Reactor state evaluation method and system based on convolutional neural network and acoustic feature variable prediction

Abstract

The application discloses a kind of based on convolutional neural network and acoustic characteristic variable prediction's reactor state evaluation method and system, the method includes: step S1: obtaining reactor box wall vibration signal and noise signal as data;Step S2: data is eliminated and completed using outlier and continuous interpolation method;Step S3: the acoustic vibration characteristic prediction model of dry-type reactor based on convolutional neural network is constructed;Step S4: evaluation index is constructed, state evaluation is completed, and finally the reactor state evaluation result based on convolutional neural network and acoustic characteristic variable prediction is obtained.The system is used to implement the above method.The application has the advantages of simple principle, high degree of intelligence, high evaluation precision, etc.

Description

Technical Field

[0001] This invention mainly relates to, and specifically relates to, a reactor state evaluation method and system based on convolutional neural networks and acoustic feature variable prediction. Background Technology

[0002] Dry-type air-core reactors are characterized by their simple structure, stable inductance, and ease of operation and maintenance. They are commonly used for current limiting, reactive power compensation, and improving line voltage distribution. During normal operation, the electrodynamic forces of a dry-type air-core reactor tend to compress the reactor, with the vibration velocity at the encapsulated ends being greater than that in the middle. In addition to electrical and thermal stresses, they also bear continuous alternating electromagnetic forces, thus remaining in a state of constant vibration. Under the combined action of multiple stresses, the vibration will continue to intensify. In severe cases, overheating and discharge defects are highly likely to occur inside the dry-type air-core reactor, making it extremely prone to inter-turn short-circuit faults. Once an inter-turn short circuit occurs in a dry-type air-core reactor, due to the huge induced current in the short-circuit turn, and influenced by the short-circuit current and the distorted magnetic field near the short-circuit turn, the electrodynamic forces on (and near) the short-circuit turn will increase sharply, leading to an increase in the vibration velocity near the short-circuit turn. The short-circuit current will cause a sharp rise in the internal and surface temperature. As a result, there have been many accidents in recent years involving reactor damage, malfunctions, and even fires. Therefore, traditional dry-type air-core reactors actually pose serious safety hazards during operation, which necessitates the research of monitoring methods for the safety of dry-type air-core reactors.

[0003] Based on the vibration and noise of power equipment, defect detection and analysis methods often fall into three categories: analytical model-based analysis, signal processing-based analysis, and artificial intelligence-based analysis. Among these, the analytical model-based method is the earliest and most systematically studied fault diagnosis method. This method requires establishing a mathematical model of the research object, and then using mathematical methods to process the information characteristics of the research object. Therefore, it has strong theoretical support and inherent sensitivity to faults that have not yet occurred, based on the theoretical foundation and system essence of the research object. However, this method has certain disadvantages: it is usually difficult to completely replace the complex system model by eliminating modeling errors and environmental interference in the mathematical model; at the same time, its robustness problem is becoming increasingly significant. Considering the complex structure of reactors, it is clear that the analytical model-based method cannot be used to predict and evaluate the characteristics of reactors under actual operating conditions. Summary of the Invention

[0004] The technical problem to be solved by this invention is: in view of the technical problems existing in the prior art, this invention provides a reactor state evaluation method and system based on convolutional neural network and acoustic feature variable prediction, which is simple in principle, highly intelligent, and highly accurate in evaluation.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A reactor state assessment method based on convolutional neural networks and acoustic feature variable prediction, comprising: Step S1: Acquire the vibration signal and noise signal of the reactor tank wall as data; Step S2: Use outlier detection and continuous interpolation to remove and complete the data; Step S3: Construct a prediction model for the acoustic and vibration characteristics of dry reactors based on convolutional neural networks; Step S4: Construct evaluation indicators, complete the state evaluation, and finally obtain the reactor state evaluation results based on convolutional neural network and acoustic feature variable prediction.

[0006] As a further improvement to the method of the present invention: step S2 includes: Step S201: Extract the vibration signal sequence X and noise signal sequence M from the box wall to obtain the signal matrix, as shown in the following formula: Wherein, data column X represents the vibration signal of the box wall, data column M represents the noise signal, and n represents the length of the collected data; Step S202: Calculate the mean and standard deviation of each column in the signal matrix, as shown in the following formula: , in, and This represents the mean and standard deviation of the vibration signal. and This represents the mean and standard deviation of the noise signal; Step S203: Using the mean and standard deviation as a new sequence, calculate the Euclidean distance between each data point within the sequence, as shown in the following formula: Where k ranges from 1 to n; the first row represents the mean and standard deviation of the vibration signal as a new sequence, and the second row represents the mean and standard deviation of the noise signal as a new sequence; Step S204: Calculate the k-th distance of data point x: Sort the distances of this point to other points in ascending order, and the k-th distance is the k-th distance; Step S205: k-distance neighborhood: All points within this neighborhood whose distance to point x is less than or equal to the k-th distance, denoted as N(x,k); if the distance is greater, it is an outlier. Step S206: If the data is removed as an outlier, it is supplemented using continuous interpolation, which replaces and completes the data by using the average of the two data points before and after the removed data.

[0007] As a further improvement to the method of the present invention: step S3 includes: Step S301: Vibration signal data acquisition and processing; Step S302: Construct and initialize the convolutional neural network model; initialize the number of iterations Nmax, kernel size, weights of fully connected layers, and weights of attention mechanism of the convolutional neural network model; Step S303: Model training and parameter tuning; throughout the training process, continuously save the parameters of the minimum loss model on the validation set until the training ends. The final minimum loss model is the optimal network model.

[0008] As a further improvement to the method of the present invention: the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1; the training set is used to train the model fitting parameters, and the validation set is used to adjust the hyperparameters of the network model.

[0009] As a further improvement to the method of the present invention: the convolutional neural network model includes: Convolutional layer: Uses multiple convolutional kernels to perform convolution operations on the input data and processes the output after convolution; the convolution operation formula is as follows: in This represents the i-th input data. The parameters of the k-th convolutional kernel in this convolutional layer are... For the corresponding bias parameter, g k Extract features of the i-th sample for the k-th convolutional kernel; Max pooling layer: It consists of downsampling and upsampling layers; after convolution, it yields many different feature maps; the expression for max pooling is: in, This represents the output of the i-th neuron in the current layer. It is a downsampling function, which is a function that takes the maximum value within a certain range. It refers to the range of pooling. It is the step size of the pooling; Fully connected layer: Features are input to the fully connected layer after passing through convolutional and pooling layers; a Softmax classifier is used in the fully connected layer; the formula is expressed as: in, l Indicates the number of layers in the network. This represents the output of the fully connected layer. This represents the input to the fully connected layer. Indicates the weighting coefficient. Indicates bias.

[0010] As a further improvement to the method of the present invention: In step S303 above, the training model uses the minimum average absolute error of the signal amplitude as the loss function to predict and determine whether the model has converged; if the number of model iterations reaches the maximum number of iterations, the training ends and the prediction result is output.

[0011] As a further improvement to the method of the present invention: step S3 further includes step S304: model evaluation; that is, loading the optimal network model, inputting test set data to evaluate the diagnostic effect of the model, calculating the loss and accuracy on the entire test set, recording the results of several experiments and taking the average value as the final result to achieve feature prediction.

[0012] As a further improvement to the method of the present invention: the evaluation index set in step S4 includes one or more of the following three evaluation indexes: mean absolute error, mean absolute percentage error, and root mean square error, namely: in, These represent the predicted value and the ideal value, respectively. The three indicators range from [0, +]. The smaller the values ​​of the three indicators, the higher the model's prediction accuracy and the more ideal the prediction effect.

[0013] As a further improvement to the method of the present invention: in step S1, when acquiring the vibration signal and noise signal of the reactor tank wall, a multi-channel acquisition detection device is used, with each measuring point dispersed, and the acoustic vibration signal of different parts of the tank is acquired remotely.

[0014] The present invention further provides a reactor state evaluation system based on convolutional neural networks and acoustic feature variable prediction, comprising: The acquisition unit is used to acquire vibration and noise signals from the reactor tank wall as data. The data processing unit is used to remove and complete data using outliers and continuous interpolation. The model building unit is used to build a predictive model of acoustic and vibration characteristics of dry reactors based on convolutional neural networks. The evaluation unit is used to construct evaluation indicators, complete the state evaluation, and finally obtain the reactor state evaluation results based on convolutional neural networks and acoustic feature variable predictions.

[0015] Compared with the prior art, the advantages of the present invention are as follows: 1. The reactor condition assessment method and system based on convolutional neural networks and acoustic feature variable prediction of the present invention is simple in principle, highly intelligent, and highly accurate in assessment, enabling reactor condition assessment based on convolutional neural networks and acoustic feature variable prediction. This invention is mainly applicable to condition detection and defect diagnosis of dry-type parallel reactors, and to a certain extent overcomes the difficulty in timely and accurate detection and diagnosis of latent mechanical defects inside dry-type parallel reactors, thereby ensuring the safe and stable operation of the power grid.

[0016] 2. The reactor condition assessment method and system based on convolutional neural networks and acoustic feature variable prediction of the present invention first acquires the vibration signal and noise signal of the reactor tank wall, thereby obtaining a continuous sequence of feature parameters over a period of time. Then, outlier detection is used to remove abnormal values ​​from the acquired data, and continuous interpolation is used to complete the removed data. Furthermore, an acoustic vibration feature prediction model for dry-type reactors based on convolutional neural networks is constructed. The model is designed, and three evaluation indicators—mean absolute error, mean absolute percentage error, and root mean square error—are constructed to evaluate the model, thereby achieving condition prediction and assessment. The vibration feature assessment method for dry-type reactors of the present invention, which can adapt to the current operating environment and service life, is of great significance. At the same time, the method and system also have a solid engineering background and broad application prospects. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the reactor state evaluation method based on convolutional neural networks and acoustic feature variable prediction according to the present invention. Detailed Implementation

[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, the reactor state evaluation method based on convolutional neural networks and acoustic feature variable prediction of the present invention includes the following steps: Step S1: Acquire the vibration signal and noise signal of the reactor tank wall as data; In the process of acquiring vibration and noise signals from the reactor tank wall, we should acquire as much vibration and noise as possible from the reactor body.

[0021] Step S2: Use outlier detection and continuous interpolation to remove and complete the data; Step S3: Construct a prediction model for the acoustic and vibration characteristics of dry reactors based on convolutional neural networks; Step S4: Construct evaluation indicators, complete the state evaluation, and finally obtain the reactor state evaluation results based on convolutional neural network and acoustic feature variable prediction.

[0022] In specific application examples, when acquiring vibration and noise signals from the reactor tank wall, the on-site detection device can use multi-channel acquisition, with the measurement points spread out as much as possible, and further adopt remote acquisition methods to acquire acoustic vibration signals from different parts of the tank.

[0023] In specific application examples, step S2 may include the following process depending on actual needs: Step S201: Extract the vibration signal sequence X and noise signal sequence M from the box wall to obtain the signal matrix, as shown in the following formula: In this data column, X represents the vibration signal of the box wall, M represents the noise signal, and n represents the length of the collected data.

[0024] Step S202: Calculate the mean and standard deviation of each column in the signal matrix, as shown in the following formula: , in, and This represents the mean and standard deviation of the vibration signal. and This represents the mean and standard deviation of the noise signal.

[0025] Step S203: Using the mean and standard deviation as a new sequence, calculate the Euclidean distance between each data point within the sequence, as shown in the following formula: Where k ranges from 1 to n; the first row represents the mean and standard deviation of the vibration signal as a new sequence, and the second row represents the mean and standard deviation of the noise signal as a new sequence.

[0026] Step S204: Calculate the k-th distance of data point x: Sort the distances of this point to other points in ascending order, and the k-th distance is the k-th distance.

[0027] Step S205: k-distance neighborhood: The distance from all points in this neighborhood to point x is less than or equal to the k-th distance, denoted as N(x,k); if it is greater, it is an outlier.

[0028] Step S206: If the data is removed as an outlier, it is supplemented using continuous interpolation, which replaces and completes the data by using the average of the two data points before and after the removed data.

[0029] In specific application examples, step S3 may include the following process depending on actual needs: Step S301: Vibration signal data acquisition and processing; That is, input the data after processing in steps S1 and S2 above.

[0030] Step S302: Build and initialize the model; Initialize the convolutional neural network model with network parameters such as the number of iterations Nmax, kernel size, weights of fully connected layers, and weights of attention mechanisms.

[0031] Furthermore, as a preferred embodiment, the present invention divides the dataset into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used to train the model's fitting parameters, and the validation set is used to adjust the hyperparameters of the network model.

[0032] Step S303: Model training and parameter tuning; Throughout the training process, the parameters of the minimum loss model on the validation set are continuously saved until the training ends. The final minimum loss model is the optimal network model.

[0033] The specific structure of the convolutional neural network model established in step S3 above is as follows: Convolutional layers: These layers use multiple convolutional kernels to perform convolution operations on the input data and process the output. Convolution with different kernels can output a feature map, which represents different features extracted from the input signal.

[0034] The formula for convolution is as follows: in This represents the i-th input data. The parameters of the k-th convolutional kernel in this convolutional layer are... For the corresponding bias parameter, g k Extract features of the i-th sample for the k-th convolutional kernel.

[0035] Max pooling layer: Pooling layers, sometimes called sampling layers, are mainly divided into downsampling layers and upsampling layers. After convolution operations, many different feature maps are obtained. The expression for max pooling is: in, This represents the output of the i-th neuron in the current layer. It is a downsampling function, which is a function that takes the maximum value within a certain range. It refers to the range of pooling. It is the step size of pooling.

[0036] Fully connected layer: Features are input to the fully connected layer after passing through convolutional and pooling layers. The most commonly used classifier in a fully connected layer is the Softmax classifier. The formula can be expressed as: in, l Indicates the number of layers in the network. This represents the output of the fully connected layer. This represents the input to the fully connected layer. Indicates the weighting coefficient. Indicates bias.

[0037] Furthermore, in step S303 above, the training model uses the minimum mean absolute error of the signal amplitude as the loss function for prediction, and determines whether the model has converged. If the model reaches the maximum number of iterations, the training ends and the prediction result is output.

[0038] In a preferred embodiment, step S3 of the present invention further includes step S304: model evaluation; that is, loading the optimal network model, inputting test set data to evaluate the diagnostic effect of the model, calculating the loss and accuracy on the entire test set, recording the results of ten experiments and taking the average value as the final result to achieve feature prediction.

[0039] In specific application examples, the main evaluation indicators set in step S4 according to actual needs include one or more of the following three evaluation indicators: mean absolute error, mean absolute percentage error, and root mean square error. in, These represent the predicted value and the ideal value, respectively. The three indicators range from [0, +]. The smaller the values ​​of the three indicators, the higher the model's prediction accuracy and the more ideal the prediction effect.

[0040] The present invention further provides a reactor state evaluation system based on convolutional neural networks and acoustic feature variable prediction, for implementing the above method, the system comprising: The acquisition unit is used to acquire vibration and noise signals from the reactor tank wall as data. The data processing unit is used to remove and complete data using outliers and continuous interpolation. The model building unit is used to build a predictive model of acoustic and vibration characteristics of dry reactors based on convolutional neural networks. The evaluation unit is used to construct evaluation indicators, complete the state evaluation, and finally obtain the reactor state evaluation results based on convolutional neural networks and acoustic feature variable predictions.

[0041] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A reactor state assessment method based on convolutional neural networks and acoustic feature variable prediction, characterized in that, include: Step S1: Acquire the vibration signal and noise signal of the reactor tank wall as data; Step S2: Use outlier detection and continuous interpolation to remove and complete the data; Step S3: Construct a prediction model for the acoustic and vibration characteristics of dry reactors based on convolutional neural networks; Step S4: Construct evaluation indicators, complete the state evaluation, and finally obtain the reactor state evaluation results based on convolutional neural network and acoustic feature variable prediction; Step S2 includes: Step S201: Extract the vibration signal sequence X and noise signal sequence M from the box wall to obtain the signal matrix, as shown in the following formula: Wherein, data column X represents the vibration signal of the box wall, data column M represents the noise signal, and n represents the length of the collected data; Step S202: Calculate the mean and standard deviation of each column in the signal matrix, as shown in the following formula: ， in, and This represents the mean and standard deviation of the vibration signal. and This represents the mean and standard deviation of the noise signal; Step S203: Using the mean and standard deviation as a new sequence, calculate the Euclidean distance between each data point within the sequence, as shown in the following formula: Where k ranges from 1 to n; the first row represents the mean and standard deviation of the vibration signal as a new sequence, and the second row represents the mean and standard deviation of the noise signal as a new sequence; Step S204: Calculate the k-th distance of data point x: Sort the distances of this point to other points in ascending order, and the k-th distance is the k-th distance; Step S205: k-distance neighborhood: All points within this neighborhood whose distance to point x is less than or equal to the k-th distance, denoted as N(x,k); if the distance is greater, it is an outlier. Step S206: If the data is removed as an outlier, it is supplemented using continuous interpolation, which replaces and completes the data by using the average of the two data points before and after the removed data.

2. The reactor state evaluation method based on convolutional neural network and acoustic feature variable prediction according to claim 1, characterized in that, Step S3 includes: Step S301: Vibration signal data acquisition and processing; Step S302: Construct and initialize the convolutional neural network model; initialize the number of iterations Nmax, kernel size, weights of fully connected layers, and weights of attention mechanism of the convolutional neural network model; Step S303: Model training and parameter tuning; throughout the training process, continuously save the parameters of the minimum loss model on the validation set until the training ends. The final minimum loss model is the optimal network model.

3. The reactor state evaluation method based on convolutional neural network and acoustic feature variable prediction according to claim 2, characterized in that, The dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio; the training set is used to train the model fitting parameters, and the validation set is used to adjust the hyperparameters of the network model.

4. The reactor state evaluation method based on convolutional neural network and acoustic feature variable prediction according to claim 2, characterized in that, The convolutional neural network model includes: Convolutional layer: Uses multiple convolutional kernels to perform convolution operations on the input data and processes the output after convolution; the convolution operation formula is as follows: in This represents the i-th input data. The parameters of the k-th convolutional kernel in this convolutional layer are... For the corresponding bias parameter, g k Extract features of the i-th sample for the k-th convolutional kernel; Max pooling layer: It consists of downsampling and upsampling layers; after convolution, it yields many different feature maps; the expression for max pooling is: in, This represents the output of the i-th neuron in the current layer. It is a downsampling function, which is a function that takes the maximum value within a certain range. It refers to the range of pooling. It is the step size of the pooling; Fully connected layer: Features are input to the fully connected layer after passing through convolutional and pooling layers; a Softmax classifier is used in the fully connected layer; the formula is expressed as: in, l Indicates the number of layers in the network. This represents the output of the fully connected layer. This represents the input to the fully connected layer. Indicates the weighting coefficient. Indicates bias.

5. The reactor state evaluation method based on convolutional neural network and acoustic feature variable prediction according to claim 2, characterized in that, In step S303 above, the training model uses the minimum average absolute error of the signal amplitude as the loss function to make predictions and determines whether the model has converged; if the number of model iterations reaches the maximum number of iterations, the training ends and the prediction result is output.

6. The reactor state evaluation method based on convolutional neural network and acoustic feature variable prediction according to claim 2, characterized in that, Step S3 also includes step S304: model evaluation; that is, loading the optimal network model, inputting test set data to evaluate the diagnostic effect of the model, calculating the loss and accuracy on the entire test set, recording the results of several experiments and taking the average value as the final result to achieve feature prediction.

7. The reactor state assessment method based on convolutional neural network and acoustic feature variable prediction according to any one of claims 1-6, characterized in that, The evaluation indicators set in step S4 include one or more of the following three evaluation indicators: mean absolute error, mean absolute percentage error, and root mean square error: in, These represent the predicted value and the ideal value, respectively. The three indicators range from [0, +]. The smaller the values ​​of the three indicators, the higher the model's prediction accuracy and the more ideal the prediction effect.

8. The reactor state assessment method based on convolutional neural networks and acoustic feature variable prediction according to any one of claims 1-6, characterized in that, In step S1, when acquiring the vibration and noise signals of the reactor tank wall, a multi-channel acquisition detection device is used, with each measuring point dispersed, and the acoustic vibration signals of different parts of the tank are acquired remotely.

9. A system for reactor condition assessment based on the convolutional neural network and acoustic feature variable prediction as described in any one of claims 1 to 8, characterized in that, include: The acquisition unit is used to acquire vibration and noise signals from the reactor tank wall as data. The data processing unit is used to remove and complete data using outliers and continuous interpolation. The model building unit is used to build a predictive model of acoustic and vibration characteristics of dry reactors based on convolutional neural networks. The evaluation unit is used to construct evaluation indicators, complete the state evaluation, and finally obtain the reactor state evaluation results based on convolutional neural networks and acoustic feature variable predictions.

Images