CN117633444A

CN117633444A - Sensor abnormal intelligent recovery method and device based on association network

Info

Publication number: CN117633444A
Application number: CN202311513562.1A
Authority: CN
Inventors: 周宏宽; 孙原理; 宋志浩; 柯志武; 刘佩; 郭晓杰; 王晨阳; 孙衢骎; 陶模; 郑伟; 冯毅; 柴文婷; 陈朝旭; 李献领; 赵振兴
Original assignee: 719th Research Institute Of China State Shipbuilding Corp
Current assignee: 719th Research Institute Of China State Shipbuilding Corp
Priority date: 2023-11-14
Filing date: 2023-11-14
Publication date: 2024-03-01

Abstract

The application provides a sensor abnormality intelligent recovery method and device based on a correlation network, and relates to the technical field of intelligent monitoring. The method comprises the following steps: acquiring fault data of a sensor with higher correlation with key thermal parameters under various working conditions; inputting the fault data into a deep learning network model, and obtaining recovered sensor data output by the deep learning network model; the deep learning network model is formed by a convolutional self-encoder and a long and short term memory network stack. The sensor abnormal intelligent recovery method and device based on the association network can accurately judge system and equipment faults, ensure normal operation of a control system and improve system safety and economy.

Description

Sensor abnormal intelligent recovery method and device based on association network

Technical Field

The application relates to the technical field of intelligent monitoring, in particular to a sensor abnormality intelligent recovery method and device based on a correlation network.

Background

The pace of industrial intelligent development is increasingly advanced, however, as the precision degree and the complexity degree of intelligent equipment are improved, and industrial equipment is subjected to vibration and impact for a long time, the problems of part abrasion, part aging and the like are easily caused, and the industrial equipment is further caused to be faulty.

During operation of the power system, when operating parameters characterizing the system and equipment are abnormal, they may be caused by either a malfunction of the system and equipment or by a malfunction of the sensors. The sensor measurement data is the basis of system and equipment state monitoring, the validity and accuracy of the measurement value of the sensor are required to be ensured, the sensor of the power system is applied to high temperature, high pressure, high humidity, high corrosion or high radioactivity, and the sensor can be aged or failed to different degrees along with the increase of the running time of the power system, so that the sensor data drift, the accuracy decline and even complete failure are caused.

The abnormality and the missing of the thermal parameter measurement data caused by the sensor failure are unfavorable for judging the faults of the system and the equipment, and the normal operation of the control system is difficult to ensure.

Disclosure of Invention

The application provides a sensor abnormal intelligent recovery method and device based on a correlation network, which can accurately judge system and equipment faults by reconstructing parameters of an abnormal sensor, ensure normal operation of a control system and improve system safety and economy.

In a first aspect, the present application provides a method for intelligently recovering abnormal sensors based on an association network, including:

Acquiring fault data of a sensor with higher correlation with key thermal parameters under various working conditions;

inputting the fault data into a deep learning network model, and obtaining recovered sensor data output by the deep learning network model; the deep learning network model is formed by a convolutional self-encoder and a long and short term memory network stack.

In one embodiment, the deep learning network model comprises an input layer, a plurality of convolution layers, a pooling layer, an up-sampling layer, a plurality of long and short memory network layers, a full connection layer and an output layer, wherein the layers are connected layer by layer;

the convolution layer and the pooling layer are combined to be used as a convolution self-encoder coding layer, and feature extraction is carried out on data;

the convolution layer is combined with the up-sampling layer to serve as a convolution self-encoder decoding layer, and original data is restored based on data characteristics;

and outputting the data to the long-short-time memory network layer through an activation function after the data decoding operation.

In one embodiment, an ELU function is employed as the activation function for the fully-connected layer.

In one embodiment, the method further comprises:

and the sensor which has higher correlation with the key thermodynamic parameters under various working conditions is determined by carrying out correlation analysis on the steady-state working condition normal operation and fault operation data of the full-range simulator acquisition and storage study object.

In one embodiment, the method for determining the sensor with higher correlation with the key thermal parameters under various working conditions by performing correlation analysis on the collected and stored steady-state working condition normal operation and fault operation data of the research object by the full-range simulator comprises the following steps:

acquiring and storing steady-state working condition normal operation and fault operation data of a research object based on a full-range simulator to determine a plurality of groups of data matrixes; the number of rows of the data matrix is the number of samples, and the number of columns is the number of data features;

determining the corresponding column vectors of the key thermal parameter sensors under different working conditions by calculating a data correlation matrix;

selecting a sensor with correlation under multiple working conditions as a reconstruction reference to be selected based on a preset correlation threshold;

and carrying out correlation coefficient significance analysis on the to-be-selected reconstruction reference, and determining the sensor with higher correlation with the key thermal parameters under various working conditions.

In one embodiment, the method further comprises:

collecting training data by a sliding window method;

and training the deep learning network model by utilizing the training data.

In one embodiment, the error of the reconstructed data of the deep learning network model is used as a fitness function, and a particle swarm algorithm is adopted to determine the super-parameters of the deep learning network model.

In a second aspect, an embodiment of the present application provides an intelligent sensor anomaly recovery device based on an association network, including:

the acquisition module is used for acquiring fault data of the sensor which has higher correlation with the key thermal parameters under various working conditions;

the output module is used for inputting the fault data into a deep learning network model and obtaining the recovered sensor data output by the deep learning network model; the deep learning network model is formed by a convolutional self-encoder and a long and short term memory network stack.

In a third aspect, the present application also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of the first aspect when executing the program.

In a fourth aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method according to the first aspect.

In a fifth aspect, the present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the method of the first aspect.

According to the sensor abnormal intelligent recovery method and device based on the association network, the convolutional self-encoder and the long-short-term memory network are stacked to form the deep learning network model, the reconstruction and reproduction of abnormal sensor data are carried out by combining other sensor parameters in a data driving mode, the speed and accuracy of sensor data recovery can be improved, the influence on a system caused by instrument detection errors can be reduced or even eliminated, the system and equipment faults are accurately judged, normal operation of a control system is guaranteed, and the safety and economical efficiency of the system are improved.

Drawings

For a clearer description of the present application or of the prior art, the drawings that are used in the description of the embodiments or of the prior art will be briefly described, it being apparent that the drawings in the description below are some embodiments of the present application, and that other drawings may be obtained from these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of a sensor anomaly intelligent restoration method based on an association network provided by the application;

FIG. 2 is a schematic diagram of specific steps of a sensor abnormal intelligent recovery method based on a correlation network;

FIG. 3 is a schematic diagram of a correlation model-based reconstruction reference selection process provided in the present application;

FIG. 4 is a schematic flow chart of a particle swarm algorithm provided in the present application;

FIG. 5 is a schematic structural diagram of the sensor anomaly intelligent restoration device based on the association network;

fig. 6 is a schematic structural diagram of an electronic device provided in the present application.

Detailed Description

For the purposes of making the objects, technical solutions and advantages of the present application more apparent, the technical solutions in the present application will be clearly and completely described below with reference to the drawings in the present application, and it is apparent that the described embodiments are some, but not all, embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.

Fig. 1 is a schematic flow chart of a method for intelligently recovering abnormal sensors based on an association network according to an embodiment of the present application. Referring to fig. 1, an embodiment of the present application provides a method for intelligently recovering abnormal sensor based on an association network, where an execution subject may be an electronic device, for example, may be a controller, and the method may include:

Step 110, acquiring fault data of a sensor with higher correlation with key thermal parameters under various working conditions;

step 120, inputting fault data into a deep learning network model, and obtaining recovered sensor data output by the deep learning network model; the deep learning network model is formed by stacking convolutional self-encoders (Convolutional Auto-Encode, CAE) and Long Short-Term Memory networks (LSTM).

According to the method, abnormal faults of sensor measurement information in the running process of the power system are researched, different types of sensors of the thermal system are taken as objects, and measurement information reconstruction is carried out on the abnormal sensors which are positioned rapidly. In step 110, the controller may screen and obtain fault data of the sensor having a high correlation with the key thermal parameters under various conditions.

In step 120, the controller may input the acquired fault data into the deep learning network model, and then acquire recovered sensor data output by the deep learning network model. Wherein the deep learning network model is formed by stacking a convolutional self-encoder and a long-short-term memory network. And based on a reconstruction model of the self-encoder and the long-term and short-term memory, reading measurement parameters with high correlation coefficients with the failure target in a certain time, and then reconstructing the measurement information of the failure sensor containing time and space relations.

In the case of no hardware redundancy, the simulation modeling is removed from the system object, and the reference for reconstructing the signal data of the failed sensor is usually the relevant sensor parameter at the same time or is inferred based on the history data of the failed sensor when in normal use. However, the system equipment is complex, and when the working condition changes, the correlation changes among the sensor data, so that the absolute safety is difficult to ensure only by the reconstruction of the historical operation data or other sensor data at the same time. If the correlation between the historical data and each sensor is utilized at the same time, the reliability of the reconstructed data can be improved to a great extent.

According to the method, a long-term memory network (LSTM) is combined with a convolution self-encoder (CAE) method to perform model training and reconstruction, the LSTM model has good processing capacity for time series data, compared with the reasoning of a traditional neural network model and a filtering method, the LSTM can effectively utilize historical operation data of a reconstructed object and related sensors thereof, long-term memory of the operation state of a failure object is achieved, and more accurate virtual values are reconstructed; and the CAE method is combined to reduce the dimension and noise of the training data, so that the overall training speed of the model is effectively improved.

Fig. 2 is a schematic diagram of specific steps of a method for intelligent recovery of sensor abnormality based on an association network according to an embodiment of the present application, which may be compared with fig. 1.

According to the sensor abnormal intelligent recovery method based on the association network, the convolutional self-encoder and the long-short-term memory network are stacked to form the deep learning network model, the reconstruction and the reproduction of abnormal sensor data are carried out by combining other sensor parameters in a data driving mode, the speed and the accuracy of the sensor data recovery can be improved, the influence on a system caused by instrument detection errors can be reduced or even eliminated, the system and equipment faults are accurately judged, the normal operation of a control system is guaranteed, and the safety and the economical efficiency of the system are improved.

In one embodiment, the deep learning network model comprises an input layer, a plurality of convolution layers, a pooling layer, an up-sampling layer, a plurality of long and short time memory network layers, a full connection layer and an output layer which are connected layer by layer;

The self encoder (AE) is an unsupervised neural network model, is a multi-layer feedforward neural network, has input and output nodes with the same size, has fewer nodes in the middle layer than the input and output nodes, and can play a role in compressing and recovering data. The calculation process of the model is divided into two parts, encoding (encod) and decoding (Decode). The input information is encoded to obtain an intermediate Code (Code) and then decoded to obtain an output similar to the original input. The network learns the network parameters by optimizing the loss of the original input and the final output, the loss function is as follows:

wherein L represents a loss function, x _i An ith sample input representing a loss function, y _i The i-th sample output representing the loss function, n representing the total number of samples, i representing the current sample.

The self-encoder, like other networks, is trained using a gradient descent method. The network contains a plurality of hidden layers, high-dimensional data is encoded and changed into lower-dimensional information, and the calculation result of the middle layer can be regarded as data characteristics to use.

There are many variations of the self-encoder method, in which convolutional self-encoders (CAE) are used to perform data encoding and decoding operations, thereby implementing data noise reduction and feature dimension reduction extraction.

Convolutional self-encoders employ convolution, pooling, deconvolution, anti-pooling, etc. among convolutional neural networks instead of fully-concatenated operations. Convolutional neural networks are a type of feed-forward neural network that, because of their shared parameters, can take advantage of image locality, and are widely used for image-dependent tasks. A convolutional neural network is a network that uses convolutional operations in the hidden layer to perform feature extraction on an image. The system mainly comprises basic units such as a convolution layer, a pooling layer, an activation layer and the like. The input image is processed by a plurality of hidden layers to become a multidimensional feature map, and the feature map can be used as tasks such as classification and the like, so that the robustness is good.

The method can establish a deep learning network model structure formed by stacking the convolution self-encoder and the long-short-time memory network under a TensorFlow framework, and utilize the encoding and decoding capabilities of the self-encoder to reduce noise and extract features of original data, and simultaneously exert the mode recognition advantages of the long-short-time memory network in time sequence data to reconstruct accurately. The model is formed by connecting an input layer (training data after pretreatment), a multi-layer convolution layer, a pooling layer, an up-sampling layer, a multi-layer long-short-time memory model layer, a full-connection layer and an output layer by layer.

Wherein the convolution layer is a process of generating a feature map by applying a filter to input data. The convolution operation is performed by sliding on the image or data matrix, and the matrix multiplication and addition operation can be completed by the convolution kernel at each position of the image or data, and finally, a feature map is generated. The convolution uses a sliding window mode to obtain local signal characteristics on the image, and compared with full connection, the number of parameters is greatly reduced. The formula is used as follows:

wherein,output feature layer for layer I, +.>For the previous layer output features +.>Weights of feature map corresponding to i position for j position convolution kernel, +.>For the channel-by-channel bias term, f is the activation function.

The feature matrix obtained through calculation of the convolution layer is not reduced in dimension, so that feature compression can be performed on the feature map for reducing calculation amount, and network calculation complexity is simplified. Pooling aggregates the local signature features into output features for that location in a sliding window fashion, using some parameter-free fashion. Common pooling methods include maximum pooling, average pooling, etc., i.e. using the maximum, average of the local signal as output, the pooling layer can provide translational invariance to the characteristic signal.

The up-sampling layer can amplify the feature matrix, i.e. insert new elements between the data by adopting a proper interpolation algorithm on the basis of the original matrix. Three methods are commonly used: bilinear interpolation, deconvolution, and anti-pooling.

The convolution layer and the pooling layer are combined to be used as a coding layer of the convolution self-coder, so that characteristic extraction can be carried out on data, one-dimensional convolution is adopted in convolution operation, and the problem of characteristic information loss caused by carrying out convolution operation on characteristic dimensions is avoided. Combining the convolutional layer with the upsampling layer as a convolutional self-encoder decoding layer, the original data can be restored based on the data characteristics. And outputting the data to the long-short-time memory network layer through an activation function after the data decoding operation. The LSTM can effectively utilize the history operation data of the reconstructed object and the related sensors thereof, thereby realizing long-term memory of the operation state of the invalid object and reconstructing a more accurate virtual value.

The method can establish a long and short time memory tuple model. The long and short time memory cell group model adds a 'cell state' which runs through the whole time sequence and a 'processor' which judges whether information is useful or not in each calculation process, and comprises an input gate (input gate), a forget gate (forget gate) and an output gate (output gate); the forgetting gate determines the content to be discarded in the information at the last moment through a Sigmoid function so as to obtain output ft; the input gate determines a part to be reserved in the information at the previous moment through the joint coordination of sigmoid and tanh functions so as to obtain an output it; and Ot represents the calculation result of the output gate, the output data Ct of the current moment is obtained after the tan function is activated, and then repeated calculation is carried out on each time step by using the push type.

According to the sensor abnormal intelligent recovery method based on the association network, a deep learning network model is built through the input layer, the multi-layer convolution layer, the pooling layer, the up-sampling layer, the multi-layer long-short-time memory network layer, the full-connection layer and the output layer, and recovery of sensor data is carried out, so that the speed and accuracy of sensor data recovery can be further improved, system and equipment faults can be accurately judged, normal operation of a control system is guaranteed, and system safety and economy are improved.

By adopting the ELU function as the model full-connection layer activation function, the nonlinear characteristics in the data can be more embodied on the basis of the ReLU activation function, and the problem of neuron death is avoided. The ELU function achieves higher accuracy and faster convergence speed than other ReLU function variants (Leaky ReLU, PReLU). Meanwhile, the model avoids the occurrence of the overfitting phenomenon by using a Dropout operation. Dropout operation is added between the LSTM layer and the full connection layer, and finally the basic structure of the long-short-time memory network of the whole self-encoder is obtained.

According to the sensor abnormal intelligent recovery method based on the association network, provided by the embodiment of the application, the ELU function is used as the model full-connection layer activation function, so that nonlinear characteristics in data can be embodied, the problem of neuron death is avoided, and the speed and accuracy of sensor data recovery can be improved.

In one embodiment, the sensor anomaly intelligent restoration method based on the association network further comprises:

The controller can acquire and store the normal operation and fault operation data of the steady-state working condition of the research object through the full-range simulator to perform correlation analysis, and can determine the sensor with higher correlation with the key thermal parameters under various working conditions. The correlation analysis method may be selected from Pearson (Pearson) correlation coefficient analysis method, spearman (Spearman) correlation coefficient analysis method, etc.

According to the sensor abnormal intelligent recovery method based on the association network, association analysis is carried out on steady-state working condition normal operation and fault operation data of a full-range simulator acquisition and storage study object, so that the sensor with higher correlation with key thermal parameters under various working conditions is determined, and the sensor with higher correlation with the key thermal parameters under various working conditions can provide reference for abnormal sensor measurement information reconstruction, so that accuracy of sensor data recovery can be further improved.

In one embodiment, by performing correlation analysis on steady-state operating condition normal operation and fault operating data of a full-range simulator acquisition and storage study object, a sensor with higher correlation with key thermal parameters under various operating conditions is determined, and the sensor comprises:

In statistics, the degree of correlation is usually measured by a correlation coefficient, and common correlation calculation methods are Pearson and Spearman correlation coefficient analysis methods.

The Pearson correlation coefficient is used for measuring the degree of correlation between two variables X and Y, and is proposed by British scholars, pearson, reflects the linear relation and the correlation direction between the variables, and has a value between-1 and 1, and represents the complete positive linear correlation of the two variables when the value is 1, -represents the complete negative linear correlation of the two variables when the value is 1, and represents the wireless correlation between the variables when the value is 0.

The Pearson correlation coefficient is defined as the product of the covariance of the two variables divided by their standard deviation, and is calculated as follows:

ρ _x，y ＝cov(X，Y)/σ _X σ _Y ＝E[(X-μx)(Y-μy)]/σ _X σ _Y

wherein ρ is _x，y Representing Pearson correlation coefficients, cov representing covariance, X, Y representing variables, σ _X Represents standard deviation, sigma, of the corresponding X variable _Y Represents the standard deviation of the corresponding Y variable, E represents the standard deviation, μx represents the X variable expectation, and μy represents the Y variable expectation.

It should be noted that the Pearson correlation coefficient is suitable for the case that the standard deviation of the variables is not zero, and the two variables are normally distributed and linearly continuous data, and the two variable observations are paired and independent from each other.

In addition, the Pearson correlation coefficient is not capable of representing the intensity of correlation in the case of curve correlation and more complex conditions; before calculation, certain processing is needed to be carried out on the extreme value of the data, so that errors caused by individual numbers are avoided.

The Pearson correlation coefficient is widely applied to correlation analysis between two variables, but has higher data requirements on the variables, and as the reactor operation data in fluctuation change, the system parameters can also rise or fall in a nonlinear way when faults are inserted, and the calculation requirements of the Pearson correlation coefficient cannot be met.

The Spearman correlation coefficient uses the rank order of two variables for linear correlation analysis, does not require the distribution of the original variables, belongs to a non-parameter statistical method, and has wider application range than the Pearson correlation coefficient. For variables that can be calculated using Pearson correlation coefficients, spearman correlation coefficients can also calculate correlation coefficients, but the statistical energy efficiency is lower than Pearson correlation coefficients.

Spearman correlation coefficients the correlation analysis results were obtained by ranking the data, calculating the difference in rank coefficients after ranking, and further processing. The method utilizes the position after parameter sequencing instead of the actual value of the parameters, so that errors caused by a normalization method due to large magnitude difference of the parameters are not considered, and even if the experimental value is suddenly changed, the obtained parameter is too large or too small, the arrangement of rank order is not influenced, and the influence on the result is small.

The Spearman rank correlation coefficient expression is:

wherein ρ is _s Representing Spearman rank correlation coefficient, R _i Representing two variables X and Y to be analyzed, i.e. X _i Or y _i After the rank order (both in ascending order and in descending order), n represents the number of parameter data.

ρ _s The value range is-1, the correlation of the two parameters is expressed by the absolute value, and the larger the absolute value is, the stronger the correlation is. ρ _s When=0, two elements are uncorrelated; ρ _s When=1 or-1, two elements are linearly related; ρ _s >A0 represents a positive correlation of two elements, whereas a negative correlation of two elements is represented.

According to the correlation network model, the Spearman correlation coefficient is used as a judgment standard, and the sensors which have higher correlation with the key thermal parameters under various working conditions are screened to be used as reconstruction references by calculating correlations of the operation parameters under normal working conditions and fault working conditions. As shown in fig. 3, the specific steps may include:

step 1: collecting and storing steady-state working condition normal operation and fault operation data of a research object by using a full-range simulator, wherein the fault is an insertion system fault, the data characteristic is n, the number of samples is m, and a plurality of groups of data matrixes of n multiplied by m are obtained;

step 2: the data matrix preprocessing, because the Spearman method does not require the data magnitude, the processing emphasis is on screening invalid data in the running process (the data does not change with time);

step 3: calculating a data correlation matrix, determining corresponding column vectors of key thermal parameter sensors under different working conditions, and establishing an analysis matrix;

step 4: taking a preset correlation threshold (for example, an absolute value of 0.3) as a judging standard of whether the correlation is related or not, and selecting a sensor with correlation under multiple working conditions as a reconstruction reference to be selected;

Step 5: carrying out correlation coefficient significance analysis on a to-be-selected reconstruction reference by combining a correlation matrix calculation result, wherein a sensor which does not pass the significance analysis shows that the correlation coefficient has no reference value, and determining whether to add reconstruction by adopting other methods;

step 6: and (3) reserving a sensor passing the significance test, determining a final reconstruction reference sensor, namely a sensor with higher correlation with key thermal parameters under various working conditions, and simultaneously determining final model input parameters by combining transient data.

According to the sensor abnormal intelligent recovery method based on the association network, the Spearman correlation coefficient is used as a judgment standard, the sensors with higher correlation with the key thermal parameters under various working conditions are screened to be used as reconstruction references by calculating correlations of the operating parameters under the normal working condition and the fault working condition in pairs, and the speed and the accuracy of sensor data recovery can be further improved.

collecting training data by a sliding window method;

training the deep learning network model by using the training data.

The step of collecting training data and training the deep learning network model may include:

Step 1: acquiring target working condition operation parameters by using a full-range simulator, primarily processing and classifying, and storing;

step 2: saving the reconstructed reference sensor data determined based on the analysis result of the association model, and eliminating other low-association-degree data;

step 3: and (3) carrying out standardized processing on the data stored in the step (2) to avoid the influence of overlarge or undersize data caused by inconsistent dimensions on the reconstruction training process. The data value may be projected between 0,1 using a maximum-minimum normalization method, and the transfer function may be selected as x= (x-min)/(max-min), where max is the sample data maximum and min is the sample data minimum.

Step 4: training data is collected by using a sliding window method, the running data interval time is 0.25s, the sliding time window length is set to be 2.5s, the two-dimensional data (N multiplied by D dimension) obtained in the step 3 is converted into three-dimensional stacked data blocks ((N-num_steps+1) multiplied by (num_steps multiplied by D)), wherein N is the total data amount, D is the characteristic parameter dimension, namely the number of sensors, overlap exists among the data in the sliding sampling process, and the final data length is (N-num_steps+1).

Step 5: training the deep learning network model by using the training data.

According to the sensor abnormal intelligent recovery method based on the association network, training data are collected through a sliding window method, and the training data are utilized to train the deep learning network model, so that the speed and the accuracy of the deep learning network model can be improved, and the speed and the accuracy of the sensor data recovery can be further improved.

In one embodiment, the error of the reconstructed data of the deep learning network model is used as a fitness function, and a particle swarm algorithm is used for determining the super-parameters of the deep learning network model.

As shown in fig. 4, the particle swarm optimization (Particle Swarm Optimization, PSO) algorithm treats the solution of the optimization problem as a bird in the search space, and abstracts it into particles without mass and volume, extending it into the N-dimensional space. The position of the particles in the N-dimensional space is represented as a vector, and all particles have a fitness, determined by the function to be optimized. Each particle flies in the search space at a speed that determines the direction and distance it flies, also represented by a vector. The particle records its current position and the best position found so far, taken as the flight experience of the particle itself. At the same time, each particle also knows the best location to find for all particles in the whole population so far. The particles dynamically adjust the self flight speed according to the self flight experience and the best flight experience of the companion, namely, each particle continuously corrects the self speed and the advancing direction through the self optimal value and the group optimal value in the statistical iteration process, thereby forming a positive feedback mechanism of group optimization.

Assuming that m particles constitute a population, the position of each particle can be expressed as an n-dimensional vector, where the position of the ith particle is expressed as:

x _i ＝(x _i1 ，x _i2 ，...，x _in )，i＝1，2，...，m

wherein x is _i Indicating the position of the ith particle, (x) _i1 ，x _i2 ，...，x _in ) Represents an n-dimensional position vector, i represents the ith particle, and m represents the total number of particles.

Will x _i Substituting the optimization function can obtain the adaptability of the particles, and the position of the particles can be measured according to the adaptability. The direction and distance of flight of each particle is determined by a velocity, which is also denoted as an n-dimensional vector.

v _i ＝(v _i1 ，v _i2 ，...，v)，i＝1，2，...，m

Wherein v is _i Represents the velocity of the ith particle, (v) _i1 ，v _i2 ,., v) represents an n-dimensional velocity vector, i represents the i-th particle, and m represents the total number of particles.

The particle may update its flight based on its current location, its best location, and the best location of the particle population. The particle update rate formula is as follows:

v(t+1)＝v(t)+c ₁ *r ₁ *(pbest-x(t))+c ₂ *c ₂ *(gbest-x(t))

wherein v (t+1) represents the velocity of the particle itself at the next moment, v (t) represents the current position of the particle itself, c ₁ 、c ₂ For the learning factor, r is a random number in the (0, 1) interval, pbest represents the best position of the particle itself, x (t) represents the current position of the particle itself, and gbest represents the best position of the particle population.

The formula for the particle update location is as follows:

x(t+1)＝x(t)+v(t+1)

Where x (t+1) represents the position of the particle itself at the next time, x (t) represents the current position of the particle itself, and v (t+1) represents the speed of the particle itself at the next time.

The execution of the method of the present embodiment may be based on the following steps:

step 1: the deep learning network model adopts a standard root mean square error loss function as a loss function, and an Adam algorithm is used as an optimizer. In the training process of the self-encoder long-short-term memory network, all data are split into a plurality of batches of training samples, and the processed data are randomly disturbed to reduce uncertainty and prevent overfitting. With the increase of the number of training rounds, the training error is gradually reduced, which indicates that the reconstruction value of the network model can be continuously approximate to the true value of the fault sensor when the encoder is memorized for a long time;

step 2: carding super parameters in a long-short-term memory network; the super parameters comprise the number of layers and the size of convolution kernels of a convolution layer, the step length of the convolution process, the size of a pooling layer, the size of an up-sampling layer, the number of stacked layers and the number of layer units of a long-time memory network layer, the number of layers of a full-connection layer, the number of neurons in each layer and the parameter proportion setting of Dropout operation;

step 3: determining a feasible solution domain of the super-parameters; training a model by adopting the loss function and parameter optimization method described in the step 1, taking the error of the model reconstruction data as an adaptability function, determining super parameters by adopting a particle swarm algorithm, and optimizing a deep learning network model;

Step 4: setting the number of particles and the maximum iteration times in a particle swarm optimization algorithm, randomly generating an initial position and an initial speed by combining a model hyper-parameter feasible solution domain, and calculating a local optimal value pBest and a global optimal value gBest;

step 5: and calculating the particle adaptive function value, wherein each particle compares the adaptive function value of the current state with the self historical local optimal solution, and updating pBest when the adaptive function value is superior to the historical local optimal value. Comparing the optimal pBest with the gBest when all the particles are calculated, and updating the gBest when the updating conditions are met;

step 6: updating the position and the speed of the particles according to the local optimum pBest and the global optimum gBest;

step 7: and (3) repeating the steps (13) and (14), ending the optimization of the round of particle swarm when the termination condition is met, and outputting the global optimal value gBestin of the round.

Step 8: if the maximum iteration number is not reached, re-initializing the particle swarm algorithm parameters, adding 1 to the iteration number, and repeating the steps 5 and 6; stopping the algorithm when the maximum iteration times are reached, and transmitting the optimal gBestin in the multi-round calculation results to the deep learning model.

Step 9: model verification and testing. The sensor fault data are generated through manual insertion, data with low association degree in the test data are removed, then standardization processing is carried out, the data values are projected between [0,1] by adopting a maximum and minimum standardization method, the consistency of the data format is ensured, the error is calculated after the reconstruction result and the water level true value are normalized, and the model is evaluated.

According to the sensor abnormal intelligent recovery method based on the association network, which is provided by the embodiment of the application, aiming at the conditions that the super parameters in the deep learning network model are relatively time-consuming, labor-consuming and poor in effect under the condition of manual selection and confirmation, the optimal super parameters are automatically obtained by utilizing the particle swarm optimization algorithm, so that the speed and the precision of the deep learning network model are ensured, and the speed and the precision of the sensor data recovery can be further improved.

The related network-based sensor abnormal intelligent restoration device provided by the application is described below, and the related network-based sensor abnormal intelligent restoration device described below and the related network-based sensor abnormal intelligent restoration method described above can be referred to correspondingly.

Fig. 5 is a schematic structural diagram of an abnormal sensor intelligent restoration device based on an association network according to an embodiment of the present application. Referring to fig. 5, an intelligent sensor abnormality recovery apparatus based on an association network provided in an embodiment of the present application may include:

the acquisition module 510 is used for acquiring fault data of the sensor which has higher correlation with the key thermal parameters under various working conditions;

the output module 520 is configured to input the fault data into a deep learning network model, and obtain recovered sensor data output by the deep learning network model; the deep learning network model is formed by a convolutional self-encoder and a long and short term memory network stack.

According to the sensor abnormal intelligent recovery device based on the association network, the convolutional self-encoder and the long-short-term memory network are stacked to form the deep learning network model, the reconstruction and reproduction of abnormal sensor data are carried out by combining other sensor parameters in a data driving mode, the speed and accuracy of the sensor data recovery can be improved, the influence on a system caused by instrument detection errors can be reduced or even eliminated, the system and equipment faults are accurately judged, the normal operation of a control system is guaranteed, and the safety and economical efficiency of the system are improved.

Specifically, the intelligent sensor abnormality recovery device based on the association network provided in the embodiment of the present application can implement all the method steps implemented by the method embodiment in which the execution body is a controller, and can achieve the same technical effects, and the same parts and beneficial effects as those of the method embodiment in the embodiment are not specifically described herein.

Fig. 6 illustrates a physical schematic diagram of an electronic device, as shown in fig. 6, which may include: processor 610, communication interface (Communications Interface) 620, memory 630, and communication bus 640, wherein processor 610, communication interface 620, and memory 630 communicate with each other via communication bus 640. The processor 610 may invoke logic instructions in the memory 630 to perform associated network-based sensor anomaly intelligent restoration methods, including, for example:

Further, the logic instructions in the memory 630 may be implemented in the form of software functional units and stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on such understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

In another aspect, the present application further provides a non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, is implemented to perform the steps of the association network-based sensor anomaly intelligent restoration method provided by the methods above, for example, including:

In yet another aspect, the present application further provides a computer program product, where the computer program product includes a computer program, where the computer program can be stored on a non-transitory computer readable storage medium, where the computer program, when executed by a processor, can perform the steps of the method for intelligent recovery of sensor anomalies based on an association network provided by the above methods, including, for example:

The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.

In addition, it should be noted that: the terms "first," "second," and the like in the embodiments of the present application are used for distinguishing between similar objects and not for describing a particular sequential or chronological order. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the application are capable of operation in sequences other than those illustrated or otherwise described herein, and that the terms "first" and "second" are generally intended to be used in a generic sense and not to limit the number of objects, for example, the first object may be one or more.

In the embodiment of the application, the term "and/or" describes the association relationship of the association objects, which means that three relationships may exist, for example, a and/or B may be represented: a exists alone, A and B exist together, and B exists alone. The character "/" generally indicates that the context-dependent object is an "or" relationship.

In the embodiment of the present application, "determining B based on a" means that a is considered when determining B. Not limited to "B can be determined based on A alone", it should also include: "B based on A and C", "B based on A, C and E", "C based on A, further B based on C", etc. Additionally, a may be included as a condition for determining B, for example, "when a satisfies a first condition, B is determined using a first method"; for another example, "when a satisfies the second condition, B" is determined, etc.; for another example, "when a satisfies the third condition, B" is determined based on the first parameter, and the like. Of course, a may be a condition in which a is a factor for determining B, for example, "when a satisfies the first condition, C is determined using the first method, and B is further determined based on C", or the like.

The term "plurality" in the embodiments of the present application means two or more, and other adjectives are similar thereto.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and are not limiting thereof; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.

Claims

1. The intelligent sensor abnormality recovery method based on the association network is characterized by comprising the following steps of:

2. The correlation network-based sensor anomaly intelligent restoration method according to claim 1, wherein the deep learning network model comprises an input layer, a multi-layer convolution layer, a pooling layer, an up-sampling layer, a multi-layer long-short-time memory network layer, a full-connection layer and an output layer, and all the layers are connected layer by layer;

3. The association network-based sensor anomaly intelligent restoration method according to claim 2, wherein an ELU function is adopted as an activation function of the full connection layer.

4. The association network-based sensor anomaly intelligent restoration method of claim 2, further comprising:

5. The method for intelligently recovering abnormal sensor based on the association network according to claim 4, wherein the step of determining the sensor having higher correlation with the key thermal parameters under various conditions by performing association analysis on the steady-state working condition normal operation and fault operation data of the full-range simulator acquired and stored by the full-range simulator comprises the following steps:

6. The association network-based sensor anomaly intelligent restoration method of claim 5, further comprising:

collecting training data by a sliding window method;

and training the deep learning network model by utilizing the training data.

7. The intelligent recovery method of sensor abnormality based on the association network according to claim 6, wherein the error of the reconstructed data of the deep learning network model is used as a fitness function, and a particle swarm algorithm is used to determine the super-parameters of the deep learning network model.

8. An intelligent sensor abnormality recovery device based on an association network, which is characterized by comprising:

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the association network based sensor anomaly intelligent restoration method of any one of claims 1 to 7 when the program is executed.

10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the association network based sensor anomaly intelligent restoration method of any one of claims 1 to 7.