Nuclear power equipment state monitoring method based on multi-modal fusion and deep learning

Abstract

The application relates to the technical field of nuclear power equipment monitoring, in particular to a nuclear power equipment state monitoring method based on multi-modal fusion and deep learning, which comprises the following steps: collecting multiple monitoring samples of multiple nuclear power units within the same period; obtaining prediction samples of the monitoring samples; obtaining sample information value degrees of the monitoring samples by coordinating the value differences of various feature indexes of the monitoring samples, and then extracting the monitoring samples to screen various state features from the feature indexes; obtaining the cooperative prediction difference degrees of the monitoring samples; constructing a feature space by using the state features, and then obtaining density consistency coefficients of the monitoring samples; obtaining comprehensive scores of the monitoring samples, screening expansion samples used for expanding various state labels; and monitoring the state of the nuclear power equipment. The application aims to improve the sensitivity and recognition rate of a state decision model to rare state faults by screening expansion samples with high information value and matching the distribution characteristics of real samples.

Description

Technical Field

[0001] This application relates to the field of nuclear power equipment monitoring technology, specifically to a nuclear power equipment condition monitoring method based on multimodal fusion and deep learning. Background Technology

[0002] Nuclear power generation, with its advantages of high energy density, uninterrupted operation, high unit efficiency, and strong stability, has gradually evolved into a fundamental pillar energy source providing clean and low-carbon electricity in the power system. As the energy conversion equipment in a nuclear power unit, the condition of the steam turbine during operation not only affects the energy conversion efficiency but also relates to the overall operational safety of the nuclear power unit. Monitoring the condition of nuclear power steam turbines helps improve the production safety of nuclear power plants.

[0003] Nuclear power turbines operate in normal conditions for the vast majority of their lifecycle, resulting in a sparse number of fault condition samples. Furthermore, the varying probabilities of different fault types within nuclear power equipment can lead to biased state decisions, resulting in low recognition rates and high false alarm / false positive rates for rare fault conditions. Current technologies use oversampling to interpolate new samples in the feature space for monitoring minor fault types in nuclear power turbines. However, this method ignores the physical operating mechanisms of nuclear power equipment, easily generating unreasonable false samples. Moreover, the number of new samples generated by this method is still far lower than the number of normal samples, leading to insufficient accuracy in identifying nuclear power equipment faults. Summary of the Invention

[0004] In view of the above, it is necessary to provide a nuclear power equipment condition monitoring method based on multimodal fusion and deep learning. Compared with the traditional nuclear power equipment condition monitoring method based on multimodal fusion and deep learning, this method improves the sensitivity and recognition rate of the condition decision model to rare condition faults by screening expanded samples with high information value and matching the distribution characteristics of real samples.

[0005] The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning proposed in this application adopts the following technical solution:

[0006] One embodiment of this application provides a method for monitoring the condition of nuclear power equipment based on multimodal fusion and deep learning. The method includes the following steps:

[0007] Multiple monitoring samples were collected from multiple nuclear power units at the same time period, including vibration signals, temperature data, and audio signals; monitoring samples from different nuclear power units at the same time interval were used as collaborative monitoring samples.

[0008] Using a pre-trained collaborative prediction model, predicted samples for each monitoring sample are obtained; effective components are extracted from the vibration signal and each feature index is calculated, and each feature index is calculated for temperature data and audio signals respectively; by comparing the differences in the values ​​of each feature index among collaborative monitoring samples, the sample information value of each monitoring sample is obtained, and then monitoring samples are extracted to screen each state feature from the feature index; by comparing the values ​​of each monitoring sample and its predicted sample in each state feature, the collaborative prediction difference of each monitoring sample is obtained; a feature space is constructed using the state features, and by comparing the local density of all monitoring samples under each monitoring sample's state label in the feature space, and the local density of each monitoring sample's predicted sample in the feature space, the density consistency coefficient of each monitoring sample is obtained; then, combined with the sample information value and the collaborative prediction difference, the comprehensive score of each monitoring sample is obtained to screen expanded samples used to expand each state label.

[0009] The expanded monitoring sample was used to train the model to monitor the status of nuclear power equipment.

[0010] In one embodiment, the expression for the value of the sample information is:

[0011] In the formula, P represents the sample information value of a single monitoring sample; I represents the total number of characteristic indicators of a single monitoring sample. L1 represents the difference in the value of the i-th feature index between a single monitoring sample and its co-monitoring samples; L2 represents the same state label for a single monitoring sample and its co-monitoring samples; L3 represents the different state labels for a single monitoring sample and its co-monitoring samples. This indicates the preset value.

[0012] In one embodiment, the process of extracting monitoring samples is as follows:

[0013] The sample information value of all monitored samples is divided into a preset number of levels. The mean of the sample information value of all monitored samples in each level is used as the exponent of an exponential function with a preset value greater than 1 as the base. The proportion of the calculation result of the exponential function in each level in the calculation results of the exponential function in all levels is calculated.

[0014] The product of the stated percentage and the preset total extraction amount is rounded to the nearest integer, which is taken as the number of monitoring samples extracted from each level.

[0015] In one embodiment, the process of filtering each state feature is as follows:

[0016] The feature weights of each feature indicator are obtained by extracting monitoring samples from all levels.

[0017] The feature indicators of vibration signal, temperature data, and audio data are arranged in descending order according to their feature weights. From the feature indicators of vibration signal, temperature data, and audio data, a first preset number, a second preset number, and a third preset number of feature indicators are selected as the state features.

[0018] In one embodiment, the process of obtaining the collaborative prediction difference is as follows:

[0019] Calculate the difference in the values ​​of each state feature between each monitored sample and its predicted sample;

[0020] The collaborative prediction difference is obtained by measuring all the differences between each monitored sample and its predicted sample.

[0021] In one embodiment, the collaborative prediction dissimilarity is the average of all the dissimilarity quantities between each monitored sample and its predicted sample.

[0022] In one embodiment, the process of obtaining the density uniformity coefficient is as follows:

[0023] Calculate the arithmetic mean of the distances between each monitoring sample and multiple preset nearest neighbor monitoring samples under the same state label in the feature space;

[0024] The average of the arithmetic mean of all monitored samples under each state label is recorded as the local distance of each state label;

[0025] The number of monitoring samples in the statistical feature space centered on each monitoring sample under each state label and with the local distance as the radius; the average of the number of all monitoring samples under each state label is denoted as the average density of each state label;

[0026] The total number of monitored samples within the range centered on the predicted sample of each monitored sample under each state label and with the local distance as the radius in the statistical feature space;

[0027] The density consistency coefficient is obtained by comparing the average density with the total amount.

[0028] In one embodiment, the density uniformity coefficient is the ratio of the total amount to the average density.

[0029] In one embodiment, the process of screening augmented samples for augmenting each state label is as follows:

[0030] Calculate the difference between the preset percentage of the total number of monitored samples under the health status label and the total number of monitored samples under each status label. Sort all monitored samples under each status label other than the health status label in descending order of comprehensive score. Use the predicted samples of the monitoring samples with the aforementioned difference as the expanded samples for expanding each status label.

[0031] In one embodiment, during the process of screening the expanded samples for expanding each status label, if the difference is negative, all monitoring samples under each status label are arranged in ascending order of comprehensive score, and the monitoring samples with the opposite difference under each status label are removed.

[0032] This application has at least the following beneficial effects:

[0033] This application assesses the information value of each monitoring sample by coordinating the differences in the values ​​of various feature indicators among the monitoring samples, reflecting the degree of contribution of each monitoring sample to the state identification task. Furthermore, it extracts monitoring samples based on the information value of the monitoring samples and filters state features based on the extracted monitoring samples. This can effectively remove feature indicators that are irrelevant to the state of nuclear power equipment or have extremely weak correlation, which helps to reduce noise interference, makes the state decision model more focused on high-value features, and improves the accuracy of nuclear power equipment state identification.

[0034] Furthermore, by analyzing the differences in the values ​​of various state features between monitored samples and their predicted samples, the co-prediction difference degree is obtained to assess the reliability of the predicted samples and determine whether they can truly reflect the operating state of nuclear power equipment. By constructing a feature space and comparing local densities, a density consistency coefficient is obtained, which can screen out predicted samples that are consistent with or have a higher density than the actual monitored samples, helping to solve the problem of imbalance in the state label categories of nuclear power equipment. In addition, by combining information value, co-prediction difference degree, and density consistency coefficient, the comprehensive score of each monitored sample can be evaluated more comprehensively. Based on the comprehensive score, expanded samples are screened to ensure that the expanded samples not only have high information value but also match the distribution characteristics of the actual samples, avoiding the generation of unreasonable false samples and greatly increasing the number of rare fault samples of nuclear power equipment, thereby improving the sensitivity and recognition rate of the state decision model for rare state faults. Attached Figure Description

[0035] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1A flowchart illustrating the steps of the nuclear power equipment condition monitoring method based on multimodal fusion and deep learning provided in this application;

[0037] Figure 2 This is a schematic diagram of the BiLSTM collaborative prediction model based on the federated averaging algorithm.

[0038] Figure 3 A schematic diagram of the screening process to expand the sample. Detailed Implementation

[0039] In the description of the embodiments in this application, the words "exemplary," "or," and "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary," "or," and "for example" is intended to present the relevant concepts in a specific manner.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. It should be understood that, unless otherwise stated, " / " in this application means "or".

[0041] It should also be noted that the terms "first" and "second" in this application are used to distinguish similar objects, rather than to describe a specific order or sequence.

[0042] The following description, in conjunction with the accompanying drawings, details the specific scheme of the nuclear power equipment condition monitoring method based on multimodal fusion and deep learning provided in this application.

[0043] This application provides an embodiment of a nuclear power equipment condition monitoring method based on multimodal fusion and deep learning. Specifically, the method is described below. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps:

[0044] Step 1: Collect multiple monitoring samples from multiple nuclear power units within the same time period, including vibration signals, temperature data, and audio signals; monitoring samples from different nuclear power units within the same time interval are mutually coordinated monitoring samples; use a pre-trained coordinated prediction model to obtain the predicted samples for each monitoring sample.

[0045] As a key piece of equipment in the secondary loop system of a nuclear power plant, the nuclear power turbine uses saturated hot steam generated by the steam generator to drive the turbine rotor to rotate, which in turn drives the generator to produce electricity through a coupling. Each nuclear power unit contains one turbine. In this embodiment, the nuclear power turbine consists of one high-pressure cylinder and two low-pressure cylinders. These cylinders and the generator are arranged in series. On the main shaft, four bearing housings are arranged from the steam end to the excitation end. The second bearing housing is designated as turbine bearing housing #2.

[0046] Using turbine bearing housing #2 of nuclear power unit 1 and turbine bearing housing #2 of nuclear power unit 2 with the same technical solution and equipment model as target equipment, multimodal operation data of the target equipment during the same operating time period was obtained using a nuclear power plant-level real-time information monitoring system. The multimodal operation data specifically included vibration signals, temperature data, and audio signals. Multiple monitoring samples were obtained from the multimodal operation data, with each monitoring sample lasting 2 minutes. Detailed information of the monitoring samples is shown in Table 1. Since nuclear power unit 1 and nuclear power unit 2 have the same technical solution and equipment model, the monitoring samples of turbine bearing housing #2 of nuclear power unit 1 and turbine bearing housing #2 of nuclear power unit 2 within the same time interval are considered as collaborative monitoring samples. The 2-minute interval is merely one embodiment of this application; the implementer can set its specific value according to actual conditions, and this application does not impose any special restrictions. In this embodiment, a total of 8000 monitoring samples were collected. The number of monitoring samples can be set by the implementer according to actual conditions, and this application does not impose any special restrictions.

[0047] Each monitoring sample is manually assigned a status label, with the following specific label types: healthy, rotor unbalanced, rotor rubbing, rotor crack, and oil film oscillation. Rotor unbalanced, rotor rubbing, rotor crack, and oil film oscillation are considered as fault status labels.

[0048] Table 1. Detailed information on the monitored samples

[0049]

[0050] For the monitoring samples of Nuclear Power Unit 1, vibration signals from half of the monitoring samples under each status label are selected to form the vibration dataset of Nuclear Power Unit 1. Following the method for obtaining the vibration dataset of Nuclear Power Unit 1, the vibration dataset of Nuclear Power Unit 2 is obtained. The vibration signals in the vibration datasets of Nuclear Power Unit 1 and Nuclear Power Unit 2 are co-sample data.

[0051] Based on the vibration datasets of Nuclear Power Unit 1 and Nuclear Power Unit 2, a BiLSTM collaborative prediction model based on the federated averaging algorithm is trained. The training of the BiLSTM collaborative prediction model based on the federated averaging algorithm is a well-known technique and will not be elaborated upon in this application. A schematic diagram of the BiLSTM collaborative prediction model based on the federated averaging algorithm is shown below. Figure 2 As shown.

[0052] Vibration signals from each monitoring sample are used as vibration samples to be predicted. A trained BiLSTM collaborative prediction model based on the federated averaging algorithm is used to obtain vibration prediction samples for each monitoring sample. The vibration prediction samples are of the same length as the vibration samples to be predicted and have the same state labels. All temperature data from each monitoring sample are used to form temperature samples to be predicted. Following the same method used to obtain vibration prediction samples from each monitoring sample, temperature prediction samples are obtained for the temperature data in each monitoring sample. These temperature prediction samples are of the same length as the temperature samples to be predicted and have the same state labels. Since audio signals are highly random and generally unpredictable, they are not predicted.

[0053] Step 2: Extract the effective components from the vibration signal and calculate each characteristic index. Calculate each characteristic index for the temperature data and audio signal respectively.

[0054] For vibration signals in monitoring samples of nuclear power equipment, modal decomposition is performed on the vibration signals in each monitoring sample to obtain each modal component. During normal operation, the turbine of a nuclear power unit generates periodic pulse signals and is also affected by noise signals. The intensity of the periodic pulse signals during normal operation is greater than the intensity of the periodic pulse signals during early faults, while the noise signal intensity is less than the fault pulse intensity. To perform fault monitoring, the following processing is performed: the L-kurtosis of each modal component of the vibration signal in each monitoring sample is obtained; all modal components of the vibration signal in each monitoring sample are arranged according to the magnitude of the L-kurtosis; and the modal component located in the middle position is selected as the effective component of the vibration signal in each monitoring sample.

[0055] In this embodiment, the Variational Mode Decomposition (VMD) algorithm is used to perform mode decomposition on the vibration signal, wherein the penalty factor... The value is set to 2000, and the number of modal components is set to 3. Each modal component of the vibration signal is obtained. The VMD algorithm is a well-known technology and will not be described in detail in this application. As other implementation methods, based on the ability to perform modal decomposition on the vibration signal, the implementer may adopt other existing feasible technologies. This application does not impose any special restrictions.

[0056] Furthermore, the standard deviation, peak value, L-kurtosis, impulse factor, margin factor, and sample entropy of the effective components of the vibration signal in each monitoring sample are calculated and normalized respectively. The normalized indicators are then used as the characteristic indicators of the vibration signal.

[0057] For temperature data in the monitoring samples of nuclear power equipment, the maximum, minimum, mean, standard deviation, and coefficient of variation of the temperature data were calculated and normalized. These normalized values ​​were then used as characteristic indicators of the temperature data. For audio signals in the monitoring samples of nuclear power equipment, the root mean square value, zero-crossing rate, spectral skewness, spectral entropy, and spectral flatness of the audio signals were calculated and normalized. These normalized values ​​were then used as characteristic indicators of the audio signals. The characteristic indicators are shown in Table 2.

[0058] In this embodiment, the Min-Max normalization method is used to normalize each index. The Min-Max normalization method is a well-known technique and will not be described in detail here.

[0059] Table 2 Feature Indicators

[0060]

[0061] Step 3: By coordinating the differences in the values ​​of various characteristic indicators among the monitoring samples, the sample information value of each monitoring sample is obtained, and then the monitoring samples are extracted to screen each state feature from the characteristic indicators.

[0062] As nuclear power units become increasingly larger and more complex, the complexity of regenerator systems and turbines also increases, making monitoring sample data highly likely to exhibit strong noise characteristics. By using operational data from nuclear power equipment with the same technical solutions and models within the same time period as collaborative monitoring samples, a higher similarity between these samples indicates less uncertainty and lower random noise from external factors affecting the nuclear power equipment.

[0063] Based on the above analysis, the sample information value of each monitoring sample is obtained by measuring the differences in the values ​​of various characteristic indicators among the collaboratively monitored samples. The expression is as follows:

[0064] In the formula, P represents the sample information value of a single monitoring sample; I represents the total number of characteristic indicators of a single monitoring sample. L1 represents the difference in the value of the i-th feature index between a single monitoring sample and its co-monitoring samples; L2 represents the same state label for a single monitoring sample and its co-monitoring samples; L3 represents the different state labels for a single monitoring sample and its co-monitoring samples. This represents a preset value. In this embodiment, the value of I is 27, and the specific calculation formula is: 2×6+2×5+1×5=27.

[0065] In this embodiment, The purpose is to ensure that the sample information value under different labels falls within different numerical ranges. Since subsequent classification is based on sample information value, and the number of monitoring samples to be drawn from each level is determined by the proportion of the sample information value within each level to the total sample information value across all levels, and an exponential function is used to amplify the sample information value during this process, it is crucial to avoid drawing too few monitoring samples from levels with lower sample information value. The value of is 1. The value is preset by the user, and the implementer can set it according to the actual situation. This application does not impose any special restrictions.

[0066] In this embodiment, the difference between the values ​​of the feature indicators is the absolute value of the difference. As other implementation methods, based on the ability to measure the degree of difference between the values ​​of the feature indicators, the implementer may use other calculation methods, such as the square of the difference, the ratio, etc. This application does not impose any special restrictions.

[0067] It should be noted that: if the state label of the monitored sample is consistent with that of its co-monitored sample, and the values ​​of the feature indicators differ... The smaller the value, the higher the synergy of the nuclear power equipment during the same operating period. In this case, the less uncertainty and random noise are injected into the nuclear power equipment due to external factors, the higher the value of the sample information. The value range is [0,1]. Nuclear power equipment with the same technical solution and equipment model cannot be in the same state in all operating periods. If the labels of the monitoring sample and its co-monitoring sample are different, and the values ​​of the characteristic indicators differ, the synergy will be affected. The larger the value, the more likely the monitoring sample and its co-monitoring sample contain potential value information about different states of nuclear power equipment. This can effectively help the state decision model identify the differences between different states of nuclear power equipment. The greater the contribution to nuclear power equipment, the greater the value of the sample information. The value range is [1,2].

[0068] Furthermore, due to the high dimensionality of the feature indicators in the monitoring samples of nuclear power equipment, some of these indicators may be irrelevant or weakly correlated with the equipment's condition. These indicators can dilute the contribution of important indicators to condition monitoring, obscure the decision boundaries of condition labels, and reduce the accuracy of condition identification. Therefore, feature indicator selection is necessary for the high-dimensional monitoring samples of nuclear power equipment. Feature indicator selection includes two stages: monitoring sample extraction and weight evaluation. Specifically:

[0069] The sample information value of all monitored samples is divided into a preset number of levels. The mean of the sample information value of all monitored samples in each level is used as the exponent of an exponential function with a preset value greater than 1 as the base. The proportion of the calculation result of the exponential function in each level in the calculation results of the exponential function in all levels is calculated.

[0070] The product of the stated percentage and the preset total sampling amount is rounded to the nearest integer, which is used as the number of monitoring samples randomly selected from each level.

[0071] In this embodiment, the preset quantity is 4, that is, the sample information value of all monitored samples is divided into 4 levels. The ranges of these 4 levels are [0,0.5], (0.5,1], (1,1.5] and (1.5,2], respectively. The preset quantity is preset by humans and the implementer can set it according to the actual situation. This application does not impose any special restrictions.

[0072] In this embodiment, the preset total extraction amount is 200. The preset total extraction amount is preset by humans, and the implementer can set it according to the actual situation. This application does not impose any special restrictions.

[0073] In this embodiment, the expression for the number of monitoring samples randomly selected from the a-th level is:

[0074] In the formula, This represents the number of monitoring samples randomly selected from the a-th level; N represents the preset total sampling quantity; exp() represents an exponential function with the natural constant as the base; A represents the number of levels; represents the mean of the sample information value of all monitored samples within the a-th level; Round[] represents the rounding function. The preset value greater than 1 is a natural constant, and the preset value of a value greater than 1 is preset by the user. The implementer can set it according to the actual situation, and this application does not impose any special restrictions.

[0075] It should be noted that the information value of a sample reflects the degree of contribution of each monitoring sample to the state recognition task. By using an exponential function to give greater extraction weight to monitoring samples with high information value, the feature selection algorithm can learn the decision boundary more clearly and improve the accuracy of weight evaluation by biasing the extraction of monitoring samples with high contribution.

[0076] Furthermore, all monitoring samples extracted from all levels will be used to form a sample set for feature selection, and the feature weight of each feature indicator will be obtained based on the sample set.

[0077] In this embodiment, the sample set is used as the input of the ReliefF algorithm, and the feature weight of each feature index is output. The ReliefF algorithm is a well-known technology and will not be described in detail in this application. As other implementation methods, based on the ability to obtain the feature weight of each feature index, implementers may adopt other existing feasible technologies, and this application does not impose any special restrictions.

[0078] Furthermore, based on feature weights, K feature indicators are selected as the state characteristics of the nuclear power equipment. Specifically, the vibration signal has 2×6 feature indicators. All feature indicators are arranged in descending order according to their feature weights, and a first number is preset before selection. Each characteristic index serves as a state characteristic of the nuclear power equipment; there are 2×5 characteristic indices for temperature data. All characteristic indices are arranged in descending order according to their feature weights, with a second preset number selected beforehand. Each characteristic index serves as a state characteristic of the nuclear power equipment; the audio signal has a total of 1×5 characteristic indices, all of which are arranged in descending order according to their feature weights, with a third preset number selected beforehand. Several characteristic indicators serve as the state characteristics of nuclear power equipment. Among them, In this embodiment, , , The values ​​are 3, 1, and 1 respectively. , , The value is preset by the user, and the implementer can set it according to the actual situation. This application does not impose any special restrictions.

[0079] Step 4: By comparing the values ​​of each monitored sample and its predicted sample in each state feature, the collaborative prediction difference of each monitored sample is obtained; a feature space is constructed using the state features; by comparing the local density of all monitored samples under the state label of each monitored sample in the feature space, and the local density of the predicted sample of each monitored sample in the feature space, the density consistency coefficient of each monitored sample is obtained; and then, by combining the sample information value and the collaborative prediction difference, the comprehensive score of each monitored sample is obtained, so as to screen the expanded samples used to expand each state label.

[0080] Nuclear power turbines are in normal condition for most of their entire life cycle, and the number of samples of fault conditions is relatively small. Moreover, the probability of various faults in nuclear power equipment varies, which can easily lead to nuclear power equipment status decisions being biased towards the majority class. This results in a low recognition rate and a high rate of false alarms and missed reports for rare fault conditions of nuclear power turbines.

[0081] Nuclear power plants have high requirements for the security of nuclear equipment data. Data isolation exists between different nuclear power plants, and operational data cannot be directly shared. Historical maintenance data of older equipment is also difficult to utilize effectively due to differences in technical solutions and models. This application uses collaborative monitoring samples for collaborative predictive modeling, which makes it easier to capture the stability characteristics of nuclear power equipment and generate a large number of predictive samples with status labels corresponding to real operational data. However, due to model errors or poor reliability of predictive samples, they cannot be blindly used to expand the dataset.

[0082] Based on the above analysis, the number of monitoring samples for nuclear power equipment under each status label was counted, and a preset percentage of the number of monitoring samples under the health status label was used as the target value to eliminate the imbalance of status label categories. Each state label with a number of monitored samples less than or equal to the target value is denoted as a sub-category, and each state label with a number of monitored samples greater than the target value is denoted as a major category.

[0083] In this embodiment, the percentage is 50%. The value of the percentage is preset by the user and the implementer can set it according to the actual situation. This application does not impose any special restrictions.

[0084] For each subcategory, since the vibration signal and temperature data in each monitoring sample were predicted in step 1, K-1 state features can be obtained based on the predicted samples of each monitoring sample in each subcategory. By comparing the values ​​of each state feature between each monitoring sample and its predicted sample, the co-prediction difference of each monitoring sample can be obtained. The specific process is as follows:

[0085] Calculate the difference in the values ​​of each state feature between each monitored sample and its predicted sample; take the average of all the differences between each monitored sample and its predicted sample as the co-prediction difference of each monitored sample.

[0086] In this embodiment, the expression for the collaborative prediction difference of each monitoring sample is:

[0087] In the formula, This represents the collaborative prediction dissimilarity of monitored sample c in the x-th subclass; K represents the total number of state features; This indicates that the monitored sample c in the x-th subclass is in the x-th subclass. The values ​​of each state feature; This indicates that the predicted sample of monitored sample c in the x-th subclass is in the x-th subclass. The values ​​of each state feature; This represents the absolute value operation. The difference between the values ​​of the state feature is... As an alternative implementation, based on the ability to measure the degree of difference between the values ​​of state characteristics, the implementer may use other calculation methods, such as the square of the difference, ratio, etc., and this application does not impose any special restrictions.

[0088] It should be noted that the smaller the difference between the values ​​of the state features, the higher the fidelity of the predicted sample of the monitoring sample and the more consistent it is with the actual operating characteristics of nuclear power equipment. Therefore, the smaller the difference in the collaborative prediction, the more important it is to use the predicted sample to expand the samples of each subclass, balance the number of label categories, and improve the generalization ability of the state decision model.

[0089] Furthermore, a feature space is constructed using state features. Since no audio signal prediction was performed in step 1, the state features corresponding to the audio prediction samples are currently lacking, so the feature space is K-1 dimensional. Taking the x-th subclass and its monitoring sample c as an example, the distance between monitoring sample c in the x-th subclass and each of the other monitoring samples in the feature space is calculated. The nearest neighbor monitoring samples belonging to the same subclass as monitoring sample c are found in the feature space, and the average distance between monitoring sample c and all its nearest neighbor monitoring samples is denoted as the nearest neighbor distance of monitoring sample c. The distance between monitoring samples is the Euclidean distance.

[0090] In this embodiment, the remaining monitoring samples in the xth subclass are arranged in ascending order according to their distance from monitoring sample c, and the first 10 monitoring samples are taken as the nearest neighbor monitoring samples of monitoring sample c. Here, 10 is only one embodiment of this application, and the implementer can set it according to the actual situation. This application does not impose any special restrictions.

[0091] Following the method for obtaining the nearest neighbor distance of monitoring sample c, the nearest neighbor distance of each monitoring sample is obtained.

[0092] Furthermore, the mean of the nearest neighbor distances of all monitored samples in the x-th subclass is denoted as the local distance of the x-th subclass. The number of monitored samples within a radius of the local distance, centered on each monitored sample in the x-th subclass, is counted in the feature space. The average number of monitored samples corresponding to all monitored samples in the x-th subclass is denoted as the average density of the x-th subclass. The total number of monitored samples within a radius of the local distance, centered on the predicted samples of each monitored sample in the x-th subclass, is used as the expanded density of each monitored sample in the x-th subclass. The ratio of the expanded density of each monitored sample in the x-th subclass to the average density of the x-th subclass is used as the density consistency coefficient of each monitored sample in the x-th subclass, which measures the quality of the predicted samples of each monitored sample in the x-th subclass as expanded samples. By filtering the predicted samples of the monitoring samples using the density consistency coefficient, the predicted samples that are consistent with the distribution of the monitoring samples in the xth subclass or have a higher density are selected as the expanded samples. This solves the problem of imbalance in the nuclear power equipment status label categories, avoids the introduction of noise or expanded samples that deviate from the distribution of the monitoring samples, and improves the sensitivity and recognition rate of the status decision model for rare status faults.

[0093] Meanwhile, the mean value of the audio signal in the state feature of all monitoring samples within the range of each monitoring sample in the xth subclass as the center and the local distance as the radius is used as the value of the audio prediction sample in the state feature of each monitoring sample in the xth subclass, thereby supplementing the state feature corresponding to the audio prediction sample of the audio signal.

[0094] Furthermore, the sample information value, collaborative prediction difference, and density consistency coefficient of all monitoring samples in the x-th subclass are used as inputs to the Topsis (Technique for Order Preference by Similarity to IdealSolution) algorithm, which outputs a comprehensive score for each monitoring sample in the x-th subclass. Here, the sample information value and density consistency coefficient are maximally large indicators, while the collaborative prediction difference is a minimally small indicator. The higher the calculated comprehensive score, the more suitable the predicted samples of the nuclear power equipment monitoring samples are as supplementary samples. The Topsis algorithm is a well-known technique and will not be described in detail here.

[0095] The comprehensive score of each monitoring sample in each sub-category is obtained according to the method for obtaining the comprehensive score of each monitoring sample in the xth sub-category.

[0096] target value Total number of monitored samples in each subcategory The difference between the total number of monitoring samples and the target value is recorded as the first difference. All monitoring samples in each sub-category are sorted in descending order of their comprehensive scores. The predicted samples from the first difference's monitoring samples are used as the expanded samples for each sub-category. The difference between the total number of monitoring samples in each major category and the target value is recorded as the second difference. All monitoring samples in each major category are sorted in ascending order of their comprehensive scores. The first two difference's monitoring samples are removed. A schematic diagram of the expanded sample selection process is shown below. Figure 3 As shown.

[0097] It should be noted that the monitoring sample expansion method based on the collaborative prediction model can only expand the sample to a maximum extent. If a monitoring sample exists In special cases where the number of monitoring samples in each subclass is extremely small, the monitoring sample expansion method based on the collaborative prediction model is insufficient to eliminate class imbalance. Therefore, this application uses the predicted samples of all monitoring samples in each subclass as the expansion samples for the first expansion. After the first expansion, the SMOTE (Synthetic Minority Over-sampling Technique) algorithm is used for the second sample expansion until the number of monitoring samples in each subclass reaches the target value. The SMOTE algorithm is a well-known technique and will not be described further in this application.

[0098] Step 5: Use the expanded monitoring samples to train the model to monitor the status of nuclear power equipment.

[0099] The state labels are converted into numerical form for easier classification. Remaining monitoring samples and extended samples under each state label of the nuclear power equipment are obtained. The K-dimensional state features of the monitoring samples or extended samples, together with the label numbers, constitute a K+1-dimensional state decision vector, constructing a state decision dataset. XGBoost ensemble learning is used as the training model, and the model parameters are optimized using the state decision dataset to construct a state decision model for nuclear power equipment. The state features of the nuclear power equipment within 2 minutes prior to the current moment are obtained and used as input to the state decision model, outputting the state label of the nuclear power equipment. During training, the learning rate is set to 0.5, the number of decision trees is set to 1000, colsample_bytree is set to 0.5, and the loss function is softmax. Implementers can set the specific values ​​of the learning rate, the number of decision trees, and colsample_bytree according to actual conditions; this application does not impose special restrictions.

[0100] In summary, this application assesses the information value of each monitoring sample by analyzing the differences in the values ​​of various feature indicators among collaborative monitoring samples, reflecting the contribution of each monitoring sample to the state identification task. Furthermore, it extracts monitoring samples based on their information value and filters state features based on the extracted monitoring samples. This effectively removes feature indicators that are irrelevant to or have very weak correlation with the state of nuclear power equipment, helps reduce noise interference, allows the state decision model to focus more on high-value features, and improves the accuracy of nuclear power equipment state identification.

[0101] Furthermore, by analyzing the differences in the values ​​of various state features between monitored samples and their predicted samples, the co-prediction difference degree is obtained to assess the reliability of the predicted samples and determine whether they can truly reflect the operating state of nuclear power equipment. By constructing a feature space and comparing local densities, a density consistency coefficient is obtained, which can screen out predicted samples that are consistent with or have a higher density than the actual monitored samples, helping to solve the problem of imbalance in the state label categories of nuclear power equipment. In addition, by combining information value, co-prediction difference degree, and density consistency coefficient, the comprehensive score of each monitored sample can be evaluated more comprehensively. Based on the comprehensive score, expanded samples are screened to ensure that the expanded samples not only have high information value but also match the distribution characteristics of the actual samples, avoiding the generation of unreasonable false samples and greatly increasing the number of rare fault samples of nuclear power equipment, thereby improving the sensitivity and recognition rate of the state decision model for rare state faults.

[0102] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0103] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from its essential characteristics. Therefore, the embodiments described above should be considered exemplary and non-limiting in all respects.

Claims

1. A method for monitoring the condition of nuclear power equipment based on multimodal fusion and deep learning, characterized in that, The method includes the following steps: Multiple monitoring samples were collected from multiple nuclear power units at the same time period, including vibration signals, temperature data, and audio signals; the monitoring samples from different nuclear power units at the same time interval were used as collaborative monitoring samples. Using a pre-trained collaborative prediction model, predicted samples for each monitoring sample are obtained; effective components are extracted from the vibration signal and each feature index is calculated, and each feature index is calculated for temperature data and audio signals respectively; by comparing the differences in the values ​​of each feature index among collaborative monitoring samples, the sample information value of each monitoring sample is obtained, and then monitoring samples are extracted to screen each state feature from the feature index; by comparing the values ​​of each monitoring sample and its predicted sample in each state feature, the collaborative prediction difference of each monitoring sample is obtained; a feature space is constructed using the state features, and by comparing the local density of all monitoring samples under each monitoring sample's state label in the feature space, and the local density of each monitoring sample's predicted sample in the feature space, the density consistency coefficient of each monitoring sample is obtained; then, combined with the sample information value and the collaborative prediction difference, the comprehensive score of each monitoring sample is obtained to screen expanded samples used to expand each state label. The expanded monitoring sample was used to train the model for monitoring the status of nuclear power equipment; The expression for the value of the sample information is: In the formula, P represents the sample information value of a single monitoring sample; I represents the total number of characteristic indicators of a single monitoring sample. L1 represents the difference in the value of the i-th feature index between a single monitoring sample and its co-monitoring samples; L2 represents the same state label for a single monitoring sample and its co-monitoring samples; L3 represents the different state labels for a single monitoring sample and its co-monitoring samples. Indicates the preset value; The process of obtaining the density uniformity coefficient is as follows: Calculate the arithmetic mean of the distances between each monitoring sample and multiple preset nearest neighbor monitoring samples under the same state label in the feature space; The average of the arithmetic mean of all monitored samples under each state label is recorded as the local distance of each state label; The number of monitoring samples in the statistical feature space centered on each monitoring sample under each state label and with the local distance as the radius; the average of the number of all monitoring samples under each state label is denoted as the average density of each state label; The total number of monitored samples within the range centered on the predicted sample of each monitored sample under each state label and with the local distance as the radius in the statistical feature space; The density uniformity coefficient is the ratio of the total amount to the average density.

2. The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning as described in claim 1, characterized in that, The process of extracting monitoring samples is as follows: The sample information value of all monitored samples is divided into a preset number of levels. The mean of the sample information value of all monitored samples in each level is used as the exponent of an exponential function with a preset value greater than 1 as the base. The proportion of the calculation result of the exponential function in each level in the calculation results of the exponential function in all levels is calculated. The product of the stated percentage and the preset total extraction amount is rounded to the nearest integer, which is taken as the number of monitoring samples extracted from each level.

3. The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning as described in claim 2, characterized in that, The process of filtering each state feature is as follows: The feature weights of each feature indicator are obtained by extracting monitoring samples from all levels. The feature indicators of vibration signal, temperature data, and audio data are arranged in descending order according to their feature weights. From the feature indicators of vibration signal, temperature data, and audio data, a first preset number, a second preset number, and a third preset number of feature indicators are selected as the state features.

4. The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning as described in claim 1, characterized in that, The process for obtaining the collaborative prediction difference is as follows: Calculate the difference in the values ​​of each state feature between each monitored sample and its predicted sample; The collaborative prediction difference is obtained by measuring all the differences between each monitored sample and its predicted sample.

5. The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning as described in claim 4, characterized in that, The collaborative prediction difference is the average of all the differences between each monitored sample and its predicted sample.

6. The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning as described in claim 1, characterized in that, The process of screening augmented samples to augment each state label is as follows: Calculate the difference between the preset percentage of the total number of monitored samples under the health status label and the total number of monitored samples under each status label. Sort all monitored samples under each status label other than the health status label in descending order of comprehensive score. Use the predicted samples of the monitoring samples with the aforementioned difference as the expanded samples for expanding each status label.

7. The nuclear power equipment condition monitoring method based on multimodal fusion and deep learning as described in claim 6, characterized in that, In the process of screening the expanded samples used to expand each status label, if the difference is negative, all monitoring samples under each status label are sorted in ascending order of comprehensive score, and the monitoring samples with the opposite difference under each status label are removed.

Images