feature vector as a second feature. The second feature extraction model comprises: a low-level feature extraction sub-network, a high-level feature extraction sub-network, and a fusion sub-network. The low-level feature extraction sub-network is based on a shallow convolutional network and extracts low-level visual features such as edges and textures of the infrared thermal imaging; the high-level feature extraction sub-network is based on a deep convolutional network and extracts high-level semantic features such as shapes and distributions of temperature distribution regions of the infrared thermal imaging; and the fusion sub-network adaptively fuses the low-level features and the high-level features by using an attention mechanism to form a final second feature. The second feature can comprehensively capture abnormal temperature rise features in the infrared thermal imaging data.

[0019] The step S40 comprises: using a pre-trained third feature extraction model to perform feature extraction on the acoustic imaging data to obtain a feature vector as a third feature. The third feature extraction model comprises: a low-frequency feature extraction sub-network, a high-frequency feature extraction sub-network, a transient feature extraction sub-network, and a fusion sub-network. The low-frequency feature extraction sub-network extracts low-frequency harmonic signals in the acoustic imaging, the high-frequency feature extraction sub-network extracts high-frequency discharge noise signals, the transient feature extraction sub-network extracts transient electrical pulse signals, and the fusion sub-network fuses the three kinds of frequency band features to form a comprehensive acoustic wave feature representation. The third feature can comprehensively reflect multi-frequency features of various electrical abnormalities in the acoustic imaging data.

[0020] The step S50 comprises: using a pre-trained fourth feature extraction model to perform feature extraction on the vibration signal group to obtain a feature vector as a fourth feature. The fourth feature extraction model comprises: a large amplitude vibration feature extraction subnetwork, a small amplitude vibration feature extraction subnetwork, a complex vibration feature extraction subnetwork, and a fusion subnetwork. The large amplitude vibration feature extraction subnetwork extracts large amplitude abnormal vibration signals on the surface of the electric reactor, the small amplitude vibration feature extraction subnetwork extracts tiny progressive vibration signals, the complex vibration feature extraction subnetwork extracts complex electromagnetic-mechanical coupling vibration signals, and the fusion subnetwork fuses the three kinds of vibration features to form a comprehensive vibration feature representation. The fourth feature can comprehensively describe various vibration features reflecting the state of the mechanical components of the electric reactor in the vibration signal group.

[0021] The step S60 comprises: inputting the extracted four kinds of features, i.e., the first feature (electrical parameter), the second feature (infrared imaging), the third feature (acoustic imaging), and the fourth feature (vibration), into a pre-trained fault diagnosis model to obtain a fault vector including a fault type and a fault occurrence probability. The fault diagnosis model comprises: a multi-modal feature encoder, an interactive attention module, a fault type classifier, and a fault probability classifier. The multi-modal feature encoder performs feature encoding on the four kinds of modal data respectively, the interactive attention module fuses different modal features through an attention mechanism, the fault type classifier outputs possible fault types, and the fault probability classifier outputs occurrence probabilities of various faults. Through the output of the fault vector, the fault type and the fault probability of the to-be-tested electric reactor can be comprehensively predicted.

[0022] The fault probability classifier can output an occurrence probability value of each possible fault type. When the occurrence probability of a certain fault type exceeds a preset probability threshold, for example, 10%, it can be determined that the fault type is likely to occur.

[0023] Specifically, the LSTM network in the first feature extraction model can effectively learn the time sequence features of the input electrical parameters and the output electrical parameters, capture the long-term dependency and time sequence patterns therein, the CNN network can extract local features and mutation points, and reflect abnormal patterns in the electrical parameters, and the fusion subnetwork adaptively combines the features extracted by the LSTM and the CNN using an attention mechanism to form a first feature with higher discriminability.

[0024] Specifically, the low-level feature extraction subnetwork in the second feature extraction model is based on a shallow convolutional network and mainly extracts low-level visual features such as edges and textures in the infrared thermal image, the high-level feature extraction subnetwork is based on a deep convolutional network and can extract high-level semantic features such as shapes and distributions of different temperature regions in the infrared thermal image, and the fusion subnetwork adaptively assigns fusion weights to the low-level features and the high-level features using an attention mechanism to form a comprehensive second feature.

[0025] Specifically, the low-frequency feature extraction sub-network in the third feature extraction model mainly extracts low-frequency harmonic signals in acoustic imaging, such as low-frequency sound waves caused by reactor winding vibration; the high-frequency feature extraction sub-network mainly extracts high-frequency discharge noise signals, such as high-frequency pulse sound waves caused by local breakdown; the transient feature extraction sub-network mainly extracts transient electrical pulse signals, such as transient discharge sound waves caused by insulation faults; and the fusion sub-network fuses the above three frequency band features to form a comprehensive acoustic wave feature representation.

[0026] Specifically, the large amplitude vibration feature extraction sub-network in the fourth feature extraction model mainly extracts large amplitude abnormal vibration signals on the surface of the reactor, such as impact vibration caused by serious faults; the small amplitude vibration feature extraction sub-network mainly extracts small progressive vibration signals, such as small vibrations caused by component wear; the complex vibration feature extraction sub-network mainly extracts complex electromagnetic-mechanical coupled vibration signals; and the fusion sub-network fuses the above three vibration features to form a comprehensive vibration feature representation.

[0027] Specifically, the multi-modal feature encoder in the fault diagnosis model includes four sub-encoders corresponding to four input modalities, which encode the features of each modality data; the interaction attention module fuses different modal features through attention mechanism, so that each modality feature can obtain supplementary information from other modality features; the fault type classifier outputs possible fault types, such as open-phase fault, overload fault, insulation breakdown, etc.; and the fault probability classifier outputs the probability value of each possible fault type.

[0028] The first feature extraction model includes a long memory sub-network, a short memory sub-network, and a first fusion sub-network; the long memory sub-network is based on a long short-term memory network structure and is used to learn the long-term dependence and time sequence pattern of input electrical parameters and output electrical parameters; the short memory sub-network is based on a convolutional neural network and an attention mechanism, and is used to capture local features and mutation points of input electrical parameters and output electrical parameters; and the first fusion sub-network fuses the outputs of the long memory sub-network and the short memory sub-network to output the first feature.

[0029] The second feature extraction model includes a low-level feature extraction sub-network, a high-level feature extraction sub-network, and a second fusion sub-network; the low-level feature extraction sub-network is based on a shallow convolutional neural network and is used to extract low-level features of infrared thermal imaging, including edges and textures; the high-level feature extraction sub-network is based on a deep convolutional neural network and is used to extract high-level semantic features of infrared thermal imaging; and the second fusion sub-network fuses the low-level and high-level features, and uses an attention mechanism to guide the fusion weight to output the second feature.

[0030] The third feature extraction model comprises a low-frequency feature extraction subnetwork, a high-frequency feature extraction subnetwork, a transient feature extraction subnetwork, and a third fusion subnetwork; the low-frequency feature extraction subnetwork is used to capture low-frequency harmonic signals in acoustic imaging; the high-frequency feature extraction subnetwork is used to capture high-frequency discharge noise; the transient feature extraction subnetwork is used to capture transient electrical pulse signals; and the third fusion subnetwork fuses the three kinds of features to obtain a comprehensive feature representation of acoustic imaging.

[0031] The fourth feature extraction model comprises a large-amplitude vibration feature extraction subnetwork, a small-amplitude vibration feature extraction subnetwork, a composite vibration feature extraction subnetwork, and a fourth fusion subnetwork; the large-amplitude vibration feature extraction subnetwork captures large-amplitude vibration signals on the surface of the reactor; the small-amplitude vibration feature extraction subnetwork captures micro-vibrations; the composite vibration feature extraction subnetwork captures combinations of multiple vibration modes; and the fourth fusion subnetwork fuses the three kinds of vibration features to obtain a comprehensive feature representation of the vibration signal group.

[0032] The fault diagnosis model is a multi-modal fusion network based on an attention mechanism, comprising a plurality of modal feature encoders, an interactive attention module, a fault type classifier, and a fault probability classifier; the plurality of modal feature encoders correspond to electrical parameters, infrared imaging, acoustic imaging, and vibration signals, respectively, and encode the features of each kind of modal data; the interactive attention module interacts with the encoded different modal features to enhance the information between different modalities; the fault type classifier outputs the specific type of fault; and the fault probability classifier outputs the occurrence probability value of each fault.

[0033] Input difference between the long memory subnetwork and the short memory subnetwork:

[0034] Long memory subnetwork: input time series length: 1 hour-7 days; sampling interval: 1 minute-1 hour;

[0035] Short memory subnetwork: input time series length: 1 second-10 minutes; sampling interval: 1 millisecond-1 second;

[0036] Input difference between the low-frequency feature extraction subnetwork and the high-frequency feature extraction subnetwork:

[0037] Low-frequency feature extraction subnetwork: frequency range: 0Hz-1000Hz; main focus range: 50 / 60Hz (power frequency) and its harmonics;

[0038] High-frequency feature extraction subnetwork: frequency range: 1kHz-1MHz; main focus range: 10kHz-500kHz (partial discharge frequency range);

[0039] The input of the large amplitude vibration feature extraction subnetwork and the small amplitude vibration feature extraction subnetwork is different:

[0040] Large amplitude vibration feature extraction subnetwork: amplitude range: > 100 pm (micrometers); typical range: 0.1 mm-10 mm;

[0041] Small amplitude vibration feature extraction subnetwork: amplitude range: ≤ 100 pm (micrometers); typical range: 0.1 pm-100 pm.

[0042] The fault type that the fault type classifier can output includes but is not limited to: insulation aging; partial discharge; overheating; mechanical looseness; short circuit; open circuit; core fault; coil deformation; ground fault; cooling system fault; fault caused by environmental pollution; manufacturing defect; overload; harmonic interference; lightning damage.

[0043] Preferably, the input of the long memory subnetwork and the short memory subnetwork comes from the "input electrical parameters, output electrical parameters" mentioned in step S10. Specifically, it includes:

[0044] Input parameters: input voltage (V); input current (A); input frequency (Hz); input power factor;

[0045] Output parameters: output voltage (V); output current (A); output frequency (Hz); output power factor;

[0046] These parameters are usually measured by voltage transformers (PT) and current transformers (CT) installed on the incoming and outgoing lines of the dry-type air-core reactor, and then collected in real time by a data acquisition system.

[0047] The input of the low-frequency feature extraction subnetwork and the high-frequency feature extraction subnetwork: from the "acoustic imaging" data mentioned in step S10. Specifically, it includes: sound pressure level (dB); sound intensity (W / m 2 ); acoustic spectrum data; these data are usually collected by a special acoustic imaging device. This device may include an acoustic camera or a microphone array, which can capture and visualize the sound field distribution around the device.

[0048] The input of the large amplitude vibration feature extraction subnetwork and the small amplitude vibration feature extraction subnetwork: from the "vibration signal group" mentioned in step S10. Specifically, it includes: vibration displacement (m); vibration velocity (m / s); vibration acceleration (m / s 2 ); vibration spectrum data; these data are collected by vibration sensors installed at multiple detection points on the surface of the dry-type air-core reactor. Acceleration sensors are usually used, and velocity and displacement information can be obtained by integration.

[0049] The input of the second feature extraction model (for infrared thermal imaging) is thermal image data (usually a two-dimensional array representing the temperature value of each pixel point), including the highest temperature (℃), the average temperature (℃), and the temperature distribution information; these data are collected by an infrared thermal imager and can capture the temperature distribution of the device surface.

[0050] All these parameters are collected in real time during the normal operation of the dry-type air core reactor. The data acquisition system integrates these information and inputs them into the corresponding feature extraction model for processing.

[0051] The training steps of the fault diagnosis model are described as follows:

[0052] 1) Prepare the training data set: including collecting a large amount of dry-type reactor operation data of known fault types, including input electrical parameters, output electrical parameters, infrared thermal imaging, acoustic imaging, and vibration signals, etc.; label the collected data, give the corresponding fault type and the probability of fault occurrence for each sample; divide the data set into training set, validation set and test set.

[0053] 2) Data preprocessing: including standardization processing of input electrical parameters, output electrical parameters and other time series data; size adjustment, data enhancement and other preprocessing of infrared thermal imaging, acoustic imaging and other image / video data; filtering, segmentation and other processing of vibration signals and other sensor data.

[0054] 3) Model training: including initializing the parameters of the first to fourth feature extraction models, training the four feature extraction models on the training set respectively; initializing the parameters of the fault diagnosis model, fixing the four trained feature extraction models, training the fault diagnosis model on the training set, evaluating the model performance on the validation set in a certain epoch unit, and selecting the model parameters with the best performance on the validation set.

[0055] 4) Model evaluation: including testing the performance indicators of the fault diagnosis model on the test set, such as fault type classification accuracy, log loss of fault probability prediction, etc.; analyzing the error samples and exploring the advantages and disadvantages of the model.

[0056] Further, the plurality of modal feature encoders correspond to electrical parameters, infrared imaging, acoustic imaging, and vibration signals, respectively, and the step of encoding the features of each modal data includes an electrical parameter modal feature encoder corresponding to the first feature, an infrared imaging modal feature encoder corresponding to the second feature, an acoustic imaging modal feature encoder corresponding to the third feature, and a vibration signal modal feature encoder corresponding to the fourth feature.

[0057] Further, when the occurrence probability of a certain fault type exceeds the preset probability threshold, it is determined that the fault of the fault type occurs.

[0058] Optionally, the failure probability classifier outputs a probability value, which can set a lower probability threshold, for example 10%, when the probability of a certain failure type exceeds the threshold, it is determined that the failure type may occur, providing the basis for subsequent failure prevention.

[0059] Optionally, the LSTM network in the first feature extraction model can capture long-term dependencies and time series patterns in input and output electrical parameters, and the CNN network can extract local features and mutation points, which can comprehensively reflect the abnormal patterns of electrical parameters.

[0060] Optionally, the low-level feature extraction sub-network and the high-level feature extraction sub-network in the second feature extraction model extract the bottom-level visual features and the high-level semantic features in the infrared thermal imaging respectively, and through the attention mechanism adaptive fusion, the multi-scale information reflecting the abnormal temperature rise of the reactor in the infrared thermal imaging data can be fully utilized.

[0061] Optionally, the low-frequency feature extraction sub-network, the high-frequency feature extraction sub-network and the transient feature extraction sub-network in the third feature extraction model extract different frequency band features in acoustic imaging, such as low-frequency vibration, high-frequency discharge and transient pulse, and through fusion, comprehensive acoustic features are formed, which can comprehensively capture the electrical abnormal information contained in the acoustic imaging data.

[0062] Optionally, the large amplitude vibration feature extraction sub-network, the small amplitude vibration feature extraction sub-network and the complex vibration feature extraction sub-network in the fourth feature extraction model extract the features of large amplitude abnormal vibration, small progressive vibration and complex coupled vibration on the surface of the reactor respectively, and through fusion, comprehensive vibration features are formed, which can fully utilize the various vibration information reflecting the state of mechanical components in the vibration signal group.

[0063] Optionally, the interactive attention module in the fault diagnosis model can effectively improve the mutual enhancement between different modal features through the attention mechanism, so that each modal feature can obtain valuable supplementary information from other modal features, thereby improving the performance of fault diagnosis.

[0064] Optionally, in the training data set of the fault diagnosis model, each sample is attached with corresponding fault type label and fault occurrence probability label, which provides reliable label information for the supervised learning of the model.

[0065] Optionally, during the training process of the fault diagnosis model, the performance on the validation set needs to be monitored in real time, and the model parameters with the best performance on the validation set are selected, which can effectively avoid overfitting and improve the generalization ability of the model on new data.

[0066] Optionally, the evaluation indexes of the fault diagnosis model on the test set include fault type classification accuracy and log loss of fault probability prediction. By analyzing these indexes, the advantages and disadvantages of the model can be comprehensively understood, and the basis for further optimization is provided.

[0067] The second aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and the program instructions are used to execute the dry-type air-core reactor indoor operation fault prediction method.

[0068] The third aspect of the present application provides a dry-type air-core reactor indoor operation fault prediction system, which comprises the computer-readable storage medium.

[0069] Compared with the prior art, the dry-type air-core reactor indoor operation fault prediction method, medium and system provided by the present application have the following advantages:

[0070] 1. Multi-modal fusion diagnosis, which comprehensively improves the fault diagnosis accuracy. The present application method collects various data sources such as electrical parameters, infrared thermal imaging, sound waves, vibration and the like during the operation of the reactor, and through feature extraction and fusion of these complementary data, the operation state of the reactor can be more comprehensively and accurately reflected, thereby greatly improving the accuracy of fault diagnosis.

[0071] 2. Automatic fault prediction, which realizes the foresight of state monitoring. The fault diagnosis model based on deep learning in the present application method can not only accurately identify the possible fault types at present, but also predict the possible faults and their occurrence probabilities of the reactor within a certain time in the future, thereby providing support for preventive maintenance and avoiding the occurrence of major faults.

[0072] 3. Continuous online monitoring, which realizes the whole life cycle management of the reactor. The present application method integrates various monitoring means together, which can realize all-weather and real-time online monitoring of the reactor, timely find potential fault hidden dangers, and provide guarantee for the whole life cycle management of the reactor, thereby greatly reducing the maintenance cost and downtime loss.

[0073] In summary, the scheme of the present application solves the technical problem that the existing method usually relies on the experience of professionals to judge the fault and is difficult to predict the future fault development trend. BRIEF DESCRIPTION OF DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the description of the embodiments of the present application will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0075] Figure 1 A flow chart of the method provided by the present application. DETAILED DESCRIPTION

[0076] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0077] As shown in Figure 1 FIG. 1 is a flow chart of a dry-type air-core reactor indoor operation fault prediction method provided by the present application. The method comprises the following steps:

[0078] S10, real-time acquisition of input electrical parameters, output electrical parameters, infrared thermal imaging, acoustic imaging and vibration signal groups in the indoor operation state of the dry-type air-core reactor to be tested, wherein the vibration signal groups comprise vibration signals collected by a plurality of vibration sensors of a detection point on the surface of the dry-type reactor to be tested;

[0079] S20, extraction of features of the input electrical parameters and the output electrical parameters by using a pre-trained first feature extraction model as first features;

[0080] S30, extraction of features of the infrared thermal imaging by using a pre-trained second feature extraction model as second features;

[0081] S40, extraction of features of the acoustic imaging by using a pre-trained third feature extraction model as third features;

[0082] S50, extraction of features of the vibration signal groups by using a pre-trained fourth feature extraction model as fourth features;

[0083] S60, input of the first features, the second features, the third features and the fourth features into a pre-trained fault diagnosis model to obtain a fault vector, including a fault type and a fault occurrence probability;

[0084] S70, prediction of a fault of the dry-type air-core reactor to be tested according to the fault vector.

[0085] The specific embodiments of the above steps will be described in detail as follows:

[0086] The specific embodiment of step S10 is:

[0087] Real-time acquisition of various operation data in the indoor operation state of the dry-type air-core reactor to be tested, including input electrical parameters, output electrical parameters, infrared thermal imaging, acoustic imaging and vibration signal groups. Among them:

[0088] 1. The input electrical parameters and the output electrical parameters refer to the voltage, current, phase, and other parameters at both ends of the reactor, which reflect the working state of the reactor as they change over time. These electrical parameters can be collected in real time using voltage and current sensors.

[0089] 2. Infrared thermography refers to using an infrared thermal imager to perform thermal imaging scanning on the surface of the reactor. By analyzing the temperature distribution of different areas on the surface of the reactor, local abnormal temperature rise in the reactor windings, core, and other parts can be found, thereby predicting potential fault risks.

[0090] 3. Acoustic imaging refers to using acoustic wave sensors to detect acoustic waves on the surface and inside of the reactor. By analyzing the frequency spectrum characteristics of the acoustic waves, abnormal sound signals such as discharge and local breakdown in the reactor windings, core, and other parts can be detected.

[0091] 4. The vibration signal group refers to arranging multiple vibration sensors on the surface of the reactor to collect vibration signals on the surface of the reactor in real time. These vibration signals reflect the working state of the mechanical components of the reactor, such as abnormal vibrations caused by bearing wear and mechanical resonance.

[0092] In summary, the purpose of step S10 is to comprehensively collect multi-modal data describing the operating state of the reactor, providing a reliable data foundation for subsequent fault prediction.

[0093] The specific implementation of step S20 is:

[0094] A pre-trained first feature extraction model is used to extract features from the input electrical parameters and the output electrical parameters collected in step S10, obtaining a feature vector as the first feature.

[0095] The structure of the first feature extraction model is as follows:

[0096] 1. Long-Short Term Memory (LSTM) network: The LSTM network can effectively learn the long-term dependencies in time series data, making it suitable for extracting time series features of input electrical parameters and output electrical parameters.

[0097] 2. Convolutional Neural Network (CNN) network: The CNN network is good at extracting local features and can capture sudden changes or abnormal patterns in input electrical parameters and output electrical parameters.

[0098] 3. First fusion sub-network: The features extracted by the LSTM network and the CNN network are fused, and an attention mechanism is used to adaptively learn the fusion weights of the two features, outputting the first feature.

[0099] The purpose of the first feature extraction model is to fully mine the abnormal patterns and time series characteristics contained in the input electrical parameters and output electrical parameters, and to provide discriminative feature representation for subsequent fault diagnosis.

[0100] The specific implementation of step S30 is:

[0101] The second feature extraction model is pre-trained, and the infrared thermal imaging data collected in step S10 is subjected to feature extraction to obtain a feature vector as the second feature.

[0102] The structure of the second feature extraction model is as follows:

[0103] 1. Low-level feature extraction subnetwork: This subnetwork uses a shallow convolutional neural network to mainly extract low-level visual features such as edges and textures in infrared thermal imaging.

[0104] 2. High-level feature extraction subnetwork: This subnetwork uses a deeper convolutional neural network to extract high-level semantic features in infrared thermal imaging, such as the shape and distribution of different temperature regions.

[0105] 3. Second fusion subnetwork: The low-level features and high-level features are fused, and an attention mechanism is used to adaptively assign fusion weights to the two types of features, and the output is the second feature.

[0106] The purpose of the second feature extraction model is to comprehensively extract multi-scale visual features reflecting the abnormal temperature rise of the reactor in the infrared thermal imaging data, and to provide more abundant information for subsequent fault diagnosis.

[0107] The specific implementation of step S40 is:

[0108] The third feature extraction model is pre-trained, and the acoustic imaging data collected in step S10 is subjected to feature extraction to obtain a feature vector as the third feature.

[0109] The structure of the third feature extraction model is as follows:

[0110] 1. Low-frequency feature extraction subnetwork: This subnetwork mainly extracts low-frequency harmonic signals in acoustic imaging, such as low-frequency sound waves caused by reactor winding vibration.

[0111] 2. High-frequency feature extraction subnetwork: This subnetwork mainly extracts high-frequency discharge noise signals in acoustic imaging, such as high-frequency pulse sound waves caused by local breakdown of the reactor.

[0112] 3. Transient feature extraction subnetwork: This subnetwork mainly extracts transient electrical pulse signals in acoustic imaging, such as transient discharge sound waves caused by reactor insulation failure.

[0113] 4. Third fusion sub-network: fuse the above three different frequency band acoustic features to form a comprehensive acoustic feature representation as the third feature.

[0114] The purpose of the third feature extraction model is to comprehensively capture various acoustic features reflecting various electrical abnormalities of the reactor in the acoustic imaging data, and to provide valuable information for subsequent fault diagnosis.

[0115] The specific implementation of step S50 is:

[0116] The fourth feature extraction model is pre-trained, and the feature extraction is performed on the vibration signal group collected in step S10 to obtain the feature vector as the fourth feature.

[0117] The structure of the fourth feature extraction model is as follows:

[0118] 1. Large amplitude vibration feature extraction sub-network: This sub-network mainly extracts large amplitude abnormal vibration signals on the surface of the reactor, such as impact vibration caused by serious faults of the reactor winding.

[0119] 2. Small amplitude vibration feature extraction sub-network: This sub-network mainly extracts small vibration signals on the surface of the reactor, such as small vibrations caused by progressive wear of the reactor winding, bearing and other components.

[0120] 3. Compound vibration feature extraction sub-network: This sub-network mainly extracts compound vibration signals on the surface of the reactor, such as complex vibration patterns generated by the coupling of electromagnetic vibration and mechanical vibration.

[0121] 4. Fourth fusion sub-network: fuse the above three vibration features to form a comprehensive vibration feature representation as the fourth feature.

[0122] The purpose of the fourth feature extraction model is to comprehensively capture various vibration features reflecting the state of mechanical components of the reactor in the vibration signal group, and to provide valuable information for subsequent fault diagnosis.

[0123] The specific implementation of step S60 is:

[0124] The four features extracted in steps S20, S30, and S40, i.e., the electrical parameter feature, the infrared imaging feature, the acoustic imaging feature, and the vibration feature, are input into the pre-trained fault diagnosis model to obtain a fault vector, including the fault type and the fault occurrence probability.

[0125] The structure of the fault diagnosis model is as follows:

[0126] 1. Multi-modal feature encoder: This module contains four sub-encoders corresponding to four input modalities, which encode the features of each modality data. The specific description is as follows:

[0127] (1) Electrical parameter modal feature encoder

[0128] Input: First features output by the first feature extraction model

[0129] Encoding steps:

[0130] 1) Data normalization: Normalize input features to the range [-1, 1].

[0131] 2) Temporal modeling: Use a bidirectional LSTM network to handle temporal information.

[0132] 3) Feature compression: Reduce feature dimensions through fully connected layers.

[0133] 4) Nonlinear transformation: Apply a ReLU activation function.

[0134] Output dimension of specific encoding: Typically a 128 or 256-dimensional vector, encoding representation: [e1, e2,..., en], where each ei is a floating-point number in the range [-1, 1];

[0135] (2) Infrared imaging modal feature encoder

[0136] Input: Second features output by the second feature extraction model

[0137] Encoding steps:

[0138] 1) Spatial feature extraction: Use a pre-trained CNN such as ResNet or EfficientNet to extract spatial features.

[0139] 2) Global pooling: Use global average pooling (GAP) to compress spatial dimensions.

[0140] 3) Feature fusion: Fuse low-level and high-level features through a concat operation.

[0141] 4) Dimension adjustment: Adjust feature dimensions through fully connected layers.

[0142] Output dimension of specific encoding: Typically a 256 or 512-dimensional vector, encoding representation: [r1, r2,..., rn], where each ri is a non-negative real number.

[0143] (3) Acoustic imaging modal feature encoder

[0144] Input: Third features output by the third feature extraction model

[0145] Encoding steps:

[0146] 1) Frequency domain transformation: Perform a Fast Fourier Transform (FFT) on the input features.

[0147] 2) Spectral feature extraction: Use 1D-CNN to extract spectral features.

[0148] 3) Time-frequency feature fusion: Combine original time-domain features and spectral features.

[0149] 4) Attention mechanism: Apply self-attention mechanism to highlight important features.

[0150] Output dimension of specific encoding: usually a 192 or 384-dimensional vector; Encoding representation: [s1, s2,..., sn], where each si can be any real number;

[0151] (4) Vibration signal modal feature encoder

[0152] Input: fourth feature extracted by the fourth feature extraction model;

[0153] Encoding steps:

[0154] 1) Multi-scale decomposition: Use wavelet transform to perform multi-scale decomposition on the vibration signal.

[0155] 2) Sub-band feature extraction: Use 1D-CNN to extract features for each sub-band.

[0156] 3) Sub-band feature fusion: Use attention mechanism to fuse features of different sub-bands.

[0157] 4) Time-frequency domain feature combination: Combine time-domain and frequency-domain features through concatenation operation.

[0158] Specific encoding: Output dimension: usually a 224 or 448-dimensional vector; Encoding representation: [v1, v2,..., vn], where each vi is a non-negative real number;

[0159] General steps for all modal encoders:

[0160] 5) Normalization: Perform L2 normalization on the encoded features to ensure consistent feature scales across different modalities.

[0161] 6) Embedding space mapping: Map features of different modalities to a common embedding space through an additional fully connected layer.

[0162] Final encoding output: the output dimensions of all modal encoders are unified, e.g., all 256 dimensions; the encoding vector of each modality is represented as: [f1, f2,..., f256], where each fi is a floating-point number in the range of [-1, 1]; these encoding vectors will be used as inputs to the interaction attention module for multi-modal information fusion and interaction. The purpose of encoding is to convert the original features of different modalities into a unified and information-rich representation, facilitating subsequent fault diagnosis models for processing and decision-making. It should be noted that the specific encoding dimensions and methods may be adjusted according to the actual data characteristics and model performance. This provides a general framework, and detailed tuning may be required in actual applications.

[0163] 2. Interaction attention module: This module uses attention mechanisms to interact and fuse different modal features, allowing each modality feature to obtain valuable complementary information from other modality features.

[0164] 3. Fault type classifier: This module outputs the possible fault types of the reactor, such as open-phase fault, overload fault, insulation breakdown, etc.

[0165] 4. Fault probability classifier: This module outputs the occurrence probability of each fault type, providing quantitative basis for subsequent fault prediction.

[0166] The training steps of the fault diagnosis model include: preparing a large training dataset with known fault types, preprocessing the training data, step-by-step training of the feature extraction model and fault diagnosis model, etc. During the training process, the performance of the model on the validation set needs to be monitored, and the model parameters with the best validation set performance are selected. Finally, the fault type classification accuracy and log loss of fault probability prediction of the model are evaluated on the test set.

[0167] The specific implementation of step S70 is:

[0168] According to the fault vector obtained in step S60, the fault of the dry-type air-core reactor under test is predicted.

[0169] Specifically:

[0170] 1. Fault type prediction: According to the output of the fault type classifier, determine the possible fault types of the reactor, such as open-phase fault, overload fault, insulation breakdown, etc.

[0171] 2. Fault probability prediction: According to the output of the fault probability classifier, obtain the occurrence probability value of each possible fault type. A lower probability threshold, e.g., 10%, can be set, and when the occurrence probability of a certain fault type exceeds the threshold, it is determined that the fault type may occur.

[0172] Through the above two steps, the possible fault types and occurrence probabilities of the to-be-tested electric reactor can be comprehensively predicted, thereby providing an important basis for state monitoring and preventive maintenance of the electric reactor.

[0173] In summary, the dry-type air-core reactor indoor operation fault prediction method provided by the present application fully utilizes the complementary advantages of multi-modal data (electrical parameters, infrared thermal image, sound wave, vibration), adopts deep learning technology to extract efficient feature representation, and realizes comprehensive prediction of the fault type and fault probability of the electric reactor based on a multi-modal feature fusion fault diagnosis model.

[0174] The second aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and the program instructions are used to execute the dry-type air-core reactor indoor operation fault prediction method.

[0175] The third aspect of the present application provides a dry-type air-core reactor indoor operation fault prediction system, which comprises the computer-readable storage medium.

[0176] In order to better understand and implement the present application, the following provides a specific embodiment 1 of the present application in a computer-readable storage medium or a computer system for a computer program: the specific implementation of step S10 is as follows:

[0177] In this step, first, the multi-modal data of the to-be-tested dry-type air-core reactor in the indoor operation state need to be collected in real time, including input electrical parameters, output electrical parameters, infrared thermal imaging, sound imaging and vibration signal groups.

[0178] For the collection of input electrical parameters and output electrical parameters, voltage and current sensors can be used to measure the voltage U and current I of the electric reactor, as well as the phase angle φ and other parameters in real time. These electrical parameters can be represented as:

[0179] E in =[U in ,I in ,φ in ]

[0180] E out =[U out ,I out ,φ out ]

[0181] Wherein, E in and E out represent the input electrical parameter vector and the output electrical parameter vector respectively. These electrical parameters change with time and can reflect the operation state of the electric reactor.

[0182] For the collection of infrared thermal imaging, an infrared thermal imager can be used to scan the surface of the electric reactor to obtain a thermal imaging image I IR of the temperature distribution on the surface of the electric reactor. These thermal imaging data can reflect temperature rise abnormalities of the electric reactor winding, core and other parts.

[0183] For the collection of acoustic imaging, an acoustic wave sensor can be used to detect acoustic waves on the surface and inside of the electric reactor to obtain acoustic wave signals S acoustic emitted by the electric reactor. These acoustic wave signals contain abnormal acoustic signals such as discharge and local breakdown that may occur in the electric reactor winding, core and other parts.

[0184] For the collection of vibration signal groups, multiple vibration sensors need to be arranged on the surface of the electric reactor to collect vibration signals V = {v1, v2,..., v n in real time, where v i represents the vibration time series data collected by the i-th vibration sensor. These vibration signals can reflect the working state of the mechanical parts of the electric reactor, such as abnormal vibration caused by bearing wear and mechanical resonance.

[0185] In summary, the purpose of step S10 is to comprehensively collect multi-modal data describing the operating state of the electric reactor, i.e., {E in , E out , I IR , S acoustic , V}, to provide a reliable data basis for subsequent fault prediction.

[0186] The specific implementation of step S20 is as follows:

[0187] In this step, a pre-trained first feature extraction model needs to be used to extract features from the input electrical parameters E in and output electrical parameters E out collected in step S10 to obtain a feature vector as the first feature f1.

[0188] The structure of the first feature extraction model is as follows:

[0189] 1. Long Short-Term Memory (LSTM) network: The LSTM network can effectively learn the long-term dependencies in time series data and can be established as:

[0190] h t ,c t = LSTM(e t ,h t-1 ,c t-1 )

[0191] where e t is the electrical parameter input at time t, h t and ct are the hidden state and cell state at time t, respectively. LSTM is able to capture the long-term temporal dependencies in E in and E out .

[0192] 2. Convolutional Neural Network (CNN): CNN network is good at extracting local features, which can be established as:

[0193] f CNN = CNN (E in , E out )

[0194] where f CNN represents the local features extracted by CNN. CNN is able to capture the mutation points or abnormal patterns in E in and E out .

[0195] 3. Fusion sub-network: the features h t and f CNN extracted by LSTM network and CNN network are fused, which can adopt attention mechanism to adaptively learn the fusion weight a:

[0196] f1 = å t a t h t + f CNN

[0197] where f1 is the final first feature vector.

[0198] The purpose of the first feature extraction model is to fully excavate the abnormal patterns and time series features contained in the input electrical parameters and output electrical parameters, and to provide discriminative feature representation for subsequent fault diagnosis.

[0199] The specific implementation of step S30 is as follows:

[0200] In this step, a pre-trained second feature extraction model needs to be used to extract features from the infrared thermal imaging data I IR collected in step S10, so as to obtain the feature vector as the second feature f2.

[0201] The structure of the second feature extraction model is as follows:

[0202] 1. Low-level feature extraction sub-network: this sub-network adopts a shallow convolutional neural network, which can be established as:

[0203] f low = ConvNet low (I IR )

[0204] where flow represent low-level visual features such as edges, textures, etc.

[0205] 2. High-level feature extraction subnetwork: This subnetwork adopts a deeper convolutional neural network, which can be established as:

[0206] f high = ConvNet high (I IR )

[0207] where f high represents high-level semantic features such as the shape and distribution of temperature distribution areas, etc.

[0208] 3. Fusion subnetwork: The low-level features f low and high-level features f high are fused, and an attention mechanism can be used to adaptively assign fusion weights β to the two types of features:

[0209] f2 = βf low + (1-β)f high

[0210] where f2 is the final second feature vector.

[0211] The purpose of the second feature extraction model is to comprehensively extract multi-scale visual features reflecting the abnormal temperature rise of the reactor in the infrared thermal imaging data, providing more abundant information for subsequent fault diagnosis.

[0212] The specific implementation of step S40 is as follows:

[0213] In this step, a pre-trained third feature extraction model is needed to extract features from the acoustic imaging data S acoustic collected in step S10, obtaining a feature vector as the third feature f3.

[0214] The structure of the third feature extraction model is as follows:

[0215] 1. Low-frequency feature extraction subnetwork: This subnetwork adopts the following form:

[0216] f low = SpecNet low (S acoustic )

[0217] where f low represents low-frequency harmonic features in acoustic imaging, such as low-frequency sound waves caused by reactor winding vibration.

[0218] 2. High-frequency feature extraction subnetwork: This subnetwork adopts the following form:

[0219] f high= SpecNet high (S acoustic )

[0220] wherein, f high represents the high-frequency discharge noise feature in acoustic imaging, such as the high-frequency pulse sound wave caused by local breakdown.

[0221] 3. Transient feature extraction subnetwork: the subnetwork can be established as follows:

[0222] f trans = TransNet(S acoustic )

[0223] wherein, f trans represents the transient electrical pulse feature in acoustic imaging, such as the transient discharge sound wave caused by insulation fault.

[0224] 4. Fusion subnetwork: the above three different frequency band acoustic wave features f low , f high and f trans are fused, which can be in the following form:

[0225] f3 = FuseNet(f low , f high , f trans )

[0226] wherein, f3 is the final third feature vector.

[0227] The purpose of the third feature extraction model is to comprehensively capture various acoustic wave features reflecting various electrical abnormalities of the reactor in acoustic imaging data, and to provide valuable information for subsequent fault diagnosis.

[0228] The specific implementation of step S50 is as follows:

[0229] In this step, the pre-trained fourth feature extraction model needs to be used to extract features from the vibration signal group V collected in step S10, to obtain the feature vector as the fourth feature f4.

[0230] The structure of the fourth feature extraction model is as follows:

[0231] 1. Large amplitude vibration feature extraction subnetwork: the subnetwork can be established as follows:

[0232] f large = VibNet large (V)

[0233] wherein, f large represents the large amplitude abnormal vibration feature on the surface of the reactor, such as the impact vibration caused by serious fault.

[0234] 2. Small vibration feature extraction subnetwork: this subnetwork can be established as:

[0235] f small = VibNet small (V)

[0236] where f small represents the micro-gradual vibration characteristics, such as the micro-vibration caused by component wear.

[0237] 3. Compound vibration feature extraction subnetwork: this subnetwork can be established as:

[0238] f complex = VibNet complex (V)

[0239] where f complex represents the complex electromagnetic-mechanical coupling vibration characteristics.

[0240] 4. Fusion subnetwork: fuse the above three vibration characteristics f large , f small and f complex , which can be in the following form:

[0241] f4 = FuseNet(f large , f small , f complex )

[0242] where f4 is the final fourth feature vector.

[0243] The purpose of the fourth feature extraction model is to comprehensively capture various vibration characteristics in the vibration signal group that reflect the mechanical component state of the reactor, and provide valuable information for subsequent fault diagnosis.

[0244] The specific implementation of step S60 is as follows:

[0245] In this step, the four features f1, f2, f3 and f4 extracted in steps S20, S30 and S40 need to be input into the pre-trained fault diagnosis model to obtain the fault vector, including the fault type y c and the fault occurrence probability y p .

[0246] The structure of the fault diagnosis model is as follows:

[0247] 1. Multi-modal feature encoder: this module contains four sub-encoders corresponding to four input modalities, which encode the features of each modality data, and can be established as:

[0248] h1 = Encoder1(f1)

[0249] h2 = Encoder2(f2)

[0250] h3 = Encoder3(f3)

[0251] h4 = Encoder4(f4)

[0252] where h = [h1, h2, h3, h4] i represents the feature encoding result of the i-th modality.

[0253] 2. Interactive attention module: This module uses attention mechanism to interact and fuse different modalities of features, so that each modality of features can obtain valuable complementary information from other modalities of features. It can be established as:

[0254] h = AttentionFusion(h1, h2, h3, h4)

[0255] where h is the fused multi-modal feature vector.

[0256] 3. Fault type classifier: This module outputs the possible fault types y of the reactor c , which can be established as:

[0257] y c = Softmax(W c h + b c )

[0258] where W c and b c are the parameters of the classifier, and Softmax(·) is the Softmax activation function, which outputs the probability distribution of each fault type.

[0259] 4. Fault probability classifier: This module outputs the occurrence probability y of each fault type p , which can be established as:

[0260] y p = Sigmoid(W p h + b p )

[0261] where W p and b p are the parameters of the regressor, and Sigmoid(·) is the Sigmoid activation function, which outputs the occurrence probability value of each fault type.

[0262] The training steps of the fault diagnosis model are as follows:

[0263] 1) Prepare the training data set:

[0264] Collect a large amount of reactor operating data with known fault types, including {E in , Eout I IR ,S acoustic ,V}, and label the data, giving each sample a fault type label y c and a fault occurrence probability label y p . The dataset is divided into training, validation, and test sets.

[0265] 2) Data preprocessing:

[0266] Z-score standardization is performed on the input electrical parameters E in and output electrical parameters E out , etc. Image / video data such as infrared thermal imaging I IR and acoustic imaging S acoustic are resized and data augmented. Band-pass filtering and segmentation are performed on the vibration signal V.

[0267] 3) Model training:

[0268] The parameters of the first to fourth feature extraction models are initialized, and the four feature extraction models are trained on the training set respectively. The parameters of the fault diagnosis model are initialized, and the four trained feature extraction models are fixed. The fault diagnosis model is trained on the training set. In a certain epoch, the model performance is evaluated on the validation set, and the model parameters with the best performance on the validation set are selected.

[0269] 4) Model evaluation:

[0270] The performance indicators of the trained fault diagnosis model are tested on the test set, such as the accuracy of fault type classification Acc(y c ) and the log loss of fault probability prediction The advantages and disadvantages of the model are analyzed.

[0271] The specific implementation of step S70 is as follows:

[0272] In this step, according to the fault vector obtained in step S60, i.e. fault type y c and fault occurrence probability y p , the fault of the dry-type air-core reactor under test is predicted. Specifically:

[0273] 1. Fault type prediction:

[0274] According to the output y c of the fault type classifier, the possible fault types of the reactor can be determined, such as open-phase fault, overload fault, insulation breakdown, etc. The fault type with the highest probability can be selected as the prediction result.

[0275] 2. Fault probability prediction:

[0276] According to the output y of the fault probability classifier p , the occurrence probability value of each possible fault type can be obtained. A lower probability threshold ε, for example 10%, can be set, and when the occurrence probability y p (i) of a certain fault type is greater than ε, it is determined that the fault type is likely to occur.

[0277] In summary, through the above two steps, the possible fault types and their occurrence probabilities of the to-be-tested reactor can be comprehensively predicted, providing an important basis for the state monitoring and preventive maintenance of the reactor.

[0278] Specifically, the principle of the present application is: making full use of the multi-source heterogeneous data generated during the operation of the reactor, including electrical parameters, infrared thermal imaging, acoustic and vibration signals, etc., extracting the rich fault features contained in these data through deep learning technology, and based on a multi-modal feature fusion fault diagnosis model, comprehensively predicting the fault types and fault probabilities of the reactor. The reason why this method can effectively solve the problems existing in the prior art is mainly due to the following key links:

[0279] 1. Multi-modal data acquisition: the present application method arranges multiple sensors on the surface and inside of the reactor, such as voltage and current sensors, infrared thermal imagers, acoustic sensors, vibration sensors, etc., which can collect multiple physical quantity data describing the operating state of the reactor in real time. These different modal data complement each other and can comprehensively reflect the operating conditions of the reactor in multiple aspects such as electricity, heat, and mechanics, providing a reliable data basis for subsequent fault diagnosis.

[0280] 2. Multi-dimensional feature extraction: the present application method uses pre-trained deep learning feature extraction models for different modal data, such as LSTM network to extract time series features of electrical parameters, convolution network to extract visual features of infrared thermal imaging, spectral analysis network to extract frequency spectrum features of acoustic signals, and vibration analysis network to extract dynamic features of vibration signals, etc., fully mining the fault-related information contained in each modal data. These discriminative features provide valuable inputs for subsequent fault diagnosis.

[0281] 3. Multi-modal feature fusion: the present application method inputs the above extracted multi-dimensional features into a multi-modal fusion network based on attention mechanism, which can adaptively learn the mutual relationship and fusion weight between different modal features, so that each modal feature can obtain valuable complementary information from other modal features, further enhancing the performance of fault diagnosis.

[0282] 4. Fault prediction modeling: The method of the present application builds a fault diagnosis model containing a fault type classifier and a fault probability regressor based on multi-modal feature fusion. This model can output the possible fault types of the reactor and their occurrence probabilities, providing quantitative basis for subsequent fault prevention. During model training, by monitoring the performance on the validation set, the optimal model parameters are selected, which can effectively avoid overfitting and improve the generalization ability of the model on new data.

[0283] Through the above several key steps, the method of the present application can fully play the complementary advantages of multi-modal data, extract fault features with high discrimination, and based on the powerful expression ability of multi-modal feature fusion, realize the comprehensive and accurate prediction of the reactor fault, thereby providing strong support for the safe and stable operation of the power system.

[0284] The following provides an embodiment 2 of a specific application scenario of the present application: A certain power company operates and manages a large substation, which includes multiple dry-type air-core reactors. These reactors are used for reactive power compensation and harmonic filtering in the power system. In order to ensure the safe and stable operation of the substation, the company decides to adopt the dry-type air-core reactor operation fault prediction method based on deep learning proposed by the present application to comprehensively monitor and predict the operation state of the reactor.

[0285] The specific implementation process is as follows:

[0286] 1. Multi-modal data acquisition

[0287] Firstly, maintenance personnel arranged various sensors on the dry-type reactor in the substation, including voltage and current sensors, infrared thermal imagers, acoustic sensors, and vibration sensors. These sensors can collect various physical quantity data describing the operation state of the reactor in real time.

[0288] For input and output electrical parameters, voltage and current sensors collect voltage, current and phase angle data every 1 second, forming time series data. Taking a 200kVA dry-type reactor as an example, the typical input and output electrical parameters of the reactor in normal operation state are shown in Table 1.

[0289] Table 1 Input and output electrical parameters of 200kVA dry-type reactor

[0290] Time Input voltage (kV) Input current (A) Input phase angle (°) Output voltage (kV) Output current (A) Output phase angle (°) 0:00 10.5 15.2 26.4 10.0 14.8 35.2 0:01 10.6 15.4 26.7 10.1 15.0 35.5 0:02 10.5 15.3 26.5 10.0 14.9 35.3 ... ... ... ... ... ... ... 23:59 10.7 15.6 27.1 10.2 15.2 35.9

[0291] The infrared thermal imager performs thermal imaging scanning on the surface of the reactor every 5 minutes, obtaining the temperature distribution image of the reactor surface.

[0292] The acoustic sensor continuously collects acoustic signals on the surface and inside of the reactor, with a frequency range of 20Hz-20kHz.

[0293] Finally, vibration sensors are arranged at multiple locations on the surface of the reactor, and real-time vibration time series data of the reactor surface is collected.

[0294] Through the above sensor network, the substation collects multi-modal operation data including electrical parameters, infrared thermal imaging, sound waves, and vibrations, providing a reliable data foundation for subsequent fault prediction.

[0295] 2. Multi-dimensional feature extraction

[0296] After collecting the above multi-modal operation data, the power company invited experts in the relevant field to jointly study how to effectively extract fault features with high discrimination from these data. The specific scheme is as follows:

[0297] For the time series data of input and output electrical parameters, a pre-trained long short-term memory (LSTM) network is used for feature extraction. The LSTM network can learn the long-term time dependence between input and output electrical parameters, such as the specific nonlinear time series pattern of input and output parameters of the reactor under certain fault conditions. At the same time, a convolutional neural network (CNN) is also used to extract local abnormal features in these electrical parameters, such as sudden points of current and phase. The features extracted by LSTM and CNN are adaptively fused through an attention mechanism to form the final electrical parameter features.

[0298] For infrared thermal imaging data, a shallow convolutional network is first used to extract low-level visual features such as edges and textures, reflecting the detailed information of the temperature distribution on the surface of the reactor. Then, a deeper convolutional network is used to extract high-level semantic features such as the shape and distribution of different temperature regions, reflecting the overall thermal characteristics of the reactor. The low-level and high-level features are adaptively fused through an attention mechanism to obtain comprehensive infrared thermal imaging features.

[0299] For sound wave signals, a spectral analysis network is first used to extract low-frequency harmonic features, such as low-frequency sound waves caused by mechanical vibration of the reactor winding. Then, a special high-frequency feature extraction network is used to extract high-frequency discharge noise features, such as high-frequency pulse sound waves caused by local breakdown. At the same time, a transient feature extraction network is also used to capture electrical pulse features in the sound wave signal, such as transient discharge sound waves caused by insulation faults. The above three different frequency band sound wave features are fused to obtain comprehensive sound wave features.

[0300] Finally, for the vibration signal, three different vibration feature extraction sub-networks are designed. Among them, the large amplitude vibration feature extraction network mainly captures the large amplitude abnormal vibration on the surface of the reactor, such as impact vibration caused by serious failure; the small amplitude vibration feature extraction network extracts the tiny progressive vibration features, such as tiny vibration caused by component wear; the composite vibration feature extraction network extracts the complex vibration mode caused by electromagnetic-mechanical coupling. The three kinds of vibration features are integrated through the fusion network to form the final vibration feature.

[0301] Through the above deep learning feature extraction for different modal data, the power company obtains four kinds of discriminative feature vectors, which reflect the electrical, thermal, acoustic and vibration features of the reactor operating state. These features will be used as input for the subsequent fault diagnosis model.

[0302] 3. Multi-modal feature fusion

[0303] The power company's expert team further designs a multi-modal feature fusion network based on attention mechanism. The network contains four feature encoders, corresponding to electrical parameter features, infrared thermal imaging features, acoustic features and vibration features. Each encoder encodes the features of the corresponding modal to generate a feature vector of that modal.

[0304] Then, the interaction attention module in the network interacts and fuses the four modal features. Specifically, this module adaptively learns the correlation and fusion weight between different modal features, so that each modal feature can obtain valuable complementary information from other modal features, thereby enhancing the performance of fault diagnosis.

[0305] The fused multi-modal feature vector is input into the subsequent fault diagnosis model. This multi-modal feature fusion method fully utilizes the complementary advantages of data from different monitoring methods, providing a comprehensive and discriminative feature representation for subsequent fault prediction.

[0306] 4. Fault prediction modeling

[0307] Based on the results of the above multi-modal feature fusion, the power company further builds a fault prediction model, which consists of two main components: fault type classifier and fault probability regressor.

[0308] The fault type classifier outputs the possible fault types of the reactor, such as open-phase fault, overload fault, insulation breakdown, etc. The classifier uses a Softmax output layer to give the probability distribution of each fault type.

[0309] The fault probability regressor is responsible for predicting the probability of occurrence of each possible fault type. The regressor uses a Sigmoid output layer to give a probability value between 0 and 1.

[0310] During the model training process, the power company collects a large amount of operating data of the reactor of known fault types, including electrical parameters, infrared thermal imaging, sound waves, and vibration and other multi-modal data, and labels these data, giving the fault type and fault probability of each sample. The labeled training data is input into the above fault prediction model for supervised learning.

[0311] During the training process, the power company also sets up a validation set to monitor the performance of the model on new data in real time, and selects the model parameters with the best validation set performance to avoid overfitting and improve the generalization ability of the model.

[0312] The final trained fault prediction model can output the possible fault types and their probabilities of the reactor. The company uses this fault prediction result as an important basis for reactor condition monitoring and preventive maintenance.

[0313] According to the post-test, the accuracy rate of fault type classification: used to evaluate the model's ability to identify different fault types. For example, on the test set, the classification accuracy rates of the model for phase failure, overload failure, and insulation breakdown were 90%, 88%, and 92%, respectively.

[0314] Log loss of fault probability prediction: used to evaluate the prediction accuracy of the model for the probability of each type of fault. The lower the index, the closer the model's prediction result is to the actual probability distribution. On the test set, the average log loss of the model was 0.21.

[0315] By regularly evaluating these performance indicators, the power company can not only understand the effectiveness of the method in actual application, but also can optimize the model structure and hyperparameters to further improve the accuracy and reliability of fault prediction, ensuring the safe operation of the reactor.

[0316] In summary, the power company has made good application results in the actual operation and management of the substation by fully utilizing the deep learning-based dry-type air-core reactor operating fault prediction method proposed by the invention.

[0317] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention.

Claims

1. A dry-type air-core reactor in-house operation fault prediction method, characterized by, Includes the following steps: S10. Real-time acquisition of input electrical parameters, output electrical parameters, infrared thermal imaging, acoustic imaging, and vibration signal set of the dry-type air-core reactor under test during its indoor operation. The vibration signal set includes vibration signals collected by multiple vibration sensors at the detection points on the surface of the dry-type air-core reactor under test. S20. Using a pre-trained first feature extraction model for analyzing electrical parameters, extract the features of the input electrical parameters and the output electrical parameters as the first features; S30. Using a pre-trained second feature extraction model for analyzing heat sources, extract the features of the infrared thermal imaging, which are denoted as the second feature. S40. Using a pre-trained third feature extraction model for noise analysis, extract the features of the acoustic imaging, denoted as the third feature; S50. Using a pre-trained fourth feature extraction model for vibration analysis, extract the features of the vibration signal group, denoted as the fourth feature; S60. Input the first feature, second feature, third feature, and fourth feature into the pre-trained fault diagnosis model to obtain a fault vector including fault type and fault occurrence probability; S70. Predict the fault of the dry-type air-core reactor under test based on the fault vector. The first feature extraction model includes a long memory sub-network, a short memory sub-network, and a first fusion sub-network; wherein, the long memory sub-network is based on a long short-term memory network structure and is used to learn the long-term dependencies and temporal patterns of input electrical parameters and output electrical parameters; the short memory sub-network is based on a convolutional neural network and an attention mechanism and is used to capture local features and abrupt changes of input electrical parameters and output electrical parameters; the first fusion sub-network fuses the outputs of the long memory sub-network and the short memory sub-network, and the output is used as the first feature; The second feature extraction model includes a low-level feature extraction sub-network, a high-level feature extraction sub-network, and a second fusion sub-network. The low-level feature extraction sub-network is based on a shallow convolutional neural network and is used to extract low-level features of infrared thermal imaging, including edges and textures. The high-level feature extraction sub-network is based on a deep convolutional neural network and is used to extract high-level semantic features of infrared thermal imaging. The second fusion sub-network fuses the low-level and high-level features, uses an attention mechanism to guide the fusion weights, and outputs the result as the second feature. The third feature extraction model includes a low-frequency feature extraction subnetwork, a high-frequency feature extraction subnetwork, a transient feature extraction subnetwork, and a third fusion subnetwork; wherein, the low-frequency feature extraction subnetwork is used to capture low-frequency harmonic signals in acoustic imaging; the high-frequency feature extraction subnetwork is used to capture high-frequency discharge noise; the transient feature extraction subnetwork is used to capture instantaneous electrical pulse signals; and the third fusion subnetwork fuses the above three features to obtain a comprehensive feature representation of acoustic imaging; The fourth feature extraction model includes a large-amplitude vibration feature extraction subnetwork, a small-amplitude vibration feature extraction subnetwork, a composite vibration feature extraction subnetwork, and a fourth fusion subnetwork. The large-amplitude vibration feature extraction subnetwork captures large-amplitude vibration signals from the reactor surface; the small-amplitude vibration feature extraction subnetwork captures minute vibrations; the composite vibration feature extraction subnetwork captures combinations of multiple vibration modes; and the fourth fusion subnetwork fuses these three vibration features to obtain a comprehensive feature representation of the vibration signal group. The fault diagnosis model is structured as a multimodal fusion network based on an attention mechanism, comprising multiple modal feature encoders, an interactive attention module, a fault type classifier, and a fault probability classifier. The multiple modal feature encoders correspond to four modalities of data: electrical parameters, infrared imaging, acoustic imaging, and vibration signals, encoding features for each modality. The interactive attention module interacts with the encoded features of different modalities, enhancing the information between them. The fault type classifier outputs the specific type of fault, and the fault probability classifier outputs the probability of occurrence for each fault.

2. The method for predicting indoor operation faults of a dry-type air-core reactor according to claim 1, characterized in that, The multiple modal feature encoders correspond to four modal data: electrical parameters, infrared imaging, acoustic imaging, and vibration signals. The step of encoding the features of each modal data includes an electrical parameter modal feature encoder corresponding to the first feature, an infrared imaging modal feature encoder corresponding to the second feature, an acoustic imaging modal feature encoder corresponding to the third feature, and a vibration signal modal feature encoder corresponding to the fourth feature.

3. The method for predicting indoor operation faults of a dry-type air-core reactor according to claim 1, characterized in that, If the probability of a certain type of fault occurring exceeds a preset probability threshold, then the fault of that type is determined to have occurred.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed, are used to perform the indoor operation fault prediction method for a dry-type air-core reactor as described in any one of claims 1-3.

5. A fault prediction system for indoor operation of a dry-type air-core reactor, characterized in that, It includes the computer-readable storage medium of claim 4.

Images