Oil-immersed reactor oil injection fault risk prediction method, medium and system

Abstract

The invention provides an oil-immersed reactor oil injection fault risk prediction method, medium and system, and belongs to the technical field of oil-immersed reactors, and the method comprises the steps: firstly collecting operation data, vibration data and noise data from a historical reactor with an oil injection fault, and recording the oil injection time; then preprocessing and feature extraction are carried out on the data to obtain various feature groups; next, the relevancy between each feature and the oil injection moment is calculated by adopting a Bayesian network model, and clustering is carried out to obtain a typical feature group; and establishing and training an oil injection fault prediction model based on the typical feature groups. And for a to-be-detected reactor, extracting operation, vibration and noise feature groups of the reactor, and inputting the feature groups into the prediction model to obtain an oil injection prediction moment so as to judge a fault risk. According to the method, the prediction model is established by using historical data, the health state of the reactor can be automatically and continuously monitored, potential oil injection faults can be timely identified, and the technical problems that the monitoring means is single and the equipment state cannot be comprehensively reflected in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil-immersed reactors, and in particular, relates to a method, a medium and a system for predicting the risk of oil injection failure of an oil-immersed reactor. Background Art

[0002] Oil-immersed reactors are widely used in power systems, playing an important role in balancing grid voltage, improving power factor, and reducing power loss. However, during long-term operation, oil-immersed reactors are prone to oil injection failures, which can seriously affect the normal operation of the equipment and even lead to serious accidents.

[0003] The main reasons for oil-immersed reactor oil injection failure include: insulation aging, overload, cooling system failure, etc. When the insulation layer ages or the insulation strength decreases, local discharge will occur under high voltage, further accelerating the aging of the insulation, and eventually causing insulation breakdown, short-circuit discharge between electrodes, and oil injection. Overload will cause the core and coil to heat up, the oil temperature will rise, and the aging and decomposition of the oil will be accelerated, leading to oil injection. Cooling system failure will cause the oil temperature to be too high, causing oil injection. Other factors such as vibration, impurities, etc. will also accelerate the aging and decomposition of the oil.

[0004] At present, there are mainly the following methods for condition monitoring and fault prediction of oil-immersed reactors:

[0005] 1. Periodic inspection method: Regularly check the equipment, measure electrical parameters, test oil samples, etc. This method is simple and direct, but it cannot monitor the status of the equipment in real time, and can only detect faults that have occurred, but cannot provide early warning.

[0006] 2. Online monitoring method: Install temperature, current, vibration and other sensors on the equipment to monitor various parameters in real time. However, this monitoring method cannot fully reflect the operating status of the equipment and cannot accurately predict failures.

[0007] 3. Fault diagnosis method: By analyzing the characteristic signals of the equipment during operation, such as vibration, sound, temperature, etc., the fault type and severity of the equipment can be diagnosed. This method requires a lot of expert experience and test data support.

[0008] To sum up, there are some problems with the existing oil-immersed reactor status monitoring and fault prediction technologies: the monitoring means are single and cannot fully reflect the equipment status; fault diagnosis relies on expert experience, and the prediction accuracy for complex faults is not high; to address these problems, there is an urgent need to study a new method that can comprehensively utilize multi-source heterogeneous data to accurately predict oil-immersed reactor injection faults. Summary of the invention

[0009] In view of this, the present invention provides a method, medium and system for predicting the risk of oil injection failure of an oil-immersed reactor, which can solve the technical problem that the existing technology has a single monitoring means and cannot fully reflect the equipment status.

[0010] The present invention is achieved in that:

[0011] A first aspect of the present invention provides a method for predicting the risk of oil injection failure of an oil-immersed reactor, comprising the following steps:

[0012] S10, acquiring historical operation data, vibration data, and noise data of multiple oil-immersed reactors with oil injection faults during operation, and recording the oil injection timing of the oil-immersed reactors;

[0013] S20, preprocessing the operation data, vibration data and noise data, and extracting features using a sliding window to obtain an operation feature group, a vibration feature group and a noise feature group;

[0014] S30, using a Bayesian network model to respectively calculate the correlation between each operating feature, each vibration feature, each noise feature in the operating feature group, the vibration feature group, and the noise feature group and the injection timing;

[0015] S40, clustering the operation feature group, the vibration feature group, and the noise feature group respectively according to the correlation to obtain a typical operation feature group, a typical vibration feature group, and a typical noise feature group;

[0016] S50, establishing and training an injection fault prediction model based on the typical operation feature group, the typical vibration feature group and the typical noise feature group;

[0017] S60, continuously acquiring operation data, vibration data and noise data of the oil-immersed reactor to be detected, and extracting an operation feature group, a vibration feature group and a noise feature group of the oil-immersed reactor to be detected according to a sliding window of a preset length, and recording them as an operation feature group to be tested, a vibration feature group to be tested and a noise feature group to be tested;

[0018] S70, inputting the operation feature group to be measured, the vibration feature group to be measured and the noise feature group to be measured into the injection fault prediction model, and outputting the injection prediction time;

[0019] S80: Determine the failure risk of the oil-immersed reactor to be detected according to the predicted injection time, and output it.

[0020] The operating data include: voltage, current, frequency, power factor, active power, reactive power, apparent power and oil temperature.

[0021] The vibration data is a vibration signal collected by a vibration sensor arranged on the oil-immersed reactor; and the noise data is an audio signal collected by an audio sensor arranged around the oil-immersed reactor.

[0022] Among them, the preset length of the sliding window is obtained according to the minimum interval of the injection time and the large cycle of each type of data, and the large cycle refers to the minimum time required for this type of data to fluctuate completely once. The step of obtaining the sliding window length specifically includes: obtaining the minimum interval of the injection time and the large cycle of each type of data, and taking the minimum value therebetween as the sliding window length.

[0023] Furthermore, the complete fluctuation refers to the process in which data passes from one extreme point through another extreme point and then returns to the original extreme point.

[0024] The specific steps of calculating the large cycle are: first determine the sampling frequency of the data, then analyze the periodic changes of the data, find out the recurring patterns, and finally calculate the shortest time required to complete a complete pattern.

[0025] Among them, the injection fault prediction model adopts a multi-branch neural network structure, including an operation feature injection associated branch network, a vibration feature injection associated branch network, a noise feature injection associated branch network, a fusion feature injection associated branch network, a timing prediction branch network, a feature correlation branch network, an anomaly detection branch network and a summary branch network.

[0026] Furthermore, the operation feature injection correlation branch network is used to analyze the correlation between the operation feature and the injection event, the input is a typical operation feature group, the output is the operation feature injection correlation, and the structure is a multi-layer perceptron;

[0027] The vibration feature injection correlation branch network is used to analyze the correlation between vibration features and injection events, the input is a typical vibration feature group, the output is the vibration feature injection correlation, and the structure is a convolutional neural network;

[0028] The noise feature injection correlation branch network is used to analyze the correlation between noise features and injection events, the input is a typical noise feature group, the output is the noise feature injection correlation, and the structure is a recurrent neural network;

[0029] The fusion feature injection correlation branch network is used to comprehensively analyze the correlation between multiple features and injection events, the input is a combination of a typical operation feature group, a vibration feature group and a noise feature group, the output is the fusion feature injection correlation, and the structure is an attention mechanism network;

[0030] The time series prediction branch network is used to predict the time series of injection events, the input is the historical injection time data, the output is the next possible injection time, and the structure is a long short-term memory network;

[0031] The feature correlation branch network is used to analyze the correlation between different features, the input is all feature groups, the output is the correlation matrix between features, and the structure is an autoencoder network;

[0032] The anomaly detection branch network is used to detect abnormal patterns in features, the input is real-time feature data, the output is anomaly scores, and the structure is an isolation forest algorithm;

[0033] The summary branch network is used to integrate the outputs of all branch networks, the input is the output results of each branch network, the output is the final injection prediction time, and the structure is a fully connected network.

[0034] Furthermore, each branch network also includes an aggregation module for extracting macro features and a refinement module for extracting micro features. The macro features refer to features that reflect overall trends, and the micro features refer to features that reflect local changes. The threshold for distinguishing between macro features and micro features is specifically determined based on the time scale and impact range of the features.

[0035] The equation for judging the fault risk according to the injection prediction time is specifically described as: comparing the difference between the prediction time and the current time with a preset threshold, if the difference is less than the threshold, it is judged as high risk, otherwise it is low risk. The risk level can be further subdivided according to the size of the difference.

[0036] 1. Threshold for distinguishing macroscopic features from microscopic features:

[0037] Discrimination threshold = average characteristic period * impact factor;

[0038] Among them: the average characteristic period refers to the average change period of all characteristics; the impact factor is a coefficient between 0 and 1, which is used to adjust the strictness of the threshold.

[0039] The specific steps are as follows:

[0040] a. Calculate the change period of each feature;

[0041] b. Take the average value of all characteristic change cycles to obtain the average characteristic cycle;

[0042] c. Determine the impact factors based on expert experience or historical data analysis;

[0043] d. Multiply the average characteristic period by the influence factor to obtain the final discrimination threshold;

[0044] Features above this threshold are considered macro features, and features below this threshold are considered micro features.

[0045] 2. Preset threshold:

[0046] Preset threshold = basic time + adjustment coefficient * historical average injection interval;

[0047] Among them: the basic time is a fixed minimum warning time; the adjustment coefficient is a coefficient used for fine-tuning, usually between 0 and 1; the historical average injection interval is the average interval time of injection events in the past period of time;

[0048] The specific steps are as follows:

[0049] a. Set the base time, such as 24 hours;

[0050] b. Analyze historical data and calculate the average injection interval;

[0051] c. Determine the adjustment coefficient according to the characteristics and requirements of the system;

[0052] d. Add the basic time and the adjustment coefficient and multiply by the historical average injection interval to get the final preset threshold

[0053] When the difference between the predicted injection time and the current time is less than this preset threshold, the system will issue a high-risk warning.

[0054] Wherein, the step S10 specifically includes: obtaining the operation data, vibration data and noise data of multiple oil-immersed reactors with injection faults during operation, and recording the injection time of the oil-immersed reactor. The operation data includes parameters such as voltage, current, frequency, power factor, active power, reactive power, apparent power, oil temperature, etc.; the vibration data is the vibration signal collected by the vibration sensor installed on the oil-immersed reactor; the noise data is the audio signal collected by the audio sensor installed around the oil-immersed reactor. The purpose of obtaining these historical data and fault time is to lay the foundation for the subsequent establishment of an injection fault prediction model.

[0055] The specific steps of step S20 include: preprocessing the acquired operation data, vibration data and noise data, including denoising, calibration and other operations to ensure data quality; and then extracting features from the preprocessed data using a sliding window method. The length of the sliding window is determined according to the actual situation, mainly considering two factors: one is the minimum time interval of historical injection events, and the other is the large cycle of each type of data. The so-called large cycle refers to the minimum time required for this type of data to fluctuate once completely. Through the sliding window, a set of operation features can be extracted from the operation data, a set of vibration features can be extracted from the vibration data, and a set of noise features can be extracted from the noise data.

[0056] Wherein, the step S30 adopts a Bayesian network model to calculate the correlation between each feature in the operation feature group, the vibration feature group and the noise feature group and the historical injection time. The Bayesian network is a probabilistic graph model that can effectively characterize the conditional dependency between variables. Through the Bayesian network, the degree of correlation between each feature and the injection event can be quantified, providing a basis for subsequent feature clustering and model training.

[0057] Wherein, the step S40 clusters the operation feature group, the vibration feature group and the noise feature group respectively according to the feature correlation calculated in the previous step. The purpose of clustering is to classify the features with higher correlation into one category to form several typical feature groups. The clustering algorithm used here is the K-Means algorithm, which is a simple and efficient unsupervised learning algorithm that can automatically discover the cluster structure in the data. Through clustering, the most typical features related to the injection event in various data types can be extracted, and these features will be used as input for subsequent model training.

[0058] Wherein, the step S50 establishes and trains an injection fault prediction model based on the typical operation feature group, typical vibration feature group and typical noise feature group extracted above. This model adopts a multi-branch neural network structure, including an operation feature injection correlation branch network, a vibration feature injection correlation branch network, a noise feature injection correlation branch network, a fusion feature injection correlation branch network, a time series prediction branch network, a feature correlation branch network, an anomaly detection branch network and a summary branch network. By training this multi-branch neural network, the complex relationship between different types of features and injection events can be fully explored to improve the accuracy of injection fault prediction.

[0059] Among them, the step S60 needs to continuously obtain the operation data, vibration data and noise data of the oil-immersed reactor to be tested, and extract the operation feature group, vibration feature group and noise feature group to be tested according to the sliding window length determined previously. These feature groups will be used as inputs and sent to the injection fault prediction model trained in step S50 for prediction.

[0060] Wherein, the step S70 is to input the operation feature group to be tested, the vibration feature group and the noise feature group extracted in step S60 into the injection fault prediction model established in step S50, and the model will output an injection prediction time.

[0061] Wherein, the step S80 calculates the time difference between the predicted injection time and the current time according to the injection prediction time outputted by the step S70, and compares it with a preset threshold value. The calculation formula of the preset threshold value is: preset threshold value = basic time + adjustment coefficient * historical average injection interval. Through this method, the preset threshold value can be dynamically adjusted so that the prediction result is more in line with the actual operation situation. Finally, the system will output the fault risk of the oil-immersed reactor to be detected according to the judgment result of high risk or low risk.

[0062] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for predicting the risk of oil injection failure of an oil-immersed reactor.

[0063] A third aspect of the present invention provides an oil-immersed reactor injection fault risk prediction system, which includes the above-mentioned computer-readable storage medium.

[0064] Compared with the prior art, the method, medium and system for predicting the risk of oil injection failure of an oil-immersed reactor provided by the present invention have the following beneficial effects:

[0065] 1. Comprehensively reflect the equipment status. This method makes full use of multi-source heterogeneous data such as operation data, vibration data and noise data, which can more comprehensively reflect the operation status of the equipment and make up for the shortcomings of a single monitoring method.

[0066] 2. Improve prediction accuracy. The multi-branch neural network structure can deeply explore the complex relationship between different features and faults, greatly improving the accuracy of injection fault prediction, which is significantly improved compared with the machine learning method based on a single feature.

[0067] 3. Realize real-time monitoring. This method can obtain equipment operation data in real time, continuously extract features, and input them into the prediction model to achieve dynamic monitoring of equipment status and abnormal warning, providing important support for fault prevention.

[0068] Therefore, the solution of the present invention solves the technical problem that the existing technology has a single monitoring means and cannot fully reflect the equipment status. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0071] like Figure 1 FIG. 1 is a flow chart of a method for predicting the risk of oil injection failure of an oil-immersed reactor provided by the present invention. The method comprises the following steps:

[0072] S10, obtaining historical operation data, vibration data, and noise data of multiple oil-immersed reactors with oil injection faults during operation, and recording the oil injection time of the oil-immersed reactors;

[0073] S20, preprocessing the operation data, vibration data and noise data, and extracting features using a sliding window to obtain an operation feature group, a vibration feature group and a noise feature group;

[0074] S30, using a Bayesian network model to calculate the correlation between each operating feature, each vibration feature, each noise feature in the operating feature group, the vibration feature group, and the noise feature group and the injection timing;

[0075] S40, clustering the operation feature group, the vibration feature group, and the noise feature group respectively according to the correlation to obtain a typical operation feature group, a typical vibration feature group, and a typical noise feature group;

[0076] S50, establishing and training an injection fault prediction model based on the typical operation feature group, the typical vibration feature group, and the typical noise feature group;

[0077] S60, continuously acquiring operation data, vibration data and noise data of the oil-immersed reactor to be detected, and extracting an operation feature group, a vibration feature group and a noise feature group of the oil-immersed reactor to be detected according to a sliding window of a preset length, and recording them as an operation feature group to be tested, a vibration feature group to be tested and a noise feature group to be tested;

[0078] S70, inputting the operation feature group to be tested, the vibration feature group to be tested, and the noise feature group to be tested into the injection fault prediction model, and outputting the injection prediction time;

[0079] S80: Determine the failure risk of the oil-immersed reactor to be detected according to the predicted injection time, and output it.

[0080] The specific implementation methods of the above steps are described in detail below:

[0081] The specific implementation method of step S10 is: first, the system needs to obtain the historical operation data, vibration data and noise data recorded during the operation of multiple oil-immersed reactors with fuel injection failures, and record the specific time when these oil-immersed reactors have fuel injection failures. The so-called operation data includes parameters such as voltage, current, frequency, power factor, active power, reactive power, apparent power, oil temperature, etc.; the so-called vibration data is the vibration signal collected by the vibration sensor installed on the oil-immersed reactor; the so-called noise data is the audio signal collected by the audio sensor installed around the oil-immersed reactor. The purpose of obtaining these historical data and fault time is to lay the foundation for the subsequent establishment of a fuel injection failure prediction model.

[0082] The specific implementation method of step S20 is: first, the acquired operation data, vibration data and noise data are preprocessed, including denoising, calibration and other operations to ensure data quality. Then, the sliding window method is used to extract features from these preprocessed data. The length of the sliding window is determined according to the actual situation, and two factors are mainly considered: one is the minimum time interval of historical injection events, and the other is the large cycle of each type of data. The so-called large cycle refers to the minimum time required for this type of data to fluctuate once completely. The specific method for calculating the length of the sliding window is: first, the minimum time interval of historical injection events and the large cycle of each type of data are obtained, and then the minimum value is taken as the length of the sliding window. Through the sliding window, a set of operation features can be extracted from the operation data, a set of vibration features can be extracted from the vibration data, and a set of noise features can be extracted from the noise data. These features are the basis for subsequent modeling.

[0083] The specific implementation method of step S30 is: using the Bayesian network model, respectively calculating the correlation between each feature in the operation feature group, the vibration feature group and the noise feature group and the historical injection time. The Bayesian network is a probabilistic graph model that can effectively characterize the conditional dependency between variables. Through the Bayesian network, the degree of correlation between each feature and the injection event can be quantified, providing a basis for subsequent feature clustering and model training.

[0084] The specific implementation of step S40 is: clustering the operation feature group, vibration feature group and noise feature group respectively according to the feature correlation calculated in the previous step. The purpose of clustering is to classify the features with higher correlation into one category to form several typical feature groups. The clustering algorithm used here is the K-Means algorithm, which is a simple and efficient unsupervised learning algorithm that can automatically discover the cluster structure in the data. Through clustering, the most typical features related to the injection event in various data types can be extracted, and these features will be used as input for subsequent model training.

[0085] The specific implementation of step S50 is: based on the typical operation feature group, typical vibration feature group and typical noise feature group extracted above, establish and train a fuel injection fault prediction model. This model adopts a multi-branch neural network structure, including the following branches:

[0086] 1) Operation feature injection association branch network: The input is a typical operation feature group, and the output is the correlation between the operation feature and the injection event, using a multi-layer perceptron structure.

[0087] 2) Vibration feature injection association branch network: The input is a typical vibration feature group, and the output is the correlation between the vibration feature and the injection event, using a convolutional neural network structure.

[0088] 3) Noise feature injection association branch network: The input is a typical noise feature group, and the output is the correlation between the noise feature and the injection event, using a recurrent neural network structure.

[0089] 4) Fusion feature injection association branch network: The input is a combination of the three typical feature groups mentioned above, and the output is the correlation between the comprehensive features and the injection event, using an attention mechanism network structure.

[0090] 5) Time series prediction branch network: The input is the historical injection time data, and the output is the next possible injection time, using a long short-term memory network structure.

[0091] 6) Feature correlation branch network: The input is all feature groups, the output is the correlation matrix between features, and the autoencoder network structure is adopted.

[0092] 7) Anomaly detection branch network: The input is real-time feature data, the output is anomaly score, and the isolation forest algorithm is used.

[0093] 8) Aggregate branch network: The input is the output results of the above branch networks, and the output is the final injection prediction time, using a fully connected network structure.

[0094] By training this multi-branch neural network, the complex relationship between different types of features and injection events can be fully explored, thereby improving the accuracy of injection fault prediction.

[0095] The specific implementation of step S60 is: in practical applications, the system needs to continuously obtain the operation data, vibration data and noise data of the oil-immersed reactor to be tested, and extract the operation feature group, vibration feature group and noise feature group to be tested according to the sliding window length determined above. These feature groups will be used as inputs and sent to the injection fault prediction model trained in step S50 for prediction.

[0096] The specific implementation method of step S70 is: input the operating feature group to be tested, the vibration feature group and the noise feature group extracted in step S60 into the injection fault prediction model established in step S50, and the model will output an injection prediction time.

[0097] The specific implementation method of step S80 is: according to the injection prediction time output by step S70, the system needs to calculate the time difference between the prediction time and the current time, and compare it with a preset threshold. If the time difference is less than the preset threshold, it is judged as high risk; otherwise it is judged as low risk. The calculation formula of the preset threshold is: preset threshold = basic time + adjustment coefficient * historical average injection interval. Among them, the basic time is a fixed minimum warning time, such as 24 hours; the adjustment coefficient is a coefficient between 0 and 1, which is used to fine-tune the threshold; the historical average injection interval is the average interval time calculated based on the actual injection events that occurred in the past period of time. Through this method, the preset threshold can be dynamically adjusted to make the prediction result more in line with the actual operation situation. Finally, the system will output the fault risk of the oil-immersed reactor to be detected based on the judgment result of high risk or low risk.

[0098] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for predicting the risk of oil injection failure of an oil-immersed reactor.

[0099] A third aspect of the present invention provides an oil-immersed reactor injection fault risk prediction system, which includes the above-mentioned computer-readable storage medium.

[0100] Specifically, the principle of the present invention is: using a deep learning model that integrates multi-source heterogeneous data can comprehensively analyze the operating status of the equipment and improve the accuracy of injection fault prediction.

[0101] Specifically, the method first obtains the operating data, vibration data and noise data of the oil-immersed reactor with historical injection faults, including operating characteristics such as voltage, current, frequency, power factor, and vibration signals and noise signals collected by vibration sensors and audio sensors installed on the equipment. These data reflect information on multiple levels, such as the operating status, mechanical status and environmental status of the equipment.

[0102] Then, the sliding window is used to extract the features of these data to obtain the operation features, vibration features and noise features. The Bayesian network model is used to analyze the correlation between these features and the historical injection time, and the typical features most related to the injection event are screened out. This provides a reliable feature input for the subsequent model establishment.

[0103] Next, based on these typical features, a multi-branch neural network model is constructed. This model includes operation feature injection correlation branches, vibration feature injection correlation branches, noise feature injection correlation branches, etc., which can fully explore the complex relationship between different types of features and injection failures. It also includes time series prediction branches, feature correlation branches, etc., which can explore the time series correlation and mutual influence between features. In addition, there is anomaly detection branches, which can monitor the abnormality of feature data in real time.

[0104] By training this multi-branch neural network model, the complex mapping relationship between the equipment status and the injection fault can be learned. In practical applications, the system can collect the operation data, vibration data and noise data of the monitored equipment in real time, extract the corresponding features, input them into the trained model, and output the injection prediction time of the equipment. According to the difference between the predicted time and the current time, the failure risk of the equipment is judged and the corresponding warning is issued.

[0105] Compared with traditional single monitoring methods and machine learning methods based on single features, this deep learning method based on multi-source heterogeneous data fusion can more comprehensively reflect the operating status of the equipment and improve the accuracy and reliability of injection fault prediction. On the one hand, it makes full use of multi-source information such as operating data, vibration data and noise data, so as to more accurately detect potential abnormalities of the equipment. On the other hand, the structure of the multi-branch neural network can mine the complex associations between different features, enhance the modeling ability of the fault mechanism, and thus improve the accuracy of the prediction. In addition, this method can also monitor the equipment status in real time, detect abnormal conditions in time, and provide support for fault warning.

[0106] In order to better understand and implement the present invention, a specific embodiment 1 of the present invention is provided below. The steps of this embodiment 1 specifically include: First, for step S10, it is necessary to obtain the historical operation data, vibration data and noise data recorded during the operation of multiple oil-immersed reactors with injection failures, and record the specific time when the injection failures of these oil-immersed reactors occur. These data can be expressed as:

[0107] Operation data:

[0108] Vibration data:

[0109] Noise data:

[0110] Injection time set: T = {t1, t2, ..., t m};

[0111] in, represents the i-th running data sample, represents the jth vibration data sample, represents the kth noise data sample; t l Indicates the time when the first injection occurs. r ,n v ,n n They represent the number of samples of operation data, vibration data and noise data respectively, and m represents the number of times injection failure occurs.

[0112] Step S20 is to preprocess and extract features from these data. Preprocessing includes operations such as denoising and calibration to ensure data quality. Then, the sliding window method is used to extract features from these preprocessed data. The determination of the sliding window length needs to consider two factors:

[0113] 1. Minimum time interval Δt between historical injection events min

[0114] 2. The maximum period T of each type of data r ,T v ,T n

[0115] The so-called large cycle refers to the minimum time required for this type of data to fluctuate completely once. The specific method for calculating the sliding window length w is:

[0116] w=min{Δt min ,T r ,T v ,T n};

[0117] Through the sliding window, a set of operating features can be extracted from the operating data Extract a set of vibration features from the vibration data Extract a set of noise features from noisy data Among them, d r ,d v ,d n They represent the dimensions of operation characteristics, vibration characteristics, and noise characteristics respectively.

[0118] Next is step S30, using the Bayesian network model to calculate the correlation between each feature in the operation feature group, vibration feature group and noise feature group and the historical injection time. The Bayesian network is a probabilistic graph model, which can be expressed as:

[0119]

[0120] Among them, P(T|F r ,F v ,F n ) represents a given feature combination (Fr ,F v ,F n ), the posterior probability of injection time T; P(F r ,F v ,F n |T) means that when the injection time T occurs, the characteristic combination (F r ,F v ,F n ) is the likelihood probability; P(T) represents the prior probability of injection time T; P(F r ,F v ,F n ) represents the feature combination (F r ,F v ,F n ). Through the Bayesian network, the degree of association between each feature and the injection event can be quantified, that is, the calculation

[0121] In step S40, the operation feature group, vibration feature group and noise feature group are clustered according to the feature correlation calculated in the previous step. The clustering algorithm used here is the K-Means algorithm, and its objective function is:

[0122]

[0123] Among them, k represents the number of cluster centers, C i represents the i-th cluster, μ i represents the i-th cluster center, and x represents the sample point. Through the K-Means algorithm, three groups of typical features can be obtained:

[0124] Typical operating characteristics group:

[0125] Typical vibration characteristics group:

[0126] Typical noise signature groups:

[0127] in, They represent the dimensions of typical operation characteristics, typical vibration characteristics, and typical noise characteristics respectively.

[0128] Step S50 is to establish and train a fuel injection fault prediction model based on the typical feature group extracted above. This model adopts a multi-branch neural network structure and can be expressed as:

[0129]

[0130] in, represents the predicted injection time, and F represents the multi-branch neural network model. Specifically, the model includes the following branches:

[0131] 1. Run the characteristic injection associated branch network:

[0132]

[0133] Among them, y r represents the correlation between the operating characteristics and the injection event, f r Represents the multi-layer perceptron network structure.

[0134] 2. Vibration characteristic injection associated branch network:

[0135]

[0136] Among them, y v represents the correlation between vibration characteristics and injection events, f v Represents the convolutional neural network structure.

[0137] 3. Noise characteristic injection associated branch network:

[0138]

[0139] Among them, y n represents the correlation between noise characteristics and injection events, f n Represents a recurrent neural network structure.

[0140] 4. Fusion feature injection association branch network:

[0141]

[0142] Among them, y f represents the correlation between the fusion feature and the injection event, f f Represents the attention mechanism network structure.

[0143] 5. Timing prediction branch network:

[0144]

[0145] in, represents the predicted next injection time, f t Represents the long short-term memory network structure.

[0146] 6. Feature Correlation Branch Network:

[0147]

[0148] Among them, C represents the correlation matrix between features, f c Represents the autoencoder network structure.

[0149] 7. Anomaly Detection Branch Network:

[0150] a=f a ({F r ,F v ,F n});

[0151] Among them, a represents the anomaly score, f a Represents the Isolation Forest Algorithm.

[0152] 8. Summarize the branch network:

[0153]

[0154] Among them, f s Represents a fully connected network structure.

[0155] By training this multi-branch neural network model, the complex relationship between different types of features and injection events can be fully explored, thereby improving the accuracy of injection fault prediction.

[0156] Step S60 is that in actual application, the system needs to continuously obtain the operating data, vibration data and noise data of the oil-immersed reactor to be tested, and extract the operating feature group to be tested according to the sliding window length w determined previously. Vibration signature group and noise feature set These feature groups will be used as input and sent to the trained injection fault prediction model in step S50 for prediction.

[0157] Finally, step S70 is to group the operating characteristics to be tested extracted in step S60. Vibration signature group and noise feature set The result is input into the fuel injection fault prediction model established in step S50, and the model will output a fuel injection prediction time.

[0158] The following is an example 2 of a specific application scenario of the present invention: A power operation and maintenance team is responsible for the operation and maintenance of a batch of oil-immersed reactors, which are distributed in different substations and play an important role in the power system. In order to improve the reliability and safety of these reactors, the operation and maintenance team decided to adopt the oil-immersed reactor oil injection failure risk prediction method proposed in the present invention to comprehensively monitor and warn the equipment status.

[0159] The operation and maintenance team first collected the historical operating data of 12 oil-immersed reactors that had experienced fuel injection failures in the past five years. These data include: operating parameters such as voltage, current, frequency, power factor, active power, reactive power, apparent power, oil temperature, etc.; vibration signals collected by vibration sensors installed on the reactors; and noise signals collected by audio sensors installed around the reactors. At the same time, the operation and maintenance team also recorded the specific time when the fuel injection failures occurred in these 12 reactors.

[0160] The following takes one of the reactors as an example to introduce the specific implementation process:

[0161] 1. Data preprocessing and feature extraction:

[0162] First, the operation and maintenance team pre-processes the acquired operation data, vibration data, and noise data, including denoising, calibration and other operations, to ensure data quality.

[0163] Then, the sliding window method is used to extract features from these data. After analysis, the average interval between injection events of this reactor is 7 days, and the periodic changes of different types of data are as follows:

[0164] -Operation data (voltage, current, etc.): 1 day;

[0165] -Vibration data: 3 days;

[0166] - Noise data: 2 days;

[0167] Therefore, the operation and maintenance team set the length of the sliding window to 2 days, which is lower than the minimum interval of historical injection events and can also cover the large cycle of various data. Through the sliding window, 100 operation features were extracted from the operation data, 80 vibration features were extracted from the vibration data, and 60 noise features were extracted from the noise data.

[0168] The specific process of feature extraction is as follows:

[0169] For operating data, the extracted features include: time domain features such as mean value, standard deviation, peak factor, waveform factor, pulse factor, margin factor, etc.; frequency domain features such as frequency, active power, reactive power, power factor, etc.; and the changing trends of these features over time.

[0170] For vibration data, the extracted features include: time domain features such as root mean square value, peak value, margin factor, waveform factor; frequency domain features such as power spectrum density, frequency center, bandwidth; and time-frequency features such as wavelet energy.

[0171] For noise data, the extracted features include: time domain features such as mean, standard deviation, peak, margin factor; frequency domain features such as frequency, power spectral density, bandwidth; and time-frequency features such as MFCC coefficients.

[0172] 2. Feature Screening

[0173] Next, the operation and maintenance team used the Bayesian network model to calculate the correlation between these 100 operating characteristics, 80 vibration characteristics, and 60 noise characteristics and the historical injection times.

[0174] For the operation characteristics, the ones with higher correlation are: average voltage, standard deviation voltage, margin factor voltage, active power, power factor, etc. This shows that these characteristics are strongly related to the injection event.

[0175] For vibration characteristics, the main ones with high correlation are: RMS value, peak value, power spectrum density, frequency center, etc. This shows that there is a certain connection between the change of mechanical state of the equipment and the injection failure.

[0176] For noise features, the ones with high correlation are: mean, standard deviation, power spectrum density, MFCC coefficient, etc. This shows that changes in environmental noise may also reflect abnormal conditions of the equipment.

[0177] Through this feature correlation analysis, the operation and maintenance team screened out 40 typical operation characteristics, 30 typical vibration characteristics and 25 typical noise characteristics as input for subsequent modeling.

[0178] 3. Fault prediction model training

[0179] Based on the typical features extracted above, the operation and maintenance team built a multi-branch neural network model for injection fault prediction. The specific structure of the model is as follows:

[0180] (1) Operational characteristic injection correlation branch network

[0181] - Input: 40 typical operation characteristics;

[0182] - Output: correlation between operating characteristics and injection events;

[0183] -Network structure: 4-layer fully connected neural network, the number of hidden layer nodes is 80, 60, 40, and 20 respectively;

[0184] (2) Vibration characteristic injection correlation branch network

[0185] - Input: 30 typical vibration characteristics;

[0186] - Output: Correlation between vibration characteristics and injection events;

[0187] -Network structure: 3-layer convolutional neural network, 3 input channels, 3x3 convolution kernel size, 2x2 pooling kernel size;

[0188] (3) Noise characteristic injection correlation branch network

[0189] - Input: 25 typical noise features;

[0190] - Output: correlation between noise characteristics and injection events;

[0191] -Network structure: 2-layer recurrent neural network, 50 hidden layer units;

[0192] (4) Fusion feature injection correlation branch network

[0193] - Input: the output results of the above three branch networks;

[0194] -Output: correlation between fusion features and injection events;

[0195] -Network structure: The attention mechanism is used to fuse the outputs of the 3 branch networks;

[0196] (5) Timing Prediction Branch Network

[0197] - Input: 12 historical injection times;

[0198] - Output: next possible injection time;

[0199] -Network structure: long short-term memory network, number of hidden layer units 100;

[0200] (6) Feature Correlation Branch Network

[0201] -Input: all features;

[0202] -Output: correlation matrix between features;

[0203] -Network structure: autoencoder network, the number of hidden layer nodes is half of the number of features;

[0204] (7) Anomaly Detection Branch Network

[0205] -Input: real-time feature data;

[0206] -Output: anomaly score;

[0207] -Network structure: Isolation Forest Algorithm;

[0208] (8) Aggregate branch network

[0209] -Input: the output results of the above branch networks;

[0210] -Output: Final injection prediction time;

[0211] -Network structure: 3-layer fully connected network, with 80, 50, and 20 hidden layer nodes;

[0212] By training this multi-branch neural network model, the operation and maintenance team can fully explore the complex relationship between different types of features and injection events and improve the accuracy of predictions.

[0213] 4. Model verification and application

[0214] After the model was trained, the operation and maintenance team used a new oil-immersed reactor to be monitored for testing and verification. The operating data, vibration data, and noise data of the reactor are shown in Tables 1 to 3 below:

[0215] Table 1 - Row Data

[0216]

[0217] Table 2 - Vibration data

[0218] time Vibration value (m / s^2) 2024-08-01 00:00:00 0.15 2024-08-01 00:10:00 0.16 2024-08-01 00:20:00 0.17 ... ...

[0219] Table 3 - Noise data

[0220] time Noise level (dB) 2024-08-01 00:00:00 55 2024-08-01 00:10:00 56 2024-08-01 00:20:00 57 ... ...

[0221] Based on the above steps, the operation and maintenance team extracted the operating characteristics, vibration characteristics, and noise characteristics of the reactor, and input them into the trained multi-branch neural network model for prediction.

[0222] The output of the model shows that the reactor may have a fuel injection failure at 14:30 on August 15, 2024. Combined with the calculation of the preset threshold, the difference between the current time and the predicted time is 7 hours, which is less than the preset threshold of 24 hours. Therefore, the reactor is judged to be at high risk.

[0223] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting the risk of oil injection failure of an oil-immersed reactor, characterized in that: The following steps are involved: S10, acquiring historical operation data, vibration data, and noise data of multiple oil-immersed reactors with oil injection faults during operation, and recording the oil injection timing of the oil-immersed reactors; S20, preprocessing the operation data, vibration data and noise data, and extracting features using a sliding window to obtain an operation feature group, a vibration feature group and a noise feature group; S30, using a Bayesian network model to respectively calculate the correlation between each operating feature, each vibration feature, each noise feature in the operating feature group, the vibration feature group, and the noise feature group and the injection timing; S40, clustering the operation feature group, the vibration feature group, and the noise feature group respectively according to the correlation to obtain a typical operation feature group, a typical vibration feature group, and a typical noise feature group; S50, establishing and training an injection fault prediction model based on the typical operation feature group, the typical vibration feature group and the typical noise feature group; S60, continuously acquiring operation data, vibration data and noise data of the oil-immersed reactor to be detected, and extracting an operation feature group, a vibration feature group and a noise feature group of the oil-immersed reactor to be detected according to a sliding window of a preset length, and recording them as an operation feature group to be tested, a vibration feature group to be tested and a noise feature group to be tested; S70, inputting the operation feature group to be measured, the vibration feature group to be measured and the noise feature group to be measured into the injection fault prediction model, and outputting the injection prediction time; S80: Determine the failure risk of the oil-immersed reactor to be detected according to the predicted injection time, and output it.

2. The method for predicting the risk of oil injection failure of an oil-immersed reactor according to claim 1, characterized in that: The operating data include: voltage, current, frequency, power factor, active power, reactive power, apparent power, and oil temperature.

3. The method for predicting the risk of oil injection failure of an oil-immersed reactor according to claim 1, characterized in that: The vibration data is a vibration signal collected by a vibration sensor arranged on the oil-immersed reactor; and the noise data is an audio signal collected by an audio sensor arranged around the oil-immersed reactor.

4. The method for predicting the risk of oil injection failure of an oil-immersed reactor according to claim 1, characterized in that: The preset length of the sliding window is obtained based on the minimum interval of the injection time and the large cycle of each type of data, wherein the large cycle refers to the minimum time required for this type of data to fully fluctuate once. The step of obtaining the sliding window length specifically includes: obtaining the minimum interval of the injection time and the large cycle of each type of data, and taking the minimum value therebetween as the sliding window length.

5. The method for predicting the risk of oil injection failure of an oil-immersed reactor according to claim 4, characterized in that: The complete fluctuation refers to the process in which data passes from one extreme point through another extreme point and then returns to the original extreme point.

6. The method for predicting the risk of oil injection failure of an oil-immersed reactor according to claim 1, characterized in that: The injection fault prediction model adopts a multi-branch neural network structure, including an operation feature injection association branch network, a vibration feature injection association branch network, a noise feature injection association branch network, a fusion feature injection association branch network, a timing prediction branch network, a feature correlation branch network, an anomaly detection branch network and a summary branch network.

7. A method for predicting oil injection failure risk of an oil-immersed reactor according to claim 6, characterized in that: The operation feature injection correlation branch network is used to analyze the correlation between the operation feature and the injection event, the input is a typical operation feature group, the output is the operation feature injection correlation, and the structure is a multi-layer perceptron; The vibration feature injection correlation branch network is used to analyze the correlation between vibration features and injection events, the input is a typical vibration feature group, the output is the vibration feature injection correlation, and the structure is a convolutional neural network; The noise feature injection correlation branch network is used to analyze the correlation between noise features and injection events, the input is a typical noise feature group, the output is the noise feature injection correlation, and the structure is a recurrent neural network; The fusion feature injection correlation branch network is used to comprehensively analyze the correlation between multiple features and injection events, the input is a combination of a typical operation feature group, a vibration feature group and a noise feature group, the output is the fusion feature injection correlation, and the structure is an attention mechanism network; The time series prediction branch network is used to predict the time series of injection events, the input is the historical injection time data, the output is the next possible injection time, and the structure is a long short-term memory network; The feature correlation branch network is used to analyze the correlation between different features, the input is all feature groups, the output is the correlation matrix between features, and the structure is an autoencoder network; The anomaly detection branch network is used to detect abnormal patterns in features, the input is real-time feature data, the output is anomaly scores, and the structure is an isolation forest algorithm; The summary branch network is used to integrate the outputs of all branch networks, the input is the output results of each branch network, the output is the final injection prediction time, and the structure is a fully connected network.

8. The method for predicting the risk of oil injection failure of an oil-immersed reactor according to claim 7, characterized in that: Each branch network also includes an aggregation module for extracting macro features and a refinement module for extracting micro features. The macro features refer to features that reflect overall trends, and the micro features refer to features that reflect local changes. The threshold for distinguishing between macro features and micro features is specifically determined based on the time scale and impact range of the features.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the method for predicting the risk of oil injection failure of an oil-immersed reactor according to any one of claims 1 to 8.

10. A system for predicting the risk of oil injection failure of an oil-immersed reactor, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Images